Building evacuation with mobile devices

Merkel, Sabrina

2014

KIT Scientific Publishing

The rapidly growing world population and increasingly dense settlements demand ever-larger and more complex buildings from today's engineers. In comparison to this technological progress, a building's equipment for emergency evacuation has been hardly developed further. This work presents a concept for a building evacuation system based on mobile devices. Furthermore, various algorithms for route planning with mobile devices and for indoor localization of mobile devices are addressed.

Business and Management

English

978-3-7315-0207-4 (print)

Building

SABRINA MERKEL

Evacuation
With Mobile

Devices

Sabrina Merkel
Building Evacuation with Mobile Devices

Building Evacuation
with Mobile Devices
by
Sabrina Merkel

Dissertation, Karlsruher Institut für Technologie (KIT)
Fakultät für Wirtschaftswissenschaften
Tag der mündlichen Prüfung: 28. Februar 2014
Referent: Prof. Dr. H. Schmeck
Korreferent: Prof. Dr. A. Oberweis

Impressum

Karlsruher Institut für Technologie (KIT)
KIT Scientific Publishing
Straße am Forum 2
D-76131 Karlsruhe
KIT Scientific Publishing is a registered trademark of Karlsruhe
Institute of Technology. Reprint using the book cover is not allowed.
www.ksp.kit.edu

This document – excluding the cover – is licensed under the
Creative Commons Attribution-Share Alike 3.0 DE License
(CC BY-SA 3.0 DE): http://creativecommons.org/licenses/by-sa/3.0/de/
The cover page is licensed under the Creative Commons
Attribution-No Derivatives 3.0 DE License (CC BY-ND 3.0 DE):
http://creativecommons.org/licenses/by-nd/3.0/de/

Print on Demand 2014
ISBN 978-3-7315-0207-4
DOI: 10.5445/KSP/1000040428

Building Evacuation with
Mobile Devices
Zur Erlangung des akademischen Grades eines
Doktors der Ingenieurwissenschaften
(Dr.-Ing.)
von der Fakultät für Wirtschaftswissenschaften
des Karlsruher Instituts für Technologie (KIT)
genehmigte
DISSERTATION
von
Dipl.-Inform.Wirt Sabrina Merkel

Tag der mündlichen Prüfung:
Referent:
Korreferent:
2014 Karlsruhe

28. Februar 2014
Prof. Dr. H. Schmeck
Prof. Dr. A. Oberweis

CONTENTS

1 Introduction
1.1 Motivation and Problem Statement
1.2 Objectives and Approach . . . . .
1.3 Major Contributions . . . . . . . .
1.4 Structure . . . . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

2 Research Context and State-of-the-Art
2.1 Mobile Ad Hoc Networks . . . . . . . . . . . . . .
2.2 Organic Computing . . . . . . . . . . . . . . . . .
2.2.1 Organic Observer/Controller Architecture .
2.2.2 Structure of Organic Systems . . . . . . . .
2.2.3 Swarm Intelligence and Nature Inspired Computing . . . . . . . . . . . . . . . . . . . . .
2.3 Evacuation Management . . . . . . . . . . . . . . .
2.3.1 Modeling . . . . . . . . . . . . . . . . . . .
2.3.2 Optimization . . . . . . . . . . . . . . . . .

1
2
4
7
10
13
13
16
18
23
24
26
26
30

Contents

2.4

2.3.3 Evacuation Systems Using Mobile Devices
Localization . . . . . . . . . . . . . . . . . . . . .
2.4.1 Algorithms . . . . . . . . . . . . . . . . .
2.4.2 Distance Estimation . . . . . . . . . . . .
2.4.3 Mobility . . . . . . . . . . . . . . . . . . .
2.4.4 Indoor Localization Systems . . . . . . . .

.
.
.
.
.
.

3 Organic Building Evacuation
3.1 Concept of an Organic Building Evacuation Support
System . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1 Central Control Unit . . . . . . . . . . . . .
3.1.2 Static Sensor Network . . . . . . . . . . . .
3.1.3 Mobile Ad Hoc Network . . . . . . . . . . .
3.2 Observer/Controller Architecture for Mobile Devices
3.2.1 Observer . . . . . . . . . . . . . . . . . . .
3.2.2 Controller . . . . . . . . . . . . . . . . . . .
3.3 Summary . . . . . . . . . . . . . . . . . . . . . . .

37
39
43
54
63
66
71
72
72
74
75
76
79
81
85

4 Swarm Evacuation Planning with Mobile Devices
87
4.1 Macroscopic Swarm Evacuation Planning . . . . . 90
4.1.1 Capacity Constrained Swarm Evacuation Planning . . . . . . . . . . . . . . . . . . . . . . 92
4.1.2 Evaluation . . . . . . . . . . . . . . . . . . 102
4.1.3 Conclusion . . . . . . . . . . . . . . . . . . 125
4.2 Dynamic Multi-objective Swarm Evacuation Planning128
4.2.1 D* Lite . . . . . . . . . . . . . . . . . . . . 131
4.2.2 Dynamic Multi-objective Swarm Evacuation
Planning . . . . . . . . . . . . . . . . . . . 134
4.2.3 Evaluation . . . . . . . . . . . . . . . . . . 140
4.2.4 Conclusion . . . . . . . . . . . . . . . . . . 148
4.3 Summary . . . . . . . . . . . . . . . . . . . . . . . 151

Contents
5 Localization in Mobile Ad Hoc Networks
5.1 Effects of Mobility on Hop Count Based Distance
Estimation . . . . . . . . . . . . . . . . . . . . . .
5.1.1 Mobility Models . . . . . . . . . . . . . . .
5.1.2 Hop Count Error Model . . . . . . . . . . .
5.1.3 Evaluation . . . . . . . . . . . . . . . . . .
5.1.4 Conclusion . . . . . . . . . . . . . . . . . .
5.2 Synchronized Hop Counting in Mobile Ad Hoc Networks . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.1 Pulse-coupled Oscillators . . . . . . . . . .
5.2.2 Firefly-inspired Hop Counting . . . . . . . .
5.2.3 Evaluation . . . . . . . . . . . . . . . . . .
5.2.4 Conclusion . . . . . . . . . . . . . . . . . .
5.3 Geometric Distance Estimation for Mobile Ad Hoc
Networks . . . . . . . . . . . . . . . . . . . . . . .
5.3.1 Connectivity-based Distance Estimation . .
5.3.2 Geometric Distance Estimation . . . . . . .
5.3.3 Evaluation . . . . . . . . . . . . . . . . . .
5.3.4 Conclusion . . . . . . . . . . . . . . . . . .
5.4 Optimization of Beacon Placement in Buildings . .
5.4.1 Optimization of Network Distributions . . .
5.4.2 Evolutionary Algorithm for Optimal Beacon
Placement . . . . . . . . . . . . . . . . . . .
5.4.3 Evaluation . . . . . . . . . . . . . . . . . .
5.4.4 Conclusion . . . . . . . . . . . . . . . . . .
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . .

155
158
159
165
175
188
192
196
198
202
209
212
212
215
222
232
236
237
241
248
256
259

6 Conclusion
265
6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . 265
6.1.1 Organic Building Evacuation Support System266
6.1.2 Swarm Evacuation Planning . . . . . . . . . 270

Contents
6.1.3

6.2

6.3

Range-free Distance Estimation in Mobile
Ad hoc Networks . . . . . . . . . . . . . . . 272
Outlook . . . . . . . . . . . . . . . . . . . . . . . . 277
6.2.1 Organic Building Evacuation Support System277
6.2.2 Swarm Evacuation Planning . . . . . . . . . 280
6.2.3 Range-free Distance Estimation in Mobile
Ad Hoc Networks . . . . . . . . . . . . . . . 281
Final Remarks . . . . . . . . . . . . . . . . . . . . 283

List of Tables

285

List of Figures

287

List of Definitions

293

List of Algorithms

295

List of Abbreviations

297

Bibliography

301

ACKNOWLEDGEMENTS

I would like to express my deepest appreciation to all those who
supported me in completing this thesis. Foremost, I would like to
express my sincere gratitude to my advisor Prof. Dr. Hartmut
Schmeck for providing me with the possibility of working and
pursuing my research at his chair. Thank you for encouraging
me to ask questions and for teaching me that it is not a sign of
ignorance but the first step to improvement and new knowledge.
Apart from my advisor, I would like to thank the rest of my thesis
committee: Prof. Dr. Andreas Oberweis, Prof. Dr. Hansjörg
Fromm, and Prof. Dr. Detlef Seese, for accepting the request to
serve as reviewers in the examination committee and for taking
the time to give insightful comments, to pose hard questions, and
to provide helpful advise.
My sincere thanks also goes to my colleagues and friends of the
research group Efficient Algorithms for the wonderful working
atmosphere, their constant encouragement during the last years,
and many enjoyable lunch and coffee breaks. In particular, I would

i

Acknowledgements
like to thank Christian Gitte, Birger Becker, and Fabian Rigoll
for many inspirational conversations and their profound support
in advancing my research and writing this thesis. Likewise, I am
especially grateful to Ingo Mauser, Sebastian Kochanneck, Dr.
Florian Allerding, and Dr. Pradyumn Shukla for their honest and
encouraging feedback and many helpful suggestions. I am deeply
grateful that I got the chance to share some ideas and time with
all of you.
Moreover, I am grateful to Christoph Brunner and Matthäus
Malkowski for reviewing parts of this thesis, for listening and
discussing many underlying ideas, and for providing good advice
and moral support.
Last but not least, I would like to express special thanks to my
family, my parents Margarete and Alwin Merkel, my sister Doreen
Merkel, and to Matthäus Malkowski for their permanent support
and for always being there for me. You have helped me to stay
calm and positive in stressful times and gave me the courage to
pursue this path until the end. Thank you for always believing in
me.

Karlsruhe, April 2014

ii

Sabrina Merkel

ABSTRACT

A rapidly growing world population and an increasing density of
settlements demand ever-larger and more complex buildings from
today’s engineers. These requirements can often be met due to a
continuous development of new building fabrics and construction
processes. In comparison to this technological progress, the development of novel equipment for emergency evacuation of buildings
has been quite stagnant in recent years. Current evacuation support facilities are mainly limited to stationary exit signage and
emergency maps displaying recommended escape routes. Such
emergency maps can be easily overlooked and are often perceived
as confusing and unclear, especially when someone is in panic. However, the main problem with contemporary building evacuation
support equipment is its inability to adapt recommended escape
routes to the ever changing environment during the evacuation
process. Neither fire outbreaks, smoke formation, blocked passages
due to debris of collapsed masonry can be considered, nor can the
current distribution of people inside the building be taken into

iii

Abstract
account when planning potential escape routes. Nevertheless, these
factors strongly affect potential congestion emergence and, hence,
the evacuation time.
The increasing propagation of mobile devices designed for wireless
communication, such as smart phones, tablet PCs, or general
devices for personal digital assistance, opens up an opportunity to
improve the support of evacuees during an emergency evacuation
of a building. In emergency cases, the mobile device can alert its
user via ringing or vibrating and display an individual escape route
on its screen. By pointing out respective directions, the device
can navigate its user to a safe exit. The advantage compared to
traditional escape route signage lies in the possibility to update the
recommended escape route according to new information about the
current evacuation situation. Such information can be gathered
by the device via wireless communication between devices and
the environment. In this thesis, the Organic Building Evacuation
Support System, or OBESS, is introduced. OBESS is a concept
for an adaptive building evacuation system based on the paradigm
of controlled self-organization from the research area of Organic
Computing. To achieve this characteristic system behavior, an
exemplary implementation of the Observer/Controller Architecture
is presented. The Observer/Controller Architecture is a generic
system architecture developed in the context of Organic Computing,
designed specifically for the purpose of realizing controlled selforganization. The support system consists in part of mobile devices,
which have the ability to establish an ad hoc network via local
communication with each other. This way, local communication
provides a mean for the dissemination of information about the
current evacuation situation in the network. This information can
then be used to find the optimal escape route for the user of each
device. In OBESS, all computations are meant to be performed

iv

Abstract
in a decentralized way, directly on the mobile devices in order to
avoid having a single point of failure in the system. This approach
increases the robustness of the evacuation system, which is a crucial
characteristic, especially for emergency applications.
There are two algorithmic challenges that have to be mastered in
order to realize such a decentralized evacuation support system
based on mobile devices. Firstly, the devices have to compute
escape routes based on uncertain and incomplete information regarding the evacuation situation. Amongst other reasons, this is
partly due to the fact that communication over an ad hoc network
is generally delayed and that the network can get disconnected
from time to time since participating devices move through the
building. So far, the task of decentralized escape route planning
based on uncertain information has received little attention from
researchers. Since it is a key prerequisite for a robust evacuation
support system based on mobile devices, a large part of this thesis
is dedicated to developing solutions for this problem. The second
crucial requirement for the realization of the building evacuation
support system proposed is the ability to determine the locations of
mobile devices in the building. Since signals of the most commonly
used localization technique, the Global Positioning System (GPS),
are usually not available indoors, this task alone can be a challenge.
For static sensor networks, which have similar characteristics to
mobile ad hoc networks, many localization techniques which are
independent from GPS-signals have been proposed over time. Since
an important difference between sensor and ad hoc networks is
the mobility of the devices in the network, this thesis investigates
the effect this mobility has on various distance estimation techniques, which ultimately form the basis for localization algorithms
in OBESS. It is shown that mobility of devices has a significant
impact on the accuracy of computed locations and, therefore, has

v

Abstract
to be dealt with explicitly. For this, several solutions are suggested
and evaluated in this thesis. In addition, this thesis presents an
optimization approach to improve the localization infrastructure
for mobile devices in specific buildings and discusses the potential
of the Observer/Controller Architecture for improving the accuracy
of indoor localization.

vi

CHAPTER

1
INTRODUCTION

A continuously increasing global population and settlement density
combined with huge progress in the building and construction
industry led to the emergence of ever-larger and more complex
buildings. Recently, the world’s tallest construction, the Burj al
Khalifa in Dubai, was brought to completion in October 2009.
This imposing structure consists of 900 apartments, which are
distributed over about 190 floors. Although probably being the
most prominent example of today’s engineering skills, there are
many more to name. The Ghery Buildings in Düsseldorf, Germany,
clearly illustrate the ability and willingness to construct buildings
which diverge from simple geometric forms. Another illustrative
example of complex architectural design is the Habitat 67, a model
community and housing complex in Montreal, Canada, where
houses are assembled in an irregular structure. In contrast to
such developments, most buildings’ precautions for emergency
evacuation are comparatively underdeveloped. Hence, they fail to

1

Chapter 1 Introduction
profit from technological progress, which has found its way into
most other areas of everyday life. The fact that an infrastructure
whose main purpose it is to save human life does not keep up with
technological achievements is surprising, at best, if not worrying.
Therefore, it is an urgent matter to investigate how new scientific
insights and findings, for example in the area of computer science,
can be exploited in order to improve emergency management in
modern buildings.

1.1 Motivation and Problem Statement
Today’s approach to prevent disasters and support a well-ordered
evacuation of buildings is to install safety devices, such as sprinkler
systems, fire or smoke alarms, fire extinguishers, exit signs, and
emergency maps. In general, such emergency maps display the
current floor, recommendations for escape routes, and the locations
of first-aid kits or fire extinguishers. Apart from that, there are
usually stationary signs installed throughout the building, which
are meant to direct people to nearby exits or safe areas. Figure
1.1 displays some examples of such signage. A major drawback of
such kind of emergency facilities is the fact that they are usually
developed when the building is constructed and are permanent
without provision for frequent changes. The route guidance which
is provided by exit signs and emergency maps is designed based on
expectations regarding the evacuation process, which are obtained
from statistical data and expert knowledge. However, the resulting
escape routes can be suboptimal in case the situation in the building
differs from the expected scenario, for example in terms of the
number of evacuees or their distribution across the rooms, et
cetera. Furthermore, the escape routes are not necessarily optimal
for people with special needs due to disabilities or other handicaps.

2

1.1 Motivation and Problem Statement
Fahy and Proulx [62] published reports from survivors of the attacks
of September 11th, 2001, which indicate that the main obstacles
during their escape were congestions in bottleneck-areas and closed
or blocked passages due to smoke formation or damaged building
fabrics. Conventional evacuation signs are unable to adapt to
the specific evacuation situation in the building, to changes in the
environment during an emergency situation, or to personal needs of
specific evacuees. Another serious problem is that navigation signs
are stationary and sometimes easily overlooked in panic situations
(cf. Morishita and Shiraishi [165]). Smoke formation and lack
of illumination can further aggravate this situation. Even people
who are familiar with the building’s layout can have difficulties to
find their way to the exit under such circumstances (cf. Fahy and
Proulx [62]). Furthermore, conventional emergency maps are often
hard to apprehend, especially when in panic. Since they are fixed
to the wall, they cannot be rotated, which is common practice for
some people in order to orient themselves on a map.
For all these reasons, today’s evacuation support equipment in
buildings should be reconsidered and replaced with more innovative
solutions, which can overcome the identified deficiencies. A modern
evacuation system should have digital illuminated screens, which
show navigation instructions towards safe areas or exits of the
building. Ideally, some of these screens are portable devices, which
can be carried by evacuees. This way, frequent searches for new
signs during the evacuation process can be avoided. Evacuation
devices should provide the possibility to assess the current evacuation situation, detect changes in the environment, and adapt the
navigation instructions accordingly. Furthermore, it is desirable
to integrate personal information about the devices’ users in the
process of finding suitable escape routes.

3

Chapter 1 Introduction

EXIT
Figure 1.1: Standard navigation signs for building evacuation.

1.2 Objectives and Approach
The widespread distribution of mobile devices, such as smart phones
or tablet PCs, provides a chance to achieve a changeover in evacuation support for buildings. These devices are portable and equipped
with digital illuminated screens, which can display a building’s layout and navigation instructions to guide their users. Mobile devices
typically possess means for wireless communication via short-range
communication modules, which could provide information about
the current situation in the building. Moreover, the devices are
able to form a so-called Mobile Ad Hoc Network (MANET), or ad
hoc network for short, i.e., a network which consists of dynamically
built-up connections between devices which are located within
sufficiently close proximity. Each device in such a network forwards
messages that it receives to all other devices in its communication
range. This allows for dissemination of information in the network,
which increases the knowledge each device can obtain about the
environment. In addition, mobile devices possess computing capacities, which are needed in order to determine navigation instructions
and to dynamically adapt them according to newly available in-

4

1.2 Objectives and Approach
formation. There is no need for a central computing unit in the
building, which makes the evacuation support system scalable and
robust against a single point of failure. The fact that each evacuee
can have his own mobile evacuation device provides the potential
to personalize evacuation route planning. Individual preferences
and needs, such as for disabled or elderly people, can be taken into
account.
However, the scalability and robustness a system gains from decentralized computation and information acquisition via local communication have their price. The lack of global knowledge poses
a special challenge for evacuation route planning. Firstly, the
available information basis can be incomplete in case the ad hoc
network is disconnected, or if there are no devices located in a
certain area of the building which can report about its state. Secondly, messages received from nearby devices can be obsolete or
outdated because of the delays which arise during the forwarding
process. Furthermore, information exchanged between the devices
could, from the outset, be incorrect either as a result of error-prone
computations or when human users report false information to
their devices, be it accidentally or even on purpose. Although
these are considerable shortcomings, the above described advantages justify a thorough investigation of the potential that mobile
devices hold for the improvement of evacuation support in future
buildings. Since the task of evacuation route planning on mobile
devices differs from standard path-finding problems which require
global and certain knowledge, the first objective of this thesis is:
1. Proposing and evaluating methods to determine optimal evacuation routes on mobile devices which regard user preferences,
take into account uncertain information gathered via local
communication, and are adaptable to detected changes in
the environment.

5

Chapter 1 Introduction
In order to be able to navigate their users, the devices have to be
able to determine their current locations in the building. Since
there is usually no GPS-signal available indoors, the localization
alone becomes an ambitious task. To tackle this problem, the
devices’ ability to organize into ad hoc networks and exchange
messages with other devices in this network comes in handy. While
localization of mobile devices in an ad hoc network has not yet
received much attention, there are a vast number of localization
techniques proposed for a Static Sensor Network (SSN) in the
literature. SSNs have similar characteristics to MANETs but differ
in the fact that the devices are stationary. Mobility of the devices,
however, can fundamentally affect the applicability of localization
algorithms designed for SSNs to MANETs, which leads to the
second objective of this thesis:
2. Investigating the applicability of localization algorithms designed for SSNs to MANETs and developing reasonable adjustments.
Due to the mobility of the devices, their surroundings change
constantly. This is why the evacuation route planning, as well
as the localization process, have to be able to adapt to new and
unforeseen situations. Such a requirement calls for self-organization
amongst the involved devices. Self-organization enables a system
to autonomously operate in unpredictable and interchanging environments. If predicting all potential system states at design time
poses a problem, it is advisable to embrace self-organization as a
design principle. At the same time, self-organization always bears
the risk of undesired system behavior emerging from local interactions between the system’s components due to the lack of global
coordination and supervision. This can be disastrous, especially
when it happens in a life-threatening situation, such as a building

6

1.3 Major Contributions
evacuation. As a consequence, it is desirable to allow for a certain
amount of control in an otherwise autonomous evacuation support
system. Organic Computing (cf. Müller-Schloer and Schmeck
[169], Müller-Schloer et al. [170], Schmeck [213], VDE/ITG/GI
[235]) deals with the design of self-organized technical systems in
order to make them flexible in their reaction to changes in the environment or the system’s objectives. At the same time the system
remains trustworthy and robust with respect to failures and disturbances. The Observer/Controller Architecture (O/C Architecture)
(cf. Branke et al. [23], Müller-Schloer [168], Richter [193], Richter
et al. [195]) is a generic design framework developed to build such
organic systems in a way that they exhibit the previously described
life-like characteristics. Pursuing the goal of a controllable and selforganizing Organic Building Evacuation Support System (OBESS),
the third objective of this thesis is:
3. Developing a concept for a controllable, self-organizing evacuation support system by applying the generic O/C Architecture
to the mobile evacuation devices in the system.
The aforementioned objectives lead directly to the major contributions this thesis provides for the research areas of evacuation route
planning, localization in MANETs, and Organic Computing.

1.3 Major Contributions
The first major contribution relates to the research area of evacuation management. A concept for an evacuation support system in
buildings is presented, which consists of a MANET, an SSN, and
a Central Control Unit (CCU). The presented system emphasizes
decentralized computations and local communications between its

7

Chapter 1 Introduction
components, which makes it scalable and robust against a single
point of failure. In addition, two algorithms for decentralized evacuation route planning on mobile devices are presented, which are
designed to integrate uncertain information about other evacuees’
locations in the building. This information is gathered via local
communication and subsequent dissemination over ad hoc network
connections to other mobile devices in the building. It is shown
that the evacuation process can be accelerated compared to a situation without communication between the devices, even though
the information can be incomplete, outdated, and error-prone due
to the characteristics of the scenario. Both algorithms are designed
in a way that they can adapt to changes in their environment
and one method regards user preferences in the optimization of
escape routes. Since the strongly expanded use of mobile devices
is a rather recent phenomenon, the idea to use mobile devices to
support the evacuation of a building has not yet been subject to
extensive investigations. The proposed methods for distributed
and adaptive escape route planning on mobile devices advance the
research in this area.
The second set of contributions relates to the task of localization,
more precisely, to determine the locations of mobile devices in
an ad hoc network without the use of GPS-receivers. While this
challenge has received much attention for static devices in SSNs,
mobility is rarely addressed. The usual proposition to account for
mobility is to constantly repeat the respective algorithm in order
to find the devices’ new locations when they moved. However,
an extensive study about the impact of mobility on localization
algorithms presented in this thesis shows that simple repetition
leads to an unpredictably high inaccuracy in the resulting locations. Experiments reveal details about the effects various mobility
models have on the quality of localization results and identify main

8

1.3 Major Contributions
influencing factors of these models, such as speed or direction of
movement. A modification of a distance estimation algorithm is
suggested, which is shown to reduce the error produced by mobility
of the devices. Additionally, two indicators are proposed which can
characterize the mobility of devices in a network based on locally
available information. As a consequence, localization results could
be corrected in order to account for the distortion brought by the
specific mobility pattern. Moreover, a distributed, nature-inspired
procedure to synchronize devices in a MANET is successfully tested
and, based on it, a method is developed for distance estimation,
which is shown to improve the mobility induced localization error.
Furthermore, a novel distance estimation approach is introduced
and experiments are presented to demonstrate that by applying
this algorithm the accuracy of localization can be improved for various network topologies and especially for MANETs. The quality
of the results produced by localization algorithms mostly depends
on the placement of specific devices in the network which possess
knowledge of their own positions. This thesis contributes an
Evolutionary Algorithm (EA), which can be used to optimize the
locations of such devices. Valuable insights about an optimal device
placement for localization during a building evacuation process are
obtained by an experimental study. These findings contribute to
both, the research field of evacuation planning and localization for
MANETs.
Moreover, a contribution is made, which relates to evacuation management, localization, and Organic Computing as well. Organic
Computing is concerned with the design of self-organizing technical
systems in order to make them adaptable and trustworthy at the
same time. A generic O/C Architecture was developed for such
systems, which provides for online and offline learning mechanisms,
i.e., learning at runtime and in a simulation, in order to achieve

9

Chapter 1 Introduction
these goals. In this thesis, a conceptual system architecture is
proposed, which applies the generic O/C Architecture from Organic Computing to the mobile devices in the building evacuation
support system. It is shown how the generic architecture can be
implemented for evacuation route planning and localization. It is
further introduced how offline learning can be used to evaluate the
quality of an evacuation instruction with respect to the current
evacuation situation in the building before it is suggested to the
user. Apart from that, the O/C Architecture offers a chance to improve localization. The study of localization algorithms presented
in this thesis reveals that various localization techniques can deliver
high quality results under different environmental circumstances.
When determining to use a specific localization algorithm at design
time, there can be circumstances under which another localization
algorithm could deliver better results. Here, a concept is proposed
which uses online and offline learning mechanisms provided by the
generic O/C Architecture in order to switch between various localization algorithms at run time. Learning mechanisms are applied
to improve this decision process over time. This is a novel approach
to the task of localization, which opens up a new perspective for
research in this area.

1.4 Structure
This thesis is structured as follows. Chapter 2 describes the research context of this thesis. A short introduction of MANETs is
given and the main research topics regarding these networks are
discussed. Subsequently, the research area of Organic Computing
and the related O/C Architecture is described. In addition, an
overview of the concepts of swarm intelligence and nature-inspired
computing is given and their role in organic systems is discussed.

10

1.4 Structure
The chapter proceeds with a description of current research in the
area of evacuation management, including evacuation modeling and
optimization approaches, as well as an overview of state-of-the-art
concepts for evacuation management systems with mobile devices.
Subsequently, the task of localization is introduced and various
approaches to the problem are described with focus on localization
of devices in MANETs and SSNs. Chapter 2 concludes with an
overview of state-of-the-art localization systems.
A concept for OBESS is presented in Chapter 3 and the main
elements of this system, as well as their functionalities, are described. An examplary application of the generic O/C Architecture
to mobile evacuation devices follows and the potentials this architecture provides for the improvement of evacuation planning and
localization are discussed.
In Chapter 4, two distributed evacuation planning algorithms are
developed, which can be employed by mobile devices to navigate
users to an exit while taking into account information about other
evacuees derived from local communication. Simulative experiments show that this information can be used to improve the overall
evacuation time compared to route planning without communication, even though the devices have only uncertain information
about their environment.
Chapter 5 investigates various methods to locate mobile devices
in ad hoc networks without the use of GPS-receives. A standard
localization algorithm, which has been proposed for static networks, is stressed under various mobility models and important
influencing factors of the mobility pattern on the localization accuracy are identified. Additionally, various methods to improve
the localization results for MANETs are presented and discussed.
The approaches are investigated thoroughly, compared, and evaluated in simulative experiments. Concluding this chapter, an EA

11

Chapter 1 Introduction
is proposed, which optimizes the placement of specific devices in
order to improve localization accuracy. This approach is tested in a
simulative evacuation scenario and provides valuable insights into
the characteristics of an optimal placement and into the factors
which influence such an optimal solution.
Chapter 6 summarizes the research presented in this thesis and
subjects it to a critical evaluation which leads to the prospects for
future research in order to advance the vision of OBESS.

12

CHAPTER

2
RESEARCH CONTEXT AND
STATE-OF-THE-ART

In this chapter, the research context of the work presented in this
thesis is outlined and fundamental principles are explained. Firstly,
a short introduction to MANETs is given. The chapter proceeds
with a description of the concept of Organic Computing and its
generic O/C Architecture, which forms the basis for the software
architecture of the mobile devices in OBESS. Subsequently, the
state-of-the-art in evacuation management research is described in
detail. The chapter concludes with an introduction to the task of
localization and state-of-the-art localization methods are described.

2.1 Mobile Ad Hoc Networks
MANETs are networks which consist of a number of mobile devices
with the ability to communicate with other nearby devices using

13

Chapter 2 Research Context and State-of-the-Art
a short-range wireless communication module, like infrared, Bluetooth [20], or ZigBee [257] technology. Also, the devices forward
messages that they receive to all other devices in their neighborhood. Hence, the devices are able to spontaneously form networks
which make long-distance communication possible (cf. Ghosekar
et al. [75]). In contrast to SSNs, the devices in a MANET are mobile and not necessarily equipped with sensors. Ad hoc networks
have been subject to research since the late 1960s (cf. Abramson
[1]) and have been widely studied in the literature since. There
is a working group founded by the Internet Engineering Task
Force (IETF) [107] to investigate MANET related issues. Apart
from being mobile, the devices in MANETs are usually assumed to
have limited resources in terms of memory, processing capabilities,
and power supply.
In the work of Nan and Li [172], an overview of the main research
areas for SSNs is given. According to these authors, the main
research areas are resource management, optimization of the devices’ lifetimes and their localization in the network, routing of
information in the network, and optimizing the coverage of the
network. It should be noted that the scientific questions in the
area of SSNs can mostly be transferred directly to MANETs due to
the similarity of SSNs and MANETs. Throughout this thesis, the
localization task is the main focus, however lifetime optimization
and optimal coverage is also touched upon.
When talking about routing in ad hoc networks, there is a variety of protocols proposed, which can mostly be categorized into
distance vector routing or link state routing protocols (cf. Ballew
[14], Tanenbaum [228]). While in distance vector routing each
device communicates all distance information it possesses to its
neighbors, in link state routing only distance information about
the links to direct neighbors is communicated to the whole net-

14

2.1 Mobile Ad Hoc Networks
work. Many implementations of these protocol types are suggested
and each protocol has its advantages and disadvantages, such as
varying packet sizes or the number of messages needed to establish
and maintain routing information. For the research presented in
this thesis, however, there is no need for direct communication
between specific devices because broadcasts are used to distribute
information over the whole network. Thus, routing is paid little
attention to. The communication model assumed for the network
throughout this thesis is similar to the Gossip Protocol (cf. Haas
et al. [81]). The Gossip Protocol is an alternative routing method
to Flooding, which is the most basic way to disseminate messages
in networks where each device simply forwards all information
available to all neighbors. The idea of the Gossip Protocol, also
called Epidemic Routing, is that each device only forwards the
most recent information it receives and, thus, the communication
overhead is reduced.
Another great advantage of decentralized ad hoc networks is that
the information exchange can still function even if certain devices
are broken. In contrast to networks which depend on a central unit
for message routing, which is often the case in wireless local area
networks, communication in ad hoc networks remains intact in case
of individual failures of devices. This lack of a single point of failure
makes the network’s communication robust, even in regions where
there is no existing communication infrastructure or where the
available infrastructure is damaged. This characteristic makes the
operation of such networks attractive in many application scenarios,
such as geographic monitoring, target tracking, or evacuation
management.

15

Chapter 2 Research Context and State-of-the-Art

2.2 Organic Computing
Another research area touched upon in this thesis is Organic Computing. The Organic Computing initiative started with a position
paper from the German Informatics Society (GI)1 and the Information Technology Society (ITG)2 in 2002 and was established as
priority program 1183 by the German Research Foundation (DFG)3
in 2004 [214]. The fundamental observation that led to the Organic
Computing initiative was that future computing systems become
smaller in size and at the same time increase in performance and
quantity. Due to this development, computing systems become
ubiquitous. Sometimes, they form possibly unlimited dynamic networks via local communication, hence, building complex, variable,
and unpredictable structures. Such computing systems will likely
be part of everyday life enhancing the functionality of our houses
(cf. Allerding et al. [6], Bing et al. [19]), offices (cf. Davidsson and
Boman [49]), cars (cf. Corona and De Schutter [45], Srovnal et al.
[222]), manufacturing industry (cf. Liana et al. [135]), health-care
facilities (cf. de Ruyter and Pelgrim [51], Soar et al. [220]), et
cetera. Because of the dynamics of the environment such systems
are deployed in, they are required to be flexible and adaptive to
unforeseen situations. Since all potential states of the environment or the system itself cannot be accounted for at the time of
design, some degree of self-organization is necessary to cope with
the dynamics, complexity, and unpredictability of such computing
systems. While there are different definitions for self-organizing
systems, “...in the most general way the essence of self-organization
is that system structure appears without explicit pressure or in1

Gesellschaft für Informatik e.V.
Informationstechnische Gesellschaft
3
Deutsche Forschungsgemeinschaft
2

16

2.2 Organic Computing
volvement from outside the system” Schmeck [212]. When external
involvement is excluded, the desired behavior has to be established
exclusively by interactions between the system’s components. A
system which adapts itself to a changing environment possesses
almost life-like properties, such as self-protection or self-healing.
This is the reason for the choice of name.
When a system acts completely self-organized, the question arises
whether it is possible to trust such a system and how it can be
guaranteed that no unexpected and possibly harmful system behavior arises. Organic Computing addresses the tradeoff between
the desire to have an adaptive, flexible, and self-organized system
which is also trustworthy and robust. While an organic system (in
the sense of following Organic Computing principles) is encouraged
to solve tasks in a self-organized way using feedback loops and
learning, the possibility for a user to intervene in the case of an
undesired emergent behavior is given to ensure correct system
behavior. Another requirement of future computing systems identified and addressed by Organic Computing is the need to design
computing systems according to human needs. This postulation
refers to the need for user-friendly interfaces. Because of the complexity such systems exhibit, it is hard for a user to tell the system
exactly what it has to do. A much more convenient and realistic
way for users to communicate with such a system is by formulating
objectives or goals instead of precise instructions.
So far, Organic Computing is used for the design of various system
types, for example the control of traffic (cf. Fekete et al. [65]),
self-organization of traffic lights (cf. Prothmann et al. [185]), design
of a robot control architecture (cf. Brockmann et al. [25]), artificial
vision (cf. Walther and Würtz [241]), energy management in smart
homes (cf. Allerding et al. [6]), and many more (cf. Müller-Schloer

17

Chapter 2 Research Context and State-of-the-Art
et al. [170]). The application to a building evacuation scenario,
however, is novel.

2.2.1 Organic Observer/Controller Architecture
One outcome of the Organic Computing initiative is a generic
O/C Architecture. It allows for controlled self-organization in distributed technical systems (cf. Branke et al. [23], Müller-Schloer
[168], Richter et al. [195]). The architecture provides means for observing, analyzing, and characterizing the current state of a System
under Observation and Control (SuOC), as well as the ability to
predict its future behavior. This information is subsequently interpreted by the controller in order to direct the system into a desired
system state and prevent unwanted emergent behavior (cf. Mnif
et al. [163], Ribock et al. [192], Richter and Mnif [194]). While this
all happens without external control, the possibility for a human
user to control the system still is provided. User input can be made
in terms of objectives and goals. Apart from the O/C Architecture,
similar concepts are presented in other scientific disciplines like
mechanical engineering (cf. Hestermeyer et al. [93]) or autonomic
computing (cf. Kephart and Chess [121]). However, these architectures do not cover all aspects of an organic system. The main,
but not only, difference is that the generic O/C Architecture emphasizes the role of an external entity to specify system-objectives.
In the next section, the generic O/C Architecture is described, as
it serves as a basis for OBESS. For a detailed introduction and
distinction between the generic O/C Architecture and similar concepts, it is referred to the respective literature by Müller-Schloer
[168], Richter et al. [195], and Branke et al. [23].
The generic O/C Architecture consists of four elements, which interact with each other: The SuOC, the observer, the controller,

18

2.2 Organic Computing
and a higher level entity specifying the system-objectives, which are
usually provided by a human user but, in theory, could also be provided by another computing system. Further on, this higher level
entity is referred to as “user”. Figure 2.1 shows the basic elements
of the architecture and their relationships. These components and
their interactions are explained in detail in the following.

Figure 2.1: The main elements in the generic O/C Architecture
and their interactions.

System under Observation and Control
The SuOC can be a set of interacting entities, as illustrated in
Figure 2.1, or a single device. The behavior of this system is
subject to the influence of the controller and is desired to satisfy
the objectives given by the user. Still, it does not depend on the
existence of neither the observer, nor the controller, nor the user.
The system gets input from its environment, i.e., parameters which
can influence the behavior of the SuOC but are uncontrollable
by the system. Furthermore, there is output from the SuOC,

19

Chapter 2 Research Context and State-of-the-Art
i.e., changes to those parameters in the environment that are
visible to external entities outside the SuOC. In addition to input
and output, the SuOC has internal parameters, i.e., parameters
changeable by the system in order to influence its own behavior.
These parameters are only accessible from the outside if specified
explicitly as configuration parameters of the SuOC. An example of
such a system could be those components of a car which are meant
to collaboratively ensure a certain driving speed. The desired
driving speed would be the objective given by the user. Input
parameters could be outside temperature, the incline of the road,
et cetera, and the output parameter is the driving speed. Internal
parameters can, for example, be a certain motor configuration,
which is regulated by the software of the car, but cannot be set
explicitly by the user.
Observer
The system is observed, analyzed, and its current state is characterized by the observer. Furthermore, the observer is responsible
for predicting the most likely future state of the system. Figure
2.2 shows the components of the observer and their interdependencies in the generic O/C Architecture. The observation model
is selected by the controller and it determines how the other components in the observer work depending on the current objectives.
It can, for example, decide which parameters of the SuOC should
be monitored or what their sampling rate is. The monitor collects
the specified parameter values from the SuOC according to the
defined sampling rate and stores these values in log files. The preprocessor takes this data and prepares it for the data analyzer and
the predictor module, for example by filtering or aggregation. The
preparations needed are also specified by the observation model. In
the data analyzer, meaningful attributes that are able to character-

20

2.2 Organic Computing
ize the current system state, e.g., stochastic values are calculated
from the pre-processed data. The observation model determines
which values are computed. The predictor has some prediction
methods at its disposal and selects a suitable one according to the
observation model. With the selected prediction model and data
from the pre-processor it forecasts the future system state. This is
an important step because it increases the chance for the controller
to prevent undesired system behavior before it actually arises. The
output from pre-processor, data analyzer, and predictor is collected
in the aggregator-module and forwarded to the controller.

Figure 2.2: The observer of the generic O/C Architecture and its
components.

21

Chapter 2 Research Context and State-of-the-Art
Controller
The main purpose of the controller in the O/C Architecture is to
select control actions, i.e., actions that influence the environment of
the SuOC, the communication between its elements, or the behavior
of the elements directly. Figure 2.3 shows the components of the
controller and their relations. The selection of an appropriate action
for a certain situation is done in the action selector. The mapping
from system state to appropriate action is subject to a learning
process (cf. Fredivianus et al. [70], Richter [193], Richter and Mnif
[194] for detailed descriptions). This process can be distinguished
into online and offline learning. Both learning mechanisms take
place in different parts of the controller. During online learning,
actions are executed which influence the SuOC. The quality of
these actions is deduced from the system’s response. For this,
the controller keeps track of the previously applied actions and
the subsequent system states in two history-stacks, namely action
history and situation history. From this information, the impact
of an action is evaluated with respect to the given objectives
and a so-called fitness value is assigned to the situation-action
pair. The result of the evaluation is reported to the adaptation
module, which alters the fitness for the corresponding mapping in
the action selector, giving this mapping a higher or lower chance,
respectively, to be selected again for execution. Simultaneously
to this online learning process, the O/C Architecture provides the
possibility for offline learning. The offline learning is realized in a
simulative model of the SuOC. The adaptation module creates new
mapping rules, for example by applying mutation and reproduction
mechanisms of EAs (cf. Weicker [246] for details), which can be
evaluated in the simulative environment using standard fitness
evaluation and selection mechanisms from EAs. Thus, only rules
that perform well in the simulation are committed to the action

22

2.2 Organic Computing
selector and available for online learning. Although simulations are
never exact reproductions of the real environment and, therefore,
cannot guarantee the impact an action has, the offline learning still
provides certain quality assurance.

Figure 2.3: The controller of the generic O/C Architecture and its
components.

2.2.2 Structure of Organic Systems
The generic O/C Architecture in the previously presented form
consists of a central observer and controller which regulate a dis-

23

Chapter 2 Research Context and State-of-the-Art
tributed system, but there are other structures as well. For example,
a completely distributed O/C Architecture, where every element of
a distributed system is an SuOC and has its own observer and controller. Also, a multi-level O/C Architecture is conceivable, where
the distributed O/C Architecture has an additional central toplevel observer and controller layer. Figure 2.4 shows these system
structures. For the organic evacuation support system, each mobile
evacuation device is designed according to the O/C Architecture.
In addition, the system could benefit from a centralized O/C unit,
such that the overall system architecture resembles the hierarchical
system architecture illustrated in Figure 2.4(c). It is important to
note that the mobile devices in a hierarchical system are perfectly
functioning, even if the superordinate O/C Architecture breaks
down. This is important in the context of emergency situations,
in which a single point of failure is to be avoided. However, an
additional central unit allows for users to communicate with the
distributed system via a unique interface, which significantly facilitates the control of such a distributed system.

2.2.3 Swarm Intelligence and Nature Inspired
Computing
Although the notion organic in Organic Computing does not imply
the use of nature-inspired algorithms, these methods often fit well
into the concept of Organic Computing. Nature-inspired methods
transfer observed behavior in natural social systems, e.g., swarms of
insects, into principles for the design of artificial technical systems.
Creatures, especially when living in swarms, tend to self-organize
in order to solve complex tasks. Prominent examples of swarms
in nature who serve as algorithmic models are ant colonies (cf.
Dorigo [56], Dorigo et al. [57]), bee swarms (cf. Karaboga and

24

2.2 Organic Computing

(a)

(b)

(c)

Figure 2.4: Different structures of O/C Architectures. The
architecture of a centralized system (a), a
decentralized system (b), and a hierarchical system
(c).

25

Chapter 2 Research Context and State-of-the-Art
Akay [116]), or fish schools (Neshat et al. [175]). The main characteristic of swarm intelligence is that a global behavior of the
entire system arises as an emergent effect from simple local interactions between its components. This emergent global behavior,
however, is unknown to the single entities (cf. Eberhart et al. [58]).
Emergence denotes the phenomenon where local interactions lead
to new global properties or structures which only arise from the
interactions between its parts and would not appear with only few
or a single individual. In the words of Aristotle: “The whole is more
than the sum of its parts!”. For OBESS, many nature-inspired
algorithms have been investigated and proposed since they are
very suitable for such distributed, decentralized systems. Also, due
to the local communication between the devices, concepts from
swarm intelligence can be transferred to mobile evacuation devices.

2.3 Evacuation Management
When it comes to evacuation management, there are generally
two distinct topics which are addressed by researchers. One research subject is evacuation modeling, which has the objective to
describe an evacuation process. The other research area concerns
the optimization of an evacuation process by systematic intervention. In the following, these two aspects of evacuation research are
described in detail.

2.3.1 Modeling
Figure 2.5 displays an overview of the research in evacuation modeling. According to Hamacher and Tjandra [82] and Schadschneider
et al. [210], evacuation models can be distinguished into microscopic
and macroscopic models. The objective of microscopic models is to

26

2.3 Evacuation Management
describe an evacuation process as realistically as possible. Hence,
the focus lies on modeling the heterogeneous character of individuals and their interactions. Microscopic models often form the basis
for evacuation simulations. Their main purpose is the assessment
of evacuation processes in order to either identify potential weak
points or to evaluate improvement measures. Macroscopic models,
on the other hand, provide the basis for optimization of route
choices for evacuees. These models capture the evacuation situation from a global perspective and describe aggregated evacuation
flows rather than interactions between individual evacuees or their
characteristics.

Figure 2.5: Overview of current research in the area of evacuation
modeling.

27

Chapter 2 Research Context and State-of-the-Art
Microscopic Models
Recorded data or systematic studies about the behavior of people
in emergency situations are rare (cf. Brown [26], Yang et al. [252]).
Therefore, microscopic evacuation modeling is not straightforward
and receives much attention from the research community. Microscopic models describe the behavior of a number of, possibly
different, evacuees and their interactions during an emergency situation. In such models, each evacuee is treated as a single agent with
special characteristics, e.g., speed, risk aversion, size, et cetera.
Microscopic models often serve as a basis for investigating the
behavior of people with different characteristics during evacuation
in a certain environment.
There are various microscopic models, which can be characterized
into discrete and continuous models. Examples of discrete models
are models which are based on cellular automaton theory (cf.
Burstedde et al. [32], Gipps and Marksjö [76], Minoru Fukui [160],
Schadschneider [209]) or lattice-gas models (cf. Guo and Huang
[79], Song et al. [221]). The evacuation environment is divided into
a grid or hexagonal patches and time is modeled in discrete steps, in
which the evacuees move from one patch to another according to the
rules defined by the model. Continuous models comprise models,
such as the social force model (cf. Helbing and Molnar [89], Helbing
et al. [90]), in which the movements are determined by forces that
act on the evacuees, or dynamic fluid models or gas-kinetic models
(cf. Helbing [88], Henderson [91, 92]), where movements follow the
physical principle which describes the dispersion of fluids or gas.
Additionally, there are so-called agent models (cf. Epstein et al.
[59], Xiong et al. [251]), which allow for a distinct modeling of
evacuees, such that two different evacuees can behave differently in
the same situation (cf. Dawei et al. [50], Lo et al. [141]). Moreover,
there are models of evacuee behavior based on game theory (cf.

28

2.3 Evacuation Management
Dawei et al. [50], Lo et al. [141]) or models inspired by nature, such
as the model based on particle swarm optimization from Izquierdo
et al. [109]. For an overview refer to Schadschneider et al. [210] and
Zheng et al. [256]. Microscopic models usually form the basis for
the simulation of an emergency situation, which can then be used
to evaluate the evacuation process and improvement strategies.
Many evacuation simulations have been developed over time. For
an overview it is referred to the website of the evacuation modelling
community [164].
Macroscopic Models
A macroscopic evacuation model is a graph representation G =
(N, E) of the evacuation environment (cf. Chalmet et al. [36]).
The nodes n ∈ N represent regions of the modeled environment
in which people can be located (e.g., rooms, squares in a city, et
cetera) and have certain capacities in terms of available space. An
edge e ∈ E between two nodes in the graph represents a path
between the two regions described by these nodes. The edges are
weighted with traveling costs, such as the distance between the
connected nodes or similar properties which are decisive for route
choices. The main purpose of macroscopic models is to find optimal
paths for evacuees towards the building’s exits. An example of
a building layout and its corresponding macroscopic evacuation
graph model is shown in Figure 2.6.
Some optimization approaches use a time-expansion of such a
macroscopic graph as a basis. A time-expansion of a graph introduces multiple copies of each original node from the evacuation
graph model. Each expanded node represents the state of the
original node at a specific time step. Edges in the time-expanded
graph connect nodes which can be reached in one time step, which
implies that the travel time for each edge in the time-expanded

29

Chapter 2 Research Context and State-of-the-Art
graph has the value 1. Capacities are adopted from the original
graph. Figure 2.7 displays the time-expansion for the macroscopic
evacuation graph model shown in Figure 2.6. For better readability,
node capacities, initial allocations, and the travel time for edges
are removed from the graph.

(a) Building layout

(b) Evacuation graph

Figure 2.6: Example of a building layout (a), a corresponding
macroscopic evacuation graph model (b).

2.3.2 Optimization
Evacuation optimization refers to approaches which are meant to
improve the evacuation process. Usually, evacuation optimization
has the objective to accelerate evacuation, but there are other
objectives which can be considered, for example reducing the
number of injuries. In general, optimization can happen in two
ways, either indirectly by influencing the evacuation environment
or directly by improving the route choice of evacuees. Figure 2.8

30

2.3 Evacuation Management

(a) Evacuation graph

(b) Time-expanded graph

Figure 2.7: Time-expansion of a macroscopic evacuation graph
model.
displays an overview of the research in evacuation optimization
presented hereafter.
The optimization of environmental factors ranges from finding an
optimal room structure (cf. Kellenberger and Müller [120], Swenne
and Bäck [225]) to defining the optimal number of doors, the
optimal placement of doors, or their optimal width (cf. Alizadeh [5],
Muhdi et al. [167], Varas et al. [234]). Furthermore, the deliberate
placement of barriers, such as pillars, is subject to investigation
in order to improve the evacuation flow (cf. Frank and Dorso [69],
Johansson and Helbing [113]). Indirect optimization is generally
evaluated via evacuation simulations based on microscopic models.
Apart from environmental factors, the optimization of route choices
is a well-studied topic. Route optimization is usually performed on
the basis of a macroscopic graph representation of the environment.

31

Chapter 2 Research Context and State-of-the-Art

Figure 2.8: Overview of current research in the area of evacuation
optimization.

32

2.3 Evacuation Management
Such a graph model can be used to perform either flow optimization
or path planning. Flow optimization calculates optimal evacuation
routes for sets of evacuees located at the same node in the graph.
Such sets of evacuees are denoted as flows and the corresponding
optimization problem reflects a network flow optimization problem. If the optimization objective is to maximize the number of
evacuated individuals in a certain time period, the corresponding
network flow problem is called Maximum Dynamic Flow (MDF)
problem (cf. Ford and Fulkerson [68]). Maximizing the number
of evacuated individuals in each time step is denoted as Universal
Maximum Flow (UMF) problem (cf. Gale [71], Jarvis and Ratliff
[111]). An optimal solution for a UMF problem corresponds to an
optimal solution for an MDF problem, not only for the time period
considered, but also for any smaller time horizon (cf. Jarvis and
Ratliff [111]). Minimizing the time needed for a certain number of
evacuees to reach the exit is referred to as the Quickest Flow (QF)
problem (cf. Burkard et al. [31]). Since capacities of edges in
the evacuation graph change during the evacuation process, these
constraints are not properly reflected by a static network model as
depicted in Figure 2.6. Therefore, flow optimization is generally
based on a time-expansion of the evacuation graph model (cf. Section 2.3.1). It should be noted, though, that time-expanded graphs
have several drawbacks. The most obvious one is that they can
get very large with increasing number of time steps. Even more
problematic is that the total time of evacuation has to be estimated
beforehand in order to know how many expansions are needed. In
the work of Lu et al. [143], a heuristic approach to solve the QF
problem is presented, which does not require a time-expansion of
the evacuation graph model. Instead, capacity reservations are
modeled as a time series for each node in the graph. A routing
protocol is adapted to schedule evacuees on paths towards the exit.

