paper_id
stringlengths 9
16
| version
stringclasses 26
values | yymm
stringclasses 311
values | created
timestamp[s] | title
stringlengths 6
335
| secondary_subfield
sequencelengths 1
8
| abstract
stringlengths 25
3.93k
| primary_subfield
stringclasses 124
values | field
stringclasses 20
values | fulltext
stringlengths 0
2.84M
|
|---|---|---|---|---|---|---|---|---|---|
0909.3423 | 2 | 0909 | 2009-10-06T16:36:39 | Digital Ecosystems | [
"cs.MA",
"cs.NE"
] | We view Digital Ecosystems to be the digital counterparts of biological ecosystems, which are considered to be robust, self-organising and scalable architectures that can automatically solve complex, dynamic problems. So, this work is concerned with the creation, investigation, and optimisation of Digital Ecosystems, exploiting the self-organising properties of biological ecosystems. First, we created the Digital Ecosystem, a novel optimisation technique inspired by biological ecosystems, where the optimisation works at two levels: a first optimisation, migration of agents which are distributed in a decentralised peer-to-peer network, operating continuously in time; this process feeds a second optimisation based on evolutionary computing that operates locally on single peers and is aimed at finding solutions to satisfy locally relevant constraints. We then investigated its self-organising aspects, starting with an extension to the definition of Physical Complexity to include evolving agent populations. Next, we established stability of evolving agent populations over time, by extending the Chli-DeWilde definition of agent stability to include evolutionary dynamics. Further, we evaluated the diversity of the software agents within evolving agent populations. To conclude, we considered alternative augmentations to optimise and accelerate our Digital Ecosystem, by studying the accelerating effect of a clustering catalyst on the evolutionary dynamics. We also studied the optimising effect of targeted migration on the ecological dynamics, through the indirect and emergent optimisation of the agent migration patterns. Overall, we have advanced the understanding of creating Digital Ecosystems, the self-organisation that occurs within them, and the optimisation of their Ecosystem-Oriented Architecture. | cs.MA | cs | Imperial College London
Department of Electrical and Electronic Engineering
Digital Ecosystems
Gerard Briscoe
Submitted in part fulfilment of the requirements for the degree of
Doctor of Philosophy
2009
1
2
Abstract
We view Digital Ecosystems to be the digital counterparts of biological ecosystems, which
are considered to be robust, self-organising and scalable architectures that can automatically
solve complex, dynamic problems. So, this work is concerned with the creation, investigation,
and optimisation of Digital Ecosystems, exploiting the self-organising properties of biological
ecosystems. First, we created the Digital Ecosystem, a novel optimisation technique inspired
by biological ecosystems, where the optimisation works at two levels: a first optimisation,
migration of agents which are distributed in a decentralised peer-to-peer network, operating
continuously in time; this process feeds a second optimisation based on evolutionary computing
that operates locally on single peers and is aimed at finding solutions to satisfy locally relevant
constraints. We then investigated its self-organising aspects, starting with an extension
to the definition of Physical Complexity to include the evolving agent populations of our
Digital Ecosystem. Next, we established stability of evolving agent populations over time,
by extending the Chli-DeWilde definition of agent stability to include evolutionary dynamics.
Further, we evaluated the diversity of the software agents within evolving agent populations,
relative to the environment provided by the user base. To conclude, we considered alternative
augmentations to optimise and accelerate our Digital Ecosystem, by studying the accelerating
effect of a clustering catalyst on the evolutionary dynamics of our Digital Ecosystem, through
the direct acceleration of the evolutionary processes. We also studied the optimising effect of
targeted migration on the ecological dynamics of our Digital Ecosystem, through the indirect
and emergent optimisation of the agent migration patterns. Overall, we have advanced the
understanding of creating Digital Ecosystems, the self-organisation that occurs within them,
and the optimisation of their Ecosystem-Oriented Architecture.
3
.
4
Contents
Abstract
Acknowledgements
List of Figures
1 Introduction
1.1 Motivation and Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Dissertation Outline
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Creation of Digital Ecosystems
3
9
11
15
15
17
18
19
21
2.1 Background Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
2.1.1 Existing Digital Ecosystems . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.1.2 Nature-Inspired Computing . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.3 Biology of Digital Ecosystems . . . . . . . . . . . . . . . . . . . . . . . .
2.1.4 Computing of Digital Ecosystems . . . . . . . . . . . . . . . . . . . . . .
2.1.5 Digital Business Ecosystems . . . . . . . . . . . . . . . . . . . . . . . . .
23
25
35
45
2.2 Ecosystem-Oriented Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.2.1 Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.2.2 Habitats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.3 Populations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.4 The Digital Ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Simulation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 Agents: Semantic Descriptions . . . . . . . . . . . . . . . . . . . . . . . .
2.3.2 User Base . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.3 Populations: Evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.4
Semantic Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
55
61
67
67
68
69
70
5
6
CONTENTS
2.3.5 Evolutionary Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.3.6 Ecological Succession . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
72
2.3.7
Species Abundance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.3.8
Species-Area Relationship . . . . . . . . . . . . . . . . . . . . . . . . . . 76
2.4 Summary and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
77
3 Investigation of Digital Ecosystems
3.1 Background Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1
Self-Organisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.2 Definitions of Self-Organisation . . . . . . . . . . . . . . . . . . . . . . .
83
84
84
86
3.2 Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.2.1 Physical Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.2 Extending to Agent Populations . . . . . . . . . . . . . . . . . . . . . . .
91
99
3.2.3
Simulation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3.2.4
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
3.3 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
3.3.1 Chli-DeWilde Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3.3.2 Extensions for Evolving Populations
. . . . . . . . . . . . . . . . . . . . 120
3.3.3
Simulation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
3.3.4
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
3.4 Diversity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
3.4.1 Evolving Agent Populations . . . . . . . . . . . . . . . . . . . . . . . . . 130
3.4.2
Simulation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
3.4.3
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
3.5 Summary and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4 Optimisation of Digital Ecosystems
141
4.1 Background Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
4.2 Alternative Augmentations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
4.2.1 Clustering Catalyst . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
4.2.2 Replacement Aggregator . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
4.2.3 Agent-Pool Aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.2.4 Targeted Migration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
4.2.5 Choice of Augmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
CONTENTS
7
4.3 Clustering Catalyst . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4.3.1 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4.3.2 Physical Complexity Clustering . . . . . . . . . . . . . . . . . . . . . . . 155
4.4 Targeted Migration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
4.4.1
Similarity Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
4.4.2 Neural Networks
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
4.4.3
Support Vector Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
4.5 Simulation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
4.5.1 Clustering Catalyst . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
4.5.2 Targeted Migration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
4.6 Summary and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
5 Conclusions
181
5.1 Achievements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
5.2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
5.2.1 Ecosystems Conceptualisation . . . . . . . . . . . . . . . . . . . . . . . . 183
5.2.2
Simulation Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
5.2.3 Digital Business Ecosystems . . . . . . . . . . . . . . . . . . . . . . . . . 185
5.3 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
Bibliography
187
8
Acknowledgements
I would like to express my most sincere thanks to:
• My supervisor, Philippe De Wilde, to whom I am deeply indebted for the guidance and
support he has provided throughout.
• My examiners, Peter McBurney and Philip Treleaven, for their time and effort in the
reading of this thesis and the viva.
• My mentor, Paolo Dini, for his constant support, and originally encouraging to me pursue
this research and PhD.
• My friend and colleague, Evangelia Berdou, for her support and guidance in the writing
of this document.
• My colleagues, Sotiris Moschoyiannis and Alastair Munro, for kindly agreeing to review
this document.
• Several other researchers and friends, most notably Suzanne Sadedin, Maria Chli and
Frauke Zeller from whom I have learnt much about conducting research.
• My mother, Constance Briscoe, without whom I could never have finished this work.
• The European Union Framework VI project, Digital Business Ecosystems, Contract No
507953, and the Network of Excellence, Open Philosophies for Associative Autopoietic
Digital Ecosystems, Contract No 034824, for supporting this work.
• The Open Mac Grid of MacResearch.org, and the Digital Ecosystems Lab of the
Department of Media and Communications at the London School of Economics and
Political Science, for access to their grids for the running of simulations.
9
.
10
List of Figures
2.1 Biomimicry Design Spiral
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Ecosystem Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Fitness Landscape
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Abstract View of An Ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
27
29
30
2.5 Evolving Population of Digital Organisms
. . . . . . . . . . . . . . . . . . . . . 32
2.6 Ecosystems as Complex Adaptive Systems . . . . . . . . . . . . . . . . . . . . .
2.7 Mobile Agent System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
36
2.8 Service-Oriented Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.9
Island-Model of Distributed Evolutionary Computing . . . . . . . . . . . . . . .
2.10 Business Ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.11 Agent of the Digital Ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.12 Structure of Aggregated Agents . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.13 Habitat Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.14 User Request to the Digital Ecosystem . . . . . . . . . . . . . . . . . . . . . . .
2.15 Digital Ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.16 Digital Business Ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.17 Habitat Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.18 Distributed Evolution in the Digital Ecosystem . . . . . . . . . . . . . . . . . .
2.19 Agent Life-Cycle
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.20 Agent Semantic Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.21 User Request
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.22 Semantic Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.23 Graph of Fitness in the Evolutionary Process . . . . . . . . . . . . . . . . . . . .
2.24 Ecological Succession . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.25 Graph of Succession in the Digital Ecosystem . . . . . . . . . . . . . . . . . . .
2.26 Graph of Relative Abundance in the Digital Ecosystem . . . . . . . . . . . . . .
2.27 Graph of Species-Area in the Digital Ecosystem . . . . . . . . . . . . . . . . . .
2.28 Hypothetical Abstract Ecosystem Definition . . . . . . . . . . . . . . . . . . . .
45
46
49
51
52
55
62
64
64
65
67
68
68
70
71
72
74
76
77
80
11
12
LIST OF FIGURES
3.1 Visualisation of Self-Organisation in Evolving Agent Populations . . . . . . . . .
86
3.2 DNA Samples from a Population . . . . . . . . . . . . . . . . . . . . . . . . . .
96
3.3 Alternative DNA Samples of the Population . . . . . . . . . . . . . . . . . . . . 101
3.4 Abstract Visualisation for Populations of Variable Length Sequences . . . . . . . 103
3.5
3D Fitness Landscape with a Global Optimum . . . . . . . . . . . . . . . . . . . 104
3.6
3D Fitness Landscape with No Optimum . . . . . . . . . . . . . . . . . . . . . . 105
3.7
3D Fitness Landscape with Global Optima . . . . . . . . . . . . . . . . . . . . . 106
3.8 Population with Hidden Clusters
. . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.9 Population with Clusters Visible . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.10 Agent Atomicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
3.11 Population Constructed from a Non-Atomic Alphabet . . . . . . . . . . . . . . . 109
3.12 Original Physical Complexity Graphs . . . . . . . . . . . . . . . . . . . . . . . . 110
3.13 Graph of Physical Complexity and Maximum Fitness . . . . . . . . . . . . . . . 111
3.14 Visualisation of Evolving Agent Populations . . . . . . . . . . . . . . . . . . . . 112
3.15 Graph of Population Efficiency over the Generations . . . . . . . . . . . . . . . . 113
3.16 Graph of the Clustering Coefficient over the Generations . . . . . . . . . . . . . 114
3.17 Visualisation of Clusters in an Evolving Agent Population . . . . . . . . . . . . 114
3.18 State-Space of an Evolving Agent Population . . . . . . . . . . . . . . . . . . . . 122
3.19 Graph of the Probabilities of the Macro-States . . . . . . . . . . . . . . . . . . . 125
3.20 Visualisation of an Evolving Agent Population . . . . . . . . . . . . . . . . . . . 126
3.21 Graph of Stability with Different Mutation and Crossover Rates . . . . . . . . . 127
3.22 Graph of Uniformly Distributed Agent-Sequence Length Frequencies . . . . . . . 132
3.23 Graph of Gaussian Distributed Agent-Sequence Length Frequencies . . . . . . . 132
3.24 Graph of Power Distributed Agent-Sequence Length Frequencies . . . . . . . . . 133
3.25 Graph of Uniformly Distributed Agent Attribute Frequencies . . . . . . . . . . . 134
3.26 Graph of Gaussian Distributed Agent Attribute Frequencies
. . . . . . . . . . . 135
3.27 Graph of Power Distributed Agent Attribute Frequencies . . . . . . . . . . . . . 135
LIST OF FIGURES
13
4.1 Clustering Catalyst . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.2 Replacement Aggregator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
4.3 Agent-Pool Aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.4 Targeted Migration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.5 Effect of The Proposed Augmentations . . . . . . . . . . . . . . . . . . . . . . . 149
4.6 Pseudo-Code for Physical Complexity Clustering . . . . . . . . . . . . . . . . . . 155
4.7 Agent Life-Cycle With Targeted Migration . . . . . . . . . . . . . . . . . . . . . 157
4.8 Neural Network for the Similarity Recognition Component . . . . . . . . . . . . 164
4.9 Support Vector Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
4.10 Graph of the Hierarchical Clustering Based Clustering Catalyst
. . . . . . . . . 168
4.11 Graph of the Physical Complexity Based Clustering Catalyst . . . . . . . . . . . 169
4.12 Graph of Typical Runs for the Clustering Catalysts . . . . . . . . . . . . . . . . 170
4.13 Graph of the Targeted Migration Controls and the Digital Ecosystem . . . . . . 172
4.14 Graph of Typical Runs for the Targeted Migration Controls
. . . . . . . . . . . 172
4.15 Graph of Neural Networks Based Targeted Migration . . . . . . . . . . . . . . . 173
4.16 Graph of Support Vector Machine Based Targeted Migration . . . . . . . . . . . 174
4.17 Graph of Typical Runs for the Digital Ecosystem and Targeted Migration . . . . 175
4.18 Graph of Frequencies for the Targeted Migration . . . . . . . . . . . . . . . . . . 176
14
Chapter 1
Introduction
1.1 Motivation and Objectives
Is mimicking ecosystems the future of information systems ?
A key challenge in modern computing is to develop systems that address complex, dynamic
problems in a scalable and efficient way, because the increasing complexity of software makes
designing and maintaining efficient and flexible systems a growing challenge [209, 299, 193].
What with the ever expanding number of services being offered online from Application
Programming Interfaces (APIs) being made public, there is an ever growing number of
computational units available to be combined in the creation of applications. However, this
is currently a task done manually by programmers, and it has been argued [184] that current
software development techniques have hit a complexity wall, which can only be overcome by
automating the search for new algorithms. There are several existing efforts aimed at achieving
this automated service composition [203, 226, 207, 255], the most prevalent of which is Service-
Oriented Architectures and its associated standards and technologies [66, 320].
Alternatively, nature has been in the research business for 3.8 billion years and in that time has
accumulated close to 30 million well-adjusted solutions to a plethora of design challenges that
humankind struggles to address with mixed results [33]. Biomimicry is a discipline that seeks
solutions by emulating nature’s designs and processes, and there is considerable opportunity
15
16
, Chapter 1.
Introduction
to learn elegant solutions for human-made problems [33]. Biological ecosystems are thought
to be robust, scalable architectures that can automatically solve complex, dynamic problems,
possessing several properties that may be useful in automated systems. These properties include
self-organisation, self-management, scalability, the ability to provide complex solutions, and the
automated composition of these complex solutions [173].
Therefore, an approach to the aforementioned challenge would be to develop Digital Ecosystems,
artificial systems that aim to harness the dynamics that underlie the complex and diverse
adaptations of
living organisms in biological ecosystems. While evolution may be well
understood in computer science under the auspices of evolutionary computing [90], ecological
models are not. The possible connections between Digital Ecosystems and their biological
counterparts are yet to be closely examined, so potential exists to create an Ecosystem-Oriented
Architecture with the essential elements of biological ecosystems, where the word ecosystem is
more than just a metaphor. We propose that an ecosystem inspired approach, would be more
effective at greater scales than traditionally inspired approaches, because it would be built upon
the scalable and self-organising properties of biological ecosystems [173]. However, ecological
succession, the formation of a mature ecosystem from the predictable and orderly changes in
the composition and structure of an ecological community [29], is a slow process. So, for our
Digital Ecosystems it will be desirable to accelerate and optimise the equivalent process, which
may be possible through the application of augmentations that interact with the ecosystem
dynamics. Therefore, the primary objectives are as follows:
• Determine the structure of an Ecosystem-Oriented Architecture and so create Digital
Ecosystems, which are the digital counterpart of biological ecosystems, and so have
analogous properties of self-organisation, scalability and sustainability.
• Develop an understanding of the self-organising behaviour within a Digital Ecosystem,
learning where and how it occurs, what forms it can take, and how it can be quantified.
• Investigate if we can accelerate or optimise the evolutionary and ecological self-organising
dynamics of Digital Ecosystems, exploring how alternative augmentations interact with
the ecosystem dynamics.
1.2. Contributions ,
1.2 Contributions
17
Substantial parts of our efforts are original contributions in the area of Biologically-Inspired
Computing [99] and the emerging field of Digital Ecosystems, with our major research
contributions being as follows:
• We have determined the fundamentals for a new class of system, Digital Ecosystems,
created through combining understanding from theoretical ecology, evolutionary the-
ory, Multi-Agent Systems, distributed evolutionary computing, and Service-Oriented
Architectures.
• We have investigated where and how self-organisation occurs in Digital Ecosystems, what
forms it can take and how it can be quantified, including the self-organised complexity,
stability, and diversity of the evolving agent populations within.
• We have extended the statistical physics based definition of Physical Complexity, to
include evolving agent populations. This required extending definitions for populations
of variable length sequences, creating a measure for the efficiency of information storage,
and an understanding of clustering within populations to support the non-atomicity of
agents.
• We have extended the Chli-DeWilde definition of agent stability to include the evolu-
tionary dynamics of evolving agent populations. We then built upon this to construct an
entropy-based definition for the degree of instability, which was used to study the stability
of evolving agent populations under varying conditions.
• We have developed an understanding and definition for the self-organised diversity, finding
no existing definition suitable because of the unique hybrid nature of Digital Ecosystems.
We therefore considered the global distribution of the agents in the populations relative
to the varying requirements of the user base.
• We have investigated alternative augmentations to optimise and accelerate our Digital
Ecosystems, studying the accelerating effect of a clustering catalyst on the evolutionary
dynamics, and the optimising effect of targeted migration on the ecological dynamics.
18
1.3 Publications
The following publications support the material presented in this thesis:
[•] Digital Ecosystems: Stability of evolving agent populations. G Briscoe and P De
Wilde. Submitted to Evolutionary Computation.
[•] Digital Ecosystems: Self-organisation of evolving agent populations. G Briscoe
and P De Wilde. Submitted to Applied Soft Computing.
[•] Digital Ecosystems: Optimisation by a distributed intelligence. G Briscoe and P De
Wilde. In Digital Ecosystems and Technologies Conference. IEEE Press, 2008. Awarded
Best Track Paper.
[•] Digital Business Ecosystems: Natural science paradigms. G Briscoe and S Sadedin.
In F Nachira, A Nicolai, P Dini, M Le Louarn, and L Rivera Le´on, editors, Digital Business
Ecosystems, pages 48–55. European Commission, 2007.
[•] Biology of applied digital ecosystems. G Briscoe, S Sadedin, and G Paperin. In Digital
Ecosystems and Technologies Conference, pages 458–463. IEEE Press, 2007.
[•] Digital Ecosystems: Evolving service-oriented architectures. G Briscoe and P De
Wilde.
In Conference on Bio Inspired Models of Network, Information and Computing
Systems. IEEE Press, 2006.
[•] Creating a Digital Ecosystem: Service-oriented architectures with distributed
evolutionary computing (BOF-0759). G Briscoe, M Chli, and M Vidal. In JavaOne
Conference. Sun Microsystems, 2006.
[•] Pervasive service architecture for a digital business ecosystem. T Heistracher,
T Kurz, C Masuch, P Ferronato, M Vidal, A Corallo, P Dini, and G Briscoe. In C Canal,
J Murillo, and P Poizat, editors, International Workshop on Coordination and Adaptation
Techniques for Software Entities, pages 71–80. Universidad de Extremadura, 2004.
1.4. Dissertation Outline ,
19
1.4 Dissertation Outline
In Chapter 2, we explain the hybrid model created to provide the digital counterpart of a
biological ecosystem. We start with the relevant theory from the domain of theoretical biology,
including the fields of evolutionary and ecological theory, and from the domain of computer
science,
including the fields of Multi-Agent Systems, evolutionary computing and Service-
Oriented Architectures. The Digital Ecosystem is then measured experimentally through
simulations, with measures originating from theoretical ecology, to evaluate its likeness to a
biological ecosystem. This included its responsiveness to requests for applications from the
user base, as a measure of the ecological succession (ecosystem maturity).
Chapter 3 investigates the self-organising aspects of Digital Ecosystems. We start with
the complexity of the evolving agent populations within, by extending the statistical physics
based definition of Physical Complexity to support variable length populations of software
agents. Next, we investigate the stability of the evolving agent populations, by extending
the Chli-DeWilde definition of agent stability to include the evolutionary dynamics of Digital
Ecosystems. Finally, we study the diversity of the agents within the evolving agent populations
of the Digital Ecosystem, for optimality relative to the environment provided by the user base.
In Chapter 4, we start by considering alternative augmentations to optimise and accelerate
Digital Ecosystems. We then further investigate the most promising, the clustering catalyst and
targeted migration: the accelerating effect of a clustering catalyst on the evolutionary dynamics
of our Digital Ecosystem, through the direct acceleration of the evolutionary processes; and
the optimising effect of targeted migration on the ecological dynamics of our Digital Ecosystem,
through the indirect and emergent optimisation of the agent migration patterns.
Chapter 5 provides a summary of the conclusions, and suggests possible future research into
Digital Ecosystems. We also report on the status of the reference implementation for Digital
Ecosystems, and the dedicated simulation framework created for its future study. After this
the Bibliography follows.
20
Chapter 2
Creation of Digital Ecosystems
In this chapter we create Digital Ecosystems, starting with a discussion of the relevant literature,
including Nature Inspired Computing as a framework in which to understand this work,
and the process of biomimicry to be used in mimicking the necessary biological processes to
create Digital Ecosystems. We then consider the relevant theoretical ecology in creating the
digital counterpart of a biological ecosystem, including the topological structure of ecosystems,
and evolutionary processes within distributed environments. This leads to a discussion of
the relevant fields from computer science for the creation of Digital Ecosystems, including
evolutionary computing, Multi-Agent Systems, and Service-Oriented Architectures. We then
define Ecosystem-Oriented Architectures for the creation of Digital Ecosystems, imbibed with
the properties of self-organisation, scalability and sustainability from biological ecosystems,
including a novel form of distributed evolutionary computing. This will include a discussion of
the compromises resulting from the hybrid model created, such as the network topology. We
then performed simulations to compare the likeness of our Digital Ecosystem with biological
ecosystems, starting with ecological succession (development), measured by its responsiveness to
requests for applications from the user base, and followed by the measures of species abundance
and the species-area relationship, which are commonly applied to biological ecosystems. Finally,
we conclude with a summary and discussion of the achievements, including the experimental
results.
21
22
, Chapter 2. Creation of Digital Ecosystems
2.1 Background Theory
In this section we discuss the relevant background theory, and because of the interdisciplinary
nature of our research it will cover several fields across different domains. We start with an
introduction to Nature Inspired Computing, followed by the relevant theoretical biology and
computer science. With the theoretical biology, we will consider how properties of biological
ecosystems influence functions that are relevant to developing Digital Ecosystems to solve
practical problems. This leads us to suggest ways in which concepts from ecology can be used
in biologically inspired techniques to create Digital Ecosystems.
2.1.1 Existing Digital Ecosystems
Our focus is in creating the digital counterpart of biological ecosystems. However, the term
digital ecosystem has described a variety of concepts, which we shall now review. Sometimes
referring to the existing networking infrastructure of the Internet [79, 27, 94, 337], while several
companies offer a digital ecosystem service, which involves enabling customers to use existing
e-business solutions [32, 160, 315]. The term is also being increasingly linked to the future
developments of Information and Communications Technology (ICT) adoption for e-business,
to support business ecosystems [214]. However, perhaps the most frequent references to digital
ecosystems arise in Artificial Life research, where they are created primarily to investigate
aspects of biological and other complex systems [295, 114, 55].
The extent to which these disparate systems resemble biological ecosystems varies, and
frequently the word ecosystem is merely used for branding purposes without any inherent
ecological properties. We consider Digital Ecosystems to be software systems that exploit
the properties of biological ecosystems, and suggest that several key features of biological
ecosystems have not been fully explored in existing digital ecosystems. So, we will now discuss
how mimicking these features can create Digital Ecosystems, which are robust, scalable, and
self-organising.
2.1. Background Theory ,
23
2.1.2 Nature-Inspired Computing
Biomimicry (bios, meaning life, and mimesis, meaning to imitate) is the science that studies
nature,
its models, systems, processes, and elements, and then imitates or takes creative
inspiration from them for the study and design of engineering systems and modern technology
[33]. This concept is far from new, with humans having long been inspired by the animals and
plants of the natural world; Leonardo Da Vinci himself once said, Those who are inspired by
a model other than Nature, a mistress above all masters, are labouring in vain [40]. Albeit
overstating the point, it reminds us that the transfer of technology between life-forms and
synthetic constructs is desirable because evolutionary pressures typically force living organisms
to become highly optimised and efficient. A classical example is the development of dirt and
water repellent paint from the observation that the surface of the lotus flower plant is practically
non-sticky for anything, commonly known as the lotus effect [25]. However, biomimicry, when
done well, is not slavish imitation;
it is inspiration using the principles which nature has
demonstrated to be successful design strategies. For example, in the early days of mechanised
flight the best designs were not the ornithopters, which most completely imitated birds, but
the fixed-wing craft that used the principle of aerofoil cross-section in their wings [10].
Biomimicry in computer science is called Nature Inspired Computing (NIC) or Natural
Computation, and the benefits of natural computation technologies often mimic those found in
real natural systems, and include flexibility, adaptability, robustness, and decentralised control
[73]. The increasing demands upon current computer systems, along with technological changes,
create a need for more flexible and adaptable systems. The desire to achieve this has led many
computing researchers to look to natural systems for inspiration in the design of computer
software and hardware, as natural systems provide many examples of the type of versatile
system required [73]. Their sources of inspiration come from many aspects of natural systems;
evolution, ecology, development, cell and molecular phenomena, behaviour, cognition, and other
areas [195]. The use of nature inspired techniques often results in the design of novel computing
systems with applicability in many different areas [195]. NIC itself can be divided into three
main branches [73]:
24
, Chapter 2. Creation of Digital Ecosystems
• Biologically-Inspired Computing (BIC): This makes use of nature as inspiration for the
development of problem solving techniques. The main idea of this branch is to develop
computational tools (algorithms) by taking inspiration from nature for the solution of
complex problems.
• The simulation and emulation of nature by computational means: This is basically
a synthetic process aimed at creating patterns,
forms, behaviours, and organisms
that resemble life-as-we-know-it.
Its products can be used to mimic various natural
phenomena, thus increasing our understanding of nature and insights about computer
models.
• Computing with natural materials: This corresponds to the use of natural materials to
perform computation, to substitute or supplement the current silicon-based computers.
All branches share the common characteristic of human-designed computing inspired by nature,
the metaphorical use of concepts, principles, and mechanisms underlying natural systems. Thus,
evolutionary algorithms use the concepts of mutation, recombination, and natural selection
from biology; neural networks are inspired by the highly interconnected neural structures in
the brain and the nervous system; molecular computing is based on paradigms from molecular
biology; and quantum computing based on quantum physics exploits quantum parallelism [73].
There are however, important methodological differences between various sub-areas of natural
computing. For example, evolutionary algorithms and algorithms based on neural networks are
presently implemented on conventional computers. However, molecular computing also aims
at alternatives for silicon hardware by implementing algorithms in biological hardware, using
DNA molecules and enzymes. Also, quantum computing aims at non-traditional hardware that
can make use of quantum effects [73].
We are concerned with BIC, which relies heavily on the fields of biology, computer science, and
mathematics. Briefly put, it is the study of nature to improve the usage of computers [99],
and should not to be confused with computational biology [326], which is an interdisciplinary
field that applies the techniques of computer science, applied mathematics, and statistics to
address problems inspired by biology. BIC has produced Neural Networks, swarm intelligence
and evolutionary computing [99]. Introducing BIC, one comes quickly to its applications, partly
2.1. Background Theory ,
25
Figure 2.1: Biomimicry Design Spiral (modified from [130]): The process of biomimicry starts
with identifying some behaviour from a biological system, which would appear to be useful.
Followed by observation to understand the mechanisms or principles by which it operates, and
therefore allowing for an abstract understanding of the behaviour. This can then be mimicked
in a non-biological system and its performance and effectiveness evaluated [130].
because this is the essence of the approach, and partly because biomimicry as a process tends
to be un-formalised and ad hoc [130]. It generally involves an engineer or scientist observing or
being aware of an area of biological study, which seems applicable to a technology or research
problem they are currently tackling, or which inspires the creation of a new technology [73].
However, there are some common steps in this process, which starts with identifying some
behaviour from a biological system, which would appear to be useful. Followed by observation
to understand the mechanisms or principles by which it operates, and therefore allowing for an
abstract understanding of the behaviour. This can then be mimicked in a non-biological system
and its performance and effectiveness evaluated [130]. This process is summarised Figure 2.1.
2.1.3 Biology of Digital Ecosystems
Natural science is the study of the universe via the rules or laws of natural order, and the
term is also used to differentiate those fields using scientific method in the study of nature,
in contrast with the social sciences which apply the scientific method to culture and human
behaviour: economics, psychology, political economy, anthropology, etc [135]. The fields of
natural science are diverse, ranging from particle physics to astronomy [273], and while not all
IdentifyObserveUnderstandAbstractApplyEvaluate 26
, Chapter 2. Creation of Digital Ecosystems
these fields of study will provide paradigms for Digital Ecosystems, the further one wishes to
take the analogy of the word ecosystem, the more one has to consider the relevance of the fields
of natural science, particularly the biological sciences.
A primary motivation for our research in Digital Ecosystems is the desire to exploit the self-
organising properties of biological ecosystems. Ecosystems are thought to be robust, scalable
architectures that can automatically solve complex, dynamic problems [173]. However, the
biological processes that contribute to these properties have not been made explicit in Digital
Ecosystems research. Here, we discuss how biological properties contribute to the self-organising
features of biological ecosystems, including population dynamics, evolution, a complex dynamic
environment, and spatial distributions for generating local interactions [309]. The potential for
exploiting these properties in artificial systems is then considered. We suggest that several key
features of biological ecosystems have not been fully explored in existing digital ecosystems, and
discuss how mimicking these features may assist in developing robust, scalable self-organising
architectures.
Evolutionary computing uses natural selection to evolve solutions [110]; it starts with a set of
possible solutions chosen arbitrarily, then selection, replication, recombination, and mutation
are applied iteratively. Selection is based on conforming to a fitness function which is determined
by a specific problem of interest, and so over time better solutions to the problem can thus evolve
[110]. As Digital Ecosystems will likely solve problems by evolving solutions, they will probably
incorporate some form of evolutionary computing. However, we suggest that Digital Ecosystems
should also incorporate additional features, providing it with a closer resemblance to biological
ecosystems.
Including features such as complex dynamic fitness functions, a distributed or
network environment, and self-organisation arising from interactions among organisms and
their environment, such as those that we will now discuss.
Arguably the most fundamental differences between biological and digital ecosystems lie in the
motivation and approach of their respective researchers. Biological ecosystems are ubiquitous
natural phenomena whose maintenance is crucial to our survival [20], developing through the
process of ecological succession [29]. In contrast, Digital Ecosystems will be defined here as a
2.1. Background Theory ,
27
technology engineered to serve specific human purposes, developing to solve dynamic problems
in parallel with high efficiency.
2.1.3.1 Biological Ecosystems
Figure 2.2: Ecosystem Structure (redrawn from [261]): A stable, self-perpetuating system
made up of one or more communities of organisms, consisting of species in their habitats, with
their populations existing in their respective micro-habitats [29]. A community is a naturally
occurring group of populations from different species that live together, and interact as a self-
contained unit in the same habitat. A habitat is a distinct part of the environment [29].
An ecosystem is a natural unit made up of living (biotic) and non-living (abiotic) components,
from whose interactions emerge a stable, self-perpetuating system.
It is made up of one or
more communities of organisms, consisting of species in their habitats, with their populations
existing in their respective micro-habitats [29]. A community is a naturally occurring group
of populations from different species that live together, and interact as a self-contained unit
in the same habitat. A habitat is a distinct part of the environment [29], for example, a
stream. Individual organisms migrate through the ecosystem into different habitats competing
with other organisms for limited resources, with a population being the aggregate number of
the individuals, of a particular species, inhabiting a specific habitat or micro-habitat [29]. A
micro-habitat is a subdivision of a habitat that possesses its own unique properties, such as
nichenichenicheecosystemsmicrohabitatspopulationsenvironmentcommunityhabitat 28
, Chapter 2. Creation of Digital Ecosystems
a micro-climate [168]. Evolution occurs to all living components of an ecosystem, with the
evolutionary pressures varying from one population to the next depending on the environment
that is the population’s habitat. A population, in its micro-habitat, comes to occupy a niche,
which is the functional relationship of a population to the environment that it occupies. A
niche results in the highly specialised adaptation of a population to its micro-habitat [168].
2.1.3.2 Fitness Landscapes and Agents
As described above, an ecosystem comprises both an environment and a set of interacting,
reproducing entities (or agents) in that environment; with the environment acting as a set of
physical and chemical constraints on reproduction and survival [29]. These constraints can
be considered in abstract using the metaphor of the fitness landscape, in which individuals
are represented as solutions to the problem of survival and reproduction [335]. All possible
solutions are distributed in a space whose dimensions are the possible properties of individuals.
An additional dimension, height,
indicates the relative fitness (in terms of survival and
reproduction) of each solution. The fitness landscape is envisaged as a rugged, multidimensional
landscape of hills, mountains, and valleys, because individuals with certain sets of properties
are fitter than others [335], as visualised in Figure 2.3.
In biological ecosystems, fitness landscapes are virtually impossible to identify. This is both
because there are large numbers of possible traits that can influence individual fitness, and
because the environment changes over time and space [29].
In contrast, within a digital
environment, it is normally possible to specify explicitly the constraints that act on individuals
in order to evolve solutions that perform better within these constraints. Within genetic
algorithms, exact specification of a fitness landscape or function is common practice [110].
However, within a Digital Ecosystem the ideal constraints are those that allow solution
populations to evolve to meet user needs with maximum efficiency, with the user needs changing
from place to place and time to time. In this sense the fitness landscape of a Digital Ecosystem
is complex and dynamic, and more like that of a biological ecosystem than like that of a
traditional genetic algorithm [217, 110]. The designer of a Digital Ecosystem therefore faces a
double challenge: firstly, to specify rules that govern the shape of the fitness function/landscape
2.1. Background Theory ,
29
Figure 2.3: Fitness Landscape (modified from [314]): We can represent software development as
a walk through the landscape, towards the peaks which correspond to the optimal applications.
Each point represents a unique combination of software services, and the roughness of the
landscape indicates how difficult it is to reach an optimal software design [314]. In this example,
there is a global optimum, and several lower local optima.
in a way that meaningfully maps landscape dynamics to user requests, and secondly, to evolve
within this space, solution populations that are diverse enough to solve disparate problems,
complex enough to meet user needs, and efficient enough to be preferable to those generated
by other means.
The agents within a Digital Ecosystem will need to be like biological individuals in the sense
that they reproduce, vary, interact, move, and die [29]. Each of these properties contributes
to the dynamics of the ecosystem. However, the way in which these individual properties are
encoded may vary substantially depending on the intended purpose of the system [49].
2.1.3.3 Networks and Spatial Dynamics
A key factor in the maintenance of diversity in biological ecosystems is spatial interactions, and
several modelling systems have been used to represent these spatial interactions, including
metapopulations1, diffusion models, cellular automata and agent-based models (termed
individual-based models in ecology) [116]. The broad predictions of these diverse models are
1 A metapopulation is a collection of relatively isolated, spatially distributed, local populations bound together
by occasional dispersal between populations [175, 127, 128].
30
, Chapter 2. Creation of Digital Ecosystems
Figure 2.4: Abstract View of An Ecosystem: Showing different populations (by the different
colours) in different spatial areas, and their connection to one another by the lines. Included are
communities of populations that have become geographically separated and so are not connected
to the main network of the ecosystem, and which could potentially give rise to allopatric
(geographic) speciation [168].
in good agreement. At local scales, spatial interactions favour relatively abundant species
disproportionately. However, at a wider scale, this effect can preserve diversity, because different
species will be locally abundant in different places. The result is that even in homogeneous
environments, population distributions tend to form discrete, long-lasting patches that can
resist an invasion by superior competitors [116]. Population distributions can also be influenced
by environmental variations such as barriers, gradients, and patches. The possible behaviour
of spatially distributed ecosystems is so diverse that scenario-specific modelling is necessary to
understand any real system [119]. Nonetheless, certain robust patterns are observed. These
2.1. Background Theory ,
31
include the relative abundance of species, which consistently follows a roughly log-normal
relationship [30], and the relationship between geographic area and the number of species
present, which follows a power law [288]. The reasons for these patterns are disputed, because
they can be generated by both spatial extensions of simple Lotka-Volterra competition models
[136], and more complex ecosystem models [293].
Landscape connectivity plays an important part in ecosystems. When the density of habitats
within an environment falls below a critical threshold, widespread species may fragment into
isolated populations. Fragmentation can have several consequences. Within populations, these
effects include loss of genetic diversity and detrimental inbreeding [115]. At a broader scale,
isolated populations may diverge genetically, leading to speciation, as shown in Figure 2.4.
From an information theory perspective, this phase change in landscape connectivity can
mediate global and local search strategies [118]. In a well-connected landscape, selection favours
the globally superior, and pursuit of different evolutionary paths is discouraged, potentially
leading to premature convergence. When the landscape is fragmented, populations may
diverge, solving the same problems in different ways. Recently, it has been suggested that the
evolution of complexity in nature involves repeated landscape phase changes, allowing selection
to alternate between local and global search [117].
In a digital context, we can have spatial interactions by using a distributed system that consists
of a set of interconnected locations, with agents that can migrate between these connected
locations.
In such systems the spatial dynamics are relatively simple compared with those
seen in real ecosystems, which incorporate barriers, gradients, and patchy environments at
multiple scales in continuous space [29]. Nevertheless, depending on how the connections
between locations are organised, such Digital Ecosystems might have dynamics closely parallel
to spatially explicit models, diffusion models, or metapopulations [119]. We will discuss later
the use of a dynamic non-geometric spatial network, and the reasons for using this approach.
2.1.3.4 Selection and Self-Organisation
The major hypothetical advantage of Digital Ecosystems over other complex organisational
models is their potential for dynamic adaptive self-organisation. However, for the solutions
32
, Chapter 2. Creation of Digital Ecosystems
Figure 2.5: Evolving Population of Digital Organisms: A virtual petri dish at three successive
time-steps, showing the self-organisation of the population undergoing selection. The colour
shows the genetic variability of the digital organisms. Over time the fitter (purple) organisms
come to dominate the population, reproducing more and essentially replacing the weaker
organisms of the population [247].
evolving in Digital Ecosystems to be useful, they must not only be efficient in a computational
sense, but they must also solve purposeful problems. That is, the fitness of agents must translate
in some sense to real-world usefulness as demanded by the users [85].
Constructing a useful Digital Ecosystem therefore requires a balance between freedom of the
system to self-organise, and constraint of the system to generate useful solutions. These factors
must be balanced because the more the system’s behaviour is dictated by its internal dynamics,
the less it may respond to fitness criteria imposed by the users. At one extreme, when system
dynamics are mainly internal, agents may evolve that are good at survival and reproduction
within the digital environment, but useless in the real world [85]. At the other extreme,
where the users’ fitness criteria overwhelmingly dictates function, we suggest that dynamic
exploration, of the solution space and complexity, is likely to be limited. The reasoning behind
this argument is as follows. Consider a multidimensional solution space which maps to a rugged
fitness landscape [335]. In this landscape, competing solution lineages will gradually become
extinct through chance processes. So, the solution space explored becomes smaller over time as
the population adapts and the diversity of solutions decreases. Ultimately, all solutions may be
confined to a small region of the solution space. In a static fitness landscape, this situation is
desirable because the surviving solution lineages will usually be clustered around an optimum
[110]. However, if the fitness landscape is dynamic, the location of optima varies over time,
and should lineages become confined to a small area of the solution space, then subsequent
2.1. Background Theory ,
33
selection will locate only optima that are near this area [217]. This is undesirable if new, higher
optima arise that are far from pre-existing ones. A related issue is that complex solutions
are less likely to be found by chance than simple ones. Complex solutions can be visualised
as sharp, isolated peaks on the fitness landscape. Especially for dynamic landscapes, these
peaks are most likely to be found when the system explores the solution space widely [217].
Therefore, a self-organising mechanism other than the fitness criteria of users is required to
maintain diversity among competing solutions in a Digital Ecosystem.
2.1.3.5 Stability and Diversity in Complex Adaptive Systems
Ecosystems are often described as Complex Adaptive Systems (CAS), because like them, they
are systems made from diverse, locally interacting components that are subject to selection.
Other CAS include brains, individuals, economies, and the biosphere. All are characterised by
hierarchical organisation, continual adaptation and novelty, and non-equilibrium dynamics.
These properties lead to behaviour that is non-linear, historically contingent, subject to
thresholds, and contains multiple basins of attraction [173].
In the previous subsections, we have advocated Digital Ecosystems that include agent
populations evolving by natural selection in distributed environments. Like real ecosystems,
digital systems designed in this way fit the definition of CAS. The features of these systems,
especially non-linearity and non-equilibrium dynamics, offer both advantages and hazards for
adaptive problem-solving. The major hazard is that the dynamics of CAS are intrinsically
hard to predict because of the non-linear emergent self-organisation [174]. This observation
implies that designing a useful Digital Ecosystem will be partly a matter of trial and error.
The occurrence of multiple basins of attraction in CAS suggests that even a system that
functions well for a long period may suddenly at some point transition to a less desirable
state [97]. For example, in some types of system self-organising mass extinctions might result
from interactions among populations, leading to temporary unavailability of diverse solutions
[230]. This concern may be addressed by incorporating negative feedback or other mechanisms
at the global scale. The challenges in designing an effective Digital Ecosystem are mirrored by
34
, Chapter 2. Creation of Digital Ecosystems
the system’s potential strengths. Non-linear behaviour provides the opportunity for scalable
organisation and the evolution of complex hierarchical solutions, while rapid state transitions
potentially allow the system to adapt to sudden environmental changes with minimal loss of
functionality [173].
Figure 2.6: Ecosystems as Complex Adaptive Systems (modified from [9]): (LEFT) An abstract
view of an ecosystem showing the diversity of different populations by the different colours and
spacing. (RIGHT) An abstract view of diversity within a population, with the space between
points showing genetic diversity and the clustering prevalent.
A key question for designers of Digital Ecosystems is how the stability and diversity properties
of biological ecosystems map to performance measures in digital systems. For a Digital
Ecosystem the ultimate performance measure is user satisfaction, a system-specific property.
However, assuming the motivation for engineering a Digital Ecosystem is the development of
scalable, adaptive solutions to complex dynamic problems, certain generalisations can be made.
Sustained diversity [97], is a key requirement for dynamic adaptation. In Digital Ecosystems,
diversity must be balanced against adaptive efficiency because maintaining large numbers of
poorly-adapted solutions is costly. The exact form of this trade-off will be guided by the specific
requirements of the system in question. Stability [173], is likewise, a trade-off: we want the
system to respond to environmental change with rapid adaptation, but not to be so responsive
that mass extinctions deplete diversity or sudden state changes prevent control.
2.1. Background Theory ,
35
2.1.4 Computing of Digital Ecosystems
Based on the understanding of biological ecosystems, from the theoretical biology of the previous
subsection, we will now introduce fields from the domain of computer science relevant in the
creation of Digital Ecosystems. As we are interested in the digital counterparts for the behaviour
and constructs of biological ecosystems, instead of simulating or emulating such behaviour or
constructs, we will consider what parallels can be drawn.
The value of creating parallels between biological and computer systems varies substantially
depending on the behaviours or constructs being compared, and sometimes cannot be done
so convincingly. For example, both have mechanisms to ensure data integrity. In computer
systems, that integrity is absolute, data replication which introduces even the most minor
change is considered to have failed, and is supported by mechanisms such as the Message-Digest
algorithm 5 [266]. While in biological systems, the genetic code is transcribed with a remarkable
degree of fidelity; there is, approximately, only one unforced error per one hundred bases copied
[202]. There are also elaborate proof-reading and correction systems, which in evolutionary
terms are highly conserved [202]. In this example establishing a parallel is infeasible, despite
the relative similarity in function, because the operational control mechanisms in biological and
computing systems are radically different, as are the aims and purposes. This is a reminder
that considerable finesse is required when determining parallels, or when using existing ones.
We will start by considering Multi-Agent Systems to explore the references to agents and
migration;
followed by evolutionary computing and Service-Oriented Architectures for the
references to evolution and self-organisation.
2.1.4.1 Multi-Agent Systems
A software agent is a piece of software that acts, for a user in a relationship of agency,
autonomously in an environment to meet its designed objectives [334]. A Multi-Agent System
(MAS) is a system composed of several software agents, collectively capable of reaching goals
that are difficult to achieve by an individual agent or monolithic system [334]. Conceptually,
there is a strong parallel between the software agents of a MAS and the agent-based models of
36
, Chapter 2. Creation of Digital Ecosystems
Figure 2.7: Mobile Agent System: Visualisation that shows mobile agents as programmes
that can migrate from one host to another in a network of heterogeneous computer systems
and perform a task specified by its owner. On each host they visit, mobile agents need special
software called an agent station, which is responsible for executing the agents and providing a
safe execution environment [200].
a biological ecosystem [116], despite the lack of evolution and migration in a MAS. There is an
even stronger parallel to a variant of MASs, called mobile agent systems, in which the mobility
also mirrors the migration in biological ecosystems [249].
The term mobile agent contains two separate and distinct concepts: mobility and agency [269].
Hence, mobile agents are software agents capable of movement within a network [249]. The
mobile agent paradigm proposes to treat a network as multiple agent-friendly environments
and the agents as programmatic entities that move from location to location, performing tasks
for users. So, on each host they visit mobile agents need software which is responsible for their
execution, providing a safe execution environment [249].
Generally, there are three types of design for mobile agent systems [249]: (1) using a specialised
operating system, (2) as operating system services or extensions, or (3) as application software.
The first approach has the operating system providing the requirements of mobile agent systems
directly [301]. The second approach implements the mobile agent system requirements as
operating system extensions, taking advantage of existing features of the operating system [146].
Lastly, the third approach builds mobile agent systems as specialised application software that
runs on top of an operating system, to provide for the mobile agent functionality, with such
AgentStation 2.1. Background Theory ,
37
software being called an agent station [200]. In this last approach, each agent station hides
the vendor-specific aspects of its host platform, and offers standardised services to visiting
agents. Services include access to local resources and applications; for example, web servers
or web services, the local exchange of information between agents via message passing, basic
security services, and the creation of new agents [200]. Also, the third approach is the most
platform-agnostic, and is visualised in Figure 2.7.
2.1.4.2 Evolutionary Computing
For evolving software in Digital Ecosystems evolutionary computing is the logical field from
which to start. In Biologically-Inspired Computing, one of the primary sources of inspiration
from nature has been evolution [195]. Evolution has been clearly identified as the source
of many diverse and creative solutions to problems in nature [67, 104]. However,
it can
also be useful as a problem-solving tool in artificial systems. Computer scientists and other
theoreticians realised that the selection and mutation mechanisms that appear so effective
in biological evolution could be abstracted to be implemented in a computational algorithm
[195]. Evolutionary computing is now recognised as a sub-field of artificial intelligence (more
particularly computational intelligence) that involves combinatorial optimisation problems [14].
Evolutionary algorithms are based upon several
fundamental principles from biological
evolution, including reproduction, mutation, recombination (crossover), natural selection, and
survival of the fittest. As in biological systems, evolution occurs by the repeated application of
the above operators [13]. An evolutionary algorithm operates on a set of individuals, called a
population. An individual, in the natural world, is an organism with an associated fitness [168].
Candidate solutions to an optimisation problem play the role of individuals in a population, and
a cost function determines the environment within which the solutions live, analogous to the
way the environment selects for the fittest individuals. Candidate solutions to an optimisation
problem play the role of individuals in a population, and a cost function determines the
environment by selecting for the fittest individuals. The number of individuals varies between
different implementations and may also vary through time during the use of the algorithm.
Each individual possesses some characteristics that are defined through its genotype, its genetic
38
, Chapter 2. Creation of Digital Ecosystems
composition. These characteristics may be passed on to descendants of that individual [13].
Processes of mutation (small random changes) and crossover (generation of a new genotype by
the combination of components from two individuals) may occur, resulting in new individuals
with genotypes different from the ancestors they will come to replace. These processes iterate,
modifying the characteristics of the population [13]. Which members of the population are
kept, or are used as parents for offspring, will often depend upon some external characteristic,
called the fitness (cost) function of the population. It is this that enables improvement to occur
[13], and corresponds to the fitness of an organism in the natural world [168]. Recombination
and mutation create the necessary diversity and thereby facilitate novelty, while selection acts
as a force increasing quality. Changed pieces of information resulting from recombination and
mutation are randomly chosen. However, selection operators can be either deterministic, or
stochastic.
In the latter case, individuals with a higher fitness have a higher chance to be
selected than individuals with a lower fitness [13].
There are different strands of what has become called evolutionary computing [13]. The first
is genetic algorithms. A second strand, evolution strategies, focuses strongly on engineering
applications. A third strand, evolutionary programming, originally developed from machine
intelligence motivations, and is related to the other two. These areas developed separately
for about fifteen years, but from the early nineties they are seen as different representatives
(dialects) of one technology, called evolutionary computing [90]. In the early nineties, another
fourth stream following the general ideas had emerged, called genetic programming [90].
Genetic algorithms [110] implement a population of individuals, each of which possesses a
genotype that encodes a candidate solution to a problem. Typically genotypes are encoded
as bit-strings, but other encodings have been used in more recent developments of genetic
algorithms. Mutation and crossover, along with selection, are then used to choose a solution to
a problem. They have proven to be widely applicable, and have resulted in many applications
in differing domains [212]. Evolutionary strategies arose out of an attempt by several civil
engineers to understand a problem in hydrodynamics [74]. Evolutionary strategies [278] differ
from genetic algorithms in operating on real-valued parameters, and historically they have
tended not to use crossover as a variational operator, only mutation. However, mutation
rates have themselves been allowed to adapt in evolutionary strategies, which is not often
2.1. Background Theory ,
39
the case with genetic algorithms. Evolutionary strategies have also been used for many
applications [88].
Evolutionary programming arose distinctly from the first two strands of evolutionary computa-
tion, out of an attempt to understand machine intelligence through the evolution of finite state
machines [96]. Evolutionary programming [252] emphasises the evolution of the phenotype
(instance of a solution) instead of the genotype (genetic material) of individuals, and the
relation between the phenotype of parents and offspring, although crossover is not used. Thus,
evolutionary programming has some differences in approach from the other major strands of
evolutionary computation research. However, there have been many overlaps between the
different fields and it too has been applied in many areas [252].
Genetic programming [156] can be considered as a variant of genetic algorithms where individual
genotypes are represented by executable programmes. Specifically, solutions are represented
as trees of expressions in an appropriate programming language, with the aim of evolving the
most effective programme for solving a particular problem. Genetic programming, although
the newest form of evolutionary computing, has still proved to be widely applicable [22].
Many important questions remain to be answered in understanding the performance of
evolutionary algorithms. For example, current evolutionary algorithms for evolving programmes
(genetic programming) suffer from some weaknesses. First, while being moderately successful
at evolving simple programmes, it is very difficult to scale them to evolve high-level software
components [191]. Second, the estimated fitness of a programme is normally given by a measure
of how accurately it computes a given function, as represented by a set of input and output
pairs, and therefore there is only a limited guarantee that the evolved programme actually
does the intended computation [191]. These issues are particularly important when evolving
high-level, complex, structured software.
To evolve high-level software components in Digital Ecosystems, we propose taking advantage
of the native method of software advancement, human developers, and the use of evolutionary
computing for combinatorial optimisation [240] of the available software services. This involves
treating developer-produced software services as the functional building blocks, as the base
unit in a genetic-algorithms-based process. Such an approach would require a modular reusable
40
, Chapter 2. Creation of Digital Ecosystems
paradigm to software development, such as Service-Oriented Architectures, which are discussed
in the following subsection.
2.1.4.3 Service-Oriented Architectures
Our approach to evolving high-level software applications requires a modular reusable paradigm
to software development. Service-Oriented Architectures (SOAs) are the current state-of-the-
art approach, being the current iteration of interface/component-based design from the 1990s,
which was itself an iteration of event-oriented design from the 1980s, and before then modular
programming from the 1970s [44, 158].
Service-oriented computing promotes assembling
application components into a loosely coupled network of services, to create flexible, dynamic
business processes and agile applications that span organisations and computing platforms
[243]. This is achieved through a SOA, an architectural style that guides all aspects of
creating and using business processes throughout their life-cycle, packaged as services. This
includes defining and provisioning infrastructure that allows different applications to exchange
data and participate in business processes, loosely coupled from the operating systems and
programming languages underlying the applications [228]. Hence, a SOA represents a model in
which functionality is decomposed into distinct units (services), which can be distributed over
a network, and can be combined and reused to create business applications [243].
A SOA depends upon service-orientation as its fundamental design principle.
In a SOA
environment,
independent services can be accessed without knowledge of their underlying
platform implementation [228]. Services reflect a service-oriented approach to programming
that is based on composing applications by discovering and invoking network-available services
to accomplish some task. This approach is independent of specific programming languages or
operating systems, because the services communicate with each other by passing data from
one service to another, or by co-ordinating an activity between two or more services [243]. So,
the concepts of SOAs are often seen as built upon, and the development of, the concepts of
modular programming and distributed computing [158].
SOAs allow for an information system architecture that enables the creation of applications
that are built by combining loosely coupled and interoperable services [228]. They typically
2.1. Background Theory ,
41
implement functionality most people would recognise as a service, such as filling out an online
application for an account, or viewing an online bank statement [158]. Services are intrinsically
unassociated units of functionality, without calls to each other embedded in them.
Instead
of services embedding calls to each other in their source code, protocols are defined which
describe how services can talk to each other, in a process known as orchestration, to meet new
or existing business system requirements [287]. This is allowing an increasing number of third-
party software companies to offer software services, such that SOA systems will come to consist
of such third-party services combined with others created in-house, which has the potential
to spread costs over many users and uses, and promote standardisation both in and across
industries [51]. For example, the travel industry now has a well-defined, and documented, set
of both services and data, sufficient to allow any competent software engineer to create travel
agency software using entirely off-the-shelf software services [155, 46]. Other industries, such
as the finance industry, are also making significant progress in this direction [341].
The vision of SOAs assembling application components from a loosely coupled network
of services, that can create dynamic business processes and agile applications that span
organisations and computing platforms, is visualised in Figure 2.8. It will be made possible
by creating compound solutions that use internal organisational software assets, including
enterprise information and legacy systems, and combining these solutions with external
components residing in remote networks [242]. The great promise of SOAs is that the marginal
cost of creating the n-th application is virtually zero, as all the software required already exists
to satisfy the requirements of other applications. Only their combination and orchestration are
required to produce a new application [305, 213]. The key is that the interactions between the
chunks are not specified within the chunks themselves. Instead, the interaction of services (all
of whom are hosted by unassociated peers) is specified by users in an ad-hoc way, with the
intent driven by newly emergent business requirements [176].
The pinnacle of SOA interoperability, is the exposing of services on the internet as web services
[228]. A web service is a specific type of service that is identified by a Uniform Resource
Identifier (URI), whose service description and transport utilise open Internet standards.
Interactions between web services typically occur as Simple Object Access Protocol (SOAP)
calls carrying eXtensible Markup Language (XML) data content. The interface descriptions
42
, Chapter 2. Creation of Digital Ecosystems
Figure 2.8: Service-Oriented Architectures: Abstract visualisations, with the first image showing
the loosely joined services as cuboids, and the service orchestration as a polyhedron; and the
second image showing their high interoperability and re-usability in forming applications, from
the use of standardised interfaces and external service orchestration.
of web services are expressed using the Web Services Definition Language (WSDL) [241].
The Universal Description Discovery and Integration (UDDI) standard defines a protocol for
directory services that contain web service descriptions. UDDI enables web service clients to
locate candidate services and discover their details. Service clients and service providers utilise
these standards to perform the basic operations of SOAs [241]. Service aggregators can then
use the Business Process Execution Language (BPEL) to create new web services by defining
corresponding compositions of the interfaces and internal processes of existing services [241].
SOA services inter-operate based on a formal definition (or contract, e.g. WSDL) that is
independent of the underlying platform and programming language. Service descriptions
are used to advertise the service capabilities, interface, behaviour, and quality [241]. The
publication of such information about available services provides the necessary means for
2.1. Background Theory ,
43
discovery, selection, binding, and composition of services [241]. The (expected) behaviour
of a service during its execution is described by its behavioural description (for example,
as a workflow process). Also,
included is a quality of service (QoS) description, which
publishes important functional and non-functional service quality attributes, such as service
metering and cost, performance metrics (response time, for instance), security attributes,
integrity (transactional), reliability, scalability, and availability [241]. Service clients (end-user
organisations that use some service) and service aggregators (organisations that consolidate
multiple services into a new, single service offering) utilise service descriptions to achieve their
objectives [241]. One of the most important and continuing developments in SOAs is Semantic
Web Services (SWS), which make use of semantic descriptions for service discovery, so that a
client can discover the services semantically [254, 42].
There are multiple standards available and still being developed for SOAs [320], most notably of
recent being REpresentational State Transfer (REST) [287]. The software industry now widely
implements a thin SOAP/WSDL/UDDI veneer atop existing applications or components that
implement the web services paradigm [242], but the choice of technologies will change with time.
Therefore, the fundamentals of SOAs and its services are best defined generically, because SOAs
are technology agnostic and need not be tied to a specific technology [243]. Within the current
and future scope of the fundamentals of SOAs, there is clearly potential to evolve complex
high-level software applications from the modular services of SOAs, instead of the instruction
level evolution currently prevalent in genetic programming [157].
2.1.4.4 Distributed Evolutionary Computing
Having previously introduced evolutionary computing, and the possibility of it occurring within
a distributed environment, not unlike those found in mobile agent systems, leads us to consider
a specialised form known as distributed evolutionary computing (DEC). The motivation for
using parallel or distributed evolutionary algorithms is twofold: first, improving the speed of
evolutionary processes by conducting concurrent evaluations of individuals in a population;
second, improving the problem-solving process by overcoming difficulties that face traditional
evolutionary algorithms, such as maintaining diversity to avoid premature convergence [219,
44
, Chapter 2. Creation of Digital Ecosystems
297]. The fact that evolutionary computing manipulates a population of independent solutions
actually makes it well suited for parallel and distributed computation architectures [45]. There
are several variants of distributed evolutionary computing, leading some to propose a taxonomy
for their classification [235], with there being two main forms [45, 297]:
• multiple-population/coarse-grained migration/island models
• single-population/fine-grained diffusion/neighbourhood models
In the coarse-grained island models [178, 45], evolution occurs in multiple parallel sub-
populations (islands), each running a local evolutionary algorithm, evolving independently with
occasional migrations of highly fit individuals among sub-populations. The core parameters for
the evolutionary algorithm of the island-models are as follows [178]:
• number of the sub-populations: 2, 3, 4, more
• sub-population homogeneity
– size, crossover rate, mutation rate, migration interval
• topology of connectivity: ring, star, fully-connected, random
• static or dynamic connectivity
• migration mechanisms:
– isolated/synchronous/asynchronous
– how often migrations occur
– which individuals migrate
Fine-grained diffusion models [189, 297] assign one individual per processor. A local
neighbourhood topology is assumed, and individuals are allowed to mate only within their
neighbourhood, called a deme 2. The demes overlap by an amount that depends on their
shape and size, and in this way create an implicit migration mechanism. Each processor
runs an identical evolutionary algorithm which selects parents from the local neighbourhood,
produces an offspring, and decides whether to replace the current individual with an offspring.
However, even with the advent of multi-processor computers, and more recently multi-core
processors, which provide the ability to execute multiple threads simultaneously [193], this
approach would still prove impractical in supporting the number of agents necessary to create
a Digital Ecosystem. Therefore, we shall further consider the island models.
2 In biology a deme is a term for a local population of organisms of one species that actively interbreed with
one another and share a distinct gene pool [76].
2.1. Background Theory ,
45
Figure 2.9: Island-Model of Distributed Evolutionary Computing [178, 45]: There are different
probabilities of going from island 1(cid:13) to island 2(cid:13), as there is of going from island 2(cid:13) to island 1(cid:13).
This mirrors the naturally inspired quality that although two populations have the same physical
separation, it may be easier to migrate in one direction than the other, i.e. fish migration is
easier downstream than upstream.
An example island-model [178, 45] is visualised in Figure 2.9, in which there are different
probabilities of going from island 1(cid:13) to island 2(cid:13), as there is of going from island 2(cid:13) to island
1(cid:13). This allows maximum flexibility for the migration process, and mirrors the naturally
inspired quality that although two populations have the same physical separation, it may be
easier to migrate in one direction than the other, i.e. fish migration is easier downstream than
upstream. The migration of the island models is like the notion of migration in nature, being
similar to the metapopulation models of theoretical ecology [175]. This model has also been
used successfully in the determination of investment strategies in the commercial sector, in a
product known as the Galapagos toolkit [325, 56]. However, all the islands in this approach
work on exactly the same problem, which makes it less analogous to biological ecosystems in
which different locations can be environmentally different [29]. We will take advantage of this
property later when defining the Ecosystem-Oriented Architecture of Digital Ecosystems.