33

Chapter 2 Research Context and State-of-the-Art
The scheduling is performed sequentially starting with evacuees
closest to the building’s exit. The occupied capacities are stored in
the time series and considered when the next flow is scheduled until
all evacuees left the building. The sequential scheduling reduces the
necessary computations while still providing evacuation planning
of similar quality as the solutions based on time-expanded graphs
(cf. Sangho et al. [204]).
In contrast to flow optimization, path planning searches for an
optimal route choice for a single evacuee instead of optimizing
the flow in a network. Since only one evacuee is regarded in path
planning, capacities can be neglected in the evacuation graph model
and the optimization can be performed on a graph without timeexpansion. Similar to flow optimization, path planning can have
various objectives. If travel distance on the route is to be minimized,
the problem is called a shortest path problem. If the optimization
objective is to minimize travel time, the problem is called the
quickest path problem. Path planning can be generalized into a
minimum cost path problem, where costs represent any desired
optimization objective. In case multiple objectives are considered
at the same time, edge costs can be a weighted sum of multiple
cost-components each representing one objective.
When global information about the costs of all edges in the graph
is available, path planning can be solved deterministically using
the algorithm described by Dijkstra [53] or the A* algorithm of
Hart et al. [84]. Dijkstra algorithm computes the lowest cost path
between all nodes in a graph and a target node, whereas the A*
algorithm concentrates the search on finding the best path starting
at a specific node in the graph, which reduces the computational
costs. The problem can also be solved by heuristic approaches.
One of the most prominent heuristic method is the Ant Colony
Optimization (ACO) from Dorigo et al. [57], a nature-inspired

34

2.3 Evacuation Management
method based on the behavior of ants. Another heuristic approach
to solve this problem is to use an EA, which are derived from
the concept of natural evolution (cf. Weicker [246] for details).
For path planning in an evacuation scenario, two approaches to
find the lowest cost path according to multiple objectives are
introduced and examined by Cheng [39] and Zong et al. [258].
The edge costs contain risk and travel time and ACO is used to
find the optimal path through the building. A similar scenario is
subject to investigation in Garrett et al. [72] and Saadatseresht
et al. [202], where an EA is applied to solve the path finding
problem for multiple evacuation objectives. It should be noted
that the quickest path problem is also known and addressed in
the evacuation modeling domain. Here, it is solved to model
pedestrians which take quicker paths in order to avoid crowded
areas which results in a more realistic walking behavior (cf. for
example Kretz [126]).
In case of variable edge costs over time, a naive approach would
be to compute the new shortest path from scratch whenever costs
have changed. The D* algorithm by Jarvis and Ratliff [111] avoids
recomputation from scratch and reuses information from previous
computations, which reduces the necessary computations significantly compared to total replanning from the beginning. The D*
algorithm is an incremental and heuristic search approach that
performs local consistency checks before edge costs are newly computed. The process starts at the edge with detected change in costs
and continues from there until there are only unaffected nodes left
or the starting node of the search is reached. This limits necessary
computations to nodes which are affected by the changed edge
costs. This algorithm was further developed to result in D* Lite
by König and Likhachev [125]. D* Lite is a less complex version of
the D* algorithm in terms of its structure. In addition, the number

35

Chapter 2 Research Context and State-of-the-Art
of computations which are required to find the new best path are
shown to be equal or less compared to the D* algorithm.
There are scenarios where path planning is performed in a distributed system. In such cases, each system component has only
a limited view on the graph, for example local information about
some edge costs. The complete optimal path is then built incrementally on a next-hop basis by assembling locally optimal subparts.
An application of distributed path planning based on local information to avoid traffic jams is presented by Prothmann et al. [186].
Here, traffic lights are equipped with means to observe the traffic
flow at intersections. This information is exchanged with other
nearby traffic lights and, similar to the distance vector or the link
state routing protocol (cf. Tanenbaum [228]), the accumulated
information is used to send waiting cars to the allegedly next best
intersection with respect to the cars’ destinations. A similar approach is presented by Filippoupolitis et al. [66] for evacuation
navigation. Here, sensors observe danger indicators, such as smoke
or heat, forward this message to stationary decision nodes, which
exchange and collect information and decide based on these local
observations to which neighboring decision node passing evacuees
are sent. These methods rely on a fixed infrastructure which constantly observes a locally limited, but fixed part of the building.
From these observations, information about arising congestions or
risk values can be derived easily.
When mobile devices have to be used for observation, however, the
observable area changes over time and the estimation of waiting
times, as well as the prediction of emerging congestions, is not
straightforward anymore. In the work of Wagoum et al. [237], a
method is presented, which is similar to distributed path planning
with mobile observers. It introduces a microscopic evacuation
model that integrates local path planning with the aim of modeling

36

2.3 Evacuation Management
a more realistic pedestrian walking behavior. A pedestrian which is
slowed down by congestion on its current path is modeled to observe
other pedestrians which are located in the same room and aim for a
different exit. The pedestrian changes his route in case the evacuees
which he observes progress faster and he starts following them.
The presented method to model pedestrian behavior during an
evacuation has similarities to route choice optimization approaches.
However, this work belongs to the domain of evacuation modeling,
since the goal is to achieve a more realistic walking behavior of
evacuees, not to find the optimal paths. Hence, the chosen escape
routes in this model can differ strongly from paths which result from
route choice optimization approaches. The problem of distributed
path planning on the basis of information collected by mobile
observers is, among others, object of this thesis.

2.3.3 Evacuation Systems Using Mobile Devices
There are several projects investigating the usage of mobile devices
in outdoor evacuation scenarios. For example, the project “REPKA:
Regional Evacuation: Planning, Control, and Adaptation” of the
German Federal Ministry for Education and Research (BMBF)4 [2].
In REPKA, mobile devices are used to build a network and guide
users to safe areas. Similar projects about the coordination of rescue
forces or evacuees with the help of mobile devices are described in
publications from Jang et al. [110], Lien et al. [138], Schau et al.
[211], and Rodriguez et al. [199]. However, little has been done, so
far, to investigate the potential of mobile devices for supporting
building evacuation. An obvious reason for this is the challenge to
localize devices indoors without being able to rely on GPS, which

4

Bundesministerium für Bildung und Forschung

37

Chapter 2 Research Context and State-of-the-Art
is the preferred localization technique for outdoor deployment of
mobile devices in emergency management.
Nevertheless, first ideas about the application of mobile devices to
support people during an emergency situation in a building are
presented by Szwedko et al. [227] and Filippoupolitis et al. [66]. In
the work from Szwedko et al. [227], reducing waiting times that
occur due to overcrowding is addressed. The authors propose an
evacuation system, where users carry mobile devices and scan QRCodes or RFID chips in order to send their current location to a
central server. This server runs a so-called Scavy algorithm, which
computes the next destination for the mobile device’s user, thereby,
trying to achieve a load balancing in the building. Although this
system is shown to prevent overcrowding at specific destinations
in the building, it has several drawbacks. Firstly, it takes time
to read a QR-Code or scan an RFID chip, which is hindering to
people leaving the building as quickly as possible in case of an
emergency. Additionally, it requires physical proximity between the
user’s device and the QR-Code or scanner, which is hardly possible
for several users at the same time and, additionally, bares the
risk of overlooking the respective terminals or codes. In addition,
the algorithm used for load balancing is executed on a central
computing system which collects all information in order to get a
global view. This server poses a potential single point of failure
which can be disastrous in case it breaks down during an emergency.
In terms of its structure, the evacuation system proposed by Filippoupolitis et al. [66] is similar to OBESS presented in this thesis.
It consists of an SSN and a dynamic network of mobile devices
which are carried by people and guide their users to nearby exits.
The static sensors collect information about risk values in the environment, such as smoke or heat. The sensors are connected to each
other and exchange the measured risk values with their communi-

38

2.4 Localization
cation neighbors. When evacuees pass by such a sensor, they are
sent to the next sensor which is closest to the building’s exit and reports the lowest risk value. In contrast to OBESS, the system lacks
the possibility to incorporate user input such as preferences for
path planning. Secondly, there is no localization method provided
that supports navigation. Thirdly, the evacuation planning does
not take into account the current distribution of evacuees in the
building or potential congestions resulting from this distribution.
Using GPS-less localization in MANETs for building evacuation,
is suggested in Inoue et al. [106]. The research presented, however,
remains on a conceptual level. In contrast to OBESS, the proposed
system architecture is not designed for controlled self-organization
and it does not account for user preferences or input. Furthermore,
incorporating information about other evacuees’ positions in the
evacuation planning is not considered in this research.

2.4 Localization
The art of localization and navigation is an inevitable prerequisite
for seafaring and, thus, dates back a long time. Since the ancient
days when people have used the positions of stars in the sky for
localization and navigation purposes, a lot of research has been
conducted in this field and many different navigation systems have
been proposed since (cf. Fallah et al. [63], Hightower and Borriello
[94]). Localization denotes the process of determining the location
of a certain subject. When talking about a location, physical and
symbolic locations as well as absolute and relative locations can be
distinguished (cf. Hightower and Borriello [94]). While a physical
location is described by precise coordinates, e.g., longitude and
latitude, a symbolic location is more an abstract place, such as “in
Germany”, “at the University”, “next to a street lamp”, or the like.

39

Chapter 2 Research Context and State-of-the-Art
Kröller et al. [127], for example, propose to determine symbolic
locations in an SSN using cluster algorithms. A localization system
can be used to provide both kinds of information, nevertheless, it
is usually built to determine a physical location, which can then be
mapped into a symbolic location using additional knowledge, for
example a map of the environment. It holds for any type of location
that the notion of location always needs a frame of reference to
be meaningful. In other words, “all positions are relative. [...]
assigning a set of coordinates to an object is meaningless without
knowledge of the coordinate system with which those coordinates
are associated” Savarese et al. [205]. These considerations lead
to the difference between absolute and relative locations. While
absolute locations share a common reference grid, relative locations
do not. The relative location of an object makes a statement
about the object’s location with respect to a certain reference
point. Knowing the relative location of an object with respect
to a reference point, however, does not provide any information
about the same object’s relative location with respect to a different
reference point. While any absolute location can be transferred
to a relative one, as soon as the absolute location of the reference
point is given, it does not work the other way around. Figure 2.9
illustrates the difference between absolute and relative locations.
On the left-hand side, the absolute locations of five different devices
are depicted, on the right hand side, a sample result of relative
localization is shown. While the devices still have the same distance
and orientation with respect to each other, the computed locations
are different from their actual positions, namely mirror-inverted and
rotated compared to the absolute locations. Prominent examples of
common reference grids are geographic coordinate systems, which
humans have globally agreed upon in order to exchange location
information. Location information which refers to such a common

40

2.4 Localization
reference grid is, for example described in terms of longitude and
latitude. Because of the common understanding about the reference
frame, everybody can point to exactly the same place on earth,
when given such a coordinate. For a localization system to be
as flexible as possible, absolute physical locations are the desired
output. These can be mapped to symbolic or relative locations if
needed.

Figure 2.9: Comparison of exemplary results after applying
absolute and relative localization methods in ad hoc
networks. By using relative localization, the
relationships between the devices stay intact, but the
locations can be mirror-inverted or rotated when
compared to the absolute locations.
The ability to determine the locations of devices plays an important
role in SSNs. Examples are the localization of event reporting in
a monitoring SSN (cf. Szewczyk et al. [226], Werner-Allen et al.
[248]), location dependent routing (cf. Liao et al. [136], Maihofer
[145]) or the assistance of group querying (cf. Gehrke and Madden
[73]) in order to save energy, security enhancements to prevent
wormhole attacks (cf. Hu et al. [98], Karlof and Wagner [117]),

41

Chapter 2 Research Context and State-of-the-Art
and many more. Therefore, many localization methods for SSNs
are proposed in the literature. SSNs differ from MANETs in the
fact that the devices are stationary. However, most algorithms
developed for SSNs are transferable to MANETs by constantly
repeating the localization process in order to account for the devices’
new locations. The simple application of algorithms which are
developed for SSNs to MANETs is a viable approach. Neglecting
the mobility aspect of the network, though, can have a major
negative impact on localization results as demonstrated by a study
presented in this thesis (cf. Section 5.1). Despite this drawback,
this aspect has received surprisingly little attention by the research
community up to this point.

Figure 2.10: Classification of localization algorithms.

42

2.4 Localization

2.4.1 Algorithms
Localization algorithms can be classified according to the scheme
shown in Figure 2.10. On the highest level, methods can be
divided into centralized and decentralized approaches, depending
on whether the location calculations are carried out on a central
computing unit or in a decentralized way directly on the devices
to be located. It should be noted, though, that decentralized
algorithms can also be executed in a centralized way by sending
all necessary information from the devices which are object of
localization to a central server where all locations are computed.
In this case, the disadvantages of centralized algorithms apply
as well. Centralized algorithms, on the other hand, cannot be
executed in a decentralized way because they require input data
from numerous devices in order to compute their locations.
Centralized Algorithms
Two well-studied centralized algorithmic concepts for localization
are Semi-definite Programming (SDP) presented by Doherty et al.
[55] and Multidimensional Scaling Map (MDS-Map) from Shang
et al. [215]. In SDP, geometric constraints, such as distances or
angles between devices, are estimated by the devices in the network
and sent to a central server. On the central server, the constraints
are collected and represented as linear matrix inequalities, which
are then solved in order to find locations for the devices in a
way that the constraints are fulfilled. In MDS-Map, a technique
from mathematical psychology is used (cf. Bachrach and Taylor
[12]). The algorithm constructs relative positions based on distance
estimates and applies concepts of linear algebra to calculate the
coordinates. The main advantage of centralized algorithms is that
information from all devices in the network can be used to compute

43

Chapter 2 Research Context and State-of-the-Art
locations which often results in a higher precision compared to
approaches which use only local information. On the other hand,
a major drawback is that they primarily stress devices close to
the central station during information collection because these
devices have to forward information from all other devices in the
network (cf. Bachrach and Taylor [12]). Moreover, when central
computation is used, the computation is usually a time consuming
process because a lot of information has to be forwarded through
the network, processed on the central server, and possibly be sent
back to the devices in case they require the location information.
This can pose problems, especially if the devices to be located are
mobile, since the devices may have already changed their locations
by the time the computation is completed. Apart from that, there
is a lot of communication involved in the process since all devices
have to exchange information with the central computing unit.
These factors make centralized algorithms hardly scalable and
unsuitable for dynamic networks.
Scene analysis, also called odometry, is a localization approach
which requires the device which is object of localization to be
able to perform odometry readings. Odometry readings are, for
example, observations of the environment which can be used to draw
conclusions about the current location. Such information can be
obtained via sensors attached to the device, such as infrared sensors
to measure the distance to nearby walls or barriers, photosensors,
which measure the incidence of light, or even cameras, which record
visual data from the device’s surroundings. A set of odometry
readings at a specific location is called fingerprint. This fingerprint
is subsequently matched with entries in a database. This database
consists of locations and corresponding fingerprint entries, which
are measured preliminary for different locations in the application
area of the system. A high storage capacity is required for the

44

2.4 Localization
database and sufficient computing power has to be available to
perform the pattern matching process, especially when cameras
are used for the odometry readings (cf. Hightower and Borriello
[94]). Hence, the database is usually located on a central server and
queried via wireless communication. In theory, it is conceivable to
have a copy of the database locally stored on the device which is
object of localization. If this is possible, the approach belongs to
decentralized localization methods. However, this is usually not
feasible due to the restrictions of the devices to be located. Scene
analysis requires a high configuration effort when collecting the
database entries. Furthermore, the sensor data has to be different
for any of the considered locations in order to distinguish them.
This usually requires more than one type of sensor reading. For
these reasons, this localization approach is rarely applied to SSNs
in which the devices usually have only one sensor at their disposal
and are limited in computing power, memory space, and power
supply. Examples of localization using a scene analysis approach
are described in Rajamäki et al. [188], Retscher [191], and Bahl
and Padmanabhan [13].
Decentralized Algorithms
There are two types of decentralized localization algorithms depending on whether or not they rely on the existence of beacons.
Beacon-based algorithms are required, whenever resulting coordinates have to relate to a common reference frame, i.e., when
absolute locations are desired. In contrast, beacon-free or nonbeacon-based algorithms can only perform relative localization,
which can result in locations being rotated or mirror-inverted with
respect to the actual coordinate system.

45

Chapter 2 Research Context and State-of-the-Art
Beacon-free Algorithms There are two different approaches for
beacon-free localization algorithms which compute physical, relative coordinates. The first concept is called relaxation-based
localization which uses a coarse algorithm to roughly determine
the devices’ locations and then iteratively adjusts each device’s
position in order to minimize some local error metric (cf. Bachrach
and Taylor [12]). An example of such an algorithm is the spring
model presented by Priyantha et al. [184]. The second procedure
is called coordinate system stitching. For this bottom-up approach,
the network is divided into small overlapping subregions. In each
subregion, a local map is created and then adjacent sub-regions
merge their local maps iteratively until a single global map is
formed. Examples of such algorithms can be found in Ji and Zha
[112] and Meertens and Fitzpatrick [149]. Relative localization
has the advantage that there is no need for specially equipped devices, i.e., beacons. In OBESS, however, the resulting coordinates
should not be rotated or mirrored with respect to the building map.
Still, such algorithms can be applied for refinement of coordinates
produced by an absolute localization process.
Beacon-based Algorithms Beacons are devices which are equipped
with similar or equal hardware as the devices to be located but
possess knowledge of their exact positions a priori and can, therefore, support the computation of coordinates for all other devices
in the network. A beacon can have a priori knowledge of its own
location either due to manual configuration, because it possesses a
working GPS-receiver, or because it has already computed its own
location. In the latter case, the device is usually not denoted as a
beacon, but can be used as such in any beacon-based algorithm.
For precise calculation of two-dimensional absolute coordinates, the
appropriate localization algorithm and at least three beacons are

46

2.4 Localization
needed, which are placed in a non-colinear way. For three dimensions, four beacons are needed respectively. In a fully connected
network and in the absence of any distance measurement errors,
these three or four reference points yield a perfect solution and no
improvement is observed from having additional reference points
(cf. Savarese et al. [205]). However, localization hardware produces
noisy measurements due to occlusion, collisions, and multipath
effects (cf. Bachrach and Taylor [12]) and as a result, beacon-based
localization normally uses several beacons to improve localization
results. It is usually aimed at minimizing the number of beacons
in a system because these devices can be expensive either due
to their additional equipment with GPS-receivers, or in terms of
preparation effort, when they have to be configured with their
exact location information. With beacons available, there are three
categories of localization algorithms which can be applied: proximity, angulation, and distance-based localization. These methods
differ in the required hardware, number of beacons, accuracy of
the derived results, and input parameters used for localization.
One group of localization approaches are algorithms which require
knowledge about the proximity of beacons to the device which is
to be located. Proximity knowledge means that a device is able
to sort nearby beacons according to their distance from itself. In
contrast to distance-based localization methods, the exact value of
the distance, however, is not required. A method which is often
used to decide about the proximity of beacons is to compare the
quality of the communication signal received from different beacons
(cf. Bulusu et al. [29]).
The most straightforward proximity based localization approach
is the so-called cellular proximity method, which requires a vast
number of beacons to achieve high accuracy results. In this method,
beacons send a signal and their location information to nearby

47

Chapter 2 Research Context and State-of-the-Art
devices. The device receives these messages and determines its own
location to be at the same location like the beacon which is closest
to the device. The accuracy of the location system increases with
the number of beacons in the system. A cellular proximity based
localization system is, for example, proposed by Want et al. [244].
The diffusion method, sometimes also referred to as centroid localization, requires fewer beacons, but still comparatively more than
angulation or distance-based localization techniques. With this
method, a device derives its location at the center of the locations
of a fixed number of beacons, which are closest to the device. In
Figure 2.11, the localization of a device (grey dot) using diffusion
and four beacons (black dots) is demonstrated. The center of the
four beacons lies at the intersection of the dashed lines which is
where the diffusion algorithm estimates the device’s position. Exemplary implementations for the diffusion approach can be found
in Almuzaini and Gulliver [7] and Kaseva et al. [118]. Meertens
and Fitzpatrick [149] refine this concept further by determining a
device’s location not only in the center of their nearest beacons, but
in the center of all neighboring devices after they have determined
their locations. The appealing advantage of this algorithm is its
simplicity since there is no estimate for the distance or angle information needed. Nevertheless, it is not always guaranteed that there
are enough beacons to achieve the desired accuracy in localization.
A more sophisticated proximity-based approach called Approximated Point-in-Triangle Test (APIT) is proposed by He et al. [86].
In APIT, each device divides its environment into different triangles, which are determined by all possible combinations of three
nearby beacon locations. Subsequently, the device uses a point-intriangle test to determine whether it is located outside or inside
of each defined triangle. The latter triangles are then aggregated
and the device’s location is computed as the center of gravity of

48

2.4 Localization

Figure 2.11: Location derived using the diffusion approach. The
black dots represent beacons while the gray dot
marks the location determined via diffusion at the
center of the beacons.
the combined triangles. The basic idea of the point-in-triangle
test is the following consideration. If there is a location next to
the current location of a device which is further or closer to all
beacons which define the triangle, the location of the device lies
outside this triangle, otherwise it is inside. Since the devices are
not assumed to be able to move in any direction in order to check
whether this changes their proximity to the beacons, the authors
propose to check whether the locations of the device’s communication neighbors fulfill this condition. Figure 2.12 illustrates the
localization process using APIT.
Angulation, as the name suggests, uses knowledge about the angles
between the device to be located and the beacons. Examples of
angulation can be found in Nasipuri and Li [173] and Zhang et al.
[255]. To compute the angle between a device and a reference
beacon, directed antennas or microphones are used which can distinguish the direction of an incoming signal. In addition, it is

49

Chapter 2 Research Context and State-of-the-Art

Figure 2.12: Example of localization using the APIT method. The
black dots symbolize beacons and the gray dot
represents the determined location. The triangles
defined by the beacons’ locations are indicated by
dashed lines. In this example, the location lies inside
all possible triangles, indicated by gray shaded areas,
and is therefore determined to be at the center of
gravity of the tetragon resulting from aggregation of
all triangles.

50

2.4 Localization
possible to derive the Angle of Arrival (AoA) from optical communication (cf. Bachrach and Taylor [12]). The angles are estimated
using the time difference between the signals’ arrivals at individual
microphones, or other receivers. The accuracy lies within a few degrees (cf. Priyantha et al. [183]). Since this method needs complex
installation and bulky hardware, the procedure is not commonly
applied to SSNs. Also, the need for spatial separation between
speakers is difficult with decreasing sensor size (cf. Bachrach and
Taylor [12]). In the work of Coore [44], an algorithm to compute
relative angles between network devices and a beacon without the
use of hardware is proposed. The computed angles are relative, i.e.,
they start at an arbitrary device with an angle value of zero and
estimate consistent angles from this point on. Niculescu and Nath
[178] suggest using AoA information in combination with distance
information to enhance localization results. From at least two
known angles, the unknown location of a device can be computed.
Figure 2.13 displays how angles between four beacons (black dots)
and the device to be located (gray dot) are used to determine the
unknown location of that fifth device.
While angulation is based on estimated angles, distance-based
localization methods require an estimate for the distances between
beacons and devices in the network. There are two algorithms
which use distance estimates to derive location information: the
bounding box approach and lateration. The bounding box method
is, for example, published by Savvides et al. [206] and Bachrach
and Taylor [12]. It assumes the position of a device to be in the
center of the overlap of square boxes which are drawn around the
beacons. The inradii of the square boxes correspond to the distance
estimates between the device and the respective beacons. Figure
2.14 illustrates how the location is determined with the bounding
box method as described. Although the bounding box procedure

51

Chapter 2 Research Context and State-of-the-Art

Figure 2.13: Example of localization using angulation. The black
dots symbolize beacons, while the gray dot represents
the unknown location. For each beacon the angle
between the x-axis and the unknown location is
determined. The unknown location is then assumed
to be at the intersection of the lines radiating from
each beacon at the predetermined angle.

52

2.4 Localization
has the advantage to be computationally relatively simple, the
derived locations can be calculated more precisely using lateration,
which requires the same kind of information as the bounding box
procedure and is, therefore, used more often.

Figure 2.14: With the bounding box method, the position
estimate (gray dot) is at the center of the bounding
boxes with inradii equal to the respective distance
estimates between the device and the beacons (black
dots).
Lateration is the most well-known localization technique (cf. Nagpal et al. [171] and Bachrach and Taylor [12]). Although it is often
referred to as triangulation, the notion of triangulation actually includes angulation and lateration (cf. Bachrach and Taylor [12]). In
contrast to angulation, lateration is based on determining distances
to beacons instead of angles. Using lateration, the position of a
device is assumed to be in the overlap of circles around beacons
with radii equal to the estimated distances between the device and
the respective beacons (cf. Figure 2.15).
In the work of Nagpal et al. [171], an iterative version of the
lateration algorithm is proposed. The method iteratively improves

53

Chapter 2 Research Context and State-of-the-Art

Figure 2.15: With lateration, the unknown position (gray dot) lies
within the overlap of circles around the beacons
(black dots) with radii equal to the respective
distance estimates.
an initial location estimate, for example random coordinates, by
stepwise minimizing the error between the estimated distances to
all beacons and the distances of the coordinates in the current
iteration.
In addition to the previously presented methods for localization,
several hybrid approaches have been suggested over time, for example in Chintalapudi et al. [40], Sun et al. [223], and Eren [60].
They usually combine some of the localization methods mentioned
previously in order to benefit from their various advantages and to
compensate for their drawbacks.

2.4.2 Distance Estimation
As mentioned in the preceding section, the determination of distances between devices and beacons is a crucial element of distancebased localization algorithms. However, since the devices to be

54

2.4 Localization
located usually have no means to explicitly measure distances,
numerous methods for distance estimation have been proposed
over time to overcome this problem. Figure 2.16 gives an overview
of these techniques, which can generally be categorized into rangebased and range-free methods. Range-based methods rely on the
analysis of physically transmitted signals. In contrast, range-free
methods abstract from the signal and derive information about the
distances of devices from the content of messages which devices
exchange in a network. The following two sections describe these
two kinds of distance estimation methods in detail.
Range-based Distance Estimation
The most common approach to distance estimation is to analyze the
strength of a transmitted radio frequency signal. This method is
commonly referred to as Radio Signal Strength Indication (RSSI)
(cf. Kai and Chun [114], Patwari and Hero [180]). The signal
strength p (d) decreases with increasing distance d. Hence, the
strength of a communication signal on receipt provides information
about the distance between sender and receiver. Equation 2.1
shows the underlying theoretical model:


d
p (d) = p (d0 ) − 10 α log
d0



+X

(2.1)

d0 denotes a reference distance, α represents the path-loss exponent,
X ∼ N (μ, σ) denotes a normally distributed random variable with
mean (μ) zero and variance σ, which describes the shadowing effect.
Although widely studied in the literature, there are some drawbacks
to this approach. RSSI ranging measurements contain noise in the
order of several meters (cf. Bahl and Padmanabhan [13]) and radio
propagation tends to be highly affected by barriers such as walls or

55

Chapter 2 Research Context and State-of-the-Art

Figure 2.16: Classification of distance estimation techniques.

56

2.4 Localization
furniture since they tend to reflect or absorb signals (cf. Bachrach
and Taylor [12]).
Another prominent method for distance estimation is called Time
of Arrival (ToA), also often referred to as Time of Flight (ToF)
(cf. Girod and Estrin [77], Meghani et al. [150]). In ToA, the
time for a physical signal to be transmitted from one device to
another is measured. ToA can be performed in a one-way or
two-way version. In the two-way version, a transceiver sends a
signal to the receiver which, in turn, sends an answer. The sender
measures the time t between sending and receiving the signal and,
together with the velocity v of the signal and the expected delay
treply which occurs between receiving a signal and answering, the
distance estimate d¯ between the two devices can be computed
using Equation 2.2. For the one-way version, the system clocks
of the transmitting and receiving device have to be synchronized.
This is usually done using radio and ultrasonic signals. The radio
signal has almost zero propagation speed indoors when compared
to the ultrasonic signal and is used for synchronization purpose (cf.
Meghani et al. [150]). Every device is equipped with a speaker and a
microphone. The transmitter (beacon) first sends a radio message,
waits for a fixed time period, and then sends a chirp sequence
over its speaker. The receiver gets the radio signal, turns on its
microphone and hears the chirp sequence. It notes the time t of
the transmission of the ultrasonic signal and uses this information
together with the transmission speed v of the audio signal to
calculate its physical distance from the transmitter using Equation
2.3. While ToA is impressively accurate for some scenarios, it has
the major drawback to require line-of-sight conditions and extensive
hardware. In addition, the speed of sound in the air varies with
air temperature and humidity (cf. Bachrach and Taylor [12]).
Furthermore, relatively fast processing capabilities are required in

57

Chapter 2 Research Context and State-of-the-Art
order to be able to resolve small time differences. Kwon et al. [131]
show that the precision can be improved by checking for reflexivity
and the triangle inequality.


d¯T oA = v ·

1
t − treply
2

d¯T oA = v · t



(2.2)
(2.3)

A similar distance estimation technique is called Time Difference
of Arrival (TDoA) (cf. Gustafsson and Gunnarsson [80], Mao et al.
[146], Meghani et al. [150]). The device to be located sends signals
to the beacons in the network and these beacons each measure the
ToA of their signal. From each set of differences between ToAs
and the known distance between the respective beacons, unique
hyperbolic curves around the respective beacons are defined. The
intersections of these curves determine the unknown location of
the device (cf. Bucher and Misra [27]).
All range-based distance estimation methods have in common
that they require special hardware in order to be able to analyze
physical signals. Since such hardware can be large and expensive,
range-based distance estimation is unsuitable for some applications. However, range-free distance estimation methods are able
to overcome this drawback.
Range-free Distance Estimation
Estimating distances in a network without having to rely on the
interpretation of the physical signal is called range-free distance
estimation. Instead, range-free distance estimation is based on messages which are exchanged between the devices in a network. There
are two types of range-free distance estimation algorithms, hop
count based and connectivity-based distance estimation algorithms.

58

2.4 Localization
The hop count based algorithm is the more established approach of
the two. A hop count denotes the minimum number of relay devices
a device needs to be able to communicate with a beacon. The
Gradient Algorithm (GA), proposed by Nagpal et al. [171], is an
algorithm for networks used to determine each device’s hop count
with respect to a beacon. It is the basis for all distance estimation
algorithms based on hop counts. With GA, each beacon initiates a
so-called gradient wave by sending a message including an integer
value of 0 to its neighbors. Each neighboring device takes the
minimum value it receives, increments it by 1, and propagates it to
its neighbors. This process is repeated in fixed intervals, such that
the value can be updated regularly, especially, in dynamic networks.
Figure 2.17 illustrates the concept of hop counts. Devices which
are able to communicate with each other are connected by straight
lines. The beacon, displayed as a black dot on the left side, initiates
the process by communicating a value of zero to its neighbors. All
devices in the network take the minimum value they receive and
add one to compute their own hop count, which is represented
by the number inside the respective dots. Figure 2.18 shows an
example of the result achieved by the GA when applied to an
ad hoc network. All devices with the same hop count value are
colorized in the same shade of Grey. All devices with the same hop
counts are located in symmetrical rings around the beacon and
each ring has approximately the width of the communication range
r (cf. Nagpal et al. [171]). Therefore, the most basic idea for hop
count based distance estimation is to use r as an approximation for
the length of each hop and r times the hop count as an estimate for
the distance between a device and beacons. However, this estimate
is only valid for perfectly dense networks, which rarely occur in
reality.

59

Chapter 2 Research Context and State-of-the-Art

Figure 2.17: Hop count determination with GA initiated by the
beacon (black dot) which sends a value of zero to its
neighbors (connected by dotted lines), which
compute their hop count as minimum value received
plus one (numbers inside the dots).
Figure 2.18 displays the perfect gradient rings, i.e., the rings centered at the beacon with radii of multiples of r, drawn as black
circular rings. It is easy to see that the gradient rings which result
from GA differ from these perfect gradient rings. Kleinrock and
Silvester [122] show that the expected length of a hop in uniformly
random distributed networks is given as a function of the local
neighborhood density, i.e., the number of communication partners.
A similar, but somewhat simpler principle is used by Wong et al.
[249] for networks with varying density. Here, density dependent
reduction rates are chosen manually depending on the local density
and applied to the naive estimate r before multiplying it with the
hop count. A different method to estimate the length of a hop
is provided by an approach called DV-HOP (cf. Huang and Selvakennedy [102], Niculescu and Nath [177]). DV-Hop uses the fact
that Euclidean distances between beacons are known because their
location information is available. By comparison of the Euclidean

60

2.4 Localization

Figure 2.18: A network situation after application of GA. Devices
with the same hop counts are displayed in the same
shade of Grey. Black rings are centered at the beacon
and have radii of multiples of r, with r being the
communication range of the devices.

61

Chapter 2 Research Context and State-of-the-Art
distances with the respective hop counts between beacons, the
average length of a hop is calculated as a fraction of Euclidean
distance and hop count. Apart from improving the hop length
estimate, various methods are developed to determine the position
of a device within its own gradient ring. For example, in the work
of Nagpal et al. [171], an average of all hop counts in a device’s
neighborhood is calculated before multiplying it by the hop length
estimate, following the principle that devices which are closer to
the inner border of their gradient ring have a higher number of
neighbors with lower hop counts. Conversely, a node which is
closer to the outer border of its gradient ring has a higher number
of neighbors with higher hop counts. Liu et al. [140] refine this
approximation further by using the exact proportions of neighbors
with lower, equal, or higher hop counts in order to improve the
estimate and Wang et al. [242] propose to refine distance estimates
through weighted interpolation.
The second type of range-free distance estimation algorithms is
called connectivity-based distance estimation. It relies on the number of shared communication partners between two devices with
respect to the total number of communication partners. The idea
behind this concept is that devices which are close to each other
share more communication neighbors than distant devices (cf. Figure 2.19). This approach was first presented by Fekete et al. [64]
and Buschmann et al. [33]. Later on, it was refined by Aslam et al.
[10], Villafuerte et al. [236] and Huang et al. [99]. While Fekete
et al. [64] and Buschmann et al. [33] use only the number of shared
neighbors for distance estimation, Aslam et al. [10], Villafuerte
et al. [236], and Huang et al. [99] expand the concept by using the
ratio of shared to total communication partners. The mapping
between this ratio and the distance estimate is generally designed
as a lookup table derived through an a priori empirical study. How-

62

2.4 Localization
ever, Huang et al. [99] apply a first order Taylor series expansion to
approximate the mapping function. The method is tested on real
hardware and is shown to deliver more precise estimates compared
to an RSSI-based distance estimation approach.

Figure 2.19: Two devices and their communication ranges. The
distance between the devices and the number of
shared communication partners (gray dots with black
border-line) is correlated.

2.4.3 Mobility
As mentioned before, there is little research about the effects mobility has on the accuracy of localization algorithms. A common
argument for why mobility is not considered explicitly is that the
execution of the respective localization algorithm can be repeated
constantly in order to update location information. However, localization requires input data, such as distances or angle information,
usually derived by various estimation techniques, which are prone
to error under mobile conditions. This is, for example, shown by
Bergamo and Mazzini [17], where the accuracy of distance estimation based on radio signal strength is investigated in dynamic
networks. Experiments reveal that Rayleigh fading due to the
motions of the sensors can introduce significant errors to the esti-

63

Chapter 2 Research Context and State-of-the-Art
mates. Similar to these findings, mobility can be expected to have
an impact on range-free distance estimation. Range-free distance
estimates are derived from a device’s communication neighborhood
which is affected by the mobility of devices in the network. Potential influences on range-free distance estimation are not primarily
hardware related and have to be examined individually. This is,
for example, done by Lim and Rao [139], who identify the mobility
of some devices in the network to have a positive impact on the
accuracy of hop count based localization. However, the underlying
assumption of the performed experiments is that the mobile devices
know when they are moved and, if so, immediately update their
hop count values and distance estimates. This leads to a general
improvement of distance estimation within the network as it resembles a scenario in which new devices are placed in the network
at all locations, where the devices move to. Therefore, the network
density is virtually increased. Liu et al. [140] show that mobility affects the probabilistic density of devices in a communication range
which can influence the computed hop count values. The authors
provide a method to account for mobility when calculating distance
estimates based on hop counts. It has to be noted, though, that for
the proposed compensation of mobility-induced density distortions,
knowledge about the underlying mobility model is required. The
question how the devices can attain knowledge about their own
mobility pattern stays unanswered though. At least for devices
which are either carried or moved by external forces and do not
move actively, the answer to this question is not straightforward.
One solution to this problem is proposed by Kumar and Das [130].
The authors derive the movement pattern from a sequence of previously computed locations. However, the assumption that mobility
in a network follows a certain pattern is not always valid.

64

2.4 Localization
In contrast to localization techniques, dead-reckoning methods,
or deduced reckoning, are specific methods to locate, or better,
track the positions of mobile objects over time. Originally, deadreckoning was used by seafarers in the fifteenth century before
more accurate celestial navigation techniques were developed. A
ship’s current location was determine using its previous location
and an estimate for the direction and distance it has sailed since
(cf. Kumar and Das [130]). The estimate of an object’s current
location using dead-reckoning is based on previously determined
locations, a model which describes the object’s dynamics, and
frequent corrections by odometry readings, for example the object’s speed and direction. For odometry readings, sensors such as
accelometers, magnetometers, compasses, and gyroscopes are used.
Kalman Filters (cf. Bartlett [15]) or Particle Filters (cf. Ristic et al.
[197]), which take into account that sensor values are usually prone
to error, can be used to compute the location estimate. Examples
of localization systems that use dead-reckoning can be found in
Fischer et al. [67], Höllerer et al. [96], Koide and Kato [123], and
Retscher [191]. If the mobile device is carried by a human user,
information about the user’s specific walking behavior, such as
his average speed, can be used instead of odometry readings in
order to direct the search towards the most likely locations (cf.
Wu et al. [250]). Hu and Evans [97] use the location of newly detected beacons, which come into range when the device moves, as
odometry readings to derive location estimates. A major drawback
of dead-reckoning systems is the necessity for odometry readings
and the required model of the object’s dynamics. Moreover, deadreckoning is subject to cumulative errors up to the point where it
receives new extrinsic location information, hence this approach
is best used in combination with other localization techniques for
error correction. A great advantage is the possibility to extrapolate

65

Chapter 2 Research Context and State-of-the-Art
locations in case there is not enough information available to apply
a localization algorithm, for example due to a temporary absence of
beacons or due to the failure of distance estimation. Additionally,
dead-reckoning allows for the prediction of likely future locations of
mobile devices, which can come in handy for several applications.
Apart from research which deals with the mobility of the devices to
be located, some researchers attend to the potential improvement of
localization by employing mobile beacons (cf. Li et al. [133], Wang
et al. [243], You et al. [253], Zhang and Yu [254].

2.4.4 Indoor Localization Systems
In the literature, numerous localization systems specifically developed for indoor applications can be found. This section aims at
providing an overview of the proposed solutions. Apart from the
localization method applied, localization systems can be distinguished with regard to various criteria (cf. Hightower and Borriello
[94]). One criterion is the wireless data transmission technique in
the system, which is usually infrared, radio frequency, or ultrasound. Data transmission by means of radio frequency signals can
further be classified according to the implemented communication
protocols, such as Bluetooth [20], Zigbee [257], or the IEEE 802.11
standards, which are commonly used in Wireless Local Area Network (WLAN) technology [104]. There are localization systems
where the devices to be located perform the necessary computations locally and systems in which the localization is undertaken
by an external, mostly centralized computing unit. The latter
approach is associated with all disadvantages that characterize
centralized localization algorithms (cf. Section 2.4). In addition,
these systems bear the risk of privacy infringement since the location information is computed on an extrinsic server. Localization

66

2.4 Localization
systems, in general, can further differ with regard to their costs and
installation expenditure, the accuracy and precision of obtained
locations, or the kind of locations they deliver, i.e. physical versus
symbolic, absolute or relative locations. There are also some indoor localization systems which use entirely different localization
methods than those presented in Section 2.4.1. Two examples of
alternative localization systems are the Smart Floor introduced
by Orr and Abowd [179] and the MotionStar concept presented
by Raab et al. [187]. SmartFloor is a localization system for buildings which derives the locations of people from embedded pressure
sensors in the floor. In MotionStar, devices which generate axial
DC magnetic-field pulses are installed at fixed locations in the
building. Magnetic sensors in the devices to be located are used
to determine the orientation of the magnetic field and derive their
locations from this information. Table 2.1 gives an overview of
various indoor localization systems. Except for the SpotOn localization system, none of the proposed indoor navigation solutions
listed in Table 2.1 specifically exploits ad hoc network connections
between the devices to be located. This is probably due to the fact
that localization systems are desired to work even if there is only a
single device to be located in the building. In localization systems
based on ad hoc networks, the quality of the locations correlates
positively with the number of devices in the building. However, in
an evacuation scenario, a small number of devices imply a reduced
danger of congestions during the evacuation process. Hence, the
importance of the drawback that there are few devices available
for localization is dampened because the reduction in evacuation
time which can be achieved by the navigation system is negligible.
Nevertheless, localization systems based on ad hoc networks can
significantly reduce the requirements of the localization infrastructure, which in turn reduces hardware costs, as well as installation

67

Chapter 2 Research Context and State-of-the-Art
and maintenance expenditure. High costs and effort have, so far,
been a major hindrance in the prevalence of indoor localization
systems, a trend which could be overcome by the use of ad hoc
network based localization systems. Since dynamic evacuation
support systems should be available in as many buildings as possible, enhancing localization support by exploiting ad hoc network
connections could advance this vision and bring it closer to reality.
For this reason, this thesis concentrates on the investigation of ad
hoc network based localization methods for mobile devices.

68

2.4 Localization
Table 2.1: Overview of indoor localization systems, their applied
localization algorithms, and the wireless
communication technique in use.
Localization
Scene analysis

Lateration

Proximity

CommunSystem
ication
Computer vi- Wearable computers [78]
sion
Indoor navigation for the
blind [103]
NAVIO [191]
Easy Living [128]
IEEE 802.11
LaureaPOP [188]
NAVIO [191]
RADAR [13]
RFID tags
Virtual leading blocks [8]
Infrared
Navigation aid system [16]
Ultrasound
Cricket [182]
Drishti [189]
UHF
Navitime [9]
IEEE 802.11
Semantic navigation system [232]
RADAR [13]
3G
NAVIO [191]
Magnetic pulse MotionStar [187]
RFID tags
RadioVirgilio/ SesamoNet
[48]
RG-I [129]
Wearable way-finding computer [201]

69

Chapter 2 Research Context and State-of-the-Art

Deadreckoning

70

Blind navigation system
[54]
Space-identifying infrastructure [18]
SpotOn [95]
Infrared
Active Badge [244]
Ultrasound
ActiveBats [85]
Cricket [182]
Bluetooth
Smart navigation environment [101]
Barcode scan- E-scavenger hunt game
ning
[227]
Context-aware wayfinding
[37]
Metronaut [218]
RFID tags
Human navigation system
[123]
3G
NAVIO [191]
Ultrasound
Ultrasound Indoor Navigation [67]

CHAPTER

3
ORGANIC BUILDING EVACUATION

Organic Computing deals with the design of complex, self-organizing
technical systems in order to make them flexible and trustworthy
at the same time. According to the definition provided by Muehl
et al. [166], a system is self-organizing when it is self-managing,
structure-adaptive, and employing decentralized control. A system
is denoted as being self-managing if it can adapt to changes in its
environment without outside control. Structure-adaptive means
that the system has to maintain a certain kind of structure which
can be spatial, temporal, or of other kind. The O/C Architecture
developed in Organic Computing supports self-organization by
feedback and learning mechanisms as explained in Section 2.2.1.
In contrast to the previously given definition of self-organization,
the O/C Architecture explicitly allows for external control in order
to make the system trustworthy and robust against failure. Especially for such a dynamic environment as a building evacuation
scenario, adaptability plays an important role since it is impos-

71

Chapter 3 Organic Building Evacuation
sible to consider all potentially arising situations at design time
of an evacuation system. However, it is even more important to
make such an evacuation system controllable, since relying completely on a technical system in case of a life-threatening situation
is undesirable for most humans. In the following, a concept for
OBESS is introduced, which is designed according to the paradigm
of controlled self-organization from Organic Computing. The main
system components and their respective functionalities are described. Furthermore, it is noted how the O/C Architecture can
be used in OBESS in order to achieve the desired system behavior.

3.1 Concept of an Organic Building Evacuation
Support System
OBESS consists of three main components, which are displayed in
Figure 3.1. There is a CCU, i.e., a central server for each building,
which is used to configure the sensors in the SSN. The SSN is
the second component of OBESS and consists of numerous sensors
distributed throughout the building. These sensors are able to
communicate with each other and the CCU. The third system
component is a MANET, established by wireless communication
between mobile devices carried by potential evacuees. Mobile
devices in the MANET can communicate with sensors of the SSN
and vice versa. The functionalities of all three system components
are explained in detail in the following.

3.1.1 Central Control Unit
One important aspect of OBESS is its decentralized structure. As
emphasized before, the lack of a single point of failure is crucial,
especially in emergency systems. Nevertheless, there is a CCU in

72

3.1 Concept of an Organic Building Evacuation Support System

Figure 3.1: Structure of OBESS, its main elements (CCU, SSN,
and MANET), and the communication links between
the components.
OBESS and it is important to note that this control unit does not
play any part in the evacuation planning process and is, therefore,
not fatal for the evacuation support service in case of failure. The
main task of the CCU is to configure the static sensors. For example, providing the building layout and corresponding information
or assigning the respective location information to each sensor in
the building, such that they can be used as beacons for localization
of the mobile devices. In order to accomplish this task, the CCU
offers an interface for the user, which can be used to configure and
control OBESS. Apart from that, the CCU collects information
from the mobile devices using the sensors in the SSN in order to
optimize and adjust OBESS to the building’s specifics over time.
Such information could be the average number of mobile devices
in the building at specific times and days, the most frequented
rooms and paths, or others. The CCU can learn building-specific
characteristics and adapt the evacuation system over time in order to improve its effectiveness. For this, the two-level learning
mechanism of the generic O/C Architecture can be used which is

73

Chapter 3 Organic Building Evacuation
described in Section 2.2.1. An example for such a building-specific
optimization of OBESS is described in Section 5.4, where the placement of the sensors is improved in order to achieve better results
during the localization process of the mobile devices.

3.1.2 Static Sensor Network
The most obvious task of the SSN is to monitor indicators for
dangerous situations, such as heat, smoke, or similar properties,
perceivable by technical sensors. In case any of the measured
values exceeds a certain threshold, an alarm can be triggered or
a human user, who is responsible for the alarm, can be notified.
The SSN, therefore, serves as a monitoring and alarm system, similar to the ones that can be found in certain buildings today. In
Section 4.2, an algorithm is presented which relies partly on such
sensor information. The sensors in the SSN are assumed to have
knowledge of their exact locations in the building due to an a priori
configuration via the CCU interface. Hence, the static sensors
serve as beacons to support the localization of the mobile devices
in the building as detailed in Section 2.4. In addition, sensors can
provide useful information as a download to the mobile devices,
for example a layout of the building and its representation as an
evacuation graph model (cf. Section 2.3.1). Normally, sensors in
the network communicate wirelessly with each other. Nevertheless,
it is conceivable that the SSN is set up as a so-called small world
network. Small world networks are characterized by having mainly
local communication links between nearby devices but occasionally
feature certain devices which are connected over longer distances by
wire. It is shown that, in small world networks, a small number of
wired long distance connections can speed up the spread of information significantly (cf. Watts and Strogatz [245]). As a consequence,

74

3.1 Concept of an Organic Building Evacuation Support System
providing some wired connections between the sensors can enhance
the information exchange between distant mobile and static sensor
devices. Another task for the SSN is to collect information from
the mobile devices which can be used to evaluate the current status
of the evacuation system in this specific building. Such information
could be available input data for localization or trajectories of the
devices’ movements through the building. Sensors in OBESS are
not required to have an extensive computing capacity, and there
is no need to implement a complex O/C Architecture on these
devices.

3.1.3 Mobile Ad Hoc Network
The devices in the MANET are assumed to run software enabling
their usage for building evacuation. Whenever such a device enters
a building which is equipped with OBESS hardware, it receives a
message from a nearby sensor, including all necessary information
about the current building. In particular a possibility for downloading the building’s layout and the respective evacuation graph
model is provided. In case of an emergency situation, the sensors
inform all nearby mobile devices about this situation. The device
then calls the attention of its user by ringing and vibrating. On the
display of the device, the building’s layout and the current location
of the user are shown and the device indicates navigation instructions to an exit door, for example by pointing out the direction on
the screen or via voice instructions.
The mobile devices in OBESS are each assumed to belong to a
specific person who carries the device. This provides the opportunity to collect information about the human user’s preferences
and characteristics. These preferences or characteristics can be
used to personalize the evacuation support service offered by the

75

Chapter 3 Organic Building Evacuation
device. For example, the device can collect information about the
age of its user or whether he has any handicaps in order to adjust
the navigation instructions accordingly. Also, rescue forces can
be informed about certain health conditions to improve the help
they can provide. Whether such personal information is manually
entered by the user himself or is even learned by the mobile device
from observing its user’s behavior, i.e., visited websites, social
networks, et cetera, is open for discussion and future research. Of
course, an important aspect in this context is the protection of the
user’s privacy.
The main focus of this thesis lies on the mobile devices in OBESS.
In order to design the software for these devices, two main algorithmic challenges have to be mastered. Firstly, the device has to
be able to compute an evacuation path from its current position to
an exit of the building, thereby integrating knowledge about the
current evacuation situation. Two algorithms which can generate
navigation instructions are presented in Chapter 4. The second
challenge is that the devices have to be capable of determining their
locations inside the building. Chapter 5 addresses exemplary ways
to master this task. An implementation of the O/C Architecture
for the mobile evacuation devices in OBESS is described in the
following.

3.2 Observer/Controller Architecture for
Mobile Devices
As described in Section 2.2.1, the O/C Architecture is a design
framework for organic technical systems, i.e., systems which are selforganizing and exhibit life-like properties, such as being adaptable
to unforeseen changes, while at the same time being trustworthy,
robust, and controllable by human users. Hereafter, an exemplary

76

3.2 Observer/Controller Architecture for Mobile Devices
implementation of this generic architecture for mobile devices used
in OBESS is presented.
In OBESS, each mobile device and the CCU is designed according
to the O/C Architecture. However, the resulting system structure
is not hierarchical because there are no means for the CCU to
control the mobile devices in the system. Hence, the structure is
rather a decentralized one as depicted in Figure 2.4(b). The SuOC
of the mobile devices is different from the standard structure in the
way that the devices cannot only observe themselves, but also all
devices located in their communication ranges. Nevertheless, only
the devices’ own actions are controllable. Figure 3.2 illustrates this
structure.
The software of a mobile device in OBESS has three modes of
operation. Firstly, it can simply display the layout of the building
on its screen without providing any further functionality. Secondly,
it can run in localization-mode, in which the software determines
the device’s location within the building and shows it on its screen.
Thirdly, the device can switch to evacuation-mode, in which it
additionally computes navigation instructions in order to support
the evacuation of its user from the building. Since all devices have
limited resources and constrained energy supply, it is reasonable to
turn-off the evacuation functionalities if they are not needed. To
do so, the O/C Architecture provides the concept of observation
models which determine what kind of data is currently being
monitored and processed. The controller can switch between these
observation models depending on the current requirements. For
example, if the device is notified of an emergency situation either
by the sensors in its environment or directly by its user, evacuationmode can be switched-on; otherwise it is turned off and only
localization or simply the building’s layout is available to the user.
Both kinds of operation modes can have numerous different sub-

77

Chapter 3 Organic Building Evacuation

Figure 3.2: Structure of the O/C Architecture for mobile devices
in OBESS.

78

3.2 Observer/Controller Architecture for Mobile Devices
level observation models, one for each evacuation or localization
algorithm which is available to the mobile device. The main
purpose of the O/C Architecture is to observe and characterize
the device’s environment and switch between different algorithms
for evacuation planning and localization in order to select the
most suitable one for the current situation of the mobile device.
Furthermore, changes in the environment are meant to be detected
which trigger an update of the evacuation route. In the following,
the O/C Architecture is described for both, the evacuation-mode,
as well as the localization-mode. It should be noted that the
evacuation-mode requires location information, which is why the
evacuation-mode constitutes an enhancement of the localizationmode observation model.

3.2.1 Observer
The observer’s task is to provide for a characterization of the current system situation as well as a prediction of probable future
system states to the controller. These characterizations are used
as a basis for making decisions in the controller, i.e. selecting
the most suitable localization and evacuation algorithm. Figure
2.2 illustrates the observer part of the generic O/C Architecture.
The monitor unit is responsible for observing all data specified
by the observation model. In case of the evacuation-mode, all
necessary input data to perform evacuation planning has to be
monitored and the same holds for localization. For evacuation
planning, the locations of other evacuees in the building could
be of interest, as well as sensor data about dangerous areas. As
mentioned before, there is one observation model for each available
evacuation or localization algorithm. Hence, depending on the
selected algorithm, different input data is monitored. For example,

79

Chapter 3 Organic Building Evacuation
when beacon-based localization is applied (cf. Section 2.4.1), any
beacon location communicated by a nearby device is monitored.
In case of distance-based localization, hop count information from
surrounding devices could be monitored, or other data depending
on the selected distance estimation technique (cf. Section 2.4.2). In
addition, information about the current environment of the device
is observed, such as indicators about the network topology, for
example how many beacons are within reach. This environmental information can be used in the controller to decide about a
necessary change of the selected algorithm. Section 3.2.2 explains
this in detail. The pre-processor computes derivative data for the
algorithms by processing or filtering the monitored data. For example, if the device receives multiple messages with contradictory
content, the pre-processor decides which message contains the most
up-to-date information by comparing age-values of the messages
received. Another example would be to find the minimum hop
count value from a set of hop count messages received by neighbors
or to evaluate a potential evacuation route by summarizing sensor
information from rooms along this path. In the data-analyzer,
meaningful attributes which are able to characterize the current
system state are derived from the pre-processed data. Examples
for such attributes are the average speed with which nearby devices
are moved and their direction of movement. These parameters
could be used to characterize the mobility in the network. For
evacuation planning, the average concentration of evacuees in the
building is an example for a computed value which characterizes
the evacuation situation. The prediction module of the generic
O/C Architecture is meant to forecast the most likely future system state. For evacuation planning, it is conceivable to assess the
emergence of congestions, for example from an increasing concentration of devices in parts of the building. In case of an active

80

3.2 Observer/Controller Architecture for Mobile Devices
localization-mode, this is the place where dead-reckoning methods,
as described in Section 2.4.3, could be used in order to predict the
next location of a device. The observer’s aggregator module passes
all this information to the controller.

3.2.2 Controller
The controller’s main function in the O/C Architecture is to choose
an appropriate action, given a certain situation assessment and a
prediction of a future system state by the observer. In the presented
implementation of the O/C Architecture for mobile evacuation devices in OBESS, choosing an appropriate action is equivalent to
selecting a suitable evacuation planning and localization algorithm.
For evacuation planning, there are two evacuation planning approaches proposed in this thesis which differ in the information
required and deliver evacuation routes of different quality. Depending on the device’s state-of-charge, it might be advisable to choose
an evacuation planning algorithm which requires fewer messages to
be exchanged, even though the computed navigation instructions
are suboptimal. Furthermore, replanning of an evacuation instruction can become necessary when there is new information about
the situation in the building available to a device. In Section 2.4,
the broad variety of localization methods for ad hoc networks is
demonstrated. The sheer quantity of different approaches to the
problem suggests that there is more than one optimal method to
find the location of devices in a network. In fact, each algorithm
has its own advantages and disadvantages and it is not always easy
to choose which localization algorithm is preferable for a given situation. Fortunately, the organic O/C Architecture offers a solution
to select between multiple evacuation and localization algorithms at
runtime, respecting the current network situation and the device’s

81

Chapter 3 Organic Building Evacuation
system conditions. Figure 2.3 illustrates the controller part of the
generic O/C Architecture, its elements, and their interactions.
The controller takes actions in order to achieve desired and prevent
unwanted system behavior. Undesired evacuation behavior occurs
when an evacuation instruction is not optimal anymore with respect
to current objectives. One reason for this can be a deviation of
the user from the suggested evacuation path or because a newly
available situation characteristic implies that another escape route
is superior in terms of current optimization objectives. In case of
localization, an undesired system behavior appears when locations
which are computed by a chosen localization algorithm deviate too
much from real locations of the device and when there is another
localization algorithm which is expected to deliver more accurate
results. The available control actions are to change the applied
evacuation or localization algorithm and to trigger replanning of
evacuation instructions. In the O/C Architecture, there are two
types of control mechanisms. An intrinsic control, where the controller chooses an action depending on the situation characteristics
and prediction of the future system state which it receives from the
observer. These characteristics are matched in the action selection
module and the situation-action mapping which has the highest
reward is chosen as described in Section 2.2.1. Alternatively, it
is also conceivable that a control action is triggered by the user
of the O/C system. In the generic O/C Architecture, the user is
intended to control the system by setting and adapting the system’s
objectives and goals. The most obvious change in system objectives
is a switch between the modes of operation, such as turning the
evacuation procedure on or off. Another conceivable change of
system objectives is to set other optimization objectives for the escape route planning in the evacuation-mode. It could be desired to
search for an escape route with minimum risk exposure instead of

82

3.2 Observer/Controller Architecture for Mobile Devices
travel time. However, in the O/C Architecture for mobile devices
in OBESS, a direct interaction between user and mobile device is
also allowed. For example, the user can correct computed locations
by tapping on the screen of the device or by choosing a different
evacuation route than the suggested one simply by walking in another direction. In addition, it is conceivable that the user selects
a compulsory localization algorithm, which has to be used by the
system, or inserts a priori available information about the building
in order to facilitate the controller’s selection of an appropriate algorithm. The O/C Architecture is, furthermore, particularly suited
for respecting user preferences and characteristics during operation.
The user can provide personal information which influences the
objectives for evacuation planning in the evacuation-mode. Such
information could be the need for accessible paths or the avoidance
of stairs in case the user is handicapped or of high age.
In the generic O/C Architecture the situation to action mapping
is intended to be learned using the two-level learning infrastructure
provided. While the first-level learning, also called online learning,
is a feedback mechanism, which evaluates results achieved by
control actions at runtime, the second-level learning, or offline
learning, is based on a simulation, in which potential control actions
are tested before they are made available to the system. While the
first-level learning is often based on a Learning Classifier System
(LCS), as described by Richter [193], the second-level learning can
be implemented as an EA (cf. Section 2.2.1). For online learning,
the LCS assigns a reward to each mapping. From different matching
situations, it applies the action with the highest reward value.
After applying an action, the system receives positive or negative
feedback, evaluates the applied mapping accordingly, and updates
its reward. In OBESS, receiving feedback for evacuation planning
or localization is not straightforward since the mobile devices have

83

Chapter 3 Organic Building Evacuation
no possibility to know whether they computed the right locations
or whether the determined evacuation route is optimal. However,
active interventions of the user can be interpreted as a negative
feedback and, hence, used to evaluate the quality of the current
localization and evacuation algorithms. Additionally, feedback can
be derived from situation parameters provided by the observer. If
a change in the environmental situation is detected, for example a
recomputation of the evacuation instruction can be triggered, which
integrates this newly available information. A sample situation
for changing the applied evacuation planning algorithm is when
the observed state-of-charge of the device’s battery falls below a
certain threshold and an evacuation algorithm which requires fewer
messages to be exchanged is chosen in order to save energy. One
way of evaluating a localization algorithm is to perform consistency
checks. Usually, two consecutively computed locations should not
lie far apart from each other. If this is the case, the currently
selected localization algorithm does not deliver high quality results
in the current network situation and the corresponding fitness value
is decreased. Moreover, the computed locations of the device’s
neighbors should lie within its communication range, and so forth.
Feedback for an evacuation route is obtained whenever the device
receives new information about the current situation in the building.
The device could receive information that there are more evacuees
located in rooms along the currently chosen route than expected, in
which case a replanning of the evacuation instruction is triggered.
Apart from the online learning described above, the O/C Architecture provides for offline learning. For this, the situation parameters
from the observer can be used to configure a simulation, which
reflects the actual network situation and evacuation and localization
algorithms can be tested offline in this simulation before they are
made available for usage during runtime. This mechanism reduces

84

3.3 Summary
the probability of choosing unsuitable actions, i.e., inappropriate
localization algorithms, for a given network environment. It is
also conceivable to use the generic O/C Architecture’s capability
for offline learning in order to improve evacuation planning. The
simulation environment can be based on any of the evacuation
models presented in Section 2.3.1 and it is initialized by choosing
parameter values that reflect the situation currently observed in
the building. With this simulation, the consequences of following a
specific evacuation instruction derived by an evacuation planning
algorithm can be assessed and evaluated. Evacuation instructions
which lead to a slow evacuation time in the simulation can be
substituted by instructions which result in quicker evacuations.
This procedure ensures that evacuation instructions are thoroughly
evaluated before they are shown to the user.

3.3 Summary
In this chapter, a concept for a building evacuation support system
using mobile devices is introduced. This concept is based on the
principles of Organic Computing and called OBESS. OBESS consists of mobile devices carried by potential evacuees, who organize
in a MANET, a stationary SSN installed in the building, and a
CCU mainly used for configuration purpose. The sensors are meant
to be able to observe danger indicators in the building and support
the localization of the mobile devices as beacons. The mobile devices guide their users to safe exits by performing localization and
subsequent evacuation planning. This chapter proceeds with presenting an exemplary application of the generic O/C Architecture
to mobile evacuation devices. The O/C Architecture supports an
adaptive behavior of the mobile devices to changes in their environment by a two-level learning mechanism. Such a behavior

85

Chapter 3 Organic Building Evacuation
is crucial in an evacuation scenario, which is unpredictable and
dynamic. At the same time, the system becomes trustworthy,
robust, and controllable by human users via intrinsic and direct
control mechanisms. This is an important characteristic in order to
make a technical system trustworthy, which humans are supposed
to depend on in life-threatening situations. Apart from that, this
generic architecture provides for several further advantages, such
as the possibility to select the data that is currently monitored
and to easily integrate user preferences into evacuation planning.
The main objective of the O/C Architecture for mobile evacuation
devices is to select an appropriate localization and evacuation planning algorithm from a set of available methods or to trigger a new
evacuation path planning. The selected actions are chosen in the
controller depending on the currently observed situation in the
building and a prediction of its likely future state. To do so, a twolevel learning structure is used which consists of an online learning
mechanism based on feedback from the user and the device’s environment and a simulation-based offline learning approach. It is
discussed what kind of feedback can be used for online learning and
how this information is used to choose an appropriate localization
and evacuation planning algorithm at runtime. Offline learning
offers the possibility for simulative testing of localization algorithms
in realistic network scenarios as well as the evaluation of a specific
evacuation instruction before admitting it to productive utilization. The main advantage compared to standard simulative based
approaches, is that the simulation can be parameterized according
to the currently observed situation inside the building instead of
being based on expected values derived from past experiences.

86

CHAPTER

4

SWARM EVACUATION PLANNING WITH
MOBILE DEVICES

If a building has to be evacuated, the evacuees, ideally, leave the
place in the shortest time possible. According to Daniels et al.
[46], the time evacuees need to escape is composed of the time
until the emergency is detected and the alarm is triggered and the
time for evacuees to react and leave the building. Apart from the
reaction time of the evacuee, all other components can be subject
to optimization in an evacuation support system. Nevertheless, in
the following, only the time for leaving the building is considered.
This time can be influenced directly by the quality of navigation
suggestions provided by mobile devices. The time which an evacuee
needs to find his way out of a building depends on several factors.
The knowledge about a building’s layout and the right choice of
directions towards an exit have the highest impact. In today’s
buildings, the layout is usually provided by stationary emergency

87

Chapter 4 Swarm Evacuation Planning with Mobile Devices
evacuation maps. Apart from that, standard evacuation signs as
well as markings on the evacuation map point out the shortest
path from a certain location to a nearby exit. Although knowing
the shortest path is helpful, the shortest path does not always
correspond to the quickest path leading outside the building. The
main reason for this is that congestions often arise when a huge
crowd tries to leave the building on the shortest path. Other causes
can be fire or gas which block passages that are part of a shortest
path. As a consequence, evacuation planning should concentrate on
finding the quickest rather than the shortest path leading outside
a building. Currently, evacuation research tackles this problem
by performing simulations and, subsequently, trying to avoid congestions by either making modifications to the building or to the
recommended exit routes (cf. Section 2.3.2). This procedure, however, does not guarantee good results for situations which differ
strongly from the simulations, for example, if there are far more or
less evacuees inside the building compared to the scenario it has
been optimized for. This is the point at which OBESS begins to
develop its full potential. Due to the ability of the mobile devices
to communicate over ad hoc network connections, the opportunity
arises to gain insights into the current situation inside the building
at the time of evacuation. Using communication, knowledge about
the number of evacuees inside the building, potential congestions,
or other useful information can be distributed over the ad hoc
network. This allows for the devices to get a view of the evacuation situation, which extends beyond their own limited range of
communication. This knowledge can, subsequently, be integrated
into evacuation path planning, enabling not only the optimization
of the distance but also of the time it takes to reach an exit. Of
course, questions arise whether the information gathered via local
communication in a MANET has enough quality to really improve

88

evacuation planning since it is potentially incomplete, delayed,
or error-prone and how evacuation planning can be performed in
a decentralized way without central coordination. This chapter
addresses these questions and proposes two algorithms, which are
evaluated in a simulative evacuation scenario. In order to be able to
realize a decentralized evacuation system like OBESS, the mobile
devices have to be enabled to compute reasonable evacuation routes
towards a building’s exit without central control. Furthermore,
it is desirable to integrate available knowledge about the current
situation from localization communication into this process.
Definition 4.1 (Swarm Evacuation Planning). Swarm Evacuation Planning (SEP) denotes the process of finding an optimal
evacuation route for users of mobile devices without the help of a
central computing unit by integrating information about the current
evacuation situation gathered via local communication.
SEP receives its name due to its similarity to the concept of Swarm
Intelligence (cf. Section 2.2.3). In Swarm Intelligence, a global
behavior, which is not achievable by an individual alone, arises as
an emergent effect from the interactions between the individuals
in the swarm. The same holds for SEP. Only a swarm of mobile
devices is able to collect enough information about the global
evacuation situation to be able to assess congestion potentials and
guide evacuees on optimal routes.
For computation of the optimal route, information about the global
situation is collected via local communication. The result of SEP
is an evacuation route for the individual user of each device such
that the overall evacuation time is minimized. Another conceivable
objective is to optimize each individual’s evacuation time. This
could possibly lead to a higher acceptance of the evacuation device but certainly could also increase the overall evacuation time.

89

Chapter 4 Swarm Evacuation Planning with Mobile Devices
OBESS, so far, optimizes the overall evacuation time because this
objective is more reasonable from a social perspective. Nevertheless, changing this objective to a more egoistic one could certainly
be up for discussion.
A great advantage of such a decentralized evacuation system is
that the information exchange works, even in case of failure of
certain devices. Such a network is robust and lacks a single point
of failure, which makes its deployment attractive, especially for
emergency situations. To avoid that each user fully depends on his
mobile device, the system can easily be complemented by installing
some static digital devices inside the building, which can serve as
a fallback in case a user does not have a portable device or if it is
broken or run out of battery. The approaches proposed hereafter
and the corresponding experimental results have been published
partially by Merkel et al. [155, 158].

4.1 Macroscopic Swarm Evacuation Planning
As introduced in Section 2.3.2, macroscopic evacuation models are
the common basis for evacuation planning. Macroscopic models
regard groups of homogeneous evacuees, i.e., evacuees who are
equal in terms of speed, size, and behavior. Such a graph model
is usually assumed to be constructed manually by experts and
can be provided to the devices beforehand. The environment is
modeled as a graph G = (N, E) with nodes, n ∈ N , representing
rooms or corridors of the building, and edges, e ∈ E, representing
doors or similar connections between the areas represented by
nodes. Nodes are typically modeled to have a capacity and a
number of initially allocated evacuees, which corresponds to the
number of evacuees currently located in that room. An edge has
two weights, c and l, with c referring to the number of evacuees

90

4.1 Macroscopic Swarm Evacuation Planning
who can travel simultaneously on this edge and l representing the
length of the edge. The edge length can, for example, be measured
as the average time needed for one evacuee to traverse it or as the
distance between the centers of the connected rooms. Figure 2.6
shows an example of a building and its corresponding macroscopic
graph model G.
The goal of traditional macroscopic evacuation planning is to determine how many evacuees (flow size) have to be sent on which
path in order to optimize the given objective. This objective can
be to minimize the overall travel time, maximize the flow size in a
certain time period, and so forth (cf. Section 2.3.2 for a detailed
description). The problem with such macroscopic evacuation planning is that it only computes how many evacuees located in a room
have to use a certain path towards the exit. There are no means
to decide which evacuee is sent on which path. Since the different
paths have different evacuation times, the allocation of evacuees to
paths is a challenge in SEP. In addition, macroscopic evacuation
planning algorithms require global information about the number of
evacuees and their locations in the building, which is not provided
in a scenario without a global observer. Nevertheless, for a first
attempt to develop an SEP algorithm it seems reasonable to take
a traditional evacuation planning approach as a role model. Thus,
a macroscopic evacuation planning approach serves as a basis for
the following capacity constrained SEP algorithm. The algorithm
is modified in order to work in a distributed and decentralized way,
i.e., to be executable on each device in the network, and to deliver
an optimal evacuation route for each evacuee.

91

Chapter 4 Swarm Evacuation Planning with Mobile Devices

4.1.1 Capacity Constrained Swarm Evacuation Planning
The evacuation scenario considered consists of m mobile devices,
which have the evacuation graph model G of the building at their
disposal. For simplicity, each device has perfect information about
its own position in the building. In order to compute an optimal
evacuation route for each individual, the mobile device uses the
Capacity Constrained Swarm Evacuation Planner (CC-SEP), which
is shown coarsely in Algorithm 4.1. Firstly, the device collects
information about the position of other evacuees in the network via
local communication. This information is then used to compute a
global evacuation plan with an appropriate macroscopic evacuation
algorithm. As a final step, one of the evacuation instructions from
the evacuation plan is selected. This instruction can be used by the
device to guide its user, for example by displaying the respective
directions on its screen. In the following, the realization of each
step in this algorithm is described in detail.
Algorithm 4.1 Capacity Constrained Swarm Evacuation Planner
Require: Evacuation graph G = (N, E)
Ensure: Evacuation instruction i
1: G = Inf ormationCollection (G)
2: Evacuation plan I = EvacuationP lanning (G)
3: Evacuation instruction i = InstructionSelection (I)
4: return i

Information Collection
In order to compute the global evacuation plan, initial allocations
of evacuees in the evacuation graph have to be known, i.e., how
many evacuees are currently located at each node. This knowledge

92

4.1 Macroscopic Swarm Evacuation Planning
is gathered via local communication. Since the optimization is
done on the abstract graph level, the exact position of each evacuee
is negligible and only the number of evacuees located at a certain
node is required. To obtain the initial allocation, each evacuee’s
device periodically sends a message containing its identification
number (ID), the information about the node n where the device
is currently located at, and two lists, which are referred to as
local-count and global-count. The global-count list maintains the
number of devices located at each node in the evacuation graph,
which is the required information for the evacuation planning
process. In order to compile this data, an auxiliary list called
local-count is used. It contains IDs from all known devices which
are located at the same node as the device to which this list
belongs. Each entry in these lists has an age-value assigned to
it and all age-values are increased by 1 before a device sends a
message. Due to asynchronous message forwarding in the network,
it can happen that a message which contains more up-to-date
information is transmitted subsequently to a message with outdated
information. The age-value ensures that each device knows which
message contains the most recent information. The lower an agevalue is, the more recent the respective information. Each device
generates an entry for its own node in the global-count list by
counting the IDs in its local-count list. The exact procedure
is described in the following. The local-count always contains
the device’s ID with an age-value of 0 assigned to it. When
a device communicates with another nearby device at the same
node, it adds the ID of its communication partner to the localcount list and assigns an age-value of 0 to this information as
well. In addition, it compares the entries from its own local-count
list with the ones in the local-count list of its communication
partner. It adds all IDs which are unknown to its local-count

93

Chapter 4 Swarm Evacuation Planning with Mobile Devices
list and assigns the same age-values which the entries have in the
communication partner’s list. If two different messages contain
the same ID, the device takes the one with the more recent agevalue. As a next step, the device counts all entries in its local-count
list and adds an entry to its global-count list which contains the
node information n, the number of entries a (n) and an assigned
age-value of 0. Further, the device compares the entries in the
global-count list of its communication partners with its own and
adds the missing entries with the age-values they have in the
communication partner’s list. If the device receives a message with
an entry in the global-count list that is already stored in its own
global-count list, it overwrites the information only in case the
age-value of the received information is lower. If the age-value of
an entry in the local-count or global-count list exceeds a predefined
threshold tmax , the entry is removed from the list. This expiration
date is necessary because it is impossible to determine whether
an evacuee has already left a room or not. An increase in the
age-value for a certain entry in the local-count list between two
updates can have two reasons. Either, the evacuee is not in the
room anymore and, therefore, there is no device which resets this
age-value, or the device receiving the information is moving away
from the source of information. In this case, the number of devices
it takes to forward the message increases and, as a consequence,
the age-value is higher when the information arrives at the device.
The right choice of the expiration date is crucial. If the value is
too small, it can happen that the information is not passed on to
all devices before it is recognized as obsolete. If an expiration date
is set too high, information about other evacuees’ whereabouts are
assumed to be up-to-date, although the devices may have already
been moved somewhere else. In order to determine an appropriate
value for the expiration date, the characteristics of the building’s

94

4.1 Macroscopic Swarm Evacuation Planning
layout should be taken into account. Using knowledge about the
building’s layout and the communication range r of the devices,
the expiration date for information on the local-count list is set as
SpEx (n)
r
where n ∈ N represents each room in the building, i.e. node in the
macroscopic graph, and SpEx (n) refers to the spatial expansion
of that room. In case of rectangular rooms this can, for example,
be the diagonal of the room. For values in the global-count list,
the expiration date is computed based on the spatial expansion
of the whole building. In case of a building with multiple floors,
the expansion of one floor would be the best reference value to
compute the expiration date.
By using the algorithm described above, the devices are able
to collect information about the number of evacuees in different
rooms (nodes) and, with time, build a local view of the evacuation
situation. This view can then be used in a macroscopic evacuation
planning approach, in the same manner as if the device had global
information. As soon as the initial allocation of evacuees is known,
the evacuation planning can be initiated. There are two conceivable
implementations for the evacuation planning. It is either performed
only once by using all the information collected up to this point,
for example right after the alarm is triggered, or the evacuation
planning is constantly repeated, hence, adapting to new information
which becomes available during the evacuation process. Both
versions are tested in simulative experiments.
maxn∈N

Evacuation Planning
In order to compute the global evacuation plan, any macroscopic
evacuation planning algorithm can be used. As detailed in Section

95

Chapter 4 Swarm Evacuation Planning with Mobile Devices
2.3.2, most of these algorithms require a time-expansion of the
evacuation graph, making evacuation planning a time and space
consuming task. Lu et al. [143] introduce a heuristic algorithm
called Capacity Constrained Routing Planner (CCRP), which does
not rely on a time-expansion of the evacuation graph and provides
solutions of similar quality (cf. Sangho et al. [204]). However,
the algorithm can also be replaced by other macroscopic planning
methods, for example, the multi-objective optimization approach
proposed by Cheng [39], Zong et al. [258], or evacuation planning
based on EA as described in Garrett et al. [72] and Saadatseresht
et al. [202].
CCRP returns an evacuation plan I which consists of a set of
evacuation instructions i. Each instruction contains a path p in
the evacuation graph, i.e., a sequence of nodes, a flow size f , i.e.,
the number of evacuees who are sent on this path, and a sequence
of starting times t. The starting times denote at which time the
corresponding node on path p is entered by evacuees who follow
this evacuation instruction. The basic idea of CCRP is to model
the used capacities of each node and each edge as a time series.
This time series can then be used to derive waiting times at certain
nodes and edges in order to find an optimal path with respect to
traveling time, while taking into account all previously scheduled
evacuees and the capacity constraints of the graph. The algorithm
used for optimization is taken from network routing theory and
assumes a homogeneous speed of the evacuees involved. CCRP
is an iterative method, which computes evacuation paths for all
evacuees located at one node before processing the next node. The
priority with which the nodes are processed decreases with their
increasing distance to an exit. In other words, firstly, all evacuees
located on the node which is closest to an exit are scheduled for
evacuation, after which the next node is processed and so forth. For

96

4.1 Macroscopic Swarm Evacuation Planning
each node, path p is determined as the path closest to an exit with
minimum travel time. The travel time includes the time needed
to cover the distance as well as the waiting times for released
capacities which are occupied by previously scheduled evacuees.
After the quickest path is found, the number of evacuees f which
can be sent simultaneously on this path is computed. In the next
step, the time series for all nodes on this path are updated and
respective capacities are reserved for the evacuees at times t when
they pass the respective node or edge. The reserved capacities
are then taken into account in the next iteration when waiting
times are computed. The set of p, f , and t form one evacuation
instruction i and are added to the global evacuation plan I. The
evacuation planning terminates when all evacuees are scheduled.
Algorithm 4.2 describes the operating principle of CCRP (a fully
detailed description can be found in Lu et al. [143]). As explained
before, nodes n ∈ N have an initial allocation a (n) and a maximum
capacity c (n). The capacity of an edge e = (ni , nj ) is denoted by
c (e). Each edge in an evacuation graph represents the connection
between the center of two rooms. An edge, therefore, has a length
l (e) which is comprised of the two parts ls (e) and le (e), with
ls (e) representing the travel distance inside the starting room and
le (e) the travel distance which corresponds to the target node of
the edge. It holds that ls (e) + le (e) = l (e). This is important
for computing the reserved capacities in each room at the time
evacuees are assumed to enter them. The evacuation graph G
contains a set of exit nodes D ⊂ N , which represent the exits of
the building. In order to find evacuation instructions with CCRP,
an additional super source node suSo and an additional super sink
node suSi are added to the evacuation graph, both with infinite
capacities. The super source node suSo is connected via edges with
infinite capacities and zero lengths to all nodes in the graph, except

97

Chapter 4 Swarm Evacuation Planning with Mobile Devices
Table 4.1: Evacuation instructions obtained by applying CCRP to
the example graph in Figure 2.6.

Index
1
2
3
4
5
6

Instruction
Flow Size Path and Timing
f=2
p = < n3 , n4 >
t = < 0, 1 >
f=1
p = < n3 , n4 >
t = < 1, 2 >
f=1
p = < n1 , n3 , n4 >
t = < 0, 1, 2 >
f=3
p = < n2 , n4 >
t = < 0, 2 >
f=2
p = < n1 , n3 , n4 >
t = < 1, 2, 3 >
f=1
p = < n2 , n4 >
t = < 1, 3 >

for the super sink node. Similarly, all exit nodes are connected
to the super sink node via edges with infinite capacities and zero
lengths. This is an auxiliary construction, which ensures that there
is only one source and only one sink node in the graph and, hence,
simplifies the wayfinding procedure. Table 4.1 shows an example of
a list of evacuation instructions produced by the CCRP given the
example graph shown in Figure 2.6 as input. It should be noted
that the evacuation instructions returned by CCRP are sorted in
ascending order with respect to their evacuation time and that
it is possible to have more than one evacuation instruction for a
specific starting node n.

98

4.1 Macroscopic Swarm Evacuation Planning
Algorithm 4.2 Capacity Constrained Routing Planner [143]
Require: Evacuation graph G = (N, E)
Ensure: Evacuation plan I (set of evacuation instructions i)
// Initialize capacity time series for all nodes and edges:
1: ∀n ∈ N : f reeCap (n, ∗) ← c (n),
f reeCap (n, 0) ← c (n) − a (n)
2: ∀e ∈ E: f reeCap (e, ∗) ← c (e)
3: while ∃n : f reeCap (n, 0) < c (n) do
4:
prev (n) ← null, minT (n) ← ∞
5:
∀n ∈ N , minT (suSo) ← 0, Q ← {suSo}
// Find next best path:
6:
while Q
= ∅ do
7:
Node u ← minn∈Q (minT (n))
8:
for Node v ∈ neighbors (u) do
9:
if (u
= suSo) ∨ (f reeCap (v, 0) < c (v)) then
10:
arrT ← minT (u), waitT ← 0
11:
Edge e ← (u, v)
12:
nCap ← f reeCap (v, arrT + waitT + ls (e))
13:
eCap ← f reeCap (e, arrT )
14:
while eCap = 0 ∨ nCap = 0 do
15:
waitT ← waitT + 1
16:
eCap ← f reeCap (e, arrT + waitT )
17:
nCap ← f reeCap (v, arrT + waitT + ls (e))
18:
end while
19:
if (arrT + waitT + l (e)) < minT (v) then
20:
minT (v) ← arrT + waitT + l (e)
21:
prev (v) ← u
22:
end if
23:
end if
24:
end for
25:
end while

99

Chapter 4 Swarm Evacuation Planning with Mobile Devices

26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:
48:
49:
50:
51:
52:
53:
54:

100

// Create path p time schedule t and flow size f :
Node n ← prev (suSi)
while n
= suSo do
p.add (n), t.add (minT (n) − l (e))
Edge e ← (prev (n) , n)
f ← min (f, f reeCap (e, minT (n) − l (e)))
n ← prev (n)
end while
reverse (p), reverse (t)
Source node s ← p.getF irstN ode ()
f ← min (f, c (s) − f reeCap (s, 0))
// Allocate capacities on path:
f reeCap (s, 0) ← f reeCap (s, 0) + f
Node n ← prev (suSi)
while prev (n)
= suSo do
Edge e ← (prev (n) , n)
schedT ← minT (n) − l (e)
waitT ← minT (n) − minT (prev (n)) − l (e)
f reeCap (e, schedT ) ← f reeCap (e, schedT ) − f
startT = minT (n)
endT = minT (prev (n)) + waitT + ls (e)
for startT ≥ time > endT do
f reeCap (n, time) ← f reeCap (n, time) − f
end for
m = prev(n)
startT = minT (m) + waitT + ls (e), endT = minT (m)
for startT ≥ time > endT do
f reeCap(m, time) ← f reeCap(m, time) − f
end for
n ← prev(n)
end while

4.1 Macroscopic Swarm Evacuation Planning
Evacuation instruction i ← (p, t, f )
56:
Add instruction i to plan I
57: end while
58: return Evacuation plan I
55:

Instruction Selection
The third step in CC-SEP is to select an appropriate evacuation
instruction for the device’s user from the global evacuation plan I.
Valid candidates are all instructions with a path p that starts at the
node where the device is located. The objective of the instruction
selection is for the number of devices which select a specific instruction to approximately correspond to the instruction’s associated
flow size f . This is necessary in order get a close match between the
real waiting times which occur during evacuation and the waiting
times used for evacuation planning. To meet this challenge, each
evacuee’s device sorts the IDs in its local-count list in an ascending
order and maps the instructions returned by CCRP successively
to the IDs in the list. Thereby, the flow-size of each instruction
determines how many successive IDs are mapped to that specific
instruction before advancing to the next one in the evacuation
plan. The mapping process can be stopped as soon as the device’s
own ID is mapped to an evacuation instruction which the device
then uses to guide its user. In CC-SEP each device has its own
local view of the evacuation situation, as a result, the computed
evacuation plan I can also differ from one device to another. Nevertheless, it is safe to assume that devices which are located in the
same room have a very similar information base and, thus, their
evacuation plans are similar as well. Although sorting according to
device identification numbers seems to be a very naive approach to
assign evacuation instructions, there is a high chance for the selec-

101

Chapter 4 Swarm Evacuation Planning with Mobile Devices
tion process to result in a distribution of evacuation instructions
amongst the devices which is proportional to the flow-sizes of the
respective instructions without explicit coordination between the
devices. Nonetheless, this selection process is a naive procedure,
which is prone to manipulation. Because the instructions in the
evacuation plan are in increasing order of evacuation time, there
is a high incentive for the mobile device user to manipulate the
device’s ID in order to receive a faster evacuation path. For a
first assessment on how distributed and decentralized evacuation
planning in a MANET can be performed, however, this approach
is sufficient.

4.1.2 Evaluation
For the experimental evaluations presented in this thesis, a Javabased simulation called Swarm Simulation is developed, which
incorporates both, an evacuation simulation and a mobile ad hoc
network simulation. The evacuation simulation is based on a simple
time-discrete evacuation model similar to the two-dimensional cellular automaton described in Burstedde et al. [32]. For the mobile
ad hoc network communication, the communication neighborhood
of a device corresponds to its physical neighborhood on the plane
within a fixed Euclidean distance r. This model corresponds to the
so-called unit disc model which is described by Clark et al. [43].
Agents in the simulation are equivalent to evacuees and the mobile
devices they carry around. The simulation runs in cycles and in
each cycle, all agents are executed once sequentially in a specified
order, which is also called scheduling. For the evaluations in this
thesis, a random scheduling is selected. One execution denotes the
process of first updating the list of neighbor’s a device currently
has, i.e. can communicate with. As a next step, the behavior of

102

4.1 Macroscopic Swarm Evacuation Planning
the agent is executed; this can, for example, be the computation
of an evacuation instruction or the current location of the device.
For such computations, the device can access any information
from its neighbors as it is available at that point of time. After
the behavior is finished, the configured movement of the agent is
executed. Table 4.2 shows the main configuration parameters of
the simulation. In addition to these parameters, there are several
more configuration options depending on the selected behavior or
movement. The simulation is built in a modular way, such that
it can be easily extended to define new behaviors, movements,
environments, agents, or scheduling-schemes.
For evacuation, a layout of the building is read from an XML-file,
which is required to follow a specified standard. The building’s
layout has to be in form of a grid with variable dimensions and each
patch in this grid can be marked to be accessible to the evacuees
or to be part of a wall or barrier. Additionally, each patch of the
discretized layout is assigned to a node in the macroscopic evacuation graph, which is also specified in the XML input file. Edges
between nodes in the macroscopic evacuation graph are required
to be specified manually in the input file. The capacity of a node
c (n) is computed as the number of patches in a room. The length
of an edge l (e) is calculated as the Manhattan distance between
the two patches at the center of each room. The edge capacity
c (e) corresponds to the width of the connecting door, measured
in the number of patches. If there is no passage, i.e. sequence of
accessible patches, between two nodes which are connected via an
edge in the graph, the simulation responds with an error. The
same holds for an initial distribution of agents in rooms where
there is not enough space.

103

Chapter 4 Swarm Evacuation Planning with Mobile Devices

Table 4.2: Main configuration parameters for the Swarm
Simulation.
Parameter
Behavior
PopSize
SignalRange
SimCycles
Torus
NumInit
Positioning
Seed
Scheduling
Environment
Colliding
Movement

104

Description
Software program run on the mobile devices.
Number of agents in the simulation.
Communication range as percentage of the sidelength of the square environment.
Number of simulation cycles run before the
simulation is terminated.
Whether the environment is a torus world.
Number of beacons in the environment.
Initial distribution of agents in the environment.
Random seed used in this simulation run.
Order in which the agents are executed.
Design of the simultion environment.
Whether agents are colliding with each other.
The type of mobility model executed.

4.1 Macroscopic Swarm Evacuation Planning
Experiment Settings
In order to evaluate the results achieved with CC-SEP, a simple
evacuation scenario is simulated and the impact of several parameters, such as the number of evacuees or their distribution inside
the building are investigated. The objective of this experiment is
to examine the effectiveness of the proposed CC-SEP approach
and to investigate whether evacuation planning can be improved
by integrating information about other evacuees in the building,
even though this information is potentially faulty, delayed, and
incomplete. For this, the overall evacuation time is compared with
the time needed in an uncoordinated panic-like situation and a situation in which all evacuees escape using the shortest path leading
outside the building. Although one of the main motivations for
OBESS is the construction of increasingly complex buildings, the
investigated scenarios in this study are comparatively simple. It
should be noted, though, that any building layout can be divided
into small two-dimensional sections, which, for example, represent
one floor. Therefore, the insights gained in this study can be
directly transferred to more complex scenarios. Due to the general
absence of central control, all algorithms presented in this thesis,
including the CC-SEP, scale well to complex scenarios.
For the experimental study, an evacuation scenario is simulated in
a simple building, which is shown in Figure 4.3. The environment
is a square field, which is divided into 25 times 25 patches. One
agent can occupy exactly one patch. Patches which are occupied by
an agent or which are a-priori marked to be part of a wall or barrier
are not accessible for the agents during the evacuation simulation.
The simulation terminates when all agents have reached the exit.
The devices start computing their evacuation instructions after a
warm up time of 30 cycles, in which they are allowed to gather
information through communication. When an evacuation instruc-

105

Chapter 4 Swarm Evacuation Planning with Mobile Devices
tion is available, the agent consequently follows this instruction
by walking on the shortest path between its current patch and
the closest patch belonging to the next room of the instruction.
The shortest path in Manhattan distance metric between these
patches is computed with the A* Algorithm presented in Hart et al.
[83]. Agents are only allowed to move to an accessible adjacent
patch in their von-Neumann neighborhood (cf. Figure 4.1(a)), i.e.
all adjacent patches without diagonal patches. This restriction
ensures a constant traveling speed of the agent because the Euclidean distance traveled in each movement is constant. However,
if there is no free patch in the von-Neumann neighborhood but
in the Moore-neighborhood (cf. Figure 4.1(b)), i.e. all adjacent
patches including the diagonal patches, and this patch is closer in
Euclidean distance to the target patch than the patch where the
agent is currently located, the individual moves in two time steps
to this closer patch in the Moore neighborhood. This has the effect
that agents cluster in front of doors in an arc-like shape instead of
lining-up in front of it.

(a) Von-Neumann neigh- (b) Moore neighborhood
borhood

Figure 4.1: Different neighborhood models for movements in the
evacuation simulation.

106

4.1 Macroscopic Swarm Evacuation Planning
Two scenarios with different initial distribution of evacuees across
two rooms are considered. In Scenario A agents are initially located
at nodes 1 and 3 in the sample building shown in Figure 4.3. In
Scenario B, evacuees are initially located in rooms 1 and 2. Simulations are run with 50 and 100 evacuees. In order to test the impact
of evacuees’ distributions in the building, different experiments are
performed in which the percentage of agents located in room 1 is
decreased from 100% to 0% and the remaining evacuees start in
the second room of the considered scneario, respectively. Node 6
indicates the exit of the building and the communication range is
set to 5%, 10%, and 15%. It is defined as a fraction of the side
length of the square plane. With a communication range of 5% the
agent’s own patch and its von-Neumann-neighborhood is covered.
Setting r to 10% and 15% increases the communication range by
one patch each. Figure 4.2 illustrates the examined communication
ranges. The communication range r is assumed to be equal for
all devices. The influence of walls on the communication range is
neglected in the simulation. Panic behavior is simulated by letting
agents randomly select their next room, omitting doors which they
have already passed through unless there is no other option. All
experiments are repeated 40 times with different initial random
agent placement according to the constraints and the results are
averaged.
Scenario A
The first scenario serves as a basis for investigating the effectiveness
of CC-SEP. The main focus lies on whether local communication
can deliver enough information in order to improve the overall
evacuation time when compared to a shortest path approach or
panic behavior. In addition, the effects of different communication
ranges, the total number of evacuees, and their distribution across

107

Chapter 4 Swarm Evacuation Planning with Mobile Devices

(a) Continuous view

(b) Rectilinear view

Figure 4.2: Different communication ranges for the experiments.

Figure 4.3: Building plan and corresponding macroscopic
representation used in the experimental study.

108

4.1 Macroscopic Swarm Evacuation Planning
the building are examined. In Scenario A, agents are distributed
equally across the two rooms which belong to node 1 and 3 of the
macroscopic graph. This way, each agent has two possibilities to
reach the building’s exit, with the shortest path being: node 1,
node 3, node 5, and node 6. The alternative route is node 1, node
2, node 4, node 5, and node 6, or node 3, node 4, node 5, and node
6 respectively. In this experiment setting, CC-SEP is expected to
deliver good results because the agents can switch to a slightly
longer but less crowded evacuation route.
Figure 4.4(a) shows the average total time for evacuation in Scenario
A using 50 agents and varying the occupation of the two rooms.
For a communication range of 5%, the overall evacuation time is
only slightly improved by CC-SEP compared to the shortest path
behavior. This can be explained by the restrictive communication
abilities and, as a consequence, the lack of information which can
be used for CC-SEP. Due to the lack of information, most evacuees
assume room 1 and 3 to be empty and, thus, choose the shortest
path towards the exit. By increasing the communication range,
however, the total time for evacuation decreases and CC-SEP starts
to deliver significantly better results because evacuees now switch
to the longer but less crowded paths. By setting the communication
range to 15% the total time for evacuation can be reduced by up
to 30% if all agents are located in either room 1 or 2, which are
the scenarios with highest congestion potential. However, the
reduction in evacuation time when increasing the communication
range from 10% to 15% is not significant anymore, which implies
that a communication range of 10% is already sufficient to gather
most information about the evacuation situation. In this scenario,
the panic behavior exhibits the worst evacuation time.
Figure 4.4(b) displays the results for 100 agents. From this experiment, it can be observed that the shortest path behavior produces

109

Chapter 4 Swarm Evacuation Planning with Mobile Devices

(a) 50 agents.

(b) 100 agents.

Figure 4.4: Comparison of the total time for evacuation with only
50 (a) and 100 agents (b) in Scenario A applying
CC-SEP, shortest path, and panic behavior.

110

4.1 Macroscopic Swarm Evacuation Planning
similar overall evacuation times when compared to the panic-like
behavior. In the case of agents being quite equally distributed
across both rooms at the beginning of the evacuation, a panic-like
behavior can yield even lower total evacuation times. The reason
for this effect is that with 100 agents in the building, the congestion
on the shortest path has become so large that randomly searching
for an exit results in a faster evacuation process when compared to
enduring the waiting time in the congestion. With CC-SEP and a
communication range of 15% the evacuation process is up to 40%
faster than in the shortest path and the panic-like behavior, which
is a higher improvement when compared to the same scenario with
50 agents inside the building. Here, the difference between shortest
and quickest path, which is emphasized at the beginning of this
section, becomes clearly apparent. So far, it can be concluded that
providing information about the shortest path to the exit is better
for most investigated situations when compared to evacuees blindly
searching for their way out. Nevertheless, leading all evacuees on
the shortest path can lead to congestions, which can result in an
evacuation performance which is even worse than the one in a
panic situation. It is shown that CC-SEP, on the other hand, is
able to collect and use information about other evacuees’ locations
in the building and, thus, reduce the overall evacuation time, even
though this information has to be collected over uncertain links of
an ad hoc network. Similar to the scenario with 50 agents, the difference between evacuation times produced by CC-SEP with 10%
communication range and a 15% range is not significant, however,
the results for a 5% communication range are noticeably different.
With 5% and a high concentration of agents in one of the two
initial rooms, the performance of CC-SEP is similar to the results
with 10% and 15%, but the more equally the agents are distributed,
the worse the results get. When agents are distributed across both

111

Chapter 4 Swarm Evacuation Planning with Mobile Devices
rooms, 50 agents are located in each room. Hence, the scenario is
similar to the first experiment where all agents are concentrated in
room number 1, except for the fact that now additional 50 agents
are located in room 3. Due to the small communication range, the
devices, again, assume that the shortest path is occupied by only
a few other evacuees and choose this path for evacuation. The
overall evacuation time is now even higher than for the scenario
with 50 agents because of the additional 50 evacuees in room 3,
who are the reason for the increase in the time it takes for evacuees
in room 1 to reach the exit on the shortest path.
Scenario B
In Scenario B, the path options for some evacuees are reduced.
When following the shortest path, agents in room 1 leave the
building over node 1, node 3, node 5, and node 6, while agents
on node 2 pass by node 4, node 5, and node 6. Agents from the
room corresponding to node 2 have only one path leading to the
exit, while agents in room 1 can still select between two choices.
As a consequence, it depends on the amount of agents in room 2
whether or not it is advisable for agents of room 1 to switch to
the route with a longer distance to the exit. With this scenario, it
is examined whether CC-SEP is capable of recognizing the right
choice for agents in room 1.
In Figure 4.5(a), the results for 50 agents are displayed, which
basically confirm the observations from Scenario A. The shortest
path behavior causes a slightly higher evacuation time compared to
CC-SEP with a communication range of 5%. When using a wider
communication range of 10% or 15%, the evacuation time is further
reduced by up to 30%. When more than half of all evacuees are
initially located in room 2, CC-SEP does not yield any additional
performance improvement when compared to the shortest path.

112

4.1 Macroscopic Swarm Evacuation Planning
The reason for this is that due to the accumulation of agents in
room 2, the escape route node 2, node 4, node 5, and node 6 does
not present an attractive path choice for agents in room 1. The
shortest path route node 3, node 5, and node 6 is more attractive
for agents in room 1. Since the design of the evacuation graph does
not provide any other choice of route for evacuees in room 2 than
the path over node 4, node 5, and node 6, CC-SEP does not bring
any additional benefit compared to the shortest path behavior in
case of such a distribution. This example shows how important
the choice of the macroscopic building model is for the evacuation
performance with CC-SEP.
In Figure 4.5(b), the experimental results with 100 evacuees are
depicted. When 100 evacuees are distributed equally across the two
starting rooms, CC-SEP becomes slightly worse than the shortest
path behavior for a 5% communication range. To understand
the reason for this effect, it has to be noted that any deviation
from the shortest path increases the evacuation time in a scenario
with precisely as many evacuees in room 1 and room 2. This is
because any evacuee which deviates from the respective shortest
path would block evacuees on the alternative path due to the
symmetrical layout of the graph. With a 5% communication range,
the chance for evacuees in room 1 to receive no information about
evacuees in room 2 is higher when compared to a situation with a
10% or 15% communication range. Hence, evacuees are more likely
to choose route node 2, node 4, node 5, and node 6 than with a
larger communication range. In contrast to Scenario A, where the
panic behavior received better results compared to the shortest
path behavior, this is not the case in Scenario B. The reason for
this is that in Scenario A, the shortest path always contains node
3 where the agents meet, while in Scenario B this happens the

113

Chapter 4 Swarm Evacuation Planning with Mobile Devices

(a) 50 agents.

(b) 100 agents.

Figure 4.5: Comparison of the total time for evacuation with 50
(a) and 100 agents (b) in Scenario B applying
CC-SEP, shortest path, and panic behavior.

114

4.1 Macroscopic Swarm Evacuation Planning
earliest in the room associated to node 5, where the capacity is
higher.
Initialization Time and Repetition
So far, CC-SEP is computed after an initialization time of 30 cycles,
in which the agents are allowed to collect information. Although
this is a sound presupposition, there could be scenarios in which
such a time is not given, thus, it is interesting to investigate the
influence of initialization time on the algorithm’s performance.
All following experiments are carried out with 100 agents and a
communication range of 15%. The results for panic-like behavior
are omitted since it has been shown that, on average, they are
significantly worse than evacuation times achieved with CC-SEP
or the shortest path following behavior. Figures 4.6(a) and 4.6(b)
display the results with and without initialization time. If CC-SEP
starts immediately, it means that information can only be gained
from devices which are one hop away before the computation of
evacuation instructions begins. The results confirm the intuitive
expectation that providing for a time period to initialize and, thus,
collect information, improves the overall evacuation time. Of course,
this is not the case if there is no possibility to choose between
paths, such as in Scenario B when most agents start at node 2. In
addition, when agents are concentrated in one room, initialization
time does not improve the evacuation time significantly because
most information can be exchanged in few rounds.
When aiming at building an adaptive evacuation system, one has
to allow for the computation of new evacuation instructions in
order to include recent information. Such adaptive route planning
is achievable when CC-SEP is computed repeatedly. Figures 4.7(a)
and 4.7(b) show the results for experiments where CC-SEP starts

115

Chapter 4 Swarm Evacuation Planning with Mobile Devices

(a) Scenario A.

(b) Scenario B.

Figure 4.6: Results for evacuation time with (I) and without (nI)
initialization time of 30 cycles.

116

4.1 Macroscopic Swarm Evacuation Planning
without initialization time and is repeated in every simulation cycle,
while taking the most current information into account.

(a) Scenario A.

(b) Scenario B.

Figure 4.7: Results for evacuation time with (R) and without (nR)
repetition of computing evacuation instructions when
allowing for an initialization time (I).