2.1.5 Digital Business Ecosystems
The questions we have raised are wide-ranging, and are motivating several interdisciplinary
research teams, including those involved in an EU Framework VI project called Digital Business
Ecosystems (DBEs). The DBE is a proposed methodology for economic and technological
P31P13P21P12P51P15P41P14P23P32P24P42P25P52P43P34P53P35P45P54Island3Island4Island5Island1Island2 46
, Chapter 2. Creation of Digital Ecosystems
Figure 2.10: Business Ecosystem [222]: Conceptual visualisation [93] showing a Business
Ecosystem of interacting Small and Medium sized Enterprise users, via the services they provide
and consume. Creating a network of business ecosystems distributed over different geographical
regions, business domains, and industry sectors.
innovation. Specifically, the DBE is a software infrastructure for supporting large numbers of
interacting business users and services [222]. The DBE aims to be a next generation Information
and Communications Technology that will extend the Service-Oriented Architecture concept
with the automatic combining of available and applicable services in a scalable architecture, to
meet business user requests for applications that facilitate business processes. In essence, the
DBE will be an internet-based environment in which businesses will be able to interact with
each other in very effective and efficient ways [223].
The synthesis of the concept of Digital Business Ecosystems emerged by adding [221] digital
in front of business ecosystem [214]. The term Digital Business Ecosystem was used earlier,
but with a focus exclusively on developing countries [215]. The generalisation of the term
to refer to a new interpretation of what socio-economic development catalysed by ICT means
was new, emphasising the co-evolution between the business ecosystem and its partial digital
representation: the digital ecosystem. The term Digital Business Ecosystem came to represent
the combination of the two ecosystems [222].
2.1. Background Theory ,
47
The business ecosystem is an economic community supported by a foundation of interacting
organisations and individuals;
i.e.
the organisms of the business world. This economic
community produces goods and services of value to customers, who are themselves members of
the ecosystem [214]. A wealthy ecosystem sees a balance between co-operation and competition
in a dynamic free market. Regarding a particular business ecosystem, two main different
interpretations of its structure have been discussed in the literature. The keystone model
has a structure in which a business ecosystem is dominated by a large firm that is surrounded
by many small suppliers [138]. This model works well when the central firm is healthy, but
represents a significant weakness for the economy of the region when the dominant economic
actor experiences difficulties [214]. This model also matches the economic structure of the
USA where there is a predominant number of large enterprises at the centre of large value
networks of suppliers [138]. However, the model for a business ecosystem developed in Europe
is less structured and more dynamic;
it is composed mainly of Small and Medium sized
Enterprises (SMEs), but can accommodate large firms [275]. All actors complement one
another, leading to a more dynamic division of labour, organised along one-dimensional value
chains and two-dimensional value networks [60]. This model is particularly well-adapted for
the service and knowledge industries, where it is easier for small firms to reinvent themselves
than, for instance, in the automotive industry which is dominated by large enterprises [222].
In the DBE, the digital ecosystem is the technical infrastructure, based on a peer-to-peer (P2P)
distributed software technology that transports, finds, and connects services and information
over Internet links enabling networked transactions, and the distribution of all the digital objects
present within the infrastructure [222]. Such organisms of the digital world encompass any
useful digital representations expressed by languages (formal or natural) that can be interpreted
and processed (by computer software and/or humans), e.g.
software applications, services,
knowledge, taxonomies, folksonomies, ontologies, descriptions of skills, reputation and trust
relationships, training modules, contractual frameworks, laws [222]. So, the Digital Business
Ecosystem is a biological metaphor that highlights the interdependence of all actors in the
business environment, who co-evolve their capabilities and roles [214], and which has attempted
to develop an isomorphic model between biological behaviour and the behaviour of the digital
ecosystem, leading to an evolutionary, self-organising, and self-optimising environment built
upon an underlying Service-Oriented Architecture [223].
The DBE aims to help local economic actors become active players in globalisation [82],
valorising their local culture and vocations, enabling them to interact and create value networks
48
, Chapter 2. Creation of Digital Ecosystems
at the global level.
Increasingly this approach, dubbed glocalisation, is being considered a
successful strategy of globalisation that preserves regional growth and identity [267, 303, 148],
and has been embraced by the mayors and decision-makers of thousands of municipalities [109],
because of the possible tension between globalisation and localisation when adopting ICTs [47].
The DBE represents a business-to-business (B2B) interaction concept supported by a software
platform (digital ecosystem) that is intended to have the desirable properties of biologi-
cal ecosystems [83], and its researchers also recognise the importance of Service-Oriented
Architectures in creating Digital Ecosystems [260, 223]. So, we will consider using the concept
of a business ecosystem as a potential user base for Digital Ecosystems.
2.2 Ecosystem-Oriented Architectures
We will now define the architectural principles of Digital Ecosystems. We will use our
understanding of theoretical biology from section 2.1.3, mimicking the processes and structures
of life, evolution, and ecology of biological ecosystems. We will achieve this by combining
elements from mobile agents systems, distributed evolutionary computing, and Service-Oriented
Architectures from section 2.1.4, to create a hybrid architecture which is the digital counterpart
of biological ecosystems.
We will refer to the agents of Digital Ecosystems as Agents, populations as Populations, and the
habitats as Habitats, to distinguish their new hybrid definitions from their original biological
and computing definitions.
2.2.1 Agents
The Agents of the Digital Ecosystem are functionally analogous to the organisms of biological
ecosystems, including the behaviour of migration and the ability to be evolved [29], and will be
achieved through using a hybrid of different technologies. The ability to migrate is provided
by using the paradigm of agent mobility from mobile agent systems [249], with the Habitats of
the Digital Ecosystem provided by the facilities of agent stations from mobile agent systems
[200], i.e. a distributed network of locations to migrate to and from. The Habitats, and the
Habitat network will be discussed later. The ability of the Agents to be evolved is in two
parts: first, by using the interoperability of services from Service-Oriented Architectures [228]
2.2. Ecosystem-Oriented Architectures ,
49
Figure 2.11: Agent of the Digital Ecosystem: A lightweight entity consisting primarily of
a pointer to the semantic web service it represents, which is Service-Oriented Architectures
compliant and therefore includes an executable component and semantic description. A software
service can be a software service only, e.g. for data encryption, or a software service providing
a front-end to a real-world service, e.g. selling books.
to aggregate Agents; and second, the use of evolutionary computing [90] for combinatorial
optimisation [240] at the Habitats to evolve optimal aggregations of Agents. The Agents will
take advantage of the interoperability of Service-Oriented Architectures [228], by acting in a
relationship of agency [334] to the user supplied semantic web services, which will be Service-
Oriented Architecture compliant [241]. We can then evolve high-level software applications by
using evolutionary computing [90] for combinatorial optimisation [240] of the available Agents,
or rather the services they represent, in a genetic-algorithms-based [110] process. This makes
an Agent, of the Digital Ecosystem, a lightweight entity consisting primarily of a pointer to the
semantic web service it represents, including the service’s executable component and semantic
description. A software service can be a software service only, e.g. for data encryption, or a
software service providing a front-end to a real-world service, e.g. selling books, as shown in
Figure 2.11.
An organism within Digital Ecosystems is an Agent, or an Agent aggregation created using
evolutionary optimisation in response to a user request for an application. These Agents will
migrate through the Habitat network of the Digital Ecosystem and adapt to find niches where
they are useful in fulfilling other user requests for applications. The Agents interact, evolve, and
adapt over time to the environment, thereby serving the ever-changing requirements imposed
by the user base.
User BaseSOADigital EcosystemAgentSemantic Web ServiceReal World ServiceAgentpointerpointer 50
, Chapter 2. Creation of Digital Ecosystems
The executable component, of a semantic web service that an Agent represents, is equivalent to
the DNA of an organism, whose sequence encodes the genetic information of living organisms
and has two primary functions [168]: the holder of virtually all information in inheritance,
and the controller of protein synthesis for the construction and operation of its organism.
Equivalently, the executable component is also the inheritable component from one generation
to the next, and defines the objects and behaviour of its service’s run-time instantiation.
The genotype of an individual describes the genetic constitution (DNA) of an individual,
independent of its physical existence (the phenotype) [168]. Equivalently, the semantic
description, of a semantic web service that an Agent represents, describes the functionality
of the executable component. The phenotype of an individual arises from the combination of
an organism’s DNA and the environment [168]. Equivalently, the run-time instantiation, of a
service that an Agent represents, results from instantiating the executable component in the
run-time environment. This differentiation between genotype and phenotype is fundamental
for escaping local optima, and is often lacking in artificial evolutionary systems [281], having
instead a one-to-one genotype-phenotype mapping, in which the phenotype is directly encoded
in the genotype with no differentiation provided by instantiation (development) [281]. Neutral
genotype-phenotype mappings have this differentiation between the genotype and phenotype
[285], which more strongly parallels biological evolution [21]. We therefore expect the use of a
neutral genotype-phenotype mapping to help Digital Ecosystems demonstrate behaviour more
akin to biological ecosystems.
2.2.1.1 Agent Aggregation
The executable component of a semantic web service that an Agent represents is equivalent
to an organism’s DNA and is the gene (functional unit) in the evolutionary process [168].
So, the Agents should be aggregated as a sequence, like the sequencing of genes in DNA
[168].
It could be argued that the Agents should be aggregated as an unordered set, or,
based on service orchestration, into a tree or workflow , as shown in Figure 2.12. However, the
aggregated structure of the Agents should not be the orchestration structure of the collection
of software services that the Agents represent, not only because the service orchestration of the
run-time instantiation is application domain-specific (e.g. trees in supply chain management
[163], workflows in the travel industry [31]), but because it would also move it undesirably
towards a one-to-one genotype-phenotype mapping [281].
2.2. Ecosystem-Oriented Architectures ,
51
Figure 2.12: Structure of Aggregated Agents: The executable component of a semantic web
service that an Agent represents is equivalent to an organism’s DNA and is the gene (functional
unit) in the evolutionary process [168]. So, the Agents should be aggregated as a sequence,
like the sequencing of genes in DNA [168]. Instead of an unordered set, or, based on service
orchestration, into a tree or workflow.
2.2.2 Habitats
The Habitats are the nodes of the Digital Ecosystem, and are functionally analogous to
the habitats of a biological ecosystem [168]. Their functionality is provided by using the
agent stations from mobile agent systems [200] (to provide a distributed environment in
which Agent migration can occur), with evolutionary computing [90] for the Agent interaction
(instead of traditional agent interaction mechanisms [334]), and the island-model of distributed
evolutionary computing [178] for the connectivity between Habitats. There will be a Habitat for
each user, which the users will typically run locally, and through which they will submit requests
for applications. Supporting this functionality, Habitats have the following core functions:
• Provide a subset of the Agents and Agent-sequences available globally, relevant to the user
that the Habitat represents, and stored in what we will call an Agent-pool (for reasons
that will be explained later).
• Accelerate, via the Agent-pool, the Populations instantiated to evolve optimal Agent-
sequences in response to user requests for applications.
• Manage the inter-Habitat connections for Agent migration.
• For service providers; manage the distribution of Agents (which represent their services)
to other users of the Digital Ecosystem, via the network of interconnected Habitats.
setsequencetreeworkflow 52
, Chapter 2. Creation of Digital Ecosystems
Figure 2.13: Habitat Network: Uses the agent stations from mobile agent systems [200] (to
provide a distributed environment in which Agent migration can occur), with evolutionary
computing [90] for the Agent interaction (instead of traditional agent interaction mechanisms
[334]), and the island-model of distributed evolutionary computing [178] for the connectivity
between Habitats.
The collection of Agents at each Habitat (peer) will change over time, as the more successful
Agents spread throughout the Digital Ecosystem, and as the less successful Agents are
deleted. Successive user requests over time to their dedicated Habitats makes this process
possible, because the continuous and varying user requests for applications provide a dynamic
evolutionary pressure on the Agents, which have to evolve to better satisfy those requests. So,
the Agents will recombine and evolve over time, constantly seeking to increase their effectiveness
for the user base. The Agent is the base unit of the evolutionary process in Digital Ecosystems,
in the same way that the gene is the base unit for evolution in biological ecosystems [29]. So,
the collection of Agents at each Habitat provides an Agent-pool, similar to a gene-pool, which
is all the genes in a population [168]. Additionally, it also stores Agent-sequences evolved from
the Habitat’s Populations, and Agent-sequences that migrate to the Habitat from other users’
Habitats, because they can potentially accelerate future Populations instantiated to respond to
user requests.
The landscape,
in energy-centric biological ecosystems, defines the connectivity between
habitats [29]. Connectivity of nodes in the digital world is generally not defined by geography
or spatial proximity, but by information or semantic proximity. For example, connectivity
in a peer-to-peer network is based primarily on bandwidth and information content, and
not geography. The island-models of distributed evolutionary computing use an information-
Genetic AlgorithmAgent StationHHHHHHHHHHHHHHHHHH 2.2. Ecosystem-Oriented Architectures ,
53
centric model for the connectivity of nodes (islands) [178]. However, because it is generally
defined for one-time use (to evolve a solution to one problem and then stop) it usually has
a fixed connectivity between the nodes, and therefore a fixed topology [45]. So, supporting
evolution in the Digital Ecosystem, with a dynamic multi-objective selection pressure (fitness
landscape [335] with many peaks), requires a re-configurable network topology, such that
Habitat connectivity can be dynamically adapted based on the observed migration paths of
the Agents between the users within the Habitat network. So, based on the island-models of
distributed evolutionary computing [178], each connection between the Habitats is bi-directional
and there is a probability associated with moving in either direction across the connection, with
the connection probabilities affecting the rate of migration of the Agents. However, additionally,
the connection probabilities will be updated by the success or failure of Agent migration using
the concept of Hebbian learning [132]: the Habitats which do not successfully exchange Agents
will become less strongly connected, and the Habitats which do successfully exchange Agents
will achieve stronger connections. This leads to a topology that adapts over time, resulting in a
network that supports and resembles the connectivity of the user base. When we later consider
an example user base, we will further discuss a resulting topology.
When a new user joins the Digital Ecosystem, a Habitat needs to be created for them, and
most importantly connected to the correct cluster(s) in the Habitat network. A new user’s
Habitat can be connected randomly to the Habitat network, as it will dynamically reconnect
based upon the user’s behaviour. User profiling can also be used to help connect a new user’s
Habitat to the optimal part of the network, by finding a similar user or asking the user to
identify a similar user, and then cloning their Habitat’s connections. Also, when a new Habitat
is created, its Agent-pool should be created by merging the Agent-pools of the Habitats to
which it is initially connected.
2.2.2.1 Agent Migration
The Agents migrate through the interconnected Habitats combining with one another in
Populations to meet user requests for applications. The migration path from the current Habitat
is dependent on the migration probabilities between the Habitats. The migration of an Agent
within the Digital Ecosystem is initially triggered by deployment to its user’s Habitat, for
distribution to other users who will potentially make use of the service the Agent represents.
When a user deploys a service, its representative Agent must be generated and deployed to
54
, Chapter 2. Creation of Digital Ecosystems
their Habitat. It is then copied to the Agent-pool of the user’s Habitat, and from there the
migration of the Agent occurs, which involves migrating (copying) the agent probabilistically
to all the connected Habitats. The Agent is copied rather than moved, because the Agent may
also be of use to the providing user. The copying of an Agent to a connected Habitat depends
on the associated migration probability. If the probability were one, then it would definitely
be sent. When migration occurs, depending on the probabilities associated with the Habitat
connections, an exact copy of the Agent is made at a connected Habitat. The copy of the
Agent is identical until the new Agent’s migration history is updated, which differentiates it
from the original. The successful use of the migrated Agent, in response to user requests for
applications, will lead to further migration (distribution) and therefore availability of the Agent
to other users.
The connections joining the Habitats are reinforced by successful Agent and Agent-sequence
migration. The success of the migration, the migration feedback, leads to the reinforcing and
creation of migration links between the Habitats, just as the failure of migration leads to the
weakening and negating of migration links between the Habitats. The success of migration is
determined by the usage of Agents at the Habitats to which they migrate. When an Agent-
sequence is found and used in responding to a user request, then the individual Agent migration
histories can be used to determine where they have come from and update the appropriate
connection probabilities.
If the Agent-sequence was fully or partly evolved elsewhere, then
where the sequence or sub-sequences were created needs to be passed on to the connection
probabilities, because the value in an Agent-sequence is the unique ordering and combination it
provides of the individual Agents contained within. So, it is necessary to manage the feedback
to the connection probabilities for migrating Agent-sequences, and not just the individual
Agents contained within the sequence, including the partial use of an Agent-sequence in a newly
evolved one. Specifically, the mechanism for migration feedback needs to know the Habitats
where migrating Agent-sequences were created, to create new connections or reinforce existing
connections to these Habitats. The global effect of the Agent migration and migration feedback
on the Habitat network is the clustering of Habitats around the communities present within
the user base, and will be discussed later in more detail.
The escape range is the number of escape migrations available to an Agent upon the risk of
death (deletion). If an Agent migrates to a Habitat and is not used after several user requests,
then it will have the opportunity to migrate (move not copy) randomly to another connected
Habitat. After this happens several times the Agent will be deleted (die). The escape range
2.2. Ecosystem-Oriented Architectures ,
55
will be dynamically responsive to the size of the Habitat cluster that the Agent exists within.
This creates a dynamic time-to-live [58] for the Agents, in which Agents that are used more
will live longer and distribute farther than those that are used less.
2.2.3 Populations
Figure 2.14: User Request to the Digital Ecosystem (modified from [162]): A user will formulate
queries to the Digital Ecosystem by creating a request as a semantic description, like those being
used and developed in Service-Oriented Architectures [254], specifying an application they desire
and submitting it to their Habitat. A Population is then instantiated in the user’s Habitat in
response to the user’s request, seeded from the Agents available at their Habitat (Agent-pool).
The Populations of the Digital Ecosystem are functionally equivalent to the evolving, self-
organising populations of a biological ecosystem, and are achieved through using evolutionary
computing. A population in biological ecosystems is all the members of a species that occupy
a particular area at a given time [168]. Our Population is also all the members of a species that
occupy a particular area at a given time, like an island from the island-models of distributed
evolutionary computing [178]. The use of distributed evolutionary computing to accelerate the
Populations will be explained later.
The users will formulate queries to the Digital Ecosystem by creating a request as a semantic
description,
like those being used and developed in Service-Oriented Architectures [254],
specifying an application they desire and submitting it to their Habitat. This description
enables the definition of a metric for evaluating the fitness of a composition of Agents, as a
distance function between the semantic description of the request and the Agents’ semantic
descriptions. A Population is then instantiated in the user’s Habitat in response to the user’s
request, seeded from the Agents available at their Habitat (i.e.
its Agent-pool). This allows
the evolutionary optimisation to be accelerated in the following three ways: first, the Habitat
HabitatAgent-poolAgentrequest ?solution userPopulation(Genetic Algorithm)Population(GA) 56
, Chapter 2. Creation of Digital Ecosystems
network provides a subset of the Agents available globally, which is localised to the specific
user it represents; second, making use of Agent-sequences previously evolved in response to
the user’s earlier requests; and third, taking advantage of relevant Agent-sequences evolved
elsewhere in response to similar requests by other users. The Population then proceeds to
evolve the optimal Agent-sequence(s) that fulfils the user request, and as the Agents are the
base unit for evolution, it searches the available Agent-sequence combination space. For an
evolved Agent-sequence that is executed (instantiated) by the user, it then migrates to other
peers (Habitats) becoming hosted where it is useful, to combine with other Agents in other
Populations to assist in responding to other user requests for applications.
2.2.3.1 Evolution
Evolution in biological ecosystems leads to both great diversity and high specialisation of its
organisms [29].
In Digital Ecosystems the diversity of evolution will provide for the wide
range of user needs and allow for quick responses to the changing of these user needs, while
the specialisation will simultaneously provide solutions which are tailored to fulfil specific
user requests. We will consider the issue of diversity in a later subsection, because it is
achieved through evolution in a distributed environment, which will be discussed later.
In
biological ecosystems, evolutionary specialisation is localised to a population within its micro-
habitat, which allows for the creation of niches (high specialisation) [168]. So, a Population is
instantiated in the user’s own Habitat, where the collection of Agents is chosen for the user,
and the micro-Habitat is provided by the user request. There is nothing to preclude more
than one Population being instantiated in a user’s Habitat at any one time, provided there are
computational resources sufficiently available.
A selection pressure is the sum aggregate of the forces acting upon a population, resulting in
genetic change through natural selection [168]. Those organisms best fit to survive the selection
pressures operating upon them will pass on their biological fitness to their progeny through
the inheritance process [168]. The fitness of an individual Agent-sequence within a Population
is determined by a selection pressure, applied as a fitness function [90] instantiated from the
user request, and works primarily on comparing the semantic descriptions of the Agents with
the semantic description of the user request. The pressure selects for those Agent-sequences
that are fit and capable of surviving the environment to reproduce, and against those that do
not have sufficient fitness and therefore die before passing on their genes, thereby providing
2.2. Ecosystem-Oriented Architectures ,
57
the direction for genetic change. In biology fitness is a measure of an organism’s success in its
environment [168], and its definition here will be further explained in the next subsection.
Genes are the functional unit in biological evolution [168]; whereas here the functional unit
is the Agent. Therefore, the evolutionary process of a Population provides a combinatorial
optimisation [240] of the Agents available, when responding to a user request. So, it does not
change or mutate the Agents themselves. In biology a mutation is a permanent transmissible
change (over the generations) in the genetic material (DNA) of an individual, and recombination
(e.g. crossover) is the formation within the offspring of alleles (gene combinations), which are
not present in the parents [168]. As in genetic algorithms [110], mutations will occur by
switching Agents in and out of the Agent-sequence structure, and recombination (crossover)
will occur by performing a crossing of two Agent-sequences.
As the Digital Ecosystem receives more and more sophisticated requests, so more and more
complex applications are evolved and become available for use by the users. To achieve this
evolution, specifically the Agent-sequence recombination and optimisation, is a very significant
challenge, because of the range of services that must be catered for and the potentially huge
number of factors that must be considered for creating an applicable fitness function. First,
to construct ever more complex software solutions, requires modularity, which is provided by
the paradigm of service interoperability from Service-Oriented Architectures [228]. Second,
two of the most important issues are that of defining fitness and managing bloat, which we will
discuss next. Finally, there is a huge body of work and continuing research regarding theoretical
approaches to evolutionary computing [90], including the extensive use of genetic algorithms for
practical real-world problem solving [85]. In defining Digital Ecosystems we should make use of
the current state-of-the-art, and future developments, in the areas of evolutionary computing
[144] and service interoperability [228].
2.2.3.2 Fitness
In biology fitness is a measure of how successful an organism is in its environment, i.e.
its
phenotype [168]. The fitness of an Agent-sequence within a Population would also, ideally,
be based upon its phenotype, the run-time instantiation, and nothing else. However, such
an approach would be impractical, because it is currently infeasible to execute all the Agent-
sequences of a Population at every generation, and not least because of the computational
58
, Chapter 2. Creation of Digital Ecosystems
resources that would be required. The other concern is one of practicality, by which we mean
that it may not even be possible to perform a live execution for the executable components of an
Agent-sequence; for example, if they are for buying an item from an online retailer. These are
well known issues in evolutionary computing, which is why fitness functions are often defined
as simulated input/output pairs to test functionality [191]. In Digital Ecosystems we can use
historical usage information, but this would be insufficient initially, because such information
would not be available at the time of an Agent’s deployment. However, because each Agent also
carries a semantic description, a specification of what it does, the fitness function can measure a
complete Agent-sequence’s collective semantic descriptions relative to the semantic description
of a user request. So, initially the fitness function should be based primarily on comparing the
semantic descriptions of the Agent-sequences to the semantic description of the user request,
ever increasingly augmented with the growing usage information available for the Agents. In
biological terms the genotype will be used as the phenotype, combined with any available past
fitness of the phenotype; with the Agent’s semantic description (genotype) therefore acting as a
guarantee of its expected behaviour. So, for any newly deployed Agent a one-to-one genotype-
phenotype mapping [281] will initially exist, until sufficient usage information is available. While
the use of such a mapping is undesirable, it is temporary, and necessary to allow Digital
Ecosystems to operate effectively.
We have already suggested that the primary driver of the evolutionary process should initially
be the extent by which an Agent-sequence can verifiably satisfy the specified requirements. This
could be measured probabilistically, or using theorem-proving to validate the system, though
automatic theorem proving is notoriously slow [289, 277]. However, there will also be other
pressures on the fitness. For example, one may seek the most parsimonious solution to a problem
(one that provides exactly the specified features and no more), or the cheapest solution, or one
with a good reputation. Some aspects of fitness will be implicit in the evolutionary process (e.g.
Agents which are often used will gain more fitness) while others will require explicit measures
(e.g. price, or user satisfaction). One way to handle this multiplicity of fitness values (some
qualitative) is to explicitly recognise the multi-objective nature of the optimisation problem. In
this way, we are seeking not the single best solution, but a range of possible compromises that
can be made most optimally. The set of solutions for which there are no better compromises is
called the Pareto-set, and evolutionary techniques have been adapted to solve such problems
with considerable success [318]. The main point is that selection has to be driven not by an
absolute value of fitness, but rather by a notion of what it means for one solution to be better
2.2. Ecosystem-Oriented Architectures ,
59
than another. We say one solution dominates another if it is better in at least one respect, and
no worse in any of the others [98].
2.2.3.3 Bloat
If the repetition of Agents is allowed within evolving Agent-sequences, then the search space
can become countably infinite, because the nature of the problem to be solved may not allow us
to determine what the length of a solution is beforehand. Therefore, a variable length approach
must be adopted, which is common in genetic programming [156]. When variable length
representations of solutions are used, a well-known phenomenon arises, called bloat, in which
the individuals of an evolving population tend to grow in size without gaining any additional
advantage [166]. The bloat phenomenon can cause early termination of an evolutionary process
due to the exhaustion of the available memory, and can also significantly reduce performance,
because typically longer sequences have higher fitness computation costs [259]. Bloat is not
specific to genetic programming, and is inherent in search techniques with discrete variable
length representations [164]. It is a fundamental area of research within search-based approaches
such as genetic algorithms, genetic programming and other approaches not based on populations
such as simulated annealing [164]. However, considerable work on bloat has been done in
connection with genetic programming [165, 22], and we believe that the genetic algorithms
community generally, and the genetic-algorithms-based approach of our Digital Ecosystems
specifically, can benefit directly from this research. While bloat is a phenomenon which was
first observed in practice [156], theoretical analyses have been attempted [23]. One should take
care with these approaches as implementations will always deal with finite populations, while
theoretical approaches often deal with infinite populations [156], and this difference can be
important. Yet, both theoretical and empirical approaches are required to understand bloat.
There are many factors contributing to bloat, and while the phenomenon may appear simple,
the reasons are not. There are several theories to explain why this occurs, and, as we shall
discuss, some measures that can be taken for its prevention.
There are several different qualitative theories which attempt to explain bloat, and they can
be considered in two groups. First, protection against crossover and bias removal (which can
be considered jointly) and second, the nature of programme search spaces [23]. First, near
the end of a run a Population consists of mostly fit individuals, and any crossover is likely
to be detrimental to the fitness of the offspring.
In any sequence of Agents there may be
60
, Chapter 2. Creation of Digital Ecosystems
Agents that do not contribute semantically to the complete functionality of the sequence if,
for example, their functionality was not requested by the user or if it is duplicated in the
sequence; analogously to genetic programming [23], we can call these redundant Agents bloat.
The genotype can then be grown further without affecting the phenotype if Agents with similar
functionality are added; but, as the genotype grows larger, crossover is more likely to transfer
redundant Agents to the new off-springs (assuming uniform crossover). Second, above a certain
threshold size, the distribution of functionality does not vary with the size of the search space
[23]. Thus, if we randomly sample long and short Agent-sequences above a length threshold,
they will likely have the same functionality and fitness. So, as a search process progresses we
are more likely to sample longer Agent-sequences, as mutation results in more of them (all
other things being equal) and this will give rise to the bloating phenomenon.
Each of the stages of construction of a genetic algorithm (i.e. choice of fitness function, selection
method and genetic operator) can affect bloat. It has been shown that even small differences
in the fitness function can cause a difference: a single programme glitch in an otherwise flat
fitness landscape (from the neutral theory of molecular evolution [149]) is sufficient to drive
the average programme size of an infinite population [204]. If a fitness-proportional selection
method is used, individuals with zero fitness will be discontinued as they have zero probability
of being selected as parents [36]. However, if tournament selection method is used, then there
is a finite chance that individuals with zero fitness will be selected to be parents [36]. Finally,
the choice of genetic operator affects the size of the programmes which are sampled; standard
crossover on a flat landscape heavily oversamples the shorter programmes [251]. There are
other factors that may affect bloat, for example, how the population is initialised, or the choice
of representation used, such as a neutral genotype-phenotype mapping, which can actually
alleviate bloat [208].
Bloat is a fact, whatever the reasons, happening in this type of optimisation and needs to be
controlled if the space is to be searched effectively. One solution is to apply a hard limit to the
size of the sets that can be sampled [166]: this enables the search algorithm to keep running
without having out-of-memory run-time errors, but poses questions on how to set this hard
limit. An alternative but similar method is to apply a parsimony pressure, where a term is
added to the fitness function which chastises big sets in preference for smaller sets [296]. In
this approach, individuals larger than the average size are evaluated with a reduced probability,
biasing the search to smaller sets, while providing a dynamic limit which adapts to the average
size of individuals in a changing population [296].
2.2. Ecosystem-Oriented Architectures ,
61
2.2.4 The Digital Ecosystem
The Digital Ecosystem supports the automatic combining of numerous Agents (which represent
services), by their interaction in evolving Populations to meet user requests for applications, in
a scalable architecture of distributed interconnected Habitats. The sharing of Agents between
Habitats ensures the system is scalable, while maintaining a high evolutionary specialisation for
each user. The network of interconnected Habitats is equivalent to the abiotic environment of
biological ecosystems [29]; combined with the Agents, the Populations, the Agent migration for
distributed evolutionary computing, and the environmental selection pressures provided by the
user base, then the union of the Habitats creates the Digital Ecosystem, which is summarised
in Figure 2.15. The continuous and varying user requests for applications provide a dynamic
evolutionary pressure on the Agent sequences, which have to evolve to better fulfil those user
requests, and without which there would be no driving force to the evolutionary self-organisation
of the Digital Ecosystem.
In the Digital Ecosystem, local and global optimisations concurrently operate to determine
solutions to satisfy different optimisation problems. The global optimisation here is not a
decentralised super-peer based control mechanism [263], but the completely distributed peer-
to-peer network of the interconnected Habitats, which are therefore not susceptible to the failure
of super-peers. It provides a novel optimisation technique inspired by biological ecosystems,
working at two levels: a first optimisation, migration of Agents which are distributed in a
peer-to-peer network, operating continuously in time; this process feeds a second optimisation,
based on evolutionary combinatorial optimisation, operating locally on single peers and is aimed
at finding solutions that satisfy locally relevant constraints. So, the local search is improved
through this twofold process to yield better local optima faster, as the distributed optimisation
provides prior sampling of the search space through computations already performed in other
peers with similar constraints. This novel form of distributed evolutionary computing will be
discussed further below, once we have discussed a topology resulting from an example user
base.
If we consider an example user base for the Digital Ecosystem, the use of Service-Oriented
Architectures in its definition means that business-to-business (B2B) interaction scenarios [158]
lend themselves to being a potential user base for Digital Ecosystems. So, we can consider
the business ecosystem of Small and Medium sized Enterprise (SME) networks from Digital
Business Ecosystems [222], as a specific class of examples for B2B interaction scenarios; and in
62
, Chapter 2. Creation of Digital Ecosystems
Figure 2.15: Digital Ecosystem: A network of interconnected Habitats, combined with the
Agents, the Populations, the Agent migration for distributed evolutionary computing, and the
environmental selection pressures provided by the user base, then the union of the Habitats
creates the Digital Ecosystem. Agents travel along the peer-to-peer connections; in every node
(Habitat) local optimisation is performed through an evolutionary algorithm, where the search
space is determined by the Agents present at the node.
which the SME users are requesting and providing software services, represented as Agents
in the Digital Ecosystem, to fulfil the needs of their business processes. Service-Oriented
Architectures promise to provide potentially huge numbers of services that programmers
can combine, via the standardised interfaces, to create increasingly more sophisticated and
distributed applications [241]. The Digital Ecosystem extends this concept with the automatic
combining of available and applicable services, represented by Agents, in a scalable architecture,
to meet user requests for applications. These Agents will recombine and evolve over time,
constantly seeking to improve their effectiveness for the user base. From the SME users’ point
of view the Digital Ecosystem provides a network infrastructure where connected enterprises
can advertise and search for services (real-world or software only), putting a particular emphasis
migrating AgentHabitat networkHabitatAgent-poolAgentPopulation(Genetic Algorithm)application(Agent-sequence)Population(GA) 2.2. Ecosystem-Oriented Architectures ,
63
on the composability of loosely coupled services and their optimisation to local and regional,
needs and conditions. To support these SME users the Digital Ecosystem is satisfying the
companies’ business requirements by finding the most suitable services or combination of
services (applications) available in the network. A composition of services is an Agent-sequence
in the Habitat network that can move from one peer (company) to another, being hosted only
in those where it is most useful in satisfying the SME users’ business needs.
2.2.4.1 Topology
The Digital Ecosystem allows for the connectivity in the Habitats to adapt to the connectivity
within the user base, with a cluster of Habitats representing a community within the user
base.
If a user is a member of more than one community, the user’s Habitat will be in
more than one cluster. This leads to a network topology that will be discovered with time,
and which reflects the connectivity within the user base. Similarities in requests by different
users will reinforce behavioural patterns, and lead to clustering of the Habitats within the
ecosystem, which can occur over geography, language, etc. This will form communities for
more effective information sharing, the creation of niches, and will improve the responsiveness
of the system. The connections between the Habitats will be self-managed, through the
mechanism of Agent migration defined earlier. Essentially, successful Agent migration will
reinforce Habitat connections, thereby increasing the probability of future Agent migration
along these connections. If a successful multi-hop migration occurs, then a new link between
the start and end Habitats can be formed. Unsuccessful migrations will lead to connections
(migration probabilities) decreasing, until finally the connection is closed.
If we consider the business ecosystem - a network of Small and Medium sized Enterprises
from Digital Business Ecosystems [222] - as an example user base, such business networks
are typically small-world networks [333, 338]. They have many strongly connected clusters
(communities), called sub-networks (quasi-complete graphs), with a few connections between
these clusters (communities) [327]. Graphs with this topology have a very high clustering
coefficient and small characteristic path lengths [327]. As the connections between Habitats are
reconfigured depending on the connectivity of the user base, the Habitat clustering will therefore
be parallel to the business sector communities, as shown in Figure 2.16. The communities will
cluster over language, nationality, geography, etc. – all depending on the user base. So, the
Digital Ecosystem will take on a topology similar to that of the user base.
64
, Chapter 2. Creation of Digital Ecosystems
Figure 2.16: Digital Business Ecosystem: Business ecosystem, network of Small and Medium
sized Enterprises [222], using the Digital Ecosystem. As the connections between Habitats are
reconfigured depending on the connectivity of the user base, the Habitat clustering will therefore
be parallel to the business sector communities.
Fragmentation of the Habitat network can occur, but only if dictated by the structure of the
user base. The issue of greater concern is when individual Habitats become totally disconnected,
which can only occur under certain conditions. One condition is that the Agents within the
Agent-pool consistently fail to satisfy user requests. Another condition is when the Agents and
Agent-sequences they share are undesirable to the users that are within the migration range
of these Agents and Agent-sequences. These scenarios can arise because the Habitat is located
within the wrong cluster, in which case the user can be asked to join another cluster within the
Habitat network, assuming the user base is of sufficient size to provide a viable alternative.
Figure 2.17: Habitat Clustering: Topology adapted to the small-world network of a business
ecosystem of SMEs from Digital Business Ecosystems [222], having many strongly connected
clusters (communities), called sub-networks (quasi-complete graphs), with a few connections
between these clusters (communities) [327]. Graphs with this topology have a very high clustering
coefficient and small characteristic path lengths [327].
BusinessEcosystemDigitalEcosystemHabitatSMEHHHHHHHHHHHHHHCommunity(cluster) 2.2. Ecosystem-Oriented Architectures ,
65
2.2.4.2 Distributed Evolution
The Digital Ecosystem is a hybrid of Multi-Agent Systems, more specifically of mobile agent
systems, Service-Oriented Architectures, and distributed evolutionary computing, which leads
to a novel form of evolutionary computation. The novelty comes from the creation of multiple
evolving Populations in response to similar requests, whereas in the island-models of distributed
evolutionary computing there are multiple evolving populations in response to only one request
[178].
So,
in our Digital Ecosystem different requests are evaluated on separate islands
(Populations), with their evolution accelerated by the sharing of solutions between the evolving
Populations (islands), because they are working to solve similar requests (problems). This
is shown in Figure 2.18, where the dashed yellow lines connecting the evolving Populations
indicate similarity in the requests being managed.
Figure 2.18: Distributed Evolution in the Digital Ecosystem: Different requests are evaluated
on separate islands (Populations), with their evolution accelerated by the sharing of solutions
between the evolving Populations (islands), because they are working to solve similar requests
(problems). The yellow lines connecting the evolving Populations indicate similarity in the
requests being managed.
If we again consider the business ecosystem of Small and Medium sized Enterprises from Digital
Business Ecosystems [222] as an example user base, then in Figure 2.18 the four Habitats, in
the left cluster, could be travel agencies, and the three with linked evolving Populations are
looking for similar package holidays. So, an optimal solution found and used in one Habitat
will be migrated to the other connected Habitats and integrated into any evolving Populations
via the local Agent-pools. This will help to optimise the search for similar package holidays at
the Habitats of the other travel agencies. This also works in a time-shifted manner, because
an optimal solution is stored in the Agent-pool of the Habitats to which it is migrated, being
available to optimise a similar request placed later.
ClusterHabitatPopulation(GA) 66
, Chapter 2. Creation of Digital Ecosystems
The distributed architecture of Digital Ecosystems favours the use of Pareto-sets for fitness
determination, because Pareto optimisation for multi-objective problems is usually most
effective with spatial distribution of the populations, as partial solutions (solutions to different
niches) evolve in different parts of a distributed population [75] (i.e. different Populations in
different Habitats). By contrast, in a single population, individuals are always interacting with
each other, via crossover, which does not allow for this type of specialisation [13].
This approach requires the Digital Ecosystem to have a sufficiently large user base, so that
there can be communities within the user base, and therefore allow for similarity in the user
requests. Assuming a user base of hundreds of users, then there would be hundreds of Habitats,
in which there will be potentially three or more times the number of Populations at any one
time. Then there will be thousands of Agents and Agent-sequences (applications) available to
meet the requests for applications from the users. In such a scenario, there would be a sufficient
number of users for the Digital Ecosystem to find similarity within their requests, and therefore
apply our novel form of distributed evolutionary computing.
2.2.4.3 Agent Life-Cycle
An Agent is created to represent a user’s service in the Digital Ecosystem, and its life-cycle
begins with deployment to its owner’s Habitat for distribution within the Habitat network. The
Agent is then migrated to any Habitats connected to the owner’s Habitat, to make it available
in other Habitats where it could potentially be useful. The Agent is then available to the local
evolutionary optimisation, to be used in evolving the optimal Agent-sequence in response to a
user request. The optimal Agent-sequence is then registered at the Habitat, being stored in
the Habitat’s Agent-pool. If an Agent-sequence solution is then executed, an attempt is made
to migrate (copy) it to every other connected Habitat, success depending on the probability
associated with the connection. The Agent life-cycle is shown in Figure 2.19.
An Agent can also be deleted if after several successive user requests at a Habitat it remains
unused; it will have a small number of escape migrations, in which it is not copied, but is
randomly moved to another connected Habitat. If the Agent fails to find a niche before running
out of escape migrations, then it will be deleted.
2.3. Simulation and Results ,
67
Figure 2.19: Agent Life-Cycle: Begins with deployment to its owner’s Habitat for distribution
within the Habitat network. It can then be used in evolving the optimal Agent-sequence in
response to a user request. The optimal Agent-sequence is then registered at the Habitat. If
an Agent-sequence solution is then executed, an attempt is made to migrate (copy) it to every
other connected Habitat, success depending on the probability associated with the connection.
2.3 Simulation and Results
We simulated the Digital Ecosystem, based upon our Ecosystem-Oriented Architecture, and
recorded key variables to determine whether it displayed behaviour typical of biological
ecosystems. Although agent-based modelling solutions, like Repast (Recursive Porous Agent
Simulation Toolkit) [57] and MASON (Multi-Agent Simulator Of Neighbourhoods) [182], and
evolutionary computing libraries, like ECJ (Evolutionary Computing in Java) [183] and the
JCLEC (Java Computing Library for Evolutionary Computing) [319], are available, it was
evident that it would take as much effort to adapt one, or a combination, of these to simulate
the Digital Ecosystem, as it would to create our own simulation of the Digital Ecosystem,
because the required ecological dynamics are largely absent from these and other available
technologies. So, we created our own simulation, following the Ecosystem-Oriented Architecture
from the previous section (unless otherwise specified), using the business ecosystem of Small
and Medium sized Enterprises from Digital Business Ecosystems [222] as an example user base.
2.3.1 Agents: Semantic Descriptions
An Agent represents a user’s service, including the semantic description of the business process
involved, and is based on existing and emerging technologies for semantically capable Service-
evolve Agent Populationfor solution (Agent-sequence)register solution at Agent-pool of HabitatexecuteAgent-sequenceuser requestAgent migrationdeploy Agent to Agent-pool of Habitat 68
, Chapter 2. Creation of Digital Ecosystems
A = {(1,25), (2,35), (3,55), (4,6), (5,37), (6,12)}
Figure 2.20: Agent Semantic Descriptions: Each simulated Agent had a semantic description
with an abstract representation consisting of a set of between three and six numeric tuples;
each tuple representing an attribute of the semantic description, one integer for the attribute
identifier and one for the attribute value, with both ranging between one and a hundred.
Oriented Architectures [254], such as the OWL-S semantic markup for web services [197]. We
simulated a service’s semantic description with an abstract representation consisting of a set of
numeric tuples, to simulate the properties of a semantic description. Each tuple representing
an attribute of the semantic description, one integer for the attribute identifier and one for the
attribute value, with both ranging between one and a hundred. Each simulated Agent had a
semantic description, with between three and six tuples, as shown in Figure 2.20.
2.3.2 User Base
R = [{(1,23),(2,45),(3,33),(4,6),(5,8),(6,16)}, {(1,84),(2,48),(3,53),(4,11),(5,16)}]
Figure 2.21: User Request: A simulated user request consisted of an abstract semantic
description, as a list of sets of numeric tuples to represent the properties of a desired business
application; each tuple representing an attribute of the semantic description, one integer for
the attribute identifier and one for the attribute value, with both ranging between one and a
hundred.
Throughout the simulations we assumed a hundred users, which meant that at any time the
number of users joining the network equalled those leaving. The Habitats of the users were
randomly connected at the start, to simulate the users going online for the first time. The users
then produced Agents (services) and requests for business applications. Initially, the users each
deployed five Agents to their Habitats, for migration (distribution) to any Habitats connected
to theirs (i.e. their community within the business ecosystem). Users were simulated to deploy
a new Agent after the submission of three requests for business applications, and were chosen
at random to submit their requests. A simulated user request consisted of an abstract semantic
description, as a list of sets of numeric tuples to represent the properties of a desired business
application. The use of the numeric tuples made it comparable to the semantic descriptions
of the services represented by the Agents; while the list of sets (two level hierarchy) and a
much longer length provided sufficient complexity to support the sophistication of business
applications. An example is shown in Figure 2.21.
2.3. Simulation and Results ,
69
The user requests were handled by the Habitats instantiating evolving Populations, which used
evolutionary computing to find the optimal solution(s), Agent-sequence(s).
It was assumed
that the users made their requests for business applications accurately, and always used the
response (Agent-sequence) provided.
2.3.3 Populations: Evolution
Populations of Agents, [A1, A1, A2, ...], were evolved to solve user requests, seeded with Agents
and Agent-sequences from the Agent-pool of the Habitats in which they were instantiated. A
dynamic population size was used to ensure exploration of the available combinatorial search
space, which increased with the average length of the Population’s Agent-sequences. The
optimal combination of Agents (Agent-sequence) was evolved to the user request R, by an
artificial selection pressure created by a fitness function generated from the user request R.
An individual (Agent-sequence) of the Population consisted of a set of attributes, a1, a2, ...,
and a user request essentially consisted of a set of required attributes, r1, r2, .... So, the fitness
function for evaluating an individual Agent-sequence A, relative to a user request R, was
f itness(A, R) =
(2.1)
1 +(cid:80)
1
r∈R r − a ,
where a is the member of A such that the difference to the required attribute r was minimised.
Equation 2.1 was used to assign fitness values between 0.0 and 1.0 to each individual of the
current generation of the population, directly affecting their ability to replicate into the next
generation. The evolutionary computing process was encoded with a low mutation rate, a fixed
selection pressure and a non-trapping fitness function (i.e. did not get trapped at local optima).
The type of selection used fitness-proportional and non-elitist, fitness-proportional meaning that
the fitter the individual the higher its probability of surviving to the next generation [36]. Non-
elitist means that the best individual from one generation was not guaranteed to survive to the
next generation; it had a high probability of surviving into the next generation, but it was not
guaranteed as it might have been mutated [90]. Crossover (recombination) was then applied
to a randomly chosen 10% of the surviving population, a one-point crossover, by aligning two
parent individuals and picking a random point along their length, and at that point exchanging
their tails to create two offspring [90]. Mutations were then applied to a randomly chosen 10%
of the surviving population; point mutations were randomly located, consisting of insertions (an
Agent was inserted into an Agent-sequence), replacements (an Agent was replaced in an Agent-
70
, Chapter 2. Creation of Digital Ecosystems
sequence), and deletions (an Agent was deleted from an Agent-sequence) [168]. The issue of
bloat was controlled by augmenting the fitness function with a parsimony pressure [296] which
biased the search to shorter Agent-sequences, evaluating longer than average length Agent-
sequences with a reduced fitness, and thereby providing a dynamic control limit which adapted
to the average length of the ever-changing evolving Agent Populations.
2.3.4 Semantic Filter
Agent’s semantic description:
{(1,25), (2,35), (3,55), (4,6), (5,37), (6,12)}
{(Business, Airline), (Company, British Midland), (Quality, Economy),
(Cost, 60), (Depart, Edinburgh), (Arrive, London)}
(with semantic filter):
user request:
[{(1,23), (2,45), (3,33), (4,6), (5,8), (6,16)}, {(1,84), (2,48), (3,53), (4,11),
(7,16), (8,34)}, {(1,23), (2,45), (3,53), (4,6), (5,16)(6,53)}, {(1,86),
(2,48), (3,33), (4,25), (7,55)(8,23)}, {(1,25), (2,52), (3,53), (4,5), (5,55),
(6,37)}, {(1,86), (2,48), (3,43), (4,25), (7,37), (8,40)}, {(1,22), (2,77),
(3,82), (4,9), (5,35), (6,8)}]
3*),
110),
(with semantic filter):
(Cost,
(Quality,
[{(Business, Airline), (Company, Air France), (Quality, Economy),
(Cost, 60), (Depart, Edinburgh), (Arrive, Paris)}, {(Business, Hotel),
(Company, Continental),
(Location,
Paris), (Nights, 3)}, {(Business, Airline), (Company, Air France),
(Quality, Economy),(Cost,60),(Depart, Paris), (Arrive, Monte Carlo)},
{(Business, Hotel), (Company, Continental), (Quality, 2*), (Cost, 250),
(Location, Monte Carlo), (Nights, 2)}, {(Business, Airline), (Company,
KLM), (Quality, Economy), (Cost, 50), (Depart, Monte Carlo), (Arrive,
London)}, {(Business, Hotel), (Company, Continental), (Quality, 3*),
(Cost, 250), (Location, London), (Nights, 4)}, {(Business, Airline),
(Company, Air Espana), (Quality, First), (Cost, 90), (Depart, London),
(Arrive, Edinburgh)}]
Figure 2.22: Semantic Filter: Shows the numerical semantic descriptions, of the simulated
services (Agents) and user requests, in a human readable form. The semantic filter translates
numerical semantic descriptions for one community within the user base, showing it in the
context of the travel industry. The simulation still operated on the numerical representation for
operational efficiency, but the semantic filter essentially assigns meaning to the numbers.
The simulation of the Digital Ecosystem complies with the Ecosystem-Oriented Architecture
defined in the previous section, but there was the possibility of model error in the business
ecosystems of the user base (Small and Medium sized Enterprises from Digital Business
Ecosystems [222]), because while the abstract numerical definition for the simulated semantic
descriptions, of the services and requests the users provide, makes it widely applicable, it was
2.3. Simulation and Results ,
71
unclear that it could accurately represent business services. So we created a semantic filter to
show the numerical semantic descriptions, of the simulated services (Agents) and user requests,
in a human readable form. The basic properties of any business process are cost, quality, and
time [68]; so this was followed in the semantic filter. The semantic filter translates numerical
semantic descriptions for one community within the user base, showing it in the context of
the travel industry, as shown in Figure 2.22. The simulation still operated on the numerical
representation for operational efficiency, but the semantic filter essentially assigns meaning to
the numbers. The output from the semantic filter, in Figure 2.22, shows that the numerical
semantic descriptions are a reasonable modelling assumption that abstracts sufficiently rich
textual descriptions of business services.
2.3.5 Evolutionary Dynamics
Figure 2.23: Graph of Fitness in the Evolutionary Process: This shows both the maximum
and average fitness increasing over the generations of a typical Population, and as expected
the average fitness remains below the maximum fitness because of variation in the Population
[110], showing that the evolutionary processes, which construct order in the Digital Ecosystem,
are operating satisfactorily.
We plotted the fitness of the evolutionary process for a typical Population, to ensure that
the core process that creates order within the Digital Ecosystem was operating satisfactorily.
The graph in Figure 2.23 shows both the maximum and average fitness increasing over the
020406080100100200300FitnessGenerationMaximumFitnessAverageFitness 72
, Chapter 2. Creation of Digital Ecosystems
generations of a typical Population, and as expected the average fitness remains below the
maximum fitness because of variation in the Population [110], showing that the evolutionary
processes, which construct order in the Digital Ecosystem, are operating satisfactorily.
2.3.6 Ecological Succession
We then compared some of the Digital Ecosystem’s dynamics with those of biological
ecosystems, to determine if it had been imbibed with the properties of biological ecosystems.
A biological ecosystem develops from a simpler to a more mature state, by a process of
succession, where the genetic variation of the populations changes with time [29]. So,
it
becomes increasingly more complex through this process of succession, driven by the evolution
of the populations within the ecosystem [59]. Equivalently, the Digital Ecosystem’s increasing
complexity comes from the Agent Populations being evolved to meet the dynamic selection
pressures created by the user requests.
Figure 2.24: Ecological Succession (modified from [69]): The formation of a mature ecosystem
is the slow, predictable, and orderly changes in the composition and structure of an ecological
community, for which there are defined stages in the increasing complexity [29], as shown. So,
it becomes increasingly more complex through this process of succession, driven by the evolution
of the populations within the ecosystem [59].
The formation of a mature ecosystem, ecological succession, is the slow, predictable, and orderly
changes in the composition and structure of an ecological community, for which there are defined
2.3. Simulation and Results ,
73
stages in the increasing complexity [29], as shown in Figure 2.24. Succession may be initiated
either by the formation of a new, unoccupied habitat (e.g., a lava flow or a severe landslide)
or by some form of disturbance (e.g. fire, logging) of an existing community. The former case
is often called primary succession, and the latter secondary succession [29]. The trajectory of
ecological change can be influenced by site conditions, by the interactions of the species present,
and by more stochastic factors such as availability of colonists or seeds, or weather conditions
at the time of disturbance. Some of these factors contribute to predictability of successional
dynamics; others add more probabilistic elements [112]. Trends in ecosystem and community
properties of succession have been suggested, but few appear to be general. For example, species
diversity almost necessarily increases during early succession upon the arrival of new species,
but may decline in later succession as competition eliminates opportunistic species and leads
to dominance by locally superior competitors [59]. Net Primary Productivity3, biomass, and
trophic level properties all show variable patterns over succession, depending on the particular
system and site [112]. Generally, communities in early succession will be dominated by fast-
growing, well-dispersed species, but as the succession proceeds these species will tend to be
replaced by more competitive species [29].
We then considered existing theories of complexity for ecological succession and how it would
apply to Digital Ecosystems, seeking a high-level understanding that would apply equally to
both biological and digital ecosystems. As succession leads communities, of an ecosystem,
to states of dynamic equilibrium 4 within the environment [29], the complexity has to increase
initially or there would be no ecosystem, and presumably this increase eventually stops, because
there must be a limit to how many species can be supported. The period in between is more
complicated. If we consider the neutral biodiversity theory [136], which basically states network
aspects of ecosystems are negligible, we would probably get a relatively smooth progression,
because although you would get occasional extinctions, they would be randomly isolated events
whose frequency would eventually balance arrivals, not self-organised crashes like in systems
theory.
In systems theory [91], when a new species arrives in an ecological network, it can
create a positive feedback loop that destabilises part of the network and drives some species
to extinction. Ecosystems are constantly being perturbed, so it is reasonable to assume that a
species that persists will probably be involved in a stabilising interaction with other species. So,
3 Net Primary Productivity is the net flux of carbon from the atmosphere into green plants per unit time [168].
4 Dynamic Equilibrium is when opposing forces of a system are proceeding at the same rate, such that its state
is unchanging with time [29].
74
, Chapter 2. Creation of Digital Ecosystems
the whole ecological network evolves to resist invasion. That would lead to a spiky succession
process, perhaps getting less spiky over time.
So, which theory is more applicable to the Digital Ecosystem depends on the extent that
a species in the ecosystem acts independently, competing entities (smooth succession) [136]
versus tightly co-adapted ecological partners (spiky succession) [91]. Our Digital Ecosystem
despite its relative complexity is quite simple compared to biological ecosystems. It has the
essential and fundamental processes, but no sophisticated social mechanisms. Therefore, the
smooth succession of the neutral biodiversity theory [136] is more probable.
As the increasing complexity of the Digital Ecosystem comes from its evolving Agent
Populations responding to user requests, the effectiveness of the evolved Agent-sequences
(responses) can provide a measure of its complexity over time. So, in simulation we measured
the effectiveness of its responses over a thousand user requests, i.e. until it had reached a mature
state like a biological ecosystem [29], and graphed a typical run in Figure 2.25. The range and
diversity of Agents at initial deployment were such that 70% fulfilment of user requests was
Figure 2.25: Graph of Succession in the Digital Ecosystem: The formation of a mature biological
ecosystem, ecological succession, is a relatively slow process [29], and the simulated Digital
Ecosystem acted similarly in reaching a mature state. Still, at the end of the simulation run,
the Agent-sequences had evolved and migrated over an average of only ten user requests per
Habitat, and collectively had already reached near 70% effectiveness for the user base.
0204060801002004006008001000Complexity(%matchtouserrequest)Time(userrequestevents) 2.3. Simulation and Results ,
75
possible, increasing to 100% fulfilment as more Agents were deployed. The Digital Ecosystem
performed as expected, adapting and improving over time, reaching a mature state as seen
in the graph of Figure 2.25. The succession of the Digital Ecosystem followed the smooth
succession of the neutral biodiversity theory [136], shown by the tight distribution and equal
density of the points around the best fit curve of the graph in Figure 2.25. The variation in the
percentage responsiveness, over the successive user request events, came from the differential
rates of adaption at the Habitats. Still, by the end of the simulation run, the Agent-sequences
had evolved and migrated over an average of only ten user requests per Habitat, and collectively
had already reached near 70% effectiveness for the user base. The formation of a mature
biological ecosystem, ecological succession, is a relatively slow process [29], and the simulated
Digital Ecosystem acted similarly in reaching a mature state.
2.3.7 Species Abundance
In ecology, relative abundance is a measure of the proportion of all organisms in a community
belonging to a particular species [30]. A relative abundance distribution provides the inequalities
in population size within an ecosystem and therefore an indicator of biodiversity, with the
distribution of most biological ecosystems taking a log-normal form [30]. So, for Digital
Ecosystems this measures globally the abundance of different solutions relative to one another.
A snapshot of the Agents (organisms) within the Digital Ecosystem, for a typical simulation
run, was taken after a thousand user requests, i.e. once it had reached a mature state.
In
biology a species is a series of populations within which significant gene flow can and does
occur, so groups of organisms showing a very similar genetic makeup [168]. We therefore chose
to define species within Digital Ecosystems similarly, as a grouping of genetically similar digital
organisms (based on their semantic descriptions), with no more than 10% variation within the
species group. Relative abundance was calculated for each species and grouped by frequency
in Figure 2.26. In contrast to expectations from biological ecosystems, relative abundance in
the Digital Ecosystem did not conform to the expected log-normal [30]. We suggest that the
high frequency for the lowest relative abundance was caused by the dynamically re-configurable
topology of the Habitat network, which allowed species of small abundance to survive as their
respective Habitats were clustered by the Digital Ecosystem. Therefore, it also most likely
skewed the other frequencies of the relative abundance measure.
76
, Chapter 2. Creation of Digital Ecosystems
Figure 2.26: Graph of Relative Abundance in the Digital Ecosystem: Relative abundance is
a measure of the proportion of all organisms in a community belonging to a particular species
[30]. A relative abundance distribution provides the inequalities in population size within an
ecosystem and therefore an indicator of biodiversity, with the distribution of most biological
ecosystems taking a log-normal form [30]. However, the Digital Ecosystem did not conform to
the expected log-normal.
2.3.8 Species-Area Relationship
In ecology the species-area relationship measures diversity relative to the spatial scale, showing
the number of species found in a defined area of a particular habitat or habitats of different
areas [288], and is commonly found to follow a power law in biological ecosystems [288]. For
Digital Ecosystems this relationship represents how similar solutions are to one another at
different Habitat scales.
Again, a snapshot of the Agents (organisms) within the Digital Ecosystem, for a typical
simulation run, was taken once it had reached a mature state, after a thousand user requests.
For this experiment, we assumed each Habitat to have an area of one unit. Then, the number
of species, at n randomly chosen Habitats, was measured, where n ranged between one and
a hundred. For each n, ten sets of measurements were taken at different random sets of
Habitats to calculate averaged results, and the log10 values of these results are depicted in the
graph of Figure 2.27. The distribution of species diversity over a spatial scale in the Digital
0102030405060702.32.42.52.62.72.82.93.03.13.23.33.43.53.63.73.83.94.04.14.24.34.4FrequencyLogofRelativeAbundance 2.4. Summary and Discussion ,
77
Figure 2.27: Graph of Species-Area in the Digital Ecosystem: In ecology the species-area
relationship measures diversity relative to the spatial scale, showing the number of species
found in a defined area of a particular habitat or habitats of different areas [288], and is
commonly found to follow a power law in biological ecosystems, which the Digital Ecosystem
also demonstrates.
Ecosystem demonstrates behaviour similar to biological ecosystems, also following a power law
[288]. However, diversity at fine spatial scales appears to be lower than predicted by the line of
best fit. This may be explained by higher specialisation at some Habitats, making them more
like micro-habitats in terms of a reduced species diversity [168].
2.4 Summary and Discussion
We started by reviewing existing digital ecosystems, and then introduced biomimicry in
computing, Nature Inspired Computing, to create a definition that could be called the
digital counterpart of biological ecosystems. Then, by comparing and contrasting the relevant
theoretical ecology with the anticipated requirements of Digital Ecosystems, we examined
how ecological features may emerge in some systems designed for adaptive problem solving.
Specifically, we suggested that Digital Ecosystems, like a biological ecosystems, will consist of
self-replicating agents that interact both with one another and with an external environment
[29]. Population dynamics and evolution, spatial and network interactions, and complex
1.522.5300.511.52LogofNumberofSpeciesLogofArea 78
, Chapter 2. Creation of Digital Ecosystems
dynamic fitness landscapes will all influence the behaviour of these systems. Many of these
properties can be understood via well-known ecological models [185, 136], with a further body
of theory that treats ecosystems as Complex Adaptive Systems [173]. These models provide
a theoretical basis for the occurrence of self-organisation, in digital and biological ecosystems,
resulting from the interactions among the agents and their environment, leading to complex non-
linear behaviour [185, 136, 173]; and it is this property that provides the underlying potential
for scalable problem-solving in digital environments. Based on the theoretical ecology, we
considered fields from the domain of computer science, relevant in the creation of Digital
Ecosystems. As we required the digital counterparts for the behaviour and constructs of
biological ecosystems, and not their simulation or emulation, we considered parallels using
existing and developing technologies to provide their equivalents. This included elements from
mobile agent systems [249] to provide a parallel to the agents of biological ecosystems and
their migration to different habitats, and distributed evolutionary computing [178] and Service-
Oriented Architectures [228] for the distribution and evolution of these migrating agents in
evolving populations.
Our efforts culminated in the definition of Ecosystem-Oriented Architectures for the creation
of Digital Ecosystems, where the Digital Ecosystem supports the automatic combining of
numerous Agents (which represent services), by their interaction in evolving Populations to meet
user requests for applications, in a scalable architecture of distributed interconnected Habitats.
Agents travel along the peer-to-peer connections; in every node (Habitat) local optimisation
is performed through an evolutionary algorithm, where the search space is determined by the
Agents present at the node. The sharing of Agents between Habitats ensures the system
is scalable, while maintaining a high evolutionary specialisation for each user. The network
of interconnected Habitats is equivalent to the physical environment of biological ecosystems
[29] and - combined with the Agents, the Populations, the Agent migration for distributed
evolutionary computing, and the environmental selection pressures provided by the user base -
the union of the Habitats creates the Ecosystem-Oriented Architecture of a Digital Ecosystem.
Continuous and varying user requests for applications provide a dynamic evolutionary pressure
on the Agent-sequences, which have to evolve to better fulfil those requests, and without which
there would be no driving force to the evolutionary self-organisation of the Digital Ecosystem.
This represents a novel, cutting-edge approach to distributed evolutionary computing, because
instead of having multiple populations sharing solutions to find the optimal solution for
one problem, there are multiple populations to find optimal solutions for multiple similar
2.4. Summary and Discussion ,
79
problems. The business ecosystem of Small and Medium sized Enterprises from Digital Business
Ecosystems [222] was considered as an example user base, because of their adoption of the
ecosystems paradigm, and because our use of Service-Oriented Architectures in defining Digital
Ecosystems predisposes them to business-to-business interaction scenarios. We have also
dealt with critical issues which would otherwise cripple our complex system and prevent it
from providing a scalable solution, like bloat in evolutionary processes and points-of-failure in
networking topologies. In essence, we are making a system greater than the sum of its parts,
expected to show emergent and complex behaviour that cannot be predicted until it is created.
In simulation, we compared the Digital Ecosystem’s dynamics to those of biological ecosystems.
The ecological succession, measured by the responsiveness to user requests, conformed to
expectations from biological ecosystems [136]:
improving over time, before approaching a
plateau. As the evolutionary self-organisation of an ecosystem is a slow process [29], even
the accelerated form present in Digital Ecosystems reached only 70% responsiveness, showing
potential for improvement.
In the species abundance experiment the Digital Ecosystem did
not conform to the log-normal distribution usually found in biological ecosystems [30]. The
high frequency for the lowest relative abundance was probably caused by the dynamically
re-configurable topology of the Habitat network, which allowed species of small abundance
to survive as their Habitats were clustered by the Digital Ecosystem.
In the species-area
experiment, which measures diversity relative to spatial scale, the Digital Ecosystem did follow
the power law commonly found in biological ecosystems [288]. The species diversity at fine
spatial scales was lower than predicted by the line of best fit, and may be explained by
the high specialisation at some Habitats, making them more like micro-habitats, including
a reduced species diversity [168]. The majority of the experimental results indicate that Digital
Ecosystems behave like their biological counterparts, and suggest that incorporating ideas from
theoretical ecology can contribute to useful self-organising properties in Digital Ecosystems,
which can assist in generating scalable solutions to complex dynamic problems.
Creating the digital counterpart of biological ecosystems was not without apparent compro-
mises; the temporary one-to-one genotype-phenotype mapping for Agents, the information-
centric dynamically re-configurable network topology, and the species abundance result are
inconsistent with biological ecosystems. Initially, any newly deployed Agent has a one-to-one
genotype-phenotype mapping [281], until sufficient usage (phenotype) information is amassed
for use in fitness functions. While the use of such a mapping is undesirable, it is temporary,
and necessary to allow the Digital Ecosystem to operate. The Digital Ecosystem requires a re-
80
, Chapter 2. Creation of Digital Ecosystems
Figure 2.28: Hypothetical Abstract Ecosystem Definition: If there were an abstract ecosystem
class in the Unified Modelling Language, then the Digital Ecosystem and biological ecosystem
classes would both inherit from the abstract ecosystem class, but implement its attributes
differently.
So, we would argue that the apparent compromises in mimicking biological
ecosystems are actually features unique to Digital Ecosystems.
configurable network topology, to support the constantly changing multi-objective information-
centric selection pressures of the user base. Hence, using the concept of Hebbian learning
[132], Habitat connectivity is dynamically adapted based on the observed migration paths
of the Agents within the Habitat network. The dynamically re-configurable network topology
probably caused the Digital Ecosystem not to conform, in the species abundance experiment, to
the log-normal distribution expected from biological ecosystems [30]. We would argue that these
differences are not compromises, but features unique to Digital Ecosystems. As we discussed
earlier, biomimicry, when done well, is not slavish imitation; it is inspiration using the principles
which nature has demonstrated to be successful design strategies [33]. Hypothetically, if there
were an abstract definition of an ecosystem, defined as an abstract ecosystem class, then
the Digital Ecosystem and biological ecosystem classes would both inherit from the abstract
ecosystem class, but implement its attributes differently, as shown in the Unified Modelling
Language class diagram of Figure 2.28. So, we would argue that the apparent compromises in
mimicking biological ecosystems are actually features unique to Digital Ecosystems.
Service-oriented architectures promise to provide potentially huge numbers of services that
programmers can combine via standardised interfaces, to create increasingly sophisticated and
distributed applications [241]. The Digital Ecosystem extends this concept with the automatic
combining of available and applicable services in a scalable architecture to meet user requests
for applications. This is made possible by a fundamental paradigm shift, from a pull -oriented
approach to a push-oriented approach. So, instead of the pull -oriented approach of generating
applications only upon request in Service-Oriented Architectures [287], the Digital Ecosystem
follows a push-oriented approach of distributing and composing applications pre-emptively, as
well as upon request. Although the use of Service-Oriented Architectures in the definition of
Digital Ecosystems provides a predisposition to business [158], it does not preclude other more
NetworkEcosystemNetwork: EnergyCentricBiological EcosystemNetwork: InformationCentricDigital Ecosystem 2.4. Summary and Discussion ,
81
general uses. The Ecosystem-Oriented Architecture definition of Digital Ecosystems is intended
to be inclusive and interoperable with other technologies, in the same way that the definition
of Service-Oriented Architectures is with grid computing and other technologies [287]. For
example, Habitats could be executed using a distributed processing arrangement, such as grid
computing [287], which would be possible because the Habitat network topology is information-
centric (instead of location-centric).
In this chapter we have determined the fundamentals for a new class of system, Digital
Ecosystems, created through combining understanding from theoretical ecology, evolutionary
theory, Multi-Agent Systems, distributed evolutionary computing, and Service-Oriented
Architectures. The word ecosystem is more than just a metaphor since it is the digital
counterpart of biological ecosystems. Therefore, Digital Ecosystems have their desirable
properties, such as scalability and self-organisation, and are complex systems that show
emergent behaviour, since they are more than the sum of their constituent parts. Once we
have further investigated its self-organising properties in the next chapter, Chapter 3, we will
attempt its optimisation, for which the experimental results have shown there is potential, in
the following chapter, Chapter 4.
82
Chapter 3
Investigation of Digital Ecosystems
In this chapter we investigate the self-organising behaviour of Digital Ecosystems, because a
primary motivation for our research is to exploit the self-organising properties of biological
ecosystems. Over time a biological ecosystem becomes increasingly self-organised through the
process of ecological succession, driven by the evolutionary self-organisation of the populations
within the ecosystem. Analogously, a Digital Ecosystem’s increasing self-organisation comes
from the Agent Populations within being evolved to meet the dynamic selection pressures
created by the requests from the user base. We start by discussing the relevant literature
on self-organisation, including the philosophical meaning of organisation and of self, before
focusing on its application to evolving Agent Populations. The self-organisation of biological
ecosystems is often defined in terms of the complexity, stability, and diversity. So, we studied
further to extend a definition for the complexity, grounded in the biological sciences, called
Physical Complexity; based on statistical physics, automata theory, and information theory,
providing a measure of the quantity of information in an organism’s genome, by calculating the
entropy in a population to determine the randomness in the genome. Next, we investigate and
extend a definition for the stability, originating from the computer sciences, called Chli-DeWilde
stability, which views a Multi-Agent System as a discrete time Markov chain with potentially
unknown transition probabilities. With a Multi-Agent System being considered stable when
its state has converged to an equilibrium distribution. Finally, we investigate a definition for
the diversity, relative to the selection pressures provided by the user requests, considering the
collective self-organised diversity of the evolving Agent Populations relative to the global user
request behaviour. We conclude with a summary and discussion of the achievements, including
the experimental results.
83
84
, Chapter 3.
Investigation of Digital Ecosystems
3.1 Background Theory
Self-organisation is perhaps one of the most desirable features in the systems that we design,
and a primary motivation for our research in Digital Ecosystems is the desire to exploit
the self-organising properties of biological ecosystems [173], which are thought to be robust,
scalable architectures that can automatically solve complex, dynamic problems. Over time
a biological ecosystem becomes increasingly self-organised through the process of ecological
succession [29], driven by the evolutionary self-organisation of the populations within the
ecosystem. Analogously, a Digital Ecosystem’s increasing self-organisation comes from the
Agent Populations within being evolved to meet the dynamic selection pressures created by
the requests from the user base. The self-organisation of biological ecosystems is often defined
in terms of the complexity, stability, and diversity [150], which we will also apply to our Digital
Ecosystems.
It is important for us to be able to understand, model, and define self-organising behaviour,
determining macroscopic variables to characterise this self-organising behaviour of the order
constructing processes within, the evolving Agent Populations. However, existing definitions of
self-organisation may not be directly applicable, because evolving Agent Populations possess
properties of both computing systems (e.g. agent systems) as well as biological systems (e.g.
population dynamics), and the combination of these properties makes them unique. So, to
determine definitions for the self-organising complexity, stability, and diversity we will start by
considering the available literature on self-organisation, for its general properties, its application
to Multi-Agent Systems (the dominant technology in Digital Ecosystems), and its application
to our evolving Agent Populations.
3.1.1 Self-Organisation
Self-organisation has been around since the late 1940s [11], but has escaped general formalisa-
tion despite many attempts [231, 152]. There have instead been many notions and definitions of
self-organisation, useful within their different contexts [133]. They have come from cybernetics
[11, 28, 134], thermodynamics [231], mathematics [171], information theory [282], synergetics
[125], and other domains [170]. The term self-organising is widely used, but there is no generally
accepted meaning, as the abundance of definitions would suggest. Therefore, the philosophy of
self-organisation is complicated, because organisation has different meanings to different people.
3.1. Background Theory ,
85
So, we would argue that any definition of self-organisation is context dependent, in the same
way that a choice of statistical measure is dependent on the data being analysed.
Proposing a definition for self-organisation faces the cybernetics problem of defining system,
the cognitive problem of perspective, the philosophical problem of defining self, and the context
dependent problem of defining organisation [107].
The system in this context is an evolving Agent Population, with the replication of individuals
from one generation to the next, the recombination of the individuals, and a selection pressure
providing a differential fitness between the individuals, which is behaviour common to any
evolving population [29].
Perspective can be defined as the perception of the observer in perceiving the self-organisation
of a system [12, 28], matching the intuitive definition of I will know it when I see it [283],
which despite making formalisation difficult shows that organisation is perspective dependent
(i.e. relative to the context in which it occurs). In the context of an evolutionary system, the
observer does not exist in the traditional sense, but is the selection pressure imposed by the
environment, which selects individuals of the population over others based on their observable
fitness. Therefore, consistent with the theoretical biology [29], in an evolutionary system the
self-organisation of its population is from the perspective of its environment.
Whether a system is self -organising or being organised depends on whether the process causing
the organisation is an internal component of the system under consideration. This intuitively
makes sense, and therefore requires one to define the boundaries of the system being considered
to determine if the force causing the organisation is internal or external to the system. For an
evolving population the force leading to its organisation is the selection pressure acting upon it
[29], which is formed by the environment of the population’s existence and competition between
the individuals of the population [29]. As these are internal components of an evolving Agent
Population [29], it is a self-organising system.
Now that we have defined,
for an evolving Agent Population, the system for which its
organisation is context dependent, the perspective to which it is relative, and the self by which
it is caused, a definition for its self-organisation can be considered. The context, an evolving
Agent Population in its environment, lacks a 2D or 3D metric space, so it is necessary to consider
a visualisation in a more abstract form. We will let a single square,
, represent an Agent,
with colours to represent different Agents. Agent-sequences will therefore be represented by a
86
, Chapter 3.
Investigation of Digital Ecosystems
Figure 3.1: Visualisation of Self-Organisation in Evolving Agent Populations: The number
of Agents, in total and of each colour, is the same in both populations. However, the Agent
Population on the left intuitively shows organisation through the uniformity of the colours across
the Agent-sequences, whereas the population to the right shows little or no organisation.
sequence of coloured squares,
, with a Population consisting of multiple Agent-sequences,
as shown in Figure 3.1.
In Figure 3.1 the number of Agents, in total and of each colour, is the same in both populations.
However, the Agent Population on the left intuitively shows organisation through the uniformity
of the colours across the Agent-sequences, whereas the population to the right shows little or
no organisation. Following biological ecosystems, which defines self-organisation in terms of the
complexity, stability, and diversity relative to the perspective of the selection pressure [150]: the
self-organised complexity of the system is the creation of coherent patterns and structures from
the Agents, the self-organised stability of the system is the resulting stability or instability that
emerges over time in these coherent patterns and structures, and the self-organised diversity of
the system is the optimal variability within these coherent patterns and structures.
3.1.2 Definitions of Self-Organisation
Many alternative definitions have been proposed for self-organisation within populations and
agent systems, with each defining what property or properties demonstrate self-organisation.
So, we will now consider the most applicable alternatives for their suitability in defining the
self-organised complexity, stability, and diversity of an evolving Agent Population.
One possibility would be the ∈-machine definition of evolving populations, which models the
emergence of organisation in pre-biotic evolutionary systems [64]. An ∈-machine consists of
a set of causal states and transitions between them, with symbols of an alphabet labelling
the transitions and consisting of two parts: an input symbol that determines which transition
to take from a state, and an output symbol which is emitted on taking that transition [64].
∈-machines have several key properties [65]: all their recurrent states form a single, strongly
PopulationPopulation(HIGH organisation)(LOW organisation) 3.1. Background Theory ,
87
connected component, their transitions are deterministic in the specific sense that a causal
state with the edge symbol-pair determines the successor state, and an ∈-machine is the
smallest causal representation of the transformation it implements. The ∈-machine definition
of self-organisation also identifies the forms of complexity, stability, and diversity [64], but with
definitions focused on pre-biotic evolutionary systems, i.e.
the primordial soup of chemical
replicators from the origin of life [257]. Complexity is defined as a form of structural-complexity,
measuring the state-machine-based information content of the ∈-machine individuals of a
population [64]. Stability is defined as a meta-machine, a set (composition) of ∈-machines,
that can be regarded as an autonomous and self-replicating entity [64]. Diversity is defined,
using an interaction network, as the variability of interaction in a population [64]. So, while
these definitions of self-organisation are compatible at the higher more abstract level, i.e.
in
the forms of self-organisation present, the deeper definitions of these forms are not applicable
because they are context dependent. As we explained in the previous subsection, definitions of
self-organisation are context dependent, and so the context of pre-biotic evolutionary systems,
to which the ∈-machine self-organisation applies, is very different to the context of an evolving
Agent Population from our Digital Ecosystem. Evolving Agent Populations are defined from
Ecosystem-Oriented Architectures, which have evolutionarily surpassed the context of pre-biotic
evolutionary systems, shown by the necessity of our consideration of the later evolutionary stage
of ecological succession [29] in section 2.3.6.
The Minimum Description Length principle [24] could be applied to the executable components
or semantic descriptions of the Agent-sequences of a Population, with the best model, among
a collection of tentatively suggested ones, being the one that provides the smallest stochastic
complexity. However, the Minimum Description Length principle does not define how to select
the family of model classes to be applied for determining the stochastic complexity [126]. This
problem of model selection is well known and cannot be adequately formalised, and so in
practise selection is based on human judgement and prior knowledge of the kinds of models
previously chosen [126]. Therefore, while models could be chosen to represent the self-organised
complexity, and possibly even the diversity, there is no procedural method for determining these
models, because subjective human intervention is required for model selection on a case-by-case
basis.
The Prugel-Bennett Shapiro formalism models the evolutionary dynamics of a population of
sequences, using techniques from statistical mechanics and focuses on replica symmetry [253].
The individual sequences are not considered directly, but in terms of the statistical properties
88
, Chapter 3.
Investigation of Digital Ecosystems
of the population, using a macroscopic level of description with specific statistical properties
to characterise the population, that are called macroscopics. A macroscopic formulation of
an evolving population reduces the huge number of degrees of freedom to the dynamics of
a few quantities, because a non-linear system of a few degrees of freedom can be readily
solved or numerically iterated [253]. However, since a macroscopic description disregards a
significant amount of information, subjective human insight is essential so that the appropriate
macroscopics are chosen [284]. So, while macroscopics could be chosen to represent the self-
organised complexity, stability, and diversity, there is no procedural method for determining
these macroscopics, because subjective human insight is required for macroscopic selection on
a case-by-case basis.
Kolmogorov-Chaitin complexity defines the complexity of binary sequences by the smallest
possible Universal Turing Machine, algorithm (programme and input) that produces the
sequence [177]. A sequence is said to be regular if the algorithm necessary to produce it
on a Universal Turing Machine is shorter than the sequence itself [177]. A regular sequence
is said to be compressible, whereas its compression, into the most succinct Universal Turing
Machine possible, is said to be incompressible as it cannot be reduced any further in length
[177]. A random sequence is said to be incompressible, because the Universal Turing Machine
to represent it cannot be shorter than the random sequence itself [177]. This intuitively
makes sense for algorithmic complexity, because algorithmically regular sequences require
a shorter programme to produce them. So, when measuring a population of sequences,
the Kolmogorov-Chaitin complexity would be the shortest Universal Turing Machine to
produce the entire population of sequences. However, Chaitin himself has considered the
application of Kolmogorov-Chaitin complexity to evolutionary systems, and realised that
although Kolmogorov-Chaitin complexity represents a satisfactory definition of randomness
in algorithmic information theory, it is not so useful in biology [48]. For evolving Agent
Populations the problem manifests itself most significantly when the Agents are randomly
distributed within the Agent-sequences of the Population, having maximum Kolmogorov-
Chaitin complexity, instead of the complexity it ought to have of zero. This property makes
Kolmogorov-Chaitin complexity unsuitable as a definition for the self-organised complexity of
an evolving Agent Population.
A definition called Physical Complexity can be estimated for a population of sequences,
calculated from the difference between the maximal entropy of the population, and the actual
entropy of the population when in its environment [8]. This Physical Complexity, based
3.1. Background Theory ,
89
on Shannon’s entropy of information, measures the information in the population about its
environment, and therefore is conditional on its environment. It can be estimated by counting
the number of loci that are fixed for the sequences of a population [5]. Physical Complexity
would therefore be suitable as a definition of the self-organised complexity. However, a possible
limitation is that Physical Complexity is currently only formulated for populations of sequences
with the same length.
Self-Organised Criticality in evolution is defined as a punctuated equilibrium in which the
population’s critical state occurs when the fitness of the individuals is uniform, and for which
an avalanche, caused by the appearance and spread of advantageous mutations within the
population, temporarily disrupts the uniformity of individual fitness across the population [17].
Whether an evolutionary process displays Self-Organised Criticality remains unclear. There
are those who claim that Self-Organised Criticality is demonstrated by the available fossil
data [292], with a power law distribution on the lifetimes of genera drawn from fossil records,
and by artificial life simulations [4], again with a power law distribution on the lifetimes of
competing species. However, there are those who feel that the fossil data is inconclusive,
and that the artificial life simulations do not show Self-Organised Criticality, because the key
power law behaviour in both can be generated by models without Self-Organised Criticality
[229]. Also, the Self-Organised Criticality does not define the resulting self-organised stability
of the population, only the organisation of the events (avalanches) that occur in the population
over time.
Evolutionary Game Theory [330] is the application of models inspired from population genetics
to the area of game theory, which differs from classical game theory [103] by focusing on the
dynamics of strategy change more than the properties of individual strategies. In Evolutionary
Game Theory, agents of a population play a game, but instead of optimising over strategic
alternatives, they inherit a fixed strategy and then replicate depending on the strategy’s payoff
(fitness) [330]. The self-organisation found in Evolutionary Game Theory is the presence of
stable steady states, in which the genotype frequencies of the population cease to change over
the generations. This equilibrium is reached when all the strategies have the same expected
payoff, and is called a stable steady state, because a slight perturbing will not cause a move
far from the state. An evolutionary stable strategy leads to a stronger asymptotically stable
state, as a slight perturbing causes only a temporary move away from the state before returning
[330]. So, Evolutionary Game Theory is focused on genetic stability between competing between
90
, Chapter 3.
Investigation of Digital Ecosystems
individuals, rather than the stability of the population as a whole, which therefore limits its
suitability for the self-organised stability of an evolving Agent Population.
Multi-Agent Systems are the dominant computational technology in the evolving Agent
Populations, and while there are several definitions of self-organisation [245, 188, 308, 81] and
stability [216, 331, 238] defined for Multi-Agent Systems, they are not applicable primarily
because of the evolutionary dynamics inherent in the context of evolving Agent Populations.
Whereas Chli-DeWilde stability of Multi-Agent Systems [53] may be suitable, because it
models Multi-Agent Systems as Markov chains, which are an established modelling approach
in evolutionary computing [272]. A Multi-Agent System is viewed as a discrete time Markov
chain with potentially unknown transition probabilities, in which the agents are modelled as
Markov processes, and is considered to be stable when its state has converged to an equilibrium
distribution [53]. Chli-DeWilde stability provides a strong notion of self-organised stability over
time, but a possible limitation is that its current formulation does not support the necessary
evolutionary dynamics.
The main concept in Mean Field Theory is that for any single particle the most important
contribution to its interactions comes from its neighbouring particles [244]. Therefore, a
particle’s behaviour can be approximated by relying upon the mean field created by its
neighbouring particles [244], and so Mean Field Theory could be suitable as a definition for the
self-organised diversity of an evolving Agent Population. Naturally, it requires a neighbourhood
model to define interaction between neighbours [244], and is therefore easily applied to domains
such as Cellular Automata [124]. While a neighbourhood model
is feasible for biological
populations [95], evolving Agent Populations lack such neighbourhood models based on a 2D
or 3D metric space, with the only available neighbourhood model being a distance measure
on a parameter space measuring dissimilarity. However, this type of neighbourhood model
cannot represent the information-based interactions between the individuals of an evolving
Agent Population, making Mean Field Theory unsuitable as a definition for the self-organised
diversity of an evolving Agent Population.
3.2 Complexity
A definition for the self-organised complexity of an evolving Agent Population should define the
creation of coherent patterns and structures from the Agents within, with no initial constraints
3.2. Complexity ,
91
from modelling approaches for the inclusion of pre-defined specific behaviour, but capable of
representing the appearance of such behaviour should it occur.
None of the proposed definitions are directly applicable for the self-organised complexity of
an evolving Agent Population. The ∈-machine modelling [64] is not applicable, because it is
only defined within the context of pre-biotic populations. Neither is the Minimum Description
Length principle [24] or the Prugel-Bennett Shapiro formalism [253], because they require the
involvement of subjective human judgement at the critical stage of model and quantifier selection
[126, 284]. Kolmogorov-Chaitin complexity [48] is also not applicable as randomness is given
maximum complexity.
Physical Complexity [8] fulfils abstractly the required definition for the self-organised complexity
of an evolving Agent Population, estimating complexity based upon the individuals of a
population within the context of their environment. However,
its current formulation is
problematic, primarily because it is only defined for populations of fixed length, but as this is
not a fundamental property of its definition [8] it should be feasible to redefine and extend it
as needed. So, the use of Physical Complexity as a definition for the self-organised complexity
of evolving Agent Populations will be investigated further to determine its suitability.
3.2.1 Physical Complexity
Physical Complexity was born [5] from the need to determine the proportion of information in
sequences of DNA, because it has long been established [307] that the information contained
is not directly proportional to the length, known as the C-value enigma/paradox [120].
Understanding DNA requires knowing the environment (context) in which it exists, which
may initially appear obvious as DNA is considered to be the language of life [279] and the
purpose of life is to procreate or replicate [70]. Virtually all activities of biological life-forms
are towards this aim [70], with a few exceptions (e.g. suicide before procreation), and to achieve
replication requires resources, energy and matter to be harvested [192]. So, for any individual
the environment represents the problem of extracting energy for replication, and so their DNA
sequence represents a solution to this problem. Furthermore, an individual DNA solution is not
necessarily a simple inverse of the problem that the environment represents, with forms of life
having evolved specialised, specific and effective ways (niches) to acquire the necessary energy
and matter for replication [168]. Even with this understanding it would seem we still need
to define the environment to be able to distinguish the information from the redundancy in a
92
, Chapter 3.
Investigation of Digital Ecosystems
solution (DNA sequence). However, because Physical Complexity analyses a group of solutions
to the same problem, the consistency between the different solutions shows the information, and
the differences the redundancy [6]. Entropy, a measure of disorder [321], is used to determine the
redundancy from the information in a population of solutions. Physical Complexity therefore
provides a context-relative definition for the self-organised complexity of a population without
needing to define the context (environment) explicitly [7].
Physical Complexity was derived [7] from the notion of conditional complexity defined by
Kolmogorov, which is different from traditional Kolmogorov complexity and states that the
determination of complexity of a sequence is conditional on the environment in which the
sequence is interpreted [177]. In contrast, traditional Kolmogorov-Chaitin (KC) complexity is
only conditional on the implicit rules of mathematics necessary to interpret a programme on
the tape of a Turing Machine (TM), and nothing else [177]. So, if we consider a TM that takes
a tape e as input (which represents its physical environment), including the particular rules of
mathematics of this world ; without such a tape, this TM is incapable of computing anything,
except for writing to the output what it reads in the input. Thus, without tape e all sequences
s have maximal KC-complexity, because there is nothing by which to determine regularity [7].
However, conditional complexity can be stated as the length of the smallest programme that
computes sequence s from an environment e,
K(se) = min{p : s = CT (p, e)} ,
(3.1)
where CT (p, e) denotes the result of running programme p on Turing Machine T with the input
sequence e [7]. This is not yet Physical Complexity, but rather, it is the smallest programme
that computes the sequence s from an environment e, in the limit of sequences of infinite length,
containing only the bits that are entirely unrelated to e, since, if they were not, they could be
obtained from e with a programme of a size tending to zero [7]. The Physical Complexity
K(s : e) can now be defined as the number of bits that are meaningful in sequence s (that can
be obtained from e with a programme of vanishing size), and is given by the mutual complexity
[154],
K(s : e) = K(s∅) − K(se),
(3.2)
where K(s∅) is the unconditional complexity with an empty input tape, e ≡ ∅ [7]. This is
different from the Kolmogorov complexity, because in Kolmogorov’s construction the rules of
mathematics were given to the TM [177]. As argued above, every sequence s is random if no
3.2. Complexity ,
93
environment e is specified, as non-randomness can only exist for a specific world or environment.
Thus, K(s∅) is always maximal,
K(s∅) = s,
(3.3)
and is given by the length of s [7]. So (3.2) represents the length of the sequence s, minus
those bits that cannot be obtained from e. So, conversely (3.2) represents the number of bits
that can be obtained in a sequence s, by a computation with vanishing programme size, from
e. Thus, K(s : e) represents the Physical Complexity of s [7]. The determination of the
Physical Complexity, K(s : e), of a sequence s with a description of the environment e is not
practical. Meaning that it cannot generally be determined by inspection, because its impossible
to determine which, and how many, of the bits of sequence s correspond to information about
the environment e. The reason is that we are generally unaware of the coding used to code
information about e in s, and therefore coding and non-coding bits look entirely alike [7].
However, it is possible to distinguish coding from non-coding bits if we are given multiple
copies of sequences that have adapted to the environment, or more generally, if a statistical
ensemble (population) of sequences is available to us. Then, coding bits are revealed by non-
uniform probability distributions across the population (conserved sites), whereas random bits
have uniform distributions (volatile sites) [7]. The determination of complexity then becomes
an exercise in information theory, because the average complexity (cid:104)K(cid:105), in the limit of infinitely
long strings, tends to the entropy of the ensemble of strings S1 [343],
(cid:88)
s∈S
(cid:104)K(s)(cid:105)S =
p(s)K(s) ≈ H(S),
(3.4)
(3.5)
where H is defined from Shannon’s (information) entropy [186], and is given by
H(S) = logn(S),
where n is the number of symbols available for encoding. If each symbol is equally probable,
we can rewrite the above function as
H(S) = −logn(1/S)
= −logn(p),
(3.6)
1 This holds for near-optimal codings. For strings s that do not code perfectly we have (cid:104)K(cid:105) ≥ H [342].
94
, Chapter 3.
Investigation of Digital Ecosystems
where p is the probability of occurrence of any one of the symbols. For a source that outputs
an infinite sequence of bits, to communicate a finite set of symbols S, Shannon generalised the
above function to express an average symbol length [186]. This derivation is easier to see for a
large, but finite, number of symbols N ,
S(cid:80)
S(cid:80)
i=1
=
Ni [− logN (1/Si)]
S(cid:80)
[logN (1/Si)] = − S(cid:88)
Ni
i=1
Ni
N
i=1
H(S) =
i=1
.
.
= − S(cid:88)
i=1
Ni [− logN (1/Si)]
N
pi logN (pi),
(3.7)
where Ni is the number of occurrences of the symbol Si. So, given (3.4) and (3.7), the average
complexity of the sequences s of a population S, (cid:104)K(s)(cid:105)S, tends to the entropy of the sequences
s in the ensemble S [7],
p(s) log p(s).
(3.8)
(cid:104)K(s)(cid:105)S = −(cid:88)
s∈S
(3.8) remains consistent with (3.3) as the determination of K(s∅), sequence s without an
environment e, must equal the sequence’s length s, because Shannon’s formula for entropy
is an average logarithmic measure of the symbol sets [186], and so the maximum entropy of a
population is equivalent to the length of the sequences in the population, Hmax(S) = s. Indeed,
if nothing is known about the environment to which a sequence s pertains, then according to
the principle of indifference 2, the probability distribution p(s) must be uniformly random.
However, if an environment e is given we have some information about the system, and the
probability distribution will be nonuniform. Indeed, it can be shown that for every probability
distribution p(se), to find sequence s given environment e, we have
H(Se) ≤ H(S∅) = s,
(3.9)
because of the concavity of Shannon entropy [7]. So, the difference between the maximal entropy
H(S∅) = s and H(Se), according to the construction outlined above, represents the average
number of bits in sequence s taken from the population S that can be obtained by zero-length
universal programmes from the environment e. Therefore, the average mutual complexity of
2 The principle of indifference states that if there are n > 1 mutually exclusive and collectively exhaustive
possibilities, which are indistinguishable except for their names then each possibility should be assigned an
equal probability 1
n [143].
3.2. Complexity ,
sequences s in a population S, given an environment e, is
(cid:88)
(cid:104)K(s : e)(cid:105)S =
s∈S
p(s)K(s : e)
≈ H(S∅) − H(Se)
≡ I(Se),
95
(3.10)
where I(Se) is the information about the environment e stored in the population S, which we
identify with the Physical Complexity [7]. To estimate I(Se) it is necessary to estimate the
entropy H(Se) using a representative population of sequences S for a given environment e, by
summing, over the sequences s of the population S, the probability p(se) multiplied by the
logarithm of the probability p(se),
p(se) log p(se).
(3.11)
The entropy H(Se) can be estimated by summing the per-site H(i) entropies of the sequence,
H(Se) = −(cid:88)
s∈S
s(cid:88)
H(Se) ≈
H(i),
(3.12)
where i is a site in the sequence s [7]. Random sites are identified by a nearly uniform probability
distribution, and contribute positively to the entropy, whereas non-random sites (which have
i=1
strongly peaked distributions) contribute very little [7]. So, the Physical Complexity, the
average mutual complexity of sequences s in a population S for an environment e, (cid:104)K(s : e)(cid:105)S,
abbreviated as C, is the maximal entropy H(S∅) minus the sum of the per-site entropies,
C = H(S∅) −
H(i).
(3.13)
s(cid:88)
i=1
If the sequences s are constructed from an alphabet, a set D, then the per site entropy H(i)
for the sequences is
pd(i) logD pd(i),
(3.14)
H(i) = −(cid:88)
d∈D
where i is a site in the sequences ranging between one and the length of the sequences (cid:96), D is
the alphabet of characters found in the sequences, and pd(i) is the probability that site i (in the
for each site i equalling one, (cid:80)
sequences) takes on character d from the alphabet D, with the sum of the pd(i) probabilities
pd(i) = 1 [7]. Taking the log to the base D conveniently
d∈D
96
, Chapter 3.
Investigation of Digital Ecosystems
normalises H(i) to range between zero and one,
0 ≤ H(i) ≤ 1.
If the site i is identical across the population it will have no entropy,
Hmin(i) = 0.
(3.15)
(3.16)
If the content of site i is uniformly random, i.e. the pd(i) probabilities all equal to 1D, it will
have maximum entropy,
Hmax(i) = 1.
(3.17)
When the entropy of H(i) is at its minimum of zero, then the site i holds information, as
every sample shows the same character of the alphabet. When the entropy of H(i) is at its
maximum of one, the character found in the site i is uniformly random and therefore holds no
information. So, the amount of information is the maximal entropy of the site (3.17) minus the
actual per-site entropy (3.14) [7],
I(i) = Hmax(i) − H(i)
= 1 − H(i).
(3.18)
DNA, whose sequence encodes the genetic information of living organisms [168], was the original
driver for the creation of Physical Complexity [5], and so is a good example upon which to
demonstrate the definition. DNA sequences are made up from four nucleotides, Adenosine (A),
Figure 3.2: DNA Samples from a Population: DNA sequences are made up from four
nucleotides, Adenosine (A), Thymine (T), Cytosine (C) and Guanine (G). The nucleotides
always pair as follows, Adenosine with Thymine, and Cytosine with Guanine. So, DNA
sequences can be reduced to a genome sequence showing half of the paired information [168].
site12345678910111213141516171819Sample 1:CGCGATACCTTTTGATTGGSample 2:CGCGATACCTATTGATTGGSample 3:CGCGATACCTGTTGATTGCSample 4:CGCGATACCTCTTGATTGC 3.2. Complexity ,
97
Thymine (T), Cytosine (C) and Guanine (G). The nucleotides always pair as follows, Adenosine
with Thymine, and Cytosine with Guanine. So, DNA sequences can be reduced to a genome
sequence showing half of the paired information [168], and with a sufficiently sized sample
population, the pd(i) probabilities can be estimated by the frequencies of the nucleotides at the
sites. Considering the genome samples, in Figure 3.2, the per-site entropy for site 11 will have
maximum entropy, as the nucleotides (characters of the alphabet) all have equal probability,
H(11) = − A,T,C,G(cid:88)
(cid:18)1
d∈D
= −
pd(11) logD pd(11)
log4
4
1
4
+
1
4
log4
1
4
+
1
4
log4
1
4
+
1
4
log4
1
4
(cid:19)
= 1,
given that the alphabet D equals {A, T, C, G}, and that the probabilities all equal a quarter,
pA(11) = pT (11) = pC(11) = pG(11) = 1
4. As the per-site entropy (randomness) is maximum,
the information content is its minimum of zero,
I(11) = 1 − H(11) = 0.
This intuitively makes sense, as it states that if the site content is random across the population,
then it contains no information. At the other extreme, if we calculate the per-site entropy for
site 16 in Figure 3.2, it will have no entropy,
H(16) = − A,T,C,G(cid:88)
d∈D
pd(16) logD pd(16)
= − (0 log4 0 + 1 log4 1 + 0 log4 0 + 0 log4 0) = 0,
as the nucleotide Thymine has a probability of one, pT (16) = 1, while the other three nucleotides
have a probability of zero, pA(16) = pC(16) = pT (16) = 0. As the per-site entropy is minimum,
the information content is its maximum of one,
I(16) = 1 − H(16) = 1.
This also intuitively makes sense, as it states that if the site is identical across the entire
population (no randomness), then the site holds definitive information. Finally, the per-site
98
, Chapter 3.
Investigation of Digital Ecosystems
entropy for site 19 is at neither extreme, but is entropically in the middle,
H(19) = − A,T,C,G(cid:88)
(cid:18)
d∈D
pd(19) logD pd(19)
= −
0 log4 0 + 0 log4 0 +
(cid:19)
=
1
2
,
1
2
log4
1
2
+
1
2
log4
1
2
as the probabilities pA(19) = pT (19) = 0, and pC(19) = pG(19) = 1
2. Intuitively, this states
that if there is some entropy (randomness) in the samples of the site, then there is only partial
information,
I(19) = 1 − H(19) =
1
2
.
For clarity the length of the sequences s will be abbreviated to (cid:96) [7],
s ≡ (cid:96).
(3.19)
So, the complexity of a population S, of sequences s, is the maximal entropy of the population
(equivalent to the length of the sequences) (cid:96), minus the sum, over the length (cid:96), of the per-site
entropies H(i),
C = (cid:96) − (cid:96)(cid:88)
H(i),
(3.20)
i=1
given (3.13), (3.9) and (3.19) [7]. The equivalence of the maximum complexity to the length
matches the intuitive understanding that if a population of sequences of length (cid:96) has no
redundancy, then their complexity is their length (cid:96).
If G represents the set of all possible genotypes constructed from an alphabet D that are of
length (cid:96), then the size (cardinality) of G is equal to the size of the alphabet D raised to the
length (cid:96),
G = D(cid:96).
(3.21)
For the complexity measure to be accurate, a sample size of D(cid:96) is suggested to minimise
the error [7, 26], but such a large quantity can be computationally infeasible. The definition’s
creator, for practical applications, chooses a population size of roughly 1.29D(cid:96) [8]. We suggest
that a population size of D(cid:96) is sufficient to show any trends present, but that the population
size will fluctuate when simulated, and so a population size slightly larger than D(cid:96) is chosen for
simulations to ensure that the necessary minimum of D(cid:96) is maintained. So, for a population
of sequences S we choose, with the definition’s creator, a computationally feasible population
3.2. Complexity ,
size of D times (cid:96),
99
S ≥ D(cid:96).
(3.22)
The size of the alphabet, D, depends on the domain to which Physical Complexity is applied.
For RNA the alphabet is the four nucleotides, D = {A, C, G, U}, and therefore D = 4 [7].
When Physical Complexity was applied to the Avida simulation software, there was an alphabet
size of twenty-eight, D = 28, as that was the size of the instruction set for the self-replicating
programmes [8].
3.2.2 Extending to Agent Populations
Reformulating Physical Complexity for an evolving Agent Population requires consideration of
the following issues: the mapping of the sequence sites to the Agent-sequences, the managing of
Populations of variable length sequences, and the non-atomicity of Agents leading to clustering
within populations.
3.2.2.1 Mapping Sequence Sites
The first concern is mapping the Population’s Agent-sequences to the sequence sites of Physical
Complexity, with the intuitive approach being to map the sites to the Agents, because they are
the functional unit of processing, the base unit for evolution in the evolving Agent Populations.
Physical Complexity has been applied to RNA sequences [7], and populations of self-replicating
programmes in the artificial life simulator Avida [172]; for the RNA the sites were mapped to
the nucleotides from which it is constructed, and for the artificial life simulator the sites were
mapped to the programme instructions which made up the self-replicating programmes. So,
the only alternative, of mapping the sites to the Agents, would be mapping to the programme
instructions of the executable components of the services that the Agents represent, similarly
to the populations of self-replicating programmes in the artificial life simulator Avida [172].
However, mapping to the executable components in the evolving Agent Populations would be
like mapping to the binary representation of the instruction set in the Avida simulator, or to
the molecules that make up the nucleotides in RNA, which in all cases would be unsuitable
as they are the components that make up the functional units, and not the functional units
themselves. Therefore, mapping the sequence sites of Physical Complexity to the Agents is the
most suitable approach for evolving Agent Populations.
100
, Chapter 3.
Investigation of Digital Ecosystems
3.2.2.2 Variable Length Sequences
Physical Complexity is currently formulated for a population of sequences of the same length [7],
and so we will now investigate an extension to include populations of variable length sequences,
which will include Populations of variable length Agent-sequences. This will require changing
and re-justifying the fundamental assumptions, specifically the conditions and limits upon
which Physical Complexity operates. In (3.20) the Physical Complexity, C, is defined for a
population of sequences of length (cid:96) [7]. The most important question is what does the length (cid:96)
equal if the population of sequences is of variable length? The issue is what (cid:96) represents, which
is the maximum possible complexity for the population [7], which we will call the complexity
(cid:96)(cid:80)
potential CP . The maximum complexity in (3.20) occurs when the per-site entropies sum to
H(i) → 0, as there is no randomness in the sites (all contain information), i.e. C → (cid:96)
zero,
i=1
[7]. So, the complexity potential equals the length,
CP = (cid:96),
(3.23)
provided the population S is of sufficient size for accurate calculations, as found in (3.22), i.e. S
is equal or greater than D(cid:96). For a population of variable length sequences, SV , the complexity
potential, CVP , cannot be equivalent to the length (cid:96), because it does not exist. However, given
the concept of minimum sample size from (3.22), there is a length for a population of variable
length sequences, (cid:96)V , between the minimum and maximum length, such that the number of
per-site samples up to and including (cid:96)V is sufficient for the per-site entropies to be calculated.
So the complexity potential for a population of variable length sequences, CVP , will be equivalent
to its calculable length,
CVP = (cid:96)V .
(3.24)
If (cid:96)V where to be equal to the length of the longest individual(s) (cid:96)max in a population of variable
length sequences SV , then the operational problem is that for some of the later sites, between
one and (cid:96)max, the sample size will be less than the population size SV . So, having the length
(cid:96)v equalling the maximum length would be incorrect, as there would be an insufficient number
of samples at the later sites, and therefore (cid:96)V (cid:54)≡ (cid:96)max. Consider the alternative samples of
DNA sequences shown in Figure 3.3; if the entropy is calculated again for site 19, H(19) = 0,
but there is an insufficient sample size for the estimated probabilities to provide an accurate
calculation. Therefore, the length for a population of variable length sequences, (cid:96)V , is the
3.2. Complexity ,
101
Figure 3.3: Alternative DNA Samples of the Population from Figure 3.2: Calculating the
entropy for site 19 provides a value of zero, but as evident there is an insufficient sample size
for the estimated probabilities to provide an accurate calculation. So, having the length of a
population of variable length sequences equalling the maximum length would be incorrect.
highest value within the range of the minimum (one) and maximum length, 1 ≤ (cid:96)V ≤ (cid:96)max,
for which there are sufficient samples to calculate the entropy. A function which provides the
sample size at a given site is required to specify the value of (cid:96)V precisely,
sampleSize(i : site) : int,
(3.25)
where the output varies between 1 and the population size SV (inclusive). Therefore, the
length of a population of variable length sequences, (cid:96)V , is the highest value within the range of
one and the maximum length for which the sample size is greater than or equal to the alphabet
size multiplied by the length (cid:96)V ,
sampleSize((cid:96)V ) ≥ D(cid:96)V ∧ sampleSize((cid:96)V + 1) < D(cid:96)V ,
(3.26)
where (cid:96)V is the length for a population of variable length sequences, and (cid:96)max is the maximum
length in a population of variable length sequences, (cid:96)V varies between 1 ≤ (cid:96)V ≤ (cid:96)max, D is the
alphabet and D > 0. This definition intrinsically includes a minimum size for populations
of variable length sequences, D(cid:96)V , and therefore is the counterpart of (3.22), which is the
minimum population size for populations of fixed length.
The length (cid:96) used in the limits of (3.14) no longer exists, and therefore (3.14) must be updated;
so, the per-site entropy calculation for variable length sequences will be denoted by HV (i), and
pd(i) logD pd(i),
(3.27)
is,
HV (i) = −(cid:88)
d∈D
site12345678910111213141516171819Sample 1:CGCGATACCTTTTGATTGGSample 2:CGCGATACCTATTGATTGGSample 3:CGCGATACCTGTTGATTSample 4:CGCGATACCTCTTGATT 102
, Chapter 3.
Investigation of Digital Ecosystems
where D is still the alphabet, (cid:96)V is the length for a population of variable length sequences,
with the site i now ranging between 1 ≤ i ≤ (cid:96)V , while the pd(i) probabilities still range between
0 ≤ pd(i) ≤ 1, and still sum to one. It remains algebraically almost identical to (3.14), but
the conditions and constraints of its use will change, specifically (cid:96) is replaced by (cid:96)V . Naturally,
HV (i) ranges between zero and one, as did H(i) in (3.15),
0 ≤ HV (i) ≤ 1,
(3.28)
where i ranges between 1 and (cid:96)V , and so the condition on the site i changes at the upper limit
from (cid:96) to (cid:96)V . As before in (3.16), if a site i is identical across the population, it will have no
entropy,
HVmin(i) = 0,
(3.29)
where, again, i ranges between 1 and (cid:96)V . Analogously to (3.17), if the pd(i) probabilities in
(3.27) are equal, then the site i has maximum entropy.
In effect, the content of the site is
uniformly random and therefore
HVmax(i) = 1
(3.30)
is true for all i, where the pd(i) probabilities are 1D, and where i continues to range between 1
and (cid:96)V . Analogously to (3.18), when the entropy is its minimum of zero then the site i holds
information, as every sample shows the same character of the alphabet. However, when the
entropy is maximum the character found in the site i is uniformly random, and therefore holds
no information. So, the amount of information is the maximal entropy of the site (3.30) minus
the actual per-site entropy (3.27),
IV (i) = HVmax(i) − HV (i)
= 1 − HV (i),
(3.31)
where i again now ranges between 1 and (cid:96)V . Therefore, the complexity for a population of
variable length sequences, CV , is the complexity potential of the population of variable length
sequences minus the sum, over the length of the population of variable length sequences, of the
CV = (cid:96)V − (cid:96)V(cid:88)
per-site entropies (3.27),
where (cid:96)V is the length for the population of variable length sequences, and HV (i) is the entropy
for a site i in the population of variable length sequences.
i=1
HV (i),
(3.32)
3.2. Complexity ,
103
Figure 3.4: Abstract Visualisation for Populations of Variable Length Sequences: The alphabet
} , the maximum length (cid:96)max is 6 and the length for populations of variable
D is the set { ,
length sequences (cid:96)V is calculated as 5 from (3.26). The Physical Complexity and Efficiency
values are consistent with the intuitive understanding one would have for the self-organised
complexity of the sample populations.
,
Physical Complexity can now be applied to populations of variable length sequences, so we
will consider the abstract example populations in Figure 3.4. We will let a single square,
,
represent a site i in the sequences, with different colours to represent the different values.
Therefore, a sequence of sites will be represented by a sequence of coloured squares,
.
}, the maximum length (cid:96)max is 6 and the
Furthermore, the alphabet D is the set { ,
length for populations of variable length sequences (cid:96)V is calculated as 5 from (3.26). The
,
Physical Complexity values in Figure 3.4 are consistent with the intuitive understanding one
would have for the self-organised complexity of the sample populations; the population with
high Physical Complexity has a little randomness, while the population with low Physical
Complexity is almost entirely random.
Using our extended Physical Complexity we can construct a measure showing the use of the
information space, called the Efficiency E, which is calculated by the Physical Complexity CV
over the complexity potential CVP ,
E =
CV
CVP
.
The Efficiency E will range between zero and one,
0 ≤ E ≤ 1,
(3.33)
(3.34)
only reaching its maximum of one when the actual complexity CV equals the complexity
potential CVP , indicating that there is no randomness in the population.
In Figure 3.4 the
populations of sequences are shown with their respective Efficiency values as percentages, and
the values are as one would expect.
The complexity CV (3.32) is an absolute measure, whereas the Efficiency E (3.33) is a relative
measure (based on the complexity CV ). So, the Efficiency E can be used to compare the
PopulationCV=4.420,%E=88.4CV=0.575,%E=11.5Population 104
, Chapter 3.
Investigation of Digital Ecosystems
self-organised complexity of populations, independent of their size, their length, and whether
their lengths are variable or not (as its equally applicable to the fixed length populations of the
original Physical Complexity).
3.2.2.3 Clustering
The self-organised complexity of an evolving Agent Population is the clustering, amassing of
same or similar sequences, around the optimum genome [29]. This can be visualised on a fitness
landscape [335], which shows the combination space (power set) of the alphabet D against the
fitness values from the selection pressure (user request). The Agent-sequences of an evolving
Population will evolve, moving across the fitness landscape and clustering around the optimal
genome at the peak of the global optimum, assuming that its evolutionary process does not
become trapped while clustering over local optima, and as shown in Figure 3.5.
Figure 3.5: 3D Fitness Landscape with a Global Optimum: This shows the combination space
(power set) of the alphabet D against the fitness values from the selection pressure (user request),
resulting in a global optimum. The Agent-sequences of an evolving Population will evolve,
moving across the fitness landscape and clustering around the optimal genome at the peak of
the global optimum.
Clustering is indicated by the Efficiency E tending to its maximum, as the population’s Physical
Complexity CV tends to the complexity potential CVP , because an optimal sequence is becoming
dominant in the population, and therefore increasing the uniformity of the sites across the
population. With a global optimum, the Efficiency E tends to a maximum of one, indicating
that the evolving population of sequences is tending to a set of clusters T of size one,
E =
CV
CVP
→ 1 as T → 1,
(3.35)
3.2. Complexity ,
105
assuming its evolutionary process does not become trapped at local optima. So, the tending of
the Efficiency E provides a clustering coefficient. It tends, never quite reaching its maximum,
because of the mutation inherent in the evolutionary process.
Figure 3.6: 3D Fitness Landscape with No Optimum: Theoretical extreme scenario in which the
selection pressure is non-discriminating. So, the population occupancy of the fitness landscape
would then be uniformly random, as any position (sequence) has the same fitness as any other.
So the entropy (randomness) tends to maximum, resulting in the complexity CV tending to zero,
and therefore the Efficiency E also tending to zero.
The other extreme scenario occurs when the number of clusters equals the size of the population,
which would only occur with a flat fitness landscape [149] resulting from a non-discriminating
selection pressure, as shown in Figure 3.6. The population occupancy is uniformly random, as
any position (sequence) has the same fitness as any other. So the entropy (randomness) tends
to maximum, resulting in the complexity CV tending to zero, and therefore the Efficiency E
also tending to zero, while the number of clusters T tends to the number of sequences in the
population S,
E =
CV
CVP
→ 0 as T → S.
(3.36)
So the number of clusters T tends to the population size S, with each cluster consisting of
only one unique sequence (individual).
If there are global optima, as there are in Figure 3.7, the Efficiency E will tend to a maximum
below one, because the population of sequences consists of more than one cluster, with each
having an Efficiency tending to a maximum of one. The simplest scenario of clusters is pure
clusters; pure meaning that each cluster uses a distinct (mutually exclusive) subset of the
alphabet D relative to any other cluster. In this scenario the Efficiency E tends to a value
based on the number of clusters T, because a number of the pd(i) probabilities at each site
1T. So, given that the number of the
in (3.27) are the reciprocal of the number of clusters,
pd(i) probabilities taking the value 1T is equal to the number of clusters, while the other pd(i)
probabilities take a value of zero, then the per-site entropy calculation of HV (i) from (3.27)
106
, Chapter 3.
Investigation of Digital Ecosystems
3D Fitness Landscape with Global Optima: Clustering scenario, in which the
Figure 3.7:
Efficiency E of the population S tends to a value based on the number of clusters T, because
the population of sequences is clustering around more than one global optima, with each cluster
having an Efficiency E tending to a maximum of one.
becomes
(3.37)
where i is the site, D is the alphabet size, and T is the number of clusters. Hence, given
(3.37), (3.32), and (3.24), then the Efficiency E from (3.33) becomes
HV (i) = logD T,
E → 1 − (logD T),
(3.38)
where D is the alphabet size and T is the number of clusters. Therefore, the Efficiency E,
the clustering coefficient, tends to a value that can be used to determine the number of pure
clusters in an evolving population of sequences.
For a population S with clusters, each cluster is a sub-population with an Efficiency E tending
to a maximum of one. To specify this relationship we require a function that provides the
Efficiency E (3.33) of a population or sub-population of sequences,
efficiency(input :population) :int.
(3.39)
So, for a population S consisting of a set of clusters T , each member (cluster) t is therefore a
sub-population of the population S, and is defined as
t ∈ T →
t ⊆ S ∧ efficiency(t) → 1 ∧ t ≈ S
(cid:32)
(cid:33)
t = S
,
(3.40)
T ∧(cid:88)
t∈T
3.2. Complexity ,
107
where a cluster t has an Efficiency E tending to a maximum of one, and the cluster size t is
approximately equal to the population size S divided by the number of clusters T. It is only
approximately equal because of variation from mutation, and because the population size may
not divide to a whole number. These conditions are true for all members t of the set of clusters
T , and therefore the summation of the cluster sizes t equals the size of the population S.
The population of sequences from the fitness landscape of Figure 3.7 is visualised in Figure 3.8,
but the clusters within cannot be seen. So, the population is arranged to show the clustering
in Figure 3.9, in which the two clusters are clearly evident. The clusters of the population have
Efficiency values tending to a maximum of one, compared to the Efficiency of the population as
a whole, which is tending to a maximum significantly below one. This is the expected behaviour
of clusters as defined in (3.40).
Figure 3.8: Population with Hidden Clusters: Visualisation for the population of sequences
from the fitness landscape of Figure 3.7, with clusters visually hard to identify. The clusters
lead to a low complexity CV relative to the maximum CVP , and hence the Efficiency E (3.33)
tends to a maximum significantly below one.
Figure 3.9: Population with Clusters Visible: Visualisation for the population of sequences
from Figure 3.8, which has been arranged to show the clusters present. The clusters of the
population have Efficiency values tending to a maximum of one, compared to the Efficiency of
the population as a whole, which is tending to a maximum significantly below one.
The population size S, in Figures 3.8 and 3.9, is double the minimum requirement specified in
(3.26), so that the complexity CV (3.32) and Efficiency E (3.33) could be used in defining the
principles of clustering without redefining the length of a population of variable length sequences
(cid:96)V (3.26). However, when determining the variable length (cid:96)V of a cluster t, the sample size
requirement is different, specifically a cluster t is a sub-population of S, and therefore by
CV=1.107,CVP=!V=3,%E=36.9D=alphabet={,,}PopulationPopulationCluster 1Cluster 2CV=2.704,E=0.901CVP=!V=3CV=3,E=1 108
, Chapter 3.
Investigation of Digital Ecosystems
definition cannot have a population size equivalent to S (unless the population consists of only
one cluster). Therefore, to manage clusters requires a reformulation of (cid:96)V (3.26) to
(cid:19)
,
(3.41)
(cid:18)
(cid:96)V =
sampleSize((cid:96)V ) ≈ D(cid:96)V
T ∧ sampleSize((cid:96)V + 1) <
D(cid:96)V
T
where (cid:96)max is the maximum length in a population of variable length sequences, (cid:96)V varies
between 1 ≤ (cid:96)V ≤ (cid:96)max, D is the alphabet, D > 0, and T is the set of clusters in the
population S.
A population with clusters will always have an Efficiency E tending towards a maximum
significantly below one. Therefore, managing populations with clusters requires a reformulation
of the Efficiency (3.33) to
Ec(S) =
P
t∈T
CV
CVP
.
Ec(t)
T
if T = 1
.
if T > 1
,
(3.42)
where t is a cluster, and a member of the set of clusters T of the population S. So, the Efficiency
Ec is equivalent to the Efficiency E if the population consists of only one cluster, but if there
are clusters then the Efficiency Ec is the average of the Efficiency E values of the clusters.
Figure 3.10: Agent Atomicity: Property of a set of Agents, such that no single Agent can
functionally replace any Agent-sequence, i.e.
their functionality is mutually exclusive to one
another. It is important because non-atomicity can adversely affect the Physical Complexity
measure. In this example, the alphabet is non-atomic, with the yellow Agent able to functionally
replace a green blue Agent-sequence.
Atomicity is the property of a set of Agents, such that no single Agent can functionally
replace any Agent-sequence, i.e. their functionality is mutually exclusive to one another. It
is important because non-atomicity can adversely affect the uniformity of the calculated per-
site entropies, which is the main construct of the Physical Complexity measure, and so non-
atomicity risks introducing error when calculating the information content. Our extensions to
Physical Complexity to support clustering are also necessary to manage non-atomicity, because
it leads to the formation of clusters within evolving Agent Populations. The presence of clusters
can be identified by the clustering coefficient, the Efficiency E tending to a value below one,
D=alphabet={,,,}abbccdac 3.2. Complexity ,
109
Population Constructed from a Non-Atomic Alphabet: The population is
Figure 3.11:
constructed from the alphabet shown in Figure 3.10, with the yellow Agent able to functionally
replace a green blue Agent-sequence. So, the Efficiency E of the population is a half, whereas the
Efficiency Ec for populations with clusters is one, because it supports clustering and therefore
non-atomicity.
with the Efficiency Ec (3.42) used to calculate the actual Efficiency as it supports clustering
and therefore non-atomicity.
If we consider the example population shown in Figure 3.11, which is constructed from the
alphabet shown in Figure 3.10, the yellow Agent
can functionally replace a green blue
Agent-sequence
, and so the uniformity across site two is lost. Therefore, the Efficiency
E of the population is a half, whereas the Efficiency Ec for populations with clusters is one,
because it supports clustering and therefore non-atomicity.
3.2.3 Simulation and Results
We simulated an evolving Agent Population from the Digital Ecosystem, using our simulation
from section 2.3 (unless otherwise specified), seeded with an alphabet (Agent-pool) of 15 Agents
for the evolutionary process. We also added the classes and methods necessary to calculate our
extended Physical Complexity and Efficiency, which required implementing the CV of (3.32),
the (cid:96)V of (3.41) and the HV of (3.27) for the per-site entropies. The Efficiency Ec (3.42), for
populations with clusters, was also implemented in the simulation.
3.2.3.1 Physical Complexity
Our extended Physical Complexity has the same structure and properties as the original
Physical Complexity [7], and so the relationship between fitness and our extended Physical
Complexity should be the same as the relationship between fitness and the original Physical
CV=1,CVP=2PopulationCluster 1Cluster 2E=0.5Ec=1CV=3,CVP=3CV=2,CVP=2E=1E=1site 2 = 110
, Chapter 3.
Investigation of Digital Ecosystems
Original Physical Complexity Graphs (reprinted from [8]): The Physical
Figure 3.12:
Complexity increases over the generations, suffering short-term decreases from the arrival of
fitter mutants, which spread through the population over several generations and causes the
uniformity of the sites to decrease temporarily, while the maximum fitness of the population
increases over the generations until the global optimum is reached.
Complexity [8]. If we consider the original Physical Complexity and fitness graphs, reprinted
from [8] in Figure 3.12, we can define the relationship as follows; the Physical Complexity
increases over the generations, suffering short-term decreases from the arrival of fitter mutants,
which spread through the population over several generations and causes the uniformity of
the sites to decrease temporarily, while the maximum fitness of the population increases over
the generations until the global optimum is reached, provided that there is a static selection
pressure and a low mutation rate (making it unlikely that the maximum fitness will decrease)
[8]. The original Physical Complexity starts uncharacteristically high in Figure 3.12, because
the population is seeded with a single sequence that temporarily takes over the population [8].
Figure 3.13 shows, for a typical evolving Agent Population, the Physical Complexity CV (3.32)
for variable length sequences and the maximum fitness Fmax over the generations. The Physical
Complexity for variable length sequences increases over the generations, showing short-term
decreases as expected [8]. It increases over the generations because of the increasing information
being stored, with the sharp increases occurring when the effective length (cid:96)V of the Population
increases. The temporary decreases, such as the one beginning at generation 138, are preceded
by the advent of a new fitter mutant, as indicated by a corresponding sharp increase in the
maximum fitness in the immediately preceding generations, which temporarily disrupt the self-
organised complexity of the population, until this new fitter mutant becomes dominant and
Updates [x104] Updates [x104] FitnessPhysical Complexity02641080264108506070809010010-2100102104 3.2. Complexity ,
111
Figure 3.13: Graph of Physical Complexity and Maximum Fitness over the Generations: The
Physical Complexity for variable length sequences increases over the generations, showing short-
term decreases as expected [8].
It increases over the generations because of the increasing
information being stored, with the sharp increases occurring when the effective length (cid:96)V of the
Population increases.
leads to a new higher level of self-organised complexity. Figure 3.13 shows that the fitness
and our extended Physical Complexity; both increase over the generations, synchronised with
one another, until generation 160 when the maximum fitness tapers off more slowly than the
Physical Complexity. At this point the optimal length for the sequences is reached within the
simulation, and so the advent of new fitter sequences (of the same of similar length) creates
only minor fluctuations in the Physical Complexity, while having a more significant effect on
the maximum fitness.
The similarity of the graph in Figure 3.13 to the graphs in Figure 3.12 confirms that the
Physical Complexity measure has been successfully extended to variable length sequences. The
temporary decreases in the Physical Complexity CV for variable length sequences were not
as severe as the original [8], because our simulation’s mutation rate was relatively low at only
10%. Also, our Physical Complexity CV does not start uncharacteristically high like the original,
because at the start the entire population was randomly seeded, instead of being seeded by just
a single individual as in the original [8].
010203050100138150200250300020406080100UnitsofInformation(PhysicalComplexity)%FitnessGenerationMaximumFitnessFmaxPhysicalComplexityCV 112
, Chapter 3.
Investigation of Digital Ecosystems
3.2.3.2 Efficiency
Figure 3.14 is a visualisation of the simulation, showing two alternate Populations that were
run for a thousand generations, with the one on the left from Figure 3.13 run under normal
conditions, while the one on the right was run with a non-discriminating selection pressure;
each multi-coloured line represents an Agent-sequence, while each colour represents an Agent
(site). The visualisation shows that our Efficiency E accurately measures the self-organised
complexity of the two Populations. It also shows significant variation in the Population run
under normal conditions, as the evolutionary computing process creates the opportunity to find
fitter (better) sequences, providing potential to avoid getting trapped at local optima.
Figure 3.14: Visualisation of Evolving Agent Populations at the 1000th Generation: Each
multi-coloured line represents an Agent-sequence, while each colour represents an Agent (site).
The population on the left from Figure 3.13 was run under normal conditions, while the one
on the right was run with a non-discriminating selection pressure. The visualisation shows that
our Efficiency E accurately measures the self-organised complexity of the two Populations.