117

Chapter 4 Swarm Evacuation Planning with Mobile Devices
Figures 4.8(a) and 4.8(b) illustrate the experimental results when
CC-SEP is repeatedly executed after an initialization time of 30
cycles. It can be observed that, when there is little time to collect
information before the instruction is computed, repetitions can
help to reduce the evacuation time in certain cases. However, when
many agents are located at node 1, repeating CC-SEP prolongs the
evacuation process, especially, when time for information collection
is given to the agents before the evacuation starts. There are two
reasons why repeated execution of CC-SEP can worsen the results.
Firstly, agents often change between evacuation instructions and
lose time in doing so. The second reason lies within the design of the
information exchange procedure. In CC-SEP, only the number of
evacuees in each room of the building is exchanged over the network
and not the IDs of the respective evacuees. Hence, information
about other rooms in the building is only available in an aggregated
and anonymized manner. Because of this design decision, it can
occur that evacuees which change rooms are counted multiple times
in different rooms before the age-value of the obsolete information
expires and the number of known evacuees in the respective previous
rooms is reduced. Although this is an undesired effect, repetition
of evacuation planning is necessary for an evacuation management
system to be adaptive. Consequently, it is a positive observation
that the evacuation time is still well below the evacuation time
needed when all agents take the shortest path.
Approaches for Reducing Frequent Decision Switching
While the problem of obsolete information is hard to deal with, a
high frequency of target changes caused by repeated execution of
CC-SEP can be counteracted. To do so, two variations of CC-SEP
are being tested subsequently. Agents are, now, only allowed to
recompute instructions after a certain patience period after a target

118

4.1 Macroscopic Swarm Evacuation Planning

(a) Scenario A.

(b) Scenario B.

Figure 4.8: Results for evacuation time with (R) and without (nR)
repetition of computing the evacuation instructions
without allowing for initialization time (nI).

119

Chapter 4 Swarm Evacuation Planning with Mobile Devices
change or if they are more than a maximum distance away from
their current target. The patience period prevents agents from
changing targets too frequently and the maximum distance avoids
target changes shortly before the current target is reached. The
results for a set-up in which 100 agents start in the room belonging
to node 1 are depicted in Figure 4.9(a) and 4.9(b). As previously
defined, distances are denoted as a fraction of the side length of
the square test environment. The patience period does not have a
significant impact on the overall evacuation time, which indicates
that frequently switching targets is not the main problem when
executing CC-SEP repeatedly. Nevertheless, it might be helpful to
introduce a certain patience period before rule changing is allowed
not only to reduce total evacuation time, but also to avoid creating
mistrust and frustration with the system amongst the users due to
a high frequency of decision changes. It is shown that prohibiting
changes in evacuation instructions for agents which are close to
their next target leads to an improvement of evacuation time if the
maximum distance is sufficiently large. However, if the maximum
distance is chosen too large, this can lead to a situation where
agents do not change instructions anymore. As a consequence,
the adaptability of the system is lost. It can be concluded from
these results that the slight performance decrease when computing
CC-SEP repeatedly is likely to be caused by the design of the
information exchange as aggregated, anonymized numbers.
To conclude the experimental study, the robustness of CC-SEP is
tested. For this, 100 agents are located in room 1 and the amount
of rule following agents is reduced gradually. Evacuees who do
not follow the instructions provided by CC-SEP are assumed to
be taking the shortest path to the exit instead. In addition to
this robustness test, the evacuation behavior with CC-SEP is
investigated in a more complex environment. Figure 4.10 depicts

120

4.1 Macroscopic Swarm Evacuation Planning

(a)

(b)

Figure 4.9: Total evacuation time when repeating CC-SEP after a
certain patience period after changing targets (a) and
when limiting repetitions to agents which are a
maximum distance away from their target (b).

121

Chapter 4 Swarm Evacuation Planning with Mobile Devices
the more complex and realistic Scenario C, in which 100 evacuees
are distributed according to a uniform random distribution across
all rooms of the building. Node 16 represents the exit of the
building.
Robustness
Figure 4.11(a) shows the results of the robustness test in the simple
scenario. Without repeating the computation of evacuation instructions, evacuation time increases linearly with a decreasing number
of rule-following agents until it is reduced to the performance of
the shortest path behavior. For repeated computations, the evacuation time even decreases slightly with a higher percentage of
non-followers and there is a tolerance for up to only 40% of rulefollowing agents before the overall evacuation time is worse than
the scenario in which all agents follow the instructions provided by
CC-SEP. Unfortunately, a similar behavior cannot be observed in
Scenario C, which indicates that this result is most likely due to the
concentration of agents in one room or due to the construction of
the macroscopic evacuation graph. However, the important result
is that in both scenarios evacuation time does not increase over
proportionally when the number of rule-followers is reduced by less
than 30%. This shows that CC-SEP is quite robust even when
some evacuees deviate from the evacuation instructions. Furthermore, these results suggest that the deterioration of performance
due to a repeated execution of CC-SEP is limited to few specific
cases. With a uniformly random initial evacuee distribution and
the more realistic Scenario C, the deterioration of performance
when CC-SEP is executed repeatedly is negligible compared to the
previous test set-up.

122

4.1 Macroscopic Swarm Evacuation Planning

Figure 4.10: Building plan of Scenario C and the corresponding
macroscopic graph representation.

123

Chapter 4 Swarm Evacuation Planning with Mobile Devices

(a)

(b)

Figure 4.11: Robustness test of CC-SEP with 100 agents placed in
room 1 of Scenario A (a) and with 100 agents placed
randomly inside the building of Scenario C (b).

124

4.1 Macroscopic Swarm Evacuation Planning

4.1.3 Conclusion
In this section, an SEP algorithm based on capacity constrained
routing is presented, which can be used to estimate the quickest
evacuation path based on uncertain information. Information
about the distribution of evacuees in the building is gained via
local communication by using ad hoc network connections between
mobile devices. A modification of the CCRP algorithm published
by Lu et al. [142, 143] is introduced, which estimates waiting times
on each path using a time series model in order to evaluate the
different path options for each evacuee. The main challenge for
any SEP algorithm is the lack of certain global knowledge. Hence,
experiments are performed in order to determine whether CC-SEP
is able to improve the total evacuation time when compared to
simply offering the shortest path as a navigation instruction to each
evacuee. In experiments, varying communication ranges (5%, 10%
and 15%), different numbers of evacuees in the building (50, 100),
as well as variations in their initial distribution over the rooms of
the building are investigated. Additionally, two different building
layouts are considered. The first layout is specifically designed to
cause congestions and the second layout is a more realistic and
complex building layout with a large number of different rooms
and valid paths towards the exit. The performance of CC-SEP is
compared to a shortest path following behavior of evacuees and a
panic model, in which evacuees are assumed to have no knowledge
about the building’s layout. Furthermore, different variations of
CC-SEP are subject to examination, such as the introduction of
an initialization time where mobile devices are allowed to collect
information, an iterative repetition of evacuation planning, as well
as a patience period and a maximum target distance which have to
be fulfilled in order to allow for changes in evacuation instructions.

125

Chapter 4 Swarm Evacuation Planning with Mobile Devices
Finally, the robustness of CC-SEP against rule deviating evacuees
is considered.
The results confirm that knowledge collected via ad hoc network
connections significantly improves evacuation time in almost all
scenarios considered by up to 40% when compared to the shortest
path behavior, even though the local view on the real evacuation
situation is mostly incomplete and sometimes even incorrect. When
comparing the results with the panic-like behavior of evacuees,
the improvement is even higher. Increasing the communication
range from 5% to 10% provides significantly more information
for CC-SEP and, hence, improves the evacuation time, while an
increase from 10% to 15%, yields only slightly better results. The
positive effect of CC-SEP on the evacuation time is stronger when
there are more evacuees in the building, when they are concentrated
in few rooms, and when there are multiple alternative paths towards
the exit. This is due to an increased congestion potential, which
CC-SEP successfully avoids. Evaluation of Scenario B reveals that
the right choice of a macroscopic evacuation model is crucial for
good performance of CC-SEP and that it is important to provide
many alternative paths in order to exploit the full potential of
CC-SEP. Especially the direction of edges in the graph is to be
selected carefully. In this context, it can be useful to provide for
the possibility to change directions of edges in the graph model
dynamically depending on the expected distribution of evacuees
across the building. In the work of Sangho et al. [204], an approach
for reconfiguring macroscopic graph edges is presented, which could
be applied to this problem. The O/C Architecture of the building’s
CCU (cf. Section 3.1.1) could be used to realize such a dynamic
adaptation of the macroscopic graph model.
Providing for some initialization time in order to collect information
is shown to improve the results of CC-SEP. When repeating

126

4.1 Macroscopic Swarm Evacuation Planning
evacuation planning in each cycle, however, evacuation time slightly
increases. The reason for this is identified to be, most likely, the fact
that information is disseminated over the network in an aggregated
form, i.e. the number of devices in a room is communicated instead
of their individual locations and, hence, outdated information is
used in the evacuation planning process. However, the performance
still exceeds the shortest path behavior in a vast majority of the
considered scenarios. It is further shown to yield even better
results compared to the scenario without repetition when there is
no initialization time given to the evacuees. While introduction
of a patience period does not lead to significant improvements,
limiting the allowed route changes to situations in which evacuees
are a certain distance away from their current target is shown to
overcome slightly deteriorated results with planning repetitions.
Furthermore, performance reductions resulting from a repeated
execution of the evacuation planning are negligible in the more
realistic scenario. Finally, CC-SEP is shown to be robust against up
to 30% of rule deviators and the induced performance decrease does
not increase overproportionally with the number of rule deviating
evacuees. A possibility to further improve results of CC-SEP could
be to distinguish between having no information about a path and
having the information that the path is empty. The latter case
could be deduced from knowing that there are many evacuees close
to the respective path which do not report about anybody using
it. Additionally, introducing a reliability measurement based on
message delays, which are encoded in the age-value, could lead to
potential improvements of CC-SEP.

127

Chapter 4 Swarm Evacuation Planning with Mobile Devices

4.2 Dynamic Multi-objective Swarm
Evacuation Planning
Distributed and decentralized evacuation planning on mobile devices with CC-SEP, presented in the previous section, is shown to
work effectively in speeding-up the evacuation process compared
to a situation in which the devices do not use a MANET for exchanging information. However, taking a macroscopic evacuation
graph model as a basis for optimization and using the presented
information model has several constraints:
1. Locations of evacuees are abstracted to be at the center of
the corresponding room.
2. Edges in the macroscopic graph model are unidirectional,
excluding feasible evacuation routes from planning.
3. Evacuees are assumed to be homogeneous and individual
characteristics of the evacuees are not considered.
4. Waiting time on a path is assumed to be increasing linearly
with an increasing amount of evacuees using the path.
5. The communication model in CC-SEP does not keep track
of the locations of specific evacuees, but rather operates on
aggregated and anonymized data.
These restrictions can lead to amendable path planning due to
the high level of abstraction and loss of information. Due to
the abstract nature of the macroscopic graph model, for example,
different locations of evacuees within a single room are neglected
by evacuation planning. Furthermore, unidirectional edges in the
graph can lead to path planning which ignores potential escape

128

4.2 Dynamic Multi-objective Swarm Evacuation Planning
routes even if they are short and empty. Additionally, homogeneous
evacuees, as well as linearity of waiting times in congestions are
far from reality (cf. Helbing et al. [90]). Another problem with
CC-SEP lies within its communication model. The reduction of the
exchanged information to the number of evacuees per room instead
of their exact locations has the advantage that messages are kept
short. However, the drawback of this feature is that the locations
of single evacuees are untraceable. This can lead to multiple
counting of the same evacuee in different rooms and can distort the
information input for evacuation planning. This has shown to have
a slightly negative effect when CC-SEP is executed repeatedly.
This is not desirable since evacuation planning is supposed to
be adaptive to new situations, such as blocked passages or the
emergence of congestions for which repeated computations of paths
are essential. For these reasons, it is worth looking into alternative
possibilities to perform SEP on mobile devices.
Inspiration for an alternative approach to the evacuation planning presented in the previous section can be found in the area of
robotics, where path planning, or motion planning as it is often
referred to, is a well-studied task (cf. for example Choset [41] or
LaValle [132]). Motion planning is often performed on the basis
of a grid-like, discretized map of the environment, which is made
available to the robot. The patches of this grid correspond to the
size of the robot and represent nodes in a graph and neighboring
patches are connected via edges. The resulting graph model is
more detailed when compared to the macroscopic graph evacuation model used in CC-SEP. In addition, locations and distances
are represented more realistically. This graph model is used for
motion planning by applying a shortest path algorithm. In the
work of König and Likhachev [125], for example, an algorithm for
robot motion planning is presented, which can be used to compute

129

Chapter 4 Swarm Evacuation Planning with Mobile Devices
lowest cost paths based on such a discretized map. This approach
is especially interesting for evacuation path planning, since it is
specifically designed for dynamic environments, where changes in
traveling costs become available to the robot while it moves towards
its target. Evacuation planning on the basis of information from
communication in an ad hoc network has similar characteristics.
The mobile devices constantly receive new information during the
evacuation process. This newly available knowledge should, ideally,
be integrated in the path planning process in order to adjust the
evacuation instructions if reasonable. Only under these circumstances can the evacuation planning be adaptable to changes in the
environment, such as blocked passages or emerging congestions.
Apart from the constraints mentioned previously, another limitation of CC-SEP has to be addressed. Since the path planning is
based on the CCRP algorithm (cf. Lu et al. [143]), it only optimizes
for evacuation time and is inflexible to integrate other objectives.
While the circumvention of potential congestions is an obvious goal
when choosing an evacuation route, there are other criteria that
should be considered. For example, the avoidance of risky areas
can be crucial for the safety of evacuees. Moreover, the personal
preference for such evaluation criteria can be different for each evacuee. For example, people with physical disabilities would probably
prefer to take a shorter path or may be willing to take a longer
path if it is better accessible. Evacuation planning should, ideally,
be capable of respecting individual characteristics and preferences,
as well as multiple objectives. The Dynamic Multi-objective Swarm
Evacuation Planner (DMO-SEP) introduced hereafter overcomes
some of the aforementioned limitations of CC-SEP and respects
user preferences during path planning. It is designed to optimize
for multiple criteria simultaneously when searching for an optimal
escape route for a specific evacuee. DMO-SEP is based on the pre-

130

4.2 Dynamic Multi-objective Swarm Evacuation Planning
viously described dynamic robot motion planning algorithm, called
D* Lite, which is introduced by König and Likhachev [125]. The
algorithm is described briefly in the next section, before DMO-SEP
is presented in detail and tested in an experimental study.

4.2.1 D* Lite
D* Lite is based on a graph representation G = (N, E) of the
environment which differs from the macroscopic evacuation graph
model. The environment is divided into square patches in a grid-like
fashion, such that each patch corresponds to the expansion of an
average person, and each patch represents a node n ∈ N of graph
G. Neighboring patches in the von-Neumann neighborhood (cf.
Figure 4.1(a)) are connected via edges E : (N × N ) in the graph
representation and each edge has traveling costs c (n, m) assigned
to it, representing the costs for traveling from patch n ∈ N to
m ∈ N . D* Lite computes the path with minimal total costs from
a specific starting patch ns ∈ N to a target patch nt ∈ N . The
starting patch corresponds to the current location of the robot.
The algorithm starts at the target patch and from there looks for
surrounding patches with minimal traveling costs to the target
patch similar to the well-known Dijkstra algorithm. The search is
directed towards the starting patch by means of a heuristic value
h (n), which denotes the Euclidean distance between each patch n
and the robot’s current position. Patches with a lower value for
h (n) are preferred in the search for the lowest cost path, which
can reduce computational costs due to a faster termination of the
search. This approach was first proposed by Hart et al. [84] in
form of the A* algorithm designed for static path finding.
Let Adj (n) return all adjacent nodes of a node n ∈ N with respect to its von-Neumann neighborhood and d (n) denote the

131

Chapter 4 Swarm Evacuation Planning with Mobile Devices
minimal traveling costs from n to nt , d¯(n) is an auxiliary variable. In D* Lite, patches which have to be processed in order
to find the lowest cost path are added to a sorted set U . All
nodes in this set
 are sorted in ascending order with respect to
key1 (n) = min d (n) , d¯(n) +h (n) and, subsequently, key2 (n) =


min d (n) , d¯(n) . Let f irst (U ) denote the first element in the
set U , i.e., the element with the smallest key value. The complete
procedure of D* Lite is described in Algorithm 4.3.
If a change in traveling costs is detected for an edge e (n, m), for
example because one of the corresponding patches is blocked, the
update procedure is called for nodes n and m and for all nodes which
are subsequently added to U , recursively, until U is empty or the
starting patch is updated and consistent. Consistent means that the
values for d (n) and d¯(n) are equal. The consistency check in line 5
of the update procedure limits the propagation of cost changes to
affected nodes and reduces the computational costs of the algorithm
when compared to replanning from scratch. Figure 4.12 illustrates
the update process with a simple example. The top row displays
the discretized environment consisting of 9 patches; the start patch
of the robot is marked with a circle and the target with a cross.
The bottom row shows the corresponding graph representations
using nodes (dots) and connecting edges. The traveling costs are
assumed to be 1 for all edges in the graph and the numbers in the
nodes denote the minimum travel costs d (n) from that node to
the target node. In the second state of the environment, the center
patch is discovered to be blocked, therefore, the connected nodes
are updated according to the procedure described in Algorithm
4.3. The gray nodes are found to be consistent, i.e., their distance
costs are not affected by the change, whereas the shaded, red node
is inconsistent and requires an update. Since the red node is the

132

4.2 Dynamic Multi-objective Swarm Evacuation Planning

Algorithm 4.3 D* Lite
Require: graph G = (N, E), start node ns , target node nt , heuristic h (n) ∀n ∈ N , costs c (e) ∀e ∈ E
1: d (n) ← ∞, ∀n ∈ N
2: d¯(n) ← ∞, ∀n ∈ N , d¯(nt ) ← 0
3: U.add (nt )
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:

procedure computePath


while (key (f irst (U )) < key (ns )) ∨ d¯(ns )
= d (ns ) do
Define u ← f irst (U )
if d (u) > d¯(u) then
d (u) ← d¯(u)
update (u)
end if
for p ∈ Adj (u) do
update (p)
end for
end while
end procedure
procedure update(n)
if n
= nt then
d¯(n) ← min∀s∈Adj(n) (c (n, s) + d (s))
U.add (n)
if d (n) = d¯(n) then
U.remove (n)
end if
end if
end procedure

133

Chapter 4 Swarm Evacuation Planning with Mobile Devices
starting node and it is consistent after the update, the process
terminates and the new cost minimal path is found by following
the nodes which have the lowest costs assigned. In this example,
there are two valid lowest cost paths, of which one is selected at
random and depicted. From this example, the advantage of D*
Lite becomes apparent. In contrast to updating all nodes in the
graph after a change in the environment is detected, only the nodes
which are affected require an update.

Figure 4.12: Update process of D* Lite when a patch is discovered
to be blocked.

4.2.2 Dynamic Multi-objective Swarm Evacuation
Planning
For DMO-SEP, each device is assumed to possess knowledge of its
current location and of the building’s layout, e.g., by downloading it
upon entering the building. Similar to CC-SEP, DMO-SEP relies
on an exchange of information between mobile devices via ad hoc

134

4.2 Dynamic Multi-objective Swarm Evacuation Planning
network links and performs decentralized path computations on
each device. It, therefore, belongs to the family of SEP algorithms
as defined in Section 4.1. The information from local communication is used to assign costs to the edges of the graph model
before applying the dynamic path planning algorithm D* Lite (cf.
Section 4.2.1) to find an optimal path towards the exit. Since
route planning with D* Lite is solely based on costs instead of an
estimation of traveling and waiting times, the approach is flexible
to incorporate multiple objectives. This can be achieved by using
a complex cost function which consists of multiple components
o ∈ O, such as risk, distance, or waiting time, each representing a
different optimization objective. These components are weighted
and summarized in order to derive the total traveling costs for
evacuee a, which can then be used for D* Lite:
ca (e) =



(woa · cao (e))

(4.1)

o∈O



with weights woa ∈ [0, 1] and O woa = 1. The weights can be
defined according to the characteristics and preferences of each
evacuee. After the costs for all edges in the discretized layout have
been computed, D* Lite finds the minimum cost path, which is a
sequence of patches < ns , ..., nt >. From this path, a sequence of
rooms is generated, by analogy with the evacuation instructions in
CC-SEP, and a navigation path can be displayed to the user on the
screen of the mobile device. The transformation from patches to a
room-based path is not necessary, but it ensures that the displayed
path is easily conceived by the user in contrast to a path which
assembles a winding sequence of patches.
To illustrate and test the functionality of DMO-SEP, three optimization criteria are considered. Since a straightforward objective
for evacuation planning is to reduce the traveling distance, the dis-

135

Chapter 4 Swarm Evacuation Planning with Mobile Devices
tance traveled constitutes one of the cost components. These costs
cdistance (e) arise for each edge e ∈ E in the graph model G and
represent the distance between the patches which are connected
by the respective edge. They can be made available to the mobile
devices along with the download of the building’s layout. Although
they are initially equal for all devices, it is conceivable that an
evacuee detects a blocked passage during his escape, reports this to
his mobile evacuation device, and this information is then spread
to other mobile devices using the ad hoc network communication.
In such a case it is possible that, due to the delays and potential network link breakages, distance costs for the patches in the
graph model diverge between mobile devices. Apart from traveling distance, two other objectives are considered when choosing
evacuation routes, the minimization of risk and the minimization
of waiting time on the paths. Since jamming queues in front of
narrow passages or doors are the main reason for waiting time, the
minimization of waiting time is considered equivalent to congestion
avoidance.
Risk Minimization
In order to be able to avoid risky paths, the devices need to know
which areas are dangerous. It is reasonable to assume that there are
sensors installed inside the building which can measure potential
danger indicators, such as gas or heat, and forward this information
to mobile devices nearby. Such a scenario is, for example, examined
by Filippoupolitis et al. [66]. For simplicity reasons, it is assumed
that the sensors communicate a certain risk level risk (R) ∈ [0, 1]
for room R instead of empirical measurement data which would
require a certain degree of interpretation. The corresponding costs
are assigned to all edges e (n, m) , n ∈ R connecting nodes in the
respective room:

136

4.2 Dynamic Multi-objective Swarm Evacuation Planning

crisk (e (n, m)) = risk (R) : n ∈ R

(4.2)

By assigning a risk value to each patch in a room, it becomes
comparatively more expensive to cross a larger room with the same
risk level. The information about the risk level of a room is shared
with other devices using ad hoc network connections. To do so,
each message contains an age-value which is set to zero at the
moment at which a risk level is reported from a sensor to a mobile
device. Then, the age-value is increased according to the system
clock of the mobile devices. When the information is forwarded
to other mobile devices, the age-value is included in the message.
Whenever a device receives contradictory information about the
risk level of a room from two different sources, it can identify and
adopt the most recent information.
Congestion Avoidance
Congestion avoidance is a well-known task in the research area of
traffic optimization. However, there is a major difference between
traffic and evacuation scenarios. While traffic jams are usually
rather well-organized situations, in which cars are lined up behind
red traffic lights or other barriers, evacuation is often accompanied
by panic. In panic situations, however, the time needed for evacuees
to traverse a narrow passage is not easy to estimate. Firstly, it does
not increase linearly with the number of waiting evacuees or the size
of the door, but is rather dependent on many other factors which
affect the forces that act on the evacuees and, hence, determine the
time needed for a congestion to dissolve (cf. Helbing et al. [90]).
This kind of situation is further aggravated when information about
the locations of all evacuees is uncertain. Similar to CC-SEP, the
mobile devices in DMO-SEP are assumed to communicate their

137

Chapter 4 Swarm Evacuation Planning with Mobile Devices
own location to other nearby devices and spread such information
in the ad hoc network. However, the main difference is that
now all known locations of other devices are reported as opposed
to solely reporting the total number of observed devices in each
room. Analogously to the exchange of risk messages, each message
concerning location information has an age-value assigned to it,
which is zero if the device’s own location is being reported and
which is increased in accordance with the device’s system clock
in any other case. Due to the delays and link breakages, which
are likely to occur in the communication network, the knowledge
about other evacuees’ positions is uncertain. As a consequence,
predicting waiting times on this basis is prone to error. Fortunately,
it is sufficient for evacuation planning to be able to compare two
routes with respect to their potential for congestion emergence,
which is why a precise prediction of waiting times for each path is
not necessarily required. In order to measure a route’s potential for
congestion emergence, two indicators are proposed. These can be
computed from the locations of evacuees in the building, and are
intuitively suitable to evaluate congestion potentials on evacuation
paths. Both indicators are introduced hereafter and tested for their
effectiveness in a subsequent experimental evaluation.
Load The first congestion indicator is based on the idea that
jamming queues are more likely to occur when a large number of
evacuees are located in a relatively small room. The potential for
such a situation to occur is measured by a parameter called load,
which is constituted of the number of devices d ∈ D in relation to
the size of the room R in which their location is determined. The
associated costs for each edge e (n, m) between nodes n and m in
room R are computed as:

138

4.2 Dynamic Multi-objective Swarm Evacuation Planning

cload (e (n, m)) =

|{d ∈ D : n (d) ∈ R}|
|{n ∈ R}|

(4.3)

with n (d) denoting the node on which device d is located. One
could argue that using the number of devices in relation to the
size of the room’s doors as an indicator would be more effective,
but a room can have multiple doors and yet usually only one of
them determines the room’s flow rate. Since it is hard to tell which
doors will be used by how many evacuees, load as defined above
seems the more general indicator for congestions.
Entropy Congestions are so-called emergent phenomena, i.e., formations of order from disorder based on self-organization. Emergence can be measured by applying the concept of entropy, a metric
to measure the amount of order in a system, as suggested by Mnif
and Müller-Schloer [162]. A system with high order corresponds to
a low entropy value and vice versa. When congestions arise in an
evacuation situation, it is usually because evacuees jam up in front
of doors or narrow exits. As a consequence, the distribution of the
locations of mobile devices in the affected room is concentrated in
front of the cause for congestion. Entropy can be used to describe
this degree of concentration as explained in the following. Patches
in a room can be organized in rows and columns reflecting their
horizontal and vertical order respectively. Let a room R contain
x × y patches, i.e., x columns and y rows. Let further num (i),
num (j) denote the number of devices d ∈ D located on column
i or row j respectively. The entropy of a room R is computed as
shown in Equation 4.4 and 4.5, with x (n), y (n) denoting the row
and column of a patch n, respectively. The value p (n) denotes the
relative frequency of evacuees on patch n of room R.

139

Chapter 4 Swarm Evacuation Planning with Mobile Devices

e (R) = −



p (n) ld p (n)

(4.4)

n∈R

num (x (n)) + num (y (n))
p (n) = x y
i=1
j=1 num (i) + num (j)

(4.5)

Figure 4.13 shows an example of the calculation of a room’s entropy.
The locations of devices are marked with four rings inside the
respective patches. Firstly, the devices in each row and column are
counted. Then, the sum of the corresponding values is assigned to
each patch. In the second step, the relative frequencies p (n) are
computed for each patch and the entropy value e (R) of room R
is derived. The entropy value can vary significantly for different
room sizes. To make rooms comparable, the entropy is normalized
as follows. The maximum entropy emax (R) = ld (x · y) for a room
R is reached when each patch is occupied by one evacuee since
the size of each patch corresponds to the expansion of an average
person (cf. Section 4.2.1). The minimum entropy value emin (R) =
2
ld (x + y) − x+y
is achieved when only one single evacuee is located
somewhere in the room. Since entropy decreases with increasing
concentration of devices, the entropy based cost component for
edge e (n, m) with n ∈ R is computed as:


centropy (e (n, m)) = 1 −



e (R) − emin (R)
,n ∈ R
emax (R) − emin (R)

(4.6)

4.2.3 Evaluation
In order to assess whether DMO-SEP can accelerate the evacuation
process, a simulative experimental study is performed. By analogy
with the evaluation of CC-SEP, agents correspond to evacuees

140

4.2 Dynamic Multi-objective Swarm Evacuation Planning

(a) Counting devices.

(b) Computing entropy.

Figure 4.13: Example of computing the entropy of a room with
sixteen patches and four evacuees (patches with
rings).

141

Chapter 4 Swarm Evacuation Planning with Mobile Devices
and their mobile devices. The first aim is to verify whether the
reduced abstraction of the underlying evacuation model helps to
improve evacuation time compared to the macroscopic graph based
CC-SEP. Furthermore, the proposed risk avoidance mechanism is
to be tested and it is subject to investigation whether the presented
congestion indicators are suited to compare routes with respect to
their congestion potential and, hence, to avoid jamming queues
and to facilitate a faster evacuation of the building. The simulation environment used in the experimental study is the same as
described in Section 4.1.2. The investigated building layout is the
same as Scenario C used for evaluation of CC-SEP, which is the
more complex building layout shown in Figure 4.10. The rooms
with number 4, 6, and 10 connect the other rooms behind to the
room with number 12, which leads directly to the building’s exit.
Hence, agents initially located in rooms behind these potential
bottlenecks are likely to get caught up in congestions. Due to this
characteristic and because there are various valid escape routes,
this layout is suitable to test the effectiveness of the congestion
indicators. Unless stated otherwise, all experiments are performed
with 60 agents and are repeated 40 times, randomly varying the
initial distribution of agents in the building. The initial locations
are chosen randomly according to a uniform distribution. The
communication range is set to r = 15%, which covers the agent’s
current patch plus three patches in horizontal and vertical direction
respectively (cf. Figure 4.2). Walls or barriers do not interfere
with the communication range. For evaluation, total and average
evacuation time is measured as well as the average time spent in
risky areas or without movement. The latter is denoted as waiting
time. Naturally, a significant difference between total and average
evacuation time is to be expected in most scenarios because total
evacuation time reflects the maximum time any agent needs to

142

4.2 Dynamic Multi-objective Swarm Evacuation Planning
leave the building. Hence, one outliner can significantly influence
this result.
Evaluation of Reduced Abstraction
Firstly, the performance of DMO-SEP is compared to CC-SEP in
a sample scenario where agents are distributed over the building
according to a uniform random distribution. In this scenario, congestions are expected to be minimal, since there are only few agents
located initially in each room. Hence, the impact of the reduced
abstractions in DMO-SEP compared to CC-SEP becomes clearly
visible. Figure 4.14 displays the results. It becomes apparent that
the differences between DMO-SEP and CC-SEP are only marginal
when there is no congestion avoidance necessary. Nevertheless,
a slight improvement in the total evacuation time by one round
on average is achieved by DMO-SEP when compared to CC-SEP.
Since congestions are unlikely in the considered scenario, the reason
for this can be found in the reduced abstraction of the evacuation
graph model. While the macroscopic graph measures all distances
from the center of the rooms, DMO-SEP takes into account the
agents’ actual locations. In addition, DMO-SEP allows for bidirectional edges which increases the number of available evacuation
routes for planning. The results of this first experiment indicate
that DMO-SEP computes slightly faster evacuation instructions for
situations in which congestions are unlikely due to several reduced
abstractions in the planning process. Subsequently, the integration
of different objectives in DMO-SEP is investigated, starting with
risk avoidance.

143

Chapter 4 Swarm Evacuation Planning with Mobile Devices

Figure 4.14: Comparison of the two proposed SEP approaches
without congestion indicators.
Evaluation of Risk Minimization
The next experiment is performed in order to evaluate the impact
of risk avoidance on the total evacuation time and the average
evacuation time per agent, as well as on the time agents spend
waiting in congestions or are exposed to risk. To accomplish this,
two classes of evacuees are defined, one class consisting of risk
aversive agents, and another class which consists of agents who
prefer short ways even if they lead through risky areas. In this
experiment, rooms 4 and 10 are assumed to have a risk level of
1 and agents which enter these rooms are informed about their
current risk exposure. Risk aversive agents optimize path costs
with weights wrisk = 0.9, wdistance = 0.1 and less careful agents
assign weights wrisk = 0.8 and wdistance = 0.2 to the respective
costs. Figure 4.15 shows a sample situation which occurred during
an experiment run. Here, the different reactions to the detection
of higher risk levels for both agent classes can be observed. A risk
aversive agent, depicted as a blue square, chooses a path with a
longer traveling distance in order to avoid the risky area in room 4.

144

4.2 Dynamic Multi-objective Swarm Evacuation Planning
Agents with a higher tolerance for risk, on the other hand, refrain
from taking a detour and pass through rooms with higher risk
levels. This exemplary situation also demonstrates how ad hoc
network communication is employed in order to reduce the time
needed for evacuation. The risk aversive device does not enter room
4 in order to realize that it is an area of high risk, but, instead, is
informed about it by other agents via communication over the ad
hoc network. Therefore, the agent starts walking directly on the
alternative path instead of wasting time checking the condition of
room 4 by himself.

Figure 4.15: Risk avoidance for two different agent classes. More
risk aversive agents are depicted in black, rooms 4
and 10 represent areas of high risk.
In Figure 4.16, findings about the impact of risk avoidance on
evacuation time are presented. In this experiment, there are only
risk-aversive agents in the environment. The numerical results of
the experiment described previously reveal a quite intuitive effect.
The reduced average time an agent spends in risky areas comes at

145

Chapter 4 Swarm Evacuation Planning with Mobile Devices
the cost of an increased evacuation time. Additionally, the average
waiting time per agent is increased because more evacuees take the
less risky routes and, thereby, the congestion potential is increased
on these paths.

Figure 4.16: Experimental evaluation of risk avoidance.

Evaluation of Congestion Avoidance
The third experiment aims at testing the effectiveness of the proposed congestion indicators. Figure 4.17 displays a sample situation
in which 40 agents are initially located in rooms number 3 and
4. This setting provokes congestions in both rooms when minimizing traveling distance is the only optimization objective. The
integration of load, as well as entropy based costs, is expected to
dissolve these jamming queues. Scenario (a) shows the situation
in which only traveling distance is being optimized, while scenario
(b) optimizes distance and load costs and scenario (c) distance
and entropy costs. The weights for scenarios (b) and (c) are chosen as wload = wentropy = 0.8 and wdistance = 0.2. The depicted

146

4.2 Dynamic Multi-objective Swarm Evacuation Planning
experimental results confirm the initial expectations. For both
congestion indicators, the jamming queues in rooms 3 and 4 are
reduced. The snapshots also show the different effect both congestion indicators have on the evacuation situation. When agents
optimize entropy-based costs, they spread over various alternative
paths, which reduces the concentration of agents in the building
significantly. This effect is much less pronounced when load costs
are being optimized. Figure 4.18 quantifies the evaluation of the
three depicted scenarios which, generally confirms the previous
observations. The integration of entropy and load costs reduces the
overall evacuation time by 23% for the scenario with entropy costs
and 19% for additional load costs respectively, when compared to
evacuation planning where minimizing traveling distance is the
only objective. A comparison with CC-SEP yields an even higher
reduction in total evacuation time of 32% considering DMO-SEP
with load costs and 26% when compared to DMO-SEP including
entropy costs. This is a significant improvement when compared
to the first experiment in which agents were distributed randomly
over the building. This is due to the higher congestion potential.
Furthermore, the average evacuation time per evacuee and the
average waiting time, i.e., cycles without movements per agent, are
reduced significantly with DMO-SEP when compared to CC-SEP.
Achieving lower waiting times is an important criterion for the
evaluation of an evacuation navigation support system. It seems
natural that people in a panic situation are less patient and, if they
are expected to wait often, are more likely to lose their trust in
the navigation support system. CC-SEP navigation instructions
obviously expect evacuees to wait more often at jammed doors
compared to DMO-SEP. Even when the total evacuation time is
similar for both approaches, it is more likely that evacuees follow
navigation instructions which keep them moving during a panic

147

Chapter 4 Swarm Evacuation Planning with Mobile Devices
situation in contrast to a system which expects them to wait either
too frequently or for extended periods of time. It can be concluded
that DMO-SEP is superior to CC-SEP in all considered criteria.

(a) Distance costs

(b) Load costs

(c) Entropy costs

Figure 4.17: Evacuation situation after 24 cycles when optimizing
for distance costs only (a), distance and load costs
(b), and distance and entropy costs (c).

4.2.4 Conclusion
In this section, DMO-SEP is presented. DMO-SEP is an alternative SEP approach for mobile devices which overcomes some
of the drawbacks of CC-SEP. DMO-SEP is based on a less abstract environment model and collects precise location information
from other mobile devices over the ad hoc network. Furthermore,
DMO-SEP computes optimal evacuation paths with respect to
several objectives at the same time and can take into account
individual preferences and characteristics of the mobile device’s
user. The application of an incremental, heuristic search algorithm
to find optimal paths allows for dynamic replanning of navigation

148

4.2 Dynamic Multi-objective Swarm Evacuation Planning

Figure 4.18: Experimental evaluation of congestion avoidance
with DMO-SEP and comparison to the performance
of CC-SEP.
instructions and reduces the necessary computations compared to
a search from scratch. This is especially useful for SEP, where
knowledge collected by devices via local communication changes
constantly over time. Fast replanning and integration of newly
available information makes the evacuation planning approach
adaptable to detected changes in the environment, such as the
emergence of congestions. DMO-SEP is able to evaluate paths
with respect to multiple objectives by assigning a weighted sum of
cost components to a path, such that each cost component reflects
one optimization objective. Four cost components are proposed
and evaluated in experiments. Two of these components are meant
to capture the congestion potential on a path. One congestion
indicator is based on the load of a path, i.e. the number of evacuees in relation to the size of the rooms on a path, the other is an
entropy-based indicator, which assesses the concentration of evacuees within rooms along the considered path. Additionally, travel
distance and danger indicators are proposed as cost components
for optimization.

149

Chapter 4 Swarm Evacuation Planning with Mobile Devices
Experiments are performed in order to evaluate DMO-SEP. For
this, a realistic scenario with various rooms in the building is taken
as a basis, which is also used in the CC-SEP experiments. The
performance of DMO-SEP is assessed by regarding the resulting
total evacuation time, the average evacuation time per evacuee, the
average time an evacuee spends in risky areas, and the average waiting time of each evacuee, i.e. the time it spends without moving.
Firstly, DMO-SEP and CC-SEP are compared in a scenario with
low congestion potential. It is shown that DMO-SEP produces
better results in terms of total and average evacuation time, as
well as waiting time than CC-SEP in this scenario, which is most
likely due to the reduced abstractions in the evacuation model used
as a basis for evacuation planning. The evaluation of DMO-SEP
proceeds with testing the risk minimization objective and the integration of different preferences of evacuees. To do so, a sample
scenario is analyzed in which two classes of evacuees are simulated,
each of which representing different risk preferences. One class of
agents is risk-aversive, the other is risk-friendly. The evacuation
instructions produced by DMO-SEP are shown to respect these
preferences and it is observable that evacuees are informed early
about risky paths, such that they do not have to go near dangerous areas before they switch to alternative routes. Additionally,
experiments with only risk-aversive agents reveal that risk-aversion
comes at the cost of evacuation time, which is intuitively understandable. Furthermore, the impact of both congestion indicators
is examined and it is shown that they reduce the overall evacuation
time effectively, even in an uncertain and incomplete information
situation. In a specific scenario, which provokes congestions, the
total evacuation time is reduced with DMO-SEP and load-based
congestion avoidance by up to 32% and with entropy-based congestion avoidance by only slightly less when compared to CC-SEP.

150

4.3 Summary
Furthermore, DMO-SEP is shown to improve average evacuation
time and waiting timer per agent significantly when compared to
CC-SEP. It can be concluded from the presented evaluation that
DMO-SEP is superior to CC-SEP in terms of evacuation as well as
waiting time and that DMO-SEP is, therefore, the better choice for
evacuation planning than CC-SEP when the number of exchanged
messages is not important. One remaining open question in context
of DMO-SEP is the determination of optimal weights for the cost
components. However, the O/C Architecture offers a solution to
this problem. The online and offline learning mechanisms could
be applied in order to learn the best weights for each optimization
criteria depending on the current evacuation scenario in a building.

4.3 Summary
This chapter defines the process of SEP, which describes distributed
and decentralized evacuation planning based on information which
is collected via local communication in order to improve evacuation
performance. The approach has its name from the characteristic
that local communication and decision making is used to improve
overall evacuation time for a swarm of mobile devices within a
building. The main challenge for SEP is an uncertain information
situation for each device, which arises from delays and link breakages in ad hoc network communication. This makes it hard to
estimate waiting times on specific escape routes. Two specific SEP
algorithms are introduced. The first approach is called CC-SEP
and optimizes only total evacuation time. To achieve this goal,
a routing algorithm is adapted, which uses capacity reserving
strategies in order to find time-optimal routes for all evacuees in
a building. Routing is based on an abstract graph model of the
building, which is created from locally available information about

151

Chapter 4 Swarm Evacuation Planning with Mobile Devices
other evacuees’ locations. A suitable route for each specific mobile
device is selected from the optimization result. An experimental
study reveals that CC-SEP is able to improve total evacuation
time significantly when compared to a situation where evacuees
follow the shortest path towards the exit. Furthermore, results
are shown to be robust even when some evacuees deviate from
the evacuation routes suggested by CC-SEP. Slight drawbacks are
identified when evacuation planning is executed repeatedly in order
to adapt to new evacuation situations. However, the performance
is still significantly better when compared to a simple shortest path
following behavior and several methods are proposed and evaluated which can help to reduce this limitation. The second SEP
approach, called DMO-SEP results in even faster evacuation times.
The reason for this is partly due to a more detailed environmental
model taken as a basis for path planning and an improved communication model. Apart from its improved performance, DMO-SEP
has the advantage that it is based on a path planning approach
instead of flow optimization in order to find escape routes. As
a consequence, multiple objectives can be regarded for optimization and can be weighted according to evacuee-specific preferences.
Apart from minimizing traveling distances, two other objectives are
investigated, namely risk minimization and congestion avoidance.
In order to avoid congestions, indicators based on the distribution
of evacuees in the building are proposed, which evaluate routes
according to their congestion potential instead of estimating waiting times, as it is done in CC-SEP. These congestion indicators
improve the results of DMO-SEP even further and are shown to
accelerate the evacuation of buildings by up to 32% when compared
to CC-SEP in a scenario with high likelihood for congestions. This
makes DMO-SEP more suitable for the computation of evacuation
instructions. However, CC-SEP requires fewer configurations from

152

4.3 Summary
the device’s user since preferences do not have to be specified. This
simplicity can be an advantage, for example, if the user is not
willing to provide detailed information to the evacuation system.
Moreover, exchanging the location information of each evacuee in
the building increases the communication overhead when compared
to exchanging only the observed number of evacuees in each room,
as it is proposed for CC-SEP. Hence, CC-SEP could be the first
choice for evacuation planning in case reducing energy consumption
is an objective.

153

CHAPTER

5

LOCALIZATION IN MOBILE AD HOC
NETWORKS

In OBESS, mobile devices in combination with a permanently
installed SSN are intended to be used as navigation support for
building evacuation. While the devices in the static network are
assumed to possess knowledge of their positions in the building due
to a priori configuration, the mobile devices have to be located first
in order to be able to find suitable navigation instructions leading
evacuees from their current positions to an exit. Since GPS is not
available indoors, it cannot be used to localize mobile devices in
OBESS. Furthermore, a single point of failure is to be avoided in
the localization process in order to make it more robust against
failure. As a consequence, decentralized computation methods for
the mobile devices’ locations are required. Because the static sensors are assumed to know their locations, beacon-based localization
algorithms, which derive unknown locations of devices from several

155

Chapter 5 Localization in Mobile Ad Hoc Networks
known locations, are applicable. Section 2.4.1 gives an overview of
beacon-based localization algorithms which can be used for this
purpose. Beacon-based algorithms allow for the computation of
absolute locations with respect to a common reference grid, which
corresponds to the building’s layout in the scenario considered here.
From the presented beacon-based localization algorithms, methods
which require information about the angle between mobile and
static devices are not suitable for usage in OBESS. As detailed
before, the estimation of angles requires complex preparation and
bulky hardware, which portable devices are usually not equipped
with. This leaves the proximity and distance-based algorithms.
While proximity-based algorithms are less complex in terms of
computational effort, high quality localization results require numerous beacons in the network. Because the installation of sensors
is costly, one objective in OBESS is to minimize the number of
beacons in order to make its installation more attractive for building owners. Therefore, distance-based localization algorithms are
generally the better choice. Nevertheless, proximity-based localization algorithms are a good fallback for the devices in case there are
not enough beacon nodes for distance-based localization in certain
areas of a building. Furthermore, proximity-based localization approaches are rather simple and, therefore, do not require complex
computations to be performed on the devices. Hence, they could
be the better choice when quality of localization results has to be
traded for reducing energy consumption in case the mobile device
runs out of battery. Although localization algorithms are studied
widely in the literature, the impact of mobility has not yet been
emphasized. As explained in Section 2.4.3, the focus of current
research about the effects of mobility on localization either lies on
examining the impairment of the physical signal or the mobility is
assumed to be actively controllable by the device itself. In contrast

156

to this assumption, mobile devices in OBESS are portable and
do not autonomously move like robots, i.e. their mobility has a
passive character.
Definition 5.1 (Passive mobility). Passive mobility describes a
state in which devices are carried by people, animals, or nature
and in which the carrier, not the device, decides where and when
to move.
Distance-based algorithms use distance estimates between beacons
and the devices to be located. These estimates can be derived
by range-based or range-free distance estimation techniques (cf.
Section 2.4.2). While range-based distance estimation requires
specific hardware to analyze the physical communication signal,
range-free methods use messages which are exchanged between
the devices in a network. Since avoiding additional hardware
for the mobile devices in OBESS is to be desired for making the
devices more lightweight and less costly, range-free methods are
the obvious choice for distance estimation in OBESS. Most rangefree distance estimation techniques require a somewhat dense and
uniform distribution of communication neighbors over a device’s
communication range. Since mobility of the devices influences
the composition and structure of the network topology, it can
be expected to have a significant impact on localization results.
So far, there is little known about the influence of mobility on
the results of range-free distance-based localization algorithms,
hence, this investigation is subject to Sections 5.1, 5.2, and 5.3 of
this chapter. In Section 5.1 localization algorithms based on hop
counts are investigated while various mobility models are applied
to the devices. The potential of synchronization for improving
distance estimation based on hop counts in dynamic networks
is examined in Section 5.2. Section 5.3 presents an alternative

157

Chapter 5 Localization in Mobile Ad Hoc Networks
approach to distance estimation based on hop counts, which delivers
promising results, especially in dynamic networks. As stated before,
minimizing the number of beacons is an objective that has to be
considered in OBESS as they are costly and need to be installed,
configured, and maintained. On the other hand, the number and
positions of beacons influence the quality of localization results.
To address this trade-off, a method which optimizes the placement
of beacons for localization of mobile devices, specifically during
building evacuation, is presented in Section 5.4 of this chapter.
Section 5.5 summarizes the findings of this chapter. Parts of the
research presented hereafter are also published by Merkel et al.
[152, 153, 154, 156, 157, 159].

5.1 Effects of Mobility on Hop Count Based
Distance Estimation
A well-studied approach to estimate distances in ad hoc networks
is based on so-called hop counts. A hop count denotes the minimum number of relay devices which two devices need to exchange
messages with each other (cf. Ghosekar et al. [75]). Although this
metric is mainly used in the context of routing (cf. Boukerche
et al. [22], Chatterjee and Das [38], Chou et al. [42]), it can also
be used to estimate distances. These distance estimates then form
the basis for localization. In Section 2.4, some methods which
transform hop counts into distance information are presented. This
section, however, focuses on investigating the impact of mobility
on hop counts and, consequently, on the quality of derived distance
estimates. When investigating the influence of mobility, it is intuitively comprehensible that the characteristics of mobility play
an important role. For example, distances between the devices do
not change if all devices are moved with the same speed and in the

158

5.1 Effects of Mobility on Hop Count Based Distance Estimation
same direction, whereas random mobility has a great impact on the
distances. Furthermore, the number of moving devices, their speed,
and direction of movement are important variables when it comes
to examining the impact of mobility. As a consequence, a large
spectrum of mobility models is examined in order to derive general
statements about the influence of mobility on distance estimation
based on hop counts.

5.1.1 Mobility Models
The mobility pattern in a dynamic network highly depends on its
application and the environment it is deployed in. Therefore, in
order to be able to make general statements about the impact of
mobility on hop count based distance estimation, a variation of
mobility models is analyzed. In the work of Camp et al. [34], some
models for describing different kinds of mobility in networks are
suggested and their impact on network connectivity and routing
is investigated. These patterns are used hereafter as a basis for
studying the effect of mobility on distance estimation based on hop
counts. In addition, new mobility models, which are not considered
by Camp et al. [34], are introduced.
Individual Mobility Models
Individual mobility models describe movements where the next
position of a device is determined independently from any other
device in the network. The Random Walk (RW) mobility model is
one of the most widely studied mobility patterns in the literature,
e.g., by Zonoozi and Dassanayake [259]. In RW, a random direction
and speed are selected from predefined ranges and the device moves
accordingly until a fixed distance is traveled or a specific time has
passed. Chaos Move (CM) denotes an RW mobility model where a

159

Chapter 5 Localization in Mobile Ad Hoc Networks
new speed and direction are selected at each step, i.e., the mobile
device is kept in a small area around its starting position. The
Random Direction Walk (RD) model defines a random target at the
border of the environment and a speed value within a certain range
and the device is moved accordingly until the target is reached.
Once the device is there, it pauses for some time before a new
target is selected. When using this model, there is a high likelihood
that devices spend most of their time somewhere in the middle of
the environment. Bounded Random Walk (BR) is similar to CM,
but speed and direction are not varied completely at random but
within a small range around the preceding values. Similar to this,
the Gauss Markov Move (GM) selects the next direction and speed
according to the following equations:
st = α · st−1 + (1 − α) · μs +
dt = α · dt−1 + (1 − α) · μd +



(1 − α2 ) · sgr



(1 − α2 ) · dgr

(5.1)
(5.2)

with st and dt denoting the new values for speed and direction
respectively, α being a random parameter (0 ≤ α ≤ 1) and sgr and
dgr being chosen from a Gaussian random distribution with zero
mean and a standard deviation of one. μs and μd are constants.
In Probabilistic Random Walk (PR), the movement is defined by
a finite state machine with fixed probabilities for state transits.
The allowed states are backward, forward or stop, and turn left or
right. The probabilities are chosen in a way that they emphasize
continuous moves in the same direction (cf. Camp et al. [34]).
Figure 5.1 shows trajectories of all individual mobility models
considered above.

160

5.1 Effects of Mobility on Hop Count Based Distance Estimation

Figure 5.1: Trajectories. From top left to bottom right: CM, RD,
BR, RW, GM, PR.
Coupled Mobility Models
Coupled mobility models are characterized by mutual influences
between the mobile devices. The Column Mobility (ColM) model
simulates children walking in a line behind their parents. Here,
all devices move randomly (according to CM) except for some
groups consisting of a leader and followers. All followers move
approximately in a row behind the leading device. Furthermore,
the Nomadic Move (NomM) model is defined in which all devices
move according to CM except for some leaders that collect followers.
Whenever a non-leader device comes within communication range of
a leader, it starts following the leader while moving randomly within
its communication range. When a leader has more than a maximum
number of followers, a random follower leaves the group. In the
Reference Point Mobility (RefM) model, each device is assumed to

161

Chapter 5 Localization in Mobile Ad Hoc Networks
have a virtual reference point around which it moves randomly while
never exceeding a maximum distance. The reference points are
assembled in a grid and move according to CM, but share the same
speed and direction as if they were connected to each other. The
Stream Mobility (StrM) model simulates devices which are moved
by regular forces such as wind or water. Each device remembers the
direction of nearby devices and chooses its own moving direction
within a certain range around the angle of the most recently moved
neighbor, which is why the devices exhibit a stream-like movement.
This mobility model was developed specifically for this study and
does not belong to the models published by Camp et al. [34].
Beacon Mobility Models
Since the results of distance estimation based on hop counts depend
on the position of the beacons in the network (cf. Bachrach and
Taylor [12], Nagpal et al. [171]), beacon mobility models are defined
additionally. Beacon mobility models define moves with a certain
angle and speed around a beacon. Beacon mobility models can be
further distinguished depending on whether individual or coupled
movements are regarded. With individual beacon mobility models,
a device moves with a predefined speed in the direction of α with
respect to a beacon (cf. Figure 5.2(a)).
The coupled beacon mobility model is similar, except for the
fact that the device is moved together with a fixed number of its
neighbors in a specified direction with respect to the beacon (cf.
Figure 5.2(b)). The same angle α is defined for all devices belonging
to the coupled group. An angle of α = 0◦ indicates a movement
leading away from the beacon, α = 180◦ directs the mobile devices
towards the beacon. A movement at an angle of α = 90◦ means that
the devices move along circular rings around the beacon. Figure 5.3
illustrates sample trajectories of the individual beacon mobility

162

5.1 Effects of Mobility on Hop Count Based Distance Estimation







(a)

(b)

Figure 5.2: Illustration of the individual (a) and coupled (b)
beacon mobility model.
model and Figure 5.4 of the coupled beacon mobility model for
various directions of movements. From these trajectories, it can be
observed that the density of devices in the environment is more
regular for individual mobility than for the coupled mobility models.
Table 5.1 summarizes the mobility models described before.

163

Chapter 5 Localization in Mobile Ad Hoc Networks
Table 5.1: Overview of mobility models investigated in this
experimental study.
Type
Individual

164

Model
RW

Short description
Movement according to randomly
chosen direction and speed until a
fixed distance or time is traveled.
CM
Similar to RW, but direction and
speed are changed in each step.
RD
Devices move towards randomly
chosen target point at the environment’s border.
BR
Similar to CM but speed and directions are varied within a range
around preceding values.
GM
Speed and direction are selected
according to Equations 5.1 and
5.2.
PR
Next movements are defined by
a finite state machine with fixed
probabilities for state transits.
Beacon Mo- Movements with fixed speed and
bility
direction with respect to a beacon.

5.1 Effects of Mobility on Hop Count Based Distance Estimation

Coupled

ColM

Groups of followers which move
approximately in a line behind a
leading device.
NomM
Groups of followers which move
according to CM but stay within
a leading device’s communication
range.
RefM
Devices move around fixed reference points which are assembled
like a grid and additionally moved
according to CM.
StrM
Directions of movements are chosen randomly within a range
around the last observed directions of the most recently moved
neighbor.
Beacon Mo- Leading devices and all devices in
their communication ranges move
bility
with the same fixed speed and direction with respect to a beacon.

5.1.2 Hop Count Error Model
Hop counts relative to a beacon can be assigned to all devices
in the network using the GA presented by Nagpal et al. [171].
As detailed in Section 2.4.2, there are various methods to derive
distance estimates from hop counts. The basic idea is to multiply
the hop count value by the communication range r in order to

165

Chapter 5 Localization in Mobile Ad Hoc Networks

Figure 5.3: Trajectories of devices moved according to the
individual beacon mobility model at an angle of
α = 0◦ , 45◦ , 90◦ , 135◦ , 180◦ with respect to the
beacon (left to right).

166

5.1 Effects of Mobility on Hop Count Based Distance Estimation

Figure 5.4: Trajectories of devices moved according to the coupled
beacon mobility model at an angle of α = 0◦ , 45◦ , 90◦ ,
135◦ , 180◦ with respect to the beacon (left to right).

167

Chapter 5 Localization in Mobile Ad Hoc Networks
get an estimate for the distance. However, because networks are
usually not perfectly dense, this method generally overestimates
the distances. As a consequence, refinements are proposed which
adjust the estimates to the network density using statistical techniques. Such algorithms include the idea of applying a reduction
rate depending on the local network density (cf. Kleinrock and
Silvester [122], Wong et al. [249]), integrating information about
the hop count distribution in the neighborhood (cf. Liu et al.
[140], Nagpal et al. [171]), or computing the average hop length
using knowledge about the distances between beacons (cf. Huang
and Selvakennedy [102], Niculescu and Nath [176, 177], Savvides
et al. [208]). It is notable that the distance estimation error is
minimal for all estimation techniques in case the network is perfectly dense. This consideration offers the possibility of evaluating
the impact of mobility on hop count based distance estimation
independently from the specific estimation method. To achieve
such an independent evaluation, the hop count of a device in the
actual network is compared to the hop count a device at the same
location would have if the network was perfectly dense. In order to
do so, the concept of ideal hop counts is introduced, which denotes
the hop count each device would have in a perfectly dense network.
Definition 5.2 (Ideal Hop Count Value). Let hideal
(n) denote the
b
ideal hop count value of a device n with respect to a beacon b. The
ideal hop count corresponds to the smallest integer value which,
multiplied by the communication range r, is not smaller than the
distance between device n and beacon b. The ideal hop count is
computed as shown in Equation 5.3, with d (n, b) representing the
Euclidean distance between device n and beacon b.
(n) := 
hideal
b

168

d (n, b)

r

(5.3)

5.1 Effects of Mobility on Hop Count Based Distance Estimation
Considering a perfectly dense and evenly distributed ad hoc network, each device n ∈ N , with N being the set of all devices in the
network, would be assigned a hop count value by the GA which
corresponds to its ideal hop count. An area A in which hideal
(n)
b
has the same value for all n ∈ A represents one of the perfect gradient rings shown in Figure 2.17. Such a ring is called a gradient
ring gi , with i being the common value for hideal
(n) for all devices
b
n ∈ N which are located in gi . The value i also corresponds to the
ordinal number of the gradient ring when counting begins at the
beacon. As explained, the deviation between the hop count values
assigned to devices by the GA and the ideal hop count values is
directly related to the quality of any hop count based distance
estimation algorithm. Therefore, the hop count error is defined as
follows.
Definition 5.3 (Hop Count Error). The hop count error Eb (n) of
a device n is defined as the difference between the hop count value
which a GA assigns to the device hb (n) and its ideal hop count
(hideal
(n)) with respect to a beacon b:
b
Eb (n) = hb (n) − hideal
(n)
b

(5.4)

The average hop count error in a gradient ring gi (b) with respect
to a beacon b can be computed as described in Equation 5.5 with
J = {j ∈ N |  d(j,b)
r  = i} referring to all the devices which are
physically located in the gradient ring gi (b) with respect to the
beacon b.
Ei (b) =

1 
(i − hb (j))
|J| j∈J

(5.5)

169

Chapter 5 Localization in Mobile Ad Hoc Networks
Positive Hop Count Error
A positive hop count error means that hb (n) is larger than hideal
(n),
b
i.e., the GA leads to an overestimation of the hop count. Overestimation occurs, for example due to low density, when the neighbor
which is closest to the beacon is not located exactly at the border
of a gradient ring (cf. Figure 5.5). Hop count overestimation is
additive, since an overestimated hop count can serve as a basis
for the hop count determination of subsequent devices. Figure 5.6
illustrates a simple example of a positive hop count error. The
beacon is located at the top left corner and each device is labeled
with its assigned hop count value h (n)1 . The error occurs in the
gradient ring g2 due to the gap in g1 . The device which is physically
located in g2 (marked in red) does not have any neighbor located
in g1 and, thus, its hop count has a value of 3. The hop count
error in gradient ring g2 , hence, results in E2 = 14 . A positive
hop count error is common in static networks because they rarely
are perfectly dense. This usually leads to overestimated distances,
which the refinement techniques, described in Section 2.4.2, intend
to compensate.
Negative Hop Count Error
In dynamic networks, density varies due to the mobility of the
devices. In general, when a device is moved, the density of the
network is decreased at its former position and increased at its
new location. Consequently, at first sight such a movement does
not seem to have any effect on the average hop count error in
a network. At a closer look, however, it becomes apparent that
movements towards the beacon increase the density in gradient
1

In case of a single beacon in the network, the reference to the beacon is
removed from the notations for better readability.

170

5.1 Effects of Mobility on Hop Count Based Distance Estimation

Figure 5.5: Shift of the gradient
border towards the
beacon due to low
density in the
network.

Figure 5.6: Illustration of the
emergence of
positive hop count
error due to low
density in the
network.

rings which are closer to the beacon. Since hop count error is
additive, the network’s overall hop count error actually can be
reduced by mobility. Figure 5.7 shows an example of a decreased
hop count error due to density enhancement close to the beacon.
This effect can reduce the positive hop count error.

Figure 5.7: Illustration of hop count error correction due to
mobility in the network.

171

Chapter 5 Localization in Mobile Ad Hoc Networks
Movements leading away from the beacon have the opposite effect.
However, this is not the only effect they have. In MANETs, computations are usually assumed to be made asynchronously because
the devices do not necessarily have synchronized system clocks.
This fact, together with the assumption that the device does not
know whether it is moved, can lead to a scenario where mobility
is accompanied by a negative hop count error. When a device is
moved to a gradient ring with a higher ordinal number, i.e., away
from the beacon, it is possible that the device’s new neighbors
wrongfully adapt their hop count values according to the newly
arrived device’s hop count before that device itself updates its hop
count. Since the newcomer has a lower hop count value, which
serves as a basis for hop count computations of its new neighbors,
the new neighbors are likely to underestimate their hop count and
a negative hop count error can occur. Additionally, the negative
error is additive to subsequent devices, similar to the positive hop
count error. Figure 5.8 illustrates the emergence of a negative hop
count error in gradient ring g3 with E3 = − 25 .

Figure 5.8: Illustration of the emergence of a negative hop count
error due to mobility in the network.

172

5.1 Effects of Mobility on Hop Count Based Distance Estimation
Assessment of Hop Count Errors
In general, overestimation and underestimation of distances are
undesired for localization algorithms, which is why, ideally, one
should simply let both effects (density and mobility) compensate
each other. But passive mobility cannot be controlled by the devices and, thus, the height of the negative error is not amenable to
influence. As a consequence, natural overestimation of distances
can turn into an unpredictably high underestimation and a balancing effect between mobility and density induced errors cannot
be guaranteed. Furthermore, the density-induced overestimation
can be counteracted with the statistical techniques for dealing
with density-induced overestimation, described in Section 2.4.2,
which are not valid anymore in dynamic networks. Due to its
unpredictable character, mobility-induced underestimation is the
more aggravating type of estimation error and should, therefore,
be eliminated. In order to tackle this problem, a device has to be
able to recognize whether and how it is moved. It has to be able to
analyze the effect which the movement has on its hop count value
and to be able to eliminate this effect by an appropriate adjustment
of its hop count value. Since with passive mobility devices do not
know whether they are moved, this objective is quite challenging.
Approaches to Reduce Hop Count Error in Dynamic Networks
To avoid the emergence of negative hop count errors, a modified
version of the GA is proposed. The underlying assumption is that
when a device moves towards a new neighborhood, it is more likely
that, both, the minimum and the maximum hop count values in its
neighborhood change. In contrast, a device which has not moved
but detects a change in its neighbors’ hop count values due to a
newly arriving device, is likely to find either a new maximum or

173

Chapter 5 Localization in Mobile Ad Hoc Networks
a new minimum hop count value instead of both. As a result, it
could be achievable to avoid that devices adapt their hop counts
to newly arrived devices in their neighborhood which have not yet
updated their hop counts and, thus, avoiding the emergence of
negative hop count error.
Definition 5.4 (Maximum-oriented Gradient Algorithm). The
Maximum-oriented Gradient Algorithm (MoGA) is defined as an
algorithm which is equal to the GA with the additional condition
that a device updates its hop count only when the maximum hop
count in its neighborhood has changed as well as the minimum hop
count value.
A second proposal to tackle the problem is that each device recognizes mobility and its characteristics by observing and analyzing
certain changes in its neighborhood. Based on this analysis, the
device subsequently makes reasonable adjustments to its hop count
value. Since communication overhead is always an issue in mobile
ad hoc networks, it is reasonable to aim at deriving information
about movements from messages which are exchanged in the GA
instead of defining additional messages required. In order to do
so, two indicators ID-change and HC-change are proposed, which
describe changes in a device’s environment, i.e. its communication
neighborhood, during certain time intervals.
Definition 5.5 (ID-change). ID-change is defined as the percentage of changed IDs in a device’s neighborhood in the time interval
[t-1, t], with IDt denoting the set of IDs in the device’s neighborhood at time t:
ID-changet =

174

|IDt \ IDt−1 | + |IDt−1 \ IDt |
|IDt−1 ∪ IDt |

(5.6)

5.1 Effects of Mobility on Hop Count Based Distance Estimation
Definition 5.6 (HC-change). HC-change is defined as the percentage change of the average hop count in a device’s neighborhood
in the time interval [t-1, t], with HC t denoting the average hop
count in the device’s neighborhood at time t:
HC-changet =

HC t − HC t−1
HC t−1

(5.7)

Ideally, these change-indicators have conclusiveness about mobility
characteristics in a device’s neighborhood. In order to verify,
whether this expectation is met, both indicators are examined in
an experimental study in Section 5.1.3.

5.1.3 Evaluation
In order to quantify the impact of various passive mobility models
on the average hop count error in a network and to evaluate
the proposed containment techniques, a simulative model is used.
In the first set of experiments, the impact of various mobility
models on the hop count error is investigated. The second set of
experiments evaluates the effectiveness of MoGA to avoid negative
hop count errors and the ability of the two indicators ID-change
and HC-change to capture mobility and its characteristics.
Experiment Settings
The ad hoc network model contains 1000 mobile devices, which are
distributed according to a uniform random distribution on a twodimensional, barrier-free plane. The communication neighborhood
of a device corresponds to its physical neighborhood on the plane
within a fixed Euclidean distance r. The communication range r
is assumed to be equal for all devices and its default value is set to
7% of the square plane’s side-length. This value corresponds to an

175

Chapter 5 Localization in Mobile Ad Hoc Networks
average neighborhood size of 14 devices, which is close to the critical
minimum size identified by Nagpal et al. [171] needed to achieve
good localization results. A static beacon, which initiates the GA,
is located at the top-left corner of the environment
2.17).
√(cf. Figure


With this setting, there are 21 gradient rings ( 2/0.07 = 21).
A simulation cycle is defined as n = 1000 randomly selected devices sequentially computing their hop counts. The hop counts are
determined on the basis of available information from neighbors at
the time of computation. This setting imitates the asynchronous
environment in ad hoc networks. After computation of its hop
count, the device is moved according to the mobility model examined. Collisions between devices, as well as communication
deficiencies, such as interferences, shadowing, fading, or multipath
effects, are not being considered. All experiments are repeated
50 times with different randomly chosen uniform distributions of
the devices across the environment and the results are averaged
for each gradient ring separately. Devices which are not able to
compute a hop count because they are disconnected from the network are not taken into account. Since, due to the nature of the
scenario, each of the gradient rings contains a different number
of devices, only the middle hops (10-15) are considered because
they contain a relatively high and similar number of devices. The
simulation environment is a torus world, meaning that devices can
leave the environment and enter again at the opposite side. When
a device leaves the environment, its hop count is set to unknown,
simulating a new device entering the environment on the other side.
The speed range for each move is selected randomly between 3.4%
and 3.6% of the plane’s side-length traveled per cycle. This way, a
device requires at least two cycles to leave its own communication
range. For RW, the maximum moving distance is selected as 60%.
In RD a move is paused for 10 cycles. For GM, α is set to 75% of

176

5.1 Effects of Mobility on Hop Count Based Distance Estimation
the plane’s side-length and an average angle of 0 degree measured
from the x-Axis of the environment (bottom border) is selected.
Angle tolerance for BR and StrM is set to 30 degrees, ColM and
NomM are initialized with 10 leaders and 10 followers per leading
device.
Average Hop Count Error for Various Mobility Models
In the first set of experiments, the error is calculated for 100 cycles
starting with cycle 30, which is the minimum time required to
elapse until hop count values are stable in a static network. A
device is moved with a probability of pm = 0.5 after each hop count
update. At first, the hop count errors are calculated for a static
network. The results are shown in table 5.2. Only positive hop
count error values are obtained. These error values intensify as the
index of the gradient ring increases, demonstrating the additive
nature of the positive hop count error, which grows larger with
increasing distance from the beacon.
Table 5.2: Error Ei for distance estimation based on hop counts in
gradient rings g1 to g10 in a static ad hoc network.
gi
Ei
g11
2.5

g1
0.0
g12
2.8

g2
0.4
g13
3.0

g3
0.6
g14
3.2

g4
0.9
g15
3.3

g5
1.1
g16
3.5

g6
1.4
g17
3.7

g7
1.7
g18
3.8

g8
1.8
g19
4.0

g9
2.1
g20
4.1

g10
2.3
g21
3.6

Figure 5.9 displays the average error for gradient rings from g10
to g15 when individual mobility models are applied. All individual
mobility models overcome the positive error of a static network.
PR shows a noticeably higher underestimation than the other
mobility models. This can be explained by the nature of this model

177

Chapter 5 Localization in Mobile Ad Hoc Networks
as all devices are moved almost into the same direction leading
away from the beacon. As an underestimation can only occur
when movements are directed away from the beacon, the negative
error induced by this mobility model is higher when compared to
the other models. BR, GM, CM, RD, and RW show very similar
tendencies in terms of error values. RD exhibits a slightly more
positive behavior, which can be explained by the reduced mobility
due to the pauses of devices. So far, underestimation does not
become dominant for all individual mobility models under the
selected settings, unless the movements are mainly directed away
from the beacon.

Figure 5.9: Hop count error with various individual mobility
models.
Figure 5.10 shows the hop count error values for group mobility
models. For StrM, as well as for NomM, the underestimation is less
pronounced when compared to other coupled mobility models. One
reason for this could be that similar movements in neighborhoods

178

5.1 Effects of Mobility on Hop Count Based Distance Estimation

Figure 5.10: Hop count error with various coupled mobility
models.
have a moderating effect on the mobility induced underestimation
since both models generate different groups of devices, which move
together through the network. As already shown for PR, another
important factor for the development of negative hop count error
is the direction of a move relative to the beacon. Both mobility
models, StrM and NomM, cause random movements while ColM
leads to movements in a more constant direction which are accompanied by higher negative error values. Apart from ColM,
RefM also shows high negative error. This can be explained by
the additional movement of the dynamic reference-point network
supplementary to the individual movements of each device, which
causes devices to move further in one time step when compared
to other mobility models. This indicates that speed, i.e. distance
traveled per unit time, is another influencing factor for the level of
negative hop count error that arises with mobility. The experiment

179

Chapter 5 Localization in Mobile Ad Hoc Networks
further confirms that mobility in the network can turn the inherent
overestimation of hop counts in a static network into underestimation. Furthermore, the impact of mobility on the estimation error
is strongly dependent on the characteristics of the specific mobility
model. So far, the experiments indicate that there are three main
characteristics of mobility which influence the hop count error:
• Direction of movements with respect to the beacon
• Speed of movements in terms of traveled distance between
hop count updates
• Similarity of movements in a neighborhood
In order to be able to investigate these factors independently, the
beacon mobility model is applied to 50 random devices in the
network. For the coupled mobility model, 5 leaders are selected
with 10 followers each. Figure 5.11 shows the average error in
the gradient rings g10 to g15 for the individual beacon mobility
model with varying values for angle α and speed. It is observable
that both direction and speed are indeed relevant factors for the
mobility induced hop count underestimation. Additionally, the
results show that both parameters are interrelated and vary in the
intensity of their impact. While speed is negatively correlated with
the error value for an angle between 0 and 90 degrees as well as
between 270 and 360 degrees, it has almost no impact when devices
move at an angle between 135 and 225 degrees. In fact, movements
towards the beacon hardly influence the hop count error at all and
the hop count distribution in the network is almost the same as
in a static network. Similar results can be observed for coupled
mobility models as shown in Figure 5.12. A considerable difference
between coupled and individual mobility models can be observed

180

5.1 Effects of Mobility on Hop Count Based Distance Estimation
around an angle of 180 degrees, where errors seem to decrease for
coupled movements when compared with individual movements.
This can be explained by the simultaneous increase in density close
to the beacon as described before. The reason why this effect does
not occur for individual movements is that the density does not
increase significantly at a specific point in time, i.e., the additive
effect is minor.

Figure 5.11: Hop count errors for individual mobility models with
different movement speeds and angles with respect to
the beacon.
Subsequently, the individual and coupled mobility models are compared for angles between 0 and 180 degrees only, due to symmetries
of error values. Figure 5.13 illustrates the hop count error averaged

181

Chapter 5 Localization in Mobile Ad Hoc Networks

Figure 5.12: Hop count errors for coupled mobility models with
different movement speeds and angles with respect to
the beacon.

182

5.1 Effects of Mobility on Hop Count Based Distance Estimation
over experiments with different angles and Figure 5.14 averaged
over experiments with different speeds respectively. In both experiments, with individual and coupled beacon mobility, the same
total number of devices is moving in the same direction with the
same speed respectively. The only difference is that in experiments
with coupled beacon mobility, devices move in clusters. The results
show that with coupled mobility, the negative hop count error is
lower when compared to individual mobility models. This confirms the previously made observation that similar movements in
a neighborhood mitigate the effect mobility has on the hop count
error. For coupled mobility models and directions between 0 and 90
degrees, the effect of simultaneous density increase near the beacon
becomes visible once more. The reason for the high dispersion of
values lies in the fact that experiments with varying speeds and
angles, respectively, are averaged.
Eliminating Mobility Induced Underestimation
While the previous section examines the influence of mobility on
hop count estimation, this section investigates different solutions
proposed for dealing with the problem of mobility induced underestimation. At first, the effectiveness of MoGA is being tested in
the same experimental setup as described previously. Results are
shown in Figures 5.15 and 5.16. A decrease in mobility induced
underestimation when compared to GA can be identified in all
examined cases. Nevertheless, negative error is not successfully
eliminated and the reduction is almost constant relative to the
error level. It can be concluded that MoGA has a mitigating effect
on the negative error but does not solve the problem entirely.
In the next experiment, ID-change and HC-change (cf. Equations
5.6 and 5.7) are examined. These indicators’ ability to recognize
whether the device itself is moving or whether changing hop count

183

Chapter 5 Localization in Mobile Ad Hoc Networks

Figure 5.13: Hop count errors for individual and coupled beacon
mobility models with different movement speeds.

184

5.1 Effects of Mobility on Hop Count Based Distance Estimation

Figure 5.14: Hop count errors for individual and coupled beacon
mobility models with different movement angles with
respect to the beacon.

185

Chapter 5 Localization in Mobile Ad Hoc Networks

Figure 5.15: Hop count errors with and without the MoGA
applied to networks with individual beacon mobility
models.

186

5.1 Effects of Mobility on Hop Count Based Distance Estimation

Figure 5.16: Hop count errors with and without the MoGA
applied to networks with coupled beacon mobility
models.

187

Chapter 5 Localization in Mobile Ad Hoc Networks
values are originated in movements of neighbors is subject to investigation. Also, it is desirable to characterize mobility and its
impact on the hop count error values. If this can be achieved,
it could be achievable to find an appropriate adaptation of the
hop count values in order to reduce negative hop count errors in
dynamic networks. Figure 5.17 shows the average values for both
metrics for static and dynamic devices separately and averaged
over all considered speed and angle values. It becomes apparent
that ID-change is a strong indicator for whether a device has been
moved or not. Although HC-change does not seem to be a good
indicator to detect mobility, it is useful for its characterization.
When considering Figure 5.18, where the average HC-change values
are depicted for different angles and speeds, it becomes apparent
that HC-change values display a similar pattern as the error values
depicted in Figure 5.11. This suggests that there is a distinctive
relationship between mobility parameters, such as speed and direction of movements, and HC-change values. Furthermore, it
indicates that by applying an appropriate mapping method, HCchange can be used to characterize these parameters and to adapt
the hop count values accordingly in order to avoid underestimation.

5.1.4 Conclusion
This section investigates the impact of mobility on hop count based
distance estimation. In order to do so, a hop count error model
is introduced, which is capable of assessing the deviation of hop
counts from an ideal hop count distribution. This avoids examining
the different hop count based distance estimation algorithms individually and facilitates the evaluation of mobility induced effects
on hop count based distance estimation algorithms in general. A
thorough experimental study is performed, in which hop count

188

5.1 Effects of Mobility on Hop Count Based Distance Estimation

Figure 5.17: ID-change (5.17) for static and dynamic devices using
the individual beacon mobility model.

189

Chapter 5 Localization in Mobile Ad Hoc Networks

Figure 5.18: HC-change (5.18) for the individual beacon mobility
model (the speed of the movement is increased from
3.5% until 28% for each angle).

190

5.1 Effects of Mobility on Hop Count Based Distance Estimation
errors are determined for networks where devices move according
to different mobility models. The mobility models include some
models proposed in the literature, as well as a set of mobility
models designed specifically for this study.
From the experiments performed, it is revealed that mobility has
the effect to turn a natural overestimation of hop counts and, thus,
distances into an unpredictably high underestimation. Experiments
indicate that mobility of a certain speed and direction with respect
to beacons can compensate naturally positive hop count errors
in a network. Nevertheless, for situations in which mobility in a
network cannot be controlled, the overcompensation can negatively
affect the accuracy of distance estimates. Furthermore, traditional
heuristics to mitigate natural overestimation are not applicable
to dynamic networks without further ado. Hence, it is concluded
that the negative impact of mobility has to be contained. First
experimental results with various mobility patterns suggest that
direction, speed, and similarity of movements in neighboring regions
not only affect the hop count error in different ways, but also display
significant interdependencies, making their impact hard to predict.
Each impact factor is investigated separately in further experiments
and an increase in speed as well as movements leading away from
beacon nodes are found to increase the negative hop count error,
while movements directed towards beacons and a similarity of
movements between nearby devices are found to have a positive
impact on the hop count error.
Additionally, the concept of MoGA, an adapted version of the
standard GA, is introduced and its performance is evaluated in an
experimental study. The study shows that MoGA reduces negative
hop count errors in dynamic networks. MoGA is demonstrated to
reduce negative hop count error but does not eliminate it entirely.
Moreover, the reduction is constant relative to the error level, such

191

Chapter 5 Localization in Mobile Ad Hoc Networks
that a change in speed or direction still leads to a difference in hop
count error values. Two indicators are proposed, which are able
to assess characteristics of mobility in a network. Both indicators
can be computed in a decentralized way using only information
which is available locally from the messages exchanged in order
to determine hop counts. These mobility indicators are shown
to be able to help discovering and characterizing mobility in a
network. Hence, they could be used in order to improve distance
estimates by adapting hop counts accordingly. Such an adaptation
could, for example, be implemented using learning mechanisms of
the O/C Architecture in order to derive a mapping between these
indicator values and the resulting hop count error in a network by
training over time. This further emphasizes the need to integrate
learning mechanisms in the localization process for MANETs.

5.2 Synchronized Hop Counting in Mobile Ad
Hoc Networks
The previous section revealed that the accuracy of hop count based
distance estimation and, hence, the accuracy of distance-based
localization is influenced by passive mobility in the network such
that underestimation of distances can occur. As discussed in
Section 5.1.2, a negative hop count error is more aggravating than
a positive error for the following two reasons: Firstly, it is hard
to predict to what extent mobility causes a negative error, which
makes it difficult to counteract it, for example by adjusting the
hop count value. Secondly, most refinement techniques proposed
in the literature deal with eliminating the positive hop count
error due to low density and are not applicable to a network with
mobility induced underestimation without further ado. As a result,

192

5.2 Synchronized Hop Counting in Mobile Ad Hoc Networks
a negative hop count error is undesired and should be eliminated
if possible.
In the previous section, asynchronous computation of hop counts is
identified to be the main reason for the occurrence of a negative hop
count error. The logical consequence of this finding is to investigate
whether a MANET can be synchronized and if this reduces negative
hop count errors. These questions are addressed in the following.
Furthermore, a procedure is introduced in order to encode hop
count information in the timing at which a signal is sent instead
of transmitting it as content of a message. This method assumes
that the network is synchronized and that the mobile devices are
able to send signals at any time. The proposed algorithm allows
for the performance of distance estimation on devices which have
only basic communication abilities. Additionally, the resource
consumption while determining hop counts in MANETs can be
reduced significantly.
Since MANETs lack a CCU, a decentralized synchronization technique has to be applied. As discussed in Section 2.2.3, nature is a
good source of inspiration when looking for algorithms which can
be applied to a distributed network without central control. For the
task of synchronization, a natural role model can be found as well.
In Southeast Asia, fireflies provide an impressive natural spectacle.
Fireflies are also called “lightning bugs” due to their usage of bioluminescence. At dusk, thousands of these insects gather in trees and
emit light-flashes to attract mates or prey, reaching almost perfect
synchrony after some time (cf. Buck [28]). This phenomenon gives
name to a class of algorithms called “firefly-algorithms”, which can
model the emergence of synchrony in distributed systems. Charles
S. Peskin was the first person to mathematically describe this phenomenon of pulse-coupled oscillators in [181]. In his model, each
oscillator emits a signal at the end of a fixed time period. This

193

Chapter 5 Localization in Mobile Ad Hoc Networks
process is referred to as firing. When another oscillator observes
such a firing signal, it reduces the remaining time until its next
own firing. Figure 5.19 illustrates this mechanism. The gray and
the black dots are two oscillators and the circle represents the
fixed time period between the firings. The time period is equal for
both oscillators but before synchronization, they start at different
points in this time period. As time passes, the oscillators move
along the circle and whenever they reach the point at the top of
the circle, marked with a small line, they fire. Such firing causes
the other oscillator to advance its phase by a certain value which
is dependent on its current phase value. After two firings of the
black oscillator and one firing of the gray oscillator, both oscillators
are synchronized. Other examples of synchronization processes
in nature following the same principle are heart pace-maker cells
(cf. Peskin [181], Torre [231]), crickets chirping in synchrony (cf.
Walker [240]), organized bursting in pancreatic beta-cells (cf. Sherman et al. [216]), and even female menstrual cycles, which tend to
synchronize (cf. McClintock [148], Russell et al. [200]).
To transfer this concept to technical devices, it has to be noted that
a device’s system clock is usually equipped with an oscillator (cf.
Sivrikaya and Yener [217]). This oscillator determines the clock’s
frequency by creating almost harmonic oscillations with period
T . The state of an oscillator is described by its phase ϕ. When
the phase is normalized between 0 and 1, it denotes the elapsed
percentage of the oscillator’s period and can, thus, be expressed in
terms of time t (cf. Figure 5.20). Oscillators are said to be identical
if their periods T have the same length. When, additionally, both
periods start at the same time, i.e., their phase ϕ is identical for
any point in time, the oscillators are called synchronous.

194

5.2 Synchronized Hop Counting in Mobile Ad Hoc Networks

Figure 5.19: Illustration of the synchronization of two oscillators
depicted as black and gray dots. The circle
represents the time between two firings (point at the
top of the circle). As time passes, the oscillators
move along the circle and the firing of one oscillator
reduces the time remaining until the firing of the
other oscillator as a function of its current phase
value (indicated by arrows). Both oscillators are
synchronized after three firings.

195

Chapter 5 Localization in Mobile Ad Hoc Networks

Time t

Figure 5.20: Relation between a normalized phase ϕ of an
oscillator and time t.

5.2.1 Pulse-coupled Oscillators
Firefly-algorithms can be used to synchronize clocks and have
already been used successfully in MANETs to implement energyefficient communication mechanisms by Taniguchi et al. [229],
Wakamiya et al. [239], and Wakamiya and Murata [238]. The
basic algorithmic model for firefly-inspired synchronization was
defined by Mirollo and Strogatz [161], in which each oscillator is
described by a state variable x ∈ [0, 1]. The phase of an oscillator
is mapped to the oscillator’s state by a function f : ϕ → x, which
is required to be increasing and strictly concave. The mapping
from phase to state is introduced in order to be able to describe the
sensitivity of the phase adjustment as a function of the oscillator’s
phase at the time of an observed firing. This function is further on
referred to as state-function. When the state x reaches the value 1,
the oscillator fires and x is reset to 0. In the work of Mirollo and
Strogatz [161], the state-function is defined as:
x (i) = 1 ⇒ x (j) ← min (1, x (j) + ) , ∀j
= i

196

(5.8)

5.2 Synchronized Hop Counting in Mobile Ad Hoc Networks
When an oscillator i fires, all other oscillators increase their state by
the value of , which is denoted as impulse intensity, until they reach
a maximum value of 1 (cf. Equation 5.8). The corresponding phaseshift is determined by the state-function. With this simple rule,
any two oscillators can be synchronized and for n > 2 oscillators,
there are few combinations of initial phase values which do not
result in synchrony if the impulse intensity is non-negative and
non-zero for all oscillators (cf. Mirollo and Strogatz [161]). In the
work of Mathar and Mattfeldt [147], convergence is also shown
for a state-function f which is non-differentiable. Further, the
concavity of the state-function is abandoned and a linear statefunction is shown to be sufficient for nearly all systems to converge
if multiple synchronous firings are accounted for by an increased
impulse intensity  and  ≥ 1/n holds (cf. Bottani [21]). Lucarelli
and Wang [144] demonstrate that neighborhood-coupling, where
only firing of neighbors induces a phase shift of the device, is
enough for the convergence of a system. For this, they use a linear
state-function f (ϕ) = ϕ and a new update model where the state
advancement is sensitive to the current state of the affected device
by assigning a higher weight to flashes received when the device
itself is close to firing:
x (i) = 1 ⇒ x (j) ← min (1, x (j) +  · x (j)) , ∀j
= i

(5.9)

Due to the characteristics of the state-function f , state x and phase
ϕ are the same and Equation 5.9 can also be written as:
ϕ (i) = 1 ⇒ ϕ (j) ← min (1, ϕ (j) +  · ϕ (j)) , ∀j
= i

(5.10)

With this setting, it has been shown experimentally that synchronization can also be achieved in dynamic networks even if

197

Chapter 5 Localization in Mobile Ad Hoc Networks
connectivity is temporarily lost. In publications of Tyrrell et al.
[233] and Werner-Allen et al. [247], delays in firing perception
are examined and two algorithms are proposed which can provide
synchronization under these circumstances. Although message
delays are an issue in MANETs, the following study is based on
the algorithm published by Lucarelli and Wang [144] as a simple synchronization algorithm is sufficient to investigate the effect
which synchronization has on distance estimates.

5.2.2 Firefly-inspired Hop Counting
In distributed computing systems, message exchange is a time and
resource consuming process, which decreases the lifetime of the
devices involved (cf. Akyildiz et al. [4], Dietrich and Dressler [52]).
Since hop counting is an essential technique for both, routing (cf.
Boukerche et al. [22], Chatterjee and Das [38], Chou et al. [42]) and
distance estimation (cf. Section 2.4.2), it is desirable to reduce the
number and length of messages exchanged. Firefly-inspired Hop
Counting (FIHC) is an approach to encode hop count information
in the timing of a binary signal which has only two states, on and
off, instead of exchanging messages with the hop count value as
it is done in GA (cf. Section 2.4.2). In addition to the resource
saving potential, this approach enables devices which only possess
simple communication capabilities to determine hop counts, which
could be useful in scenarios where devices are only equipped with
rudimentary hardware, such as in smart dust applications (cf. Ilyas
and Mahgoub [105]).
FIHC requires the network to be synchronous. For FIHC, the
devices are assumed to have identical oscillators, i.e., the period of
their timers is equal. This assumption is reasonable, since ad hoc
networks are often described as containing only identical devices.

198

5.2 Synchronized Hop Counting in Mobile Ad Hoc Networks
As detailed before, the firefly-algorithm described by Lucarelli and
Wang [144] can be used for synchronization of such a network. For
FIHC, a second signal is used, which is sent by the device with
a specific delay to a synchronized signal in order communicate
the device’s encoded hop count information. When the network is
synchronized, the delay of the second signal with respect to the
synchronized signal represents the hop count of the device. Hence,
all devices with the same hop count emit this signal at the same
time, devices with a lower hop count emit the signal beforehand
and devices with higher hop counts emit the signal afterwards.
Further on, the signal used for synchronization is called sync-signal
and the delayed signal used to encode hop count information is
called HC-signal. In order to be able to use the timing of a signal
to transport information, a method for encoding and decoding has
to be provided.
Encoding
A device i with a hop count value of h (i) sends its HC-signal, when
the phase of the sync-signal has the value
ϕsync =

h (i)
.
h (i) + 1

(5.11)

For example, a device which is two hops away from the beacon
emits its HC-signal when the phase of its sync-signal has the value
2
of 2+1
= 23 . Devices with different hop counts are required to emit
their HC-signal at different times, which holds for this encoding.
Let h (i) ≥ 0, h (j) ≥ 0 be the hop counts of device i and j, and
h (i)
= h (j). It then holds that:

199

Chapter 5 Localization in Mobile Ad Hoc Networks

h (i)
= h (j)
⇒ h (i) + (h (i) · h (j))
= h (j) + (h (i) · h (j))
⇒ h (i) · (1 + h (j))
= h (j) · (1 + h (i))
h (i)
h (j)
⇒

=
h (i) + 1
h (j) + 1
Decoding
For decoding, the following equation is used. Let 0 ≤ ϕsync < 1 be
the value of the common sync-signal’s phase when device j receives
the first HC-signal from its neighbor i. From this information,
device j can compute its own hop count as:
h (j) =
Because of ϕsync =

h(i)
h(i)+1 ,

ϕsync
+1
1 − ϕsync

(5.12)

it holds that:
h(i)

ϕsync
h(i)+1
+1⇔
+ 1 ⇔ h (i) + 1
h (j) ⇔
h(i)
1 − ϕsync
1 − h(i)+1
Figure 5.21 shows an example of four devices determining their hop
counts, assuming that they are already synchronized. The beacon
starts by sending an HC-signal when the sync-signal phase has
the value ϕsync = 0. When the next device receives this signal its
sync-signal phase is also at a value of 0 (since they are synchronous)
and the device, therefore, has a hop count value of 1 and emits
its HC-signal at sync-phase ϕsync = 12 . The next device calculates
its own hop count to have a value of 2 and emits its signal at

200

5.2 Synchronized Hop Counting in Mobile Ad Hoc Networks

Beacon

Time

Time

Time

Time

Figure 5.21: Example of the FIHC over three hops via intentional
phase shifting of a signal with respect to a
synchronized reference signal.

201

Chapter 5 Localization in Mobile Ad Hoc Networks
ϕsync = 23 . Analogously, the fourth device will send its HC-signal
at sync-phase value ϕsync = 34 .
One benefit of using timed signals to encode hop counts is that the
order in which the signals are sent corresponds to the ascending
order of hop counts. Assuming that there is no delay or messageloss, each device can determine its hop count value within one
period. The asynchronous exchange of messages in the GA, on
the other hand, can take many periods (depending on the network
size) until all devices in the network have determined their hop
counts. Additionally, each device knows its hop count immediately
after the first reception of an HC-signal from its neighbors and can
ignore all other signals for the rest of the period. During this time,
listening to signals is no longer required and energy can be saved.
In contrast, the GA requires all devices to be permanently receptive
to new messages sent by neighbors. The FIHC algorithm can be
designed in two variants, FIHC or Firefly-inspired Hop Counting
with Delay (FIHCd). With FIHC each device emits an HC-signal
when the phase of its timer corresponds to the device’s encoded
hop count. The firing is conducted whether or not the device has
already received a signal from a neighbor in its current period.
In case it receives a signal after its own firing, the hop count is
adapted and the device fires again. With FIHCd, a device only
fires after having received a signal from another device, except for
the beacon, which always fires at phase 0. In the delayed version,
the firing frequency is limited to one HC-signal per device per
period, which reduces the communication overhead even further.

5.2.3 Evaluation
The first set of experiments investigates the synchronization success
and duration using the model proposed by Lucarelli and Wang [144]

202

5.2 Synchronized Hop Counting in Mobile Ad Hoc Networks
in a neighborhood-coupling setting. In a second set of experiments
the distance estimates in a synchronized network using FIHC are
compared to results using the asynchronous, message-based GA.
Scenario and Settings
The experiment setting is the same as described in Section 5.2.3,
except for the placement of the static beacon, which is now placed
in the center of the environment. 1000 devices are randomly placed
in the environment and the impulse intensity  is tested for values
 = 0.05, 0.10, and 0.15. The communication range r is varied from
5% to 20%. One simulation cycle is equivalent to the period T ,
which is divided into 100 time slots. Devices listen to signals and
messages during the whole period and for the GA they send their
hop count messages at the end of their timers’ periods. Figure 5.22
shows the simulated scenario. Non-beacons in the network move
according to the CM mobility model described in Section 5.1.1,
where each device moves with a probability of pm = 0.01 in each
time slot. Movements which would lead outside of the plane are
not executed.
Synchronization Success
In this section, the effectiveness of the synchronization algorithm
is investigated in terms of its ability to successfully synchronize
the entire network. For this purpose, an experiment is classified as
unsuccessful if complete network synchrony is not achieved within
300 cycles. Each experiment is repeated 40 times and the results
are averaged.
Figure 5.23(a) and 5.23(b) show the percentage of successfully
synchronized experiments for static and dynamic networks respec-

203

Chapter 5 Localization in Mobile Ad Hoc Networks

Figure 5.22: Network model for the simulative experimental study.
The beacon is placed in the center of the square
environment. Devices with the same color have the
same hop count value.

204

5.2 Synchronized Hop Counting in Mobile Ad Hoc Networks
tively. Figure 5.24 displays the time elapsed until the networks are
in synchrony for both, static and dynamic networks.
It is noticeable that the impulse intensity has a strong impact
on the synchronization success in both cases. For a value of
 = 0.1, the highest success rates are achieved with up to 100% for
communication ranges larger than 9% in static networks and 16% in
dynamic networks. The experiments indicate that the performance
decrease for  = 0.15 when compared to  = 0.1 could be caused
by an increase in synchronization time. This is not the case for
 = 0.05, however. In addition, synchronization time has a high
dispersion for  = 0.15, which confirms that this impulse intensity
is not suitable to synchronize the regarded network. It seems that
 = 0.1 is the best choice in both scenarios because it guarantees
a high success rate and simultaneously low synchronization time,
which is why this setting is used for the experiments investigating
FIHC. An increase in the communication range r improves the
time until successful synchronization for all impulse intensities
considered. Mobility in the network makes the synchronization
less likely. Nevertheless, with the right choice for impulse intensity
and a sufficiently large communication range the network can be
synchronized.
When comparing the results of static and dynamic networks, it
becomes apparent that the percentage of successfully synchronized
experiments in dynamic networks is lower than in static networks
when considering smaller communication ranges. The explanation
for this is straight-forward. Synchronization is achieved by mutual influences between neighbors. In dynamic networks, devices
sometimes move before this process is finished and then engage in
a synchronization process which is possibly at a different state as
the previous one. This can have the effect that the signal of the
moved device interferes with this new synchronization process and

205

Chapter 5 Localization in Mobile Ad Hoc Networks

(a) Static networks.

(b) Dynamic networks.

Figure 5.23: Percentage of successful experiments in static (a) and
dynamic (b) networks with 1000 devices and a
varying communication range r.

206

5.2 Synchronized Hop Counting in Mobile Ad Hoc Networks

(a) Static networks.

(b) Dynamic networks.

Figure 5.24: Average time until successful synchronization of
static (a) and dynamic (b) networks with 1000
devices and a varying communication range r.

207

Chapter 5 Localization in Mobile Ad Hoc Networks
causes delays. On the other hand, the absence of the moved device
in its old neighborhood can also be the cause for a slowdown in the
previous synchronization process. When the communication range
increases, a newly added or removed device in a neighborhood has
comparatively less impact since more devices are involved in each
synchronization process.
Comparison of Firefly-inspired Hop Counting and
Asynchronous Gradient Algorithm
The aim of the next experiment is to evaluate the performance
of the FIHC and FIHCd algorithms and to compare the resulting
distance estimates with the asynchronous GA. For this, the average
hop count error is computed over 40 cycles starting after 10 cycles in
order to exclude the phase in which hop count values are propagated
through the network for the first time. The devices are simulated
with a communication range of r = 7%, which corresponds to the
settings in the experiments conducted in Section 5.1.3. Results are
presented in Figure 5.25. FIHC has the lowest hop count error,
followed by GA, and the highest error occurs when using FIHCd.
Although this result creates the impression that FIHC should be
the algorithm of choice for hop count based distance estimation in
MANETs, this is true only to a certain extent.
In order to get a better insight into the deviations in performance,
Figure 5.26 shows the average percentage of devices with which a
negative or positive hop count error occurs. It becomes apparent
that the performance of the algorithms is reflected by the amount
of underestimation in the network. The higher the underestimation
is, the lower is the overall hop count error. This is due to the fact
that underestimation compensates for the natural density induced
overestimation in networks. However, as discussed in Section 5.1.2,
underestimation is the more aggravating type of error for localiza-

208

5.2 Synchronized Hop Counting in Mobile Ad Hoc Networks
tion since it cannot be predicted and prevents the application of
traditional refinement methods without adjustment, which are proposed to encounter low density issues. As a consequence, FIHCd
is the best option in dynamic networks because it results in the
lowest amount of negative hop count error. As the experiments in
Section 5.1.3 show, movements leading away from the beacon into
higher gradient rings create high negative hop count error due to
the distortion the moved devices create in their new neighborhoods.
FIHCd counteracts this effect because hop count signals are emitted only after a signal was received first. Thus, a device which
moves away from the beacon does not affect the new neighborhood
since it does not communicate its underestimated hop count before
updating. The remaining underestimation can be explained by the
fact that a device which has moved to a higher gradient ring underestimates its own distance before updating. Nevertheless, with
FIHCd the negative hop count error can be reduced significantly,
which has the advantage of being able to treat a dynamic network
just like a static one.

5.2.4 Conclusion
In this section, FIHC and FIHCd are proposed, two algorithms
for hop counting in MANETs, which are based on synchronized
timers and, based on that, timed sending of signals. The basic
idea of FIHC is to use an intentional phase shift with respect to a
synchronous base signal in order to encode information about the
device’s hop count with respect to a beacon. To realize this idea, the
nature-inspired firefly algorithm by Lucarelli and Wang [144] is used
for synchronization. Experiments are performed, which investigate
the success rate and duration of synchronization of this method
under the constraint of neighborhood-coupling, i.e., only the firing

209

Chapter 5 Localization in Mobile Ad Hoc Networks

Figure 5.25: Comparison of hop count errors in distance
estimation using FIHC, FIHCd, and asynchronous
GA (async GA).

Figure 5.26: Percentage of the devices in the network which
overestimate and underestimate their distance to the
beacon for FIHC, FIHCd, and asynchronous GA
(async GA).

210

5.2 Synchronized Hop Counting in Mobile Ad Hoc Networks
of direct communication partners has an impact on a device’s
synchronization process. The communication range is varied in
order to determine its impact on synchronization performance and
three different impulse intensities are testes. Subsequently, both
variants of the proposed signal-based hop counting approach, FIHC
and FIHCd, are evaluated and compared to results achieved by
the standard GA. The difference between FIHC and FIHCd is
that in the latter variant devices have to wait before sending an
own signal until after having received a signal from any nearby
device, whereas with FIHC, the device simply sends its signal at
the determined phase shift.
Synchronization success is shown to be 100% on average when
impulse intensity is set to  = 0.1 and the communication range
is at least 9% in static and 16% in dynamic networks. In general,
an increase in the communication range positively influences the
synchronization success and the time it takes for the network to
be in synchrony. Although mobility in the network slightly hinders synchronization, the produced synchronization delay, can be
compensated by an increase of the communication range. Experiments further demonstrate that FIHCd decreases the emergence
of negative hop count error in dynamic networks. As discussed in
the previous section, this is essential for an improved accuracy of
localization based on hop counts in dynamic networks since underestimation cannot be handled as well as the natural occurrence of
overestimated distances. FIHCd, thus, provides an alternative for
the previously presented MoGA algorithm in order to eliminate
underestimation in dynamic networks if synchronization can be
achieved.

211

Chapter 5 Localization in Mobile Ad Hoc Networks

5.3 Geometric Distance Estimation for Mobile
Ad Hoc Networks
In the previous two sections, the behavior of distance estimation
algorithms based on hop counts in mobile networks is subject to
investigation. It is found that mobility has a negative impact on the
quality of the derived distance estimates and that the height of this
negative impact is hard to predict. In this section, an alternative
distance estimation approach is presented which refrains from hop
counts as a basis. The approach is called Geometric Distance
Estimation (GeoDE) and belongs to the second type of rangefree distance estimation algorithms addressed in Section 2.4.2, the
connectivity-based distance estimation techniques.

5.3.1 Connectivity-based Distance Estimation
In order to understand the concept of connectivity-based distance
estimation, the notions of shared and individual neighbors have to
be introduced, which are graphically illustrated in Figure 5.27.
Definition 5.7 (Classification of Neighbors). Let i, j be two adjacent devices and N (i), N (j) the sets of devices situated in the
neighborhood of i and j respectively. The neighbors of i can be
categorized with respect to j as:
shared neighbors: S (i, j) := N (i) ∩ N (j)
individual neighbors: Ij (i) = N (i) \S (i, j)
In connectivity-based distance estimation, the number of shared
neighbors between two devices is used to approximate the intersection area A of the two circles which represent the communication

212

5.3 Geometric Distance Estimation for Mobile Ad Hoc Networks

Figure 5.27: An example of the communication range of two
adjacent devices i, j. Devices with dotted lines are
communication partners of device i and belong to the
neighborhood N (i). Gray filled devices are
communication partners of device j and,
consequently, belong to the neighborhood N (j).
Devices in the shaded area are shared neighbors and
belong to the set of shared neighbors S (i, j).