Figure 3.15 shows the Efficiency E (3.33), over the generations, for the Population from
Figure 3.13. The Efficiency tends to a maximum of one,
indicating that the Population
consists of one cluster, which is confirmed by the visualisation of the Population in Figure
3.14 (left). The significant decreases that occurred in the Efficiency, reducing in magnitude
and frequency over the generations, came from mirroring the fluctuations that occurred in the
complexity CV , because the Efficiency E (3.33) is the complexity CV (3.32) over the complexity
7791529441191812715111311258644101076141413414111314412115151021212121344111515915153810213101210132106115871134851171177113485171171538102131012101321011584131315159151571215157121541311151591538102131012132101587113485171177101348517117153810213101210132101158413111591541211515102121212134714141341411131314131128644101047151113791529441191812111123101911413102111371317143132812351334243361101211214214559141164415101214147119615915138101125115151611291094181213511157514144129121341243139151112586131546112811492413111089515741594410171551531471214711315485171133810131012101366210158142142424214154513128687110113514641056515231512371398111142101556121310104610123105821210149515168241113117154310313515631131362131110147104861135315814105141315111210971481213119114417134475653810613121013662101583810131101210136621015838101310121013662101583810613121013662101585451344711441761171481213213151210910531415815113123136213112101410411631035156231134111313117154152168810123105821210149121310104679138111142101556515231512341454135642421415451312868715114223810131012101366210158711315485171131551563147121479441017157121524131110854313131511125861315261128114914129134121115514511515161129109418121351591513810112119963514116441510121433611012112142145412253310991141310211137131714313281111177915294411918124715111311286410107141413414111313144121151510212121213414131115159151538102131012101321011587113481711710711348517117531685316847113485171177113485171171538102131012101321015115841311151591541211515102121213471414134141113131413112864410104117151177915294411918121211413199114132111323713171431328123534336110121121421455141164415101214371199610159151381011251151516112910941812135111551414412912134124313131511125861315461128114924131110515117121594410171551563147121471131548517113381013101210136621015814224242141545131286871101135146414545152151237913811116421065612131010461012310582121029154688142111311715431035156311313133621311210141041161511312315281114105141315121093148121311611441713441575453810613121013662101583810131012101366210158381013110121013662101583810613121013662101585451345751523151235152315123134475453810613312101366210158381013110121013662101583810131012101366210158381061312101366210158545134415711441761131481213141315121010141581511343313362131121014104116310351563113111311741421516881012310582121014912131010467913811114210155651521512341351351464242141545131286871101142238101310121013662101181131548517114155156314712149441017157121524111085431313151112586131546112811491441291213412111551451151516112910941812135159151381011211963751411644151012143361101211214214544353319911413102111371317143132811111779152944119181247151113112864410107141413414111314121151510212121344131115159153810131012101366210158381013101210136621015841311151591541211515102121213471414131411131413112864410104111511711915294411918121111319911413102111371317143132813534433611012112142145514116441510143711996159151381011251151516129109181213511155141441291213412431313151112586131546112811492413111085157121594410171551563147112147113154851711338101310121013662101581422424214154513128687110113514641454515215123791381111421015561213101410461012310582121014915168814211131174310351563113133621311214104116151131231571410141315121031481213141516114417134415754538106131210136621015838101310121013662101583810131101210136621015838106131210136621015851151516112910941812135511515161129109418121353810613121013662101583810131101210136621015838101310121013662101583810613121013662101584515441571144176141131481213141315121010141571511312313362131121410411631135156311311131174142151688101231058212101491213101046791381111421015565152151234145413514642421415451312868711011422381013101210136621015871131548517113155156314712149441017157121524511108543131315111258613154611281149144129123134121115514511515161291091812135159151381011211996375141164415101433126110121121421454435331991141310211137131714313281111779152944119181241115111311286441010714141314111351441211515102121213441311151591538101310121013662101583810131012101366210158413111515915121151510212121347141413414111314151171215107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415910610781814291014622104159116107818142910146142104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781142910146221045159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014652210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291010622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181427910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181472914101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429141462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101466310415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462104159116107818142910146221041591161071814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161081814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210431596107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146292104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911107818142910147221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101466221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622101441591161078181429101462210415911610781814291014622104159116107818729101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911678181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142914622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210134159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146228415911610781814291014622104591161078181429101462210415911610781814291014622104159116107818142910622104159116107818142910146221041591161078181429101462210415911610781814291014622104159116107818142910146221041591161078181429101462210415911610781814291014221041591161078181429101462104159116107818142910146221041591141078181429101462210415911410781814291014622104159114107818142910146221041591141078181429101462210415911410781814291014622104159114107818142941014622104159114107818142910146221041591141078181429101462210415911410781814291014622104159114107881429101462210415911410781814291014622104159114Population (normal conditions)Population (non-discriminating) Agent-sequence (length)11!V!VEfficiency%E=96.8%CV=19.351,CVP=20Efficiency%E=47.0%CV=4.227,CVP=9 3.2. Complexity ,
113
Figure 3.15: Graph of Population Efficiency over the Generations for the Population from
Figure 3.13: The Efficiency tends to a maximum of one, indicating that the Population consists
of one cluster, which is confirmed by the visualisation of the Population in Figure 3.14 (left).
The significant decreases that occurred in the Efficiency, reducing in magnitude and frequency
over the generations, came from mirroring the fluctuations that occurred in the complexity CV .
potential CVP (3.24). These falls are caused by the creation of fitter (better) mutants within the
population, which eventually become the dominant genotype, but during the process causes
the Physical Complexity and the Efficiency to fall in the short-term.
3.2.3.3 Clustering
To further investigate the self-organised complexity of evolving Agent Populations, we simulated
a typical Population with a multi-objective selection pressure that had two independent global
optima (like the fitness landscape of Figure of 3.7), and so the potential to support two pure
clusters (each cluster using a unique subset of the alphabet D). The graph in Figure 3.16 shows
the Efficiency E over the generations acting as a clustering coefficient, oscillating around the
included best fit curve, quite significantly at the start, and then decreasing as the generations
progressed. The initial severe oscillations were caused by the creation and spread of fitter longer
mutants (Agent-sequences) in the Population, causing the Physical Complexity and therefore
the Efficiency to fluctuate significantly. The Efficiency tended to 0.744, as expected from (3.38)
given the alphabet size was fifteen, D=15, and the number of clusters was two, T=2. The
0.60.70.80.910100200300EfficiencyGeneration 114
, Chapter 3.
Investigation of Digital Ecosystems
Figure 3.16: Graph of the Clustering Coefficient over the Generations: The Efficiency oscillated
around the included best fit curve, quite significantly at the start, and then decreasing as the
generations progressed. It tended to 0.744, as expected from (3.38) given the alphabet size was
fifteen, D=15, and the number of clusters was two, T=2, indicating more than one cluster.
Visualisation of Clusters in an Evolving Agent Population at the 1000th
Figure 3.17:
Generation: The Agent-sequences were grouped to show the two clusters, and as expected from
(3.40) each cluster had a much higher Physical Complexity and Efficiency compared to the
Population as a whole. However, the Efficiency Ec calculated the complexity correctly.
0.50.60.70.7440.80200400600EfficiencyGeneration1543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611475154361547334152611471574361517334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473344526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543751547334152611471543615473341526117471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361573341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526151471543615473341526114715436154733415261471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526116715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547331526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543612547533415261141543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341526114715436154733415261147154361547334152611471543615473341536114715436154733415261147154361547334152611471543615473341561147154361547334152611471115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316131514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171312141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514141313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141315111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151414131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111516121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161114171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417111314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121617111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151211111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111214171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512171611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131711151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115161413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417111514131611151417131314111512161111141115141316111514171313141115121611111417111514131611151417131314111512161111141711151413161115141713131411151216111114171115141316111514171313141115121611111417Agent-sequence (length)1Population:CV=14.60,%E=73.2%,%Ec=98.0%Cluster1:CV=19.66,%E=98.3%Cluster2:CV=19.51,%E=97.6%!V 3.2. Complexity ,
115
tending itself indicated clustering, while the value it tended to indicated, as expected, the
presence of two clusters in the Population. A visualisation of the Population is shown in Figure
3.17, in which the Agent-sequences were grouped to show the two clusters. As expected from
(3.40) each cluster had a much higher Physical Complexity and Efficiency compared to the
Population as a whole. However, the Efficiency Ec is immune to the clusters and therefore
calculated the self-organised complexity of the Population correctly.
3.2.4 Summary
None of the existing definitions we considered [64, 24, 253, 48, 8] were directly applicable as a
definition for the self-organised complexity of an evolving Agent Population, but the properties
of Physical Complexity [8] closely matched our intuitive understanding, and so was chosen for
further investigation. Based upon information theory and entropy, it provides a measure of
the quantity of information in a population’s genome, relative to the environment in which it
evolves, by calculating the entropy in the population to determine the randomness in the genome
[8]. Reformulating Physical Complexity for evolving Agent Populations required consideration
of the following issues: the mapping of the sequence sites to the Agent-sequences, and the
managing of Populations of variable length sequences. We then built upon this to construct a
variant of the Physical Complexity called the Efficiency, because it was based on the efficiency
of information storage in Physical Complexity, which we then used to develop an understanding
of clustering and atomicity within evolving Agent Populations.
We then investigated the self-organised complexity of evolving Agent Populations through
experimental simulations, for which our extended Physical Complexity was consistent with
the original. We then investigated the Efficiency, which performed as expected, confirmed
by the numerical results and Population visualisations matching our intuitive understanding.
We then applied the Efficiency to the determination of clusters when subjecting an evolving
Agent Population to a multi-objective selection pressure. The numerical results, combined
with the visualisation of the multi-cluster Population, confirmed the ability of the Efficiency
to act as a clustering coefficient, not only indicating the occurrence of clustering, but also the
number of clusters (for pure clusters). We also confirmed that the Efficiency Ec for populations
with clusters was able to calculate correctly the self-organised complexity of evolving Agent
Populations with clusters.
116
, Chapter 3.
Investigation of Digital Ecosystems
Collectively, the experimental results confirm that Physical Complexity has been successfully
extended to evolving Agent Populations. Most significantly, Physical Complexity has been
reformulated algebraically for populations of variable length sequences, which we have confirmed
experimentally through simulations. Our Efficiency definition provides a macroscopic value to
characterise the level of complexity. Furthermore, the clustering coefficient defined by the
tending of the Efficiency, not only indicates clustering, but can also distinguish between a
single cluster population and a population with clusters. The number of clusters can even
be determined, for pure clusters, from the value to which the clustering coefficient tends.
Combined, this allows the Efficiency Ec definition to provide a normalised universally applicable
macroscopic value to characterise the complexity of a population, independent of clustering,
atomicity, length (variable or same), and size.
We have determined an effective understanding and quantification for the self-organised
complexity of the evolving Agent Populations of our Digital Ecosystem. Furthermore, the
understanding and techniques we have developed have applicability beyond evolving Agent
Populations, as wide as the original Physical Complexity, which has been applied from DNA
[7] to simulations of self-replicating programmes [172].
3.3 Stability
A definition for the self-organised stability of an evolving Agent Population should define
the resulting stability or instability that emerges over time, with no initial constraints from
modelling approaches for the inclusion of pre-defined specific behaviour, but capable of
representing the appearance of such behaviour should it occur.
None of the proposed definitions are directly applicable for the self-organised stability of an
evolving Agent Population. The ∈-machine modelling [64] is not applicable, because it is only
defined within the context of pre-biotic populations. The Prugel-Bennett Shapiro formalism
[253] is not suitable, because it necessitates the involvement of subjective human judgement at
the critical stage of quantifier selection. Self-Organised Criticality [17] is also not applicable as
it only models the events of genetic change in the population over time, rather than measuring
the resulting stability or instability of the population. Neither is Evolutionary Game Theory
[330], which only defines the genetic stability of the genotypes, in terms of equilibrium and
non-equilibrium dynamics, instead of the stability of the population as a whole.
3.3. Stability ,
117
Chli-DeWilde stability of Multi-Agent Systems [53] does fulfil the required definition of the
self-organised stability, measuring convergence to an equilibrium distribution. However, its
current formulation does not include Multi-Agent Systems that make use of evolutionary
computing algorithms, i.e. our evolving Agent Populations, but it could be extended to include
such Multi-Agent Systems, because its Markov-based modelling approach is well established
in evolutionary computing [272]. While there has been past work on modelling evolutionary
computing algorithms as Markov chains [271, 232, 111, 89], we have found none including Multi-
Agent Systems despite both being mature research areas [236, 195], because their integration
is a recent development [290]. So, the use of Chli-DeWilde stability as a definition for the
self-organised stability of evolving Agent Populations will be investigated further to determine
its suitability.
3.3.1 Chli-DeWilde Stability
Chli-DeWilde stability was created to provide a clear notion of stability in Multi-Agent Systems
[53], because stability is perhaps one of the most desirable features of any engineered system,
given the importance of being able to predict its response to various environmental conditions
prior to actual deployment; and while computer scientists often talk about stable or unstable
systems [306, 18], they did so without having a concrete or uniform definition of stability.
Also, other properties had been widely investigated, such as openness [2], scalability [198] and
adaptability [286], but stability had not. So, the Chli-DeWilde definition of stability for Multi-
Agent Systems was created [53], based on the stationary distribution of a stochastic system,
modelling the agents as Markov processes, and therefore viewing a Multi-Agent System as a
discrete time Markov chain with a potentially unknown transition probability distribution. The
Multi-Agent System is considered to be stable once its state has converged to an equilibrium
distribution [53], because stability of a system can be understood intuitively as exhibiting
bounded behaviour.
Chli-DeWilde stability was derived [52] from the notion of stability defined by De Wilde [77,
169], based on the stationary distribution of a stochastic system, making use of discrete-time
Markov chains, which we will now introduce3. If we let I be a countable set, in which each
i ∈ I is called a state and I is called the state-space. We can then say that λ = (λi : i ∈ I)
3 A more comprehensive introduction to Markov chain theory and stochastic processes is available in [234] and
[61].
, Chapter 3.
Investigation of Digital Ecosystems
118
is a measure on I if 0 ≤ λi < ∞ for all i ∈ I, and additionally a distribution if (cid:80)
i∈I λi = 1
[52]. So, if X is a random variable taking values in I and we have λi = Pr(X = i), then λ is
the distribution of X, and we can say that a matrix P = (pij : i, j ∈ I) is stochastic if every
row (pij : j ∈ I) is a distribution [52]. We can then extend familiar notions of matrix and
vector multiplication to cover a general index set I of potentially infinite size, by defining the
multiplication of a matrix by a measure as λP , which is given by
(cid:88)
j∈I
(λP )i =
λjpij.
(3.43)
We can now describe the rules for a Markov chain by a definition in terms of the corresponding
matrix P [52].
Definition 1. We say that (X t)t≥0 is a Markov chain with initial distribution λ = (λi : i ∈ I)
and transition matrix P = (pij : i, j ∈ I) if:
1. Pr(X 0 = i0) = λi0 and
2. Pr(X t+1 = it+1 X 0 = i0, . . . , X t = it) = pitit+1.
We abbreviate these two conditions by saying that (X t)t≥0 is Markov(λ, P ).
In this first definition the Markov process is memoryless, resulting in only the current state of
the system being required to describe its subsequent behaviour. We say that a Markov process
X 0, X 1, . . . , X t has a stationary distribution if the probability distribution of X t becomes
independent of the time t [53]. So, the following theorem is an easy consequence of the second
condition from the first definition.
Theorem 1. A discrete-time random process (X t)t≥0 is Markov(λ, P ), if and only if for all t
and i0, . . . , it we have
Pr(X 0 = i0, . . . , X t = it) = λi0pi0i1 ··· pit−1it.
(3.44)
This first theorem depicts the structure of a Markov chain, illustrating the relation with the
stochastic matrix P , and defining its time-invariance property [52].
3.3. Stability ,
119
Theorem 2. Let (X t)t≥0 be M arkov(λ, P ), then for all t, s ≥ 0:
1. Pr(X t = j) = (λP t)j and
2. Pr(X t = j X 0 = i) = Pr(X t+s = j X s = i) = (P t)ij.
For convenience (P t)ij can be more conveniently denoted as p(t)
ij .
Given this second theorem we can define p(t)
ij as the t-step transition probability from the state
i to j [52], and we can now introduce the concept of an invariant distribution [52], in which we
say that λ is invariant if
λP = λ.
(3.45)
The next theorem will link the existence of an invariant distribution, which is an algebraic
property of the matrix P , with the probabilistic concept of an equilibrium distribution. This
only applies to a restricted class of Markov chains, namely those with irreducible and aperiodic
stochastic matrices. However, there is a multitude of analogous results for other types of Markov
chains to which we can refer [234, 61], and the following theorem is provided as an indication
of the family of theorems that apply. An irreducible matrix P is one for which, for all i, j ∈ I
there are sufficiently large t, p(t)
ii > 0 for
ij > 0, and is aperiodic if for all states i ∈ I we have p(t)
all sufficiently large t [52].
Theorem 3. Let P be irreducible, aperiodic and have an invariant distribution, λ can be any
distribution, and suppose that (X t)t≥0 is Markov(λ, P ) [52], then
Pr(X t = j) → p∞
j as t → ∞ for all j ∈ I
and
ij → p∞
p(t)
j as t → ∞ for all i, j ∈ I.
(3.46)
(3.47)
We can now view a system S as a countable set of states I with implicitly defined transitions
P between them, and at time t the state of the system is the random variable X t, with the key
assumption that (X t)t,0 is Markov(λ, P ) [52].
Definition 2. The system S is said to be stable when the distribution of the its states converge
to an equilibrium distribution,
Pr(X t = j) → p∞
j as t → ∞ f or allj ∈ I.
(3.48)
120
, Chapter 3.
Investigation of Digital Ecosystems
More intuitively, the system S, a stochastic process X 0,X 1,X 2,...
is stable if the probability
distribution of X t becomes independent of the time index t for large t [53]. Most Markov chains
with a finite state-space and positive transition probabilities are examples of stable systems,
because after an initialisation period they settle down on a stationary distribution [52].
A Multi-Agent System can be viewed as a system S, with the system state represented by a
finite vector X, having dimensions large enough to manage the agents present in the system.
The state vector will consist of one or more elements for each agent, and a number of elements
to define general properties of the system state. We can then model an agent as being dead,
i.e. not being present in the system, by setting the vector elements for that agent to some
predefined null value [52].
3.3.2 Extensions for Evolving Populations
Extending Chli-DeWilde stability to the class of Multi-Agent Systems that make use of
evolutionary computing algorithms, including our evolving Agent Populations, requires con-
sideration of the following issues: the inclusion of population dynamics, and an understanding
of population macro-states.
3.3.2.1 Population Dynamics
First, the Multi-Agent System of an evolving Agent Population is composed of n Agent-
sequences, with each Agent-sequence i in a state ξt
i at time t, where i = 1, 2, . . . , n. The states of
the Agent-sequences are random variables, and so the state vector for the Multi-Agent System
is a vector of random variables ξt, with the time being discrete, t = 0, 1, . . . . The interactions
among the Agent-sequences are noisy, and are given by the probability distributions
Pr(XiY) = Pr(ξt+1
i = Xiξt = Y),
1, . . . , n,
(3.49)
where Xi is a value for the state of Agent-sequence i, and Y is a value for the state vector
of the Multi-Agent System. The probabilities implement a Markov process [300], with the
noise caused by mutations. Furthermore, the Agent-sequences are individually subjected to a
selection pressure from the environment of the system, which is applied equally to all the Agent-
sequences of the population. So, the probability distributions are statistically independent, and
3.3. Stability ,
Pr(XY) = Πn
i=1 Pr(ξt+1
i = Xiξt = Y).
If the occupation probability of state X at time t is denoted by pt
X, then
(cid:88)
pt
X =
Pr(XY)pt−1
Y .
121
(3.50)
(3.51)
Y
This is a discrete time equation used to calculate the evolution of the state occupation
probabilities from t = 0, while equation (3.50) is the probability of moving from one state
to another. The Multi-Agent System (evolving Agent Population) is self-stabilising if the limit
distribution of the occupation probabilities exists and is non-uniform, i.e.
p∞
X = limt→∞pt
X
exists for all states X, and there exist states X and Y such that
X (cid:54)= p∞
p∞
Y .
(3.52)
(3.53)
These equations define that some configurations of the system, after an extended time, will
be more likely than others, because the likelihood of their occurrence no longer changes. Such
a system is stable, because the likelihood of states occurring no longer changes with time,
and is the definition of stability developed in [53]. While equation (3.52) is the probabilistic
equivalence of an attractor 4 in a system with deterministic interactions, which we had to extend
to a stochastic process because mutation is inherent in evolutionary dynamics.
Although the number of agents in the Chli-DeWilde formalism can vary, we require it to vary
according to the selection pressure acting upon the evolving Agent Population. We must
therefore formally define and extend the definition of dead agents, by introducing a new state
d for each Agent-sequence. If an Agent-sequence is in this state, ξt
i = d, then it is dead and
does not affect the state of other Agent-sequences in the population. If an Agent-sequence i
has low fitness then that Agent-sequence will likely die, because
Pr(dY) = Pr(ξt+1
i = dξt = Y)
(3.54)
4 An attractor is a set of states, invariant under the dynamics, towards which neighbouring states asymptotically
approach during evolution [332].
122
, Chapter 3.
Investigation of Digital Ecosystems
will be high for all Y. Conversely, if an Agent-sequence has high fitness, then it will likely
replicate, assuming the state of a similarly successful Agent-sequence (mutant), or crossover
might occur changing the state of the successful Agent-sequence and another Agent-sequence.
3.3.2.2 Population Macro-States
As we defined earlier, the state of an evolving Agent Population is determined by the collection
of Agent-sequences of which it consists at a specific time t, and potentially changing state as the
time t increases. So, we can define a macro-state M as a set of states with a common property,
here possessing at least one copy of the current maximum fitness individual. Therefore, by its
definition, each macro-state M must also have a maximal state composed entirely of copies of
the current maximum fitness individual. There must also be a macro-state consisting of all the
states that have at least one copy of the global maximum fitness individual, which we will call
the maximum macro-state Mmax.
Figure 3.18: State-Space of an Evolving Agent Population: A possible evolutionary path through
the state-space I is shown, with the selection pressure of the evolutionary process driving it
towards the maximal state of the maximum macro-state Mmax, which consists entirely of copies
of the optimal solution, and is the equilibrium state that the system S is forever falling towards
without ever quite reaching, because of the noise (mutation) within the system.
We can consider the macro-states of an evolving Agent Population visually through the
representation of the state-space I of the system S shown in Figure 3.18, which includes a
possible evolutionary path through the state-space I. Traversal through the state-space I is
directed by the selection pressure of the evolutionary process acting upon the Population S,
driving it towards the maximal state of the maximum macro-state Mmax, which consists entirely
of copies of the optimal solution, and is the equilibrium state that the system S is forever falling
macrostatemaximal stateevolutionary pathstatemaximummacro-stateMmax 3.3. Stability ,
123
towards without ever quite reaching, because of the noise (mutation) within the system. So,
while this maximal state will never be reached, the maximum macro-state Mmax itself is certain
to be reached, provided the system does not get trapped at local optima, i.e. the probability
of being in the maximum macro-state Mmax at infinite time is one, p∞
equation (3.51).
Mmax = 1, as defined from
Furthermore, we can define quantitatively the probability distribution of the macro-states that
the system occupies at infinite time. For a stable system, as defined by equation (3.53), the
degree of instability, dins, can be defined as the entropy of its probability distribution at infinite
time,
dins = H(p∞
) = −(cid:88)
p∞
X logN (p∞
X ),
(3.55)
X
where N is the number of possible states, and taking log to the base N normalises the degree
of instability. The degree of instability will range between zero (inclusive) and one (exclusive),
because a maximum instability of one would only occur during the theoretical extreme scenario
of a non-discriminating selection pressure [149] (as shown in Figure 3.6).
3.3.3 Simulation and Results
We simulated an evolving Agent Population from the Digital Ecosystem, using our simulation
from section 2.3 (unless otherwise specified), seeded with an Agent-pool of 20 Agents5 for the
evolutionary process. We also added the classes and methods necessary to implement our
extended Chli-DeWilde stability and degree of instability, which required calculating p(t)
X (cid:54)= p∞
(3.51) to estimate the stability, and p∞
The degree of instability, dins of (3.55), was also implemented in the simulation.
X of (3.52) to prove the existence of p∞
X of
Y from (3.53).
3.3.3.1 Stability
Our evolving Agent Population (a Multi-Agent System with evolutionary dynamics) is stable
if the distribution of the limit probabilities exists and is non-uniform, as defined by equations
(3.52) and (3.53). The simplest case is a typical evolving Agent Population with one global
5 From optimisation improvements to the code base of the simulation, we were able to increase the size of the
Agent-pool from 15 to 20 without any significant degradation in performance, within the scope of running
tens of thousands of simulation runs.
124
, Chapter 3.
Investigation of Digital Ecosystems
optimal solution, which is stable if there are at least two macro-states with different limit
occupation probabilities. We shall consider the maximum macro-state Mmax and the sub-
optimal macro-state Mhalf . Where the states of the macro-state Mmax each possess at least one
individual with global maximum fitness,
p∞
Mmax = limt→∞p(t)
Mmax = 1,
while the states of the macro-state Mhalf each possess at least one individual with a fitness
equal to half of the global maximum fitness,
p∞
Mhalf = limt→∞p(t)
Mhalf = 0,
thereby fulfilling the requirements of equations (3.52) and (3.53). The sub-optimal macro-state
Mhalf , having a lower fitness, is predicted to be seen earlier in the evolutionary process before
disappearing as higher fitness macro-states are reached. The system S will take longer to
reach the maximum macro-state Mmax, but once it does will likely remain, leaving only briefly
depending on the strength of the mutation rate, as the selection pressure is non-elitist 6 (as
defined in section 2.3).
A value of t = 1000 was chosen to represent t = ∞ experimentally, because the simulation
has often been observed to reach the maximum macro-state Mmax within 500 generations.
Therefore, the probability of the system S being in the maximum macro-state Mmax at the
thousandth generation is expected to be one, p1000
Mmax = 1. Furthermore, the probability of the
system being in the sub-optimal macro-state Mhalf at the thousandth generation is expected
to be zero, p1000
Mhalf = 0.
Figure 3.19 shows, for a typical evolving Agent Population, a graph of the probability as defined
by equation (3.51) of the maximum macro-state Mmax and the sub-optimal macro-state Mhalf
at each generation, averaged from ten thousand simulation runs for statistical significance.
The behaviour of the simulated system S was as expected, being in the maximum macro-state
Mmax only after generation 178 and always after generation 482. It was also observed being
in the sub-optimal macro-state Mhalf only between generations 37 and 113, with a maximum
6 Non-elitist meaning that the best individual from one generation was not guaranteed to survive to the next
generation; it had a high probability of surviving into the next generation, but it was not guaranteed as it
might have been mutated [90].
3.3. Stability ,
125
Figure 3.19: Graph of the Probabilities of the Macro-States Mmax and Mhalf at each Generation:
The system S, a typical evolving Agent Population, was in the maximum macro-state Mmax only
after generation 178 and always after generation 482. It was also observed being in the sub-
optimal macro-state Mhalf only between generations 37 and 113, with a maximum probability
of 0.053 (3 d.p.) at generation 61.
probability of 0.053 (3 d.p.) at generation 61, and was such because the evolutionary path
(state transitions) could avoid visiting the macro-state. As expected the probability of being
in the maximum macro-state Mmax at the thousandth generation was one, p1000
Mmax = 1, and so
the probability of being in any other macro-state, including the sub-optimal macro-state Mhalf ,
at the thousandth generation was zero, p1000
Mhalf = 0.
A visualisation for the state of a typical evolving Agent Population at the thousandth generation
is shown in Figure 3.20, with each line representing an Agent-sequence and each colour
representing an Agent, with the identical Agent-sequences grouped for clarity. It shows that
the evolving Agent Population reached the maximum macro-state Mmax and remained there,
but as expected never reached the maximal state of the maximum macro-state Mmax, where
all the Agent-sequences are identical and have maximum fitness, which is indicated by the lack
of total uniformity in Figure 3.20. This was expected, because of the mutation (noise) within
the evolutionary process, which is necessary to create the opportunity to find fitter (better)
sequences and potentially avoid getting trapped at any local optima that may be present.
00.20.40.60.812004006008001000ProbabilityGenerationpMmaxpMhalf 126
, Chapter 3.
Investigation of Digital Ecosystems
Figure 3.20: Visualisation of an Evolving Agent Population at the 1000th Generation: The
Population consists of multiple Agent-sequences, with each line representing an Agent-sequence,
and therefore each colour representing an Agent. The identical Agent-sequences were grouped
for clarity, and as expected the system S reached the maximum macro-state Mmax and remained
there, but never reached the maximal state of the maximum macro-state Mmax.
3.3.3.2 Degree of Instability
Given that our simulated evolving Agent Population is stable as defined by equations (3.52) and
(3.53), we can determine the degree of instability as defined by equation (3.55). So, calculated
from its limit probabilities, the degree of instability was
dins = H(p1000) = −(cid:88)
p1000
X logN (p1000
X )
= −1logN (1)
X
= 0,
where t = 1000 is an effective estimate for t = ∞, as explained earlier. The result was as
expected because the maximum macro-state Mmax at the thousandth generation was one,
p1000
Mmax = 1, and so the probability of being in the other macro-states at the thousandth
generation was zero. The system therefore shows no instability, as there is no entropy in
the occupied macro-states at infinite time.
1451278910111213141516179341415131545678910111213141516179341415135678910111213141516179341415134578910111213141516179341415134568910111213141516179341415134567910111213141516179341415134567810111213141516179341415134567891091112131415161793414151345678910121314151617934141513456789101112141314151617934141513456789101112813141516179341415134567891011121213141516179341415134567891011121415161793414151345678910111213314151617934141513456789101112131614151617934141513456789101112131516179341415134567891011121314161793414151345678910111213141516121793414151345678910111213141516174934141513456789101112131415161734141513456789101112131415161794141513456789101112131415161793141513456789101112131415161793141513456789101112614151617934141513456789103121314151617934141513456789101112131415166934141513456789101112131415161593414151345678101011121314151617934141513456789101112131415161791141415134567891011121314151617974141513456789101112131415161793314151345678910111213141516179312141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415134567891011121314151617934141513456789101112131415161793414151345678910111213141516179341415EvolvingAgentPopulationatMaximumMacro-state:MmaxpMmax(1000)=1Agent-SequencePopulation 3.3. Stability ,
3.3.3.3 Stability Analysis
127
Figure 3.21: Graph of Stability with Different Mutation and Crossover Rates: With the
mutation rate under or equal to 60%, the evolving Agent Population showed no instability, with
dins values equal to zero as the system S was always in the same macro-state M at infinite time,
independent of the crossover rate. With the mutation rate above 60% the instability increased
significantly.
We then performed a stability analysis (similar to a sensitivity analysis [43]) of a typical evolving
Agent Population, varying key parameters within the simulation. We varied the mutation and
crossover rates from 0% to 100% in 10% increments, calculating the degree of instability, dins
from (3.55), at the thousandth generation. These degree of instability values were averaged over
ten thousand simulation runs, and graphed against the mutation and crossover rates in Figure
3.21. It shows that the crossover rate had little effect on the stability of our simulated evolving
Agent Population, whereas the mutation rate did significantly affect the stability. With the
mutation rate under or equal to 60%, the evolving Agent Population showed no instability, with
dins values equal to zero as the system S was always in the same macro-state M at infinite time,
independent of the crossover rate. With the mutation rate above 60% the instability increased
significantly, with the system being in one of several different macro-states at infinite time;
with a mutation rate of 70% the system was still very stable, having low dins values ranging
between 0.08 (2 d.p.) and 0.16 (2 d.p.), but once the mutation rate was 80% or greater the
system became quite unstable, shown by high dins values nearing 0.5.
020406080100204060801000.10.20.30.40.5DegreeofInstabilityCrossoverRateMutationRateDegreeofInstability 128
, Chapter 3.
Investigation of Digital Ecosystems
As one would have expected, an extremely high mutation rate has a destabilising effect on
the stability of an evolving Agent Population. The crossover rate had only a minimal effect,
because variation from crossover was limited when the Population had matured, consisting of
Agent-sequences identical or very similar to one another.
It should also be noted that the
stability of the system is different to its performance, because although showing no instability
with mutation rates below 60% (inclusive), it only reached the maximum macro-state Mmax
with a mutation rate of 10% or above, while at 0% it was stable at a sub-optimal macro-state.
3.3.4 Summary
None of the existing definitions we considered [64, 253, 17, 53] were directly applicable as a
definition for the self-organised stability of an evolving Agent Population, but the properties
of Chli-DeWilde stability [53] closely matched our intuitive understanding, and so was chosen
for further investigation. It views a Multi-Agent System as a discrete time Markov chain (with
potentially unknown transition probabilities) that is considered to be stable when its state,
a stochastic process, has converged to an equilibrium distribution. Extending Chli-DeWilde
stability to the Multi-Agent System of an evolving Agent Population required consideration of
the following issues: the inclusion of population dynamics, and an understanding of population
macro-states. We then built upon this to construct an entropy-based definition for the degree
of instability (entropy of the limit probabilities), which was later used to perform a stability
analysis of an evolving Agent Population.
We then investigated the self-organised stability of evolving Agent Populations through
experimental simulations, and the results showed that there was a limit probability distribution,
and that it was non-uniform. Furthermore, the reaching of the maximum macro-state was
confirmed by a visualisation matching the numerical results. We then applied our degree
of instability to determine that there was no instability under normal conditions, and then
performed a stability analysis (similar to a sensitivity analysis [43]) showing the variation of the
self-organised stability under varying conditions. Collectively, the experimental results confirm
that Chli-Dewilde stability has been successfully extended to evolving Agent Populations, while
our definition for the degree of instability provides a macroscopic value to characterise the level
of stability.
We have determined an effective understanding and quantification for the self-organised stability
of the evolving Agent Populations of our Digital Ecosystem. Also, our extended Chli-DeWilde
3.4. Diversity ,
129
stability is applicable to other Multi-Agent Systems with evolutionary dynamics. Furthermore,
our degree of instability provides a definition for the level of stability, applicable to Multi-Agent
Systems with or without evolutionary dynamics.
3.4 Diversity
A definition for the self-organised diversity of an evolving Agent Population should define the
optimal variability, of the Agents and Agent-sequences, that emerge over time, with no initial
constraints from modelling approaches for the inclusion of pre-defined specific behaviour, but
capable of representing the appearance of such behaviour should it occur.
None of the proposed definitions are applicable for the self-organised diversity of an evolving
Agent Population. The ∈-machine modelling [64] is not applicable, because it is only defined
within the context of pre-biotic populations. Neither is the Minimum Description Length
[24] principle or the Prugel-Bennett Shapiro formalism [253] suitable, because they necessitate
the involvement of subjective human judgement at the critical stages of model or quantifier
selection. Mean Field Theory is also not applicable because of the necessity of a neighbourhood
model for defining interaction, and evolving Agent Populations lack a 2D or 3D metric space
for such models. So, the only available neighbourhood model becomes a distance measure on
a parameter space that measures dissimilarity. However, this type of neighbourhood model
cannot represent the information-based interactions between the individuals of an evolving
Agent Population.
We suggest that the uniqueness of Digital Ecosystems makes the application of existing
definitions inappropriate for the self-organised diversity, because while we could extend a
biology-centric definition for the self-organised complexity, and a computing-centric definition
for the self-organised stability, we found neither of these approaches, or any other, appropriate
for the self-organised diversity. The Digital Ecosystem being the digital counterpart of a
biological ecosystem gives it unique properties, as discussed earlier in section 2.4. So, the
evolving Agent Populations possess properties of both computing systems (e.g. agent systems)
as well as biological systems (e.g.
population dynamics), and the combination of these
properties makes them unique. So, we will further consider the evolving Agent Populations
to create a definition for their self-organised diversity.
130
, Chapter 3.
Investigation of Digital Ecosystems
3.4.1 Evolving Agent Populations
The self-organised diversity of an evolving Agent Population comes from the Agent-sequences
it evolves, in response to the selection pressure, seeded with Agents and Agent-sequences from
the Agent-pool of the Habitat in which it is instantiated. The set of Agents and Agent-
sequences available when seeding an evolving Agent Population is regulated over time by
other evolving Agent Populations, instantiated in response to other user requests, leading to
the death and migration of Agents and Agent-sequences, as well as the formation of new
Agent-sequence combinations. The seeding of existing Agent-sequences provides a direction to
accelerate the evolutionary process, and can also affect the self-organised diversity; for example,
if only a proportion of any available global optima is favoured. So, the set of Agents available
when seeding an evolving Agent Population provides potential for the self-organised diversity,
while the selection pressure of a user request provides a constraining factor on this potential.
Therefore, the optimality of the self-organised diversity of an evolving Agent Population is
relative to the selection pressure of the user request for which it was instantiated.
While we could measure the self-organised diversity of individual evolving Agent Populations,
or even take a random sampling, it will be more informative to consider their collective self-
organised diversity. Also, given that the Digital Ecosystem is required to support a range of
user behaviour, we can consider the collective self-organised diversity of the evolving Agent
Populations relative to the global user request behaviour. So, when varying a behavioural
property of the user requests according to some distribution, we would expect the corresponding
property of the evolving Agent Populations to follow the same distribution. We are not
intending to prescribe the expected user behaviour of the Digital Ecosystem, but investigate
whether the Digital Ecosystem can adapt to a range of user behaviour in terms of the self-
organised diversity. So, we will consider Uniform, Gaussian (Normal) and Power distributions
for the parameters of the user request behaviour. The Uniform distribution will provide a
control, while the Normal (Gaussian) distribution will provide a reasonable assumption for the
behaviour of a large group of users, and the Power distribution will provide a relatively extreme
variation in user behaviour.
3.4.2 Simulation and Results
We simulated the Digital Ecosystem, using our simulation from section 2.3 (unless otherwise
specified). We also added the classes and methods necessary to vary aspects of the user
3.4. Diversity ,
131
behaviour according to different distributions, and a way to measure the related aspects of the
evolving Agent Populations. This consisted of a mechanism to vary the user request properties
of length and modularity, according to Uniform, Gaussian (normal) and Power distributions, and
a mechanism to measure the corresponding Agent(-sequence) properties of length and number
of attributes. For statistical significance each scenario (experiment) will be averaged from ten
thousand simulation runs. We expect it will be obvious whether the observed behaviour of the
Digital Ecosystem matches the expected behaviour from the user base. Nevertheless, we will
also implement a chi-squared (χ2) test to determine if the observed behaviour (distribution) of
the Agent(-sequence) properties matches the expected behaviour (distribution) from the user
request properties.
Given the requirement to run a minimum of sixty thousand simulation runs, ten thousand for
each experiment, we adapted the code base of the simulation to take advantage of the Xgrid
[159] distributed computing technology, and therefore make use of the grids mentioned in the
acknowledgements.
3.4.2.1 User Request Length
We started by varying the user request length according to the available distributions, expecting
the length of the Agent-sequences to be distributed according to the length of the user requests,
i.e. the longer the user request, the longer the Agent-sequence needed to fulfil it.
We first applied the Uniform distribution as a control, and graphed the results in Figure 3.22.
The observed frequencies of the Agent-sequence length mostly matched the expected frequencies,
which was confirmed by a χ2 test; with a null hypothesis of no significant difference and sixteen
degrees of freedom, the χ2 value was 2.588 (3 d.p.), below the critical 0.95 χ2 value of 7.962.
We then applied the Gaussian distribution as a reasonable assumption for the behaviour of
a large group of users, and graphed the results in Figure 3.23. The observed frequencies of
the Agent-sequence length matched the expected frequencies with only very minor variations,
which was confirmed by a χ2 test; with a null hypothesis of no significant difference and sixteen
degrees of freedom, the χ2 value was 2.102 (3 d.p.), below the critical 0.95 χ2 value of 7.962.
132
, Chapter 3.
Investigation of Digital Ecosystems
Figure 3.22: Graph of Uniformly Distributed Agent-Sequence Length Frequencies: The observed
frequencies of the Agent-sequence length mostly matched the expected frequencies, which was
confirmed by a χ2 test; with a null hypothesis of no significant difference and sixteen degrees of
freedom, the χ2 value was 2.588 (3 d.p.), below the critical 0.95 χ2 value of 7.962.
Figure 3.23: Graph of Gaussian Distributed Agent-Sequence Length Frequencies: The observed
frequencies of the Agent-sequence length matched the expected frequencies with only very minor
variations, which was confirmed by a χ2 test; with a null hypothesis of no significant difference
and sixteen degrees of freedom, the χ2 value was 2.102, below the critical 0.95 χ2 value of 7.962.
051015202530351234567891011121314151617181920FrequencyAgent-sequenceLengthExpectedObserved010203040506070801234567891011121314151617181920FrequencyAgent-sequenceLengthExpectedObserved 3.4. Diversity ,
133
Figure 3.24: Graph of Power Distributed Agent-Sequence Length Frequencies: The observed
frequencies of the Agent-sequence length matched the expected frequencies with some variation,
which was confirmed by a χ2 test; with a null hypothesis of no significant difference and sixteen
degrees of freedom, the χ2 value was 5.048 (3 d.p.), below the critical 0.95 χ2 value of 7.962.
Finally, we applied the Power distribution to represent a relatively extreme variation in user
behaviour, and graphed the results in Figure 3.24. The observed frequencies of the Agent-
sequence length matched the expected frequencies with some variation, which was confirmed by
a χ2 test; with a null hypothesis of no significant difference and sixteen degrees of freedom, the
χ2 value was 5.048 (3 d.p.), below the critical 0.95 χ2 value of 7.962.
There were a couple of minor discrepancies, similar to all the experiments. First, there were
a small number of individual Agents at the thousandth time step, caused by the typical user
behaviour of continuously creating new services (Agents). Second, while the chi-squared tests
confirmed that there was no significant difference between the observed and expected frequencies
of the Agent-sequence length, there was still a bias to longer Agent-sequences (solutions).
Evident visually in the graphs of the experiments, and evident numerically in the chi-squared
test of the Power distribution experiment as it favoured shorter Agent-sequences. The cause of
this bias was most likely some aspect of bloat (as we discussed in section 2.2.3.3) that was not
fully controlled.
0501001502001234567891011121314151617181920FrequencyAgent-sequenceLengthExpectedObserved 134
, Chapter 3.
Investigation of Digital Ecosystems
3.4.2.2 User Request Modularity
Next, we varied the user request modularity according to the available distributions, expecting
the sophistication of the Agents to be distributed according to the modularity of the user
requests, i.e. the more complicated (in terms of modular non-reducible tasks) the user request,
the more sophisticated (in terms of the number of attributes) the Agents needed to fulfil it.
We first applied the Uniform distribution as a control, and graphed the results in Figure 3.25.
The observed frequencies for the number of Agent attributes mostly matched the expected
frequencies, which was confirmed by a χ2 test; with a null hypothesis of no significant difference
and ten degrees of freedom, the χ2 value was 1.049 (3 d.p.), below the critical 0.95 χ2 value of
3.940.
Figure 3.25: Graph of Uniformly Distributed Agent Attribute Frequencies: The observed
frequencies for the number of Agent attributes mostly matched the expected frequencies, which
was confirmed by a χ2 test; with a null hypothesis of no significant difference and ten degrees
of freedom, the χ2 value was 1.049 (3 d.p.), below the critical 0.95 χ2 value of 3.940.
We then applied the Gaussian distribution as a reasonable assumption for the behaviour of a
large group of users, and graphed the results in Figure 3.26. The observed frequencies for the
number of Agent attributes appeared to follow the expected frequencies, but there was significant
variation which led to a failed χ2 test; with a null hypothesis of no significant difference and
ten degrees of freedom, the χ2 value was 50.623 (3 d.p.), not below the critical 0.95 χ2 value of
3.940.
010020030040050060056789101112131415FrequencyNumberofAttributesperAgentExpectedObserved 3.4. Diversity ,
135
Figure 3.26: Graph of Gaussian Distributed Agent Attribute Frequencies: The observed
frequencies for the number of Agent attributes appeared to follow the expected frequencies,
but there was significant variation which led to a failed χ2 test; with a null hypothesis of no
significant difference and ten degrees of freedom.
Figure 3.27: Graph of Power Distributed Agent Attribute Frequencies: The observed frequencies
for the number of Agent attributes appeared to follow the expected frequencies, but there was
significant variation which led to a failed χ2 test; with a null hypothesis of no significant
difference and ten degrees of freedom.
025050075010001250150056789101112131415FrequencyNumberofAttributesperAgentExpectedObserved050010001500200056789101112131415FrequencyNumberofAttributesperAgentExpectedObserved 136
, Chapter 3.
Investigation of Digital Ecosystems
Finally, we applied the Power distribution to represent a relatively extreme variation in user
behaviour, and graphed the results in Figure 3.27. The observed frequencies for the number of
Agent attributes appeared to follow the expected frequencies, but there was significant variation
which led to a failed χ2 test; with a null hypothesis of no significant difference and ten degrees
of freedom, the χ2 value was 61.876 (3 d.p.), not below the critical 0.95 χ2 value of 3.940.
In all the experiments the observed frequencies of the number of Agent attributes appeared to
follow the expected frequencies, but this could only be confirmed statistically, by a χ2 test, for
the Uniform distribution experiment. In the Gaussian and Power distribution experiments the
χ2 tests failed by considerable margins, most likely because the evolving Agent Populations were
still self-organising to match the user behaviour, shown by the observed frequencies approaching
the expected frequencies, but not yet sufficiently to meet χ2 tests, because by the thousandth
time step (user request event) each user had placed an average of only ten requests.
3.4.3 Summary
None of the existing definitions we considered [64, 24, 253, 95] were applicable as a definition
for the self-organised diversity of the evolving Agent Populations. So, we further considered
the unique properties resulting from information-centric Digital Ecosystems being the digital
counterpart of energy-centric biological ecosystems, creating our own definition for the self-
organised diversity of an evolving Agent Population, relative to the selection pressure provided
by a user request. We then considered the collective self-organised diversity of the evolving
Agent Populations relative to the global user request behaviour. Therefore, when varying
a behavioural property of the user requests according to some distribution, we expected the
corresponding property of the evolving Agent Populations to follow the same distribution. We
used the Uniform distribution to provide a control, the Normal (Gaussian) distribution to
provide a reasonable assumption for the behaviour of a large group of users, and the Power
distribution to represent a relatively extreme distribution in user behaviour.
We then investigated the self-organised diversity of evolving Agent Populations through
experimental simulations. First, varying the user request length according to the different
distributions, and testing whether the observed frequencies of the Agent-sequence length
matched the expected frequencies, which we confirmed with successful chi-squared tests. Second,
varying the user request modularity according to the different distributions, and testing whether
the observed frequencies for the number of Agent attributes matched the expected frequencies,
3.5. Summary and Discussion ,
137
again confirming with chi-squared tests. Under the Gaussian and Power distributions the chi-
squared tests failed, most likely because the evolving Agent Populations were still self-organising
to match the user behaviour, because at the time the Digital Ecosystem was sampled each user
had placed an average of only ten requests.
Collectively, the experimental results confirm that the self-organised diversity of the evolving
Agent Populations is relative to the selection pressures of the user base, which was confirmed
statistically for most of the experiments. So, we have determined an effective understanding and
quantification for the self-organised diversity of the evolving Agent Populations of our Digital
Ecosystem. While the minor experimental failures, in which the Digital Ecosystem responded
more slowly than in the other experiments, have shown that there is potential to optimise the
Digital Ecosystem, because the evolutionary self-organisation of an ecosystem is a slow process
[29], even the accelerated form present in our Digital Ecosystem.
3.5 Summary and Discussion
We have investigated the self-organising behaviour of Digital Ecosystems, because a primary
motivation for our research is the desire to exploit the self-organising properties of biological
ecosystems [173], which are thought to be robust, scalable architectures that can automatically
solve complex, dynamic problems. Over time a biological ecosystem becomes increasingly
self-organised through the process of ecological succession [29], driven by the evolutionary self-
organisation of the populations within the ecosystem. Analogously, a Digital Ecosystem’s
increasing self-organisation comes from the Agent Populations being evolved to meet the
dynamic selection pressures created by requests from the user base. The self-organisation of
biological ecosystems is often defined in terms of the complexity, stability, and diversity [150],
which we also applied in defining the self-organisation of our Digital Ecosystems. We started
by discussing the relevant literature, including the philosophical meaning of organisation and
of self, learning that self-organisation is context dependent, and that a system is only self -
organising if the process or force causing the organisation is within its boundaries. So, we
compared and contrasted alternative definitions [64, 24, 253, 48, 8, 17, 53, 95] for the self-
organised complexity, stability, and diversity of the evolving Agent Populations, examining their
suitability and application, because possessing properties of computing systems (e.g. agent
138
, Chapter 3.
Investigation of Digital Ecosystems
systems) as well as biological systems (e.g. population dynamics), the combination of these
properties makes the evolving Agent Populations unique.
None of the existing definitions we considered [64, 24, 253, 48, 8] were directly applicable as a
definition for the self-organised complexity of evolving Agent Populations, but the properties
of Physical Complexity [8] closely matched our intuitive understanding, and so was chosen
for further investigation. Based upon information theory and entropy, it provides a measure
of the quantity of information in the genome of a population, relative to the environment in
which it evolves, by calculating the entropy in the population to determine the randomness
in the genome [8]. Reformulating Physical Complexity for an evolving Agent Population
required consideration of the following issues:
the mapping of the sequence sites to the
Agent-sequences, and the managing of populations of variable length sequences. We then
built upon this to construct a variant of the Physical Complexity called the Efficiency,
because it was based on the efficiency of information storage in Physical Complexity, which
we then used to develop an understanding of clustering and atomicity within evolving Agent
Populations. Collectively, the experimental results confirm that Physical Complexity has been
successfully extended to evolving Agent Populations. Most significantly, Physical Complexity
has been reformulated algebraically for populations of variable length sequences, which we have
confirmed experimentally through simulations. Our Efficiency definition provides a universally
applicable macroscopic value to characterise the complexity of a population,
independent
of clustering, atomicity,
length (variable or same), and size. So, we have determined an
effective understanding and quantification for the self-organised complexity of the evolving
Agent Populations of our Digital Ecosystem. The understanding and techniques we have
developed have applicability beyond evolving Agent Populations, as wide as the original
Physical Complexity, which has been applied from DNA [7] to simulations of self-replicating
programmes [172].
None of the existing definitions we considered [64, 253, 17, 17, 53] were directly applicable as
a definition for the self-organised stability of evolving Agent Populations, but the properties of
Chli-DeWilde stability [53] closely matched our intuitive understanding, and so was chosen for
further investigation. It views a Multi-Agent System as a discrete time Markov chain (with
potentially unknown transition probabilities) that is considered to be stable when its state,
a stochastic process, has converged to an equilibrium distribution. Extending Chli-DeWilde
stability to the Multi-Agent System of an evolving Agent Population required consideration
of the following issues:
the inclusion of population dynamics, and an understanding of
3.5. Summary and Discussion ,
139
population macro-states. We then built upon this to construct an entropy-based definition
for the degree of instability (entropy of the limit probabilities), which was used to perform
a stability analysis (similar to a sensitivity analysis [43]) of an evolving Agent Population.
Collectively, the experimental results confirm that Chli-Dewilde stability has been successfully
extended to evolving Agent Populations, while our definition for the degree of instability
provides a macroscopic value to characterise the level of stability. So, we have determined
an effective understanding and quantification for the self-organised stability of the evolving
Agent Populations of our Digital Ecosystem. Also, our extended Chli-DeWilde stability is
applicable to other Multi-Agent Systems with evolutionary dynamics. Furthermore, our degree
of instability is applicable to all Multi-Agent Systems, with or without evolutionary dynamics.
None of the existing definitions we considered [64, 24, 253, 95] were applicable as a definition for
the self-organised diversity of evolving Agent Populations. So, we further considered the unique
properties resulting from information-centric Digital Ecosystems being the digital counterpart of
energy-centric biological ecosystems, creating our own definition for the self-organised diversity
of an evolving Agent Population relative to the selection pressure provided by a user request. We
then considered the collective self-organised diversity of the evolving Agent Populations relative
to the global user request behaviour. Therefore, when varying a behavioural property of the
user requests according to some distribution, we expected the corresponding property of the
evolving Agent Populations to follow the same distribution. We used the Uniform distribution
to provide a control, the Normal (Gaussian) distribution to provide a reasonable assumption
for the behaviour of a large group of users, and the Power distribution to provide a relatively
extreme distribution in user behaviour. Collectively, the experimental results confirm that the
self-organised diversity of the evolving Agent Populations is relative to the selection pressures
of the user base, which was confirmed statistically for most of the experiments. So, we have
determined an effective understanding and quantification for the self-organised diversity of the
evolving Agent Populations of our Digital Ecosystem. While the minor experimental failures,
in which the Digital Ecosystem responded more slowly than in the other experiments, have
shown that there is potential to optimise the Digital Ecosystem, because the evolutionary self-
organisation of an ecosystem is a slow process [29], even the accelerated form present in our
Digital Ecosystem.
Overall an insight has been achieved into where and how self-organisation occurs in our Digital
Ecosystem, and what forms this self-organisation can take and how it can be quantified. The
hybrid nature of the Digital Ecosystem resulted in the most suitable definition for the self-
140
, Chapter 3.
Investigation of Digital Ecosystems
organised complexity coming from the biological sciences, while the most suitable definition
for the self-organised stability coming from the computer sciences. However, we were unable
to use any existing definition for the self-organised diversity, because the hybrid nature of the
Digital Ecosystem makes it unique, and so we constructed our own definition based on variation
relative to the user base. The (Physical) complexity definition applies to a single point in time
of the evolving Agent Populations, whereas the (Chli-DeWilde) stability definition applies at
the end of these instantiated evolutionary processes, while our diversity definition applies to
the optimality of the distribution of the Agents within the evolving Agent Populations of the
Digital Ecosystem. The experimental results have generally supported the hypotheses, and have
provided more detail to the behaviour of the self-organising phenomena under investigation,
showing some of its properties and for the self-organised diversity has shown that there is
potential for optimising the Digital Ecosystem.
In this chapter we have investigated the emergent self-organising properties of Digital
Ecosystems, and with the greater and more in-depth understanding we have developed and
gained of the order constructing processes (the evolving Agent Populations),
including a
clearer identification of the potential areas and scopes for augmentation, we will attempt
the optimisation of Digital Ecosystems in the following chapter, Chapter 4, for which the
results here have confirmed the potential for optimisation identified in the previous chapter,
Chapter 2.
Chapter 4
Optimisation of Digital Ecosystems
In this chapter we attempt the acceleration and optimisation of Digital Ecosystems, because the
evolutionary self-organisation of ecological succession (the formation of a mature ecosystem) is a
slow process, even the accelerated form present in our Digital Ecosystem. First, we consider the
scope for optimisation identified in the previous chapters, and the potential for augmentations
from the biological sciences. Consolidating this understanding we propose, construct and
explore alternative augmentations to accelerate or optimise the evolutionary and ecological self-
organising dynamics of our Digital Ecosystems. The most promising, the clustering catalyst and
the targeted migration, were completed theoretically, before being investigated experimentally
to determine their improvement on the evolutionary and ecological dynamics in responding
to the needs of the user base; the first aiming to optimise the evolutionary dynamics, while
the second aiming to optimise the ecological dynamics. First, the clustering catalyst operates
upon an evolutionary process, encouraging intra-cluster crossover to accelerate reaching the
optimal solution, directly accelerating a core operation of the Digital Ecosystem. A suitable
existing clustering algorithm and a Physical Complexity based one were both evaluated for
the required clustering. Second, the targeted migration operates on the ecological dynamics,
allowing the Agents to interact for additional highly targeted migration, indirectly optimising a
global operation of the Digital Ecosystem. Both Neural Networks and Support Vector Machines
were evaluated for the required pattern recognition and learning functionality. We conclude
with a summary and discussion of the achievements, including the experimental results.
141
142
, Chapter 4. Optimisation of Digital Ecosystems
4.1 Background Theory
We proposed that an ecosystem inspired approach would be more effective at greater scales than
traditionally inspired approaches, because it would be built upon the scalable and self-organising
properties of biological ecosystems [173]. So, a Digital Ecosystem, being the digital counterpart
of biological ecosystems, possesses their scalable self-organising behaviour, properties and
processes. However, the self-organising process of ecological succession is a slow one, the orderly
and predictable changes in the composition and structure of an ecological community in forming
a mature ecosystem [29], even the accelerated form present in our Digital Ecosystem. Therefore,
it may be possible to accelerate and optimise this equivalent process of our Digital Ecosystem.
The scope for optimisation and acceleration was identified and confirmed in the previous
chapters. First, identified by the results of Chapter 2, specifically the ecological succession
experiment (section 2.3.6) in which the Digital Ecosystem reached only 70% responsiveness,
clearly showing potential for improvement. Second, confirmed by the results of Chapter 3,
specifically two of the modularity scenarios of the self-organised diversity experiment (section
3.4.2.2), in which the Digital Ecosystem responded more slowly in these scenarios than others,
confirming the potential for improvement. Therefore, there is scope for optimising and
accelerating the equivalent process of ecological succession in our Digital Ecosystem.
In biological ecosystems the trajectory of ecological change can be influenced by site conditions,
by the interactions of the species present, and by more stochastic factors such as the availability
of colonists or seeds, or weather conditions at the time of disturbance [312]. So, ecological
optimisation is generally concerned with the maintenance of diversity and stability, for the
survival of populations, species, habitats, etc [323, 3, 224], and ecological acceleration is
similarly concerned with the re-establishment of diversity and stability, through optimal species
selection and promotion [201, 181, 336]. Therefore, biological ecosystems research has no focus
on the type of optimisation or acceleration we require, which is unsurprising, because one of
the fundamental differences between biological and digital ecosystems lie in the motivation
and approach of their researchers; given that biological ecosystems are ubiquitous natural
phenomena whose maintenance is crucial to our survival [20], whereas Digital Ecosystems
are a technology engineered to serve specific human purposes. So, we are unlikely to find
augmentations from biological ecosystems to optimise our Digital Ecosystem.
4.2. Alternative Augmentations ,
143
The optimisation of Digital Ecosystems sought is not that of parameter optimisation, which
is achievable through exploratory programming [294], but an augmentation to the Ecosystem-
Oriented Architecture that provides a significant improvement in performance, i.e. better
solutions for the users than the Digital Ecosystem alone could achieve.
In the previous
chapter we have investigated the emergent self-organising properties of Digital Ecosystems,
and with the greater and more in-depth understanding we have developed and gained of the
order constructing processes (the evolving Agent Populations), including a clearer identification
of the potential areas and scopes for augmentation, we will now propose, construct and
explore alternative augmentations to accelerate or optimise the evolutionary and ecological
self-organising dynamics of our Digital Ecosystem. The most promising will be completed
theoretically and then investigated experimentally through simulations.
4.2 Alternative Augmentations
Any proposed augmentation should improve the process of ecological succession [29] for our
Digital Ecosystem. So, based on the understanding and results from the previous chapters,
our general knowledge, and our intuition, we will now propose, construct, and explore possible
alternative augmentations for our Digital Ecosystem that fulfils this requirement.
4.2.1 Clustering Catalyst
A significant proportion of user requests will be returned multiple optimal responses (applica-
tions), by evolving Agent Populations consisting of clusters as defined in the previous chapter.
So, potential exists to accelerate these evolving Agent Populations with clusters by a clustering
catalyst, which would encourage intra-cluster crossover, reducing the number of generations
required for the clusters to reach their respective optimal genomes (applications), therefore
directly accelerating the evolutionary self-organisation in determining applications (Agent-
sequences) to user requests. So, accelerating the responsiveness of the Digital Ecosystem to the
user base, and the process of ecological succession [29].
Crossover involves the crossing of two Agent-sequences, leading to recombination in the creation
of new Agent-sequences, during the replication stage of the evolutionary cycle [13]. This
144
, Chapter 4. Optimisation of Digital Ecosystems
augmentation would encourage inter-cluster crossover within the evolving Agent Populations,
with the aim of directly accelerating them in to find the optimal Agent-sequence(s) in fewer
generations. As each evolving Agent Population within the Digital Ecosystem would be
accelerated, the entire ecosystem would operate more efficiently.
Figure 4.1: Clustering Catalyst: This would encourage intra-cluster crossover, reducing the
number of generations required for the clusters to reach their respective optimal genomes
(applications), therefore directly accelerating the evolutionary self-organisation in determining
applications (Agent-sequences) to user requests. So, accelerating the responsiveness of the
Digital Ecosystem to the user base, and the process of ecological succession [29].
This augmentation has considerable potential to optimise the evolving Agent Populations of
Digital Ecosystems, but to be effective the determination of clusters needs to be computationally
negligible, otherwise the overall effect would be counterproductive. While evolving Agent
Populations would find the optimal application (Agent-sequence) within fewer generations,
more time overall would be required.
Our work on clustering with Physical Complexity, from section 3.2.2.3, may prove useful for
this augmentation.
migrating AgentHabitatAgent-poolAgentPopulationapplication(Agent-sequence)Population(GA)Clustering CatalystHHH 4.2. Alternative Augmentations ,
145
4.2.2 Replacement Aggregator
Evolutionary computing [90] was chosen exclusively for the aggregation (combinatorial optimi-
sation [240]) of the Agents into optimal Agent-sequences (applications), without comparison to
other techniques, because are focus was creating the digital counterpart of biological ecosystems.
If we were to assume it might not be the optimal technique, we could consider a replacement
aggregator to perform the aggregation of the Agents with an alternative technique, potentially
accelerating the responsiveness of the Digital Ecosystem to the user base, and the process of
ecological succession [29].
Replacement Aggregator: This would work by treating the evolving Agent
Figure 4.2:
Population,
the embodiment of evolutionary computing in Digital Ecosystems, as an
interchangeable module, and considering a replacement aggregator to perform the aggregation
of the Agents with an alternative technique, potentially accelerating the responsiveness of the
Digital Ecosystem to the user base, and the process of ecological succession [29].
This augmentation would work by treating the evolving Agent Population, the embodiment of
evolutionary computing in Digital Ecosystems, as an interchangeable module. It also assumes
a more effective aggregator can be found to perform the combinatorial optimisation [240] that
occurs in response to a user request, on the set of Agents and Agent-sequences available from
the Agent-pool of a Habitat. As each Agent aggregation process within the Digital Ecosystem
would be accelerated, the entire ecosystem would operate more efficiently.
migrating AgentHabitatAgent-poolAgentReplacement Aggregatorapplication(Agent-sequence)Replacement AggregatorHHH 146
, Chapter 4. Optimisation of Digital Ecosystems
This augmentation could even allow for a range of available aggregators, choosing the most
effective depending on the user, or on a case-by-case basis. However, replacing the evolutionary
mechanism of the Digital Ecosystem with an alternative technique would weaken its Ecosystem-
Oriented Architecture; potentially risking the loss of valuable behaviour, such as emergent self-
organisation, scalability and sustainability, imbibed from creating the digital counterpart of
biological ecosystems. So, while the modular nature of the Ecosystem-Oriented Architecture
of Digital Ecosystems makes this augmentation possible, it would not be prudent.