213

Chapter 5 Localization in Mobile Ad Hoc Networks
ranges of both devices. Figure 5.28 displays the geometric conditions of intersecting circles. In reality, the communication range
is not exactly circular, but this assumption is a commonly used
simplification in the modeling of ad hoc networks, which is called
the unit-disc-graph model (cf. Aspnes et al. [11], Breu and Kirkpatrick [24]). The approximation of the intersection area follows
the principle of a Monte Carlo integration. Monte Carlo integration
approximates the size of a shape’s surface by randomly choosing
points and determining the proportion of these random points
which lie inside the shape relative to the points outside the shape
(cf. Evans and Swartz [61]). The standard Circle-Circle Intersection Equation 5.13 establishes a relation between the distance d
of the two circles’ centers and the intersection area A. Solving
Equation 5.13 for d would allow to derive an estimate for the
distance d from a given approximation of the intersection surface
A and the range r. GeoDE provides a method for deriving such
a mapping between an approximation of A and an estimate for
the distance d. It should be noted that similar connectivity-based
distance estimation approaches have been pursued by Aslam et al.
[10], Buschmann et al. [33], Fekete et al. [64], Villafuerte et al.
[236], and Huang et al. [99]. However, they differ in the way how
the mapping function is derived and do not consider multi-hop distance estimation which is required in the context of beacon-based
localization.


d
A = 2r2 arccos
2r

214





− d r2 −

d2
4

(5.13)

5.3 Geometric Distance Estimation for Mobile Ad Hoc Networks

Figure 5.28: Geometric characteristics of two overlapping circles.
The intersection area A corresponds to the shaded
area, the two black dots represent the circles’ centers.

5.3.2 Geometric Distance Estimation
In order to be able to approximate a mapping between the intersection area A and the distance d which is independent of the
communication range r, the following considerations are made.
The distance d between the centers of two overlapping circles can
be described as a ratio θ of the circles’ diameters according to
Equation 5.14. The intersection area A can be described as a ratio
∆ of the circle’s surface as shown in Equation 5.15.
θ =1−

d
2r

(5.14)

A
(5.15)
πr2
Standard equations (5.16) and (5.17) are alternative descriptions
of A and d which express both values using the communication
range r and the segment angle α (cf. Figure 5.28). Substituting A
and d in Equations 5.14 and 5.15 by the expressions from Equation
5.17 and 5.16, it becomes apparent that ∆ and θ only depend on
∆=

215

Chapter 5 Localization in Mobile Ad Hoc Networks
the segment angle α, which is independent of the communication
range r.
 

d = 2r cos

α
2

(5.16)

A = r2 (α − sin (α))

(5.17)

As a consequence, the relation between θ and Δ is independent of
the value r and can be approximated as a third-degree polynomial
function using linear regression. Figure 5.29 shows the approximation of the mapping function f : Δ → θ (dotted line). Putting it
all together, the distance d can be calculated from Δ according to
Equation 5.18 and Δ is estimated for device i as shown in Equation
5.19. This results in the final Equation 5.20, which device i uses
in order to compute an estimate for its distance to device j.
d = 2r (1 − f (Δ))
Δ≈

(5.18)

|S (i, j) |
|N (i) |



(5.19)


d¯GeoDE (i, j) = r · a · Δ̄3 + b · Δ̄2 + c · Δ̄ + e
With Δ̄ =

|S(i,j)|
|N (i)|

a = 3.90

216

and corresponding coefficients:
b = −4.16

c = 3.04

e = 0.04.

(5.20)

5.3 Geometric Distance Estimation for Mobile Ad Hoc Networks

Figure 5.29: Relation of θ to Δ and the approximated third-degree
polynomial function f , which maps Δ to θ.

217

Chapter 5 Localization in Mobile Ad Hoc Networks
Approximation Error
There are two approximations in GeoDE which influence the accuracy of the computed distance estimates. Firstly, the approximation
of Δ as the ratio of shared to total communication partners and,
secondly, the approximation of the function f through polynomial
regression. Considering the first approximation, there are two
sources of error in this approximation. Monte Carlo algorithms
base their concept on random numbers that are selected uniformly.
Networks, however, do not necessarily have an even distribution
of devices. A shift in the positions of the neighbors into a certain
direction, shown exemplarily in Figure 5.30, can distort the approximation results. The influence of various device distributions on
the estimation accuracy is, therefore, examined in experiments in
Section 5.3.3. A second problem is the sampling rate. For |N (i) |
neighbors, there are only |N (i) | + 1 different values for Δ. As a
consequence, the absolute error for the estimation of Δ, assuming
the ratio of neighbors indeed reflects the overlap size, lies within
the interval of [0, |N1(i)| ).

Figure 5.30: Example of an irregular distribution of neighboring
devices over the communication range.

218

5.3 Geometric Distance Estimation for Mobile Ad Hoc Networks
The second approximation is that of function f . Figure 5.31 shows
the deviation between f (Δ) and θ. The approximation error with
a first order Taylor series expansion, as it is used by Huang et al.
[99], is depicted for comparison. For θ < 0.9 the error using
polynomial regression is smaller compared to the first order Taylor
series approximation. The approximation error is at most 0.04,
which leads to a maximum absolute distance estimation error of
0.16r. The error value depends on the exact value of θ, which
in turn depends on r. It follows that GeoDE estimates the same
distance with slight variations in accuracy when the communication
range r changes.

Figure 5.31: The approximation error for function f , which maps
Δ to θ, derived by a third-degree polynomial
regression compared to the approximation error when
f is derived by a first order Taylor series expansion.

219

Chapter 5 Localization in Mobile Ad Hoc Networks
Geometric Distance Estimation over Multiple Hops
Although GeoDE could be applied directly to two-hop neighbors,
as well as one-hop neighbors, this is not pursued here due to
the additional communication overhead. The two devices which
need to estimate their distance would exchange all the necessary
information via a common communication partner, which burdens
that device’s resources. Nevertheless, two-hop estimation can be
performed analogously and could be useful in situations in which
accuracy of distance estimates is prior to resource saving. So far,
distance estimates d¯(i, j) can range between 0 and 2r. The first
refinement technique used in GeoDE is to limit calculated estimates
to a maximum value of r. The second refinement results from the
consideration that the number of neighbors |N (i) | and |N (j) |
of the two devices i and j can differ, even in uniform-randomly
distributed networks (cf. Figure 5.27 for an example). As a result,
device i and device j can compute different estimates of their
distance using Equation 5.20. Hence, device i and j exchange their
estimates via communication and compute the average value of
the estimates d¯(i, j) and d¯(j, i).
Algorithm 5.1 shows how a device i estimates its distance to a
neighbor j using the GeoDE approach (note that because of i and
j being neighbors it holds that |N (i) | > 0).
As stated earlier, the distance between neighboring devices is not
enough for the computation of coordinates with beacon-based
localization algorithms. Instead, the distance to beacons has to
be known, which can be multiple hops away. Algorithm 5.2 shows
how GeoDE is expanded to allow for distance estimations between
devices in a network and beacons. The algorithm can be executed
repeatedly and the necessary iterations for all devices in the network
to derive a distance estimate are subject to the neighborhood size
and the total number of devices in the network. GeoDE is similar

220

5.3 Geometric Distance Estimation for Mobile Ad Hoc Networks

Algorithm 5.1 Geometric Distance Estimation to Neighbors
Require: N (i), neighbors from device i, a neighbor j, and communication range r
Ensure: Estimated distance d¯(i, j)
1: Ask list of neighbors N (j) from device j.
2: S (i, j) ← N (i) ∩ N (j)
|S(i,j)|
3: x ← |N (i)|


4: d¯(i, j) ← r · 3.90 · x3 − 4.16 · x2 + 3.04 · x + 0.04
//Limitation:


5: if d¯(i, j) > r then
d¯(i, j) ← r
6:
end if
//Averaging:
8: Ask j for its 
non-averaged distance
estimate d¯(j, i)

1
9: d¯(i, j) ← 2 · d¯(i, j) + d¯(j, i)
10: return d¯(i, j)
7:

221

Chapter 5 Localization in Mobile Ad Hoc Networks
to the distance estimation approach based on hop counts. The
difference is that each device computes its own distance estimate
to a beacon on the basis of the minimum value any of its neighbors
has computed for its distance to the beacon. In contrast to hop
count based distance estimation, where the distance estimate is
computed on the basis of the minimum hop count value in the
neighborhood.

5.3.3 Evaluation
The objective of the following experiments is to evaluate GeoDE in
comparison with distance estimation based on hop counts. Firstly,
the sensitivity of GeoDE to variations in the communication range
r and the network’s distribution is examined in static networks.
In this first experiment, the results of GeoDE are compared to a
simple distance estimation approach based on hop counts proposed
by Nagpal et al. [171]. The equation to derive distance estimates
with this approach is shown in Equation 5.21, with hb (i) denoting
the hop count of device i with respect to beacon b.


d¯SGM (i, b) = r

j∈N (i) hb (j)



+ hb (i)
− 0.5
|N (i) | + 1

(5.21)

In the second set of experiments, the performance of GeoDE is
investigated in dynamic networks and compared to a state-ofthe-art distance estimation method based on hop counts, which
is more complex than the one presented by Nagpal et al. [171].
This method is called Gradient-based Distance Estimation (GDE)
and is published by Liu et al. [140]. The computations include
additional statistical considerations, which are based on the work
from Nagpal et al. [171] and Kleinrock and Silvester [122]. The

222

5.3 Geometric Distance Estimation for Mobile Ad Hoc Networks

Algorithm 5.2 Geometric Distance Estimation to Beacons
Require: N (i) neighbors of device i, beacon b, communication
range r
Ensure: estimated distance d¯(i, b)
1: if (b ∈ N (i)) then
d¯(i, b) ← compute d¯(i, b) using Algorithm 5.1
2:
3: else
//search for neighbor k closest to b:
D̄ ← { }
4:
5:
for j ∈ N (i) do
6:
ask j for d¯(j, b)
7:
if d¯(j, b)
= null then
8:
add d¯(j, b) to D̄
9:
end if
10:
end for
11:
if (D̄ ← { }) then
12:
return null
13:
else
 
14:
Select neighbor m: d¯(m, b) ← min D̄ closest to b
d¯(i, m) ← compute d¯(i, m) using Algorithm 5.1
15:
d¯(i, b) ← d¯(i, m) + d¯(m, b)
16:
end if
end if
19: return d¯(i, b)
17:
18:

223

Chapter 5 Localization in Mobile Ad Hoc Networks
necessary calculations to derive a distance estimate d¯GDE (i, b)
between beacon b and device i according to Liu et al. [140] are
shown in Equation 5.22.
out
in
d¯GDE (i, b) = r · (Gout
(i) + Gin
b (i) · R
b (i) · R (i)

(5.22)

+(hb (i) − 1) · dhop (i) · Δ(R))
with:

1 ¯SM G
(i, b)−(hb (i) − 1)·(1 − dhop (i))−0.5·(1 − dhop (i))
Gout
b (i) = d
r
Rout (i) = 1 − Rin (i)
out
Gin
b (i) = Gb (i) + 1

ΔR (i) =

 1
0



x

Rin (i) = 1 − dhop (i)
e(x−1)

|N (i) | − 2x2 + 4x − 1
√
√
2
2x−x +|N (i)| arccos(1−x)

|N (i)|

dhop (i) = 1 + e

−

 1
−1

e−

|N (i)|
π



2x − x2

√

(arccos t−t

1−t2 )

dx

dt

In the work of Liu et al. [140], a method is proposed to account
for mobility when computing distance estimates. However, for this
method the expected mobility pattern has to be known, which is
not the case in the scenario of passive mobility considered here.
Hence, this method is not applied in the experiments. GDE relies
on the computation of numerical integrations. Different methods
have been proposed for approximation of numerical integration
in the literature (cf. Isaacson and Keller [108]). Here, Riemann
integration (cf. Riemann [196]) is used, where the surface area is
approximated by the sum of the surfaces of multiple boxes. For
the experiments, the numerical integration is approximated using
1000 integration steps (boxes).

224

5.3 Geometric Distance Estimation for Mobile Ad Hoc Networks
Geometric Distance Estimation in Static Networks
For the experiments, the settings described in Section 5.1.3 are
used with the difference that a beacon is randomly chosen from
the set of devices in each repetition of an experiment. This modification is necessary, in order to increase the chance of creating
a connected network, especially for distorted device distributions.
Since communication range, neighborhood size, and the distribution of devices are identified to influence the quality of GeoDE,
three static network scenarios are considered. In Scenario 1, the
devices are distributed according to a uniform random distribution.
In Scenario 2, the devices are distributed using a Gaussian random
distribution and in Scenario 3, all devices are placed like in a
grid (cf. Figure 5.32). In addition, the communication range r
is varied to show the influence of the neighborhood size on the
estimation error (cf. Section 5.3.2). Mobility is not considered yet.
To evaluate the quality of the estimates, Mean absolute percentage
error (MAPE) is computed as described in Equation 5.23.


|d (i, j) − d¯(i, j) |
M AP E d¯(i, j) =
(5.23)
d (i, j)
where d (i, j) denotes the Euclidean distance between a device i
and its neighbor j and d¯(i, j) the estimate of that distance. MAPE
gives information about the relative deviation of the estimate with
respect to the real distance.
Figure 5.33 shows the MAPE for Scenario 1, Figure 5.34 shows
the results for Scenario 2, and Figure 5.35 for Scenario 3. It can
be observed that GeoDE leads to less error-prone estimates than
the estimation based on hop counts for all considered distributions
and communication ranges. Furthermore, it can be noted that
even the sample standard deviation is less or equal to the MAPE
of estimates based on hop counts. This confirms that the GeoDE

225

Chapter 5 Localization in Mobile Ad Hoc Networks

(a) Scenario 1

(b) Scenario 2

(c) Scenario 3

Figure 5.32: Different network structures used for the
experimental study. Devices are placed in the
environment according to a uniform random
distribution (a), a Gaussian random distribution (b),
and are evenly distributed in a grid-like fashion (c).
approach is a consistent improvement for distance estimation in
static networks compared to distance estimation based on the
approach from Nagpal et al. [171].

Figure 5.33: MAPE for GeoDE compared to distance estimation
based on hop counts on long distance estimation
including standard sample deviation in Scenario 1.

226

5.3 Geometric Distance Estimation for Mobile Ad Hoc Networks

Figure 5.34: MAPE for GeoDE compared to distance estimation
based on hop counts on long distance estimation
including standard sample deviation in Scenario 2.

Figure 5.35: MAPE for GeoDE compared to distance estimation
based on hop counts on long distance estimation
including standard sample deviation in Scenario 3.
Despite the imbalanced distribution of devices, GeoDE performs
slightly better in Gaussian networks compared to networks which
are distributed according to uniform random distributions. This
can be explained by two observations. Firstly, the averaging of

227

Chapter 5 Localization in Mobile Ad Hoc Networks
estimates from both involved devices could help to overcome the
bias introduced by distorted device distributions. An unbalanced
distribution of devices leads to an overestimation in one device
and an underestimation in the other device, which can provide a
good estimate on average. Another factor is the larger average
neighborhood size for most devices due to the concentration of
devices in the center of the environment. As discussed above, this
increases the number of values for Δ and, as a result, improves
the estimates. Furthermore, it can be observed that the behavior
of the estimation error is different from Scenario 2 for increasing
communication range. After an initial decline, the error starts
increasing again. This characteristic can be explained by a higher
shift-sensitivity with larger communication range. A small communication range only covers a small area of the network in which
the distribution of devices is not as distorted as when looking at a
larger area. Nevertheless, the proposed averaging technique seems
to be able to keep the overall error to a similar level compared to
uniform-randomly distributed networks.
For the grid-like distribution shown in Figure 5.35, one would
expect a similar behavior as in Scenario 1, because the distribution
of devices is even in both networks. The trend of the error behavior
with increasing communication range indeed is similar to the error
behavior in uniform-randomly distributed networks. The oscillation
can be explained by the step-like increase of the neighborhood size.
Due to the grid-like distribution, increasing the communication
range does not change the neighborhood density until suddenly
several new neighbors are included. As a consequence, the error
of GeoDE, as well as the hop count based distance estimation,
changes erratically.

228

5.3 Geometric Distance Estimation for Mobile Ad Hoc Networks
Geometric Distance Estimation in Mobile Networks
In order to compare localization results on the basis of hop counts
with results based on GeoDE, lateration is used to determine locations from the distance estimates to beacons as described in
Section 2.4.2. Algorithm 5.3 shows the procedure of lateration.
Let i be the device which is to located and B (i) the set of known
beacons, C (i) = {c (b)}, ∀b ∈ B (i), with c (b) = (x (b) , y (b)) denotes the set of all known two-dimensional beacon coordinates, and
D̄ (i) = {d¯(i, b)}, ∀b ∈ B (i) is the set of all corresponding distance
estimates. Algorithm 5.3 shows how the coordinates c (i) for device
i are computed.
The iterations are stopped, when the error of the next coordinatecandidate is not a significant reduction to the previous candidate.
The significance is determined by a parameter . α controls the
step size of the search for good coordinates, i.e., a smaller α makes
coordination more precise but increases the number of necessary
iterations. In the experiment, one beacon is placed in each corner
of the plane. This corresponds to a beacon placement in a convex
hull around the network, which is the optimal beacon placement for
localization according to Bachrach and Taylor [12]. The coordinates
of the beacons are initialized with (0,0), (0,1), (1,0) and (1,1) and
the communication range is set to 15% of the environment’s sidelength. The devices are assumed to be asynchronous. One cycle is
defined as the random execution of 1000 devices. When selected
for execution, a device performs the following steps:
1. Ask all neighbors for known beacon coordinates
2. Ask all neighbors for the necessary information to estimate
distances (list of neighbors for GeoDE and hop counts for
GDE respectively)

229

Chapter 5 Localization in Mobile Ad Hoc Networks

Algorithm 5.3 Lateration
Require: beacon coordinates C (i), distance estimates D̄ (i)
Ensure: coordinates of device i: c (i) ← (x (i), y (i))
//Select closest beacon m: d¯(i, m) ← min D̄ (i)
//Initialize:
1: c̄ (i) ← c (m) ∈ C (i)
2: Δ (E) ← ∞
3: while Δ (E) >  do
4:
c (i) ← c̄ (i)
5:
Δx (i) ← 0
6:
Δy (i) ← 0
7:
for c (b) ∈ C (i) do
8:
d (i, b) ← EuclideanDistance (c (i) , c (b))

2
9:
E ← E + d (i, b) − d¯(i, b)


10:
11:



end for
//Calculate new coordinates:
13:
c̄ (i) ← (x (i)
 − α Δx (i) , y (i) − α Δy (i))


Ē ← C(i) EuclideanDistance (c̄ (i) , c (b)) − dˆ(i, b)
14:
12:

Δ (E) ← Ē − E
16: end while
17: return coordinates c (i)

15:

230



Δx (i) ← Δx (i) + (x (i) − x (b)) 1 − d (i, b) /d¯(i, b)



Δy (i) ← Δx (i) + (y (i) − y (b)) 1 − d (i, b) /d¯(i, b)

5.3 Geometric Distance Estimation for Mobile Ad Hoc Networks
3. Calculate distance estimates to all known beacons
4. If at least three beacon locations are known, calculate coordinates with lateration
5. Move according to the applied movement pattern
The examined mobility models are CM, RW, and StrM, which are
described in Section 5.1.1. RW is chosen because it is one of the
most widely researched mobility models in the literature, e.g., in
Zonoozi and Dassanayake [259]. CM is chosen as a contrast to RW,
since the trajectories of these two movements differ significantly in
range and step size. As a representative for coupled mobility, StrM
is selected because it is designed to model stream-like movements,
such as caused by natural forces like wind or water. Moreover,
movements of evacuees are often modeled as liquid or gas dispersal
(cf. Section 2.3.1), which is why the investigation of the StrM model
is especially interesting considering the context of this thesis.
For evaluation of the experiments, the average location error E
is calculated using Equation 5.24. With N denoting the set of
devices in the network, c̄ (i) are the estimated coordinates of device
i and c (i) are its real coordinates. The function d (·, ·) refers to
the Euclidean distance.


d (c̄ (i) , c (i))
(5.24)
|N |
Figure 5.36 shows the results for localization using GeoDE and
GDE in static and dynamic networks. The experiment reveals that
GDE delivers better results for static networks but is significantly
worse in dynamic environments, which confirms the findings from
Section 5.3.3. GDE is also based on hop counts and, as a consequence, suffers from the same weakness of underestimated hop
counts under mobility described thoroughly in Section 5.1.
E (c̄ (i) , c (i)) =

i∈N

231

Chapter 5 Localization in Mobile Ad Hoc Networks

Figure 5.36: Localization error with GeoDE and GDE for static
and dynamic networks.
The quality of GDE depends on the approximation of the numerical
integration, which influences the computational costs to derive
estimates. Since computational resources are limited in MANETs,
it is interesting to know how many integration steps (or boxes in
case of the Riemann integration method) are necessary to achieve a
better performance of GDE compared to GeoDE. Figure 5.37 shows
the results for varying approximation steps. Apparently, more
than 50 approximation steps are required for GDE to outperform
GeoDE.

5.3.4 Conclusion
GeoDE is a range-free distance estimation approach, which does
not rely on hop counts but the ratio of shared to total communication partners in order to compute distances. GeoDE offers
the possibility to estimate multi-hop distances in a network. It

232

5.3 Geometric Distance Estimation for Mobile Ad Hoc Networks

Figure 5.37: Localization error for the GDE for increasingly
fine-grained approximation of the necessary
integration computations.

233

Chapter 5 Localization in Mobile Ad Hoc Networks
is shown that using a third-degree polynomial regression in order
to approximate the mapping from neighbors to distances is less
error-prone when compared to an approximation via a first-order
Taylor series expansion. Furthermore, experiments are performed
in order to compare the distance estimation using GeoDE with two
hop count based distance estimation methods proposed in the literature. The first method is presented in 2003 by Nagpal et al. [171]
and the second choice is a more recently published approach from
Liu et al. in 2011 [140], which is called GDE. Three different static
network distributions are used for the comparison of GeoDE with
the first hop count based method: a uniformly random distribution,
a Gaussian random distribution, and a grid-like distribution of
devices. The communication range is varied from 5% to 20% and
the MAPE of distance estimates is used for evaluation. The more
recent hop count based approach is tested in a uniform-randomly
distributed static network and in dynamic networks with three
different mobility patterns from the study presented in Section
5.1.1. The investigated mobility models are CM and RW, which
belong to the group of individual movements, and StrM, which
is a coupled mobility model. For evaluation, the derived distance
estimates are used to compute locations with an iterative lateration
algorithm and the localization error, i.e. the deviation between
real and computed coordinates, is calculated for evaluation.
The experiments reveal that GeoDE significantly outperforms the
first hop count based distance estimation approach in all considered
static networks. Even the standard deviation of the error with
GeoDE is mostly lower than the error produced with the hop count
based distance estimation approach. Surprisingly, the performance
in a Gaussian-randomly distributed network is even better when
compared to a uniform-randomly distributed network for certain
communication ranges. This indicates that averaging the computed

234

5.3 Geometric Distance Estimation for Mobile Ad Hoc Networks
estimates from both involved devices is successful in mitigating the
negative effect of distorted neighborhoods. Additionally, the distance estimation approach profits from the higher average number
of neighbors in Gaussian-randomly distributed networks, which is
another reason why the error is well below that of uniformly random networks. In Gaussian networks, similar to uniform-randomly
or evenly distributed networks, the error decreases first with increasing communication range. At a communication range of about
10%, a low point is reached and the error starts increasing again.
This can be ascribed to the increased distortion of devices within
a neighborhood with higher communication ranges. In evenly distributed networks, GeoDE performs similar to uniform-randomly
distributed networks, except for some oscillations in the error behavior for increased communication ranges due to the step-wise
increase in the neighborhood size.
Although GeoDE is found to be slightly outperformed by GDE in
static networks, it is notably superior in all considered dynamic
networks. This is due to the negative effect of mobility identified in
Section 5.1, which all hop count based approaches suffer from. The
GDE algorithm relies on the computation of numerical integrations
which can be computed by different approximation methods. It is
shown that with an approximation method for numerical integration from Riemann [196], at least 50 integration steps are required
for GDE in order to outperform GeoDE.
A slight change in computing the average estimate of two neighboring devices could be useful in order to prevent an increase in error
for higher communication ranges in Gaussian-randomly distributed
networks. The higher the distortions of device locations in the
neighborhood are, the greater the difference between the total
number of neighbors of both involved devices. Since the device
with higher number of neighbors has better input data for distance

235

Chapter 5 Localization in Mobile Ad Hoc Networks
estimation, its resulting estimate is likely to be more precise than
the one from its counterpart. This difference could be regarded by
weighting both estimates according to the respective number of
total neighbors before computing the average distance estimate.

5.4 Optimization of Beacon Placement in
Buildings
As discussed before, beacon-based localization algorithms are the
only reasonable choice to localize mobile evacuation devices in
OBESS because absolute coordinates of devices with respect to
a common reference grid, e.g., the building map, are required.
Beacons, i.e., devices which know their own locations, for example
due to a-priori configuration, are used to derive the locations of
all other devices in the network. The placement of beacons is
essential to the accuracy of the derived locations (cf. Bachrach
and Taylor [12], Nagpal et al. [171]), hence, it is important to
think about where optimal locations for beacons are. So far,
this problem is considered mainly for SSNs. In static networks,
an optimal beacon placement is mostly handled as an optimal
coverage problem where beacons are placed in such a way that their
communication ranges cover the largest possible area. However,
new challenges arise when considering mobile devices because the
network’s topology is changing constantly and often substantially
over time, compared to SSNs where devices only occasionally fail
or are newly added. Considering an evacuation scenario makes the
problem even more specific. During evacuation, the devices move
simultaneously towards certain targets (exits) in the building. As a
result, not only the network’s topology is changing, but its intrinsic
structure changes from a more or less evenly distributed network to
a concentrated and fragmented network. Consequently, the ability

236

5.4 Optimization of Beacon Placement in Buildings
to communicate with certain static beacons varies strongly over
time. Intuitively, a concentration of beacons in higher frequented
areas of the building seems advisable. However, in such regions, the
density of the MANET is also increased, which makes it more likely
to establish connections to beacons over multiple hops, even if there
are few of them. This tradeoff has to be taken into account when
searching for an optimal solution. In addition, for lateration at least
three beacons have to be known to a device in order to compute
its position (cf. Nagpal et al. [171]). This further distinguishes the
problem from a simple optimal coverage problem. Little is known
about the characteristics of a good solution and with operating in
a two- or even three-dimensional environment, the search space is
large. These are criteria which point to a heuristic optimization
approach as a valid strategy for problem solving (cf. Gerdes et al.
[74]). In the following, an EA, i.e., a heuristic optimization and
search method based on the principles of natural evolution, is
introduced to tackle the problem. The EA is used to optimize the
placement of static beacons for localization of devices in a MANET
during evacuation. A multi-agent evacuation simulation serves as
a tool to evaluate the fitness of a specific placement.

5.4.1 Optimization of Network Distributions
As mentioned before, optimal device placement has received attention mainly in the context of SSNs. Also, the research often
focuses on achieving an optimal coverage of a specific area with
a minimal number of devices (cf. for example Cardei and Wu
[35], Heidari and Movaghar [87], Kaplan et al. [115], Katz and
Morgenstern [119], So and Ye [219]). In general, this problem is
referred to as minimum disc coverage problem and can be solved
in time O (n log n) with n denoting the number of devices in the

237

Chapter 5 Localization in Mobile Ad Hoc Networks
network (cf. Sun et al. [224]). Research which concerns the optimal
placement of beacons for localization in static networks can be
found in Akl et al. [3], Savvides et al. [207], and Tatham and Kunz
[230]. In the work of Savvides et al. [207], placing beacons at the
perimeter of an SSN is recommended. Akl et al. [3] and Tatham
and Kunz [230] propose guidelines for beacon placements in the
context of specific localization algorithms. Moreover, Bulusu et al.
[30] introduce an approach for adaptive beacon placement in order
to encounter failure of devices in the network by reorganization.
A lower bound for localization accuracy is shown by Salman et al.
[203] and the impact of beacon placement on this boundary is examined. Another research area is the improvement of localization
results by using mobile beacons, which is, for example, discussed
in Liao et al. [137] and Li et al. [134].
Evolutionary Algorithms
Before describing the proposed solution to the problem treated
here, some basic information about an EA has to be introduced.
EAs describe heuristic optimization algorithms which follow the
principles of natural evolution based on Darwin’s theory (cf. Darwin [47]). An EA is composed of genetic operators, which are
known as reproduction, mutation, and selection. The repeated
execution of these operators represents the search process for good
solutions to a given problem. Figure 5.38 illustrates the process.
At the beginning, a set of so-called individuals, which represent
valid solutions to a given problem, is chosen. This set of individuals is called a population. The initial population can be selected
randomly or by choice if certain prior knowledge about the solution
is available. The core of an EA is the fitness evaluation. Here,
a given solution, or individual, is evaluated with respect to the
optimization objective and a fitness value is assigned to it. The

238

5.4 Optimization of Beacon Placement in Buildings

Figure 5.38: Illustration of the general procedure of an EA.

239

Chapter 5 Localization in Mobile Ad Hoc Networks
choice of appropriate fitness evaluation criteria is crucial and not
always very intuitive. Especially in the research field of Evolutionary Robotics (cf. for example König et al. [124], Merkel et al.
[151], Nelson et al. [174]), which is concerned with the evolution
of controllers for robots, the performance evaluation of an evolved
controller is not straightforward. The scenario of beacon placement
optimization faces similar challenges because the criteria which
distinct a good beacon placement from a bad one are not obvious.
Therefore, some thought has to be put into the design of the fitness
evaluation.
After evaluating each individual, some of them are chosen to be
recombined. During recombination, their genome is merged to form
new individuals, which are called children. The children are then
slightly altered; this process is commonly described as mutation.
The iteration is concluded by forming a new population before the
procedure starts from the beginning. Various configurations of the
genetic operators are conceivable, which have a different impact
on the progression of the EA. The termination criteria for an EA
can either be a predefined number of repetitions (cycles) or the
achievement of a specified fitness level in the population. For more
details about EAs, it is referred to Weicker [246] and Gerdes et al.
[74].
In the following section, an algorithm proposed by Huang and
Tseng [100] is introduced, which can be used to optimize coverage
in SSNs. Later, this algorithm is, amongst other methods, used for
fitness evaluation in the EA to optimize beacon placements.
Perimeter Coverage Approach
Huang and Tseng [100] propose an approach to compute the socalled coverage-degree of an area which is occupied by static devices.
Basically, the algorithm decides if and how often the perimeter

240

5.4 Optimization of Beacon Placement in Buildings
of a device’s communication range is intersected by other devices’
communication ranges. If the perimeter of a device is completely
contained in the communication range of other devices, it is denoted
as covered. Figure 5.39(a) illustrates an example of a first-degree
covered beacon perimeter (the middle device).
In order to use this concept for fitness evaluation, some changes are
made as follows. An area which is not contained in the communication range of any beacon is denoted as uncovered or zero-degreecovered. An area which is contained in the communication range of
a beacon is denoted first-degree covered and an area which is inside
the communication range of a beacon, which, in turn, is covered n
times by other beacons is denoted as (n + 1)-degree covered. For
example, the area covered by the middle beacon in Figure 5.39(a)
is a two-degree covered area, while the region which is only covered
by the surrounding beacons is one-degree covered and the rest is
defined as uncovered. If two beacons cover the same slice of the
perimeter of a third beacon, the overlapping part is counted for
the next coverage degree. Figure 5.39(b) illustrates this procedure.

5.4.2 Evolutionary Algorithm for Optimal Beacon
Placement
To find a good placement of beacons for the localization of mobile
devices in a MANET, an EA is designed as follows. Firstly, the
genome representation of a solution is defined. Here, a solution is
represented by a set of two-dimensional coordinates, each denoting
the position of a respective beacon. The coordinates of one beacon
in a genome is called a gene. In Figure 5.40, an example individual
and its genetic representation are illustrated.

241

Chapter 5 Localization in Mobile Ad Hoc Networks

(a) The central beacon
2nd-degree-covers
an
area.

(b) Treatment of overlapping perimeter
covers.

Figure 5.39: Illustration of the process to determine the perimeter
coverage-degree of an area.
Selection
The selection process describes the procedure of selecting a certain number nparents of individuals for recombination. Here, the
binary tournament selection is used, which is a standard method
where two individuals are randomly chosen and then compared in
terms of their fitness value. The one with the higher fitness value
is selected for recombination. Both solutions stay available for
further selections. Figure 5.41 describes the tournament-selection
graphically.
By changing the size of a tournament, the tournament-selection
allows for an easy control of the selection pressure, i.e., the pressure
towards keeping good solutions in the population and disposing
bad ones. When selection pressure is too high, the algorithm
can get stuck in a local optimum because it does not explore

242

5.4 Optimization of Beacon Placement in Buildings

Figure 5.40: An example solution for the placement of beacons
inside a building and its corresponding genetic
representation.

243

Chapter 5 Localization in Mobile Ad Hoc Networks
solutions with low fitness well enough. On the other hand, when
the selection pressure is too low, the algorithm could be prevented
from converging to an optimal solution. A binary tournamentselection corresponds to a relatively low selection pressure (cf.
Weicker [246]).

Figure 5.41: Illustration of the tournament selection operator,
which is used to select individuals for recombination.

Crossover
All selected individuals belong to a mating pool. From this pool,
two random individuals are selected and recombined such that
they produce two new individuals. For this process two standard
methods are implemented and tested in the experiments. Firstly,
the uniform crossover method is used, in which it is decided
randomly for each gene in the genome whether it is part of the
first or the second child. The remaining empty genes are taken
from the second parent. Figure 5.42(a) illustrates the method.
The second standard method is called one-point crossover. Both
parents are cut in half at a random position in the genome. The first
child is composed by the first half of the first parent and the second
half of the second parent. The second child is created analogously.
Figure 5.42(b) shows an example of one-point crossover.

244

5.4 Optimization of Beacon Placement in Buildings

(a) Uniform crossover.

(b) One-point Crossover.

Figure 5.42: Illustration of the two recombination operators used
in the experimental study: uniform crossing and
one-point crossover.

245

Chapter 5 Localization in Mobile Ad Hoc Networks
Mutation
After recombination, nmut random genes from the newly created
individuals are slightly altered by mutation. For each gene, it is
decided with a probability of pmut whether the mutation is actually
performed. The mutated genes, i.e., the altered coordinates of a
beacon, are computed according to equation 5.25.
c = (N (x, σ) , N (y, σ))

(5.25)

with c denoting the new coordinates of the mutated beacon, x
and y represent the two-dimensional coordinates of the gene before
mutation and N (m, σ) is a normally distributed random value with
mean m and standard deviation σ. With this mutation method,
the new coordinates are selected within a certain range around the
old ones. The standard deviation of the normal distribution can
be used to adjust the amount of change caused by the mutation.
New Population
The new population is created using the standard (μ + λ)-approach,
with μ denoting the size of the old population and λ describing the
number of children. To build the next generation’s population, μ
individuals with the highest fitness are selected from the combined
set of old population and children (cf. Weicker [246]).
Fitness Evaluation
As mentioned before, selecting a good fitness evaluation method
is not trivial but has an important impact on the quality of the
derived solutions. Hence, various fitness evaluation criteria are
proposed in the following, which are then compared in experiments.
To evaluate the fitness of a given solution, a multi agent simulation

246

5.4 Optimization of Beacon Placement in Buildings
is used, in which the agents, i.e., evacuees who carry a mobile
device, compute their locations while performing an evacuation.
For this, lateration is chosen as a localization algorithm (cf. Section
2.4.2) and distance estimation based on hop counts according to
Nagpal et al. [171], as well as GeoDE (cf. Algorithm 5.3) are
applied. The localization technique based on hop counts is further
denoted as HC. The average deviation between real and estimated
positions of the mobile devices throughout the simulation period
T is computed. For this, the simulation duration T is divided into
time steps t, in which localization is performed and the agents
move towards the designated exit. The fitness is defined as shown
in Equation 5.26. It is the reciprocal term of the average deviation
between real and estimated coordinates in one simulation run, with
N being the set of all devices, ct (n) denoting the real position of
device n at time step t, and c̄t (n) its estimated position. In order
to navigate people to a safe exit during evacuation, the devices
have to know the right room which they are located in rather
than their exact locations. This consideration leads to the next
suggested fitness criteria shown in Equation 5.27. The percentage
of devices which estimate their positions to be in the correct room
of the considered building is computed on the basis of hop count
based distance estimation and lateration.
FP os (HC/GeoDE) =  

|T | · |N |
|ct (n) − c̄t (n) |

(5.26)

t∈T n∈N

FRoom =

1 
|{n ∈ N : room (ct (n)) = room (c̄t (n))}|
1−
|T | t∈T
|N |
(5.27)

Apart from the simulative approach, the perimeter coverage algo-

247

Chapter 5 Localization in Mobile Ad Hoc Networks
rithm introduced in Section 5.4.1 is used as fitness criteria. For
this, the environment is partitioned into a set of squares S. Then,
the coverage-degree of each square s ∈ S is computed and the
fitness value is derived by Equation 5.28, with i-cover (s) = 1 if
square s is i-th-degree covered. The reason for computing a maximum of third-degree coverage lies in the nature of the localization
algorithm. As stated before, lateration requires information from
at least three beacons. As a consequence, a third-degree covered
area seems to be most valuable and is, therefore, weighted more.
FP C =

3

i=1



i·

s∈S

i-cover (s)
|S|

(5.28)

5.4.3 Evaluation
To test the effectiveness of the presented algorithm and the various
fitness criteria, a simulative experiment is performed. For this,
the building evacuation described in Section 4.1.2 is repeated with
100 agents, which are distributed randomly across a building (cf.
Figure 5.43). The communication range is fixed at 10%. The
parameter values for the EA are listed in table 5.3.
To be able to assess how much two individuals differ from each
other, the Hausdorff-distance (cf. Rockafellar and Wets [198]) is
used as shown in Equation 5.29. The two individuals are denoted
as i and j, B (i) refers to the set of beacons from individual i. The
Hausdorff-distance reflects the maximum distance there is between
any two closest pairs of beacons from both individuals, thus, a
Hausdorff-distance of value zero indicates identical individuals.
Since hd (i, j) and hd (j, i) are not necessarily the same, the mean
value is computed as hdij = 12 (hd (i, j) + hd (j, i)), d (·, ·) refers to
the Euclidean distance.

248

5.4 Optimization of Beacon Placement in Buildings

Figure 5.43: Example evacuation scenario used for the
experimental study. The mobile devices are depicted
as gray circles with arrows pointing to the building’s
exit. The beacons are depicted as black circles.

249

Chapter 5 Localization in Mobile Ad Hoc Networks
Table 5.3: Parameter settings for the simulative experiments to
test the EA for beacon placement.
Parameter Name
Population size (μ)
Probability for mutation
Standard deviation for mutation
Mating pool size (nparents )
Number of agents
Evolutionary iterations


Value
10
0.3
0.05
10
100
1000


hd (i, j) = max∀bi ∈B(i) minbj ∈B(j) d (bi , bj )

(5.29)

The first experiment compares the progress of the average fitness
value in the population over the course of evolution for all four
fitness criteria. Figure 5.44 shows the results for both recombination
operators described in Section 5.4.2 and a mutation rate of 1 and
5 genes per iteration. The corresponding standard deviations are
depicted in Figure 5.45. It can be observed that the fitness increases
steadily over time for all considered evaluation criteria. This
indicates that the beacon placements are continuously optimized
with respect to the given criteria. The uniform crossover with
a mutation rate of 5 genomes yields the best results followed by
uniform crossover with a mutation rate of 1 gene per iteration, onepoint crossover with a mutation rate of 5 genes per iteration, and
one-point crossover with a mutation rate of 1 gene per interation.
Except for FP C where one-point crossover with a mutation rate
of 5 overtakes the uniform crossover with a mutation rate of 1
after about 200 iterations. Figure 5.45 shows that the standard
deviation of fitness values become relatively stable towards the end

250

5.4 Optimization of Beacon Placement in Buildings
of the experiments for almost all considered settings, except for
FP os (GeoDE) with uniform crossover and a mutation rate of 5,
FP os (HC) with uniform crossover and a mutation rate of 1, and
FRoom for all considered settings except uniform crossover and a
mutation rate of 1. It should be noted that only the fitness value
levels FP os (GeoDE) and FP os (HC) are directly comparable.
The most unexpected observation one can make from these results
is the exceeding performance of FP os (HC) when compared to
FP os (GeoDE). As shown in Section 5.3.3, the approach based on
hop counts usually delivers localization results with lower quality
than GeoDE. Hence, the superior fitness of solutions which are
produced by the hop count based fitness criterion is a surprising
discovery. One possible reason for this could be that finding an
optimal beacon placement is more difficult for GeoDE compared
to the approach based on hop counts. This seems reasonable when
considering that GeoDE depends much more on the distribution of
a device’s neighbors compared to the approach based on hop counts.
Figure 5.46 displays the individuals from the final population with
the highest fitness values FP os (HC) and FP os (GeoDE), which are
simultaneously the individuals with lowest localization error. It
becomes obvious that a good beacon placement for localization
based on GeoDE looks differently from a good beacon placement
for localization based on hop counts. The corresponding average
Hausdorff-distance is hdHC,GeoDE = 0.25. From these observations,
it can be concluded that the applied localization method has a
strong impact on the result of the EA. Consequently, it can
be assumed that different localization algorithms have different
requirements on the beacon placement in a network.
As mentioned before, the results in terms of fitness cannot be
compared directly to each other. To overcome this issue, the final
beacon placements from all four experiments are evaluated in one

251

Chapter 5 Localization in Mobile Ad Hoc Networks

(a)

(b)

(c)

(d)

Figure 5.44: Progress of the fitness value during the evolution for
perimeter coverage, GeoDE, HC, and the room
mapping fitness criteria.

252

5.4 Optimization of Beacon Placement in Buildings

(a)

(b)

(c)

(d)

Figure 5.45: Standard deviation of the fitness values during the
evolution for perimeter coverage, GeoDE, HC, and
the room mapping fitness criteria.

253

Chapter 5 Localization in Mobile Ad Hoc Networks

(a)

(b)

Figure 5.46: Solutions with highest fitness value evolved using the
simulative GeoDE based localization (a) and
localization based on hop counts (b) fitness criteria.
run of the multi-agent evacuation simulation, computing the average localization error as FP os1(HC) . Figure 5.47 shows these results.
The beacon placements evolved with the simulative localization
approach based on hop counts are best in terms of localization
error followed by the beacon placements evolved with the Perimeter
Coverage fitness criteria, and the results produced by GeoDE-based
fitness evaluation. The fact that a beacon placement evolved with
a hop count based fitness criterion yields the lowest localization
error when hop count based localization is applied for evaluation is
not surprising. Also, it is understandable that a beacon placement
which was optimized for room mapping performs worst in terms
of average localization error. However, it is unexpected that the
Perimeter Coverage fitness criterion delivers similarly low localization error when compared to hop count based fitness criterion.
This does not necessarily mean that an optimal beacon placement

254

5.4 Optimization of Beacon Placement in Buildings
for localization is the same as a beacon placement with optimal
coverage. The fact that a third-degree perimeter coverage was
weighted more heavily than a first-degree coverage could play an
important role in the high performance of this approach. This is
confirmed by comparing the two best individuals evolved with the
perimeter coverage approach in terms of fitness and localization
error shown in Figure 5.48. While the fitter individual has a wider
area covered by beacons, the individual with lower localization error has a denser beacon placement leaving more squares uncovered.
From this it can be concluded that the emphasis on a third-degree
covered area in the fitness function is likely to be the reason for
the good performance of the perimeter coverage evolution. Nevertheless, it should be noted that the results are in fact similar, while
the Perimeter Coverage approach is much less computationally
complex since it does not require an evacuation simulation.
Another important discovery is that the room mapping objective
obviously delivers different results compared to the criteria which
consider localization errors. It becomes apparent that a low localization error is not necessarily the same objective than a good
room mapping of coordinates and it has to be thought about which
goal priority has before starting the optimization. When looking
at the best individual in terms of localization error evolved using
FP os (HC) in Figure 5.46(b), it becomes obvious that placing beacons along a path to the exit seems advisable. When comparing
this beacon placement with the best one evolved with perimeter
coverage in terms of fitness, which is displayed in Figure 5.48 (b),
they look rather different. In fact, their average Hausdorff distance
hdP C,HC is 0.25. However, the average localization error for both
individuals is very similar with 0.11 for perimeter coverage and
0.10 for the hop count based approach. This indicates that an even
better placement could be found when beacons are located close

255

Chapter 5 Localization in Mobile Ad Hoc Networks
to the path leading towards an exit and, at the same time, provide
good third-degree coverage.

Figure 5.47: Comparison of the last populations evolved with
various fitness criteria in terms of localization error.
The importance of a dense beacon placement is reinforced when
looking at the best individual in terms of fitness evolved with the
room mapping fitness criteria illustrated in Figure 5.49(b). Obviously, the computed coordinates map well to their corresponding
room, when beacons are placed densely, even though the localization error is comparatively high in this case (cf. Figure 5.47).

5.4.4 Conclusion
In this section, an EA is introduced with the objective to optimize
beacon placements for localization of mobile devices in an ad hoc
network during building evacuation. Since mobility of the devices

256

5.4 Optimization of Beacon Placement in Buildings

(a)

(b)

Figure 5.48: Solutions with highest fitness value (a) and lowest
localization error (b) evolved using the perimeter
coverage fitness criteria.

(a)

(b)

Figure 5.49: Solutions with lowest localization error evolved using
simulative localization based on hop counts (a) and
correct room mapping (b) fitness criteria.

257

Chapter 5 Localization in Mobile Ad Hoc Networks
and the evacuation scenario affect the network topology strongly,
it is argued that simply increasing the area which is covered by
beacons is not the optimal strategy to find good beacon placements.
To evaluate the EA, experiments are performed with uniform and
one-point crossover as recombination methods and a mutation rate
of 1 or 5 genomes per iteration of the algorithm. Four different
fitness evaluation criteria are proposed, three of them evaluate a
beacon placement by simulating an evacuation scenario in which the
devices perform localization. Two of these simulative fitness evaluations measure the average localization error during the simulation
by applying lateration to distance estimates produced by a simple
hop count based approach and by the GeoDE method proposed in
Section 5.3. The third simulative fitness evaluation measures how
often locations derived by lateration and hop count based distance
estimation lie within the correct room of the building since this
is especially important for computing evacuation instructions. In
order to make the results comparable, the final populations are
evaluated in terms of localization error in one simulation run using
lateration and hop count based distance estimation. Additionally,
the beacon placements with highest fitness and lowest localization
error are compared and their difference is evaluated by applying
the Hausdorff distance metric introduced by Rockafellar and Wets
[198].
Surprisingly, the distance estimation approach based on hop counts
performed best in terms of fitness and localization error closely
followed by the Perimeter Coverage method. Subsequently there
are GeoDE and room mapping evolution results. Unexpectedly,
GeoDE is outperformed in terms of localization error by hop count
based localization, even though it is found to be superior in previously investigated network scenarios (cf. Section 5.3.3). A likely
reason for this is that it is more difficult to optimize beacon place-

258

5.5 Summary
ments for GeoDE based localization compared to hop count based
localization. In any case, the selected localization error clearly
affects the optimal beacon placement in a network. It is further
shown that minimizing the localization error leads to different
results when compared to the objective of finding a good room
mapping of calculated coordinates. In summary, beacon placements alongside a path which leads towards an exit, as well as a
high third-degree coverage, are identified to be essential criteria
for low localization error.
It is favorable to have a beacon placement which serves as a basis
for a constantly low localization error in contrast to a low average
error which is highly volatile. Hence, a further improvement of the
presented approach could be to integrate the standard deviation
of localization errors in the fitness evaluation.