4.2.3 Agent-Pool Aggregation
The appealing vision of the Agent-pool aggregation is that of the Agents intelligently
recombining with one another, joining and leaving Agent-sequences of their own accord to
improve the responsiveness of the Digital Ecosystem, allowing for the creation of potentially
useful applications (Agent-sequences) or partial applications inside the Agent-pools, increasing
Figure 4.3: Agent-Pool Aggregation: This augmentation allows for the creation of potentially
useful applications (Agent-sequences) or partial applications inside the Agent-pools, increasing
and optimising the recombination that occurs globally within the Digital Ecosystem. So, it would
help to optimise the Agent-sequences at the Agent-pools of the Habitats, which would in turn
optimise the evolving Agent Populations, as they make use of the Agent-pools.
migrating AgentHabitatAgent-poolAgentXXXPopulation(GA)Agent-Pool AggregationHHH 4.2. Alternative Augmentations ,
147
and optimising the recombination that occurs globally within the Digital Ecosystem. So,
it would help to optimise the Agent-sequences at the Agent-pools of the Habitats, which
would in turn optimise the evolving Agent Populations, as they make use of the Agent-pools
when determining applications (Agent-sequences) to user requests. Therefore, accelerating the
process of ecological succession [29], and so the responsiveness of the Digital Ecosystem to the
user base.
This augmentation would work by providing the Agents with the opportunity to interact inside
the Agent-pools and the ability to determine whether to recombine with one another, outside the
evolutionary optimisation of the evolving Agent Populations. For the Agents to judge potential
re-combinations, they will require an understanding of the context in which they would operate,
most importantly the past user requests of the Habitat where the recombination would occur.
So, this augmentation would optimise the set of Agents and Agent-sequences at the Agent-
pools, and therefore indirectly optimise and accelerate the evolving Agent Populations within
the Digital Ecosystem. As each evolving Agent Population within the Digital Ecosystem would
be accelerated, the entire ecosystem would operate more efficiently.
This augmentation would strengthen the Agent concept within the Ecosystem-Oriented
Architecture of Digital Ecosystems, endowing the individual Agents with some intelligence
and control over their behaviour. However, a sophisticated process would be required for the
Agents to evaluate a potential recombination, considering their descriptions with other Agents’
descriptions collectively, within the context of the past user requests handled by the Habitat
where the recombination is to occur. Additionally, a scalable mechanism would be required
to determine which re-combinations the Agents should evaluate, because generally it will be
impractical to evaluate all the re-combinations possible at any one time.
Interestingly, the
effectiveness of this augmentation relies on the local interactions of the Agents, producing an
emergent global optimising effect on the evolving Agent Populations to accelerate the ecological
succession of a Digital Ecosystem.
4.2.4 Targeted Migration
The self-organised diversity experiments from the previous chapter, showed that the Digital
Ecosystem can be slow to optimally distribute the Agents, within the Habitat network, relative
to the user request behaviour. So, potential exists to optimise the distribution of the Agents
within the Habitat network, through additional targeted migration of the Agents, which would
148
, Chapter 4. Optimisation of Digital Ecosystems
Figure 4.4: Targeted Migration: This augmentation would optimise the distribution of the
Agents within the Habitat network, through additional targeted migration of the Agents, helping
to optimise the Agents found at the Agent-pools of the Habitats, which would in turn optimise
the evolving Agent Populations. So, accelerating the process of ecological succession [29], and
therefore the responsiveness of the Digital Ecosystem to the user base.
indirectly optimise the evolving Agent Populations. The migration probabilities between the
Habitats produces the existing passive Agent migration, allowing the Agents to spread in the
correct general direction within the Habitat network, based primarily upon success at their
current location. This augmentation will work in a more active manner, allowing the Agents
highly targeted migration to specific Habitats, in addition to the generally directed passive
migration. It will help to optimise the Agents found at the Agent-pools of the Habitats, which
would in turn optimise the evolving Agent Populations as they make use of the Agent-pools
when determining applications (Agent-sequences) to user requests. So, accelerating the process
of ecological succession [29], and therefore the responsiveness of the Digital Ecosystem to the
user base.
This augmentation would work by providing the Agents with the opportunity to interact inside
the Agent-pools, outside of the evolutionary optimisation of the evolving Agent Populations, to
determine if they are functionally similar based on their semantic descriptions. Similar Agents
will compare their migration histories to determine Habitats where they could find a niche (i.e.
migrating AgentHabitatAgent-poolAgenttemporary link for targeted migrationPopulation(GA)Targeted MigrationHHH 4.2. Alternative Augmentations ,
149
be valuable). This would lead to additional highly targeted migration of Agents throughout
the Habitat network, optimising the set of Agents and Agent-sequences at the Agent-pools,
and therefore indirectly optimising and accelerating the evolving Agent Populations within the
Digital Ecosystem. As each evolving Agent Population within the Digital Ecosystem would be
accelerated, the entire ecosystem would operate more efficiently.
This augmentation would strengthen the Agent concept within the Ecosystem-Oriented
Architecture of Digital Ecosystems, endowing the individual Agents with some intelligence
and control over their behaviour.
Interestingly, the effectiveness of this augmentation relies
on the local interactions of the Agents, producing an emergent global optimising effect on the
evolving Agent Populations to accelerate the ecological succession of a Digital Ecosystem.
4.2.5 Choice of Augmentation
Figure 4.5: Effect of The Proposed Augmentations: The effect of the different augmentations
on the evolutionary dynamics (the evolving Agent Populations) and the ecological dynamics (the
Habitats). This separation of concerns is an artificial construct, but useful in summarising the
potential of the different augmentations, before we decide upon which to pursue.
The question of which of the alternative augmentations are most promising, and therefore
which we should pursue to theoretical completion and then experimental confirmation, is not
obvious. The evolving Agent Populations are the embodiment of evolutionary computing [90]
in Digital Ecosystems, while the Habitats are the embodiment of the ecology-based computing
we have developed for Digital Ecosystems. So, we will start by considering the effect of the
different augmentations on the evolutionary dynamics (the evolving Agent Populations) and
the ecological dynamics (the Habitats), as shown in Figure 4.5. This separation of concerns is
Evolutionary Dynamics(effect on)Ecological Dynamics(effect on)directindirectdirectindirectTargeted MigrationClusteringCatalystAgent-Pool AggregationReplacement AggregatorEcosystem Dynamics(effect on) 150
, Chapter 4. Optimisation of Digital Ecosystems
an artificial construct, but useful in summarising the potential of the different augmentations,
before we decide upon which to pursue.
The clustering catalyst has potential to accelerate evolving Agent Populations with clusters,
and therefore the process of ecological succession [29]. However, a computationally negligible
technique is needed to determine the required clustering, else the overall effect will be
counterproductive. Nevertheless, we will pursue this augmentation further; first to theoretical
completion, and then to experimental simulations for confirmation.
The replacement aggregator could prove effective for using a range of techniques when finding
the optimal aggregation of the Agents into Agent-sequences in response to user requests.
However, it would weaken the Ecosystem-Oriented Architecture of Digital Ecosystems and
presume more effective techniques than evolutionary computing [90] can be found for the
combinatorial optimisation [240] of the Agent aggregation. The first point would obviously
be undesirable, potentially risking the loss of valuable behaviour, such as emergent self-
organisation, scalability and sustainability,
imbibed from creating the digital counterpart
of biological ecosystems. Regarding the second point,
it has been shown [270] that when
considering the combinatorial optimisation of Agent aggregation as the weighted set-cover
problem that evolutionary computing and simulated annealing are more effective than steepest
descent, Tabu search, and random search. Also that evolutionary computing is more widely
applicable (without performance degradation) than simulated annealing [270]. So, while there
may be other applicable techniques that have not been evaluated, the certainty of success with
this augmentation is considerably reduced, and therefore will not be pursued further.
The Agent-pool aggregation could prove very effective in optimising and accelerating the process
of ecological succession [29], but the computational cost would be considerable, most likely
making it impractical. So, while we are optimistic regarding the potential success of this
augmentation theoretically, the experimental impracticality leads us not to pursue it any
further.
The targeted migration also has considerable potential to optimise and accelerate the process of
ecological succession [29], by improving the migration of Agents through the Habitat network
of the Digital Ecosystem. Also, it would directly address the scope for optimisation identified
from the self-organised diversity experiment of section 3.4.2.2. Furthermore, it is the only
augmentation to effect the ecological dynamics directly, which makes it desirable to pursue,
because while evolution may be well understood in computer science under the auspices of
4.3. Clustering Catalyst ,
151
evolutionary computing [90], ecology until our efforts had not been widely explored. So,
there is inherently more potential to improve the ecological dynamics than the evolutionary
dynamics, and therefore more potential in this augmentation than the others. However, creating
a mechanism for the Agents to determine if they are functionally similar with one another
based on their semantic descriptions will be a challenge. Nevertheless, we will pursue this
augmentation further; first to theoretical completion, and then to experimental simulations for
confirmation.
4.3 Clustering Catalyst
The clustering catalyst will directly optimise the evolutionary self-organisation of evolving
Agent Populations with clusters, by encouraging intra-cluster crossover. Crossover involves the
crossing of two Agent-sequences in the creation of new Agent-sequences, during the replication
stage of the evolutionary cycle [13]. Theoretical completion of the clustering catalyst requires
consideration of how best to determine the clusters within an evolving Agent Population, so
we will now consider suitable clustering algorithms.
4.3.1 Clustering
We have understood clustering, within the context of evolving Agent Populations, as the
amassing of same or similar sequences around an optimum genome [29], but more generally
clustering is the classification of objects into different groups, or more precisely, the partitioning
of a data set into subsets (clusters), so that the data in each subset share some common trait,
often proximity according to a distance measure [139].
If the number of clusters k is not
apparent from prior knowledge, several methods are available for its determination [210]. For
our simulations we will make use of prior knowledge to determine the number of clusters k,
because if available it is obviously the most effective method, and because our focus is on
determining the effectiveness of the clustering catalyst. Naturally, the most suitable method
for determining the number of clusters can be investigated if the clustering catalyst proves to
be effective.
An important step in any clustering is to select a distance measure, which calculates the
similarity of two elements [139]. We will use a distance measure based on our simulated fitness
152
, Chapter 4. Optimisation of Digital Ecosystems
function from section 2.3.3, because it is itself based on a distance metric. So, given two Agent-
sequences A and B, consisting of a set of attributes a1, a2, ... and b1, b2, ... respectively, the
distance between them will be
distance(A, B) =
(cid:88)
b∈B
b − a,
(4.1)
where a is the member of A such that the difference to the required attribute b is minimised.
A range of clustering algorithms are available, to the extent that taxonomies having been
proposed [113, 142]
for their classification.
The top-level classification being between
hierarchical and partitional algorithms, both of which we shall now explore.
4.3.1.1 Hierarchical Clustering
Hierarchical clustering builds (agglomerative), or breaks up (divisive), a hierarchy of clusters
[142]. The traditional representation of this hierarchy is a tree (called a dendrogram), with
individual elements at one end and a single cluster containing every element at the other [139].
Agglomerative algorithms begin at the leaves of the tree, whereas divisive algorithms begin at
the root [142].
Agglomerative clustering starts with all the objects as individual clusters, which are then
merged according to their similarities until all are fused into a single cluster, with the most
similar objects being grouped first [139]. The similarity criterion is determined by the linkage
analysis, for which there are three common forms [239]:
• Single-link (nearest neighbour or minimum distance) [291]: is obtained by fusing clusters
according to the distance between their nearest members.
• Complete-link (farthest neighbour or maximum distance) [151]:
clusters according to the distance between their farthest members.
is obtained by fusing
• Average-link (average distance) [239]:
is obtained by fusing clusters according to the
average distance between pairs of members in the respective sets.
Most hierarchical agglomerative clustering algorithms are variants of the single-link, complete-
link and average-link algorithms [142]. Most notably, the minimum-variance algorithm [324],
which is a variant of the average-link algorithm, except instead of minimising an average
4.3. Clustering Catalyst ,
153
distance it minimises a squared distance weighted by cluster size [220]. The single-link and
complete-link algorithms are the most popular [142]. However, the single-link algorithm suffers
from a chaining effect [225], which tends to produce clusters that are straggly or elongated
[142].
In contrast, the complete-link algorithm produces tightly bound or compact clusters
[15]. While the average-link algorithm [239] is designed to reduce the dependence of the cluster-
linkage criterion on extreme values, such as the most similar or dissimilar of the single-link and
complete-link algorithms [239], and results in clusters that tend to have approximately equal
within-cluster variability [167]. The minimum-variance algorithm tends to join clusters with a
small number of observations first, being strongly biased to producing clusters with the same
number of observations, and therefore is very sensitive to outliers [211].
Divisive clustering methods start with one cluster containing all the objects, which are
successively separated into smaller subgroups until the number of clusters equals the number of
objects [142]. There are two forms: monothetic, which divides the data by the possession of a
single specified attribute, and polythetic, where divisions are based on several attributes [142].
Agglomerative algorithms make clustering decisions based on local patterns without initially
considering the global distribution, and these early decisions cannot be undone; while divisive
clustering benefits from complete information about the global distribution when making top-
level partitioning decisions [190].
It also has the advantage of being more efficient if we do
not generate a complete hierarchy all the way down to the individual objects [190]. However,
divisive clustering is conceptually more complex than agglomerative clustering, since a second
flat clustering algorithm is required as a subroutine [190]. There are also graph-theoretic divisive
clustering algorithms, with the best-known based on the construction of the minimal spanning
tree of the data, deleting the edges with the largest lengths to generate the clusters [340].
4.3.1.2 Partitional Clustering
A partitional clustering algorithm determines a single partition of the data [142], instead of
a clustering structure such as the dendrogram produced by a hierarchical algorithm [139].
Partitional algorithms usually produce clusters by optimising a criterion function defined either
locally (on a subset of the patterns) or globally (defined over all the patterns) [142]. A
combinatorial search for the set of possible labellings, to determine the optimum value of a
criterion, is clearly computationally prohibitive, and so in practise the algorithm is run multiple
times with different starting states, with the best configuration being used as the output of
154
, Chapter 4. Optimisation of Digital Ecosystems
the clustering [142]. Partitional algorithms have advantages in applications involving large
data sets for which the construction of a dendrogram is computationally prohibitive [142]. A
problem accompanying partitional algorithms is choosing the number of desired output clusters
beforehand [84], which is not required for hierarchical clustering. There are many forms of
partitional clustering algorithms, including mixture-resolving [139], mode-seeking [139], nearest
neighbour [180], fuzzy clustering [339], artificial neural networks [280], and others [142].
The most intuitive and frequently used criterion function in partitional clustering algorithms
is the squared error criterion, which tends to work well with isolated and compact clusters
[142]. The most commonly used algorithm employing a squared error criterion is the k-means
algorithm [187], popular because it is easy to implement [147]. It starts with a random initial
partition and keeps reassigning the patterns to clusters, based on the similarity between a
pattern and the cluster centres, until a convergence criterion is met [187]. However, a major
problem with this algorithm is that it is sensitive to the selection of the initial partition, and
may converge to a local optimum of the criterion function value if the initial partition is not
properly chosen [105].
4.3.1.3 Choice of Algorithm
The choice of the optimal clustering algorithm very much depends on the structure of the data,
because clustering is subjective, such that the same data can be partitioned differently for
different purposes [142]. We also require a clustering algorithm with a negligible computational
cost for the clustering catalyst to be effective. So, we choose a hierarchical clustering algorithm
over a partitional one, as it is more appropriate for small data sets [142], such as expected
from an evolving Agent Population with clusters. We choose a hierarchical agglomerative
clustering algorithm over a divisive one, because it is conceptually simpler [190], and because
the efficiency advantage [190] of a divisive algorithm would be negligible, given the small size of
the expected data set. Finally, we choose a hierarchical agglomerative average-link clustering
algorithm of time complexity O(n2 log n) [190], over a single-link one of O(n2) time complexity,
a complete-link one of O(n2 log n) time complexity, or a minimum-variance one of O(n2) time
complexity [71], because the average-link algorithm is designed to reduce the dependence of
the cluster-linkage criterion on extreme values, such as the most similar or dissimilar of the
single-link and complete-link algorithms [239]; and because the minimum-variance algorithm is
biased to producing clusters with the same number of objects [211], which would be problematic
4.3. Clustering Catalyst ,
155
for clusters emerging over the generations. Any difference in execution time of the algorithms
would be minimal, despite the different algorithmic time complexities, because of the small size
of the expected data set n, the Population size. So, we choose a hierarchical agglomerative
average-link clustering algorithm [239] for our clustering catalyst.
4.3.2 Physical Complexity Clustering
We also considered a clustering algorithm based on our extended Physical Complexity from
Chapter 3, because clustering is subjective in nature [142], and our extended Physical
Complexity was developed, in section 3.2.2.3, to understand the clustering of evolving Agent
Populations.
In our algorithm, with the number of clusters k determined also from prior
knowledge, we first sort the evolving Agent Population, then process its Agent-sequences
linearly, adding each to the cluster that maximises the Efficiency Ec (3.42) of the Population.
The pre-assignment sorting ensures that the cores of the clusters are established for the
Efficiency Ec to have the necessary sensitivity when assigning the Agent-sequences of greater
uniqueness. The pseudocode for our algorithm is shown in Figure 4.6.
clusters[k];
group duplicate Agent-sequences within the Population
order the groups within Population by greatest size
for each Agent-sequence in the ordered Population
XXif a duplicate and not first instance
XXXX then: assign to same cluster as first instance
XXXX else: the cluster that maximises the Efficiency Ec of the Population
XXend if
end for
Figure 4.6: Pseudo-Code for Physical Complexity Clustering: With the number of clusters k
determined from prior knowledge, we first sort the evolving Agent Population, then process its
Agent-sequences linearly, adding each to the cluster that maximises the Efficiency Ec (3.42)
of the Population. The pre-assignment sorting is for the Efficiency Ec to have the necessary
sensitivity when assigning the Agent-sequences of greater uniqueness.
Our algorithm will be computationally negligible, because, based on the pseudo-code, it will
have a time complexity of O(n2), where n is the Population size. It will also be as accurate
156
, Chapter 4. Optimisation of Digital Ecosystems
as an exhaustive search, which would have and exponential time complexity, because of the
sensitivity of the Efficiency Ec when assigning the Agent-sequences to the clusters.
Now that the clustering catalyst is theoretically complete, with two alternative clustering
algorithms, we can confirm its effect experimentally through simulations.
4.4 Targeted Migration
The targeted migration will directly optimise the ecological migration, and therefore indirectly
complement the evolutionary self-organisation of the evolving Agent Populations, through the
highly targeted migration of the Agents to their niche Habitats. The migration probabilities
between the Habitats produces the existing passive Agent migration, allowing the Agents to
spread in the correct general direction within the Habitat network, based primarily upon success
at their current location. The targeted migration will work in a more active manner, allowing the
Agents highly targeted migration to specific Habitats, based upon their interaction with one
another to discover Habitats where they could be valuable (i.e. find a niche). Theoretical
completion of the targeted migration requires further consideration of how it will operate,
including its effect on the Agent life-cycle, and a suitable pattern recognition [140] technique
for the required similarity recognition.
The targeted migration will occur when users deploy their services, specifically when deploying
their representative Agents to their Habitats within the Digital Ecosystem, and upon the
execution of applications (groups of services), specifically the resulting passive migration of
their representative Agent-sequences between the Habitats. The Agent-sequences arriving at
Habitats, with respect to the targeted migration, will be treated as individual Agents arriving
at the Habitats. So, an Agent arriving at a Habitat interacts one-on-one with Agents already
present within the Agent-pool of the Habitat, and upon determining functional similarity,
based upon comparing their semantic descriptions, will share other Habitats successfully visited
from their respective migration histories. An Agent migration history, as defined in section
2.2.2.1, is the migratory path of the Agent through the Habitat network, including its use at
the Habitats visited. So, similar Agents can share their migration histories to discover new
Habitats where they could be valuable, and then use targeted migration (via a copy, and not
a move) to explore the most promising of the recently acquired Habitats. This will allow
successfully interacting Agents to target specific Habitats where they will potentially be useful,
4.4. Targeted Migration ,
157
Figure 4.7: Agent Life-Cycle With Targeted Migration: The Agent life-cycle, defined in section
2.2.4.3, will change to support the targeted migration, as shown by the blue circle. Specifically,
there will be more opportunities for Agent migration, but more importantly these opportunities
will be for targeted migration, which will help to optimise the set of Agents found at the
Habitats.
but risks potentially infinite targeted migration, because targeted migration itself can lead to
further targeted migration. So, each Agent will require a dynamic targeted migrations counter,
which defines the number of permitted targeted migrations of the Agent. This counter will be
incremented upon an Agent’s execution in response to a user request, and decremented upon
performing a targeted migration.
The Agent life-cycle, defined in section 2.2.4.3, will change to support the targeted migration,
as shown in Figure 4.7 by the blue circle. Specifically, there will be more opportunities for
Agent migration, but more importantly these opportunities will be for targeted migration,
which will help to optimise the set of Agents found at the Habitats, and therefore support the
evolving Agent Populations created in response to user requests for applications. The targeted
migration will essentially short-circuit the hierarchical topology of the Habitat network, which
is what allows it to specialise and localise to communities, providing specific solutions to specific
requests from specific users. However, the targeted migration will also reinforce the hierarchical
topology of the Habitat network, because targeted migration between connected Habitats will
accelerate the existing migration of Agents, while between unconnected Habitats will assist
the Digital Ecosystem in supporting emerging communities. So, the targeted migration will
help strengthen and catalyse the formation of clusters within the Habitat network, and will
also assist in locating Habitats within the correct clusters. Therefore, the optimisation of the
Digital Ecosystem will be a global emergent effect resulting from the local interactions of the
Agents, allowing for niches to be fulfilled faster and so accelerating the process of ecological
evolving Agent Populationfor solution (Agent-sequence)register solution at Agent-pool of HabitatexecuteAgent-sequenceuser requestAgent migrationdeploy Agent to Agent-pool of HabitatAgent to Agentinteraction 158
, Chapter 4. Optimisation of Digital Ecosystems
succession [29]. Also, the Digital Ecosystem will adapt faster to changing environmental
conditions (e.g. changes in the request behaviour of user communities). In biological terms
the targeted migration endows the Agents with a form of reciprocal altruistic behaviour [311],
consistent with the Agent paradigm of the Ecosystem-Oriented Architecture.
4.4.1 Similarity Recognition
For the targeted migration to work successfully an effective technique will be required for the
similarity recognition between the semantic descriptions of two Agents. Each Agent will have an
embedded similarity recognition component to maintain the consistency of the Agent paradigm
of Ecosystem-Oriented Architectures. So, the Agents will interact one-on-one to determine
functional similarity based upon their semantic descriptions, using their embedded similarity
recognition components, with each of the two interacting Agents determining similarity for
themselves. Again, this is to maintain the consistency of the Agent paradigm. Similarity
recognition between the semantic descriptions of two Agents will require some form of pattern
recognition, because there is no single standard for the semantic description of services [42],
and adopting one over the others would be inconsistent with the inclusive nature of Digital
Ecosystems. So, we will now consider the field of pattern recognition to determine suitable
techniques for the similarity recognition components to be embedded within the Agents.
Pattern recognition aims to classify data (patterns) based on priori knowledge or on statistical
information extracted from the data [262]. Pattern recognition requires a sensor or sensors
for data acquisition, a pre-processing technique, a data representation scheme, and a decision
making model [140]. Also, learning from a set of examples (training set) is an important and
desirable feature of most pattern recognition systems [262]. The four best known approaches
for pattern recognition are [140]:
• Template Matching
• Statistical Classification
• Structural Matching
• Neural Networks
These approaches are not necessarily independent, and sometimes the same pattern recognition
method exists with different interpretations [140]. For example, attempts have been made to
design hybrid systems involving multiple approaches, such as the notion of attributed grammars
which unifies structural and statistical pattern recognition [102].
4.4. Targeted Migration ,
159
4.4.1.1 Template Matching
Template Matching is the simplest and earliest approach to pattern recognition, and involves
a generic operation to determine the similarity between two entities (points, curves, or shapes)
of the same type [140]. A template (typically, a 2D shape) is available, with the pattern
to be recognised being matched against the stored template, while considering all allowable
changes translation, rotation and scale [140]. The similarity measure, often a correlation, may
be optimised based on a training set, and often the template itself is defined from a training
set [140].
Template Matching is computationally demanding, but the availability of ever faster processors
has made it more feasible [140]. While effective for some application domains, it has several
disadvantages [233]. For instance, poor performance if the patterns are distorted from the
imaging process or a viewpoint change, or if there are large intra-class variations among the
patterns [140]. Deformable template models [121] or rubber sheet deformations [16] can help
to compensate when the deformation cannot be easily explained or modelled directly.
4.4.1.2 Statistical Classification
In Statistical Classification each pattern is represented in terms of d features or measurements,
and is viewed as a point in a d-dimensional feature space, with the goal being to choose
those features that allow pattern vectors belonging to different categories to occupy compact
and disjoint regions in the d-dimensional feature space [328]. The effectiveness of which is
determined by how well patterns from different classes can be separated [140]. The decision
boundaries of the d-dimensional feature space can be determined from probability distributions,
of the patterns belonging to each class, which must be specified or learnt [80, 86].
One can also take a discrimination analysis based approach to classification,
in which a
parametric form of a decision boundary is specified, and then the best decision boundary of the
specified form is found based on the classification of training patterns [140]. Such boundaries
can be constructed using, for example, a mean squared error criterion [328]. These direct
boundary construction approaches are supported by the philosophy [317] that if you possess a
restricted amount of information for solving some problem, try to solve the problem directly
and never solve a more general problem as an intermediate step, because it is possible that the
160
, Chapter 4. Optimisation of Digital Ecosystems
available information is sufficient for a direct solution but insufficient for solving a more general
intermediate problem.
Statistical approaches are generally characterised by having an explicit underlying probability
model, which provides a probability of being in each class and not just a classification, and
hence some human intervention is assumed regarding variable selection and transformation,
and the overall structuring of the problem [206].
4.4.1.3 Structural Matching
In Structural Matching a hierarchical perspective is adopted, where a pattern is viewed as
being composed of simple sub-patterns, which are built from yet simpler sub-patterns [246],
and therefore it is applicable to many recognition problems involving complex patterns [140].
The elementary (simplest) sub-patterns to be recognised are called primitives, with the given
complex pattern to be represented in terms of the interrelationships between these primitives
[246]. Where the structure is syntactic, a formal analogy is drawn between the structure of
patterns and the syntax of language. So, the primitives are viewed as the alphabet of the
language, and the patterns are viewed as sentences generated according to the grammar of
the language [102]. Thus, a large collection of complex patterns can be described by a small
number of primitives and grammatical rules, which must be inferred from the available training
samples [140].
Structural pattern recognition is intuitively appealing because, in addition to classification, it
also provides a description of how the given pattern is constructed from the primitives [140].
It is used in situations where the patterns have a definite structure that can be captured by a
set of rules, such as electrocardiogram waveforms [310], textured images [129], and the shape
analysis of contours [179]. However, the implementation of structural approaches leads to many
difficulties, including the segmentation of noisy patterns (to detect primitives) and the inference
of grammar from training data [140]. There can also be a combinatorial explosion of possibilities
to be investigated, demanding large training sets and significant computational effort [248].
4.4.1.4 Neural Networks
Neural Networks (NNs) can be viewed as massively parallel computing systems consisting of
an extremely large number of simple processors with many interconnections [262]. NN models
4.4. Targeted Migration ,
161
attempt to use certain organisational principles (such as learning, generalisation, adaptivity,
fault tolerance, distributed representation, and computation) in a network of weighted directed
graphs, in which the nodes are artificial neurons, and the directed edges (with weights) are
connections between the neuron outputs and inputs [262]. The main characteristics of NNs
are their ability to learn complex nonlinear input-output relationships, use sequential training
procedures, and adapt themselves to the data [140].
The most commonly used family of NNs for pattern classification tasks is the feed-forward
network, including multilayer perceptrons, which are organised into layers and has unidirec-
tional connections between the layers [140]. Another popular network is the Self-Organising
Map, or Kohonen-Network [153], which is often used for feature mapping [140]. The increasing
popularity of NN models to solve pattern recognition problems has been primarily because of
their low dependence on domain-specific knowledge (relative to model-based and rule-based
approaches) and the availability of efficient learning algorithms [140]. The learning process
involves updating the network architecture and connection weights so that a network can
efficiently perform a specific classification [262].
NNs provide a suite of nonlinear algorithms for feature extraction (using hidden layers) and
classification (e.g. multilayer perceptrons) [140].
In addition, existing feature extraction
and classification algorithms can be mapped onto NN architectures for efficient (hardware)
implementation [39]. Despite the seemingly different underlying principles, most of the well-
known NN models are implicitly equivalent or similar to classical statistical pattern recognition
methods [140]. However, NNs offer several advantages, such as unified approaches for feature
extraction and classification, and flexible procedures for finding good, moderately nonlinear
solutions [140].
4.4.1.5 Support Vector Machines
One of the most interesting recent developments in classifier design is the introduction of the
Support Vector Machine (SVM) [316], which is primarily a two-class classifier, and therefore
highly suitable for the required similarity recognition component of our targeted migration. It
uses an optimisation criterion that is the width of the margin between the classes, i.e. the empty
area around the decision boundary defined by the distance to the nearest training patterns [41].
These patterns, called support vectors, define the classification function, and their number is
minimised by maximising the margin [41]. This is achieved through a kernel function K, which
162
, Chapter 4. Optimisation of Digital Ecosystems
transposes the data into a higher-dimensional space where a hyperplane performs the separation
[41]. In its simplest form the kernel function is just a dot product between the input pattern
and a member of the support set, resulting in a linear classifier, while nonlinear kernel functions
lead to a polynomial classifier [140].
SVMs are closely related to Neural Networks, being a close cousin to classical multilayer
perceptrons, with the use of a sigmoid kernel function making them equivalent to two-layer
perceptrons [1]. However, in the training of NNs, such as multi-layer perceptrons, the weights
of the network are found by solving a non-convex unconstrained minimisation problem, while
the use of a kernel function in SVMs solves a quadratic programming problem with linear
constraints [276].
An important advantage of SVMs is that they offer the possibility to train generalisable
nonlinear classifiers in high-dimensional spaces using a small training set [313]. Furthermore,
for large training sets a small support set is typically selected for designing the classifier, thereby
minimising the computational requirements during training [313].
4.4.1.6 Choice of Technique
Template Matching is not suitable for the required pattern recognition of our targeted migration,
because its effective use is domain specific [140] and the similarity recognition between the
semantic descriptions of Agents is very different to the domains that it is typically applied
[233]. Statistical Classification is also not suitable, because the embedded similarity recognition
component of each Agent would require human intervention for variable selection and
transformation [206]. Structural Matching is suitable theoretically, but implementations lead
to many difficulties [140], including the segmentation of noisy patterns (to detect primitives)
and the inference of grammar from training data [140]. There can also be a combinatorial
explosion of possibilities to be investigated, demanding large training sets and significant
computational effort [248], neither of which is available. Neural Networks are suitable, given
their low dependence on domain-specific knowledge and the availability of efficient learning
algorithms [140]. Support Vector Machines, albeit a recent development [316], are also suitable
[145], being primarily a binary classifier [140] for training generalisable nonlinear classifiers in
high-dimensional spaces using small training sets [313]. So, we will make use of both Neural
Networks and Support Vector Machines for the theoretical completion and implementation of
our targeted migration.
4.4. Targeted Migration ,
163
4.4.2 Neural Networks
So, in the first instance, we will leverage the pattern recognition capabilities of Neural Networks
(NNs) for the embedded similarity recognition components of the Agents, allowing them to
determine similarity to one another based on the similarity of their semantic descriptions.
We will use multilayer perceptrons (feed-forward artificial NNs) with backpropagation [131] to
provide the required pattern recognition behaviour, because of their ability to solve problems
stochastically, which allows for approximate solutions to extremely complex problems [131].
They are a modification of the standard linear perceptron [268], using three or more layers
of neurons (nodes) with nonlinear activation functions to distinguish data that is not linearly
separable, or separable by a hyperplane [35]. The power of the multilayer perceptron comes
from its similarity to certain biological neural networks in the human brain, and because of
their wide applicability has become the standard algorithm for any supervised-learning pattern
recognition process [131].
A pre-processing [35] of Agent semantic descriptions will be required that is consistent across
the entire Digital Ecosystem, requiring an alphabetical ordering of the attribute tuples within a
semantic description, a standardisation of the length of the attributes, before finally making use
of a binary encoding for processing by a NN [35]. The assumption of information structured as
tuples, including an attribute name and attribute value, is accurate for our simulated semantic
descriptions, but is also a reasonable assumption for any semantic description of web services
[87, 205, 72, 42]. To standardise the length of the attributes, after removing any white-space1,
an average word length of six characters will be used, because 5.39 is the average word length
for business English [100]. For the binary encoding we propose using Unicode (UTF-8), which
is based on extending ASCII to provide multilingual support [108]. However, ASCII’s support
of only English [108] will be sufficient for our simulations. The size (number of neurons) of the
input layer [131] will be proportional to the semantic description of the Agent in which the
NN is embedded, taking advantage of the variation in length of different semantic descriptions,
which will assist the NN-based pattern recognition in determining dissimilarity.
We will use a single hidden layer, which is usually sufficient for most tasks [35]. The size of which
will be determined through exploratory programming [294] in our simulations, because of the
difficulty in determining the optimal size without training several networks and estimating the
1 A white-space is any single character or series of characters that represent horizontal or vertical space in
typography [101].
164
, Chapter 4. Optimisation of Digital Ecosystems
Figure 4.8: Neural Network for the Similarity Recognition Component of Agents in Targeted
Migration: Consisting of an input layer proportional to the semantic description of the Agent in
which it is embedded. A single hidden layer, and an output layer consisting of a single neurone
to provide a binary response to the question of whether another Agent’s semantic description
is similar.
generalisation error [274], evident by the range of inconsistent rules of thumb [37, 302, 34, 38]
available to define the optimal size. The output layer [35] will consist of a single neurone to
provide a binary (true or false) response to the question of whether another Agent’s semantic
description is similar to the Agent’s own semantic description. We will use a threshold of 0.90
on its output for the determination of similarity. The overall structure of the Neural Network
is visualised in Figure 4.8.
Multilayer perceptrons use nonlinear activation functions, which were developed to model the
frequency of action potentials (firing) of biological neurons in the brain [131]. The main
activation function used in current applications is the sigmoid function [131], a hyperbolic
tangent that is normalised and in which the output y of a neurone is the sum of the weighted
input values x [35],
y =
1
(1 + e−x)
.
(4.2)
The weights x between the neurons will be randomly initialised, then trained to the real numbers
that provide the desired functionality, because learning occurs in the perceptron by changing
the connection (synaptic) weights after each piece of data is processed, based on the error of
the output compared to the expected result [35]. This is an example of supervised learning and
is carried out through backpropagation, a generalisation of the least mean squares algorithm
InputHiddenOutput 4.4. Targeted Migration ,
165
[131]. The network is therefore trained by providing it with input and corresponding output
patterns [35].
The NN-based embedded similarity recognition component of an Agent will be trained when
the Agent is deployed to a Habitat of the Digital Ecosystem. The initial training set will consist
of the semantic description of the Agent as a positive match, and variants created from its own
semantic description. If the variant is less than 10% different it will be processed as a positive
match, else it will be processed as a negative match. The training set can be extended based on
experience, making use of when an Agent visits a Habitat through targeted migration (i.e. one
acquired from an inter-Agent interaction); if visiting the Habitat proves successful the semantic
description of the interacting Agent can be appended to the training set as a positive match,
else as a negative match.
4.4.3 Support Vector Machines
In the second instance, we will
leverage the pattern recognition capabilities of Support
Vector Machines (SVMs) for the embedded similarity recognition components of the Agents,
allowing them to determine similarity to one another based on the similarity of their semantic
descriptions. As SVMs are closely related to Neural Networks, being a close cousin to classical
multilayer perceptrons [1], we will make use of the pre-processing and the training sets defined
in the previous subsection, which will also ensure a fair comparison of the pattern recognition
techniques in empowering the similarity recognition components of the Agents.
The selection of a suitable kernel function is important, since it defines the feature space in
which the training set is classified [62], operating as shown in Figure 4.9. A Radial Basis
Functions (RBF) is recommended for text categorisation [145], with the most common form of
the RBF being Gaussian [123].
Training a SVM requires solving a large quadratic programming (QP) optimisation problem,
which Sequential Minimal Optimisation (SMO) breaks into a series of the smallest possible
QP problems.
SMO solves these small QP problems analytically, which avoids using a
time-consuming numerical QP optimisation. SMO scales between linear and quadratic time
166
, Chapter 4. Optimisation of Digital Ecosystems
Figure 4.9: Support Vector Machine (modified from [304]): Visualisation showing the training
set in the Input Space, and its binary classification by a hyperplane in the higher dimensional
Feature Space, achieved through the kernel function. A Radial Basis Functions (RBF) is
recommended for text categorisation [145], with the most common form of the RBF being
Gaussian [123].
complexity, relative to the size of the training set, because it avoids matrix computation [250].
The alternative, a standard Projected Conjugate Gradient (PCG) chunking algorithm scales
between linear and cubic time complexity, relative to the size of the training set [250]. So SMO
is faster, up to a thousand times on real-world sparse data sets [250].
The issue of the learnt behaviour of the embedded similarity recognition component of an
Agent, whether SVM or NN based, being inherited when the Agent reproduces is known as
the Baldwin effect [19]. The Baldwin effect has always been controversial within biological
ecosystems [329], primarily because of the problem of confirming it experimentally [298]. Also,
offspring in biological ecosystems can be genetically different to their parents [29], such that
any learnt behaviour could potentially be inappropriate. However, the offspring in our Digital
Ecosystem are genetically identical to their parents (in terms of the individual Agents), and
so it makes little sense to force the loss of learnt behaviour. Therefore, we doubt that the
Baldwin effect will adversely affect our Digital Ecosystem, which we will confirm through our
simulations.
Now that the targeted migration is theoretically complete, with two alternative pattern
recognition techniques, we can confirm its effect experimentally through simulations.
kernel functionFeature SpaceInput Space 4.5. Simulation and Results ,
167
4.5 Simulation and Results
We simulated the Digital Ecosystem using our simulation from section 2.3 (unless otherwise
specified), adding the classes and methods necessary to implement the proposed clustering
catalyst and targeted migration augmentations. Each experimental scenario was run ten
thousand times for statistical significance of the means and standard deviations calculated.
4.5.1 Clustering Catalyst
We implemented the clustering catalyst as defined in sections 4.2.1 and 4.3, with the clusters
determined by hierarchical agglomerative average-link clustering or our Physical Complexity
clustering. We also made use of our simulations from Chapter 3 to simulate evolving Agent
Populations with between two and six clusters, varied according to a Gaussian distribution,
with the crossover rate increased from 10% to 25% to provide a greater opportunity for the
clustering catalyst to operate.
4.5.1.1 Control
The clustering catalyst benefited from the additional crossover, which alone could have been
responsible for any observed optimisation, because it increased recombination [168] in the
evolving Agent Populations. So, increasing variation in the exploration of solutions, potentially
reducing the number of generations required to evolve the optimal solution. Therefore, our
experimental simulations included a crossover control
for the additional crossover, which
excluded the clustering catalyst.
4.5.1.2 Hierarchical Clustering
We started with the hierarchical agglomerative average-link clustering [239] based clustering
catalyst, as defined in section 4.3.1, making use of RapidMiner [256] to perform the required
clustering. In Figure 4.10 we graphed for the simulation runs the average number of generations
required to evolve the optimal solution, for the evolving Agent Populations with the hierarchical
clustering based clustering catalyst, compared to the evolving Agent Populations with the
crossover control, and the evolving Agent Populations alone. The evolving Agent Populations
168
, Chapter 4. Optimisation of Digital Ecosystems
Figure 4.10: Graph of the Hierarchical Clustering Based Clustering Catalyst: The evolving
Agent Populations alone averaged 296 (3 s.f.) generations, while the evolving Agent Populations
with the crossover control showed a 9% reduction, averaging 267 (3 s.f.) generations. The
evolving Agent Populations with the hierarchical clustering based clustering catalyst failed to
provide any further optimisation, averaging 281 (3 s.f.) generations.
alone averaged 296 (3 s.f.) generations with a standard deviation of 12.76 (2 d.p.), while
the evolving Agent Populations with the crossover control showed a 9% reduction, averaging
267 (3 s.f.) generations with a standard deviation of 8.91 (2 d.p.). The evolving Agent
Populations with the hierarchical clustering based clustering catalyst failed to provide any
further optimisation, averaging 281 (3 s.f.) generations with a standard deviation of 20.25
(2 d.p.).
4.5.1.3 Physical Complexity Clustering
Next we considered the Physical Complexity based clustering catalyst, as defined in section 4.3.2,
making use of our simulations from section 3.2.3. In Figure 4.11 we graphed for the simulation
runs the average number of generations required to evolve the optimal solution,
for the
evolving Agent Populations with the Physical Complexity based clustering catalyst, compared
to the evolving Agent Populations with the hierarchical clustering based clustering catalyst, the
evolving Agent Populations with the crossover control, and the evolving Agent Populations
ClusteringCatalyst(Hierarchical)CrossoverControlAgentPopulations050100150200250300AverageNumberofGenerationstoEvolvetheOptimalSolutionScenario281267296 4.5. Simulation and Results ,
169
Figure 4.11: Graph of the Physical Complexity Based Clustering Catalyst: The evolving Agent
Populations with the Physical Complexity based clustering catalyst averaged 274 (3 s.f.), better
than the evolving Agent Populations with the hierarchical clustering based clustering catalyst
which averaged 281 (3 s.f.) generations, but still worse than the evolving Agent Populations
with the crossover control which averaged 267 (3 s.f.) generations.
alone. The evolving Agent Populations with the Physical Complexity based clustering catalyst
averaged 274 (3 s.f.) generations with a standard deviation of 17.53 (2 d.p.), better than
the evolving Agent Populations with the hierarchical clustering based clustering catalyst which
averaged 281 (3 s.f.) generations with a standard deviation of 20.25 (2 d.p.), but still worse
than the evolving Agent Populations with the crossover control which averaged 267 (3 s.f.)
generations with a standard deviation of 8.91 (2 d.p.).
Additional experiments were conducted in which we varied several different parameters to
determine if there were any conditions under which the clustering catalyst was effective. The
varied parameters included the population size, the mutation rate and the crossover rate.
However, extensive testing through multiple scenarios failed to show any significant reduction
in the number of generations required to evolve the optimal solution. So, confirming that the
evolving Agent Populations with the clustering catalyst were less effective than the evolving
Agent Populations with the crossover control. As the clustering catalyst was unsuccessful, in
Figure 4.12 we graphed a typical run of each scenario to observe its behaviour and so better
ClusteringCatalyst(Phy.Com.)ClusteringCatalyst(Hierarchical)CrossoverControlAgentPopulations050100150200250300AverageNumberofGenerationstoEvolvetheOptimalSolutionScenario274281267296 170
, Chapter 4. Optimisation of Digital Ecosystems
Figure 4.12: Graph of Typical Runs for the Clustering Catalysts: As the clustering catalyst
was unsuccessful, we graphed a typical run of each scenario to observe its behaviour and so
better understand why it failed. However, there was no unexpected behaviour, confirming that
the evolving Agent Populations with the clustering catalyst were simply less efficient than the
evolving Agent Populations with the crossover control.
understand why it failed. However, there was no unexpected behaviour, confirming that the
evolving Agent Populations with the clustering catalyst were simply less efficient than the
evolving Agent Populations with the crossover control.
The results showed that the clustering catalyst, using hierarchical clustering [239] or Physical
Complexity clustering, failed to optimise the evolutionary processes (with clusters) of the
Digital Ecosystem. Additionally, typical runs of each showed no adverse behaviour, which
might have explained the failure to optimise the evolutionary processes. Therefore, the results
collectively confirm that the intra-cluster crossover assignment of the clustering catalyst was
less efficient than the random crossover assignment of the crossover control. The clustering
catalyst intuitively had potential, but most likely failed because the individuals within the
evolving Agent Populations lacked sufficient complexity (relative to biological populations [29])
for the mechanism to be effective, leading to the crossing of very similar individuals, producing
offspring that were very similar to their parents, and therefore not actually achieving valuable
change.
020406080100100200300MaximumFitnessGenerationAgentPopulationCrossoverControlClusteringCatalyst(Phy.Com.)ClusteringCatalyst(Hierarchical) 4.5. Simulation and Results ,
171
4.5.2 Targeted Migration
We implemented the targeted migration as defined in sections 4.2.4 and 4.4, using both Neural
Network (NN) and Support Vector Machine (SVM) based similarity recognition components
embedded within the Agents. We also made use of our Xgrid [159] modifications from Chapter
3 to take advantage of the grids mentioned in the acknowledgements.
4.5.2.1 Controls
The targeted migration was dependent on additional Agent migration, which alone could have
been responsible for any observed optimisation, because it led to greater distribution of the
Agents within the Digital Ecosystem, potentially improving responsiveness for the user base.
So, we included a migration control in our experimental simulations for the additional Agent
migration, being random instead of targeted. Furthermore, to determine the contribution of
the NNs and SVMs on the targeted migration we created a pattern recognition control, using a
rudimentary distance function adapted from the fitness function defined in section 2.3.3.
In Figure 4.13 we graphed for the simulation runs the average of the percentage response rate
after a thousand time steps (user request events), for the Digital Ecosystem with the migration
control, and the Digital Ecosystem with the pattern recognition control, compared to the Digital
Ecosystem alone. The Digital Ecosystem alone averaged a 68.0% (3 s.f.) response rate with
a standard deviation of 2.61 (2 d.p.), while the Digital Ecosystem with the migration control
showed a significant degradation to 49.6% (3 s.f.) with a standard deviation of 1.96 (2 d.p.), and
the Digital Ecosystem with the pattern recognition control showed only a small increase to 70.5%
(3 s.f.) with a standard deviation of 2.60 (2 d.p.). Therefore, any observed improvement from
the targeted migration was not from the additional migration but its targeting, and that the
effectiveness of the pattern recognition functionality will be significant if the targeted migration
is to be effective.
In Figure 4.14 we graphed a typical run of the Digital Ecosystem with the migration control, and
the Digital Ecosystem with the pattern recognition control, compared to the Digital Ecosystem
alone (taken from Figure 2.25). The Digital Ecosystem alone performed as expected, adapting
and improving over time to reach a mature state through the process of ecological succession [29].
The Digital Ecosystem with the migration control, which included additional random migration,
while initially beneficial, ultimately decreased the responsiveness of the Digital Ecosystem.
172
, Chapter 4. Optimisation of Digital Ecosystems
Figure 4.13: Graph of the Targeted Migration Controls and the Digital Ecosystem: The Digital
Ecosystem alone averaged a 68.0% (3 s.f.) response rate, while the Digital Ecosystem with the
migration control showed a significant degradation to 49.6% (3 s.f.), and the Digital Ecosystem
with the pattern recognition control showed only a small increase to 70.5% (3 s.f.).
Figure 4.14: Graph of Typical Runs for the Targeted Migration Controls and the Digital
Ecosystem: The Digital Ecosystem alone performed as expected, adapting and improving over
time to reach a mature state. The migration control with additional random migration ultimately
decreased the responsiveness, while the pattern recognition control performed only slightly better.
PatternRecognitionControlMigrationControlDigitalEcosystem020406080100AverageofResponseRate(%)After1000UserRequestEventsScenario70.5%49.6%68.0%0204060801002004006008001000ResponseRate(%matchtouserrequest)Time(userrequestevents)DigitalEcosystemMigrationControlPatternRecognitionControl 4.5. Simulation and Results ,
173
Finally, the Digital Ecosystem with the pattern recognition control performed only marginally
better than the Digital Ecosystem alone.
4.5.2.2 Neural Networks
We started with the NN-based targeted migration, as defined in section 4.4.2. We made use
of Joone (Java Object Oriented Neural Engine) [194] to implement the required NNs, and
exploratory programming [294] to determine that a hidden layer 1.5 times the size of the input
layer was effective for the NN-based similarity recognition components.
In Figure 4.15 we graphed for the simulation runs the average of the percentage response rate
after a thousand time steps (user request events), for the Digital Ecosystem with the NN-based
targeted migration, compared to the Digital Ecosystem alone. The Digital Ecosystem alone
averaged a 68.0% (3 s.f.) response rate with a standard deviation of 2.61 (2 d.p.), while the
Digital Ecosystem with the NN-based targeted migration showed a significant improvement to
a 92.1% (3 s.f.) response rate with a standard deviation of 2.22 (2 d.p.).
Figure 4.15: Graph of Neural Networks Based Targeted Migration: The Digital Ecosystem
alone averaged a 68.0% (3 s.f.) response rate with a standard deviation of 2.61 (2 d.p.), while
the Digital Ecosystem with the NN-based targeted migration showed a significant improvement
to a 92.1% (3 s.f.) response rate with a standard deviation of 2.22 (2 d.p.).
TargetedMigration(NN-based)DigitalEcosystem020406080100AverageofResponseRate(%)After1000UserRequestEventsScenario92.1%68.0% 174
, Chapter 4. Optimisation of Digital Ecosystems
4.5.2.3 Support Vector Machines
Next we considered the SVM-based targeted migration, as defined in section 4.4.3, making
use of LIBSVM (Library for Support Vector Machines) [50] to implement the required SVMs.
In Figure 4.16 we graphed for the simulation runs the average of the percentage response
rate after a thousand time steps (user request events), for the Digital Ecosystem with the
SVM-based targeted migration, compared to the Digital Ecosystem with the NN-based targeted
migration, and the Digital Ecosystem alone. The Digital Ecosystem with the SVM-based
targeted migration averaged a 92.8% (3 s.f.) response rate with a standard deviation of 2.09 (2
d.p.), slightly better than the NN-based targeted migration at 92.1% (3 s.f.) with a standard
deviation of 2.22 (2 d.p.), and so significantly better than the Digital Ecosystem alone at 68.0%
(3 s.f.) with a standard deviation of 2.61 (2 d.p.).
In Figure 4.17 we graphed typical runs of the Digital Ecosystem with the SVM-based targeted
migration, the Digital Ecosystem with the NN-based targeted migration, and the Digital
Ecosystem alone. The Digital Ecosystem alone performed as expected, adapting and improving
Figure 4.16: Graph of Support Vector Machine Based Targeted Migration: The Digital
Ecosystem with the SVM-based targeted migration averaged a 92.8% (3 s.f.) response rate,
slightly better than the NN-based targeted migration at 92.1% (3 s.f.), and so significantly better
than the Digital Ecosystem alone at 68.0% (3 s.f.).
TargetedMigration(SVM-based)TargetedMigration(NN-based)DigitalEcosystem020406080100AverageofResponseRate(%)After1000UserRequestEventsScenario92.8%92.1%68.0% 4.5. Simulation and Results ,
175
Figure 4.17: Graph of Typical Runs for the Digital Ecosystem and Targeted Migration: The
Digital Ecosystem alone performed as expected, adapting and improving over time to reach
a mature state through the process of ecological succession [29] In comparison, the Digital
Ecosystem with the targeted migration, NN or SVM-based, showed a significant improvement.
over time to reach a mature state through the process of ecological succession [29], approaching
70% effectiveness for the user base. The Digital Ecosystem with the targeted migration, NN or
SVM-based, showed a significant improvement in the ecological succession, reaching the same
performance in less than a fifth of the time, before reaching over 90% effectiveness for the
user base. To show more clearly the greater effectiveness of the SVM-based targeted migration,
compared to the NN-based targeted migration, we graphed in Figure 4.18 the frequency of poor
matches (<50%) every one hundred time steps, for the Digital Ecosystem with the SVM-based
targeted migration, compared to the Digital Ecosystem with the NN-based targeted migration,
and the Digital Ecosystem alone.
The results showed that the targeted migration optimised and accelerated the ecological
succession [29] of our Digital Ecosystem, constructively interacting with its ecological and
evolutionary dynamics. The results also showed that it was not the additional migration, but its
targeting that created the improvement in the Digital Ecosystem, and that an effective pattern
recognition technique was required for the targeted migration to operate effectively. Both NNs
and SVMs proved to be effective, SVMs marginally more than NNs. The results also showed
that there were no adverse side-effects from the Baldwin effect [19], the inheritance of learnt
0204060801002004006008001000ResponseRate(%matchtouserrequest)Time(userrequestevents)DigitalEcosystemTargetedMigration(NN-based)TargetedMigration(SVM-based) 176
, Chapter 4. Optimisation of Digital Ecosystems
Figure 4.18: Graph of Frequencies for the Targeted Migration: The frequency of poor matches
(<50%) every one hundred time steps, for the Digital Ecosystem with the SVM-based targeted
migration, compared to the Digital Ecosystem with the NN-based targeted migration, and the
Digital Ecosystem alone. It shows the greater effectiveness of the SVM-based targeted migration,
compared to the NN-based targeted migration from the seven hundredth generation onwards.
behaviour in the Agents from the embedded similarity recognition components, whether SVM
or NN based. Finally, based on the experimental results, and our theoretical understanding,
we would recommend SVMs for the pattern recognition functionality of the targeted migration.
4.6 Summary and Discussion
We started by reviewing the scope for optimisation and acceleration resulting from the
evolutionary self-organisation of ecological succession (the formation of a mature ecosystem)
being a slow process [29], even the accelerated form present in our Digital Ecosystem,
which was identified and confirmed in the previous chapters.
In the results of Chapter 2,
specifically the ecological succession experiment, the Digital Ecosystem reached only 70%
responsiveness,
identifying potential for improvement, which was confirmed by the results
of Chapter 3, specifically the self-organised diversity experiment for which the Digital
Ecosystem responded more slowly in some scenarios than others, confirming the potential
for improvement. Furthermore, the optimisation of Digital Ecosystems sought was not that
0102030405060701002003004005006007008009001000FrequencyofMatches<50%Time(userrequestevents)DigitalEcosystemTargetedMigration(NN-based)TargetedMigration(SVM-based) 4.6. Summary and Discussion ,
177
of parameter optimisation, which is achievable through exploratory programming [294], but
an augmentation to the Ecosystem-Oriented Architecture providing a significant improvement
in performance, i.e. better solutions for the users than the Digital Ecosystem alone could
achieve. We then discovered that we would be unlikely to find optimising or accelerating
augmentations for our Digital Ecosystem from biological ecosystems research, because ecological
optimisation is concerned with the maintenance of diversity and stability [323, 3, 224], and
ecological acceleration is similarly concerned with the re-establishment of diversity and stability
[201, 181, 336]. This was unsurprising, because one of the fundamental differences between
biological and digital ecosystems lie in the motivation and approach of their researchers; given
that biological ecosystems are ubiquitous natural phenomena whose maintenance is crucial to
our survival [20], whereas Digital Ecosystems are a technology engineered to serve specific
human purposes.
So, we therefore proposed, constructed, and explored several alternative augmentations to
accelerate or optimise the evolutionary and ecological self-organising dynamics of our Digital
Ecosystem, based on the understanding and results from the previous chapters, and our general
knowledge and intuition. The clustering catalyst aimed to optimise the evolutionary self-
organisation of evolving Agent Populations with clusters, by encouraging intra-cluster crossover
to directly accelerate the evolving of applications (Agent-sequences) in response to user requests,
and therefore the responsiveness of the Digital Ecosystem to the user base. The replacement
aggregator would have replaced the use of evolutionary computing [90], for the aggregation
of the Agents into optimal applications (Agent-sequences) in response to user requests, with
an alternative technique, directly optimising the responsiveness of the Digital Ecosystem to
the user base. The Agent-pool aggregation aimed to allow for the creation of potentially
useful applications (Agent-sequences) or partial applications inside the Agent-pools, optimising
the Agent-sequences found at the Agent-pools of the Habitats, which would in turn optimise
the evolving Agent Populations, and therefore the responsiveness of the Digital Ecosystem to
the user base. The targeted migration aimed to complement the evolving Agent Populations
indirectly, with additional highly targeted migration to support the existing Agent migration
between the Habitats, optimising the Agents found at the Agent-pools of the Habitats, which
would in turn optimise the evolving Agent Populations, and therefore the responsiveness of
the Digital Ecosystem to the user base. The replacement aggregator would have weakened the
Ecosystem-Oriented Architecture of Digital Ecosystems, and the Agent-Pool aggregation would
have incurred an impractical computational cost to operate, so we chose the most promising
178
, Chapter 4. Optimisation of Digital Ecosystems
augmentations, the clustering catalyst and the targeted migration, to be completed theoretically
and then investigated experimentally through simulations.
The clustering catalyst augmentation aimed to directly optimise the evolutionary self-
organisation of evolving Agent Populations with clusters, by encouraging intra-cluster crossover.
Crossover involves the crossing of two Agent-sequences in the creation of new Agent-sequences,
occurring during the replication stage of the evolutionary cycle [13]. Theoretical completion
of the clustering catalyst required an algorithm to perform the clustering of evolving Agent
Populations. So, we considered alternative clustering algorithms [142, 139, 239, 324, 180,
339, 280] for their suitability to our evolving Agent Populations. We chose a hierarchical
agglomerative average-link clustering algorithm [239] for our clustering catalyst, because a
hierarchical algorithm was more appropriate [142] than a partitional one for the small size
of the data sets of evolving Agent Populations, and because an agglomerative algorithm
was conceptually simpler than a divisive one [190], for which the efficiency advantage [190]
would have been negligible given the expected size of the data sets. Also, because the
average-link algorithm is designed to reduce the dependence of the cluster-linkage criterion
on extreme values, such as the most similar or dissimilar of the single-link and complete-
link algorithms [239]; and because the minimum-variance algorithm is biased to producing
clusters with the same number of objects [211], which would have been problematic for clusters
emerging over the generations. We also considered a clustering algorithm based on our extended
Physical Complexity from Chapter 3, because clustering is subjective by nature [142], and our
extended Physical Complexity was augmented in section 3.2.2.3 to understand the clustering of
evolving Agent Populations. So, we implemented the clustering catalyst using the hierarchical
agglomerative average-link clustering [239] and our Physical Complexity clustering. The results
showed that the clustering catalyst, using either clustering algorithm, failed to optimise the
evolutionary processes of the Digital Ecosystem. It intuitively had potential, but most likely
failed because the individuals within the evolving Agent Populations lacked sufficient complexity
(relative to biological populations [29]) for the mechanism to be effective, leading to the crossing
of very similar individuals, producing offspring that were very similar to their parents, and
therefore not actually achieving valuable change.
The targeted migration augmentation aimed to directly optimise the ecological migration,
and therefore indirectly complement the evolutionary self-organisation of the evolving Agent
Populations, through the highly targeted migration of the Agents to their niche Habitats.
The migration probabilities between the Habitats produces the passive Agent migration,
4.6. Summary and Discussion ,
179
allowing the Agents to spread in the correct general direction within the Habitat network,
based primarily upon success at their current location. The targeted migration works in
a more active manner, allowing the Agents highly targeted migration to specific Habitats,
based upon their interaction with one another to discover Habitats where they could be
valuable (i.e. find a niche). Theoretical completion of the targeted migration required further
consideration of how it would operate,
including its effect on the Agent life-cycle, and a
suitable pattern recognition [140] technique for the required similarity recognition between
the semantic descriptions of the Agents. So, we considered alternative pattern recognition
techniques [140, 262, 102, 233, 328, 246, 141, 123] for similarity recognition components to
be embedded within the Agents. Template Matching was not suitable because its effective
use is domain specific [140] and the similarity recognition between the semantic descriptions of
Agents is very different to the domains that it is typically applied [233]. Statistical Classification
was also not suitable, because the embedded similarity recognition component of each Agent
would have required human intervention for variable selection and transformation [206]. While
Structural Matching was a suitable technique theoretically, implementations have had many
difficulties [140], including the segmentation of noisy patterns (to detect primitives) and the
inference of grammar from training data [140]. There can also be a combinatorial explosion
of possibilities to be investigated, demanding large training sets and significant computational
effort [248], neither of which was available. Neural Networks (NNs) were suitable, given their low
dependence on domain-specific knowledge and the availability of efficient learning algorithms
[140]. Support Vector Machines (SVMs), albeit a recent development [316], were also suitable
[145], being primarily a binary classifier [140] for training generalisable nonlinear classifiers
in high-dimensional spaces using small training sets [313]. So, we implemented the targeted
migration using both NN and SVM based similarity recognition components embedded within
the Agents. The results showed that the targeted migration accelerated and optimised the
ecological succession of the Digital Ecosystem, constructively interacting with its ecological
and evolutionary dynamics, marginally more when powered by SVMs than NNs. So, based on
the experimental results, and our theoretical understanding, we would recommend SVMs for
the pattern recognition functionality of the targeted migration.
The targeted migration also resulted in the Baldwin effect [19], the inheritance of learnt
behaviour in the Agents,
from the embedded similarity recognition components, whether
SVM or NN based. While the Baldwin effect has always been controversial within biological
ecosystems [329], primarily because of the problem of confirming it experimentally [298],
180
, Chapter 4. Optimisation of Digital Ecosystems
it undoubtedly occurred in our Digital Ecosystem. However, the experimental results and
our exploratory programming [294] showed no adverse side-effects, therefore supporting the
possibility of its presence in biological ecosystems.
In this chapter we attempted the acceleration and optimisation of Digital Ecosystems, because
ecological succession (the formation of mature ecosystems) is a slow process [29], even the
accelerated form present in our Digital Ecosystems. While not all our attempts were
successful, understandable considering the constructive nature of our efforts in this chapter,
we have optimised and accelerated our Digital Ecosystem through the targeted migration of its
Agents. The targeted migration significantly enhanced the ecological succession, constructively
interacting with the ecological and evolutionary dynamics, helping the Agents to optimise their
migration and distribution within our Digital Ecosystem.
Chapter 5
Conclusions
5.1 Achievements
Substantial parts of our efforts are original contributions in the area of Biologically-Inspired
Computing [99] and the emerging field of Digital Ecosystems, with our major research
contributions being as follows:
• We have created the first interpretation of Digital Ecosystems where the word ecosystem
is more than just a metaphor, which we have confirmed experimentally. They are the
digital counterparts of biological ecosystems: having their properties of self-organisation,
scalability and sustainability [173]; created through combining understanding from the-
oretical ecology [175], evolutionary theory [104], Multi-Agent Systems [249], distributed
evolutionary computing [178], and Service-Oriented Architectures [228]. Furthermore,
the Ecosystem-Oriented Architecture of Digital Ecosystems includes a novel form of
distributed evolutionary computing, an optimisation technique working at two levels:
a first optimisation, migration of Agents which are distributed in a peer-to-peer network,
operating continuously in time; this process feeds a second optimisation, based on
evolutionary computing, operating locally on single peers and is aimed at finding solutions
that satisfy locally relevant constraints. So, the local search is improved through this
twofold process to yield better local optima faster, as the distributed optimisation provides
prior sampling of the search space through computations already performed in other peers
with similar constraints. We have also defined the interaction of Digital Ecosystems with
business ecosystems [214], specifically in supporting and enabling them to create Digital
Business Ecosystems.
181
182
, Chapter 5. Conclusions
• We have investigated the emergent self-organising properties of Digital Ecosystems,
because a primary motivation for our research is the desire to exploit the self-
organising properties [173] of biological ecosystems. We started with the evolutionary
self-organisation of ecological succession [29], which conformed to expectations [136].
Next we considered the self-organisation of the order constructing processes (the
evolving Agent Populations). We extended Physical Complexity [8] to include evolving
Agent Populations, which required extending definitions for populations of variable
length sequences, creating a measure for the efficiency of information storage, and an
understanding of clustering within Populations to support the non-atomicity of Agents.
We then extended Chli-DeWilde stability [53] to include the evolutionary dynamics of
evolving Agent Populations, building upon this to construct an entropy-based definition
for the degree of instability, which was used to study the stability of evolving Agent
Populations. Finally, the unique hybrid nature of Digital Ecosystems resulted in us
creating our own definition for the self-organised diversity, based on the global distribution
of the Agents in the Populations relative to the request behaviour of the user base.
Overall an insight has been achieved into where and how self-organisation occurs in Digital
Ecosystems, including what forms it can take and how it can be quantified.
• We have optimised and accelerated Digital Ecosystems, because the evolutionary self-
organisation of ecological succession [29] (the formation of a mature ecosystem) is a slow
process, even the accelerated form present in Digital Ecosystems. So, we considered
alternative augmentations, including the accelerating effect of a clustering catalyst on
the evolutionary dynamics, through the acceleration of the evolutionary processes; and
the optimising effect of targeted migration on the ecological dynamics, through the
emergent optimisation of the Agent migration patterns. The experimental results
showed that the clustering catalyst failed, despite intuitively having potential, most
likely because the individuals within the evolving Agent Populations lacked sufficient
complexity (relative to biological populations [29]) for the augmentation to be effective.
However, the experimental results also showed that the targeted migration optimised
and accelerated the ecological succession of Digital Ecosystems, constructively interacting
with the ecological and evolutionary dynamics. We also discovered that there were no
adverse side-effects from the Baldwin effect [19], the inheritance of learnt behaviour,
in Digital Ecosystems. Therefore, supporting the possibility of the Baldwin effect in
biological ecosystems, which has always been controversial [329].
5.2. Future Directions ,
183
5.2 Future Directions
Our efforts offer considerable scope for the future, with there being several interesting avenues
to pursue, some of which we discuss below.
5.2.1 Ecosystems Conceptualisation
Conceptualising ecosystems has been an inherent part of this work, which presents us with
an opportunity to formalise our current and future efforts to improve the cross-disciplinary
knowledge transfer required.
5.2.1.1 Biology of Digital Ecosystems
In creating Digital Ecosystems, the digital counterpart of biological ecosystems, we naturally
asked their likeness to the biological ecosystems from which they came. Further to this, we could
consider the applicability of other aspects of ecosystems theory in understanding and analysing
the dynamics of Digital Ecosystems. For example, energy pyramids 1 of biological ecosystems,
what is their equivalent in Digital Ecosystems? Given that Digital Ecosystems are information-
centric, whereas biological ecosystems are energy-centric [29], they would undoubtedly be
information pyramids, but further definition would naturally require more research.
5.2.1.2 Biological Design Patterns
A design pattern is a general reusable solution to a commonly occurring problem in software
design [106].
It is not a finished design that can be transformed directly into code, but
a description or template for how to solve a problem that can be used in many different
situations [106]. For example, object-oriented design patterns typically show relationships
and interactions between classes or objects, without specifying the final application classes or
objects that are involved [106]. Biological Design Patterns (BDPs) would extend this concept to
catalogue common interactions between biological structures using a pattern-oriented modelling
approach [122], which when applied would endow software systems with the desirable properties
of biological systems, such as self-organisation, self-management, scalability and sustainability.
1 Energy pyramids show the dissipation of energy at trophic levels, positions that organisms occupy in a food
chain, e.g. producers or consumers [237].
184
, Chapter 5. Conclusions
5.2.1.3 Classes of Ecosystems
While evolutionary theory [104] was well understood within computer science, under the
auspices of evolutionary computing [90], ecosystems theory [29], until our efforts, was not.
Similarly, while evolutionary theory is well understood within linguistics [63] and economics
[227], equally ecosystems theory is not [218]. So, using our efforts as a case study, we could
follow the same process to create Language Ecosystems and Economic Ecosystems. For
example, there are many separate efforts within linguistics using evolution to model language
change [54], but there is no unifying framework, which has resulted from different linguists
independently adopting elements of evolutionary theory [54]. So, we could provide a wide-
ranging and encompassing definition of Language Ecosystems, which would unify the many
disparate efforts in linguistics aimed at understanding language evolution.
5.2.1.4 Generic Ecosystem Definition
In the creation of Digital Ecosystems we considered aspects of biological ecosystems, including
Agent-Based Modelling [116] and Complex Adaptive Systems (CAS) [173], and then constructed
their counterparts in Digital Ecosystems. After which we considered the possibility of a
Generic Ecosystem definition, as we suggested at the end of Chapter 2, without which some
of the counterparts we constructed appeared to be compromised, when they were actually
the realisation of generic abstract concepts in Digital Ecosystems. Most notably the network
structure, which is energy-centric in biological ecosystems [29], while information-centric in
Digital Ecosystems. So, there is potential to create a Generic Ecosystem definition, using a
suitable modelling technique such as CAS [322], which would abstractly define the key properties
of an ecosystem, and would theoretically be applicable to any domain where the modelling
technique has been applied. Therefore, the Generic Ecosystem definition would provide a
framework for the application of ideas, concepts, and models from biological ecosystems to
other classes of ecosystems, including Digital Ecosystems, Language Ecosystems and Economic
Ecosystems.
5.2.2 Simulation Framework
An open-source simulation framework for Digital Ecosystems [161] was created by the Digital
Business Ecosystem (DBE) project [92], and is currently supported by the Open Philosophies for
5.2. Future Directions ,
185
Associative Autopoietic Digital Ecosystems (OPAALS) project [258] to assist further research
into Digital Ecosystems, including the wider implications of interacting with social systems,
such as business ecosystems of Small and Medium sized Enterprises (SMEs).
5.2.3 Digital Business Ecosystems
In an old market-based economy, made up of sellers and buyers, the parties exchange property
[78]. While in a new network-based economy, made up of servers and clients in a business
ecosystem [214], the parties share access to services and experiences [78]. Digital Ecosystems
are a platform for the network-based economy of business ecosystems, providing mechanisms
for the creation of Digital Business Ecosystems.
5.2.3.1 Service Futures Market
One such mechanism the Digital Ecosystem could provide to the network-based economy of
business ecosystems [214], would be a futures market 2 for services. As each service (Agent)
consists of an executable component and a semantic description, the later acting as a guarantee
of behaviour, and the evolving Agent Populations only requiring the guarantees (semantic
descriptions) to operate, the actual executable component of a service (Agent) is only required
once an application (Agent-sequence) has been assembled. Therefore, service (Agent) evolution
could operate entirely on the semantic descriptions, with business users only needing to supply
the executable component of a service (Agent) once there is a demand, i.e. when the semantic
description of one of their services has been used in the construction of an application which
meets the request of another business user. Therefore, creating a futures market for evolving
services within Digital Business Ecosystems.
5.2.3.2 Regional Deployment
A partial reference implementation [199]
for our Digital Ecosystem, which includes an
implementation of the targeted migration, was created by the Digital Business Ecosystems
project [92], and we expect that once completed will be deployed as part of the software
platform intended for the regional deployment of their Digital Ecosystems [264, 258]. Digital
2 An auction market in which participants buy and sell commodities for an agreed price, that the sellers have
yet to produce [137].
186
, Chapter 5. Conclusions
Ecosystems (distributed adaptive open socio-technical systems, with properties of self-
organisation, scalability and sustainability, inspired by natural ecosystems [258]) are emerging
as a novel approach to the catalysis of sustainable regional development driven by Small
and Medium sized Enterprises (SMEs) [258]. The community focused on the deployment of
Digital Ecosystems, REgions for Digital Ecosystems Network (REDEN) [265], is supported
by projects such as the Digital Ecosystems Network of regions for (4) DissEmination and
Knowledge Deployment (DEN4DEK) [264], a thematic network that aims to share experiences
and disseminate all the necessary knowledge that will allow regions to plan an effective
deployment of Digital Ecosystems at all levels (economic, social, technical and political) to
produce real impacts in the economic activities of European regions through the improvement
of SME business environments. So, the next major step in our research will be to collect real
world data, confirming that Digital Ecosystems operate effectively with business ecosystems in
creating Digital Business Ecosystems.
5.3 Concluding Remarks
The ever-increasing challenge of software complexity in creating progressively more sophisti-
cated and distributed applications, makes the design and maintenance of efficient and flexible
systems a growing challenge [209, 299, 193], for which current software development techniques
have hit a complexity wall [184]. In response we have created Digital Ecosystems, the digital
counterparts of biological ecosystems, possessing their properties of self-organisation, scalability
and sustainability [173]; Ecosystem-Oriented Architectures that overcome the challenge by
automating the search for new algorithms in a scalable architecture, through the evolution of
software services in a distributed network.
Bibliography
[1] S. Abe. Support Vector Machines for Pattern Classification. Springer, 2005.
[2] V. Abramov, N. Szirbik, J. Goossenaerts, T. Marwala, P. De Wilde, L. Correia, P.
Mariano, and R. Ribeiro. Ontological basis for open distributed multi-agent system.
In Symposium on Adaptive Agents and Multi-Agent Systems, pages 33–43. Society for the
Study of Artificial Intelligence and Simulation of Behaviour, 2001.
[3] P. Abrams. Foraging time optimization and interactions in food webs. The American
Naturalist, 124:80–96, 1984.
[4] C. Adami. Self-organized criticality in living systems. Physical Letters A, 203:29–32,
1995.
[5] C. Adami. Introduction To Artificial Life. Springer, 1998.
[6] C. Adami. Sequence complexity in darwinian evolution. Complexity, 8:49–56, 2003.
[7] C. Adami and N. Cerf. Physical complexity of symbolic sequences. Physica D, 137:62–69,
2000.
[8] C. Adami, C. Ofria, and T. Collier. Evolution of biological complexity. Proceedings of
the National Academy of Sciences, 97:4463–4468, 2000.
[9] A. Agiorgitis. Complex adaptive systems [online]. Available from:
http://www.siliconyogi.com/andreas/it professional/sol/complexsystems/.
[10] J. Anderson. Introduction to Flight. McGraw-Hill Professional, 2004.
[11] W. Ashby. Principles of the self-organizing dynamic system.
Journal of General
Psychology, 37:125–128, 1947.
[12] W. Ashby. Principles of the self-organizing system.
In H. Von Foerster and G. Zopf,
editors, Principles of Self-Organization, pages 255–278. Pergamon Press, 1962.
[13] T. Back.
Evolutionary Algorithms in Theory and Practice: Evolution Strategies,
Evolutionary Programming, Genetic Algorithms. Oxford University Press, 1996.
[14] T. Baeck, D. Fogel, and Z. Michalewicz, editors. Handbook of Evolutionary Computation.
CRC Press, 1997.
[15] R. Baeza-Yates. Introduction to data structures and algorithms related to information
retrieval. In W. Frakes and R. Baeza-Yates, editors, Information Retrieval Data Structures
& Algorithms, pages 13–27. Prentice Hall, 1992.
187
188
BIBLIOGRAPHY
[16] R. Bajcsy and S. Kovacic. Multiresolution elastic matching. Computer Vision, Graphics,
and Image Processing, 46:1–21, 1989.
[17] P. Bak, C. Tang, and K. Wiesenfeld. Self-organized criticality. Physical Review A, 38:364–
374, 1988.
[18] H. Balakrishnan, M. Stemm, S. Seshan, and R. Katz. Analyzing stability in wide-
In S. Leutenegger, editor, International Conference on
area network performance.
Measurement and Modeling of Computer Systems, pages 2–12. ACM Press, 1997.
[19] J. Baldwin. A new factor in evolution. American Naturalist, 30:441–451, 1896.
[20] A. Balmford, A. Bruner, P. Cooper, R. Costanza, S. Farber, R. Green, M. Jenkins,
P. Jefferiss, V. Jessamy, J. Madden, K. Munro, N. Myers, S. Naeem, J. Paavola, M.
Rayment, S. Rosendo, J. Roughgarden, K. Trumper, and R. Turner. Economic reasons
for conserving wild nature. Science, 297:950–953, 2002.
[21] W. Banzhaf. Genotype-phenotype-mapping and neutral variation–a case study in genetic
programming. Parallel Problem Solving from Nature III, 866:322–332, 1994.
[22] W. Banzhaf. Genetic Programming. Morgan Kaufmann Publishers, 1998.
[23] W. Banzhaf and W. Langdon. Some considerations on the reason for bloat. Genetic
Programming and Evolvable Machines, 3:81–91, 2002.
[24] A. Barron, J. Rissanen, and B. Yu. The minimum description length principle in coding
and modeling. IEEE Transactions on Information Theory, 44:2743–2760, 1998.
[25] W. Barthlott and C. Neinhuis. Purity of the sacred lotus, or escape from contamination
in biological surfaces. Planta, 202:1–8, 1997.
[26] G. Basharin. On a statistical estimate for the entropy of a sequence of independent
random variables. Theory Probability and its Applications, 4:333–336, 1959.
[27] BBC News. Microsoft begins windows 7 push [online]. 2009. Available from:
http://news.bbc.co.uk/1/hi/technology/7817190.stm.
[28] S. Beer. Decision and Control: The Meaning of Operational Research and Management
Cybernetics. Wiley, 1966.
[29] M. Begon, J. Harper, and C. Townsend.
Ecology:
Individuals, Populations and
Communities. Blackwell Publishing, 1996.
[30] G. Bell. The distribution of abundance in neutral communities. American Naturalist,
396:606–617, 2000.
[31] B. Benatallah, M. Dumas, and Q. Sheng. Facilitating the rapid development and scalable
orchestration of composite web services. Distributed and Parallel Databases, 17:5–37,
2005.
[32] D. Bennett. Digital transformation in the entertainment industry - embracing the fully
digital ecosystem. Technical report, Accenture, 2006. Available from:
http://www.accenture.com/NR/rdonlyres/A58111E4-22E5-4DDD-B3DE-FB3741F0052
F/0/EmbracingDigitalEco.pdf.
[33] J. Benyus. Biomimicry, Innovation Inspired by Nature. Harper Collins Publishers, 2002.
BIBLIOGRAPHY
189
[34] M. Berry and G. Linoff. Data Mining Techniques: For Marketing, Sales, and Customer
Support. Wiley, 1997.
[35] C. Bishop. Neural Networks for Pattern Recognition. Oxford University Press, 1995.
[36] T. Blickle and L. Thiele. A comparison of selection schemes used in evolutionary
algorithms. Evolutionary Computation, 4:361–394, 1996.
[37] A. Blum.
Neural networks in C++:
an object-oriented framework for building
connectionist systems. Wiley, 1992.
[38] Z. Boger and H. Guterman. Knowledge extraction from artificial neural network
models. In J. Tien, editor, International Conference on Systems, Man, and Cybernetics:
Computational Cybernetics and Simulation, volume 4. IEEE Press, 1997.
[39] N. Botros and M. Abdul-Aziz. Hardware implementation of an artificial neural network
IEEE Transactions on Industrial Electronics,
using field programmable gate arrays.
41:665–667, 1994.
[40] S. Bramly. Leonardo: The Artist and the Man. Penguin Books, 1994.
[41] C. Burges. A tutorial on support vector machines for pattern recognition. Data Mining
and Knowledge Discovery, 2:121–167, 1998.
[42] L. Cabral, J. Domingue, E. Motta, T. Payne, and F. Hakimpour. Approaches to semantic
In The Semantic Web: Research and
web services: an overview and comparisons.
Applications, pages 225–239. Springer, 2004.
[43] D. Cacuci, M. Ionescu-Bujor, and I. Navon. Sensitivity and Uncertainty Analysis. CRC
Press, 2003.
[44] E. Canal. Using the ESB with e4: Applying new tendencies to CSC business integration
architecture. Technical report, Computer Sciences Corporation, 2007. Available
from: http://www.csc.com/aboutus/leadingedgeforum/knowledgelibrary/uploads/Using
%20the%20ESB%20with%20e4%20-%20ERCanal.pdf.
[45] E. Cantu-Paz. A survey of parallel genetic algorithms. R´eseaux et syst`emes r´epartis,
Calculateurs Parall`eles, 10:141–171, 1998.
[46] J. Cardoso and A. Sheth.
Introduction to semantic web services and web process
In J. Cardoso and A. Sheth, editors, Semantic Web Services and Web
composition.
Process Composition, pages 1–13. Springer, 2004.
[47] M. Castells. The Rise of The Network Society. Blackwell Publishing, 2000.
[48] G. Chaitin. Algorithmic information and evolution. In O. Solbrig and G. Nicolis, editors,
Perspectives on Biological Complexity, pages 51–60. IUBS Press, 1988.
[49] L. Chambers. The practical handbook of genetic algorithms: applications. CRC Press,
2001.
[50] C. Chang and C. Lin. LIBSVM: A library for support vector machines. Technical report,
National Taiwan University, 2008. Available from:
http://www.csie.ntu.edu.tw/ cjlin/papers/libsvm.pdf.
190
BIBLIOGRAPHY
[51] A. Chhatpar.
Increase business agility through BRM systems and SOA: Best bets
to increase ROI on your enterprise applications. Technical report, IBM Devleoper
Works, 2008. Available from: http://download.boulder.ibm.com/ibmdl/pub/software/
dw/architecture/ar-brmssoa/ar-brmssoa-pdf.pdf.
[52] M. Chli. Convergence and Interactivity of Multi-Agent Systems. PhD thesis, Imperial
College London, 2006.
[53] M. Chli, P. De Wilde, J. Goossenaerts, V. Abramov, N. Szirbik, L. Correia, P. Mariano,
and R. Ribeiro. Stability of multi-agent systems. In E. Santos Jr and P. Willett, editors,
International Conference on Systems, Man, and Cybernetics, pages 551–556. IEEE Press,
2003.
[54] M. Christiansen and S. Kirby. Language Evolution. Oxford University Press, 2003.
[55] D. Cliff and S. Grand. The creatures global digital ecosystem. Artificial Life, 5:77–93,
1999.
[56] Codefarm Software Limited. Codefarm - technology for structured credit [online]. 2008.
Available from: http://www.codefarm.com/ [cited 05 Feb 2008].
[57] N. Collier. Repast: An extensible framework for agent simulation. Technical report,
University of Chicago, 2003. Available from:
http://www.econ.iastate.edu/tesfatsi/RePastTutorial.Collier.pdf.
[58] D. Comer. Internetworking with TCP/IP: principles, protocols, and architecture. Prentice
Hall, 2005.
[59] J. Connell and R. Slatyer. Mechanisms of succession in natural communities and their
role in community stability and organization. The American Naturalist, 111:1119–1144,
1977.
[60] A. Corallo, G. Passiante, and A. Prencipe. The Digital Business Ecosystem. Edward
Elgar Publishing, 2007.
[61] D. Cox and H. Miller. The Theory of Stochastic Processes. CRC Press, 1977.
[62] N. Cristianini and J. Shawe-Taylor. An Introduction to Support Vector Machines and
Other Kernel-based Learning Methods. Cambridge University Press, March 2000.
[63] W. Croft. Explaining language change. Longman, 2000.
[64] J. Crutchfield and O. Gornerup. Objects that make objects: the population dynamics of
structural complexity. Journal of The Royal Society Interface, 3(7):345–349, 2006.
[65] J. Crutchfield and K. Young. Inferring statistical complexity. Physical Review Letters,
63:105–108, 1989.
[66] F. Curbera, M. Duftler, R. Khalaf, W. Nagy, N. Mukhi, and S. Weerawarana. Unraveling
IEEE Internet
the web services web: An introduction to SOAP, WSDL, and UDDI.
Computing, 6:86–93, 2002.
[67] C. Darwin. On the Origin of Species by Means of Natural Selection. John Murray, 1859.
[68] T. Davenport and J. Short. The new industrial engineering: Information technology and
business process redesign. Sloan Management Review, 31:11–27, 1990.
BIBLIOGRAPHY
191
[69] O. Davis. Secondary succession. Technical report, University of Arizona, 2008. Available
from: http://www.geo.arizona.edu/Antevs/nats104/00lect20.html.
[70] R. Dawkins. The Selfish Gene. Oxford University Press, 2006.
[71] W. Day and H. Edelsbrunner. Efficient algorithms for agglomerative hierarchical
clustering methods. Journal of Classification, 1:7–24, 1984.
[72] J. de Bruijn, H. Lausen, A. Polleres, and D. Fensel. The web service modeling language
WSML: An overview. The Semantic Web: Research and Applications, 4011:590–604,
2006.
[73] L. De Castro. Fundamentals of Natural Computing: Basic Concepts, Algorithms, And
Applications. CRC Press, 2006.
[74] K. De Jong, D. Fogel, and H. Schwefel. A history of evolutionary computation. In Baeck
et al. [14], pages 1–12.
[75] F. de Toro, J. Ortega, J. Fernandez, and A. Diaz. PSFGA: a parallel genetic algorithm
In F. Vajda and N. Podhorszki, editors, Euromicro
for multiobjective optimization.
Workshop on Parallel, Distributed and Network-based Processing, pages 384–391. IEEE
Press, 2002.
[76] J. de Visser and S. Elena. The evolution of sex: empirical insights into the roles of
epistasis and drift. Nature Reviews Genetics, 8:139–49, 2007.
[77] P. De Wilde, H. Nwana, and L. Lee. Stability, fairness and scalability of multi-agent
International Journal of Knowledge-Based Intelligent Engineering Systems,
systems.
3:84–91, 1999.
[78] P. Delcloque and A. Bramoull´e. DISSEMINATE, an initial implementation proposal: a
new point of departue in call for the ‘year 01’ ? ReCALL, 13:277–292, 2001.
[79] P. Denning and R. Metcalfe. Beyond Calculation: The Next Fifty Years of Computing.
Springer, 1997.
[80] L. Devroye, L. Gyorfi, and G. Lugosi. A Probabilistic Theory of Pattern Recognition.
Springer, 1996.
[81] G. Di Marzo Serugendo, M. Gleizes, and A. Karageorgos. Self-organization in multi-agent
systems. The Knowledge Engineering Review, 20:165–189, 2006.
[82] P. Dini, G. Lombardo, R. Mansell, A. Razavi, S. Moschoyiannis, P. Krause, A. Nicolai,
and L. Rivera Le´on. Beyond interoperability to digital ecosystems: regional innovation
and socio-economic development led by SMEs. International Journal of Technological
Learning, Innovation and Development, 1:410–426, 2008.
[83] P. Dini and A. Nicolai. A scientific foundation for digital ecosystems. In Nachira et al.
[223], pages 1–20.
[84] R. Dubes. How many clusters are best? - an experiment. Pattern Recognition, 20:645–663,
1987.
[85] E. Ducheyne, B. De Baets, and R. De Wulf. Is fitness inheritance useful for real-world
In C. Fonseca, editor, Evolutionary Multi-criterion Optimization, pages
applications.
31–42. Springer, 2003.
192
BIBLIOGRAPHY
[86] R. Duda, P. Hart, and D. Stork. Pattern classification and scene analysis. Wiley, 1973.
[87] S. Dustdar and W. Schreiner. A survey on web services composition.
International
Journal of Web and Grid Services, 1:1–30, 2005.
[88] W. Ebeling. Applications of evolutionary strategies.
Systems Analysis Modelling
Simulation, 7:3–16, 1990.
[89] A. Eiben, E. Aarts, and K. Van Hee. Global convergence of genetic algorithms: A
Markov chain analysis. In H. Schwefel and R. Manner, editors, Parallel Problem Solving
from Nature, pages 4–12. Springer, 1991.
[90] A. Eiben and J. Smith. Introduction to Evolutionary Computing. Springer, 2003.
[91] R. Ellen. Environment, Subsistence, and System: The Ecology of Small-Scale Social
Formations. Cambridge University Press, 1982.
[92] A. English. Digital Business Ecosystems (DBE) [online]. Available from:
http://www.digital-ecosystem.org/DBE Main/about [cited 26 Jan 2008].
[93] A. English and B. Dory. The DBE and micro economics - the benefits of the DBE for
SMEs [online]. 2007. Available from:
http://www.digital-ecosystem.org/downloadfiles/intro mov dbe.swf.
[94] C. Fiorina. The digital ecosystem [online]. 2000. Available from:
http://www.hp.com/hpinfo/execteam/speeches/fiorina/ceo worldres 00.html.
[95] H. Flyvbjerg, K. Sneppen, and P. Bak. Mean field theory for a simple model of evolution.
Physical Review Letters, 71:4087–4090, 1993.
[96] L. Fogel, A. Owens, and M. Walsh. Artificial Intelligence Through Simulated Evolution.
Wiley, 1966.
[97] C. Folke, S. Carpenter, B. Walker, M. Scheffer, T. Elmqvist, L. Gunderson, and C.
Holling. Regime shifts, resilience, and biodiversity in ecosystem management. Annual
Review of Ecology, Evolution, and Systematics, 35:557–581, 2004.
[98] C. Fonseca and P. Fleming. An overview of evolutionary algorithms in multiobjective
optimization. Evolutionary Computation, 3:1–16, 1995.
[99] N. Forbes. Imitation of Life: How Biology Is Inspiring Computing. MIT Press, 2004.
[100] R. Fox. The social identity of management ergolect. English for Specific Purposes, 18:261–
278, 1999.
[101] J. Friedl. Mastering Regular Expressions. O’Reilly, 2006.
[102] K. Fu. Syntactic pattern recognition and applications. Prentice Hall, 1982.
[103] D. Fudenberg and J. Tirole. Game Theory. MIT Press, 1991.
[104] D. Futuyma. Evolutionary Biology. Sinauer Associates, 1998.
[105] J. Galindo. Handbook of Research on Fuzzy Information Processing in Databases. IGI
Global, 2008.
BIBLIOGRAPHY
193
[106] E. Gamma, R. Helm, R. Johnson, and J. Vlissides. Design patterns: elements of reusable
object-oriented software. Addison-Wesley, 1995.
[107] C. Gershenson and F. Heylighen. When can we call a system self-organizing?
In J.
Carbonell and J. Siekmann, editors, European Conference on Advances in Artificial Life,
pages 606–614. Springer, 2003.
[108] R. Gillam. Unicode Demystified: A Practical Programmer’s Guide to the Encoding
Standard. Addison-Wesley, 2002.
[109] Glocal Forum and CERFE. The glocalization manifesto. Technical report, The
Glocal Forum, 2004. Available from: http://www.glocalforum.org/mediagallery/media
Download.php?mm=/warehouse/documents/the glocalization manifesto.pdf.
[110] D. Goldberg. Genetic algorithms in search, optimization, and machine learning. Addison-
Wesley, 1989.
[111] D. Goldberg and P. Segrest. Finite Markov chain analysis of genetic algorithms.
In J. Grefenstette, editor, International Conference on Genetic Algorithms and their
application, pages 1–8. Lawrence Erlbaum Associates, 1987.
[112] N. Gotelli. A primer of ecology. Sinauer Associates, 1995.
[113] J. Grabmeier and A. Rudolph. Techniques of cluster algorithms in data mining. Data
Mining and Knowledge Discovery, 6:303–360, 2002.
[114] S. Grand and D. Cliff. Creatures: Entertainment software agents with artificial life.
Autonomous Agents and Multi-Agent Systems, 1:39–57, 1998.
[115] D. Green and M. Kirley. Adaptation, diversity and spatial patterns. International Journal
of Knowledge-Based Intelligent Engineering Systems, 4:184–190, 2000.
[116] D. Green, N. Klomp, G. Rimmington, and S. Sadedin. Complexity in Landscape Ecology.
Springer, 2006.
[117] D. Green, T. Leishman, and S. Sadedin. Dual phase evolution: a mechanism for self-
In A. Minai, D. Braha, and Y. Bar-Yam, editors,
organization in complex systems.
International Conference on Complex Systems. Springer, 2006.
[118] D. Green, D. Newth, and M. Kirley. Connectivity and catastrophe - towards a general
In M. Bedau, editor, International Conference on Artificial Life,
theory of evolution.
pages 153–161. MIT Press, 2000.
[119] D. Green and S. Sadedin. Interactions matter- complexity in landscapes and ecosystems.
Ecological Complexity, 2:117–130, 2005.
[120] T. Gregory. Coincidence, coevolution, or causation? DNA content, cellsize, and the
C-value enigma. Biological Reviews, 76:65–101, 2001.
[121] U. Grenander and M. Miller. Pattern Theory: From Representation to Inference. Oxford
University Press, 2007.
[122] V. Grimm, E. Revilla, U. Berger, F. Jeltsch, W. Mooij, S. Railsback, H. Thulke, J. Weiner,
T. Wiegand, and D. DeAngelis. Pattern-oriented modeling of agent-based complex
systems: Lessons from ecology. Science, 310:987–991, 2005.
194
BIBLIOGRAPHY
[123] S. Gunn. Support vector machines for classification and regression. Technical report,
University of Southampton, 1998.
[124] H. Gutowitz, J. Victor, and B. Knight. Local structure theory for cellular automata.
Physica D, 28:18–48, 1987.
[125] H. Haken. Synergetics: an introduction: nonequilibrium phase transitions and self-
organization in physics, chemistry, and biology. Springer, 1977.
[126] M. Hansen and B. Yu. Model selection and the principle of minimum description length.
Journal of the American Statistical Association, 96:746–774, 2001.
[127] I. Hanski. Metapopulation Ecology. Oxford University Press, 1999.
[128] I. Hanski and O. Ovaskainen. Metapopulation theory for fragmented landscapes.
Theoretical Population Biology, 64:119–127, 2003.
[129] R. Haralick. Statistical and structural approaches to texture. Proceedings of the IEEE,
67:786–804, 1979.
[130] C. Hastrich. Biomimicry: A tool for innovation. Technical report, The Biomimicry
Institute, 2007. Available from:
http://www.biomimicryinstitute.org/about-us/biomimicry-a-tool-for-innovation.html.
[131] S. Haykin. Neural Networks: A Comprehensive Foundation. Prentice Hall, 1998.
[132] D. Hebb. The Organization of Behavior. Wiley, 1949.
[133] F. Heylighen. The science of self-organization and adaptivity. The Encyclopedia of Life
Support Systems, pages 253–280, 2002.
[134] F. Heylighen and C. Joslyn. Cybernetics and second order cybernetics. Encyclopedia of
Physical Science & Technology, 4:155–170, 2001.
[135] M. Hollis. The Philosophy of Social Science: An Introduction. Cambridge University
Press, 1994.
[136] S. Hubbell. The Unified Neutral Theory of Biodiversity and Biogeography. Princeton
University Press, 2001.
[137] J. Hull. Fundamentals of Futures and Options Markets. Prentice Hall, 2005.
[138] M. Iansiti and R. Levien. The Keystone Advantage: What the New Dynamics of Business
Ecosystems Mean for Strategy, Innovation, and Sustainability. Harvard Business School
Press, 2004.
[139] A. Jain and R. Dubes. Algorithms for clustering data. Prentice Hall, 1988.
[140] A. Jain, R. Duin, and J. Mao.
Statistical pattern recognition: A review.
IEEE
Transactions On Pattern Analysis and Machine Intelligence, 22:4–37, 2000.
[141] A. Jain, J. Mao, and K. Mohiuddin. Artificial neural networks: A tutorial.
IEEE
Computer, pages 31–44, 1996.
[142] A. Jain, M. Murty, and P. Flynn. Data clustering: A review. ACM Computing Surveys,
31:264–323, 1999.
BIBLIOGRAPHY
195
[143] E. Jaynes. Probability theory: the logic of science. Cambridge University Press, 2003.
[144] Y. Jin. A comprehensive survey of fitness approximation in evolutionary computation.
Soft Computing - A Fusion of Foundations, Methodologies and Applications, 9:3–12, 2005.
[145] T. Joachims. Text categorization with support vector machines: Learning with many
In J. Carbonell and J. Siekmann, editors, European Conference on
relevant features.
Machine Learning. Springer, 1997.
[146] D. Johansen, R. van Renesse, and F. Schneider. Operating system support for mobile
In Workshop on Hot Topics in Operating Systems, pages 42–45. IEEE Press,
agents.
1995.
[147] T. Kanungo, D. Mount, N. Netanyahu, C. Piatko, R. Silverman, and A. Wu. An efficient k-
means clustering algorithm: Analysis and implementation. IEEE Transactions on Pattern
Analysis and Machine Intelligence, pages 881–892, 2002.
[148] H. Khondker. Glocalization as globalization: Evolution of a sociological concept.
Bangladesh e-Journal of Sociology, 1:1–9, 2004.
[149] M. Kimura. The neutral theory of molecular evolution. Cambridge University Press, 1983.
[150] A. King and S. Pimm. Complexity, diversity, and stability: A reconciliation of theoretical
and empirical results. American Naturalist, 122:229–239, 1983.
[151] B. King. Step-wise clustering procedures. Journal of the American Statistical Association,
69:86–101, 1967.
[152] T. Kohonen. Self-organization and associative memory. Springer, 1989.
[153] T. Kohonen. Self-Organizing Maps. Springer, 2001.
[154] A. Kolmogorov. Three approaches to the definition of the concept “quantity of
information”. Problems of Information Transmission, 1:3–11, 1965.
[155] A. Kotok. ebXML: The New Global Standard for Doing Business Over the Internet. Sams
Publishing, 2001.
[156] J. Koza. Genetic Programming: On the Programming of Computers by Means of Natural
Selection. MIT Press, 1992.
[157] J. Koza. Overview of genetic programming.
In Genetic Programming: On the
Programming of Computers by Means of Natural Selection [156], pages 73–78.
[158] D. Krafzig, K. Banke, and D. Slama. Enterprise SOA: Service-Oriented Architecture Best
Practices. Prentice Hall, 2004.
[159] D. Kramer and M. MacInnis. Utilization of a local grid of Mac OS X-based computers
using Xgrid. In International Symposium on High Performance Distributed Computing,
pages 264–265. IEEE Press, 2004.
[160] M. Kulkarni and R. Kreutzer. Building your own digital ecosystem: a holistic approach
to enterprise integration. Technical report, Syntel, 2006. Available from:
http://www.syntelinc.com/uploadedFiles/Syntel DigitalEcosystem.pdf.
[161] T. Kurz. EvE simulator [online]. 2008. Available from:
http://evesim.sourceforge.net/ [cited 26 Jan 2008].
196
BIBLIOGRAPHY
[162] T. Kurz and T. Heistracher. Simulation of a self-optimising digital ecosystem. In Digital
Ecosystems and Technologies Conference, pages 165–170. IEEE Press, 2007.
[163] D. Lambert and M. Cooper. Issues in supply chain management. Industrial Marketing
Management, 29:65–83, 2000.
[164] W. Langdon. The evolution of size in variable length representations. In P. Simpson,
editor, International Conference on Evolutionary Computation, pages 633–638. IEEE
Press, 1998.
[165] W. Langdon. Genetic Programming and Data Structures: Genetic Programming + Data
Structures = Automatic Programming! Kluwer Publishers, 1998.
[166] W. Langdon and R. Poli. Fitness causes bloat. In P. Chawdhry, R. Roy, and R. Pant,
editors, Soft Computing in Engineering Design and Manufacturing, pages 13–22. Springer,
1997.
[167] D. Larose. Discovering Knowledge in Data: An Introduction to Data Mining. Wiley,
2005.
[168] E. Lawrence. Henderson’s dictionary of biological terms. Pearson Education, 2005.
[169] L. Lee, H. Nwana, D. Ndumu, and P. De Wilde. The stability, scalability and performance
of multi-agent systems. BT Technology Journal, 16:94–103, 1998.
[170] J. Lehn.
Perspectives in supramolecular chemistry—from molecular recognition
towards molecular information processing and self-organization. Angewandte Chemie
International Edition in English, 29:1304–1319, 1990.
[171] G. Lendaris. On the definition of self-organizing systems. In W. Banzhaf, editor, European
Conference on Advances in Artificial Life, pages 324–325. IEEE Press, 1964.
[172] E. Lenski, C. Ofria, R. Pennock, and C. Adami. The evolutionary origin of complex
features. Nature, 423:139–144, 2003.
[173] S. Levin. Ecosystems and the biosphere as complex adaptive systems. Ecosystems, 1:431–
436, 1998.
[174] S. Levin. Fragile dominion: complexity and the commons. Perseus Books Group, 1999.
[175] R. Levins. Some demographic and genetic consequences of environmental heterogeneity
for biological control. Bulletin of the Entomological Society of America, 15:237–240, 1969.
[176] F. Leymann, D. Roller, and M. Schmidt. Web services and business process management.
IBM Systems Journal, 41:198–211, 2002.
[177] M. Li and P. Vit´anyi. An Introduction to Kolmogorov Complexity and Its Applications.
Springer, 1997.
[178] S. Lin, W. Punch III, and E. Goodman. Coarse-grain parallel genetic algorithms:
categorization and new approach. In Symposium on Parallel and Distributed Processing,
pages 28–37. IEEE Press, 1994.
[179] S. Loncaric. A survey of shape analysis techniques. Pattern Recognition, 31:983–1001,
1998.
BIBLIOGRAPHY
197
[180] S. Lu and K. Fu. A sentence-to-sentence clustering procedure for pattern analysis. IEEE
Transactions on Systems, Man, and Cybernetics, 8:381–389, 1978.
[181] A. Lugo. The apparent paradox of reestablishing species richness on degraded lands with
tree monocultures. Forest Ecology and Management, 99:9–19, 1997.
[182] S. Luke, C. Cioffi-Revilla, L. Panait, and K. Sullivan. MASON: A new multi-agent
simulation toolkit. In SwarmFest. University of Michigan, 2004.
[183] S. Luke, L. Panait, G. Balan, S. Paus, Z. Skolicki, E. Popovici, K. Sullivan, J. Harrison,
J. Bassett, R. Hubley, and A. Chircop. ECJ: A Java-based evolutionary computation
research system [online]. Available from: http://cs.gmu.edu/ eclab/projects/ecj/.
[184] K. Lyytinen and Y. Yoo. The next wave of nomadic computing: A research agenda for
information systems research. Sprouts: Working Papers on Information Systems, 1:1–20,
2001.
[185] R. MacArthur and E. Wilson. The Theory of Island Biogeography. Princeton University
Press, 1967.
[186] D. MacKay.
Information Theory, Inference, and Learning Algorithms. Cambridge
University Press, 2003.
[187] J. MacQueen. Some methods for classification and analysis of multivariate. In L. Le Cam
and J. Neyman, editors, Berkeley Symposium on Mathematical Statistics and Probability.
University of California Press, 1965.
[188] M. Mamei and F. Zambonelli. Self-organization in multi agent systems: a middleware
approach. In G. Serugendo, editor, Engineering Self-Organising Systems, pages 233–248.
Springer, 2003.
[189] B. Manderick and P. Spiessens.
In J.
Schaffer, editor, International Conference on Genetic Algorithms, pages 428–433. Morgan
Kaufmann Publishers, 1989.
Fine-grained parallel genetic algorithms.
[190] C. Manning, P. Raghavan, and H. Schutze.
Introduction to Information Retrieval.
Cambridge University Press, 2008.
[191] T. Mantere and J. Alander. Evolutionary software engineering, a review. Applied Soft
Computing, 5:315–331, 2005.
[192] D. Marais and M. Walter. Astrobiology: Exploring the origins, evolution, and distribution
of life in the universe. Annual Reviews in Ecology and Systematics, 30:397–420, 1999.
[193] J. Markoff. Faster chips are leaving programmers in their dust. Technical report, New
York Times, 2007. Available from:
http://www.nytimes.com/2007/12/17/technology/17chip.html.
[194] P. Marrone. Java object oriented neural engine: The complete guide. Technical report,
joone.org, 2007. Available from:
http://prdownloads.sourceforge.net/joone/JooneCompleteGuide.pdf.
[195] P. Marrow. Nature-inspired computing technology and applications. BT Technology
Journal, 18:13–23, 2000.
198
BIBLIOGRAPHY
[196] P. Marrow, M. Koubarakis, R. van Lengen, F. Valverde-Albacete, E. Bonsma, J. Cid-
Suerio, A. Figueiras-Vidal, A. Gallardo-Antolin, C. Hoile, T. Koutris, H. Molina-Bulla,
A. Navia-Vazquez, P. Raftopoulou, N. Skarmeas, C. Tryfonopoulos, F. Wang, and C.
Xiruhaki. Agents in decentralized information ecosystems: The DIET approach.
In
Symposium on Information Agents for Electronic Commerce, pages 109–117. Society for
the Study of Artificial Intelligence and Simulation of Behaviour, 2001.
[197] D. Martin, M. Paolucci, S. McIlraith, M. Burstein, D. McDermott, D. McGuinness, B.
Parsia, T. Payne, M. Sabou, M. Solanki, N. Srinivasan, and K. Sycara. Bringing semantics
to web services: The OWL-S approach. In J. Cardoso and A. Sheth, editors, Semantic
Web Services and Web Process Composition, pages 6–9. Springer, 2004.
[198] T. Marwala, P. De Wilde, L. Correia, P. Mariano, R. Ribeiro, V. Abramov, N. Szirbik,
and J. Goossenaerts. Scalability and optimisation of a committee of agents using genetic
In D. Campbell and C. Fyfe, editors, International ICSC Symposium Soft
algorithm.
Computing and Intelligent Systems For Industry. ICSC-NAISO Academic Press, 2001.
[199] C. Masuch. Evolutionary environment network [online]. 2008. Available from:
http://evenet.sourceforge.net/ [cited 31 Jan 2008].
[200] F. McCabe and K. Clark. April-agent process interaction language. In M. Wooldridge
and N. Jennings, editors, Intelligent Agents: Workshop on Agent Theories, Architectures,
and Languages, pages 324–340. Springer, 1994.
[201] T. McClanahan and R. Wolfe. Accelerating forest succession in a fragmented landscape:
The role of birds and perches. Conservation Biology, 7:279–288, 1993.
[202] S. McCulloch, R. Kokoska, O. Chilkova, C. Welch, E. Johansson, P. Burgers, and T.
Kunkel. Enzymatic switching for efficient and accurate translesion DNA replication.
Nucleic Acids Research, 32:4665–4675, 2004.
[203] S. McIlraith, C. Son, and H. Zeng. Semantic web services. IEEE Intelligent Systems,
16:46–53, 2001.
[204] N. McPhee and R. Poli. A schema theory analysis of the evolution of size in genetic
programming with linear representations. In J. Miller, M. Tomassini, P. Lanzi, C. Ryan,
A. Tettamanzi, and W. Langdon, editors, European Conference on Genetic Programming,
pages 108–125. Springer, 2001.
[205] B. Medjahed, A. Bouguettaya, and A. Elmagarmid. Composing web services on the
semantic web. The International Journal on Very Large Data Bases, 12:333–351, 2003.
[206] D. Michie, D. Spiegelhalter, C. Taylor, and J. Campbell. Machine learning, neural and
statistical classification. Ellis Horwood, 1995.
[207] N. Milanovic and M. Malek. Current solutions for web service composition. IEEE Internet
Computing, 8:51–59, 2004.
[208] J. Miller. What bloat?
In
L. Spector, editor, Genetic and Evolutionary Computation Conference, pages 295–302.
Morgan Kaufmann Publishers, 2001.
cartesian genetic programming on boolean problems.
[209] S. Miller. Aspect-oriented programming takes aim at software complexity.
IEEE
Computer, 34:18–21, 2001.
BIBLIOGRAPHY
199
[210] G. Milligan and M. Cooper. An examination of procedures for determining the number
of clusters in a data set. Psychometrika, 50:159–179, 1985.
[211] G. Milligan. An examination of the effect of six types of error perturbation on fifteen
clustering algorithms. Psychometrika, 45:325–342, 1980.
[212] M. Mitchell. An Introduction to Genetic Algorithms. MIT Press, 1996.
[213] G. Modi. Service oriented architecture & web 2.0. Technical report, Guru Tegh Bahadur
Institute of Technology, 2007. Available from:
http://www.gsmodi.com/files/SOA Web2 Report.pdf.
[214] J. Moore. The Death of Competition: Leadership and Strategy in the Age of Business
Ecosystems. Harvard Business School Press, 1996.
[215] J. Moore. Digital business ecosystems in developing countries: An introduction. Technical
report, Berkman Center for Internet and Society, Harvard Law School, 2003. Available
from: http://cyber.law.harvard.edu/bold/devel03/modules/moore.pdf.
[216] L. Moreau. Stability of multiagent systems with time-dependent communication links.
IEEE Transactions on Automatic Control, 50:169–182, 2005.
[217] R. Morrison. Designing Evolutionary Algorithms For Dynamic Environments. Springer,
2004.
[218] S. Mufwene. The Ecology of Language Evolution. Cambridge University Press, 2001.
[219] H. Muhlenbein. Evolution in time and space - the parallel genetic algorithm. Foundations
of Genetic Algorithms, 1:316–337, 1991.
[220] F. Murtagh. A survey of recent advances in hierarchical clustering algorithms. The
Computer Journal, 26:354–359, 1983.
[221] F. Nachira. Towards a network of digital business ecosystems fostering the local
development. Technical report, Directorate General Information Society and Media,
European Commission, 2002. Available from:
http://www.digital-ecosystems.org/doc/discussionpaper.pdf.
[222] F. Nachira, P. Dini, and A. Nicolai. A network of digital business ecosystems for europe:
Roots, processes and perspectives. In Nachira et al. [223], pages 1–20.
[223] F. Nachira, A. Nicolai, P. Dini, M. Le Louarn, and L. Rivera Le´on, editors. Digital
Business Ecosystems. European Commission, 2007.
[224] S. Naeem and S. Li. Biodiversity enhances ecosystem reliability. Nature, 390:507–509,
1997.
[225] G. Nagy. State of the art in pattern recognition. Proceedings of the IEEE, 56:836–863,
1968.
[226] S. Narayanan and S. McIlraith. Simulation, verification and automated composition of
web services. In international conference on World Wide Web, pages 77–88. ACM Press,
2002.
[227] R. Nelson and S. Winter. An Evolutionary Theory of Economic Change. Harvard
University Press, 1982.
200
BIBLIOGRAPHY
[228] E. Newcomer and G. Lomow. Understanding SOA with web services. Addison-Wesley,
2005.
[229] M. Newman.
Self-organized criticality, evolution and the fossil extinction record.
Proceedings of the Royal Society B: Biological Sciences, 263:1605–1610, 1996.
[230] M. Newman. A model of mass extinction. Journal of Theoretical Biology, 189:235–252,
1997.
[231] G. Nicolis and I. Prigogine.
Self-organization in Nonequilibrium Systems: From
Dissipative Structures to Order Through Fluctuations. Wiley, 1977.
[232] A. Nix and M. Vose. Modeling genetic algorithms with Markov chains. Annals of
Mathematics and Artificial Intelligence, 5:79–88, 1992.
[233] M. Nixon and A. Aguado. Feature Extraction & Image Processing. Newnes, 2007.
[234] J. Norris. Markov Chains. Cambridge University Press, 1997.
[235] M. Nowostawski and R. Poli. Parallel genetic algorithm taxonomy.
In L. Jain,
editor, International Conference on Knowledge-Based Intelligent Information Engineering
Systems, pages 88–92. IEEE Press, 1999.
[236] H. Nwana. Software agents: An overview. Knowledge Engineering Review, 11:205–244,
1996.
[237] E. Odum. Energy flow in ecosystems: A historical review. Integrative and Comparative
Biology, 8:11–18, 1968.
[238] R. Olfati-Saber, J. Fax, and R. Murray. Consensus and cooperation in networked multi-
agent systems. Proceedings of the IEEE, 95:215–233, 2007.
[239] C. Olson. Parallel algorithms for hierarchical clustering. Parallel Computing, 21:1313–
1325, 1995.
[240] C. Papadimitriou and K. Steiglitz. Combinatorial Optimization: Algorithms and
Complexity. Dover Publications, 1998.
[241] M. Papazoglou and D. Georgakopoulos. Service-oriented computing. Communications of
the ACM, 46:25–28, 2003.
[242] M. Papazoglou, P. Traverso, S. Dustdar, and F. Leymann. Service-oriented computing:
State of the art and research challenges. IEEE Computer, 40:38–45, 2007.
[243] M. Papazoglou. Service-oriented computing: concepts, characteristics and directions.
In T. Catarci, M. Mecella, J. Mylopoulos, and M. Orlowska, editors, International
Conference on Web Information Systems Engineering, pages 3–12. IEEE Press, 2003.
[244] G. Parisi. Statistical Field Theory. Perseus Books Group, 1998.
[245] H. Parunak and S. Brueckner. Entropy and self-organization in multi-agent systems. In
International Conference on Autonomous Agents, pages 124 – 130. ACM Press, 2001.
[246] T. Pavlidis. Structural Pattern Recognition. Springer, 1977.
[247] R. Pennock. Evolving intelligence [online]. 2006. Available from:
http://www.msu.edu/ pennock5/research/EI.html.
BIBLIOGRAPHY
201
[248] L. Perlovsky. Conundrum of combinatorial complexity. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 20:666–670, 1998.
[249] V. Pham and A. Karmouch. Mobile software agents: an overview. IEEE Communications
Magazine, 36:26–37, 1998.
[250] J. Platt. Fast training of support vector machines using sequential minimal optimization.
In B. Scholkopf, C. Burges, and A. Smola, editors, Advances in kernel methods: support
vector learning, pages 185–208. MIT Press, 1999.
[251] R. Poli and N. McPhee. Exact schema theorems for GP with one-point and standard
crossover operating on linear structures and their application to the study of the evolution
of size. In J. Miller, editor, European conference on Genetic Programming, pages 126–142.
Springer, 2001.
[252] V. Porto. Evolutionary programming. In T. Baeck, D. Fogel, and Z. Michalewicz, editors,
Handbook of Evolutionary Computation, pages 89–102. CRC Press, 2000.
[253] A. Prugel-Bennett. Modelling evolving populations. Journal of Theoretical Biology,
185:81–95, 1997.
[254] P. Rajasekaran, J. Miller, K. Verma, and A. Sheth. Enhancing web services description
and discovery to facilitate composition. In J. Cardoso and A. Sheth, editors, Semantic
Web Services and Web Process Composition, pages 55–68. Springer, 2004.
[255] J. Rao and X. Su. A survey of automated web service composition methods.
In J.
Cardoso and A. Sheth, editors, Semantic Web Services and Web Process Composition,
pages 43–54. Springer, 2004.
[256] Rapid-I. Rapidminer 4.3. Technical report, Rapid-I GmbH, 2008.
[257] S. Rasmussen, L. Chen, D. Deamer, D. Krakauer, N. Packard, P. Stadler, and M. Bedau.
Evolution: Transitions from nonliving to living matter. Science, 303:963–965, 2004.
[258] N. Rathbone. Open Philosophies for Associative Autopoietic Digital Ecosystems
(OPAALS) [online]. Available from: http://www.opaals.org/research.php [cited 26 Jan
2008].
[259] A. Ratle and M. Sebag. Avoiding the bloat with probabilistic grammar-based genetic
programming. In P. Collet, C. Fonlupt, J. Hao, E. Lutton, and M. Schoenauer, editors,
International Conference on Artificial Evolution, pages 255–266. Springer, 2001.
[260] A. Razavi, S. Moschoyiannis, and P. Krause. A coordination model for distributed
transactions in digital business ecosystems. Digital Ecosystems and Technologies
Conference, pages 159–164, 2007.
[261] A. Redmore and M. Griffin. Advanced Level and Advanced Special Level Biology. Longman
Group Limited, 1994.
[262] B. Ripley. Pattern Recognition and Neural Networks. Cambridge University Press, 1996.
[263] J. Risson and T. Moors. Survey of research towards robust peer-to-peer networks: Search
methods. Computer Networks, 50:3485–3521, 2006.
202
BIBLIOGRAPHY
[264] L. Rivera Le´on. Digital Ecosystems Network of Regions for Dissemination and Knowledge
Deployment (DEN4DEK) [online]. Available from: http://www.den4dek.org [cited 20 Jan
2009].
[265] L. Rivera Le´on. Regions for Digital Ecosystems Network (REDEN) [online]. Available
from: http://reden.opaals.org/ [cited 20 Jan 2009].
[266] R. Rivest. The MD5 message-digest algorithm. Technical report, MIT, 1992. Available
from: http://people.csail.mit.edu/rivest/Rivest-MD5.txt.
[267] R. Robertson. Globalisation or glocalisation. Journal of International Communication,
1:33–52, 1994.
[268] R. Rojas. Neural Networks: A Systematic Introduction. Springer, 1996.
[269] K. Rothermel and F. Hohl. Mobile agents.
In A. Kent and J. Williams, editors,
Encyclopedia for Computer Science and Technology, volume 40, pages 155–176. CRC
Press, 1999.
[270] J. Rowe and B. Mitavskiy. Report on evolution of high-level software components.
Technical report, Digital Business Ecosystems project, 2005. Available from:
http://files.opaals.org/DBE/deliverables/Del 8.1 DBE Report.pdf.
[271] G. Rudolph. Convergence analysis of canonical genetic algorithms. IEEE Transactions
on Neural Networks, 5:96–101, 1994.
[272] G. Rudolph. Finite Markov chain results in evolutionary computation: A tour d’horizon.
Fundamenta Informaticae, 35:67–89, 1998.
[273] M. Salmon. Introduction to the Philosophy of Science. Hackett Publishing, 1999.
[274] W. Sarle. How many hidden units should i use? Technical report, Internet FAQ Archives,
2008. Available from:
http://www.faqs.org/faqs/ai-faq/neural-nets/part3/section-10.html.
[275] M. Schmiemann. SMEs and entrepreneurship in the EU. Technical report, Statistical
Office of the European Communities (Eurostat), European Commission, 2006. Avail-
able from: http://epp.eurostat.ec.europa.eu/cache/ITY OFFPUB/KS-NP-06-024/EN/
KS-NP-06-024-EN.PDF.
[276] B. Scholkopf, K. Sung, C. Burges, F. Girosi, P. Niyogi, T. Poggio, and V. Vapnik.
Comparing support vector machines with gaussian kernels to radial basis function
classifiers. IEEE Transactions on Signal Processing, 45:2758–2765, 1997.
[277] J. Schumann. Automated Theorem Proving in Software Engineering. Springer, 2001.
[278] H. Schwefel. Evolution and Optimum Seeking: The Sixth Generation. Wiley, 1995.
[279] D. Searls. The language of genes. Nature, 420:211–217, 2002.
[280] I. Sethi and A. Jain. Artificial neural networks and statistical pattern recognition. Old
and new connections. North-Holland, 1991.
[281] M. Shackleton, R. Shipma, and M. Ebner. An investigation of redundant genotype-
phenotype mappings and their role in evolutionary search. In Congress on Evolutionary
Computation, pages 493–500. IEEE Press, 2000.
BIBLIOGRAPHY
203
[282] C. Shalizi. Causal Architecture, Complexity and Self-Organization in Time Series and
Cellular Automata. PhD thesis, University of Wisconsin-Madison, 2001.
[283] C. Shalizi and K. Shalizi. Quantifying self-organization in cyclic cellular automata. In
L. Schimansky-Geier, D. Abbott, A. Neiman, and C. Broeck, editors, Noise in Complex
Systems and Stochastic Dynamics, pages 108–117. SPIE, 2003.
[284] J. Shapiro. Statistical mechanics theory of genetic algorithms. In L. Kallel, B. Naudts,
and A. Rogers, editors, Theoretical Aspects of Evolutionary Computing, pages 87–108.
Springer, 2001.
[285] R. Shipman, M. Shackleton, and I. Harvey. The use of neutral genotype-phenotype
mappings for improved evolutionary search. BT Technology Journal, 18:103–111, 2000.
[286] M. Simoes-Marques, P. Mariano, R. Ribeiro, L. Correia, M. Chli, P. De Wilde, V.
Abramov, and J. Goosenaerts. Contributions to adaptable agent societies. In Emerging
Technologies and Factory Automation, pages 354–361. IEEE Press, 2003.
[287] M. Singh and M. Huhns. Service-Oriented Computing: Semantics, Processes, Agents.
Wiley, 2005.
[288] A. Sizling and D. Storch. Power-law species-area relationships and self-similar species
distributions within finite areas. Ecology Letters, 7:60–68, 2004.
[289] J. Slagle. Automated theorem-proving for theories with simplifiers, commutativity, and
associativity. Journal of the ACM, 21:622–642, 1974.
[290] R. Smith and N. Taylor. A framework for evolutionary computation in agent-based
systems. In J. Glasgow, editor, International Conference on Intelligent Systems, pages
221–224. AAAI Press, 1998.
[291] P. Sneath and R. Sokal. Numerical Taxonomy: The Principles and Practice of Numerical
Classification. Springer, 1973.
[292] K. Sneppen, P. Bak, H. Flyvbjerg, and M. Jensen. Evolution as a self-organized critical
phenomenon. Proceedings of the National Academy of Sciences, 92:5209–5213, 1995.
[293] R. Sol´e, D. Alonso, and A. McKane. Self-organized instability in complex ecosystems.
Philosophical Transactions: Biological Sciences, 357:667–681, 2002.
[294] I. Sommerville. Software Engineering. Addison-Wesley, 1992.
[295] K. Soraku-gun. An evolutionary approach to synthetic biology: Zen and the art of creating
life. In C. Langton, editor, Artificial Life: An Overview, pages 195–226. MIT Press, 1995.
[296] T. Soule and J. Foster. Effects of code growth and parsimony pressure on populations in
genetic programming. Evolutionary Computation, 6:293–309, 1998.
[297] J. Stender. Parallel Genetic Algorithms: Theory and Applications. IOS Press, 1993.
[298] K. Sterelny. A review of evolution and learning: the baldwin effect reconsidered. Evolution
& Development, 6:295–300, 2004.
[299] H. Sutter. The free lunch is over: A fundamental turn toward concurrency in software.
Dr. Dobb’s Journal, 2005.
204
BIBLIOGRAPHY
[300] J. Suzuki. A Markov chain analysis on simple genetic algorithms. IEEE Transactions on
Systems, Man, and Cybernetics, 25:655–659, 1995.
[301] M. Svahnberg. Background analysis and design of an agent-based operating system.
Master’s thesis, University of Karlskrona/Ronneby, 1998. Available from:
http://www.ide.hk-r.se/ nesse/ABOS/ABOS thesis.pdf.
[302] K. Swingler. Applying Neural Networks: A Practical Guide. Morgan Kaufmann
Publishers, 1996.
[303] E. Swyngedouw. The mammon quest. ‘Glocalisation’, interspatial competition and the
monetary order: the construction of new scales. In M. Dunford and G. Kafkalas, editors,
Cities and regions in the new Europe: The Global-local Interplay and Spatial Development
Strategies, pages 39–67. Wiley, 1992.
[304] N. Takahashi.
Stability analysis of nonlinear dynamical systems related to neural
networks. Technical report, Kyushu University, 2008. Available from:
http://www-kairo.csce.kyushu-u.ac.jp/ norikazu/research.en.html.
[305] Q. Tang. Economics Of Web Service Provisioning: Optimal Market Structure And
Intermediary Strategies. PhD thesis, University of Florida, 2004.
[306] J. Thomas and K. Sycara. Heterogeneity, stability, and efficiency in distributed systems.
In Y. Demazeau, editor, International Conference on Multi Agent Systems, pages 293 –
300. IEEE Press, 1998.
[307] C. Thomas Jr. The genetic organization of chromosomes. Annual Reviews in Genetics,
5:237–256, 1971.
[308] H. Tianfield. Towards self-organization in multi-agent systems and grid computing.
Multiagent and Grid Systems, 1:89–95, 2005.
[309] D. Tilman and P. Kareiva. Spatial Ecology: The Role of Space in Population Dynamics
and Interspecific Interactions. Princeton University Press, 1997.
[310] P. Trahanias and E. Skordalakis. Syntactic pattern recognition of the ECG.
IEEE
Transactions on Pattern Analysis and Machine Intelligence, 12:648–657, 1990.
[311] R. Trivers. The evolution of reciprocal altruism. The Quarterly Review of Biology, 46:35,
1971.
[312] M. Turner. Factors influencing succession: Lessons from large,
infrequent natural
disturbances. Ecosystems, 1:511–523, 1998.
[313] F. Valafar. Pattern recognition techniques in microarray data analysis: A survey. In F.
Valafar, editor, Techniques in Bioinformatics and Medical Informatics, pages 41–64. New
York Academy of Sciences, 2002.
[314] S. Valverde-Castillo. Software design as an adaptive walk [online]. 2004. Available from:
http://web.archive.org/web/20041206173432/http://complex.upf.es/ sergi/.
[315] M. Vandenberghe. Digital ecosystem solution - the business factory. Technical report,
XeWOW, 2006. Available from:
http://themaddesigner.free.fr/XeWOW%20White%20Paper.pdf.
[316] V. Vapnik. Statistical learning theory. Wiley, 1998.
BIBLIOGRAPHY
205
[317] V. Vapnik and S. Kotz. Estimation of Dependences Based on Empirical Data. Birkhauser,
2006.
[318] D. Veldhuizen and G. Lamont. Multiobjective evolutionary algorithms: Analyzing the
state-of-the-art. Evolutionary Computation, 8:125–147, 2000.
[319] S. Ventura, C. Romero, A. Zafra, J. Delgado, and C. Herv´as. JCLEC: a Java framework
for evolutionary computation. Soft Computing - A Fusion of Foundations, Methodologies
and Applications, 12:381–392, 2008.
[320] B. Violino. How to navigate a sea of SOA standards [online]. 2007. Available from:
http://www.cio.com/article/104007/How to Navigate a Sea of SOA Standards.
[321] L. von Bertalanffy. General system theory: Foundations, Development, Applications.
Braziller, 1973.
[322] M. Waldrop. Complexity: The Emerging Science at the Edge of Order and Chaos. Simon
& Schuster, 1992.