5.5 Summary
In this chapter, range-free distance estimation, which can be used
as a basis to localize devices in ad hoc networks via lateration,
is subject to investigation. The focus lies on the influence which
mobility of the involved devices has on the accuracy of distance
estimation, whether adjustments have to be made in order to
account for this impact, and which adjustments could be reasonable. Firstly, a study is presented which examines hop count based
distance estimation in MANETs. A so-called hop count error measurement is introduced. It measures the deviation between actual
hop counts in a network and an ideal hop count distribution, which
would be obtained if the network was perfectly dense and the
devices were stationary. This allows deriving general statements
about the influence of mobility on any hop count based distance
estimation method. Various mobility models from the literature

259

Chapter 5 Localization in Mobile Ad Hoc Networks
and some novel models, which have been specifically designed for
this study, are described and their impact on hop count errors
is tested. Experiments reveal that mobility turns naturally positive hop count errors into unpredictably high negative hop count
errors, which leads to underestimation of distances. This effect
is explained by devices which move and then communicate their
hop count values before updating them according to their new
locations. This can affect devices at their new locations, which in
turn underestimate their distances as well. An increased speed,
directions leading away from beacons, and increased heterogeneity
in the movements of nearby devices are found to be reinforcing
factors for underestimation. MoGA is presented, which is a variant
of the standard GA used to determine hop counts, in order to
reduce negative hop count errors. Experiments show that this is
successfully done; however, a complete elimination is not achieved.
As a next step, two indicators are presented called HC-change and
ID-change, which are able to identify and characterize mobility
of devices in a network. Additionally, they are computable in a
decentralized manner solely based on local information. The aim is
to use these indicators in context of learning mechanisms provided
by the O/C Architecture in order to reduce hop count errors by
adapting the hop count values according to the current network
dynamics.
Asynchronous communication and updates of hop counts are identified to be the main reasons for the emergence of negative hop
count errors in dynamic networks. Hence, it is subsequently proposed to synchronize the devices in the network by applying a
nature-inspired synchronization algorithm. Before investigating
the benefit of synchronization for the determination of hop counts
in dynamic networks, an assessment of the proposed synchronization algorithm is performed. In order to evaluate its effectiveness

260

5.5 Summary
in MANETs, it is tested under different parameter constellations.
Synchronization duration and synchronization success are measured. It is demonstrated that reliable synchronization success can
be achieved for both, static and dynamic networks if the algorithm
is initialized with a suitable parameter configuration. Furthermore,
increasing the communication ranges of the devices is proven to
increase the synchronization success while simultaneously reducing the time duration. An algorithm called FIHC is presented,
which can be used to encode hop count information in the time
difference between two signals in a synchronized network. Due
to the synchrony in the network, now a device can wait until it
receives a signal from its neighbors, which confirms its own hop
count value, before it reports this value to other nearby devices.
Waiting for confirmation before hop count communication aims
at reducing negative hop count errors. Experiments confirm that
this simple mechanism indeed eliminates negative hop count errors
almost entirely.
Apart from hop count based distance estimation, GeoDE is presented in this chapter. This method derives distance estimates
between neighboring devices from the ratio of shared to total communication partners. An algorithm is presented which extends this
method in order to estimate distances over multiple hops between
devices in the network and beacons. The mapping from communication partners to distances is approximated with a third-degree
polynomial function, which is shown to be less error-prone when
compared to an approximation with a first-order Taylor series
expansion. Additionally, a short discussion about potential sources
of error in GeoDE is given. The proposed algorithm is evaluated in
an experimental study and compared to the results achieved by two
different hop count based distance estimation approaches. GeoDE
significantly outperforms the more established hop count based

261

Chapter 5 Localization in Mobile Ad Hoc Networks
distance estimation technique for all considered communication
ranges. This result is consistent for all three regarded stationary network distributions, namely a uniformly random network, a
Gaussian random network and a network with grid-like distributed
devices. Even the standard deviation of the MAPE with GeoDE
lies below the error of the hop count based approach. However,
when compared with a more sophisticated recently published hop
count based approach called GDE, GeoDE produces estimation
results of similar, marginally worse, quality in static networks.
Nevertheless, in four different dynamic networks, the superiority of
GeoDE over GDE is shown to be significant. The main reason for
this can be assumed to lie in the previously mentioned sensitivity to
negative hop count error under mobile conditions, which hop count
based distance estimation methods suffer from. GDE is based on
numerical integration and an additional experiment shows that it
requires at least 50 steps of the Riemann integration approximation
method in order to deliver more accurate distance estimates than
GeoDE in static networks.
Because the placement of beacons affects the quality of localization
results, an EA is presented to optimize such placements, specifically
for localization during an evacuation scenario. Various fitness criteria are proposed in order to evaluate a certain beacon placement.
Two fitness criteria are based on the average localization error
produced with lateration during a simulated evacuation scenario.
One determines locations on the basis of hop counts, the other uses
GeoDE. A third fitness evaluation method measures the average
percentage of devices which locate themselves in the correct room
using lateration and hop count based distance estimation. The last
proposed fitness evaluation is based on a method proposed in the
literature in order to determine the perimeter coverage-degree of
a beacon placement. According to this method, the perimeter of

262

5.5 Summary
a beacon is denoted as one-degree covered when it’s perimeter is
completely covered by other beacons’ perimeters. The fitness value
is computed in a way that third-degree covered beacons are valued
the most since at least three beacons are required to compute
two-dimensional locations using lateration. An experimental study
is conducted to evaluate the proposed EA. It is shown that beacon
placements are indeed optimized over time with respect to the
individual fitness functions. Two recombination methods, namely
uniform and one-point crossing, and a mutation rate of 1 and 5
genes per genome are tested. Uniform crossing with a mutation
rate of 5 delivers best results, independently of the underlying
fitness evaluation method. While lowest localization errors are
produced by beacon placements evolved according to hop count
based fitness, perimeter coverage fitness evaluation follows closely
behind. Subsequently, there are beacon placements produced by
GeoDE based evolution and placements evolved based on the room
mapping fitness criterion. Comparison of some evolved beacon
placements from final populations reveal that an optimal placement
for hop count based distance estimation is rather different from one
evolved to be used with GeoDE. Moreover, minimizing localization
error yields different beacon placements when compared to optimizing the room mapping of derived locations. Furthermore, the
experiments indicated that finding an optimal beacon placement is
more challenging for GeoDE than hop count based methods. In
general, a beacon placement with a high third-degree coverage and
where beacons are located alongside paths towards the building’s
exit seem to be desirable.
As a summary, it can be concluded from the research presented in
this chapter that localization algorithms have huge performance
deviations depending on the characteristics of the network, the
mobility of the devices in the network, and the placement of

263

Chapter 5 Localization in Mobile Ad Hoc Networks
beacons. This confirms that learning mechanisms should be applied
in order to improve localization in OBESS. Firstly, to the mobile
devices such that they learn how to select appropriate localization
algorithms at runtime depending on the current environmental state
and, secondly, to the CCU in order to optimize beacon placements
for localization in that specific building.

264

CHAPTER

6
CONCLUSION

Increasingly large and complex buildings are the main motivation
for a long overdue overhaul of today’s building evacuation support.
In order to achieve this goal, OBESS has been proposed in this
thesis, a system consisting of partly mobile devices used to navigate potential evacuees during an emergency evacuation. This
chapter concludes the presented work by summarizing the major
contributions in Section 6.1 and describing aspects which remain
open for future research in Section 6.2. Section 6.3 gives some final
remarks on the work presented.

6.1 Summary
This thesis has made three major contributions to advance the
evacuation support equipment in buildings. Firstly, main problems
of current emergency equipment have been identified and a concept

265

Chapter 6 Conclusion
for an evacuation support system has been presented in order to
overcome these limitations. Section 6.1.1 summarizes these findings.
Secondly, the concept of decentralized and distributed evacuation
route planning with mobile devices has been investigated. The
developments regarding this topic are described in Section 6.1.2.
Thirdly, localization of mobile devices in ad hoc networks has been
addressed. Section 6.1.3 concludes this summary by presenting the
research performed in this domain.

6.1.1 Organic Building Evacuation Support System
The preparations of today’s buildings for an emergency evacuation
mainly consist of stationary, analogous emergency evacuation maps
and exit signs installed for the purpose of guiding evacuees towards
safe areas or exits. This route guidance system is designed by
experts at the time of construction. The objective is to provide
for a well-ordered evacuation in case of an emergency situation as
it is most likely to occur in that specific building. This includes
the estimation of the average number of people located inside that
building and their most probable distribution across the rooms.
However, this evaluation can be far from reality of an actual
evacuation scenario. Furthermore, potential blockages of passages
due to collapsed masonry or congestions of evacuees in front of
narrow passages or doors can change the optimal route completely.
Another problem with today’s evacuation support is its stationary
nature. Exit signs and evacuation maps are likely to be overlooked,
especially when in panic. Consequently, an ideal evacuation support
system is desired to react in an adaptive manner to changes in
its environment while also being portable. In addition, an ideal
evacuation support system should be designed in a way such that
the optimal escape route can be personalized according to the

266

6.1 Summary
specific needs of a particular evacuee. Such individual needs can
result, for example from high age or a physical disability of the
evacuee, which could make it necessary to avoid narrow passages or
stairs. In order to achieve the aforementioned objectives, this thesis
has proposed to use mobile devices, such as smart phones or tablet
PCs, for evacuation support in buildings. Mobile devices possess
computing capacities and digital screens which enable adaptive
route guidance. Furthermore, they are usually equipped with
means for wireless communication, which allow for the collection of
information about the current evacuation situation inside a building.
This knowledge can then be incorporated in the escape route
planning process. Furthermore, mobile devices are portable and
belong to specific users, thereby, fulfilling all previously mentioned
desired characteristics for an emergency navigation device.
In this thesis, OBESS has been proposed as an evacuation support system that consists of three main components, a CCU, an
SSN, and mobile devices which are capable of establishing ad hoc
network connections via local communication. The CCU is used
for the configuration of the static sensors which are distributed
across the building. The CCU communicates the layout of the
building to the sensors, as well as all other information required
to support the evacuation route planning and localization process
of the mobile devices. This can, for example, be a graph model
of the building’s layout or information about the sensors’ own
locations inside the building in order to serve as beacons for the
localization of mobile devices. Moreover, the CCU is intended to
collect information about the performance of OBESS which can
then be used for building-specific optimization. An example of
such an optimization is given in Section 5.4 of this thesis, where
the placement of sensors in the building is optimized in order to
improve localization accuracy. Sensors are able to communicate

267

Chapter 6 Conclusion
with each other and with mobile devices whenever they are within
reach as well as to provide necessary input for evacuation route
planning and localization. Evacuation route planning is performed
directly on the mobile devices, which use their ad hoc network
connections in order to exchange information about the current
evacuation situation and use it to improve evacuation route planning. The computed evacuation paths are then displayed on the
devices’ screens or provided as voice instructions to the user.
Although there is a CCU in this system architecture, OBESS is
designed as a decentralized system in order to avoid having a single
point of failure. Therefore, in case the CCU fails to work, the evacuation system still continues to function. In addition, occasional
sensors can be defective without rendering the entire evacuation
support system useless. It is even conceivable to install some navigation devices inside the building in order to provide guidance
for evacuees without or with broken mobile devices. Consequently,
due to this design, OBESS is quite robust against failures. This
is of special importance as building evacuation is usually a lifethreatening situation where a support system is expected to work
robustly.
Another crucial virtue of OBESS is the concept of self-organization
between the mobile devices in order to employ adaptive evacuation route planning and localization, such that the ability to react
to unforeseen changes in the environment is provided for. Selforganization, however, can lead to undesired emergent system behavior, which can be avoided by using the generic O/C Architecture
from Organic Computing. This architecture provides means for
controlled self-organization and, hence, allows a system to be flexible as well as robust and trustworthy at the same time. This
thesis has shown how the generic O/C Architecture can be applied
to mobile evacuation devices in order to allow for controllable

268

6.1 Summary
self-organization in evacuation route planning, as well as in the localization procedure. For evacuation planning, the online learning
mechanism can be used to decide if an evacuation instruction has
to be updated due to changes in the environment. Additionally,
offline-learning allows for validating an evacuation instruction in
a simulative environment before it is made available to the evacuee. This quality management is meant to improve evacuation
instructions and, thus, the resulting evacuation process. Moreover,
it has been described how the two-level learning mechanism of
the O/C Architecture can be used for the selection of an appropriate localization algorithm from a set of different algorithms
available to the device. The choice is made by taking into account
the current topology of the MANET or other criteria, such as
the state-of-charge of the mobile device’s battery. The intrinsic
control mechanism of the O/C Architecture uses feedback mechanisms such as consistency checks and dead-reckoning methods for
evaluating the performance of the selected algorithm and decides
about an appropriate control action, which can be a change of
the applied localization algorithm. Furthermore, a simulation of
the current network topology can be set-up and algorithms can
be tested in this simulative environment before they are applied
during runtime. In addition to these intrinsic control mechanisms,
possibilities for direct user control have been presented, such as
correcting computed locations by tapping on the device’s screen
or triggering an update of the evacuation instruction by simply
walking in another direction than the one suggested by the current
navigation instruction.

269

Chapter 6 Conclusion

6.1.2 Swarm Evacuation Planning
After having introduced the architecture for a self-organizing and
robust evacuation support system, this thesis has addressed the
task of decentralized evacuation route planning with mobile devices.
In contrast to standard path planning approaches, which use global
knowledge of the evacuation situation in order to optimize escape
routes, the main challenge for OBESS is the uncertainty of the
information basis available to each mobile device for finding an
optimal evacuation route for its user. Due to the decentralized
system architecture, the mobile devices can only rely on local
communication with devices within their proximity and information
dissemination over ad hoc network connections for generating and
updating their knowledge base. SEP refers to such decentralized
evacuation planning methods which use only local information as
a basis. Two SEP algorithms have been developed and evaluated
in this thesis. Both algorithms are adaptive to changes in this
knowledge base and are shown to accelerate the evacuation process
when compared to a situation in which evacuees’ choose the shortest
path towards an exit of the building.
The first method is called CC-SEP. It uses a macroscopic graph
model of the building and a flow optimization approach to find optimal evacuation routes for each evacuee. This is done by scheduling
all evacuees such that the overall evacuation time is minimized
and by subsequently selecting an appropriate path for the device’s
specific user. This approach has been subject to thorough investigation. Varying numbers of evacuees, distributions inside the
building, communication ranges of the mobile devices, and several configuration parameters of CC-SEP have been investigated.
Moreover, different building layouts have been used for evaluation
in order to confirm the generality of the results. Additionally, a robustness test has been performed, in which the number of evacuees

270

6.1 Summary
who follow the provided navigation instructions have been varied.
In summary, it has been shown that CC-SEP leads to a lower
overall evacuation time in almost all considered scenarios when
compared to a situation in which all evacuees follow the shortest
path. Furthermore, the approach has been proven to be robust in
scenarios with up to 30% of evacuees which deviate from suggested
evacuation routes. Although the evacuation time has been shown
to increase slightly when the evacuation planning is performed
repeatedly in order to adapt to changes in the knowledge base, it
is still faster than in the scenario where evacuees simply follow the
shortest paths.
DMO-SEP is the second SEP algorithm, which has been proposed
in this thesis. This algorithm uses a discretized version of the building layout instead of a macroscopic graph model as a basis for route
planning, a complex cost function to evaluate different routes, and
a dynamic path finding algorithm. This dynamic algorithm adapts
evacuation paths to newly available information, which reduces the
necessary computations when compared to methods which compute
the optimal path from scratch. Moreover, DMO-SEP is suited to
take into account user specific objectives in the evacuation route
optimization process, such as an individual level of risk aversion
or others. For the evaluation of different evacuation paths with
regard to their potential for congestions, two congestion indicators
have been proposed. The first one is called load and it is based
on the number of evacuees with respect to the size of the room
in which they are located in. The second indicator focuses on the
entropy of evacuees’ locations in rooms, i.e., their concentration.
An experimental study has been performed in which DMO-SEP
is compared to CC-SEP and the different optimization objectives
are investigated. It has been shown that DMO-SEP can further
improve the overall evacuation time when compared to CC-SEP.

271

Chapter 6 Conclusion
Additionally, DMO-SEP is demonstrated to be able to consider
different objectives when optimizing the evacuation path, such
as risk aversion or congestion avoidance. Moreover, both congestion indicators have been compared and it is revealed that load is
an indicator better suited to improve the overall evacuation time
than entropy, although both indicators lead to an overall faster
evacuation when compared to minimizing travel distance only.

6.1.3 Range-free Distance Estimation in Mobile Ad hoc
Networks
Distance-based localization has been identified to be the first
choice for localization of mobile devices in OBESS since such
algorithms are decentralized, deliver absolute localization results,
do not necessarily require bulky hardware, and are more accurate
than proximity-based localization approaches. In order to apply
distance-based localization, distance estimates between the devices
to be located and beacons, i.e., devices which know their own
locations, have to be determined. For this purpose, many distance
estimation techniques have been proposed in the literature and
described in this thesis. Range-free distance estimation is especially suited for an application in OBESS because it exploits ad hoc
network connections instead of relying on the analysis of a physical
communication signal. This way, the hardware requirements of
the mobile devices, as well as the number of required beacons in
the system are reduced, hence, reducing the total installation costs
of such an evacuation support system. However, most range-free
distance estimation concepts have not yet been subject to research
when applied to mobile devices instead of static ones. Since the
devices in OBESS are mobile, the application of such algorithms
to MANETs has been addressed in this thesis.

272

6.1 Summary
In the first instance, an extensive study has been performed to
investigate the effect various mobility models have on distance
estimation based on hop counts, i.e., the minimum number of
relay devices between the device to be located and the beacon.
This study has revealed that mobility has a negative impact on
the accuracy of hop count based distance estimation due to asynchronous computation and communication of hop counts. While
hop count based distance overestimation is a common problem
in static networks, to which many solutions have already been
proposed in the literature, mobility of devices can turn this overestimation into an unpredictably high underestimation. It has been
argued thoroughly in this thesis that it is crucial to avoid such
an underestimation in order to improve the accuracy of localization in mobile networks. In this study, speed and direction of a
device’s movement with respect to beacons have been shown to
have a significant impact on the underestimation. Additionally,
similarity in the movements of devices which are close together has
been identified to have an impact on this kind of underestimation.
MoGA, a modified version of a standard algorithm for determining
hop counts has been proposed in this thesis and has been shown
to reduce mobility induced underestimation. Furthermore, two
indicators have been introduced, which can be computed based
on local information from the hop counting algorithm. These indicators have been shown to be suited for identifying whether a
device moves and which characteristics its movement has. In the
context of O/C Architecture, these indicators could be used for
letting the devices learn how to adjust hop count based distance
estimation results in order to compensate for mobility induced
underestimation.
Since asynchronous hop count computation and communication
have been identified to be the main reasons for the negative impact

273

Chapter 6 Conclusion
of mobility, this thesis has examined whether a MANET can be
synchronized in order to reduce mobility induced underestimation.
A nature-inspired, decentralized synchronization algorithm has
been investigated in static and dynamic network scenarios with
varying communication ranges and configuration parameters. It
has been shown that synchronization can be achieved reliably in
both, static and dynamic networks, when initialized with suitable
parameter values. Based on synchronized networks, FIHC, an
algorithm for scheduling the communication of hop count values,
has been introduced. Two variants of this hop counting algorithm
have been investigated and compared to the standard hop count
algorithm. Experiments have revealed that this scheduling of hop
count messages in a network can significantly reduce mobility
induced underestimation, up to an almost complete elimination of
underestimation in the scenario considered.
Apart from hop count based distance estimation, a connectivitybased distance estimation approach called GeoDE has been developed in this thesis. This approach utilizes the ratio of shared to
total communication partners between two devices for estimating
the distance between them. It has been shown how this algorithm
can be used for the purpose of estimating distances between beacons and devices in the network. As for the mapping of ratios
to distance estimates, two approximation approaches have been
compared, and it has been shown that the linear regression yields
a lower approximation error when compared to a first order Taylor
series expansion. GeoDE has been evaluated for varying communication ranges of the mobile devices and for different static network
topologies. Additionally, dynamic networks using different mobility
models have been investigated. It has been demonstrated that
GeoDE outperforms two different state-of-the-art hop count based
distance estimation techniques in static and dynamic networks for

274

6.1 Summary
almost all cases considered. Only when compared to GDE from
Liu et al. [140] in a uniformly random distributed static network,
GeoDE does not significantly outperform the hop count based
estimation method, but rather delivers quite similar localization
errors. However, an additional experiment has shown that this is
only the case when numerical integration in GDE is approximated
with a sufficiently high accuracy. In any other case, GeoDE is
superior to this estimation method as well.
The final contribution of this thesis to the research areas of evacuation management and localization alike is an EA which optimizes
the placement of beacons for the application of mobile device localization during building evacuation. Building evacuation leads to a
change in the distribution of mobile devices in the building over
time, with devices temporarily concentrating at bottleneck-areas,
such as doors or narrow passages, and tending to cluster around
the exits of the building after a certain time. This dynamically
changing topology poses a special challenge for finding an optimal
beacon placement to support localization during the whole evacuation process. Since only little is known about the characteristics of
a potentially good beacon placement for this specific application
scenario and the search space is quite large, a heuristic optimization
approach has been chosen to tackle the problem. Four different
evaluation criteria for the fitness evaluation process have been proposed in this thesis. Three of these evaluation criteria are based on
a simulative building evacuation, during which the mobile devices
determine their locations via distance-based localization. The first
variant uses hop count based distance estimation for localization,
the second variant is based on GeoDE. The average localization
error of all devices during the evacuation has been taken as an evaluation criterion for determining the fitness, i.e., quality, of a specific
beacon distribution. The third simulative evaluation method also

275

Chapter 6 Conclusion
uses hop count based distance estimation for the localization of the
devices, but a room mapping criterion, instead of the average localization error, forms the basis for fitness evaluation. This criterion
measures how often the determined locations match the correct
room in which the devices are actually located in. As a fourth
fitness evaluation criterion a so-called perimeter coverage approach
has been examined. A beacon distribution has high perimeter coverage when a large area is covered by the communication ranges of
the beacons and the perimeters of their communication ranges have
many overlaps. The experimental evaluation has delivered some
valuable insights into an optimal beacon placement for localization
during a building evacuation scenario. Beacon placements with a
high perimeter coverage placed alongside a path towards the exit
of the building have been found suitable to support localization in
the considered scenario. Furthermore, it has been shown that a
beacon placement which results in a low average localization error
is not necessarily optimal when it comes to determining the correct
room in which devices are located in and vice versa. Hence, it
is important to determine the appropriate optimization objective
before choosing a beacon placement. In addition, the study has
revealed that GeoDE, although previously proven to be superior
to hop count based distance estimation in mobile networks, can
under certain conditions result in a higher average localization
error when applied to an evacuation scenario. This finding indicates that alternating between localization algorithms during
runtime in response to changing conditions of the environment can
improve localization results. This finding confirms the strategy for
an adaptive evacuation and localization approach proposed in this
thesis.

276

6.2 Outlook

6.2 Outlook
With the introduction of the concept of OBESS and the proposal
for decentralized path planning and localization, this thesis has
laid the foundation for a long overdue changeover in building evacuation support. By applying the O/C Architecture from Organic
Computing to the mobile devices used in OBESS, the devices are
capable of autonomously computing their locations in the building and of finding optimal evacuation instructions while at the
same time maintaining their controllability by the user. With the
introduction of this concept, the development of decentralized evacuation planning methods, and the new insights about range-free
localization in dynamic networks, the research objectives of this
thesis have been fulfilled. However, the main ambition of this
thesis is to initiate a process, in which researchers improve and
reconsider the presented ideas in order to further develop the vision
of building evacuation support via mobile devices. In the following,
some initial thoughts on potential improvements are presented.

6.2.1 Organic Building Evacuation Support System
While the focus of this thesis has been laid on the mobile evacuation devices used in OBESS, the design of the CCU is another
important factor. As already mentioned, the CCU could also be
implemented according to the generic O/C Architecture, allowing
for the learning of building-specific characteristics in order to improve the performance of OBESS over time. Information about the
number of people inside the building at specific days or about the
most frequented rooms, et cetera, can be collected from the mobile
devices using the SSN connections, transferred to the CCU and
incorporated in the building-specific optimization process. Using
this data, potential bottleneck-passages in the building could be

277

Chapter 6 Conclusion
identified and removed. In Section 2.3.2, several publications are
presented which address the optimization of buildings to enable
faster or less dangerous evacuation. Such optimization includes the
deliberate placement of barriers inside the building or additional
doors in order to prevent congestions. This optimization could be
performed on the CCU in OBESS, with the great benefit of being
able to use building-specific characteristics as a basis for evaluating
measures before they are applied to the building. Since adaptation
of buildings via constructional measures can be cost intensive, this
is a major advantage of OBESS. Additionally, the CCU can use
collected data in order to improve the localization of mobile devices
in OBESS. One example of such an optimization approach has
been presented in Section 5.4 of this thesis, where an EA is used to
find appropriate placements of beacons in a building with the aim
of supporting localization during an evacuation scenario. This idea
could be expanded even further by collecting data from multiple
CCUs in OBESS-ready buildings on external servers in order to
provide an even broader learning basis.
Apart from developing building-specific optimization concepts, the
general design of OBESS needs to be considered. For example,
the user interfaces provided by the CCU or the mobile devices
have to be defined. Questions have to be addressed, such as
which data should be displayed to the user and in which form
should the layout of the building be provided to OBESS, et cetera.
When talking about the building layout, it is essential to consider
certain standards for layouts, as well as a way to make them
automatically readable by OBESS. Moreover, further research is
required which focuses on the automatic transformation of building
layouts to macroscopic evacuation graphs, which can be used as
a basis for CC-SEP. Additionally, it would be useful to enhance
the layout by providing information about the location of first-

278

6.2 Outlook
aid kits, fire extinguishers, or others. It is also conceivable to
add information about other points-of-interest inside the building.
Such information could be anything, from the locations of paintings
in a museum accompanied by additional background information,
over the locations of printers or conference rooms in an office
building, up to the location of restrooms, check-in desks, or gates
at an airport. Similar to computing evacuation instructions, the
mobile devices in OBESS could be used for navigating users to
such points-of-interest in non-hazardous situations. Furthermore,
OBESS could be used to help users to find each other by sending
a search request via the SSN and waiting for a response from the
user that is searched for.
Another important feature of the CCU in OBESS is the configuration of all sensors in the SSN as to their respective locations inside
the building. The question arises how this can be done in a way
that the manual configuration to be carried out by the user is kept
to a minimum. One approach to reduce this overhead could be
to configure only a small amount of sensors manually and derive
the locations of all other sensors by applying any of the proposed
localization algorithms in Section 2.4.1. In order to do so, it has
to be investigated which sensors in the network require manual
configuration and which localization algorithm has to be applied
to determine the locations of the remaining sensors.
So far, it has been argued in this thesis that a central approach to
evacuation support is undesirable due to the high risk of failure.
However, it could be reasonable to rely on centralized evacuation
planning, as long as the CCU is unimpaired, and to use decentralized mechanisms either supplementary or as a fallback. This
could reduce the amount of communication and computation which
mobile devices have to perform and, hence, increase the lifetime
of their batteries. When it comes to the development of software

279

Chapter 6 Conclusion
running on mobile evacuation devices in OBESS, a challenge which
has to be faced are well-known technical issues which arise with
wireless communication between a multitude of devices, e.g., signal
collisions, interferences, et cetera (cf. Rappaport [190]). However,
there are already several proposals to remedy such problems, for
example published by Priyantha et al. [183].

6.2.2 Swarm Evacuation Planning
In the proposed SEP approaches, the age-value of a message is
solely used to resolve conflicts, although it provides additional
knowledge about how much time has passed since the location
has been computed. Evacuation planning is likely to deliver even
better results, when reported locations of other evacuees in the
building are extrapolated using the age-values of the respective
messages and knowledge about the evacuees’ traveling speeds and
directions, which can be computed using historic location information. Another field of application is to utilize the age-value for
evaluating the reliability of a certain piece of information. When
two paths are only slightly different with respect to their congestion
potential, it could be reasonable to select the path for which the
obtained information is more recent and, hence, more reliable. For
DMO-SEP, additional work is to be done in order to answer the
question how to determine the right weights of different objectives.
Ideally, the weights are derived automatically from user preferences.
However, since a user cannot be expected to define the degree of its
own risk aversion, learning a user’s preferences and characteristics
from its behavior could be one direction of future research in the
area of evacuation management. Smart phones are often connected
to social media platforms, where personal information about the
user can be stored. This could be of help when approaching the

280

6.2 Outlook
task of identifying a user’s preferences. Nevertheless, this remains
an open challenge and, surely, the protection of the user’s privacy
has to be considered when pursuing such an approach.

6.2.3 Range-free Distance Estimation in Mobile Ad Hoc
Networks
In the study about the performance of hop count based distance
estimation under mobile conditions, which has been presented in
this thesis, two indicators have been proposed in order to assess the
mobility in a network. In a next step, these indicators have to be
mapped to a specific adaptation of the hop count values such that
underestimation is compensated and the accuracy of the hop count
values is improved. Such a mapping could, for example, be derived
using a supervised machine learning approach. The multitude
of mobility models described in this thesis forms a rich basis for
generating suitable training data. If it were possible to derive a
fixed mapping, this mapping could then be used to improve the
results of any hop count based distance estimation technique when
applied to dynamic networks.
Apart from refining the standard GA by using hop count adaptation, a synchronization based hop counting method called FIHCd
has been presented in this thesis. One limitation of this method
is that it assumes devices to be able to send signals at arbitrarily
short time intervals. There are cases, however, in which this assumption does not hold. Especially in large networks, the difference
between two successive hop count signals becomes very small with
the proposed method. In order to overcome this drawback and
preserve the benefit of negative hop count error reduction, FIHCd
can be modified as follows. The network is still assumed to be
synchronous but mobile devices send their hop count information

281

Chapter 6 Conclusion
in the form of a message instead of simple signals at the end of
their timers’ periods, similar to standard GA. However, the devices now alternate between a listening and a sending period and
they are only allowed to send messages during the sending period
after having received a message during the listening period which
confirms their hop count value or leads to an update respectively.
In any other case, the sending period turns into a listening period
until a message is received. This method can avoid having devices
send messages before they managed to update their hop counts
after having moved away from their current gradient ring similar
to FIHCd.
As for GeoDE, a potential improvement could be to use a weighted
average of the computed distance estimates between two neighbors.
A device which has a higher number of neighbors in its communication range can be expected to deliver a more accurate estimate;
hence, its initial estimate should be assigned a higher weight when
computing the average value.
Finally, there are several potential improvements of the EA for
optimizing beacon placements, which has been presented in the last
section of Chapter 5. Firstly, the evolutions could be initialized
with regular beacon placements, such as grid-like distributions, in
order to examine whether such simple structures leave room for
improvements. Another improvement could be to refine the fitness
function of the EA. So far, only the average localization error
was taken as a basis for evaluating a particular beacon placement.
Another objective, however, could be to obtain localization errors
which are constant for the time of the entire evacuation and similar
for all evacuees in the building. Furthermore, it could be interesting
to look at a combination of the fitness evaluation criteria proposed.
In general, EAs have many more configuration parameters, which
have not yet been examined, such as the population size, the number

282

6.3 Final Remarks
of parents or children during recombination, the mutation operator
employed, and many more. Before an EA is used productively
in an OBESS system, finding an optimal configuration should be
subject to thorough investigation.

6.3 Final Remarks
While certainly not aiming at providing final solutions to the
challenges of modern and adaptive evacuation management for
buildings, this thesis has demonstrated the huge potential that
lies within the usage of mobile devices for navigation support,
especially during an emergency. The ability of the devices to communicate with each other offers the opportunity to collect relevant
and up-to-date information about the evacuation situation, which
can be used for dynamic and adaptive evacuation route planning.
Two approaches are provided in this thesis for accomplishing this
goal. Moreover, it has been shown how an evacuation system
can be designed in a decentralized and self-organizing manner by
applying concepts from Organic Computing. This provides for a
system which exhibits an almost life-like adaptability and flexibility
towards unforeseen changes in its environment and, at the same
time, stays robust, trustworthy, and controllable. Especially, in
an unpredictable and dynamic situation such as a life-threatening
emergency evacuation, these traits are crucial system properties.
Apart from suggesting a concept for future evacuation management,
the challenge of localizing mobile device inside a building, where
GPS-signals are not available, has been addressed. Problems which
arise due to the mobility of the devices have been identified and
various solution approaches have been suggested and successfully
evaluated in experiments. It is, therefore, believed that the insights
gained in this thesis have the potential to significantly improve

283

Chapter 6 Conclusion
indoor navigation systems. In conclusion, this thesis has laid the
foundation for a changeover in the evacuation support of modern buildings, meeting the standards of the technological progress
which has already found its way into most other areas of human
life.

284

LIST OF TABLES

2.1

Overview of indoor localization systems . . . . . .

4.1

Sample evacuation instructions produced by Capacity Constrained Swarm Evacuation Planning . . . 98
Main configuration parameters for Swarm Simulation104

4.2
5.1
5.2
5.3

69

Overview of investigated mobility models . . . . . 164
Hop count error for each gradient ring of a sample
network . . . . . . . . . . . . . . . . . . . . . . . . 177
Overview of the parameter settings in the simulative experiments of the Evolutionary Algorithm for
optimized beacon placement . . . . . . . . . . . . . 250

285

LIST OF FIGURES

1.1
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8

Navigation signs that are used for evacuation support in current buildings . . . . . . . . . . . . . . .
Main elements of the generic Observer/Controller
Architecture . . . . . . . . . . . . . . . . . . . . . .
Structure of the observer in the Observer/Controller
Architecture . . . . . . . . . . . . . . . . . . . . . .
Structure of the controller in the Observer/Controller Architecture . . . . . . . . . . . . . . . . . . . .
Various structures of Observer/Controller Architectures . . . . . . . . . . . . . . . . . . . . . . . . . .
Overview of research in evacuation optimization . .
Example of a building layout and a corresponding
macroscopic evacuation graph model . . . . . . . .
Time-expanded macroscopic evacuation graph model
Overview of research in evacuation optimization . .

4
19
21
23
25
27
30
31
32

287

List of Figures
2.9
2.10
2.11
2.12
2.13
2.14
2.15
2.16
2.17

Absolute versus relative locations . . . . . . . . . .
Classification of localization algorithms . . . . . . .
Localization via diffusion . . . . . . . . . . . . . .
Localization via APIT method . . . . . . . . . . .
Localization via angulation . . . . . . . . . . . . .
Localization via bounding box . . . . . . . . . . . .
Localization via lateration . . . . . . . . . . . . . .
Classification of distance estimation techniques . .
Hop count determination using the Gradient Algorithm . . . . . . . . . . . . . . . . . . . . . . . . .
2.18 Sample hop count distribution in an ad hoc network
produced by the Gradient Algorithm . . . . . . . .
2.19 Distribution of devices in the communication neighborhood of two adjacent devices . . . . . . . . . .
3.1
3.2
4.1
4.2
4.3
4.4
4.5
4.6

288

Structure of the Organic Building Evacuation Support System . . . . . . . . . . . . . . . . . . . . . .
Observer/Controller Architecture for mobile evacuation devices . . . . . . . . . . . . . . . . . . . . .
Patches which are accessible for an agent during the
evacuation simulation . . . . . . . . . . . . . . . .
Coverage of the different investigated communication ranges . . . . . . . . . . . . . . . . . . . . . .
Sample building plan and a corresponding macroscopic evacuation graph representation . . . . . . .
Total evacuation time for 50 and 100 agents in evacuation Scenario A . . . . . . . . . . . . . . . . . . .
Total evacuation time for 50 and 100 agents in evacuation Scenario B . . . . . . . . . . . . . . . . . . .
Total evacuation time with and without initialization time . . . . . . . . . . . . . . . . . . . . . . .

41
42
49
50
52
53
54
56
60
61
63
73
78
106
108
108
110
114
116

List of Figures
4.7

4.8

4.9

4.10
4.11
4.12
4.13
4.14
4.15
4.16
4.17
4.18
5.1
5.2
5.3

Total evacuation time with initialization time and
with or without repetition of the evacuation planning
algorithm . . . . . . . . . . . . . . . . . . . . . . .
Total evacuation time without initialization time
and with or without repetition of the evacuation
planning algorithm . . . . . . . . . . . . . . . . . .
Total evacuation time with repeated evacuation
planning including patience periods and a maximum target distance in order to allow for route
changes . . . . . . . . . . . . . . . . . . . . . . . .
More complex building plan and its corresponding
macroscopic evacuation graph representation . . .
Robustness test with a simple and a complex building layout . . . . . . . . . . . . . . . . . . . . . . .
Recomputation of shortest path using D*Lite after
discovering a change in the environment . . . . . .
Example of entropy computation . . . . . . . . . .
Comparison of the proposed Swarm Evacuation
Planning approaches . . . . . . . . . . . . . . . . .
Illustration of risk avoidance during evacuation planning . . . . . . . . . . . . . . . . . . . . . . . . . .
Evaluation of risk avoidance experiments . . . . . .
Comparison of evacuation optimized for travel distance, load costs, and entropy costs . . . . . . . . .
Experimental evaluation of congestion avoidance .

117

119

121
123
124
134
141
144
145
146
148
149

Trajectories of devices which are moved according
to the individual mobility models . . . . . . . . . . 161
Illustration of the concept of beacon mobility models163
Trajectories of devices which are moved by individual beacon mobility models . . . . . . . . . . . . . 166

289

List of Figures
5.4
5.5
5.6
5.7
5.8
5.9
5.10
5.11
5.12
5.13
5.14
5.15
5.16
5.17
5.18
5.19
5.20
5.21
5.22
5.23

290

Trajectories of devices which are moved by coupled
beacon mobility models . . . . . . . . . . . . . . . 167
Shift of the gradient border due to low density . . 171
Emergence of positive hop count error due to low
density . . . . . . . . . . . . . . . . . . . . . . . . . 171
Hop count error correction due to mobility . . . . . 171
Emergence of negative hop count errors due to mobility172
Hop count error with various individual mobility
models . . . . . . . . . . . . . . . . . . . . . . . . . 178
Hop count error with various coupled mobility models179
Hop count error with the individual beacon mobility
models . . . . . . . . . . . . . . . . . . . . . . . . . 181
Hop count error with the coupled beacon mobility
models . . . . . . . . . . . . . . . . . . . . . . . . . 182
Influence of speed on the hop count errors . . . . . 184
Influence of direction on the hop count errors . . . 185
Comparison of the Maximum-oriented Gradient Algorithm and the standard Gradient Algorithm . . . 186
Comparison of the Maximum-oriented Gradient Algorithm and the standard Gradient Algorithm . . . 187
Frequency of changes in device-IDs in the neighborhood of dynamic and static devices . . . . . . . . . 189
Frequency of changes in the average hop count in
the neighborhood of dynamic and static devices . . 190
Synchronization of two pulse-coupled oscillators . . 195
Relation between phase of an oscillator and time . 196
Demonstration of Firefly-inspired Hop Counting
over three hops . . . . . . . . . . . . . . . . . . . . 201
Network model for simulative experiments . . . . . 204
Successful synchronizations in static and dynamic
networks . . . . . . . . . . . . . . . . . . . . . . . . 206

List of Figures
5.24 Synchronization time in static and dynamic networks207
5.25 Hop count error in distance estimation using Fireflyinspired Hop Counting compared to the Gradient
Algorithm . . . . . . . . . . . . . . . . . . . . . . . 210
5.26 Proportion of overestimated and underestimated
distances in the network . . . . . . . . . . . . . . . 210
5.27 Classification in shared and individual communication partners of two neighboring devices . . . . . . 213
5.28 Geometric characteristics of two overlapping circles 215
5.29 Approximation of the mapping function used to
derive distances from the classification of communication neighbors . . . . . . . . . . . . . . . . . . . 217
5.30 Irregular distribution of neighbors over the communication ranges of two adjacent devices . . . . . . . 218
5.31 Approximation error of the mapping function used
in Geometric Distance Estimation . . . . . . . . . 219
5.32 Network structures investigated in the experiments 226
5.33 Mean absolute percentage error for Geometric Distance Estimation in uniform-randomly distributed
networks . . . . . . . . . . . . . . . . . . . . . . . . 226
5.34 Mean absolute percentage error for Geometric Distance Estimation in Gaussian-randomly distributed
networks . . . . . . . . . . . . . . . . . . . . . . . . 227
5.35 Mean absolute percentage error for Geometric Distance Estimation in evenly distributed networks . . 227
5.36 Comparison of the localization error with Geometric
Distance Estimation and with the Gradient-based
Distance Estimation for distance estimation . . . . 232
5.37 Influence of the integration approximation quality on
the Gradient-based Distance Estimation for distance
estimation . . . . . . . . . . . . . . . . . . . . . . . 233

291

List of Figures
5.38 Procedure of an Evolutionary Algorithm . . . . . . 239
5.39 Process to determine the perimeter coverage-degree
of an area . . . . . . . . . . . . . . . . . . . . . . . 242
5.40 Sample placement of beacons in a building and the
corresponding genetic representation . . . . . . . . 243
5.41 Tournament selection operator in the Evolutionary
Algorithm for beacon placement optimization . . . 244
5.42 Uniform and one-point crossover operators in the
Evolutionary Algorithm for beacon placement optimization . . . . . . . . . . . . . . . . . . . . . . . . 245
5.43 Evacuation scenario used in the experimental study 249
5.44 Progress of fitness values with various fitness criteria 252
5.45 Standard deviation of fitness values with various
fitness criteria . . . . . . . . . . . . . . . . . . . . . 253
5.46 Evolved solutions with highest fitness value for simulative localization fitness criteria . . . . . . . . . . 254
5.47 Localization error of beacon placements evolved with
various fitness criteria . . . . . . . . . . . . . . . . 256
5.48 Evolved solutions with highest fitness value and lowest localization error using the perimeter coverage
fitness criterion . . . . . . . . . . . . . . . . . . . . 257
5.49 Evolved solutions with lowest localization error for
the hop count based localization and the correct
room mapping fitness criteria . . . . . . . . . . . . 257

292

LIST OF DEFINITIONS

4.1

Definition (Swarm Evacuation Planning) . . . . . .

5.1
5.2
5.3
5.4
5.5
5.6
5.7

Definition
Definition
Definition
Definition
Definition
Definition
Definition

(Passive mobility) . . . . . . . . . . . .
(Ideal Hop Count Value) . . . . . . . . .
(Hop Count Error) . . . . . . . . . . . .
(Maximum-oriented Gradient Algorithm)
(ID-change) . . . . . . . . . . . . . . . .
(HC-change) . . . . . . . . . . . . . . .
(Classification of Neighbors) . . . . . . .

89
157
168
169
174
174
175
212

293

LIST OF ALGORITHMS

4.1
4.2
4.3

Capacity Constrained Swarm Evacuation Planner . 92
Capacity Constrained Routing Planner [143] . . . . 99
D* Lite . . . . . . . . . . . . . . . . . . . . . . . . 133

5.1
5.2
5.3

Geometric Distance Estimation to Neighbors . . . 221
Geometric Distance Estimation to Beacons . . . . 223
Lateration . . . . . . . . . . . . . . . . . . . . . . . 230

295

LIST OF ABBREVIATIONS

ACO Ant Colony Optimization
AoA Angle of Arrival
APIT Approximated Point-in-Triangle Test
BMBF German Federal Ministry for Education and Research
BR Bounded Random Walk
CM Chaos Move
ColM Column Mobility
CCRP Capacity Constrained Routing Planner
CC-SEP Capacity Constrained Swarm Evacuation Planner
CCU Central Control Unit
DFG German Research Foundation

297

List of Abbreviations
DMO-SEP Dynamic Multi-objective Swarm Evacuation Planner
EA Evolutionary Algorithm
FIHC Firefly-inspired Hop Counting
FIHCd Firefly-inspired Hop Counting with Delay
GA Gradient Algorithm
GDE Gradient-based Distance Estimation
GeoDE Geometric Distance Estimation
GI German Informatics Society
GM Gauss Markov Move
GPS Global Positioning System
IETF Internet Engineering Task Force
ITG Information Technology Society
LCS Learning Classifier System
MANET Mobile Ad Hoc Network
MAPE Mean absolute percentage error
MDF Maximum Dynamic Flow
MDS-Map Multidimensional Scaling Map
MoGA Maximum-oriented Gradient Algorithm
NomM Nomadic Move

298

OBESS Organic Building Evacuation Support System
O/C Architecture Observer/Controller Architecture
PR Probabilistic Random Walk
QF Quickest Flow
RD Random Direction Walk
RefM Reference Point Mobility
RSSI Radio Signal Strength Indication
RW Random Walk
SDP Semi-definite Programming
SEP Swarm Evacuation Planning
SSN Static Sensor Network
StrM Stream Mobility
SuOC System under Observation and Control
TDoA Time Difference of Arrival
ToA Time of Arrival
ToF Time of Flight
UMF Universal Maximum Flow
WLAN Wireless Local Area Network

299

BIBLIOGRAPHY

[1] N. Abramson. The ALOHA system: Another alternative
for computer communications. In Proceedings of the 1970
Fall Joint Computer Conference, pages 281–285. ACM Press,
1970.
[2] AG Optimierung - TU Kaiserslautern. REPKA - Regionale
Evakuierung: Planung, Kontrolle und Anpassung, January
2013. http://www.repka-evakuierung.de.
[3] R. Akl, K. Pasupathy, and M. Haidar. Anchor nodes placement for effective passive localization. In Proceedings of the
2011 International Conference on Selected Topics in Mobile
and Wireless Networking, pages 127–132. IEEE, 2011.
[4] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and
E. Cayirci. Wireless sensor networks: A survey. Computer
Networks, 38:393–422, 2002.

301

Bibliography
[5] R. Alizadeh. A dynamic cellular automaton model for evacuation process with obstacles. Safety Science, 49(2):315–323,
2011.
[6] F. Allerding, B. Becker, and H. Schmeck. Decentralised
energy management for smart homes. In C. Müller-Schloer,
H. Schmeck, and T. Ungerer, editors, Organic Computing A Paradigm Shift for Complex Systems, Autonomic Systems,
pages 605–607. Springer, 2011.
[7] K. K. Almuzaini and T. A. Gulliver. A new distributed
range-free localization algorithm for wireless networks. In
Proceedings of the 70th IEEE Vehicular Technology Conference, pages 174–178. IEEE, 2009.
[8] T. Amemiya, J. Yamashita, K. Hirota, and M. Hirose. Virtual
leading blocks for the deaf-blind: a real-time way-finder by
verbal-nonverbal hybrid interface and high-density RFID
tag space. In Proceedings of the 2004 IEEE Virtual Reality
Conference, pages 165–287. IEEE, 2004.
[9] M. Arikawa, S. Konomi, and K. Ohnishi. Navitime: Supporting pedestrian navigation in the real world. IEEE Pervasive
Computing, 6(3):21–29, 2007.
[10] F. Aslam, C. Schindelhauer, and A. Vater. Improving geometric distance estimation for sensor networks and unit disk
graphs. In Proceedings of the 2009 International Conference
on Ultra Modern Telecommunications Workshops, pages 1–5,
2009.
[11] J. Aspnes, D. Goldenberg, and Y. R. Yang. On the computational complexity of sensor network localization. In

302

Bibliography
Proceedings of 1st International Workshop on Algorithmic
Aspects of Wireless Sensor Networks, pages 32–44. Springer,
2004.
[12] J. Bachrach and C. Taylor. Localization in sensor networks.
In Handbook of Sensor Networks, pages 277–310. John Wiley
& Sons, 2005.
[13] P. Bahl and V. N. Padmanabhan. RADAR: an in-building
RF-based user location and tracking system. In Proceedings of
the 19th Annual Joint Conference of the IEEE Computer and
Communications Societies, pages 775–784. IEEE Computer
Society, 2000.
[14] S. M. Ballew. Managing IP networks with Cisco Routers.
O’Reilly, 1997.
[15] D. Bartlett. Essentials of Positioning and Location Technology. Cambridge University Press, 2013.
[16] J. Baus, A. Krüger, and W. Wahlster. A resource-adaptive
mobile navigation system. In Proceedings of the 7th International Conference on Intelligent User Interfaces, pages 15–22.
ACM Press, 2002.
[17] P. Bergamo and G. Mazzini. Localization in sensor networks
with fading and mobility. In Proceedings of the 13th IEEE
International Symposium on Personal, Indoor and Mobile
Radio Communications, pages 750–754. IEEE, 2002.
[18] M. Bessho, S. Kobayashi, N. Koshizuka, and K. Sakamura. A
space-identifying ubiquitous infrastructure and its application
for tour-guiding service. In Proceedings of the 2008 ACM

303

Bibliography
symposium on Applied computing, pages 1616–1621. ACM
Press, 2008.
[19] K. Bing, L. Fu, Y. Zhuo, and L. Yanlei. Design of an internet of things-based smart home system. In Proceedings of
the 2nd International Conference on Intelligent Control and
Information Processing, pages 921–924. Curran Associates,
2011.
[20] Bluetooth SIG, Inc. Bluetooth technology, July 2013.
http://www.bluetooth.com.
[21] S. Bottani. Pulse-coupled relaxation oscillators: From biological synchronization to self-organized criticality. Physical
Review Letters, 74(21):4189–4192, 1995.
[22] A. Boukerche, B. Turgut, N. Aydin, M. Z. Ahmad, L. Bölöni,
and D. Turgut. Routing protocols in ad hoc networks: A
survey. Computer Networks, 55(13):3032–3080, 2011.
[23] J. Branke, M. Mnif, C. Müller-Schloer, H. Prothmann,
U. Richter, F. Rochner, and H. Schmeck. Organic Computing
- addressing complexity by controlled self-organization. In
Proceedings of the 2nd International Symposium on Leveraging Applications of Formal Methods, Verification and Validation, pages 185–191. IEEE, 2006.
[24] H. Breu and D. G. Kirkpatrick. Unit disk graph recognition
is NP-hard. Computational Geometry, 9(1-2):3–24, 1998.
[25] W. Brockmann, E. Maehle, K.-E. Grosspietsch, N. Rosemann,
and B. Jakimovski. ORCA: An organic robot control architecture. In C. Müller-Schloer, H. Schmeck, and T. Ungerer,

304

Bibliography
editors, Organic Computing - A Paradigm Shift for Complex
Systems, Autonomic Systems, pages 385–398. Springer, 2011.
[26] R. Brown. Social Psychology, 2nd Ed. Free Press, 2003.
[27] R. Bucher and D. Misra. A synthesizable VHDL model of the
exact solution for three-dimensional hyperbolic positioning
system. VLSI Design, 15(2):507–520, 2002.
[28] J. B. Buck. Synchronous rhythmic flashing of fireflies. The
Quarterly Review of Biology, 13(3):301–314, 1938.
[29] N. Bulusu, J. Heidemann, and D. Estrin. GPS-less low-cost
outdoor localization for very small devices. IEEE Personal
Communications, 7(5):28–34, 2000.
[30] N. Bulusu, V. Bychkovskiy, D. Estrin, and J. Heidemann.
Scalable, ad hoc deployable RF-based localization. Proceedings of the Grace Hopper Conference on Celebration of
Women in Computing, 2002.
[31] R. Burkard, K. Dlaska, and B. Klinz. The quickest flow
problem. Mathematical Methods of Operations Research, 37
(1):31–58, 1993.
[32] C Burstedde, K. Klauck, A. Schadschneider, and J. Zittartz.
Simulation of pedestrian dynamics using a two-dimensional
cellular automaton. Physica A: Statistical Mechanics and its
Applications, 295(3-4):507–525, 2001.
[33] C. Buschmann, H. Hellbrück, S. Fischer, A. Kröller, and
S. P. Fekete. Radio propagation-aware distance estimation
based on neighborhood comparison. In Proceedings of the 4th
European Conference on Wireless Sensor Networks, pages
325–340. Springer, 2007.