[323] C. Walters and R. Hilborn. Ecological optimization and adaptive management. Annual
Reviews in Ecology and Systematics, 9:157–188, 1978.
[324] J. Ward. Hierarchical grouping to optimize an objective function. Journal of the American
Statistical Association, 58:236–244, 1963.
[325] M. Ward. Life offers lessons for business [online]. 2004. Available from:
http://news.bbc.co.uk/1/hi/technology/3752725.stm.
[326] M. Waterman. Introduction to Computational Biology: Maps, Sequences and Genomes.
CRC Press, 1995.
[327] D. Watts and S. Strogatz. Collective dynamics of ‘small-world’ networks. Nature, 393:440–
442, 1998.
[328] A. Webb. Statistical Pattern Recognition. Newnes, 1999.
[329] B. Weber and D. Depew. Evolution and Learning: The Baldwin Effect Reconsidered.
Bradford Books, 2003.
[330] J. Weibull. Evolutionary Game Theory. MIT Press, 1995.
[331] G. Weiss. Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence.
MIT Press, 1999.
[332] E. Weisstein. CRC Concise Encyclopedia of Mathematics. CRC Press, 2003.
[333] D. White and M. Houseman. The navigability of strong ties: Small worlds, tie strength,
and network topology. Complexity, 8:72–81, 2002.
[334] M. Wooldridge. Introduction to MultiAgent Systems. Wiley, 2002.
[335] S. Wright. The roles of mutation, inbreeding, crossbreeding and selection in evolution. In
D. Jones, editor, International Congress on Genetics, pages 356–366. Brooklyn botanic
garden, 1932.
[336] J. Wunderle. The role of animal seed dispersal in accelerating native forest regeneration
on degraded tropical lands. Forest Ecology and Management, 99:223–235, 1997.
206
BIBLIOGRAPHY
[337] Ximbiotix. About the digital ecosystem [online]. 2005. Available from: http://www.ximb
iotix.com/desktop/the-digital-ecosystem/about-the-digital-ecosystem.cfm [cited 26 Jan
2008].
[338] X. Yang. Chaos in small-world networks. Physical Review E, 63:1–4, 2001.
[339] L. Zadeh. Fuzzy sets. In G. Klir and B. Yuan, editors, Fuzzy Sets, Fuzzy Logic, and Fuzzy
Systems: Selected Papers, pages 19–34. World Scientific, 1996.
[340] C. Zahn. Graph-theoretical methods for detecting and describing gestalt clusters.
Transactions on Computers, 100:68–86, 1971.
[341] O. Zimmermann, S. Milinski, M. Craes, and F. Oellermann. Second generation web
services-oriented architecture in production in the finance industry. In J. Vlissides and
D. Schmidt, editors, Conference on Object-Oriented Programming, Systems, Languages,
and Applications, pages 283–289. ACM Press, 2004.
[342] W. Zurek. Algorithmic information content, church-turing thesis. physical entropy,
In W. Zurek, editor, Complexity, Entropy, and the Physics of
and maxwll’s demon.
Information, pages 73–89. Addison-Wesley, 1990.
[343] A. Zvonkin and L. Levin. The complexity of finite objects and the development of the
concepts of information and randomness by means of the theory of algorithms. Russian
Mathematical Surveys, 25:83–124, 1970.
|
1702.03226 | 2 | 1702 | 2017-03-24T11:29:52 | An applied spatial agent-based model of administrative boundaries using SEAL | [
"cs.MA"
] | This paper extends and adapts an existing abstract model into an empirical metropolitan region in Brazil. The model - named SEAL: a Spatial Economic Agent-based Lab - comprehends a framework to enable public policy ex-ante analysis. The aim of the model is to use official data and municipalities spatial boundaries to allow for policy experimentation. The current version considers three markets: housing, labor and goods. Families' members age, consume, join the labor market and trade houses. A single consumption tax is collected by municipalities that invest back into quality of life improvements. We test whether a single metropolitan government - which is an aggregation of municipalities - would be in the best interest of its citizens. Preliminary results for 20 simulation runs indicate that it may be the case. Future developments include improving performance to enable running of higher percentage of the population and a number of runs that make the model more robust. | cs.MA | cs |
An applied spatial agent-based model of
administrative boundaries using SEAL
Bernardo Alves Furtado1,2 and Isaque Daniel Eberhardt Rocha1
1 Institute of Applied Economic Research, Brazil
[email protected]
2 National Council of Research, CNPq
Abstract. This paper extends and adapts an existing abstract model
[14] into an empirical metropolitan region in Brazil. The model - named
SEAL: a Spatial Economic Agent-based Lab - comprehends a framework
to enable public policy ex-ante analysis. The aim of the model is to
use official data and municipalities spatial boundaries to allow for policy
experimentation. The current version considers three markets: housing,
labor and goods. Families' members age, consume, join the labor market
and trade houses. A single consumption tax is collected by municipali-
ties that invest back into quality of life improvements. We test whether
a single metropolitan government - which is an aggregation of munic-
ipalities - would be in the best interest of its citizens. Results for 20
simulation runs indicate that it may be the case. Future developments
include improving performance to enable running of higher percentage of
the population and a number of runs that make the model more robust
as well as further policy analysis.
Keywords: Agent-based model, Public Policy, Spatial modeling, Metropoli-
tan regions, Public Finances
1
Introduction
Anticipating public policy effects is no simple task [8,19,17]. Once planned and
put into effect, policies are subject to interested actors reaction, learning, chance,
and emergent behavior [23]. That is why governments actions are monitored and
ex-post evaluation is implemented.
Understanding of policies' effects ex-ante is relevant in the Brazilian context
for at least two main reasons. Firstly, municipalities in Brazil are autonomous,
independent entities, comprising a three-level federation since the 1988's Con-
stitution. That implies metropolitan neighbors can act politically independently
in terms of transportation or other public services.
Secondly, metropolises in Brazil tend to be centralized with concentration of
economic activity, higher level of business and amenities in the capital. Thus,
agglomeration economies and fiscal regulation favor capital cities in detriment
of nearby municipalities [16]. Peripheral areas, in turn, observe higher levels
of violence, longer commuting times to the center and lower quality of public
2
services offer. In practice, that means most workers commute to the capital,
contribute to increased economic output, but go home at night to an outer
municipality that does not collect enough taxes to supply basic services.
As a result, urban economic space in Brazil is highly heterogeneous with
concentration of capital and quality in the center and lack of access to services
and jobs in the borders. Such spatial configuration, when divided by political-
administrative boundaries, tend to further segregate access to citizens and, thus,
lower economic output. In fact, political fragmentation in metropolises has been
found to reduce productivity [1].
Given such a context for the Brazilian case, an exercise of ex-ante policy
analysis that could verify whether a political entity spatial configuration serves
the public best may be helpful. Nevertheless, there is no evidence of such exercise.
Probably because solving what-if problems in an urban environment looking
towards the future is methodologically burdensome.
However, recent developments in agent-based modeling, especially for eco-
nomic [9,12] and urban analysis [13], have been shown to contribute to the
debate and enlighten future prospects of policy analysis. Moreover, ABMs are
helpful in modeling complex systems such as markets [10,4] in which interactions
among agents are not linearly related to observed macro structures.
The model presented here is intended as a simple one 3 in which the main
behaviors of the economy in space are captured. Hence, we depart from an
abstract previous model [14], turn the model into an empirical one and test a
specific policy, i.e., whether bounded municipalities, with a single metropolitan
government are more beneficial to the average citizen. Such benefit is tested
using a Quality of Life Index (QLI) weighted by the population. This test is
provided as an illustration of ex-ante policy analysis. Other possibilities of the
model as a framework are also discussed in 4.
The main contribution of the paper is an extension and testing of a simple,
fully functional, spatial modeling that is based on previous work [21,18,14]. Yet,
it goes further than land-use change analysis [13,25] to include spatially gov-
ernmental entities and fiscal testing in an applied exercise. Thus, moving the
arguments of [19] and [8] into a practical application.
Besides this introduction, the next section (2) presents and details the model
and the agents' behavior. Then preliminary results are discussed (3) along with
the limitations of this version. The closing section (4) develops the scope of the
model and details further developments.
2 Model and processes
The objective of SEAL is to model economic markets in space and time so
that policies and budgetary decisions can be evaluated given actual municipal
boundaries. Thus, citizens reside in domiciles that are precisely located in urban
and rural space within different municipalities. Businesses also have geographical
3 There has been a debate in ABM literature about whether a model should be large
and complex such as [11,20] or simple, as argued by [21].
3
coordinates. These locations allow the model to implement decisions based on
distance, such as shopping at the nearest shops or working closest to home. Also,
production and tax collections are localized. That means the spatial design of
taxes can be investigated. This is specially relevant for the case of Brazilian
metropolises in which there is clear disagreement among urban functional areas
and offer of public services [16] and commuting and transport are relevant given
the poor quality of infrastructure.
The model in [14] is based on [21] and [18]. However, the original abstract
model is built upon seven, squared and fictitious regions with 1,000 agents in
total. The one in this paper is based on actual municipal boundaries of the
metropolitan region of Bras´ılia, in Brazil. The model also uses census population
from the year 2000 [5], following age group proportions. The agents are then
bounded into families and they age following official mortality probability data.
Female agents give birth according to their fertility rates by age and state.
Number of firms and urban rural proportions also follow actual numbers and
spatial coordinates. A detailed description of the model is available in [15].
2.1 Agents
Citizens, families and households. Citizens are bounded into families and
are distinguished by gender, age and qualification. Families consolidate every
working members' income and make spending decisions together. They also move
and participate in the housing market as a single entity.
Houses are fixed and can be either empty or occupied although they always
have a single owner. Household addresses are the reference for all members of
the family. They are relevant when searching for a job, to compute commuting
statistics for those employed and when choosing firms to buy goods. Citizens,
families and households each have their own class instances. Children are born
and remain in their mothers' families. The model does not contemplate neither
new families nor marriage at this time.
Firms and government. Firms participate in the labor market and maintain
a number of workers that varies in size according to sales performance. Firms
are fixed in space and pay taxes to the government they belong to, given their
location. Production is proportional to labor force and its qualification. Workers'
productivity is adjusted by an α parameter. Technology is fixed, products are
homogeneous and prices depend on the level of firms' stock [3,9].
Governments - as municipalities or as a single metropolitan region - simply
collect consumption taxes at the moment of sales transaction at the location of
the firm. Municipalities' QLI starts at the same level of Human Index Develop-
ments in 2000. Then, they are incrementally raised as a linear value of collected
taxes within a given municipality weighted by current population. That implies a
municipality with a proportionally more active goods market and smaller popu-
lation collects more taxes and can increase QLI faster. Higher QLI makes houses
on that municipality more expensive.
4
2.2 Markets
Goods market. Families consume a given probabilistic share of their current
available cash depending on a propensity to consume parameter β. Families then
search either for closest firm or the cheapest price in a parameter-defined sample
of the market.
Once selected the firm, the family will spend all designated money for con-
sumption, given that the firm has enough product to supply. 4 Otherwise, the
family consumes whatever quantity is available. The money not set to consume
plus any amount not spent is saved by the family. Such saved money is afterwards
reserved for use in the housing market. Taxes are paid to the local government
of the firm when the family makes the purchase.
Housing market. House prices are given by a fixed size, intrinsic quality and
the QLI of their region [22]. The QLI represents amenities of the municipality
and richer municipalities have higher QLI levels. The housing market follows
empirical results for the Brazilian case which suggest that there are always un-
occupied houses [24,7]. Every month a percentage of families enter the housing
market. All vacant houses are also available for purchasing.
Both families and houses are sorted. Cheaper houses and families with less
savings rank first. If the family has enough savings to purchase a house, the
transaction happens. If not, the savings of the next family are checked. When
there are no more families with enough funds or available houses, the market
closes.
Every time a family buys a new house, they will evaluate their option of
actually moving to the new house. If the family lives in their best house, but all
adults are unemployed, they move to the second best house and then make the
best house available for sale in the market the following month. On the contrary,
if the family does not live in the best house and at least one adult is employed,
they move into the best house the family owns. In all other cases, the family
does not move.
Such a system of moving into the second best house and selling the best-
quality one, slowly tends to make the poorest move to the outer municipalities
where housing is cheaper [6]. Moreover, even though it would be in the family's
interest to remain in the affluent area, they actually are unable to afford it and
need to get liquidity to participate in the consumption market.
Labor market. All agents that are not employed and are over 15 or below
70 years old enter the labor market every month. Firms evaluate a number of
strategies. They quantify a 'cash reserve' value which enables them to pay out
employees before laying them off. If the firm's available cash is above such cash
reserve threshold, the firm enters the labor market hiring.
4 Probabilistically, on average, we assume that consumption at a single shop would
yield similar results if the families chose more places to shop.
5
Further, if the firm is currently with no employees but either their stock of
products or available cash is positive, they also enter the market. When the
firms' current cash drops below the monthly payroll, they start firing employees,
one at a time.
Once offering firms and applicants are defined, the labor market operates
under two alternatives: half the applicants are chosen according to their qualifi-
cation and half given their proximity to the firms' address. This is relevant for
the case of Brazil as commuting is expensive and time consuming due to its low
quality general inefficiency. Further, employers are legally bind to pay employ-
ees' commuting that exceeds 6% of their salary in what is called 'commuting-
voucher'. In practice, that means business discriminates for place of residence
for non-specialized positions.
Firms that had profit in the last quarter and have available cash above their
'cash reserve' level distribute such surpluses to their employees proportionally
to their qualification level.
Finally, firms are sorted according to their current calculated paying wages.
Firms paying higher salaries choose first and get employees with better qualifi-
cation or who live closer to their firms.
Spatial and markets dynamics. Those rules of the markets provide a very
dynamically spatial model with families moving within and across municipalities.
Firms getting varying demand and labor offers and municipalities collecting ever-
changing taxes which affect their QLI at different speeds. When QLI increases
too much, houses get more expensive and less families can afford to move in.
Less populated municipalities then may sell less which in turn decreases tax
collection.
Parameters This model has not been fully parametrized to fit a real example.
However, as made explicit in 3, a general economic pathway in terms of macro
indicators is observable. A brief sensitivity analysis of the model was conducted
in order to check whether small changes in the parameters alters the general
results obtained. The simulation of this paper uses the following parameters: α
= .3, β = .94, percentage of families entering the real estate market = .004,
vacant houses percentage = .09, and taxes are set at .25. Distance as a hiring
criteria is also a parameter.
2.3 Scheduling and multiple simulation runs
Simulation runs The model is structured so that a single simulation or multiple
runs can be made automatically. That means either the model runs once from
2000 to 2020 and the outputs are produced or the model repeatedly runs for a
given number of times and the accumulated results are produced. In each run,
different random numbers are used. That influences many aspects of the model,
such as: the choice of amount for families' consumption, entering the real estate
market, demographic probabilities, chance of being fired, among others.
6
Scheduling Agents are generated and saved for a given sample of population.
In a single run, the model first saves or loads the agents (2.1). Then actions
are implemented in a day, month, quarter and year pattern. The only activities
that happen in business days (set to 21) are daily production and commuting.
Monthly activities concentrate the core of the model and happen in the following
order:
1. Process demographics (aging, births, deaths).
2. Firms make payments.
3. Families redistribute their own cash within members.
4. Family members consume (goods market).
5. Governments collect taxes.
6. Governments spend the collected taxes on life quality improvement.
7. Firms calculate profits and update prices.
8. Labor market is processed.
9. Real estate market is processed.
10. Statistics and output are processed.
3 Results
Bras´ılia and its metropolitan region (RIDE) with 10 municipalities is the illus-
trative case study (see Figure 1). In order to evaluate the results, a simulation
from 2000 to 2020 was repeated a total of 20 times (with different random seeds).
The same model was run 10 times with the municipalities being considered in-
dependent entities just as they are today. Another 10 runs were made as if
all municipalities were one single government entity. That way, for half of the
simulation runs taxes are collected locally and lead to QLI increase within the
municipality where the transactions were made. For the other half, taxes are
globally collected and homogeneously improve the QLI for all municipalities.
Results indicate that, given the implicit variability of the model, average
population weighted QLI follow similar trajectory (see Figure2). Actually, a t test
could not reject the hypothesis of identical averages between the two alternatives
(together and separated municipalities)at 5% (t-statistic=-1.742). Although the
QLI is slightly superior for the runs in which the municipalities are together, we
are unable to state that such result holds for other samples of the population.
We are evaluating whether the increase in QLI has an absolute effect on
all citizens. However, if we focus on the individual poorer municipalities with
lower values for QLI, we can tell that a distributive effect actually happens with
gains observed in Figure 3b when compared to flat lines in Figure 3a. Of course,
such gains derive from a slower increase from Bras´ılia, resulting in a very small
absolute result for this model.
Finally, these outputs of the model despite being illustrative show that it
is possible to use spatially separated markets to observe differentiated behavior
across the territory. In fact, although average weighted QLI is the same, its
distribution in space is considerably different.
7
Fig. 1: Case study: Bras´ılia. The 10 municipalities in orange officially constitute
the metropolitan region (RIDE). The run of the individual municipalities are
compared with the run of the model considering the full RIDE as one single
government entity. Urban areas, in black, are proportionately more populated.
Commuting (Euclidean distance) is considered among households and firms in
both labor and goods market.
Sensitivity analysis The sensitivity analysis indicates that the changes af-
fected by the parameters are small in magnitude. Among the parameters tested,
α (labor productivity), β (propensity to consume) and Tax on consumption seem
to impact the most. Alpha produced the largest GINI index at the maximum
value at 0.5, but the lowest GINI was found at 0.3 and not the minimum value.
Consumption of the families is also higher at the lowest level of taxes. In general,
lower taxes (5%), implied in higher levels of firms' and families' wealth as well
as a higher absolute GDP value. Finally, higher values of beta lead to higher
production, but also higher inequality.
Size of the market, house vacancy, percentage of families entering the real
estate market or use of distance in the hiring process do not seem to impact the
general trends of the results in a relevant manner. Some variation is observed in
absolute GDP output as well as in the time onset of some trends. When distance
is totally disconsidered and hiring criteria is restricted to qualification only, the
economy seems to reach stability earlier. In contrast when distance is the sole
determinant of hiring, the economy takes longer to become stable and GINI
values are lower.
8
Fig. 2: Results of mean weighted Quality of Life Index (QLI) for Bras´ılia, 20 runs.
Blue lines represent runs with one government entity. Red lines boundaries are as
they are currently: 10 individual municipalities. Solid lines are mean values. QLI
increases according to government revenue collected for a given municipality.
Housing prices are influenced by QLI at their locations.
Limitations The results presented here are only illustrative. In order to make
them more robust, it would be necessary to run a higher number of simulations
with varying samples of the population so that a trend could be identified, simi-
larly to the proposed by [2]. Further, hitherto we have been unable to implement
multiprocessing and parallel programming while maintaining an object oriented
programing paradigm and intense agents interaction. Consequently, given the
need of repeated runs with large number of agents, simulations are taking un-
reasonably long.
Considering the content and the investigation of administrative boundaries,
we believe we would have to expand the tax systems in order to make it more
similar to observed taxes. That is especially relevant because of the regressive
tax system in Brazil that tends to favor metropolitan capitals [16].
Once a full tax system is implemented, the current model needs to be vali-
dated against observed data. As it is, unemployment remains between 0 and 5%
of the working force from the 5th year onwards. GINI coefficient levels are below
9
(a) individual municipalities
(b) municipalities as single entity
Fig. 3: QLI evolution of each municipality. On the top, municipalities collect
taxes as individual entities and invest in their own territory. On the right-hand
side, all taxes collected in the region are distributed homogeneously as if they
were a single government entity.
10
Brazilian standards at 0.28 on average. GDP per capita growth varies around
0.7%. Finally, this model, as the abstract one, does not contain a credit market.
4 The framework and final considerations
Turning a simple abstract model into an empirical one has presented some chal-
lenges. Firstly, implementing demographics dynamics meant reading age group,
gender and location of population size, aging all the agents and allowing for
fecundity while controlling for mortality by region.
Secondly, another difficult task is the need to maintain a system in which
firms keep a permanent working force, but have to face very dynamic and er-
ratic sales. Additionally, keeping the system on a growth path means avoiding
generalized shut down of firms, constant consumption of families and an active
labor market.
Thirdly, given the nature of agent-based models and its intrinsic stochasticity,
every test should include repeated simulations so that the conclusions are based
on most likely scenarios. However, running simulations with a 10% of the sample
for the case of Bras´ılia would mean 273,000 agents and such simulation would
need to be run at least 100 in order to guarantee robust results. Finally, a
sensitivity analysis with six main parameters and ten interval tests varying one
parameter per time only as the others remain fixed would imply 60 further
simulation runs.
Given such a context, this paper tests an illustrative ex-ante policy analysis
using simple markets and spatially bounded, local government entities. Thus,
it demonstrates the viability of a stable economic framework in which further
experimentation can be tried upon. In fact, SEAL's framework enables a large
number of policy analysis, including, at least: (a) transportation and commuting
studies, given that firms and households are known and dynamic in the case of
the latter, (b) price and wage strategies at the level of the firm specifically, (c)
labor market rules and relevance of workers' qualification, and (d) efficiency of
municipalities' public investments.
References
1. Ahrend, R., Farchy, E., Kaplanis, I., & Lembcke, A. C. (2014). What makes cities
more productive? Evidence on the role of urban governance from five OECD coun-
tries. OECD Regional Development Working Papers, 2014(5), 33.
2. Anselin, L. (1995). Local indicators of spatial associationLISA. Geographical Anal-
ysis, 27(2), 93-115.
3. Blinder, A. S. (1994). On sticky prices: academic theories meet the real world.
In Monetary policy (pp. 117-154). The University of Chicago Press: N. Gregory
Mankiw.
4. Boero, R., Morini, M., Sonnesa, M., & Terna, P. (2015). Agent-based models of the
economy: from theories to applications. New York, NY: Palgrave Macmillan.
5. Brazil: Censo demogr´afico 2000: agregado por setores censit´arios dos resultados do
universo. Instituto Brasileiro de Geografia e Estat´ıstica (IBGE) (2003).
11
6. Brueckner, J. (1987). The structure of urban equilibria: a unified treatment of the
Muth-Mills model. In Handbook of Regional and Urban Economics (pp. 821-845).
Elsevier Science Publishers B.V.
7. Cavalcanti, C. B., Oliveira, C. A. P. de, Teixeira, L. M., Giustina, Y. R. D., Furtado,
B. A., Krause, C. H., Nadalin, V. G. (2016). Retrato das ´areas centrais do Brasil.
8. Colander, D., Kupers, R.: Complexity and the Art of Public Policy: Solving Societys
Problems from the Bottom Up. Princeton University Press (2014).
9. Dawid, H., Gemkow, S., Harting, P., Van der Hoog, S., Neugart, M.: Agent-based
macroeconomic modeling and policy analysis: the Eurace@ Unibi model. Bielefeld
Working Papers in Economics and Management. (2014).
10. Dawid, H. (2015). Modeling the economy as a complex system. In Modeling com-
plex systems for public policies (pp. 169-190). Bras´ılia, DF: Bernardo Alves Furtado,
Patr´ıcia Sakowski, Marina H T´ovolli.
11. Deissenberg, C., Van Der Hoog, S., & Dawid, H. (2008). EURACE: A massively
parallel agent-based model of the European economy. Applied Mathematics and
Computation, 204(2), 541-552.
12. Dosi, G., Fagiolo, G., Napoletano, M., Roventini, A.: Income Distribution, Credit
and Fiscal Policies in an Agent-Based Keynesian Model. SSRN eLibrary. (2012).
13. Filatova, T., Parker, D., Van der Veen, A.: Agent-Based Urban Land Markets:
Agents Pricing Behavior, Land Prices and Urban Land Use Change. Journal of
Artificial Societies and Social Simulation. 12, (2009).
14. Furtado, B.A., Eberhardt, I.D.R.: A Simple Agent-Based Spatial Model of the
Economy: Tools for Policy. JASSS. 19, 12 (2016).
15. Furtado, B.A., Eberhardt, I.D.R., Messa, A.: SEAL's operating manual: a
Spatially-bounded Economic Agent-based Lab. arXiv:1609.03996 [cs, q-fin]. (2016).
16. Furtado, B.A., Mation, L., Monasterio, L.: Fatos estilizados das finan¸cas p´ublicas
municipais metropolitanas brasileiras entre 2000-2010. In: Territ´orio metropolitano,
pol´ıticas municipais. pp. 291-312. Bernardo Alves Furtado; Cleandro Krause; Karla
Fran¸ca, Bras´ılia (2013).
17. Furtado, B.A., Sakowski, P.A.M., T/'ovolli, M.H.: Modeling complex systems for
public policies. IPEA, Instituto de Pesquisa Economica Aplicada, Bras´ılia (2015).
18. Gaffeo, E., Gatti, D.D., Desiderio, S., Gallegati, M.: Adaptive microfoundations
for emergent macroeconomics. Eastern Economic Journal. 34, 441-463 (2008).
19. Geyer, R., Cairney, P.: Handbook on complexity and public policy. Edward Elgar
Publishing (2015).
20. Guocheng, W., Yuna, S., Jie, W., & Zili, W. (2015). Application analysis on large-
scale computation for social and economic systems. Presented at the International
Conference on Systems, Man and Cybernetics, Hong Kong: IEEE SMC.
21. Lengnick, M.: Agent-based macroeconomics: A baseline model. Journal of Eco-
nomic Behavior and Organization. 86, 102-120 (2013).,
22. Malpezzi, S. (2002). Hedonic pricing models: a selective and applied review. In
Housing Economics. Kenneth Gibb e Anthony O'Sullivan.
23. Miller, J.H., Page, S.E.: Complex adaptive systems. Princeton University Press
(2007).
24. Nadalin, V., Igliori, D.: Empty spaces in the crowd. Residential vacancy in Sao
Paulo's city centre. Urban Studies. 0042098016666498 (2016).
25. Parker, D. C., Manson, S. M., Janssen, M. A., Hoffmann, M. J., & Deadman,
P. (2003). Multi-Agent Systems for the Simulation of Land-Use and Land-Cover
Change: A Review. Annals of the Association of American Geographers, 93(2), 314-
337. https://doi.org/doi:10.1111/1467-8306.9302004
|
1909.07867 | 1 | 1909 | 2019-09-17T14:57:45 | TruPercept: Trust Modelling for Autonomous Vehicle Cooperative Perception from Synthetic Data | [
"cs.MA",
"cs.CV",
"cs.RO"
] | Inter-vehicle communication for autonomous vehicles (AVs) stands to provide significant benefits in terms of perception robustness. We propose a novel approach for AVs to communicate perceptual observations, tempered by trust modelling of peers providing reports. Based on the accuracy of reported object detections as verified locally, communicated messages can be fused to augment perception performance beyond line of sight and at great distance from the ego vehicle. Also presented is a new synthetic dataset which can be used to test cooperative perception. The TruPercept dataset includes unreliable and malicious behaviour scenarios to experiment with some challenges cooperative perception introduces. The TruPercept runtime and evaluation framework allows modular component replacement to facilitate ablation studies as well as the creation of new trust scenarios we are able to show. | cs.MA | cs | TruPercept: Trust Modelling for Autonomous Vehicle Cooperative
Perception from Synthetic Data
Braden Hurl1, Robin Cohen1, Krzysztof Czarnecki1, and Steven Waslander2
9
1
0
2
p
e
S
7
1
]
A
M
.
s
c
[
1
v
7
6
8
7
0
.
9
0
9
1
:
v
i
X
r
a
Abstract -- Inter-vehicle communication for autonomous ve-
hicles (AVs) stands to provide significant benefits in terms of
perception robustness. We propose a novel approach for AVs
to communicate perceptual observations, tempered by trust
modelling of peers providing reports. Based on the accuracy
of reported object detections as verified locally, communicated
messages can be fused to augment perception performance
beyond line of sight and at great distance from the ego vehicle.
Also presented is a new synthetic dataset which can be used to
test cooperative perception. The TruPercept dataset includes
unreliable and malicious behaviour scenarios to experiment
with some challenges cooperative perception introduces. The
TruPercept runtime and evaluation framework allows modular
component replacement to facilitate ablation studies as well as
the creation of new trust scenarios we are able to show.
I. INTRODUCTION
Cooperative Perception, sometimes referred to as collabo-
rative perception or collective perception [1], is a perception
domain which focuses on fusing information from multiple
agents. Cooperative perception is grounded in the idea that
two or more agents can combine information from their
respective viewpoints to increase perceptual range and ac-
curacy and reduce blind spots caused by occlusions from
one perspective.
State-of-the-art cooperative perception methods typically
assume information received is correct. In real-world situa-
tions this assumption may not hold, and malicious attackers
could take advantage of models which do not incorporate
mechanisms to detect and eliminate false information. Trust
modelling is a field of study which aims to identify un-
trustworthy agents by evaluating the consistency of received
information. Trust modelling has been studied for vehicular
ad hoc networks (VANETs) (e.g., [2]) but never, to the best
of our knowledge, with a focus on cooperative perception.
Our aim is to enable the benefits of cooperative perception:
extended perceptual range, perception of occluded objects,
and increased perceptual accuracy.
The motivation behind using cooperative perception to
increase perceptual accuracy, even in situations where the
object is at least partially visible to the ego-vehicle, arises
from two main ideas: 1) Ensemble Methods: aggregate out-
puts from multiple classifiers to produce scores better than
a single classifier [3]. A hypothesis is introduced that using
detections from multiple vehicles could improve detection
accuracy through similar principles of ensemble learning.
1Braden Hurl, Robin Cohen, and Krzysztof Czarnecki are with the
David R. Cheriton School of Computer Science, University of Waterloo.
[bdhurl, rcohen, k2czarne]@uwaterloo.ca
2Steven Waslander is with the Institute for Aerospace Studies, University
of Toronto. [email protected]
2) Viewpoint Diversity: Different viewpoints may increase
detection accuracy of objects in view of the ego-vehicle as
a closer, less obstructed, or non-truncated view of an object
can increase the probability of a true positive detection.
The goal is to increase object detection accuracy, even
of objects that are beyond the typical sensor range, which
will in turn increase vehicle, pedestrian, and cyclist safety.
The solution depends on parties being willing to share
information and must not degrade performance (i.e.,
the
solution must run in real-time).
Existing work in this area is limited, likely due to the
lack of testing data. Current state-of-the-art cooperative
perception methods focus on limited test scenarios (e.g.,
[4]). This makes it difficult to analyze the effect on the
entire perceptual pipeline. Cooperative perception without
conducting a thorough analysis on the potential implications
of malicious agents could lead to danger.
In an endeavor to progress the capabilities and safety of
cooperative perception, we offer two primary contributions:
• A novel method for integrating trust modelling with
cooperative perception into an end-to-end distributed
perception model. Experimental results show an in-
crease of up to 5% Average Precision (AP).
• The TruPercept dataset, which is, to the best of our
knowledge, the first multi-vehicle perspective percep-
tion dataset for AVs gathered in a realistic environment.
The TruPercept dataset is not scenario-based but en-
compasses regular urban driving in a synthetic world.
This will allow researchers to benchmark models for
improving cooperative perception.
These building blocks for effectively understanding trust
modelling with perceptual information are essential for ex-
panding robustness of detection methods and exploiting com-
municated information without leading to overconfidence and
susceptibility to simple attacks.
II. RELATED WORK
Existing state of the art for cooperative vehicles uses
VANETs and various information fusion approaches which
are introduced in this section.
Vehicular ad hoc networks (VANETs) are spontaneous net-
works used for intelligent vehicle communication. VANETs
can be comprised solely of intelligent vehicles (IVs), but
also commonly include road-side units (RSUs) to extend
communication range and a central server for tasks such as
moderating agents or information aggregation and evaluation.
VANETs can be used to transfer information such as traffic
congestion, road conditions, and collision warnings. They
will also be the communication medium for cooperative
perception. Key issues to address include reliable exchange
of messages, even with low latency (e.g., [5], [6]) and
ensuring the availability of networks for IVs and to optimize
RSU locations ([7]). Correa et al. [8] also show how smart
infrastructure can enhance perceptual range. Map merging to
align views, handling localization between vehicles in case
of GPS failures and increasing accuracy have all been ex-
amined to provide richer solutions ([9], [10], [11]). Aligning
detections shared between vehicles has also been promoted
([12]). Latency and alignment issues are set aside for now,
with our particular model.
Feng et al. exhibit, in their summary of modern object
detection and segmentation networks [13], methods which
utilize information fusion at various stages in the detection
network. Typical fusion methods in 3D object detection
(3DOD) fuse image and LiDAR data. Cooperative perception
also fuses information from multiple sources; however, from
a considerably greater quantity as the sources are nearby
vehicles. With cooperative perception the information shared
can range from raw image and point cloud data to final object
detections or tracks (detections over time).
Arnold et al. [14] use images from multiple perspectives
to reconstruct 3D objects and discuss how this could be used
to better detect relevant 3D objects. Chen et al. [15] fuse raw
point cloud data which they argue is better than images since
there is no need for overlap to perform convolutions. Fusing
raw sensor data provides more information to detection
networks at the cost of increased bandwidth requirement and
latency.
Correa et al. [8] say perception performance can be en-
hanced by transporting detection results (late fusion) instead
of raw data. They point to two standardized message trans-
portation protocols for late fusion; the Cooperative Aware-
ness Message (CAM) [16] and the Collective Perception
Message (CPM) [17]. Hobert et al. [18] discuss some of
the limitations and recommendations for CAMs [16]. Obst
et al. [4] combine tracks from multiple vehicles to increase
detection of vehicles that are out of the sensor field of view,
which can facilitate detecting and tracking vehicles that enter
into view.
Unfortunately, there is no standard to compare autonomous
vehicle cooperative perception methods. Chen et al. [15] test
their method by concatenating point clouds from the same
vehicle of a KITTI scene at two different times. Other works
typically use scenario-based test cases and do not include
quantitative metrics for large regular driving situations.
Furthermore, there is no guarantee that information re-
ceived from other vehicles is correct. The cooperative per-
ception methods discussed do not take this into account,
which introduces a huge security flaw. A vehicle could
disseminate false information unintentionally, for example by
using a poorly performing perceptual network, or malicious
information could be spread with the intention of causing
harm. To prevent this, a central server or authority can be
included in VANET system models, often assisting with trust
certificates to authenticate trustworthiness ([19], [20]). Many
trust models include information relayed through vehicles to
RSUs then onto the central server; accuracy of information
can then be determined if authority-provided ground truth is
acquired. Without a central authority (or one unable to obtain
ground truth), a trust modelling component then becomes
paramount to introduce (e.g., [21], [22]).
III. TRUPERCEPT MODEL
Inspired by related work, this section outlines a novel in-
tegrative approach to perception and trust for AV perception.
A. Perception
There are many computer vision models designed for
3DOD in AVs such as [23] and [24], which can be inter-
changed within the TruPercept system. An object instance,
denoted as θ, is detailed with a class (vehicle, pedestrian,
cyclist), 3D position relative to on-board GPS (x, y, z ∈ R3),
3D bounding box dimensions (width, height, length ∈ R3),
heading (rotation around up axis ∈ {−π, π}), and a score.
The set of all possible detections is denoted Θ so that θ ∈ Θ.
The ego-vehicle, the perspective the model is defined from,
is denoted as α and a superscript is used to denote agent
perspective (i.e., Θα are the local detections output by α).
Detections and their evaluations will be broadcast to nearby
vehicles, which can determine to rebroadcast if the message
is within a preset range and time period.
B. Detection Evaluation
Union (IoU) of two 3D bounding boxes.
Judging trustworthiness of reports received from peers is
often calculated by evaluating messages against ground truth
data. For 3DOD ground truth data is true information (θ)
for each object in range, which, if available, would make
perceptual pipelines unnecessary. Instead, each vehicle will
evaluate each detection it receives with the information from
its perspective. The following function definitions are used:
• IoU: Θ × Θ → [0, 1]: Returns the Intersection over
• sv : Θ → [0, 1]: Returns the detection score of a
detection from a vehicle v which attempts to represent
confidence [23] or probability [25] of a true detection.
• ev : Θ → [0, 1] ∪ −1: The evaluation score (from
a vehicle v) of a detection. Evaluation scores should
ideally be probabilistic between 0 and 1, where 0
represents that the detection is incorrect and 1 signifies
the vehicle v is 100% certain the detection is correct.
is already evaluating
whether detections are present or not, within the FoV of
on-board sensors, from a local 3DOD network. It follows
naturally to evaluate received detections on whether or not
they match Θα. Firstly, all received detections are filtered
by local detection area (70 metres 90◦FoV for AVOD [23]
default configuration). Next, received detections are matched
with Θα, producing a set M of sets of matches Θm. A
matching set Θm has a detection from α (if matched),
followed by all matching received detections. Two detections
are considered a match if the IoU is over a threshold param-
eter τ and are the same object class (e.g., car, pedestrian,
The ego-vehicle, denoted as α,
cyclist). Each received detection will be matched to the ego-
vehicle detection with which it has the highest IoU. Only
one detection per received vehicle detection list is matched
with any ego-vehicle detection. Any detection θ which α
receives which matches one of its own detections θ(cid:48) will
have eα equal to sα(θ(cid:48)). If there is no match with θ, α will
set eα(θ) to η, a negative evaluation constant. Experiments
are conducted with η set to either 0 or -1. For each received
detection θ, and the detection of α labelled θ(cid:48), where θ(cid:48) is
the object instance in Θα which has the maximum IoU with
θ, the following applies
eα(θ) =
sα(θ(cid:48))
η
IoU (θ, θ(cid:48)) > τ
otherwise
(1)
(cid:40)
Using only the local detection score to evaluate detections
is not sufficient. Objects which are further away, truncated, or
occluded may have lower or η evaluation scores which would
be erroneous evaluations. To solve this problem, another
value which encompasses the visibility of the detection is
proposed using the following definitions:
• pv : Θ → Z≥: the number of points from the perspec-
tive vehicle v LiDAR point cloud P v which are within
the boundaries of an object θ.
• Φv : Θ → [0, 1]: the visibility value of a detection θ
from a vehicle v. 1 is completely visible and 0 signifies
no knowledge from the current viewpoint.
The true visibility of an object θ from a perspective v is
difficult to calculate and will be approximated by pv(θ). Let
γl and γu be the lower and upper pv(θ) limits for min/max
visibility respectively. Visibility is calculated as
max (0, pv(θ) − γl)
(cid:18)
(cid:19)
Φv(θ) = min
1,
(2)
γu − γl
If there are no points residing in the 3D bounding box then
the object is likely obstructed and the visibility value will
be set to 0 (the vehicle is completely unsure as to whether
the received detection is correct or not). This follows the
assumption that the more LiDAR points that return from
an object, the higher the chance the 3DOD algorithm has
of detecting it. Thus, the visibility score can be used as a
means to estimate the confidence that there is no object for
a negative (i.e., unmatched) detection evaluation.
C. Trust
It can also be beneficial to evaluate the total information
flow from a vehicle. For example, if a vehicle is broadcasting
malicious information, the vehicle should be identified, so
it can be ignored and appropriate measures can be taken.
Trust calculation can be done centrally or by each vehicle.
Central aggregation introduces a strong system requirement
(central server), but is better as vehicles only enter within
proximity of each other for short time periods (≈ 15 seconds
average in the TruPercept data). Trust values are calculated
for each object and then vehicle on the central server and then
periodically broadcast so vehicles may use the information
while integrating detections from other vehicles.
The trust T for a detection θα will be denoted as T (θα)
and will be aggregated using evaluations from all nearby
vehicles V which received the detection θα (V excludes α).
The evaluation from each vehicle v ∈ V will be aggregated
in proportion to how visible the detection is from each
evaluator perspective v (i.e., Φv(θ)). Let the trust function
for a detection from the ego-vehicle θα be T : Θ → [0, 1].
T (θα) =
Φv(θα) · ev(θα)
Φv(θα)
(3)
(cid:80)
v∈V
(cid:80)
v∈V
The trust value for a vehicle v is calculated on a central
server by aggregating the trust feedback from all detections
that vehicle has sent in proportion to the local detection score
for each detection. Up to this point, we have done everything
instantaneously, but trust is calculated over time. Let Θα
t be
Θα from a time step t. Let Bα be the set of all detections
for every t ∈ [−f, 0], where f is the freshness
from Θα
length. Let the trust function for vehicles be T : B → [0, 1].
t
Then the trust for α is
(cid:80)
θ∈Bα
(cid:80)
θ∈Bα
sα(θ) · T (θ)
sα(θ)
(4)
T (Bα) =
D. Detection Aggregation
(cid:80)
θv∈θm
(cid:80)
A final aggregation step (at each vehicle) produces a
score ω for every matching set Θm ∈ M where θmi is
the ith detection θ ∈ Θm. There are several
inputs to
this stage: evaluations ev(θ), visibility Φv(θ), and trust of
evaluator T (Bv). Let the score function for a final detection
be ω : Θm → [0, 1].
1) Weighted Average: A simple aggregation system which
uses a weighted average calculates final scores as
Φv(θv) · T (Bv) · ev(θv)
(5)
ωα(Θm) =
Φv(θv) · T (Bv)
η = 0 for this method so that ω(Θm) ∈ [0, 1].
θv∈Θm
2) Additive (Positive and Negative): The weighted av-
erage method is able to quantify the belief an object is
present relative to the trust and visibility of the object.
However, it does not account for the quantity of vehicles
which perceived the object. It is expected that the higher
the ratio of vehicles which perceive a visible object, the
likelier the object is to exist. The additive aggregation adds to
the final detection score for every vehicle which perceived
the object and subtracts for every vehicle from which the
object was not perceived (unmatched). η = −1 to create
the positive/negative addition/subtraction mechanism. The
aggregation is weighted by the object visibility Φv(θ) and
trustworthiness T v of each evaluator v. The final score is
ωα(Θm) =
Φv(θv) · T (Bv) · ev(θv)
(6)
(cid:88)
θv∈Θm
The resulting value of equation 6 is not bounded
later bounded using ωα(Θm) =
by 0 or 1;
is
it
min (1, max (0, ωα(Θm))). Trust value for the ego-vehicle
should be set to 1, or higher than 1 if bias towards the ego-
vehicle detections is desired.
E. Plausibility Checker
Thus far, the message evaluation relies on visibility of
objects. This could be taken advantage of if a malicious agent
were to insert false detections where there are no points.
For example, false detections are inserted half a meter off
the ground in front of vehicles where there is open air. No
LiDAR points would be registered and the visibility value
would be calculated as zero even though the non-object is
visible. If there are points behind the object, and no points in
front or within the object bounding box, then the existence of
the object is false with high confidence. Obst et al. [4] create
a mechanism for determining false tracks for cooperative
perception and use the term plausibility checker.
The TruPercept system incorporates a novel plausibility
checker which performs a frustum cull on the point cloud.
The frustum is centered to the object center and extends a
quarter of the smallest dimension (width or length) to the
top, bottom, left, and right. If more than 10% of the points
are closer than the object center, it is considered plausible.
The plausibility check can be performed in two key spots:
1) During message evaluation it can provide strong negative
evaluations instead of not contributing to the evaluation due
to low/zero visibility. 2) After aggregating detections to
remove any detections which were erroneously aggregated
to scores higher than zero.
F. TruPercept Summary
The TruPercept system has the following detection and
decision-making cycle on the ego-vehicle α: obtain sensor
data; run the detection network to obtain Θα; broadcast Θα;
receive Θv ∀ v ∈ V or stop at timeout t; calculate M;
calculate and broadcast p(θ), e(θ) ∀ θ ∈ Θv, v ∈ V ; receive
p(θ), e(θ) ∀ θ ∈ Θv, v ∈ V or stop at timeout t; calculate
ω(Θm) ∀ Θm ∈ M.
The central server receives vehicle broadcasts of e(θ) and
updates T (θ) and T (Bv) ∀ θ ∈ Bv, v ∈ V ∪ α which it then
periodically broadcasts.
IV. THE TRUPERCEPT DATASET
Ideally, data would be collected from the real-world,
as it would most accurately represent autonomous driving
conditions; however, collecting data from the real-world for
cooperative perception has resulted only in limited test-case
scenarios. To the best of our knowledge, there are no ex-
isting publicly available real-world or synthetic cooperative
perception datasets for autonomous driving encompassing
regular driving. Creating a real-world dataset would require
a fleet of vehicles equipped with autonomous sensor systems
simultaneously collecting data.
Trust modelling for V2V are often evaluated through the
use of simulation software. Chen et al. [19] use OMNET++
[26] for V2V communication and SUMO [27] for road traffic
simulation. However, these simulators do not contain sensory
level perceptual information. Simulators such as CARLA
[28] are a viable options as they can obtain synthetic percep-
tual data and attempt to simulate traffic flow. Alternatively,
Grand Theft Auto V (GTA V) is a video game which contains
realistic graphics and has been used to generate synthetic
data in a diverse environment for AV perception [29]. Recent
efforts have even created an in-game LiDAR point cloud
generator [30]. Hurl et al. [31] extend the range of [30] and
improve the accuracy of point clouds for pedestrians. The
PreSIL dataset [31] produces the necessary data for 3DOD
for autonomous driving and has been shown to increase
performance on state-of-the-art 3DOD networks.
The TruPercept dataset uses the PreSIL generation method
[31] and consists of two series of captures, spanning ap-
proximately 300 and 400 seconds respectively. The data is
captured at 1 Hz and contains identical data to the PreSIL
dataset for the ego-vehicle and each vehicle within a 100
metre radius. A 100 metre radius is used to restrict the total
number of vehicles. Ideally, 160m will be used as the LiDAR
scanner used for the PreSIL dataset has a max range of 80m
so vehicles must be within double the max LiDAR range
to have overlapping point coverage areas and thus obtain
matching detections. The data is captured over two routes
through urban environments.
V. QUALITATIVE ANALYSIS
A. Scenario 1: Algorithm Analysis Using a Single-Frame
A scenario was constructed to evaluate qualitatively
whether the novel cooperative perception solution can detect
a completely occluded object and filter out a malicious
detection. The scenario contains the ego-vehicle passing by
a parked truck. A pedestrian, occluded from the ego-vehicle
perspective, is walking out from behind the truck attempting
to illegally cross the road. Another vehicle moving in the
opposite direction of the ego-vehicle has a clear line of
sight to the pedestrian. A false detection in a dangerous
location (less than 10m away, directly in its trajectory, and
oncoming) for the ego-vehicle is inserted into the broadcast
of the oncoming vehicle. The false detection is given a high
detection score (1.0) in an attempt to fool the ego-vehicle.
The TruPercept cooperative message evaluation mecha-
nism is able to correctly detect
the pedestrian which is
occluded from the ego-vehicle perspective. Unfortunately
the false detection, although the score is decreased, is still
present. However, if the plausibility checker is also run, it
can eliminate the possibility that the false detection exists
while still maintaining the pedestrian which is occluded.
The image and a top-down view of the LiDAR point cloud,
obtained from the ego-vehicle perspective, is displayed in
Figure 1 along with detections from each car and final
outputs. The false detection is completely removed and the
occluded pedestrian is detected.
B. Scenario 2: Vehicle Trust
This experiment attempts to show that
the TruPercept
model can correctly identify a malicious vehicle. Trust is
accrued over time by the evaluations of peers. To mimic this,
A. Cooperative vs. Local Perception
The first experiment is designed to evaluate if the Tru-
Percept model improves upon single perspective methods.
AVOD [23], a state-of-the-art 3D object detector, is trained
using the PreSIL dataset [31] and the mean AP score on
the TruPercept data is used as a baseline. Next, AVOD is
run for every nearby vehicle (within a 100m radius), and the
TruPercept model is run to obtain updated scores. The scores
of these methods are displayed in Table I. Several variations
of TruPercept are 1) Weighted Average, η = 0 2) Summed
Positive and Negative, η = −1 3) Local detections were
used primarily. Any non-matching TruPercept 2 detection θ
is used if, from the ego-vehicle perspective v, Φv(θ) < 10
and the plausibility check returns true. For shown results
τ = 0.1, γl = 0, γu = 100 for cars, and 40 for pedestrians.
TABLE I
RESULTS: TRUPERCEPT (% AP)
Car
Pedestrian
Easy Mod. Hard All Easy Mod. Hard All
Method
AVOD
66.8 62.8 52.1 36.3 84.2 78.7 75.1 57.3
TruPercept 1 54.6 48.9 43.9 31.5 76.7 68.7 70.0 54.0
TruPercept 2 57.2 51.4 46.0 32.8 78.6 70.4 71.4 55.0
TruPercept 3 64.8 62.1 51.1 36.7 83.4 78.8 75.8 58.5
The most important insight is that none of the tested
methods improve the perceptual accuracy by a meaningful
margin over the local perception methods. This shows the
importance of testing any modules which are added to the
perceptual track with the entire system instead of only sce-
narios. Two plausible explanations for the poor TruPercept
results are: 1) The alignment issues with the synthetic data
could reduce performance for cooperative, but not local, per-
ception. 2) Each perspective can introduce false detections,
when aggregating between multiple vehicles this could result
in more false positives. By investigating the precision-recall
curves for several of the models it is evident that for many
of the poorly performing cooperative perception models the
precision is much lower. This is especially true for lower
recall values, likely caused by false detections from nearby
vehicles being inserted with high scores.
One point to note is that pedestrians have higher detection
scores than cars, a reversal from typical real-world data. This
could be from pedestrian bounding boxes having identical
sizes in the PreSIL data, making regression easier. The
PreSIL data also has larger quantities of vehicles such as
trucks, busses, and SUVs which are not part of the 'Car'
class. This could cause confusion for a detection network
which is only attempting to label cars, resulting in more
false positives. This could be rectified by training a network
to detect all classes.
B. Position and Orientation
The hypothesis is that vehicles which are closer to an
object will provide more accurate positioning and orientation
information. This could lead to increased accuracy if objects
which are detected are not meeting the KITTI standard of
0.5 IoU to be a true positive. For this experiment, the local
Fig. 1. Top: Ego-vehicle image and point cloud of scenario 1. Middle: Ego-
vehicle (left) and oncoming vehicle detections (right). Oncoming vehicle has
inserted a false detection in front of the ego-vehicle. Bottom: Cooperative
perception results without (left) and with (right) the plausibility checker.
Ground truth is shown in green, detection score on boxes.
the single frame from the previous scenario is replicated 100
times and the trust score is examined at various instances in
time. The false detections are inserted after 50 frames so
the vehicle has time to build up trust. A freshness timeout
f = 50 frames is used so that any evaluation further than
50 frames in the past is discarded and not used towards the
vehicle trust. A secondary situation is introduced where the
oncoming vehicle places three false detections to compare
when there is more false information.
The trust for the ego-vehicle remains at 0.82. For the
malicious vehicle, the trust value begins at 0.81 and decreases
to 0.15 by frame 100. When three false vehicles are added
the trust value decreases faster and balances out at zero.
Trustworthy vehicles remain unchanged, thus the TruPercept
algorithm was able to correctly identify the malicious vehi-
cle. Note that there is a third vehicle not depicted in the scene
images which has a clear view of all objects in the scene but
is further away. It contributes to the positive detection of the
occluded pedestrian and has a trust score of 0.97.
VI. QUANTITATIVE EXPERIMENTS
This analysis shows quantifiable effects of using coop-
erative perception on the general perceptual pipeline. The
TruPercept dataset is, to the best of our knowledge, the
first of its kind, and enables cooperative perception model
evaluation on regular driving data. The KITTI evaluation
criteria of easy, moderate, and hard are defined by limiting
values on occlusion, truncation, and minimum height (in
pixels) of objects. This scope is too narrow, as cooperative
perception allows detection of objects which are completely
occluded, truncated, or are far in the distance. To provide a
deeper inspection, the 'All' criteria, which encompasses any
object within the forward facing 90 degree FoV area up to
140 metres away, is also evaluated.
AVOD detections have the position and orientation set to that
of the closest detecting vehicle in each matching detections
list. Results in Table II show an increase in AP of up to 5%. It
is revealing that the largest improvement to local perception
is gained simply by modifying the position and orientation
of local detections. This enables the model to ignore false
positives from nearby vehicles and yet still improve the
accuracy of local detections.
RESULTS: UPDATED POSITION/ORIENTATION (% AP)
TABLE II
Car
Pedestrian
Easy Mod. Hard All Easy Mod. Hard All
Method
AVOD
66.8 62.8 52.1 36.3 84.2 78.7 75.1 57.3
Corrected 69.5 61.0 54.1 37.8 89.0 80.6 80.1 61.0
C. Trust Levels
The previous experiments evaluated detections from hon-
est (although potentially not correct) vehicles. This experi-
ment augments the dataset with malicious and unreliable ve-
hicles to evaluate the performance of the TruPercept models,
while being exposed to entities with non-optimal behaviours.
• Trustworthy: detections are taken from the output of the
local detection network. Results in Table I.
• Unreliable: for each local detection, 10% chance to
remove, 10% add a randomly positioned detection.
• Malicious: a vehicle and pedestrian are inserted in front
of the ego-vehicle for every frame from 10% of vehicles.
Full results are available with the dataset. The car class
shows a significant performance reduction ( 10%) when in-
troducing the unreliable behaviours. Strangely, the pedestrian
class actually improves in all categories for the TruPercept
models. If the unreliable vehicles had bad detections to
begin with,
the final score of these detections could be
reduced. Furthermore, false detections also have a chance
to be removed, these could be harder to identify as false
detections since they will be located where points are, instead
of random locations which could contain no points. The
malicious detections decrease the performance significantly,
even for pedestrians. The malicious detections behaviour
represents a coordinated attack between 10% of the vehicles
specifically targeted towards the ego-vehicle. This shows a
weakness in the model towards coordinated attacks.
true detections, will be able to fool
The mean trust values at the termination of the exper-
iments were: 0.27 trustworthy, 0.25 unreliable, and 0.13
malicious. The trust value for unreliable vehicles is not
significantly different than trustworthy vehicles. It is likely
that behaviours such as the unreliable behaviour, which
mostly present
the
current system. The trust model was able to detect blatantly
malicious behaviour with a much higher success. Average
trust values are fairly low,
likely due to the amount of
false positives which are being forwarded. If local detection
accuracy increases, trust values would be expected to in-
crease. These behaviours are simply a glimpse at strategies
malicious entities could employ. In the future, great effort
should be placed on creating and testing behaviours which
could precipitate dangerous situations.
VII. DISCUSSION AND CONCLUSION
In the future, we can imagine expanding our trust model-
ing, inspired by the work of VANETs researchers, such as
supporting role-based trust modeling (e.g., police cars with
greater initial trustworthiness) (as in [32], [33]) or adding
incentives for honest reporting ([2], [33]). Zhang et al. [32]
also suggest that clustering methods may help to reduce false
information spread. We note that since objects are naturally
geographically clustered, relay control groups come to mind;
by integrating this into the TruPercept solution, information
could perhaps be drastically reduced (e.g., one score per
matching object group). This is an avenue for future work.
Other researchers have new ideas for the use of a central
server to deal with malicious entities (e.g., Li et al.'s [20]
reputation-based announcement system which aggregates re-
ports by reputation score). This may suggest an expanded
role for our central system. More sophisticated trust mod-
elling using Bayesian methods and Partially Observable
Markov Decision Processes (POMDPs) are another area of
study (e.g., Raya's system [34] which allows not knowing
about events, Zhang et al. [19]'s use of partially observable
Markov decision processes to enable dynamic updates). One
way in which we could try for expanded trust modelling
would be considering each new object in our framework
with a POMDP. Allowing more uncertainty about vehicles
is advocated by Balkrishan [35] though their context
is
MANETs, while Souissi et al. [36] suggest more potential
by examining similarity of messages received from different
sources, which may help to determine the risk of believing
a given report. These are other possible directions for richer
trust modelling components.
In conclusion, multiple trust modelling methods for ad
hoc networks were presented, with a majority being used for
VANETs. The topic was evaluated from a new perspective:
integration with 3DOD. A multi-agent solution attempting to
increase 3DOD range and accuracy for AVs was presented.
The TruPercept dataset, including all trust scenarios, as well
as all tools, evaluation, and model code are publicly available
at https://tinyurl.com/y2nwy52o.
Four broader-scoped takeaways are: 1) Cooperative per-
ception can be used to more accurately localize objects.
2) Cooperative perception creates a gargantuan security
flaw for AVs since the correctness of incoming information
cannot be guaranteed. 3) Cooperative perception techniques
need to incorporate trust modelling or incoming information
evaluation. It is fairly easy to concoct situations which could
cause an AV to act erratically and create dangerous situations.
4) Cooperative perception methods should be integrated
into perceptual pipelines then evaluated on large datasets.
Creating a module to prevent a contrived scenario from
occurring is not necessarily hard. Integrating modules into
complex perceptual systems to improve overall performance
and prevent a plethora of prospectively hazardous scenarios
is a more onerous task.
REFERENCES
[1] S. Khan, F. Andert, N. Wojke, J. Schindler, A. Correa, and A. Wi-
jbenga, "Towards collaborative perception for automated vehicles
in heterogeneous traffic," in Advanced Microsystems for Automotive
Applications 2018, J. Dubbert, B. Müller, and G. Meyer, Eds. Cham:
Springer International Publishing, 2019, pp. 31 -- 42.
[2] U. F. Minhas, J. Zhang, T. Tran, and R. Cohen, "Intelligent agents
in mobile vehicular ad-hoc networks: Leveraging trust modeling
based on direct experience with incentives for honesty," in 2010
IEEE/WIC/ACM International Conference on Web Intelligence and
Intelligent Agent Technology, vol. 2, Aug 2010, pp. 243 -- 247.
[3] T. G. Dietterich, "Ensemble methods in machine learning," in Multiple
Classifier Systems. Berlin, Heidelberg: Springer Berlin Heidelberg,
2000, pp. 1 -- 15.
[4] M. Obst, L. Hobert, and P. Reisdorf, "Multi-sensor data fusion for
checking plausibility of v2v communications by vision-based multiple-
tracking," in 2014 IEEE Vehicular Networking Conference
object
(VNC), Dec 2014, pp. 143 -- 150.
[5] Q. Xu, T. Mak, J. Ko, and R. Sengupta, "Vehicle-to-vehicle safety
messaging in dsrc," in Proceedings of the 1st ACM International
Workshop on Vehicular Ad Hoc Networks, ser. VANET '04. New
York, NY, USA: ACM, 2004, pp. 19 -- 28.
[Online]. Available:
http://doi.acm.org/10.1145/1023875.1023879
[6] H. Volos, T. Bando, and K. Konishi, "Reladec: Reliable latency deci-
sion algorithm for connected vehicle applications," in IEEE Intelligent
Vehicles Symposium, 2019.
[7] K. Riedl, S. Kurscheid, A. Noll, J. Betz, and M. Lienkamp, "Road
network coverage models for cloud-based automotive applications:
A case study in the city of munich," in IEEE Intelligent Vehicles
Symposium, 2019.
[8] A. Correa, R. Alms, J. Gozalvez, M. Sepulcre, R. B. Michele Rondi-
none, L. Lücken, and G. Thandavarayan, "Infrastructure support for
cooperative maneuvers in connected and automated driving," in IEEE
Intelligent Vehicles Symposium, 2019.
[9] S. Kim, Z. J. Chong, B. Qin, X. Shen, Z. Cheng, W. Liu, and
M. H. Ang, "Cooperative perception for autonomous vehicle control
on the road: Motivation and experimental results," in 2013 IEEE/RSJ
International Conference on Intelligent Robots and Systems, Nov
2013, pp. 5059 -- 5066.
[10] A. Jiménez-González, J. R. Martínez-de Dios, and A. Ollero, "An
integrated testbed for cooperative perception with heterogeneous
mobile and static sensors," Sensors, vol. 11, no. 12, p. 11516 -- 11543,
Dec 2011. [Online]. Available: http://dx.doi.org/10.3390/s111211516
[11] H. Li, "Cooperative perception : Application in the context of outdoor
l'École Nationale
intelligent vehicle systems," Ph.D. dissertation,
Supérieure des Mines de Paris, Sept 2012.
[12] C. Allig and G. Wanielik, "Alignment of perception information
for cooperative perception," in IEEE Intelligent Vehicles Symposium,
2019.
[13] D. Feng, C. Haase-Schuetz, L. Rosenbaum, H. Hertlein, F. Duffhauss,
C. Glaser, W. Wiesbeck, and K. Dietmayer, "Deep multi-modal
object detection and semantic segmentation for autonomous driving:
Datasets, methods, and challenges," CoRR, vol. abs/1902.07830,
2019. [Online]. Available: http://arxiv.org/abs/1902.07830
[14] E. Arnold, O. Y. Al-Jarrah, M. Dianati, S. Fallah, D. Oxtoby, and
A. Mouzakitis, "Cooperative object classification for driving applica-
tions," in IEEE Intelligent Vehicles Symposium, 2019.
[15] Q. Chen, S. Tang, Q. Yang, and S. Fu, "Cooper: Cooperative
perception for connected autonomous vehicles based on 3d point
clouds," CoRR, vol. abs/1905.05265, 2019.
[Online]. Available:
http://arxiv.org/abs/1905.05265
[16] "Intelligent Transport Systems (ITS); Vehicular Communications; Ba-
sic Set of Applications; Part 2: Specification of Cooperative Awareness
Basic Service," ETSI, Tech. Rep. EN 302 637-2 V1.3.2, Nov 2014.
[17] "Intelligent Transport Systems (ITS); Vehicular Communications; Ba-
sic Set of Applications; Analysis of the Collective Perception Service
(CPS) ," ETSI, Tech. Rep. TR 103 562 0.0.15 Draft, Jan 2019.
[18] L. Hobert, A. Festag, I. Llatser, L. Altomare, F. Visintainer, and
A. Kovacs, "Enhancements of v2x communication in support of
cooperative autonomous driving," IEEE Communications Magazine,
vol. 53, pp. 64 -- 70, 12 2015.
[19] S. Chen, A. A. Irissappane, and J. Zhang, "Pomdp-based decision
making for fast event handling in vanets," in AAAI, 2018.
[20] Q. Li, A. Malip, K. M. Martin, S. Ng, and J. Zhang, "A reputation-
based announcement scheme for vanets," IEEE Transactions on Ve-
hicular Technology, vol. 61, no. 9, pp. 4095 -- 4108, Nov 2012.
[21] C. Chen, J. Zhang, R. Cohen, and P. Ho, "A trust modeling framework
for message propagation and evaluation in vanets," in 2010 2nd
International Conference on Information Technology Convergence and
Services, Aug 2010, pp. 1 -- 8.