305

Bibliography
[34] T. Camp, J. Boleng, and V. Davies. A survey of mobility
models for ad hoc network research. Wireless Communications and Mobile Computing, Special Issue On Mobile Ad Hoc
Networking: Research, Trends and Applications, 2:483–502,
2002.
[35] M. Cardei and J. Wu. Energy-efficient coverage problems in
wireless ad-hoc sensor networks. Computer Communications,
29(4):413–420, 2006.
[36] L. G. Chalmet, R. L. Francis, and P. B. Saunders. Network
models for building evacuation. Fire Technology, 18(1):90–
113, 1982.
[37] Y.-J. Chang, S.-K. Tsai, and T.-Y. Wang. A context aware
handheld wayfinding system for individuals with cognitive
impairments. In Proceedings of the 10th International ACM
SIGACCESS Conference on Computers and Accessibility,
pages 27–34. ACM Press, 2008.
[38] P. Chatterjee and N. Das. A distributed algorithm for loadbalanced routing in multihop wireless sensor networks. In Proceedings of the 9th International Conference on Distributed
Computing and Networking, pages 332–338. Springer, 2008.
[39] N. Cheng. An optimization method for dynamic evacuation
route programming based on improved ant colony algorithm.
In Proceedings of the 2010 International Conference on Intelligent System Design and Engineering Application, pages
265–267. IEEE, 2010.
[40] K. K. Chintalapudi, A. Dhariwal, R. Govindan, and
G. Sukhatme. Ad-hoc localization using ranging and sectoring. In Proceedings of the 23rd Annual Joint Conference of

306

Bibliography
the IEEE Computer and Communications Societies, pages
2662–2672. IEEE, 2004.
[41] H. M. Choset. Principles of Robot Motion: Theory, Algorithms, and Implementation. MIT Press, 2005.
[42] C.-H. A. Chou, K.-F. Ssu, H. C. Jiau, W.-T. Wang, and
C. Wang. A dead-end free topology maintenance protocol
for geographic forwarding in wireless sensor networks. IEEE
Transactions on Computers, 60(11):1610–1621, 2011.
[43] B. N. Clark, C. J. Colbourn, and D. S. Johnson. Unit disk
graphs. Discrete Mathematics, 86(1–3):165–177, 1990.
[44] D. Coore. Establishing a coordinate system on an amorphous
computer. Techreport MIT/LCS/TR-737, MIT Laboratory
for Computer Science, 1998.
[45] D. Corona and B. De Schutter. Adaptive cruise control for a
smart car: A comparison benchmark for MPC-PWA control
methods. IEEE Transactions on Control Systems Technology,
16(2):365–372, 2008.
[46] G. Daniels, K. Wallasch, and B. Stock. BS 9999 - Neue
Vorgehensweise bei der risikobasierten Auslegung des Brandschutzes. vfdb-Zeitschrift Forschung, Technik und Management im Brandschutz, 58:93–101, 2009.
[47] C. Darwin. On the Origin of the Species by Means of Natural
Selection: Or, The Preservation of Favoured Races in the
Struggle for Life. John Murray, 1859.
[48] E. D’Atri, C. M. Medaglia, A. Serbanati, and U. B. Ceipidor.
A system to aid blind people in the mobility: A usability

307

Bibliography
test and its results. In Proceedings of the 2nd International
Conference on Systems, pages 35–35. IEEE, 2007.
[49] P. Davidsson and M. Boman. Distributed monitoring and
control of office buildings by embedded agents. Information
Sciences, 171(4):293–307, 2005.
[50] L. Dawei, G. Lihua, and W. Li. Emergency evacuation based
on the nested game analysis. In Proceedings of the 29th
Chinese Control Conference, pages 1733–1737. IEEE, 2010.
[51] B. de Ruyter and E. Pelgrim. Ambient assisted-living research
in carlab. Interactions, 14(4):30–33, 2007.
[52] I. Dietrich and F. Dressler. On the lifetime of wireless sensor
networks. ACM Transactions on Sensor Networking, 5(1):
1–39, 2009.
[53] E. W. Dijkstra. A note on two problems in connexion with
graphs. Numerische Mathematik, 1(1):269–271, 1959.
[54] B. Ding, H. Yuan, X. Zang, and L. Jiang. The research on
blind navigation system based on RFID. In Proceedings of the
2007 International Conference on Wireless Communications,
Networking and Mobile Computing, pages 2058–2061. IEEE,
2007.
[55] L. Doherty, K. S. J. Pister, and L. El Ghaoui. Convex position
estimation in wireless sensor networks. In Proceedings of the
20th Annual Joint Conference of the IEEE Computer and
Communications Societies, pages 1655–1663. IEEE, 2001.
[56] M. Dorigo. Optimization, learning and natural algorithms.
Ph.D. dissertation, Politecnico di Milano, Milano, Italy, 1992.

308

Bibliography
[57] M. Dorigo, V. Maniezzo, and A. Colorni. Ant system: Optimization by a colony of cooperating agents. IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics,
26(1):29–41, 1996.
[58] R. C. Eberhart, Y. Shi, and J. Kennedy. Swarm Intelligence.
Morgan Kaufmann, 2001.
[59] J. M. Epstein, R. Pankajakshan, and R. A. Hammond. Combining computational fluid dynamics and agent-based modeling: A new approach to evacuation planning. PLOS ONE, 6
(5):e20139, 2011.
[60] T. Eren. Cooperative localization in wireless ad hoc and
sensor networks using hybrid distance and learing (angle
of arrival) measurements. EURASIP Journal on Wireless
Communications and Networking, 2011(73):1–18, 2011.
[61] M. Evans and T. Swartz. Approximating Integrals via Monte
Carlo and Deterministic Methods. Oxford University Press,
2000.
[62] R. F. Fahy and G. Proulx. Analysis of published accounts
of the world trade center evacuation. federal building and
fire safety investigation of the world trade center disaster.
Federal Building and Fire Safety Investigation Report (NIST
NCSTAR 1-7A), US Department of Commerce, NIST, 2010.
[63] N. Fallah, I. Apostolopoulos, K. Bekris, and E. Folmer. Indoor human navigation systems: A survey. Interacting with
Computers, 25(1):21–33, 2013.
[64] S. P. Fekete, A. Kröller, C. Buschmann, and S. Fischer.
Geometric distance estimation for sensor networks and unit

309

Bibliography
disk graphs. In Proceedings of the 16th International Fall
Workshop on Computational Geometry, 2006.
[65] S. P. Fekete, B. Hendriks, C. Tessars, A. Wegener, H. Hellbrück, S. Fischer, and S. Ebers. Methods for improving
the flow of traffic. In C. Müller-Schloer, H. Schmeck, and
T. Ungerer, editors, Organic Computing - A Paradigm Shift
for Complex Systems, Autonomic Systems, pages 447–460.
Springer, 2011.
[66] A. Filippoupolitis, G. Gorbil, and E. Gelenbe. Spatial computers for emergency management. In Proceedings of the
5th IEEE Conference on Self-Adaptive and Self-Organizing
Systems Workshops, pages 61–66. IEEE, 2011.
[67] C. Fischer, K. Muthukrishnan, M. Hazas, and H. Gellersen.
Ultrasound-aided pedestrian dead reckoning for indoor navigation. In Proceedings of the 1st ACM International Workshop on Mobile Entity Localization and Tracking in GPS-less
Environments, pages 31–36. ACM Press, 2008.
[68] L. R. Ford and D. R. Fulkerson. Flows in Networks. Princeton
University Press, 1962.
[69] G. A. Frank and C. O. Dorso. Room evacuation in the
presence of an obstacle. Physica A: Statistical Mechanics
and its Applications, 390(11):2135–2145, 2011.
[70] N. Fredivianus, U. Richter, and H. Schmeck. Collaborating
and learning predators on a pursuit scenario. In M. Hinchey,
B. Kleinjohann, L. Kleinjohann, P. A. Lindsay, F. J. Rammig,
J. Timmis, and M. Wolf, editors, Distributed, Parallel and
Biologically Inspired Systems, volume 329 of IFIP Advances

310

Bibliography
in Information and Communication Technology, pages 290–
301. Springer, 2010.
[71] D. Gale. Transient flows in networks. The Michigan Mathematical Journal, 6(1):59–63, 1959.
[72] A. Garrett, B. Carnahan, R. Muhdi, J. Davis, G. Dozier,
M. P. SanSoucie, P. V. Hull, and M. L. Tinker. Evacuation
planning via evolutionary computation. In Proceedings of the
2006 IEEE Congress on Evolutionary Computation, pages
157–164. IEEE, 2006.
[73] J. Gehrke and S. Madden. Query processing in sensor networks. IEEE Pervasive Computing, 3(1):46–55, 2004.
[74] I. Gerdes, F. Klawonn, and R. Kruse. Evolutionäre Algorithmen: Genetische Algorithmen, Strategien und Optimierungsverfahren, Beispielanwendungen. Vieweg und Teubner, 1st edition, 2004.
[75] P. Ghosekar, G. Katkar, and P. Ghorpade. Mobile ad hoc
networking: Imperatives and challenges. International Journal of Computer Applications, Special Issue on MANETs, 3:
153–158, 2010.
[76] P. G. Gipps and B. Marksjö. A micro-simulation model for
pedestrian flows. Mathematics and Computers in Simulation,
27(2-3):95–105, 1985.
[77] L. Girod and D. Estrin. Robust range estimation using
acoustic and multimodal sensing. In Proceedings of the 2001
IEEE/RSJ International Conference on Intelligent Robots
and Systems, pages 1312–1320. IEEE, 2001.

311

Bibliography
[78] A. R. Golding and N. Lesh. Indoor navigation using a diverse
set of cheap, wearable sensors. In The Third International
Symposium on Wearable Computers, pages 29–36. IEEE,
1999.
[79] R. Y. Guo and H. J. Huang. A mobile lattice gas model
for simulating pedestrian evacuation. Physica A: Statistical
Mechanics and its Applications, 387(2-3):580–586, 2008.
[80] F. Gustafsson and F. Gunnarsson. Positioning using timedifference of arrival measurements. In Proceedings of the
2003 IEEE International Conference on Acoustics, Speech,
and Signal Processing, pages 553–556. IEEE, 2003.
[81] Z. J. Haas, J. Y. Halpern, and L. Li. Gossip-based ad hoc
routing. IEEE/ACM Transactions on Networking, 14(3):
479–491, 2006.
[82] H. W. Hamacher and S. A. Tjandra. Mathematical modelling
of evacuation problems : A state of art. Pedestrian and
Evacuation Dynamics, 24:227–266, 2001.
[83] P. E. Hart, N. J. Nilsson, and B. Raphael. A formal basis for
the heuristic determination of minimum cost paths. IEEE
Transactions on Systems Science and Cybernetics, 4(2):100–
107, 1968.
[84] P. E. Hart, N. J. Nilsson, and B. Raphael. Correction to "A
formal basis for the heuristic determination of minimum cost
paths". SIGART Bulletin, 37:28–29, 1972.
[85] A. Harter, A. Hopper, P. Steggles, A. Ward, and P. Webster.
The anatomy of a context-aware application. In Proceedings
of the 5th Annual ACM/IEEE International Conference on

312

Bibliography
Mobile Computing and Networking, pages 59–68. ACM Press,
1999.
[86] T. He, C. Huang, B. M. Blum, J. A. Stankovic, and T. Abdelzaher. Range-free localization schemes for large scale
sensor networks. In Proceedings of the 9th Annual International Conference on Mobile Computing and Networking,
pages 81–95. ACM Press, 2003.
[87] E. Heidari and A. Movaghar. An efficient method based
on genetic algorithms to solve sensor network optimization
problem. International Journal on Applications of Graph
Theory in Wireless Ad Hoc Networks and Sensor Networks,
3(1):18–33, 2011.
[88] D. Helbing. A fluid dynamic model for the movement of
pedestrians. Complex Systems, 6:391–415, 1998.
[89] D. Helbing and P. Molnar. Social force model for pedestrian
dynamics. Physical Review E, 51(5):4282–4286, 1995.
[90] D. Helbing, I. Farkas, and T. Vicsek. Simulating dynamical
features of escape panic. Nature, 407(6803):487–490, 2000.
[91] L. F. Henderson. The statistics of crowd fluids. Nature, 229
(5284):381–383, 1971.
[92] L. F. Henderson. On the fluid mechanics of human crowd
motion. Transportation Research, 8(6):509–515, 1974.
[93] T. Hestermeyer, O. Oberschelp, and H. Giese. Structured
information processing for self-optimizing mechatronic systems. In H. Araújo, A. Vieira, J. Braz, B. Encarnação,

313

Bibliography
and M. Carvalho, editors, Proceedings of the 2004 International Conference on Informatics in Control, Automation
and Robotics, pages 230–237. IEEE, 2004.
[94] J. Hightower and G. Borriello. Location systems for ubiquitous computing. Computer, 34(8):57–66, 2001.
[95] J. Hightower, G. Borriello, and R. Want. SpotON: An indoor
3D location sensing technology based on RF signal strength.
Techreport 2000-02-02, University of Washington, 2000.
[96] T. Höllerer, D. Hallaway, N. Tinna, and S. Feiner. Steps
toward accommodating variable position tracking accuracy
in a mobile augmented reality system. In In Proceedings of
the 2nd International Workshop on Artificial Intelligence in
Mobile Systems, pages 31–37. AAAI Press, 2001.
[97] L. Hu and D. Evans. Localization for mobile sensor networks.
In Proceedings of the 10th Annual International Conference
on Mobile Computing and Networking, pages 45–57. ACM
Press, 2004.
[98] Y.-C. Hu, A. Perrig, and D. B. Johnson. Packet leashes: A
defense against wormhole attacks in wireless networks. In
Proceedings of the 22nd Annual Joint Conference of the IEEE
Computer and Communications, volume 3, pages 1976–1986.
IEEE, 2003.
[99] B. Huang, C. Yu, B. D. O. Anderson, and G. Mao.
Connectivity-based distance estimation in wireless sensor
networks. In Proceedings of the 2010 Global Telecommunications Conference, pages 4565–4569. IEEE, 2010.

314

Bibliography
[100] C.-F. Huang and Y.-C. Tseng. The coverage problem in a
wireless sensor network. In Proceedings of the 2nd ACM
International Conference on Wireless Sensor Networks and
Applications, pages 115–121. ACM Press, 2003.
[101] H. Huang, G. Gartner, M. Schmidt, and Y. Li. Smart environment for ubiquitous indoor navigation. In Proceedings of the
2009 International Conference on New Trends in Information
and Service Science, pages 176–180. IEEE, 2009.
[102] Q. Huang and S. Selvakennedy. A range-free localization
algorithm for wireless sensor networks. In Proceedings of
the 63rd IEEE Conference on Vehicular Technology, pages
349–353. IEEE, 2006.
[103] A. Hub, J. Diepstraten, and T. Ertl. Design and development of an indoor navigation and object identification
system for the blind. In Proceedings of the 6th international
ACM SIGACCESS conference on Computers and accessibility, pages 147–152. ACM Press, 2003.
[104] IEEE
Standards
Association.
IEEE
802.11
Standards,
October
2013.
http://standards.ieee.org/about/get/802/802.11.html.
[105] M. Ilyas and I. Mahgoub. Smart Dust: Sensor Network
Applications, Architecture And Design. CRC Taylor & Francis,
2006.
[106] Y. Inoue, A. Sashima, T. Ikeda, and K. Kurumatani. Indoor emergency evacuation service on autonomous navigation
system using mobile phone. In Proceedings of the 2nd International Symposium on Universal Communication, pages
79–85. IEEE, 2008.

315

Bibliography
[107] Internet Engineering Task Force (IETF). Mobile adhoc networks (MANET) - charter, November 2011.
http://datatracker.ietf.org/wg/manet/charter.
[108] E. Isaacson and H. B. Keller. Analysis of Numerical Methods.
Dover Publications, reprint edition, 1994. ISBN 0486680290.
[109] J. Izquierdo, I. Montalvo, R. Pérez, and V.S. Fuertes. Forecasting pedestrian evacuation times by using swarm intelligence. Physica A: Statistical Mechanics and its Applications,
388(7):1213–1220, 2009.
[110] H.-C. Jang, Y.-N. Lien, and T.-C. Tsai. Rescue information
system for earthquake disasters based on MANET emergency
communication platform. In Proceedings of the 2009 International Conference on Wireless Communications and Mobile
Computing: Connecting the World Wirelessly, pages 623–627.
ACM Press, 2009.
[111] J. J. Jarvis and H. D. Ratliff. Some equivalent objectives for
dynamic network flow problems. Management Science, 28
(1):106–109, 1982.
[112] X. Ji and H. Zha. Sensor positioning in wireless ad-hoc sensor
networks using multidimensional scaling. In Proceedings of
the 23rd Annual Joint Conference of the IEEE Computer and
Communications Societies, pages 2652–2661. IEEE, 2004.
[113] A. Johansson and D. Helbing. Pedestrian flow optimization
with a genetic algorithm based on boolean grids. In N. Waldau, P. Gattermann, H. Knoflacher, and M. Schreckenberg,
editors, Pedestrian and Evacuation Dynamics 2005, pages
267–272. Springer, 2007.

316

Bibliography
[114] W. Kai and C. Chun. Using RSS with difference method in
localization algorithm for sensor networks. In Proceedings
of the 2nd International Conference of Information Science
and Engineering, pages 2500–2502. IEEE, 2010.
[115] H. Kaplan, M. J. Katz, G. Morgenstern, and M. Sharir.
Optimal cover of points by disks in a simple polygon. In
M. Berg and U. Meyer, editors, European Symposium on
Algorithms, volume 6346 of LNCS, pages 475–486. Springer,
2010.
[116] D. Karaboga and B. Akay. A survey: Algorithms simulating
bee swarm intelligence. Artificial Intelligence Review, 31(1-4):
61–85, 2009.
[117] C. Karlof and D. Wagner. Secure routing in wireless sensor
networks: attacks and countermeasures. In Proceedings of
the 1st IEEE International Workshop on Sensor Network
Protocols and Applications, pages 113–127. IEEE, 2003.
[118] V. Kaseva, T. D. Hamalainen, and M. Hannikainen. Rangefree algorithm for energy-efficient indoor localization in wireless sensor networks. In Proceedings of the 2011 Conference
on Design Architectures for Signal Image Processing, pages
1–8. IEEE, 2011.
[119] M. J. Katz and G. Morgenstern. A scheme for computing
minimum covers within simple regions. In Proceedings of
the 11th International Symposium on Algorithms and Data
Structures, pages 447–458. Springer, 2009.
[120] L. Kellenberger and R. Müller. A genetic algorithm module
for spatial optimization in pedestrian simulation. In W.W.F.

317

Bibliography
Klingsch, C. Rogsch, A. Schadschneider, and M. Schreckenberg, editors, Pedestrian and Evacuation Dynamics 2008,
pages 359–370. Springer, 2010. ISBN 978-3-642-04503-5,
978-3-642-04504-2.
[121] J. O. Kephart and D. M. Chess. The vision of autonomic
computing. Computer, 36(1):41–50, 2003.
[122] L. Kleinrock and J. Silvester. Optimum transmission radii
for packet radio networks or why six is a magic number.
In Proceedings of the 1978 National Telecommunications
Conference, pages 431–435, 1978.
[123] S. Koide and M. Kato. 3-d human navigation system considering various transition preferences. In Proceedings of the
2005 IEEE International Conference on Systems, Man and
Cybernetics, pages 859–864. IEEE, 2005.
[124] L. König, S. Mostaghim, and S. Schmeck. A markov-chainbased model for success prediction of evolution in complex
environments. In Proceedings of the 3rd International Joint
Conference on Computational Intelligence, pages 90–102.
IEEE Computer Society, 2011.
[125] S. König and M. Likhachev. Improved fast replanning for
robot navigation in unknown terrain. In Proceedings of
the 2002 IEEE International Conference on Robotics and
Automation, volume 1, pages 968–975, 2002.
[126] T. Kretz. Pedestrian traffic: On the quickest path. Journal
of Statistical Mechanics: Theory and Experiment, 2009(3):
P03012, 2009.

318

Bibliography
[127] A. Kröller, S. P. Fekete, C. Buschmann, S. Fischer, and
D. Pfisterer. Koordinatenfreies Lokationsbewusstsein (localization without coordinates). it Information Technology, 47
(2):15, 2005.
[128] J. Krumm, S. Harris, B. Meyers, B. Brumitt, M. Hale, and
S. Shafer. Multi-camera multi-person tracking for EasyLiving.
In Proceedings of the 3rd IEEE International Workshop on
Visual Surveillance, pages 3–10. IEEE, 2000.
[129] V. Kulyukin, C. Gharpure, J. Nicholson, and G. Osborne.
Robot-assisted wayfinding for the visually impaired in structured indoor environments. Autonomous Robots, 21(1):29–41,
2006.
[130] V. Kumar and S. R. Das. Performance of dead reckoningbased location service for mobile ad hoc networks: Research
articles. Wireless Communications and Mobile Computing, 4
(2):189–202, 2004.
[131] Y. Kwon, K. Mechitov, S. Sundresh, W. Kim, and G. Agha.
Resilient localization for sensor networks in outdoor environments. In Proceedings of the 25th IEEE International
Conference on Distributed Computing Systems, pages 643–
652. IEEE, 2005.
[132] S. M. LaValle. Planning Algorithms. Cambridge University
Press, 2006.
[133] H. Li, J. Wang, X. Li, and H. Ma. Real-time path planning
of mobile anchor node in localization for wireless sensor
networks. In Proceedings of the 2008 IEEE International
Conference on Information and Automation, pages 384–389.
IEEE, 2008.

319

Bibliography
[134] X. Li, H. Shi, and Y. Shang. Selective anchor placement
algorithm for ad-hoc wireless sensor networks. In Proceedings
of the 2008 International Conference on Communications,
pages 2359–2363. IEEE, 2008.
[135] S. Y. Liana, R. L. Hecker, and R. G. Landers. Machining
process monitoring and control: The state-of-the-art. Journal
of manufacturing science and engineering, 126(2):297–310,
2004.
[136] W.-H. Liao, Y.-C. Tseng, and J.-P. Sheu. GRID: a fully
location-aware routing protocol for mobile ad hoc networks.
Telecommunication Systems, 18(1-3):37–60, 2001.
[137] W.-H. Liao, Y.-C. Lee, and S. P. Kedia. Mobile anchor positioning for wireless sensor networks. IET Communications,
5(7):914–921, 2011.
[138] Y.-N. Lien, H.-C. Jang, and T.-C. Tsai. A MANET based
emergency communication and information system for catastrophic natural disasters. In Proceedings of the 29th IEEE
International Conference on Distributed Computing Systems
Workshops, pages 412–417. IEEE, 2009.
[139] J. G. Lim and S. V. Rao. Mobility-enhanced positioning in ad
hoc networks. In Proceedings of the 2003 IEEE Conference on
Wireless Communications and Networking, volume 3, pages
1832–1837, 2003.
[140] Q. Liu, A. Pruteanu, and S. Dulman. GDE: a distributed
gradient-based algorithm for distance estimation in largescale networks. In Proceedings of the 14th ACM International
Conference on Modeling, Analysis and Simulation of Wireless
and Mobile Systems, pages 151–158. ACM Press, 2011.

320

Bibliography
[141] S. M. Lo, H. C. Huang, P. Wang, and K. K. Yuen. A game
theory based exit selection model for evacuation. Fire Safety
Journal, 41(5):364–369, 2006.
[142] Q. Lu, Y. Huang, and S. Shekhar. Evacuation planning: A capacity constrained routing approach. In H. Chen, R. Miranda,
D. D. Zeng, C. Demchak, J. Schroeder, and T. Madhusudan,
editors, Intelligence and Security Informatics, volume 2665
of LNCS, pages 111–125. Springer, 2003.
[143] Q. Lu, B. George, and S. Shekhar. Capacity constrained
routing algorithms for evacuation planning: A summary of
results. In C. B. Medeiros, M. J. Egenhofer, and E. Bertino,
editors, Advances in Spatial and Temporal Databases, volume
3633 of LNCS, pages 291–307. Springer, 2005.
[144] D. Lucarelli and I.-J. Wang. Decentralized synchronization
protocols with nearest neighbor communication. In Proceedings of the 2nd International Conference on Embedded
Networked Sensor Systems, pages 62–68. ACM Press, 2004.
[145] C. Maihofer. A survey of geocast routing protocols. IEEE
Communications Surveys Tutorials, 6(2):32–42, 2004.
[146] G. Mao, B. Fidan, and B. D. O. Anderson. Wireless sensor
network localization techniques. Computer Networks, 51(10):
252–2553, 2007.
[147] R. Mathar and J. Mattfeldt. Pulse-coupled decentral synchronization. SIAM Journal on Applied Mathematics, 56(4):
1094–1106, 1996.
[148] M. K. McClintock. Menstrual synchrony and suppression.
Nature, 229(5282):244–245, 1971.

321

Bibliography
[149] L. Meertens and S. Fitzpatrick. The distributed construction of a global coordinate system in a network of static
computational nodes from inter-node distances. Techreport
KES.U.04.04, Kestrel Institute, 2004.
[150] S. K. Meghani, M. Asif, and S. Amir. Localization of WSN
node based on time of arrival using ultra wide band spectrum.
In Proceedings of the 13th Annual Wireless and Microwave
Technology Conference, pages 1–4. IEEE, 2012.
[151] S. Merkel, L. König, and H. Schmeck. Age based controller
stabilization in evolutionary robotics. In Proceedings of the
2nd World Congress on Nature and Biologically Inspired
Computing, pages 84–91. IEEE, 2010.
[152] S. Merkel, C. W. Becker, and H. Schmeck. Firefly-inspired
synchronization for energy-efficient distance estimation in mobile ad-hoc networks. In Proceedings of the 31st International
Performance Computing and Communications Conference,
pages 205–214. IEEE, 2012.
[153] S. Merkel, S. Mostaghim, and H. Schmeck. Distributed geometric distance estimation in ad hoc networks. In D. Hutchison, T. Kanade, J. Kittler, J. M. Kleinberg, F. Mattern,
J. C. Mitchell, M. Naor, O. Nierstrasz, C. Pandu Rangan,
B. Steffen, M. Sudan, D. Terzopoulos, D. Tygar, M. Y. Vardi,
G. Weikum, X.-Y. Li, S. Papavassiliou, and S. Ruehrup, editors, Ad-hoc, Mobile, and Wireless Networks, volume 7363
of LNCS, pages 28–41. Springer, 2012.
[154] S. Merkel, S. Mostaghim, and H. Schmeck. A study of mobility in ad hoc networks and its effects on a hop count

322

Bibliography
based distance estimation. In Proceedings of the 5th International Conference on New Technologies, Mobility and
Security, pages 39–43. IEEE, 2012.
[155] S. Merkel, S. Mostaghim, D. Blum, and H. Schmeck. Distributed swarm evacuation planning. In Proceedings of the
2013 IEEE Symposium on Swarm Intelligence, pages 276–283.
IEEE, 2013.
[156] S. Merkel, S. Mostaghim, and H. Schmeck. Hop count based
distance estimation in mobile ad hoc networks - challenges
and consequences. Ad Hoc Networks, 11(8), 2013. (to appear).
[157] S. Merkel, P. Unger, and H. Schmeck. Evolutionary algorithm for optimal anchor node placement to localize devices
in a mobile ad hoc network during building evacuation. In
Proceeding of the 15th Annual Conference Companion on Genetic and Evolutionary Computation Conference Companion,
pages 1407–1414. ACM Press, 2013.
[158] S. Merkel, A. Kesseler, and H. Schmeck. Dynamic multiobjective evacuation path planning in mobile ad hoc networks.
Techreport 3044, Karlsruhe Institute of Technology (KIT),
2014.
[159] S. Merkel, S. Mostaghim, and H. Schmeck. Self-organized
swarm display. International Journal of Swarm Intelligence,
2014. revision submitted on 10th of December 2013.
[160] Yoshihiro Ishibashi Minoru Fukui. Jamming transition in
cellular automaton models for pedestrians on passageway.
Journal of The Physical Society of Japan, 68(11):3738–3739,
1999.

323

Bibliography
[161] R. E. Mirollo and S. H. Strogatz. Synchronization of pulsecoupled biological oscillators. SIAM Journal on Applied
Mathematics, 50:1645–1662, 1990.
[162] M. Mnif and C. Müller-Schloer. Quantitative emergence.
In C. Müller-Schloer, H. Schmeck, and T. Ungerer, editors,
Organic Computing - A Paradigm Shift for Complex Systems,
Autonomic Systems, pages 39–52. Springer, 2011.
[163] M. Mnif, U. Richter, J. Branke, H. Schmeck, and C. MüllerSchloer. Measurement and control of self-organised behaviour
in robot swarms. In P. Lukowicz, L. Thiele, and G. Tröster,
editors, Architecture of Computing Systems, volume 4415 of
LNCS, pages 209–223. Springer, 2007.
[164] Evacuation modelling community. Evacuation modelling
portal, January 2013. http://evacmod.net.
[165] S. Morishita and T. Shiraishi. Evaluation of billboards based
on pedestrian flow in the concourse of the station. In S. E.
Yacoubi, B. Chopard, and S. Bandini, editors, Cellular Automata, volume 4173 of LNCS, pages 716–719. Springer, 2006.
[166] G. Muehl, M. Werner, M. A. Jaeger, K. Herrmann, and
H. Parzyjegla. On the definitions of self-managing and selforganizing systems. In Proceedings of the 2007 ITG-GI
Conference on Communication in Distributed Systems, pages
1–11. IEEE, 2007.
[167] R. Muhdi, A. Garrett, R. Agarwal, J. Davis, G. Dozier, and
D. Umphress. The application of evolutionary computation
in evacuation planning. In Proceedings of the 2006 IEEE
Intelligent Transportation Systems Conference, pages 600–
605. IEEE, 2006.

324

Bibliography
[168] C. Müller-Schloer. Organic computing - on the feasibility
of controlled emergence. In Proceedings of the 2004 International Conference on Hardware/Software Codesign and
System Synthesis, pages 2–5, 2004.
[169] C. Müller-Schloer and H. Schmeck. Organic Computing: A
grand challenge for mastering complex systems. it - Information Technology, 52(3):135–141, 2010.
[170] C. Müller-Schloer, H. Schmeck, and T. Ungerer. Organic
Computing - A Paradigm Shift for Complex Systems. Springer,
2011.
[171] R. Nagpal, H. Shrobe, and J. Bachrach. Organizing a global
coordinate system from local information on an ad hoc sensor
network. In F. Zhao and L. Guibas, editors, Information
Processing in Sensor Networks, volume 2634, pages 333–348.
Springer, 2003.
[172] G. Nan and M. Li. Evolutionary based approaches in wireless sensor networks: A survey. In Proceedings of the 4th
International Conference on Natural Computation, volume 5,
pages 217–222, 2008.
[173] A. Nasipuri and K. Li. A directionality based location discovery scheme for wireless sensor networks. In Proceedings
of the 1st ACM International Workshop on Wireless Sensor
Networks and Applications, pages 105–111. ACM Press, 2002.
[174] A. L. Nelson, G. J. Barlow, and L. Doitsidis. Fitness functions
in evolutionary robotics: A survey and analysis. Robotics
and Autonomous Systems, 57(4):345–370, 2009.

325

Bibliography
[175] M. Neshat, G. Sepidnam, M. Sargolzaei, and A. N. Toosi.
Artificial fish swarm algorithm: A survey of the state-of-theart, hybridization, combinatorial and indicative applications.
Artificial Intelligence Review, pages 1–33, 2012.
[176] D. Niculescu and B. Nath. Ad hoc positioning system (APS).
In Proceedings of the 2001 IEEE Global Telecommunications
Conference, pages 2926–2931. IEEE, 2001.
[177] D. Niculescu and B. Nath. DV based positioning in ad hoc
networks. Telecommunication Systems, 22(1):267–280, 2003.
[178] D. Niculescu and Badri Nath. Ad hoc positioning system
(APS) using AOA. In Proceedings of the 22nd Annual Joint
Conference of the IEEE Computer and Communications,
volume 3, pages 1734–1743. IEEE, 2003.
[179] R. J. Orr and G. D. Abowd. The smart floor: A mechanism
for natural user identification and tracking. In Extended
Abstracts on Human Factors in Computing Systems, pages
275–276. ACM Press, 2000.
[180] N. Patwari and A. O. Hero. Using proximity and quantized
RSS for sensor localization in wireless networks. In Proceedings of the 2nd ACM International Conference on Wireless
Sensor Networks and Applications, pages 20–29. ACM Press,
2003.
[181] C. S. Peskin. Mathematical aspects of heart physiology. New
York University Press, 1975.
[182] N. B. Priyantha. The Cricket Indoor Location System. Ph.D.
dissertation, Massachusetts Institute of Technology, Massachusetts, USA, 2005.

326

Bibliography
[183] N. B. Priyantha, A. K. L. Miu, H. Balakrishnan, and S. Teller.
The cricket compass for context-aware mobile applications.
In Proceedings of the 7th Annual International Conference on
Mobile Computing and Networking, pages 1–14. ACM Press,
2001.
[184] N. B. Priyantha, H. Balakrishnan, E. Demaine, and S. Teller.
Anchor-free distributed localization in sensor networks.
Techreport 892, MIT Laboratory for Computer Science, 2003.
[185] H. Prothmann, S. Tomforde, J. Branke, J. Hähner, C. MüllerSchloer, and H. Schmeck. Organic traffic control. In C. MüllerSchloer, H. Schmeck, and T. Ungerer, editors, Organic Computing - A Paradigm Shift for Complex Systems, Autonomic
Systems, pages 431–446. Springer, 2011.
[186] H. Prothmann, S. Tomforde, H. Lyda, H. Branke, H. Hähner,
C. Müller-Schloer, and H. Schmeck. Self-organised routing
for road networks. In Fernando A. Kuipers and Poul E.
Heegaard, editors, Self-Organizing Systems, volume 7166 of
LNCS, pages 48–59. Springer, 2012.
[187] F. H. Raab, E. B. Blood, T. O. Steiner, and H. R. Jones.
Magnetic position and orientation tracking system. IEEE
Transactions on Aerospace and Electronic Systems, AES-15
(5):709–718, 1979.
[188] J. Rajamäki, P. Viinikainen, J. Tuomisto, T. Sederholm, and
M. Säämänen. LaureaPOP indoor navigation service for the
visually impaired in a WLAN environment. In Proceedings
of the 6th WSEAS International Conference on Electronics,
Hardware, Wireless and Optical Communications, pages 96–
101. World Scientific and Engineering Academy and Society
(WSEAS), 2007.

327

Bibliography
[189] L. Ran, S. Helal, and S. Moore. Drishti: an integrated indoor/outdoor blind navigation system and service. In Proceedings of the 2nd Annual Conference on Pervasive Computing
and Communications, pages 23–30. IEEE, 2004.
[190] T. Rappaport. Wireless Communications: Principles and
Practice. Prentice Hall, 2 edition, 2002.
[191] G. Retscher. Pedestrian navigation systems and locationbased services. In Proceedings of the 5th IEEE International
Conference on 3G Mobile Communication Technologies, pages
359–363. IEEE, 2004.
[192] O. Ribock, U. Richter, and H. Schmeck. Using Organic
Computing to control bunching effects. In U. Brinkschulte,
T. Ungerer, C. Hochberger, and R. G. Spallek, editors, Architecture of Computing Systems, volume 4934 of LNCS, pages
232–244. Springer, 2008.
[193] U. Richter. Controlled self-organisation using learning classifier systems. Ph.D. dissertation, Karlsruhe Institute of
Technology, Karlsruhe, Germany, 2009.
[194] U. Richter and M. Mnif. Learning to control the emergent
behaviour of a multi-agent system. In F. Klügl, K. Tuyls,
and S. Sen, editors, Proceedings of the 2008 Workshop on
Adaptive Learning Agents and Multi-Agent Systems, pages
33–40, 2008.
[195] U. Richter, M. Mnif, J. Branke, C. Müller-Schloer, and
H. Schmeck. Towards a generic Observer/Controller architecture for Organic Computing. In C. Hochberger and
R. Liskowsky, editors, GI Jahrestagung, volume 93 of LNI,
pages 112–119. Gesselschaft für Informatik, 2006.

328

Bibliography
[196] B. Riemann. Über die Darstellbarkeit einer Funktion
durch eine trigonometrische Reihe. In Abhandlungen der
Königlichen Gesellschaft der Wissenschaften in Göttingen.
Weidmann, 1868.
[197] B. Ristic, S. Arulampalam, and N. Gordon. Beyond the
Kalman Filter: Particle Filters for Tracking Applications.
Artech House, 2004.
[198] R. T. Rockafellar and R. J.-B. Wets. Variational Analysis.
Springer, 2011.
[199] D. Rodriguez, S. Heuer, C. Kunze, and B. Weber. Management of mass casualty of incidents using an autonomous
sensor network. In Proceedings of the 14th International Symposium on Wireless Personal Multimedia Communications,
pages 1–5. IEEE, 2011.
[200] M. J. Russell, G. M. Switz, and K. Thompson. Olfactory
influences on the human menstrual cycle. Pharmacology,
Biochemistry, and Behavior, 13(5):737–738, 1980.
[201] S. Helal S. Willis. RFID information grid for blind navigation
and wayfinding. In Proceeding of the 9th IEEE International
Symposium on Wearable Computers, pages 34–37. IEEE,
2005.
[202] M. Saadatseresht, A. Mansourian, and M. Taleai. Evacuation planning using multiobjective evolutionary optimization
approach. European Journal of Operational Research, 198
(1):305–314, 2009.
[203] N. Salman, H. K. Maheshwari, A. H. Kemp, and M. Ghogho.
Effects of anchor placement on mean-CRB for localization.

329

Bibliography
In Proceedings of the 10th IFIP Annual Mediterranean Ad
Hoc Networking Workshop, pages 115–118. IEEE, 2011.
[204] K. Sangho, S. Shekhar, and M. Min. Contraflow transportation network reconfiguration for evacuation route planning.
IEEE Transactions on Knowledge and Data Engineering, 20
(8):1115–1129, 2008.
[205] C. Savarese, J. M. Rabaey, and K. Langendoen. Robust
positioning algorithms for distributed ad-hoc wireless sensor networks. In Proceedings of the 2002 USENIX Annual
Technical Conference, pages 317–327. USENIX Association,
2002.
[206] A. Savvides, C.-C. Han, and M. B. Strivastava. Dynamic
fine-grained localization in ad-hoc networks of sensors. In
Proceedings of the 7th Annual International Conference on
Mobile Computing and Networking, pages 166–179. ACM
Press, 2001.
[207] A. Savvides, H. Park, and M. B. Srivastava. The bits and
flops of the n-hop multilateration primitive for node localization problems. In Proceedings of the 1st ACM International
Workshop on Wireless Sensor Networks and Applications,
pages 112–121. ACM Press, 2002.
[208] A. Savvides, H. Park, and M. B. Srivastava. The n-hop multilateration primitive for node localization problems. Mobile
Network Applications, 8(4):443–451, 2003.
[209] A. Schadschneider. Cellular automaton approach to pedestrian dynamics - theory. In M. Schreckenberg and S. D.
Sharm, editors, Pedestrian and Evacuation Dynamics, pages
75–86. Springer, 2001.

330

Bibliography
[210] A. Schadschneider, W. Klingsch, H. Klüpfel, T. Kretz,
C. Rogsch, and A. Seyfried. Evacuation dynamics: Empirical
results, modeling and applications. In R. A. Meyers, editor, Encyclopedia of Complexity and Systems Science, pages
3142–3176. Springer, 2009.
[211] V. Schau, C. Erfurth, G. Eichler, S. Späthe, and W. Rossak.
Geolocated communication support in rescue management.
In Proceedings of the 8th International Conference on Information Systems for Crisis Response and Management,
2011.
[212] H. Schmeck. Organic Computing. Künstliche Intelligenz, 5
(3):68–69, 2005.
[213] H. Schmeck. Organic Computing - a new vision for distributed embedded systems. In Proceedings of the 8th IEEE
International Symposium on Object-Oriented Real-Time Distributed Computing, pages 201–203. IEEE, 2005.
[214] H. Schmeck. DFG priority program 1183 Organic Computing,
January 2013. http://www.organic-computing.de/spp.
[215] Y. Shang, W. Ruml, Y. Zhang, and M. P. J. Fromherz.
Localization from mere connectivity. In Proceedings of the 4th
ACM International Symposium on Mobile Ad Hoc Networking
and Computing, pages 201–212. ACM Press, 2003.
[216] A. Sherman, J. Rinzel, and J. Keizer. Emergence of organized bursting in clusters of pancreatic beta-cells by channel
sharing. Biophysical Journal, 54(3):411–425, 1988.
[217] F. Sivrikaya and B. Yener. Time synchronization in sensor
networks: A survey. IEEE Network, 18(4):45–50, 2004.

331

Bibliography
[218] A. Smailagic and R. Martin. Metronaut: a wearable computer with sensing and global communication capabilities. In
Proceedings of the 1st International Symposium on Wearable
Computers, pages 116–122. IEEE, 1997.
[219] A. M.-C. So and Y. Ye. On solving coverage problems in
a wireless sensor network using voronoi diagrams. In Proceedings of the 1st International Conference on Internet and
Network Economics, pages 584–593. Springer, 2005.
[220] J. Soar, A. Livingstone, and S.-Y. Wang. A case study
of an ambient living and wellness management health care
model in australia. In M. Mokhtari, I. Khalil, J. Bauchet,
D. Zhang, and C. Nugent, editors, Ambient Assistive Health
and Wellness Management in the Heart of the City, volume
5597 of LNCS, pages 48–56. Springer, 2009.
[221] W. Song, X. Xu, B.-H. Wang, and S. Ni. Simulation of
evacuation processes using a multi-grid model for pedestrian
dynamics. Physica A: Statistical Mechanics and its Applications, 363(2):492–500, 2006.
[222] V. Srovnal, Z. Machacek, R. Hercik, R. Slaby, and V. Srovnal.
Intelligent car control and recognition embedded system. In
Proceedings of the 2010 International Multiconference on
Computer Science and Information Technology, pages 831–
836. Polskie Towarzystwo Informatyczne, 2010.
[223] H. Sun, Z. Fang, T. Wang, Y. Ma, and N. Ren. CDHL: a
hybrid range-free localization algorithm in wireless sensor
networks. In Proceedings of the 5th International Conference
on Frontier of Computer Science and Technology, pages 180–
183. IEEE, 2010.

332

Bibliography
[224] M.-T. Sun, C.-W. Yi, C.-K. Yang, and T.-H. Lai. An optimal
algorithm for the minimum disc cover problem. Algorithmica,
50(1):58–71, 2007.
[225] M. Swenne and T. Bäck. Optimizing pedestrian environments
with evolutionary strategies. In U. Weidmann, U. Kirsch,
and M. Schreckenberg, editors, Pedestrian and Evacuation
Dynamics. Springer, 2012. to appear.
[226] R. Szewczyk, E. Osterweil, J. Polastre, M. Hamilton, A. Mainwaring, and D. Estrin. Habitat monitoring with sensor networks. Communications of the ACM - Wireless sensor networks, 47(6):34–40, 2004.
[227] J. Szwedko, C. Shaw, A. G. Connor, A. Labrinidis, and
P. K. Chrysan. Demonstrating an evacuation algorithm
with mobile devices using an e-scavenger hunt game. In
Proceedings of the 8th ACM International Workshop on Data
Engineering for Wireless and Mobile Access, pages 49–52,
2009.
[228] A. S. Tanenbaum. Computer networks. Prentice Hall, 2002.
[229] Y. Taniguchi, N. Wakamiya, and M. Murata. A communication mechanism using traveling wave phenomena for wireless
sensor networks. In Proceedings of the 2007 IEEE International Symposium on a World of Wireless, Mobile and
Multimedia Networks, pages 1 –6. IEEE, 2007.
[230] B. Tatham and T. Kunz. Anchor node placement for localization in wireless sensor networks. In Proceedings of the 7th
International Conference on Wireless and Mobile Computing, Networking and Communications, pages 180–187. IEEE,
2011.

333

Bibliography
[231] V. Torre. A theory of synchronization of heart pace-maker
cells. Journal of Theoretical Biology, 61(1):55–71, 1976.
[232] V. Tsetsos, C. Anagnostopoulos, P. Kikiras, and S. Hadjiefthymiades. Semantically enriched navigation for indoor
environments. International Journal of Web Grid Services, 2
(4):453–478, 2006.
[233] A. Tyrrell, G. Auer, and C. Bettstetter. Fireflies as role
models for synchronization in ad hoc networks. In Proceedings
of the 1st International Conference on Bio-inspired Models of
Network, Information and Computing Systems. ACM Press,
2006.
[234] A. Varas, M. D. Cornejo, D. Mainemer, B. Toledo, J. Rogan,
V. Muñoz, and J. A. Valdivia. Cellular automaton model
for evacuation process with obstacles. Physica A: Statistical
Mechanics and its Applications, 382(2):631–642, 2007.
[235] VDE/ITG/GI. Organic Computing.
VDE/ITG/GI, 2003.

Positionspapier,

[236] F. L. Villafuerte, K. Terfloth, and J. Schiller. Using network
density as a new parameter to estimate distance. In Proceedings of the 7th International Conference on Networking,
pages 30–35. IEEE, 2008.
[237] A. U. Kemloh Wagoum, A. Seyfried, and S. Holl. Modelling
dynamic route choice of pedestrians to assess the criticality
of building evacuation. Techreport arXiv e-print 1103.4080,
Forschungszentrum Jülich GmbH and Bergische Universität
Wuppertal, 2011.

334

Bibliography
[238] N. Wakamiya and M. Murata. Synchronization-based data
gathering scheme for sensor networks. IEICE Transactions
on Communications, 88(3):873–881, 2005.
[239] N. Wakamiya, S. Kashihara, and M. Murata.
A
synchronization-based data gathering scheme in unstable
radio environments. In Proceedings of the 4th International
Conference on Networked Sensing Systems, pages 10–18.
IEEE, 2007.
[240] T. J. Walker. Acoustic synchrony: Two mechanisms in the
snowy tree cricket. Science, 166(3907):891–894, 1969.
[241] T. Walther and R. P. Würtz. Learning to look at humans.
In C. Müller-Schloer, H. Schmeck, and T. Ungerer, editors,
Organic Computing - A Paradigm Shift for Complex Systems,
Autonomic Systems, pages 309–322. Springer, 2011.
[242] C.-L. Wang, Y.-W. Hong, and Y.-S. Dai. A decentralized
positioning method for wireless sensor networks based on
weighted interpolation. In Proceedings of 2007 IEEE International Conference on Communications, pages 3167–3172.
IEEE, 2007.
[243] Y.-H. Wang, C.-P. Hsu, C.-M. Lee, K.-F. Huang, and T.-W.
Chang. A location mechanism with mobile reference nodes
in wireless sensor networks. In Proceedings of the 21st International Conference on Advanced Information Networking
and Applications Workshops, pages 48–55. IEEE, 2007.
[244] R. Want, A. Hopper, V. Falcão, and J. Gibbons. The active
badge location system. Transactions on Information Systems,
10(1):91–102, 1992.

335

Bibliography
[245] D. J. Watts and S. H. Strogatz. Collective dynamics of
small-world networks. Nature, 393(6684):440–442, 1998.
[246] K. Weicker. Evolutionäre Algorithmen. Vieweg und Teubner,
2nd edition, 2007.
[247] G. Werner-Allen, G. Tewari, A. Patel, M. Welsh, and R. Nagpal. Firefly-inspired sensor network synchronicity with realistic radio effects. In Proceedings of the 3rd International
Conference on Embedded Networked Sensor Systems, pages
142–153. ACM Press, 2005.
[248] G. Werner-Allen, K. Lorincz, J. Johnson, J. Lees, and
M. Welsh. Fidelity and yield in a volcano monitoring sensor
network. In Proceedings of the 7th Symposium on Operating Systems Design and Implementation, pages 381–396.
USENIX Association, 2006.
[249] S. Y. Wong, J. G. Lim, S. V. Rao, and W. K. G. Seah.
Density-aware hop-count localization (DHL) in wireless sensor networks with variable density. In Proceedings of 2005
IEEE Wireless Communications and Networking Conference,
pages 1848–1853. IEEE, 2005.
[250] H. Wu, A. Marshall, and W. Yu. Path planning and following algorithms in an indoor navigation model for visually
impaired. In Proceedings of the 2nd International Conference
on Internet Monitoring and Protection, pages 38–45. IEEE,
2007.
[251] M. Xiong, Y. Chen, H. Wang, and M. Hu. An agent-based
model for simulating human-like crowd in dense places. In

336

Bibliography
Z. Li, X. Li, Y. Liu, and Z. Cai, editors, Computational Intelligence and Intelligent Systems, Communications in Computer
and Information Science, pages 8–19. Springer, 2012.
[252] L. Z. Yang, D. L. Zhao, J. Li, and T. Y. Fang. Simulation
of the kin behavior in building occupant evacuation based
on cellular automaton. Building and Environment, 40(3):
411–415, 2005.
[253] Z. You, M. Q.-H. Meng, H. Liang, S. Li, Y. Li, W. Chen,
Y. Zhou, S. Miao, K. Jiang, and Q. Guo. A localization
algorithm nin wireless sensor networks using a mobile beacon
node. In Proceedings of 2007 International Conference on
Information Acquisition, pages 420–426, 2007.
[254] B. Zhang and F. Yu. An energy efficient localization algorithm for wireless sensor networks using a mobile anchor
node. In Proceedings of the 2008 International Conference
on Information and Automation, pages 215–219. IEEE, 2008.
[255] B. Zhang, F. Yu, and Z. Zhang. A high energy efficient
localization algorithm for wireless sensor networks using
directional antenna. In Proceedings of the 11th IEEE International Conference on High Performance Computing and
Communications, pages 230–236. IEEE, 2009.
[256] X. Zheng, T. Zhong, and M. Liu. Modeling crowd evacuation
of a building based on seven methodological approaches.
Building and Environment, 44(3):437–445, 2009.
[257] ZigBee Alliance.
ZigBee
http://www.zigbee.org.

technology,

July

2013.

337

Bibliography
[258] X. Zong, S. Xiong, Z. Fang, and Q. Li. Multi-objective
optimization for massive pedestrian evacuation using ant
colony algorithm. In Y. Tan, Y. Shi, and K. C. Tan, editors,
Advances in Swarm Intelligence, volume 6145 of LNCS, pages
636–642. Springer, 2010.
[259] M. M. Zonoozi and P. Dassanayake. User mobility modeling
and characterization of mobility patterns. IEEE Journal on
Selected Areas in Communications, 15(7):1239–1251, 1997.

338

The rapidly growing world population and increasingly dense settlements demand
ev er-larger and more complex buildings from today‘s engineers. In comparison to
this technological progress, a building‘s equipment for emergency evacuation has
been hardly dev eloped further. Today‘s evacuation support facilities are mainly
limited to stationary exit signage and emergency maps, which display recommended
escape routes. Emergency maps can be easily ov erlooked and are often perceiv ed
as confusing and unclear, especially when in panic. Another serious problem with
contemporary evacuation support equipment of buildings is its inability to adapt the
recommended escape routes to the current situation which occurs during the evacuation process, such as fire outbreaks or congested evacuation routes. Howev er, the
increasingly wide spread of mobile dev ices with the ability for wireless communication, such as smart phones, tablet PCs, or general dev ices for personal digital
assistance, opens up an opportunity to improv e the support of evacuees for a fast
emergency evacuation of buildings. In case of emergency, a mobile dev ice can alert
its user v ia ringing or v ibrating and display an indiv idual escape route on its screen.
By pointing out the respectiv e directions, the dev ice can nav igate its user to a safe
exit. This work presents a concept for a building evacuation system based on mobile
dev ices. Furthermore, various algorithms for route planning with mobile dev ices and
for indoor localization of mobile dev ices are addressed.

ISBN 978-3-7315-0207-4

9 783731 502074