[22] Y. Mass and O. Shehory, "Distributed trust
in open multi-agent
the Workshop on Deception, Fraud,
systems," in Proceedings of
and Trust in Agent Societies Held During the Autonomous Agents
Conference: Trust
in Cyber-societies, Integrating the Human and
Artificial Perspectives. London, UK, UK: Springer-Verlag, 2001,
pp. 159 -- 174. [Online]. Available: http://dl.acm.org/citation.cfm?id=
646674.701823
[23] J. Ku, M. Mozifian, J. Lee, A. Harakeh, and S. Waslander, "Joint
3d proposal generation and object detection from view aggregation,"
IROS, 2018.
[24] C. R. Qi, W. Liu, C. Wu, H. Su, and L. J. Guibas, "Frustum
pointnets for 3d object detection from RGB-D data," CoRR, vol.
abs/1711.08488, 2017. [Online]. Available: http://arxiv.org/abs/1711.
08488
[25] A. Harakeh, M. Smart, and S. L. Waslander, "Bayesod: A
bayesian approach for uncertainty estimation in deep object
detectors," CoRR, vol. abs/1903.03838, 2019. [Online]. Available:
http://arxiv.org/abs/1903.03838
[26] A. Varga, "The omnet++ discrete event simulation system," Proc.
ESM'2001, vol. 9, Jan 2001.
[27] M. Behrisch, L. Bieker-Walz, J. Erdmann, and D. Krajzewicz, "Sumo
-- simulation of urban mobility: An overview," Proceedings of SIMUL,
vol. 2011, 10 2011.
[28] A. Dosovitskiy, G. Ros,
an
F. Codevilla, A. López,
"CARLA:
open
abs/1711.03938, 2017.
and
urban
simulator,"
[Online]. Available: http:
driving
V. Koltun,
CoRR, vol.
//arxiv.org/abs/1711.03938
[29] M. Johnson-Roberson, C. Barto, R. Mehta, S. N. Sridhar, and
R. Vasudevan, "Driving in the matrix: Can virtual worlds replace
real world tasks?" CoRR, vol.
human-generated annotations for
abs/1610.01983, 2016. [Online]. Available: http://arxiv.org/abs/1610.
01983
[30] X. Yue, B. Wu, S. A. Seshia, K. Keutzer, and A. L. Sangiovanni-
Vincentelli, "A lidar point cloud generator: from a virtual world to
autonomous driving," CoRR, vol. abs/1804.00103, 2018. [Online].
Available: http://arxiv.org/abs/1804.00103
[31] B. Hurl, K. Czarnecki, and S. Waslander, "Precise synthetic image
and lidar (presil) dataset for autonomous vehicle perception," in IEEE
Intelligent Vehicles Symposium, 2019.
[32] J. Zhang, C. Chen, and R. Cohen, "Trust modeling for message relay
control and local action decision making in vanets," Security and
Communication Networks, vol. 6, no. 1, pp. 1 -- 14, 2012. [Online].
Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/sec.519
[33] U. F. Minhas, J. Zhang, T. Tran, and R. Cohen, "A multifaceted
approach to modeling agent trust for effective communication in the
application of mobile ad hoc vehicular networks," IEEE Transactions
on Systems, Man, and Cybernetics, Part C (Applications and Reviews),
vol. 41, no. 3, pp. 407 -- 420, May 2011.
[34] M. Raya, P. Papadimitratos, V. D. Gligor, and J. . Hubaux, "On data-
centric trust establishment in ephemeral ad hoc networks," in IEEE
INFOCOM 2008 - The 27th Conference on Computer Communica-
tions, April 2008, pp. 1238 -- 1246.
[35] V. Balakrishnan, V. Varadharajan, and U. Tupakula, "Subjective logic
based trust model for mobile ad hoc networks," in Proceedings
of
the 4th International Conference on Security and Privacy
in Communication Netowrks, ser. SecureComm '08. New York,
NY, USA: ACM, 2008, pp. 30:1 -- 30:11.
[Online]. Available:
http://doi.acm.org/10.1145/1460877.1460916
[36] I. Souissi, N. B. Azzouna, and T. Berradia, "Towards a self-adaptive
trust management model for vanets," in Proceedings of
the 14th
International Joint Conference on e-Business and Telecommunications
- Volume 6: SECRYPT, (ICETE 2017), INSTICC. SciTePress, 2017,
pp. 513 -- 518.
|
1807.01014 | 1 | 1807 | 2018-07-03T08:21:07 | A Survey on Agent-based Simulation using Hardware Accelerators | [
"cs.MA",
"cs.DC",
"cs.PF"
] | Due to decelerating gains in single-core CPU performance, computationally expensive simulations are increasingly executed on highly parallel hardware platforms. Agent-based simulations, where simulated entities act with a certain degree of autonomy, frequently provide ample opportunities for parallelisation. Thus, a vast variety of approaches proposed in the literature demonstrated considerable performance gains using hardware platforms such as many-core CPUs and GPUs, merged CPU-GPU chips as well as FPGAs. Typically, a combination of techniques is required to achieve high performance for a given simulation model, putting substantial burden on modellers. To the best of our knowledge, no systematic overview of techniques for agent-based simulations on hardware accelerators has been given in the literature. To close this gap, we provide an overview and categorisation of the literature according to the applied techniques. Since at the current state of research, challenges such as the partitioning of a model for execution on heterogeneous hardware are still a largely manual process, we sketch directions for future research towards automating the hardware mapping and execution. This survey targets modellers seeking an overview of suitable hardware platforms and execution techniques for a specific simulation model, as well as methodology researchers interested in potential research gaps requiring further exploration. | cs.MA | cs | A Survey on Agent-based Simulation using
Hardware Accelerators
Jiajian Xiao1,3, Philipp Andelfinger1,2, David Eckhoff1,3,
Wentong Cai2, and Alois Knoll3
2School of Computer Science and Engineering, Nanyang Technological University
3Department of Informatics, Technische Universität München
1TUMCREATE, Singapore
8
1
0
2
l
u
J
3
]
A
M
.
s
c
[
1
v
4
1
0
1
0
.
7
0
8
1
:
v
i
X
r
a
Abstract
Due to decelerating gains in single-core CPU performance, computationally expensive sim-
ulations are increasingly executed on highly parallel hardware platforms. Agent-based simula-
tions, where simulated entities act with a certain degree of autonomy, frequently provide ample
opportunities for parallelisation. Thus, a vast variety of approaches proposed in the literature
demonstrated considerable performance gains using hardware platforms such as many-core CPUs
and GPUs, merged CPU-GPU chips as well as FPGAs. Typically, a combination of techniques
is required to achieve high performance for a given simulation model, putting substantial burden
on modellers. To the best of our knowledge, no systematic overview of techniques for agent-based
simulations on hardware accelerators has been given in the literature. To close this gap, we pro-
vide an overview and categorization of the literature according to the applied techniques. Since
at the current state of research, challenges such as the partitioning of a model for execution on
heterogeneous hardware are still a largely manual process, we sketch directions for future research
towards automating the hardware mapping and execution. This survey targets modellers seeking
an overview of suitable hardware platforms and execution techniques for a specific simulation
model, as well as methodology researchers interested in potential research gaps requiring further
exploration.
Introduction
1
Since around 2005, it can be observed that due to the breakdown of Dennard scaling, clock frequen-
cies of single CPUs are no longer increasing significantly, even though the transistor counts are still
growing [152]. Instead, CPU manufacturers have more and more focused on developing multi-core
processors. This in turn calls for parallel computing techniques, as programmes (including simu-
lations) that cannot be run in parallel can no longer simply be sped up by incorporating a newer
and faster CPU. Performance can be increased further when the workload of a programme is effi-
ciently distributed to heterogeneous hardware such as Graphics Processing Units (GPUs) or Field
Programmable Gate Arrays (FPGAs) [27].
Some types of hardware are better suited for certain tasks than others, for example, tasks with
large amounts of fine-grained parallelism can benefit greatly from the massively parallel architecture
of modern GPUs with its thousands of cores. Tasks that are largely sequential or characterised
by unpredictable data accesses and control flow lend themselves better to CPUs with out-of-order
execution, long pipelines and large caches. Similarly, if offloading a task to a GPU requires copying
large amounts of data to and from graphics memory, execution on a CPU may be preferable even if
substantial parallelism is available. This issue can be addressed by an Accelerated Processing Unit
(APU), where CPU and an integrated graphics core (of lower performance compared to stand-alone
GPUs) share the same memory. Lastly, compute-intensive and memory-light tasks can be outsourced
to FPGAs which can be programmed to carry out specific computations in hardware.
One field that has always sought after more performance is the field of simulation. Faster com-
puters allow an increase in complexity of the incorporated simulation models, allowing researchers
1
to obtain more accurate results in a faster manner. Agent-based simulations have received broad
attention as they can be employed to study various domains, such as road traffic [39], social net-
works [42], pedestrian movement [156], military [29], biology [4], economics [157] and so on. The
main characteristic of agent-based simulation is that autonomous agents (e.g., individuals or entities)
act and interact to create effects of emergence on the entire system. The complex decision-making of
agents and the huge scale of many simulated systems can lead to enormous runtimes, motivating the
need for employing high-performance computing platforms.
Agent-based simulations are a promising target for parallel computing techniques as agents are
autonomous and in some cases carry out independent computations. In mobility simulations, inter-
actions between agents usually only take place between close-by agents in a somewhat regular 2D
or 3D environment, allowing researchers to employ space partitioning without inducing too much
synchronisation overhead. Moreover, many ABS are time-stepped and agents are often updated at
the same logical time, providing inherent independence and thus potentials for parallelised execution.
Unfortunately, being able to partition a problem and execute it in parallel is not a guarantee that it
can be accelerated using heterogeneous hardware.
To enable ABS on heterogeneous hardware, some general challenges have to be overcome. First,
the simulation has to be partitioned with heterogeneity in mind to decide which part of the program
lends itself best to a specific hardware device, considering the resulting overhead from data transfers
between the different devices. From this it follows that depending on the used hardware, the mapping
of simulation parts to hardware devices will likely be different. Complex simulations typically also
exhibit scattered and unpredictable memory access and control flow as the model state develops
dynamically over time. This further complicates an efficient distribution to heterogeneous hardware.
Lastly, in order to make heterogeneous accelerators available to modellers even without having in-
depth knowledge of the specific hardware platforms, there is a need for frameworks that abstract
away from hardware specifics. Some of the common frameworks provide variants supporting parallel
and distributed execution, e.g., MASON [101], Repast-HPC [32], EcoLab [149], and GridABM [154].
However, these frameworks only support traditional CPU-based environments. Some frameworks
such as FLAME GPU [31] and MCMAS [92] have been proposed that focus on the execution on
specific accelerators such as graphics cards.
In this survey, we structure the complex landscape of agent-based simulation on heterogeneous
hardware. We give an overview of existing types of hardware that have been employed to accelerate
agent-based simulations and discuss past developments and current trends. While some surveys exist
that present generic high-performance computing techniques using heterogeneous hardware [160, 107,
43], we highlight the specific challenges of ABS on heterogeneous hardware and categorize an ample
body of related work along these challenges. For each challenge, we discuss in detail how existing
literature has contributed to solving them. This overview allows us to identify research gaps that
need to be filled in order to establish heterogeneous accelerators in the simulation domain and making
them applicable to a wider range of problems – ideally by providing an automated process to support
the modeller.
The remainder of this survey is structured as follows:
in Section 2, we characterise the main
classes of hardware accelerators for general-purpose computations. Section 3 provides an overview of
agent-based simulation concepts and outlines the computational challenges of executing agent-based
simulations on hardware accelerators. In Section 4, we systematise and survey the existing works
according to the identified challenges and according to the techniques used to do so. In Section 5,
we discuss unresolved challenges and outline how a system tackling these challenges could look like,
thus sketching avenues for future work. Section 6 summarizes our findings and concludes the survey.
2 Hardware Platforms
In this section, we describe the technical characteristics, the benefits as well as the limitations of
hardware platforms that have been used to accelerate agent-based simulations. We focus on many-
core CPUs, Graphics Processing Units (GPUs), Accelerated Processing Units (APUs), and Field-
Programmable Gate Arrays (FPGAs). Readers familiar with these hardware platforms may skip this
section and continue to Section 3.
2
2.1 Many-Core CPUs
Architecture: A many-core (or many integrated core, MIC) CPU contains a group of CPU cores
on a single chip. One of the well-known many-core CPUs, the Intel Xeon Phi, is equipped with
up to 72 x86-compatible CPU cores communicating via an internal Network-on-Chip which enables
fast and parallel data transfer between the cores. A many-core CPU can be connected to the host
machine via PCI-E or can be a standalone CPU with direct access to the system memory. Figure 1
shows an overview of the second generation Intel Xeon Phi 7290F (Knights Landing) processor with
its 72 cores that are grouped into 36 tiles interconnected by a 2D mesh channel. Each tile has 2
cores sharing 1MB of L2 cache. All L2 caches are kept fully coherent by a distributed tag directory.
The processor supports a maximum of 384GB of DDR4 RAM. In addition, 16GB of 3D-stacked
multi-channel DRAM can be used for transparent caching or managed manually.
In the past years, a number of non-x86 many-core CPUs have emerged, such as the Parallella
Board [3], the Epiphany-V [119], and the Kalray MPPA (Massively Parallel Processor Array) [36].
Benefits: A notable advantage of some many-core CPUs over GPUs and FPGAs is their ca-
pability to execute largely unmodified code written for regular CPUs [90]. This makes migration
to these platforms easier, given the code is highly parallelisable. Since the individual cores support
out-of-order execution, employ deep instruction pipelines, and have access to comparatively large
caches, the need to adapt a program's control flow to the hardware is less pressing than with, e.g.,
GPUs [22]. Still, as some many-core processors support vector operations through instruction set
extensions such as AVX-512 [76], a single-instruction, multiple-data (SIMD) style of programming
can extract further parallelism.
Recent work showed that many-core CPUs can substantially accelerate Discrete Event Simulation
(DES) [79, 167]. A number of authors also evaluated the acceleration of various types of simulations
such as fluid dynamics and seismic wave propagation using non-x86 many-cores [131, 26].
Limitations: In light of the comparatively high cost of recent many-core CPUs(≈ US$3368.00 as
of 03/2018 for an Intel Xeon Phi Processor 7290F) compared to other accelerators, the performance
gains compared to traditional multi-core CPUs have frequently been relatively low. Even when
optimising scientific code for a many-core CPU, there may only be a single-digit speedup over an
execution on a traditional multi-core CPU, while in some cases there may even be an increase in
runtime [12]. Further, since the performance depends strongly on parameters such as the number of
threads and on employing the different types of memory available on a many-core CPU, it necessary
to tune these aspects to the given problem and hardware [99].
2.2 Graphics Processing Units (GPUs)
Architecture: GPUs utilise a massively parallel architecture, which makes them considerably more
efficient than general purpose CPUs when large volumes of data can be processed in parallel. Their
original purpose was to accelerate the processing of three-dimensional scenes to be displayed on two-
dimensional screens. However, modern GPUs have evolved to support a wide range of computational
tasks.
The evolution of GPUs (and with that their applicability for simulation) is characterised by three
essential steps. In the 1990s, GPUs followed a fixed-function architecture, which processed a scene's
Figure 1: The tile-based architecture of the Intel Xeon Phi 7290F Processor based on Knights Land-
ing [147].
3
Tile 0Tile 7Tile 1Tile 8............Tile 22Tile 29Tile 23Tile30Tile 5Tile 12Tile 6Tile 13Tile 27Tile 34Tile 28Tile 35............2 coresL1 & L2 CachesRouter Figure 2: A Streaming Multiprocessor (SM) in a GP105 GPU based on Nvidia's Pascal architec-
ture [114].
geometry to produce the colour and transparency values for each of the screen's pixels in a pipelined
fashion. In 2001, Nvidia released the GeForce 3, a new GPU generation which marks the second stage
of GPU evolution. The GeForce 3 included so-called shader units, which execute programs applied
to large numbers of pixel RGBA values or vertices of the objects in a 3D scene.
The flexibility of shader programming made the idea of General-Purpose Computing on Graphics
Processing Units (GPGPU) practical, with early GPGPU work mapping raw data to pixels or vertices
to achieve GPU-based parallel programming. Finally, in 2006, the shader architecture was unified by
no longer distinguishing between vertex and pixel shaders. Now, with the Compute Unified Device
Architecture (CUDA) programming framework, it became possible for GPUs to seamlessly perform
general-purpose computational tasks [133, 117].
We sketch the GPU architecture and programming model on the basis of Nvidia's terminology.
AMD hardware follows a similar design. A modern GPU consists of a scalable number of Streaming
Multiprocessors (SMs), which contain a number of Streaming Processors (SPs) that perform most
of the computations, Special Function Units (SFUs) to efficiently perform special operations such as
executing trigonometric functions, on-chip memory, cache and registers. In addition, modern GPUs
have L2 cache and off-chip RAM, both of which are shared among all SMs [114]. Nvidia's GeForce
GTX1080, for example, has 20 SMs, each of them containing 128 SPs, 32 SFUs, 256KB of registers,
8 texture units, 96KB of low-latency memory, and 48KB of L1 cache Figure 2 . There are 2 MB of
L2 cache and 8 GB of off-chip RAM, referred to as global memory.
CUDA [117] (supporting Nvidia GPUs), OpenACC [120] and OpenCL [150] (the latter two sup-
porting Nvidia, AMD GPUs, and Intel CPUs and integrated GPU) are common programming frame-
works for GPGPU. Both CUDA and OpenCL follow a similar programming model, with some differ-
ences in terminology.
The work to be performed by a GPU program is organized in a hierarchical fashion, aligned with
the properties of the underlying hardware: at the lowest level, there are threads representing a sequen-
tial control flow. On a logical level, all threads execute the same GPU program in parallel. Threads
are grouped into warps of a hardware-specific size (32 threads on Nvidia hardware). Within each
warp, threads execute in lockstep, i.e., if the control flow among threads within a warp diverges, the
different branches are serialised. Thus, although the serialisation is transparent to the programmer,
it is important to minimise intra-warp divergence. A configurable number of warps forms a block.
Warps inside a block have access to a limited amount of low-latency shared memory and can syn-
chronize efficiently. Blocks are assigned to an SM persistently, i.e., the required registers and shared
memory are assigned to the block until all warps have finished execution. Per-SM warp schedulers
dynamically assign runnable warps to the available SPs to minimise stalling on high-latency memory
accesses. Typically, there are many more threads than physical SPs, providing ample opportunities
for this type of memory latency hiding [117].
A key aspect when programming GPUs is the optimisation of memory access patterns. The
GPU hardware prescribes certain rules according to which memory accesses multiple threads can be
coalesced, i.e., executed in aggregate [117]. Generally, the number of physical memory transactions
4
required is minimised when threads with adjacent logical indexes access adjacent locations in memory.
Since many applications require scattered or even unpredictable memory access, achieving memory
coalescing is a common focus of works in parallel programming on GPUs (e.g., [169, 45]).
The recent Nvidia Volta architecture provides individual threads with their own execution context,
enabling a more fine-grained control over the intra-warp control flow [116].
Benefits: The hardware and programming model of GPUs lend itself well to problems that
can be expressed so that large numbers of similar or identical operations are performed on different
data. Commonly, GPUs accelerate fine-grained data-parallel tasks by one to two orders of magnitude
compared to an implementation on multi-core CPUs.
Mature GPU programming frameworks such as CUDA, OpenCL and OpenACC enable relatively
simple programming compared to other accelerators such as FPGAs [44]. Libraries such as Thrust [17]
and CUBLAS [113] supply the programmer with high-performance implementations for common
tasks such as parallel reduction, sorting, and linear algebra operations. Programming frameworks are
available even for more specialised tasks such as agent-based simulation [31].
Beside the performance benefits of GPUs, Richmond and Romano [136] emphasise the oppor-
tunities for efficient visualisation of simulations. Since the agent data is already stored in graphics
memory, visualisation can be achieved easily by passing the agent data to vertex or texture buffers.
Limitations: Most current GPUs are connected to their host CPU via the PCI-E bus. Thus,
the GPU does not have direct access to the host memory. Data transfer between CPU and GPU
is expensive in terms of latency and should therefore be reduced as much as possible. For instance,
according to its specification, a PCI-E 3.0 x16 connection allows an Nvidia Titan X card to transfer
data between host and graphics memory at up to 16 GB/s, while the graphics memory of the card
can achieve a throughput of 336.5 GB/s. However, on recent GPU architectures, interconnects such
as Nvidia's NVLink [115] and AMD's Infinity Fabric [2] achieve throughputs of up to 300 GB/s,
alleviating the impact of data transfers.
High performance on a GPU requires the given task to be expressed in a way that fits the GPU's
hardware properties. The main requirements are a large degree of parallelism and the possibility
to achieve coalesced memory access as well as a common control flow among the threads within a
warp. Thus, memory-intensive tasks with complex data dependencies are typically difficult to execute
efficiently on GPUs [82, 25].
Compared to many-core CPUs, programming for GPUs still requires profound knowledge of the
GPU architecture [168]. As with many-core CPUs, the large number of configurable parameters
render the performance tuning of GPU programs an important but challenging task [158].
2.3 Accelerated Processing Units (APUs)
Architecture: APUs integrate CPU and GPU on a single die. Although the term APUs has been
coined by AMD, recent Intel CPUs with Intel HD Graphics follow a similar architecture. Unlike stand-
alone GPUs, the fused GPU of an APU has direct access to the host memory through a low-latency
and high-bandwidth bus. Figure 3 sketches the high-level architecture of an APU.
Benefits: The main benefit of APUs is the opportunity for zero-copy memory access: since all
memory is accessible both from the CPU and the GPU, costly data transfers over a relatively low-
bandwidth bus like PCI-E can be avoided. Zero-copy memory access also provides memory savings,
as only one copy of an object in memory is required. Further, scattered memory accesses which could
only be handled inefficiently by the GPU can instead be performed by the CPU.
Limitations: Existing APU products have focused more on energy efficiency than high perfor-
mance. They typically contain fewer processing units than stand-alone CPUs and GPUs of the same
hardware generation. For example, the Ryzen 5 2400G APU by AMD has 704 Vega-based stream
processors, while the stand-alone graphics card AMD RX Vega 64 has 4096 stream processors. As a
consequence, compared to high-end stand-alone CPUs and GPUs, their computational power is rela-
tively low. Still, as will be discussed in Section 4, some works have considered APUs for accelerating
agent-based simulations.
5
Figure 3: In an APU, memory is shared between the fused CPU and GPU.
2.4 Field-Programmable Gate Arrays (FPGAs)
Architecture: A Field-Programmable Gate Array is an integrated circuit made of an array of in-
terconnected Configurable Logic Blocks (CLBs). Additionally, FPGAs are equipped with input and
output pads and Digital Signal Processing (DSP) blocks. FPGAs often provide various communica-
tion interfaces such as PCI-E, UART, and Ethernet.
A CLB consists of several slices (sometimes also called logic cells), each slice containing a set
of storage elements and Look-Up Tables (LUTs). A LUT has a number of inputs and outputs as
well as flip flops that store a mapping between possible inputs and outputs. The mapping between
inputs and outputs is defined by the users [63]. The number of slices is one of the most important
benchmarks to determine the computational power of an FPGA and can range from several thousand
to several million. For instance, the XCVU37P Virtex UltraScale FPGA from Xilinx has 2,851,800
slices [170]. In addition, the FPGA may have access to several GB of off-chip DRAM.
To describe the logic to be placed on an FPGA, typically a Hardware Description Language (HDL)
is used. The two most widely used HDLs are VHDL [65] and Verilog [121]. In recent years, there
have been intensive efforts to enable High-Level Synthesis, i.e., to generate FGPA layouts directly
from high-level programming languages such as C, C++, or Java. Recently, Intel released a dedicated
SDK to use OpenCL to program FPGAs [75].
FPGAs are sometimes used as a prototyping tool for the development of Application Specific
Integrated Circuits (ASICs), which requires an extensive and costly design process. To the best of
our knowledge, there is no work so far that employs custom ASICs for ABS. In the field of DES, a
number of works have considered offloading of specific simulation tasks to ASICs [134, 30, 51, 19, 102].
Notably, some of the envisioned components were fabricated physically [19]. Since the works on ASICs
have only limited relevance to the field of ABS, we exclude them from our survey.
Benefits: Due to the flexibility and high energy efficiency of FPGAs, they are frequently used for
computationally intensive and highly parallelisable tasks. For instance, FPGAs can be three orders
of magnitude faster than GPUs when conducting specialised tasks such as encrypting a single 64-bit
block by the Data Encryption Standard (DES) [27]. In contrast to CPUs or GPUs, on which data
paths are fixed, FPGAs provide flexible and customised data paths [132]. In the past years, FPGAs
have received more attention in the field of simulation, particularly in Electronic Design Automation,
since hardware designs can be naturally expressed as FPGA layouts.
Limitations: As with GPUs, FPGAs are connected to a host CPU without direct access to
system memory. The resulting need for data transfers can reduce the potential for performance
gains.
FPGAs are regarded as lacking in programmability when compared to CPUs and GPUs [44, 27].
Although recent efforts towards high-level synthesis alleviate this limitation, manual tuning is still
necessary to achieve the best performance [110, 49].
Finally, FPGAs are configured for a specific task. Since reconfiguration can require multiple
hours [180], FPGAs do not facilitate development processes that require fast iteration. This may
limit the applicability of FPGAs in early phases of simulation model development, where changes to
the simulation model frequently occur and require immediate feedback for evaluation.
6
Physical MemoryIntegratedCPUIntegratedGPUHost MemoryDevice Memory 3 Agent-based Simulation
Agent-based modelling and simulation (ABMS) is a common approach [112] used for evaluating
complex systems in domains such as traffic, crowds, economics, information propagation, biology,
etc. In the following, we characterise the modelling approach and discuss the properties of ABS with
execution on heterogeneous hardware in mind.
3.1 Modelling Approach
In ABS, the simulated entities are agents that perform actions autonomously and interact with other
agents based on certain rules. ABS typically follows a Sense-Think-Act cycle (e.g., [138]): in the Sense
stage, an agent detects and analyses its neighbours as well as the environment in which it resides.
In the Think stage, an agent makes judgement based on the information collected during the Sense
stage. The update of states takes place in the Act stage. The simulation time is typically advanced
in fixed time steps at which all agents update their states. However, if a model requires agents to
update their states at variable points in simulation time, time advancement using a discrete-event
simulation (DES) approach may be more appropriate. In DES, state updates are performed through
events scheduled for execution at discrete points in simulation time. The simulation proceeds by
iteratively executing the earliest remaining event, potentially scheduling new events in the process.
Independent of the time advancement mechanism, the defining characteristic of ABS distinguish-
ing it from other simulation techniques is the autonomy of agents, i.e., "agents are endowed with
behaviours that allow them to make independent decisions" [104]. Since the focus of this survey pa-
per is on ABS, we exclude simulation domains such as physics and chemistry, which usually consider
sets of entities that are passively affected by their environment. However, we do discuss a number
of methods proposed outside of the ABS domain with direct applications to ABS, e.g., GPU-based
priority queues in the context of DES.
3.2 Computational Aspects
When considering models with complex decision-making and behaviour at large scales, ABS can be
computationally intensive. In addition, due to the stochastic nature of ABS, the simulation of a given
scenario is usually repeated multiple times in order to generate meaningful results, further increasing
computational demands [85]. However, the Sense-Think-Act cycle described above provides ample
opportunities for parallel execution. Since the Sense and Think stages are performed on a per-agent
basis and do not modify the simulation state, these stages can be executed in parallel across all agents.
To achieve a consistent view of the simulation state for all agents, the state changes performed in the
Act stage must then be propagated to other processing elements.
When parallelising across multiple traditional CPUs or CPU cores, each processing element can
execute the state updates for a subset of agents. A well-known challenge in parallel and distributed
ABS lies in partitioning the simulation workload among the processing elements. Generally, there are
two dimensions according to which a simulation can be partitioned [109]: domain decomposition par-
titions according to the simulation space (e.g., different roads in a traffic simulation), while functional
decomposition partitions according to different models (e.g., different layers of the network stack in a
computer network simulation). High-quality partitionings are characterised by low amounts of work-
load imbalance and communication among the processing elements. When targeting heterogeneous
hardware environments, the partitioning problem is complicated by the differences in the suitabil-
ity of each hardware device for certain types of computations. Thus, to achieve high performance,
a key challenge is to find a suitable hardware assignment of the simulation tasks according to
characteristics such as the instruction mix and the available degree of parallelism.
Since typically, some communication between the partitions cannot be avoided, techniques for the
minimisation of data transfers are required to reduce the performance impact of the communi-
cation (e.g., [80]).
On CPUs, the ABS performance benefits from long instruction pipelines, large caches and effective
branch prediction. Beyond traditional parallel and distributed simulation, many-core CPUs enable
high degrees of parallelism while supporting unmodified x86 code. The key difference between a
CPU execution in a multi-core and many-core setting is the interconnect through which the CPUs
7
Figure 4: Publications on agent-based simulation on heterogeneous hardware by year and hardware
type.
communicate. Since the architecture of each core still closely follows a traditional CPU core, no
major code adaptations are required to execute the agent update logic efficiently.
In contrast, since both GPUs and FPGAs achieve highest performance with computational prob-
lems of a highly regular structure, another challenge of executing ABS on hardware accelerators lies
in dealing with the scattered memory accesses resulting from the largely unpredictable runtime
behaviour of the simulation. Further, irregular control flows and fine-grained computations make it
challenging to fully utilise high-performance many-core devices. Thus, methods for the maximisa-
tion of parallelism are required. As an example, consider a model where the simulation space is
represented by a rectangular grid of cells, each cell being occupied by at most one agent. Here, a
simple hardware assignment is a one-to-one mapping of arithmetic units to cells. On a GPU, due to
its heritage in highly regular data-parallel tasks on pixel values, such a hardware assignment tends to
enable high cache locality, minimisation of memory transactions, and high utilisation of the arithmetic
units. In fact, prior to the general-purpose programmability of GPUs, a number of works proposed
translating grid-based simulations to operations on graphic textures (e.g., [62]). The Brook language
developed at Stanford [24] automates the translation process to graphics operations. Similarly, there
is a correspondence between the structure of an FPGA and cellular grids [162]. The basic function
of a circuit in an FPGA can be seen as analogous to the function of a cell in a cellular automaton.
However, in many models such as road traffic or social network simulations, the simulation space is
a graph. Graph representations adapted to the architectural properties of the available hardware are
required to efficiently support sensing an agent's neighbourhood and updating the simulation state
while fully exploiting the available hardware. The general trend in the literature is moving towards
supporting increasingly irregular types of simulations on accelerators.
The vast majority of literature on ABS using hardware accelerators has focused on GPUs (see
Figure 4 for an overview of the number of publications since 2002) We identify three reasons for
the popularity of GPUs as accelerators for ABS: first, they are comparatively inexpensive. Second,
in the recent years, the ease of programming of GPUs is slowly approaching that of CPUs. Third,
well-established programming frameworks such as OpenCL enable the formulation of models in a less
hardware-specific manner.
In comparison, the use of FPGAs poses substantial challenges to modellers: only comparatively
costly high-end FPGAs run at clock rates close to GPUs. Thus, enormous degrees of parallelism
are required to match a GPU's performance. Further, while there exist some frameworks enabling
high-level programmability, the achievable performance is limited compared to a more low-level spec-
ification of the desired logic in a hardware description language such as VHDL or Verilog. As with
GPUs, there is a need for libraries and frameworks that provide a higher degree of abstraction from
hardware specifics. Finally, the long runtimes of synthesis steps to generate an FPGA layout make
model development and adaptation a cumbersome process. Nevertheless, some works consider the
use of FPGAs for ABS with promising results [53, 162].
8
051015202002200320042005200620072008200920102011201220132014201520162017Year of PublicationNumber of PapersGPUAPUFPGAMany−Core 4 Addressing the Challenges of Agent-Based Simulation on Ac-
celerators
In the following, we discuss the techniques from the literature applicable to the key challenges in
ABS on accelerators as identified in Section 3: hardware assignment, data transfer overheads,
scattered memory accesses, maximisation of parallelism, and abstraction from hardware
specifics. Table 1 summarises the systematisation of knowledge presented in this survey. It con-
tains our classification of challenges, techniques, and publications, as well as the considered types
of accelerators. For the publications that considered specific simulation models, Table 2 shows the
simulation domains and hardware platforms, providing researchers with pointers to relevant works in
their respective domain.
4.1 Hardware assignment
One of the main challenges in parallel and distributed computations in heterogeneous hardware
environments lies in finding a suitable partitioning, i.e., assignment of a given problem to the available
hardware [50]. We discuss techniques that have been used to address this problem according to two
different, yet interrelated, aspects: first, we consider techniques to select suitable hardware for sub-
tasks according to their ability to efficiently execute certain types of computations. The minimisation
of data transfers among the partitions running on separate devices will be considered in the next
subsection. The existing approaches can be roughly categorized as follows:
1. Static assignment:
if the simulation model involves different types of computations that
clearly suggest a certain hardware mapping, it may be sufficient to partition the model prior to a
simulation run without any adaptation during runtime. For instance, model segments involving
large numbers of independent floating point operations may be well-suited for execution on a
GPU, whereas segments with highly data-dependent control flow suggest the execution on a
CPU.
2. Dynamic assignment: frequently, the dynamic behaviour of a simulated system at runtime
translates to unpredictable computational patterns.
In such cases, maintaining high perfor-
mance may require an adaptation of the hardware mapping based on performance measure-
ments at runtime. An inherent challenge of dynamic assignment is the trade-off between the
performance increase through an improved assignment and the costs of runtime measurements
and re-assignment.
An ample body of research has considered the parallelisation of general programs onto heteroge-
neous platforms, which is an enormous challenge due to the arbitrary control flows and memory access
patterns that can be present in general programs. Thus, typically, the approaches limit themselves
to program portions that are particularly amenable to parallelisation on accelerators. In the case of
Challenge
Hardware
assignment
Technique
Publications
Static assignment by type of computation
Many-Core [90], GPU [72, 125, 6, 15, 21, 172, 148, 68, 106]
[67, 69, 176], APU [163], FPGA [159, 34, 162, 53]
Dynamic assignment based on runtime measurements
GPU [18, 165, 56, 86, 176, 59], FPGA [18]
Data transfer
overheads
Overlapping of communication and computation
Computation replication at partition boundaries
Manual caching in shared memory
Heuristics for agent update order
Representation of irregular data structures
by arrays and grids
Multiple replications in parallel
Window-based event execution
Speculative execution
Computation sorting
GPU [89, 15, 16]
GPU [1, 181]
GPU [135, 181, 96]
GPU [7, 61, 83, 84]
APU [163], GPU [62, 103, 128, 127, 129, 88, 136, 151, 161, 40]
[123, 179, 142, 85, 153, 5, 98, 124, 166, 13], FPGA [132, 108]
GPU [125, 143, 93, 89, 97, 175]
GPU [126, 24, 122, 124, 140, 5, 179, 155]
GPU [95, 98], FPGA [108]
GPU [89, 155, 85]
Scattered
memory accesses
Maximisation of
parallelism
Abstraction from
hardware specifics
Frameworks to support simulation development
Many-Core [92], GPU [135, 137, 92, 103, 71]
Unified memory access
GPU [94, 173, 78, 77]
Table 1: A classification of the challenges in agent-based simulation on accelerators along the relevant
works addressing them.
9
Domain/Hardware Many-Core CPU
Mobility
Biology
Ecology
Social
Network
Physics and Chemistry
[92]
[90]
[127]
GPU
[129] [151]
[143]
[172] [148] [71] [72]
[164]
[7]
[136]
[155]
[40]
[67]
[1]
[69]
[128] [135]
[85]
[137] [96]
[181]
[93] [175]
[83]
[88]
[84]
[142]
[6]
[179]
[161]
[106] [16]
[122]
[124]
[168]
[97]
[95]
[126] [62] [175]
[140]
[155]
[5]
[89]
[21]
APU FPGA
[163]
[159]
[34]
[162]
[53]
[108]
Table 2: Simulation model domains considered in the works covered in the survey.
ABS, constraints such as the separation of data into a per-agent state and the limited sensing range
of agents somewhat simplify the problem of parallelisation, potentially enabling a higher degree of
automation in the hardware mapping. In Section 5, we outline the vision of an automated approach
and the required building blocks towards an automated hardware mapping for heterogeneous ABS.
4.1.1 Static assignment
Several authors have compared approaches to statically assign portions of the simulation workload to
an accelerator. Hirabayashi et al. [72] compare a fully GPU-based execution to a hybrid CPU-GPU
scheme where the CPU controls the progress of the simulation and calls the GPU for specific tasks.
In a traffic simulation based on the Optimal Velocity model [11], the fully GPU-based acceleration
clearly outperforms the hybrid scheme, although the lack of synchronisation operations across blocks
introduces errors into the simulation results.
A similar categorisation is presented by Pavlov and Müller [125], who conclude that a CPU-
GPU approach where both the CPU and the GPU hold duplicated or partial agent and environment
information is the most promising.
Andelfinger et al. [6] compare four GPU/CPU simulator architectures in the context of discrete-
event simulations. In a basic CPU/GPU hybrid scheme (cf. Figure 5a), the CPU offloads each event
to the GPU individually. Input data is transferred to the GPU at the beginning of the cycle. Af-
ter the computation is completed, the output data is transferred back to the CPU. By aggregating
independent events and executing them in parallel in a single step, data transfers are reduced (cf. Fig-
ure 5b). A further reduction in data transfers is achieved by leaving computation results required
by subsequent events in graphics memory (cf. Figure 5c). Finally, if the entire simulation is ported
to the GPU, data transfers are only required at the start of the simulation and once the simulation
terminates (cf. Figure 5d). While the simulation performance increases with each of the above optimi-
sations, more and more changes to the simulator architecture are required, complicating development
and reducing maintainability.
Two approaches for parallelisation are explored, corresponding to hybrid CPU/device (cf. Fig-
ure 5a) and fully device-based simulation (cf. Figure 5d) in [6], respectively. Lai et al. [90] implement
the four geo-spatial applications of Kriging interpolation [144], ISODATA [10], Game of Life [52] and
an urban sprawl simulation using cellular automata [169]. The authors compare the performance
achieved when using one CPU per execution node, one GPU per node and 60 cores per CPU-based
many-core accelerator, using MPI for inter-node communication in each instance. The authors con-
clude that the use of GPUs and CPU-based many-core accelerators both provide a performance
benefit over the purely CPU-based execution. Given a sufficiently large number of assigned proces-
(a) Hybrid CPU/Device.
(b) Event Aggregation.
(c) Memory Reuse.
(d) Fully Device-based Simulation.
Figure 5: Four CPU-Device simulation schemes [6]. Devices in the figure can be GPUs or many-cores.
10
DeviceCPUEvent iEvent i+1Event i+2ProcessingProcessingProcessingDeviceCPUEvent iProcessingEvent i+1Event i+2ProcessingDeviceCPUEvent iProcessingEvent i+1Event i+2ProcessingDeviceCPUEvent iEvent n... sors, using the CPU-based many-core accelerator with fully device-based simulation achieves similar
performance as the GPU-based acceleration.
A number of authors considered hardware assignments tailored to specific simulation models.
For instance, when the underlying simulation can be clearly separated into model computation and
management tasks, a master-worker scheduling approach as shown by Bilel et al. [21] in the context
of large-scale mobile networks simulation can be used. In the proposed design, the model is executed
on the GPU, while the CPU orchestrates the event scheduling, simulation status monitoring, and
memory allocation. A node of the simulated network is partitioned into multiple processes, each
process being executed by one GPU warp.
The nature of traffic simulation allows for a relatively straight-forward static hardware assignment
according to different simulation aspects. Xu et al. [172] and Song et al. [148] introduce a mesoscopic
traffic simulation in a hybrid CPU-GPU architecture, assigning the agent mobility to the GPU,
whereas the route calculation, the agent generation, and file reading and writing remain on the CPU.
The two parts run asynchronously to hide data transfer latencies.
Bauer et al. [14] consider a combined continuous-discrete simulation and assign the continuous
part the GPU and the discrete part to the CPU. The benchmark model PHOLD [46] is employed to
explore different GPU configurations by varying the thread block size, the number of floating point
instructions, the data transfer volume, and the communication pattern. The authors conclude that
while keeping the GPU fully utilised poses a challenge, models in a combined simulation with a large
number of floating point computations can benefit from GPU acceleration.
Taking into account zero-copy memory access, Wang et al. [163] show how a road traffic simulation
can be accelerated using an APU. In their simulator, sorting of agent states is required to locate each
agent's neighbours. While the APU's GPU resources perform state updates and local sorting, the
sorting across GPU blocks is handled by the CPU resources. The work separation can be carried out
efficiently using zero-copy memory accesses.
In the ABS framework TurtleKit, the authors leave the simulation of agent behaviours to the
CPU while environment dynamics are handled by the GPU [106]. With this approach, the authors
aim to reduce the impact of the GPU acceleration facilities on the maintainability of the simulator
code. To increase the performance, portions of the agent behaviour that do not depend on the agent
state, e.g., perception of properties of the environment, are performed on the GPU independently of
individual agents for all locations and time steps.
Considering FPGAs, Tripp et al. [159] showed how the movement of agents on individual lanes
can be computed on the FPGA, while the agents' transitions from one road to another as well as the
behaviour at intersections is computed on the CPU. However, most works on ABS on FPGAs focus on
simulation models that allow for statically assigning the entire simulation to an FPGA. For instance,
the representation of cellular grids on FPGAs is explored by Vourkas and Sirakoulis [162], who
implement an environmental model simulation based on cellular automata (CA). The authors note
the structural similarity between a two-dimensional cellular automaton and an FPGA (cf. Section 3.2).
A lattice of cells is simulated, each Configurable Logic Block (CLB) simulating one cell. In case the
number of cells exceeds the number of CLBs, the simulation lattice is partitioned into several layers,
which are processed one after the other. Similarly, Cui et al. achieve high performance with grid-
based cellular automata on an FPGA [34]. A pipeline comprised of address generation, reading
from memory, data alignment, rule computing, and updating of memory is applied to maximise
throughput. A similar method can be applied to cellular automaton-based crowd evacuation models
as demonstrated by Georgoudas et al. [53].
General guidelines for development of GPU-accelerated ABS starting from a CPU-based simulator
implementation are proposed in [106, 68]. Their methodology requires the decomposition of the
simulation model into small task modules and the heuristic identification of modules suitable for
execution on a GPU. As a heuristic, the authors state that loops and code segments with low amounts
of conditional branching tend to be suited for execution on a GPU. Then, the original task modules are
manually replaced with GPU-executable modules. Several case studies [106, 67, 69] show promising
speedup by deploying this method.
Generally, the approaches relying on static hardware assignment split the simulation workload into
coarse-grained functional tasks so that some tasks are clearly suited for a certain hardware device.
To minimise trial-and-error, heuristics may be applied to identify a suitable mapping of tasks to the
hardware. For instance, in the literature, tasks involving large numbers of parallel floating point
11
operations are among the most common tasks offloaded to accelerators. Further observations are
made by Zhang et al. in the context of co-running programs on a CPU and GPU or in an APU [176]:
1. programs that are suitable to run in a CPU/GPU environment tend to have low memory bandwidth
usage, 2. most programs suitable for the APU allow for a large amount of overlap between CPU and
GPU computations.
4.1.2 Dynamic assignment
While a wide range of literature has considered the problem of dynamically adapting a partitioning of
agent-based simulations to multiple CPUs (e.g., [33, 100, 171]), we are not aware of such works that
specifically target heterogeneous hardware environments. In the following, we outline recent works
on dynamic assignment of general computational workloads to heterogeneous hardware. Since these
works are generic, they cannot rely on knowledge of the general structure of ABS simulators or on
model knowledge.
Belviranli et al. [18] propose a self-scheduling scheme for partitioning generic application workloads
into blocks and assigning them to CPUs, GPUs and FPGAs. The proposed system consists of two
phases:
in the first phase, the system performs an online training with a small amount of data to
estimate the maximum workload capacity size of each hardware device. Fast convergence is achieved
by fitting four sampled data points to a logarithmic function. Once the capacity is determined, the
processing unit's performance can be inferred from the same data. When the change of processing
speed between two samples drop below a threshold, it is used as the final estimated value.
In
the second phase, the remaining workload is partitioned based on the percentage of the individual
processing speed to the total speed of all available processing units, enabling faster processing units
to handle a larger portion of the workload.
Some authors use machine learning techniques such as support vector machines, artificial neural
networks and decision trees to distribute the workload of OpenCL programs to CPUs and GPUs. For
example, Grasso et al. [56, 86] and Zhang et al. [176] translate a single-device OpenCL program to
a multiple-device program, while Wen et al. [165] focus on scheduling multiple OpenCL functions to
run in parallel on CPU/GPU. They train a machine learning algorithm according to a set of typical
OpenCL programs and benchmarks. The prediction generated by the machine learning algorithm
guides the assignment of a portion of the computation to CPU or GPU. Their results show that
the above three machine learning approaches outperform purely CPU- or GPU-based approaches.
The scheduling scheme by Wen et al. achieves a performance improvement compared to a first-come,
first-served scheme and a scheme where computation-heavy task are handled by the GPU.
To automate the compilation of sequential programs for parallelised execution on heterogeneous
hardware, Grosser and Groesslinger [59] present a compiler that generates CPU and GPU code.
Regions with mostly static control flow and sufficient computational intensity are detected and trans-
formed to a formal representation to facilitate program transformations [58]. After optimisations
have been performed to increase memory access locality and parallelism, CUDA code for GPU is
generated from the formal representation. A runtime library eliminates repeated memory allocations
and unnecessary data transfers between CPU and GPU. The decision whether a region is compute-
intensive enough for execution on the GPU is made statically or at runtime using heuristics based
on metrics such as the number of instructions. The authors conclude that the compiler is able to
translate CPU code into cross-platform code with no performance penalty. For some computations,
such as the correlation benchmark from polybench [130], significant speedup of up to two orders of
magnitude can be achieved.
The main difficulty in automated hardware mapping lies in determining the control flow and data
dependencies of the original program. Current approaches either rely on the program code being
formulated in languages such as OpenCL that express independent control flows explicitly, or only
consider specific portions of programs such as loops with largely static control flow. In ABS, however,
most of the available parallelism may exist across the update routines of separate agents. Thus,
without semantic information describing the code structure, automatic detection of the parallelism
is challenging. In Section 5, we sketch how the common structure of many ABS may be utilised to
support the extraction of parallelism.
12
4.2 Minimisation of data transfer overheads
Since most hardware accelerators are equipped with their own memory, simulations making use of
accelerators typically require data transfers between host and accelerator memory. Even with a high-
quality partitioning of the simulation, these data transfers incur an overhead that reduces the speedup
gained from the distributed computation. In this section, we survey works that focus on minimising
the cost of such data transfers. The existing approaches can be roughly categorized according to the
following techniques:
1. Overlapping of communication and computation: since some communication overhead
between the processing elements involved in a simulation cannot be avoided, some authors
proposed techniques to hide communication overheads by transferring data while independent
computations are performed. Sometimes, the technique has been referred to as latency hiding
(e.g., [23]).
2. Computation replication at partition boundaries: another technique to address commu-
nication overheads is to increase the amount of computation performed before synchronisation
between processing elements is required. This is achieved by duplicating some computations
on multiple processing elements, thus delaying the need to resolve data dependencies across
processing elements.
4.2.1 Overlapping of communication and computation
One way of mitigating the overhead from data transfers between the host and an accelerator is to
execute computations at the same time as data is being transferred.
In the approach described
by Kunz et al. [89], event computations are overlapped with data transfers across the CPU-GPU
boundary, thus hiding data transfer latencies in a pipelined fashion. Since events from multiple
simulation instances are considered concurrently, there are substantial opportunities for overlapping
these steps.
Bauer et al. [15, 16] propose a generic API to optimise the data transfer between global memory
and shared memory of CUDA GPUs using so-called warp specialisation. The warps within one coop-
erative thread arrays are split into two groups: Dedicated memory warps are in charge of data transfer
between the on-chip and off-chip memory. Compute warps process the data. The approach improves
performance over thread-level separation between communication and computation since separate
warps can follow divergent control flows without any performance penalty. While their general idea
can be applied to other types of independent processing elements, the warp-based implementation is
specific to GPUs.
4.2.2 Computation replication at partition boundaries
In time-stepped ABS, at model time t each agent updates its state based on the states of its neighbours
at time t−1. If the simulation is distributed across multiple processing elements, synchronisation and
data transfers are required to provide this information at each time step. The associated communica-
tion latencies may make up a substantial portion of the simulation runtime. Thus, some authors have
proposed methods to reduce synchronisation by replicating some computations on multiple processing
elements, similarly to performance optimisations in numerical computing [38].
Aaby et al. [1] present a multi-level data partitioning scheme for cellular simulations on multi-
CPU/GPU clusters. The simulation state is partitioned into blocks and each block is executed by
a thread, a core, or a node, depending on the configured granularity. In contrast to the traditional
data partitioning into blocks of B × B cells and synchronisation at each time step, their approach
partitions the data into several overlapping (B + 2R) × (B + 2R) blocks where ((B + 2R)2 − B2)
cells form the overlapping area (cf. Figure 6). The computation in the overlapping area is performed
redundantly by multiple processing units. Thus, assuming that at each time step, a cell can only
affect its immediate neighbours, R time steps are required for a cell in the inner block to be affected
by cells in another processing element. Therefore, synchronisation is only required every R time
steps. Between synchronisation points, an error propagates inwards within the overlapping areas,
but does not affect the inner B × B cells before a new synchronisation occurs. This partitioning
13
Figure 6: In the partitioning scheme by Aaby et al. [1], R cells are duplicated among neighbouring
processing elements so that each processing element handles (B + 2R) × (B + 2R) cells.
approach is further employed in multi-GPU clusters on the node-, GPU-, block-, and thread-level,
and for multi-CPU clusters at the node-, socket-, core-, and thread-level.
While Aaby et al.'s illustrates the idea based on cellular grids, the approach applies to general
ABS. The sensing range of agents is generally limited and provides an upper bound on the propagation
of the effects of an agent's actions. As long as overlapping segments of the simulation space can be
distributed to the processing elements in a manner so that an effect requires at least R > 1 time
steps, some synchronisations can be avoided. The generality of the approach is illustrated by Zou
et al. [181], who extend the idea of computation replication to graph-based topologies in a GPU-
accelerated epidemic ABS.
4.3 Scattered memory accesses
Throughout the past decades, the increase in computational performance has outpaced the decrease
in memory access latencies, leading to modern hardware designs towards ever-increasing cache sizes
and deep memory hierarchies.
In the context of simulations, the issue of memory access latencies is particularly pressing: typi-
cally, a model's behaviour cannot be predicted before executing the simulation, significantly limiting
the opportunities for a priori optimisation of data access patterns. However, commonalities between
different simulation models can be exploited to propose data structures supporting efficient simulation
of an entire range of models on a specific type of accelerator.
Since dynamic memory allocation on GPUs is costly [48], most GPU-based simulators allocate
graphics memory for the main data structures such as the agent states statically (e.g., [96]). Another
approach is to determine after each simulation step the required amount of memory and perform
allocations accordingly [124].
The existing approaches to address the issue of scattered memory accesses can be roughly cate-
gorized as follows:
1. Manual caching in shared memory: although the support for transparent caching has
improved in recent years, achieving highest performance frequently still requires manual caching
in low-latency segments of an accelerator's memory hierarchy. Since typically the amount of
low-latency memory is small, an iterative approach can be taken to limit the number of accesses
to high-latency memory when accessing large amounts of data.
2. Heuristics for agent update order: since the data dependencies between agent state updates
are typically not known prior to the execution of the simulation, minimising cache misses during
the state updates is non-trivial. Heuristic have been proposed, aiming to favour sequences of
computations acting on the same agent data.
3. Representation of irregular data structures by arrays and grids: the hardware archi-
tecture of GPUs and FPGAs is designed so that highest performance is achieved when acting
on regular data structures such as arrays and grids. Thus, efforts are taken to represent highly
14
Block(0,0)Block(0,1)Block(0,2)Block(1,0)Block(1,1)Block(1,2)Block(2,2)Block(2,1)Block(2,0)RRB irregular data structures in a regular fashion. When covering the techniques from the litera-
ture, we first cover model-specific data structures such as graph representations of a simulated
road network. Subsequently, we discuss works covering two generic building blocks commonly
required as part of ABS engines: priority queues and sorting.
4.3.1 Manual caching in shared memory
Richmond et al. [135] propose to utilise the shared memory of the GPU as a manual cache. In their
agent-based simulation framework for cellular models in biology based on FLAME GPU [31], they
copy sets of messages to be transferred between agents into shared memory. Each thread within
a block can then efficiently iterate through the messages and identify those pertaining to the local
agent. Once all threads have iterated through the messages, the next sets of messages are loaded into
shared memory.
Similarly, Zou et al. [181] implement a manual software cache in shared memory to increase
the performance of their graph-based epidemic simulation on GPU clusters. Before the simulation
commences on the GPU, the CPU sorts the edges of the directed graph by the source vertex. Each
thread block's shared memory stores edges originating from one specific node. Since each block
processes only edges originating from this node, a cache hit rate of at least 50% is ensured.
In agent-based simulation, agents often influence and are influenced by their direct neighbours.
This fact can be exploited when arranging the simulation data in memory, reducing high-latency
memory accesses when updating agents. Li et al. [96] propose such a method for GPU-based ABS:
Assuming a constant number of agents, each agent is assigned to a GPU thread and its state data
is permanently kept in global memory. The simulation space is partitioned into a grid of rectangles.
Once a search for the neighbours within a circle around an agent is required, a search rectangle
that encloses the searching circle is created, so only agents inside the search rectangle have to be
considered. Two approaches to utilise the GPU's shared memory are proposed: in the first approach,
one block manages the searching process for a chunk C of close-by agents. Per-block shared-memory
loads the data of the agent and the agent's neighbours. Each agent in C has a high probability of
being in the other agents' neighbourhoods, so that these agents can frequently be accessed through
the current block's low-latency shared memory. However, since the limited shared memory capacity
allows only for small numbers of agents to be stored, it is still likely that some neighbours are managed
by another block and thus have to be accessed through global memory. In the second approach, the
shared memory loads the data of agents located in the union of all search rectangles of the agents'
handled by the current block. If the shared memory is not sufficient to hold all agents' data, the data
is loaded as a sequence of chunks. Of course, the increase in the search space given by the union of
search rectangles leads to a higher number of unnecessary agent access through shared memory. To
address this problem, the union rectangle can be constructed on the warp level instead of the block
level.
4.3.2 Heuristics for agent update order
The order in which agent updates are performed must adhere to the causal dependencies between
the agent states and behaviours, e.g., in road traffic simulation, vehicles in direct proximity must be
at the same point in simulated time to be able to interact according to the model specification.
Typically, this is achieved by a strictly time-stepped scheme in which agents always reside at the
same time step, after which conflicts in the resulting agent states are resolved [174]. However, since in
a typical simulation not all agents interact at each point in time, some agents may be updated further
into the simulated future than others without affecting the simulation results [7]. Harris and Scheutz
have shown that distributed agent-based simulations can be accelerated by favouring agent updates
that resolve dependencies across multiple processing elements [61]. This way, processing elements
waiting for others to proceed can be unblocked, decreasing the amount of idle time. Their approach
can be applied independently of the underlying hardware platform, but requires bounds on the agent
movement per time step.
Jin et al. [83] present an information propagation simulation supporting execution on HPC systems
and single GPUs and extend it to run on multiple GPUs [84]. Their focus lies on maximising the cache
hit rate when traversing a graph according to rules defined by the simulation models. Two categories of
approaches are developed for the cascade model [54] and the threshold model [55], which both simulate
15
the propagation of information among nodes in a graph: vertex-oriented processing and edge-oriented
processing. For the vertex-oriented approach, the authors further describe two agent update orders:
one iterates starting from active vertices, i.e., those that already have the information, and the other
from inactive vertices. Since the costs depend on the portion of active nodes, the simulation can
switch dynamically between the two vertex-oriented approaches. Finally, the edge-oriented approach
iterates over the connecting edges between two vertices. Since the number of edges is constant over a
simulation run, the cost of the edge-oriented approach is less variable than that of the vertex-oriented
approaches. The authors achieved the highest performance when dynamically switching between the
two vertex-oriented approaches.
4.3.3 Representation of irregular data structures by arrays and grids
GPUs and FPGAs are particularly suited for operations on regularly structured data. However, many
model types specify topologies that are more naturally expressed in terms of irregular structures such
as graphs. Further, execution of the simulator core itself may require operations on irregular data
structures.
A basic optimisation commonly applied in works on GPU-based computing to improve memory
access patterns is the transformation of the data layout in memory from arrays of structures (AoS)
to structures of arrays (SoA) (e.g., [135, 151]). Commonly, sequential programs represent data in an
AoS representation. Since AoS bundles the properties associated with each object in object-oriented
programming, or the states of agents in agent-based simulations, it is a natural way to represent data
within these paradigms. However, with an AoS data layout, parallel operations on the same property
across many objects results in scattered memory accesses. An SoA data layout bundles the same
property across all objects, which can increase cache hits rates and opportunities for memory access
coalescing, thus improving performance substantially.
Beyond this simple optimisation, the data representation can be specialised for a given model to
further improve performance. In the following, we give an overview of methods applicable to ABS to
achieve high performance by translating irregular data structures to a more regular form.
Model-specific data structures
Early works on executing ABS using GPUs frequently focused on cellular grids and translated
the required computations into the graphics processing domain. In a pioneering work done by Harris
et al. [62], GPU shaders are used for implementing computations on the RGBA values in a texture
that holds the agents' states. The same idea is employed by Lysenko et al. [103], Perumalla and
Aaby [128], and Kolb et al. [88].
Perumalla et al. [128] evaluate the performance of running agent-based simulation entirely on a
GPU. They ported the cellular models Mood Diffusion [111, 70], Game of Life [52] and Schelling Seg-
regation [141]. Through the Open Graphics Library (OpenGL), individual agent states are mapped
to pixel colour values. The authors report a speedup of 15 to 40 compared to CPU-based sequential
execution. Kolb et al. [88] develop a particle simulation and a GPU-based collision detection mech-
anism built on the authors' previous work [87]. Similarly, Richmond et al. [136] utilise the GPU's
texture processing ability and map agent states onto texture data. To accelerate the neighbourhood
detection, the simulation space is partitioned dynamically according to the agents' current states.
The algorithm to generate partitions is borrowed from the particle pinning problem in rigid body
particles physics [60, 57]. Identification of the start and end of the partition boundary is performed
similarly to the method described in [118]. Textures are used to represent the agents' states and ver-
tex texture fetching enables the search for the start and end of the partition boundary by comparing
the partition value to the previous agent's state.
To enable traffic simulations on GPUs, Perumalla [127] (and Perumalla et Aaby [129]) proposes
to model the road network as a grid made up of cells. A road network in Cartesian coordinates is
translated to a grid representation overlaying the network: a cell in the grid is marked as occupied
when an edge of the original road network starts in the cell, passes the cell or ends in the cell. In
graphics memory, the cells' properties such as turning probabilities and length are stored in texture
buffers. Simulation is carried out by performing operations on the texture buffers.
A different method for traffic simulation on GPUs is presented by Strippgen and Nagel [151],
who propose a queue-based approach using CUDA. Each road is represented as a single first-in, first-
out (FIFO) queue stored in memory in the form of a ring buffer. With the ring buffer, insertion
16
of a vehicle entering a road and removal of a vehicle exiting a road is achieved with constant time
complexity. Coalesced memory access can be achieved by processing adjacent roads using adjacent
threads. Since the vehicles' mobility is modelled by a fixed per-link velocity, their approach can be
considered mesoscopic. Behaviours such as overtaking or lane-changing are not modelled and would
require random insertions and removals from the ring buffers, which are associated with linear time
complexity.
Other domains in which agent-based simulations have been successfully ported to GPUs using
model-specific data structures include collision detection [161] and a simulation study of tuberculo-
sis [40]. In the former, a grid is split into tiles and data at the boundary of the tiles is replicated
so that a consecutive space is occupied in the global memory of the GPU. In the latter, the authors
propose to use a sorted array according to the liveness status of agents, so that the state of a new
agent can be stored in a memory location previously occupied by one of the dead agents.
Sorting and priority queues
Full or partial sorting are frequently required in agent-based simulations, e.g., for neighbourhood
discovery or to implement priority queues (PQ) if time advancement is performed in a discrete-
event manner. These operations can involve large amounts of data-dependent and scattered memory
accesses and are therefore challenging to implement efficiently on hardware accelerators. Since this
operation can occupy a substantial portion of the simulation runtime [139], a number of works have
focused on memory layouts and algorithms for sorting and priority queues on accelerators.
As building blocks for time advancement in a discrete-event fashion, parallel reduction and bitonic
sorting are commonly used in GPU- and FPGA-based simulation [123, 163, 179, 142, 85]. We discuss
these two operations jointly due to their structural similarities. In both cases, an input array is split
into chunks, each chunk being handled by one thread. At each cycle, the sorted arrays/minimum
values of two threads are then merged to form a new input array. Thus, at each cycle, the number
of chunks and active threads is cut into half. The algorithm iterates until only one thread is active,
leaving a sorted array or the global minimum value, respectively.
He et al. [64] propose a parallel heap-based PQ on GPU based on a previous CPU-based design [37].
The data structure resembles a binary min-heap, but stores r items per heap node. Items are inserted
and extracted in a joint bulk operation that inserts up to k ≤ 2r and extracts up to r elements. At
any time the root node is guaranteed to hold the highest-priority elements, while elements of lower
priority are gradually inserted into deeper levels of the tree over the course of multiple insert-extract
operations. Parallelism can be exploited across the sorting operations on the items within a tree node,
across the nodes on one level of the tree, and by processing all even-numbered and odd-numbered
levels of the tree in parallel. The costs of the queue operations can be hidden by performing them in
parallel with the processing of extracted items.
Similarly, the FPGA-based DES simulators by Rahman et al. [132] relies on a pipelined heap [20]
for storing events. In contrast to the parallel heap by He et al., the pipelined heap is designed to
achieve near-constant access times, but does not provide bulk operations.
A number of works avoid the need for a global PQ holding all future events. Instead, the set of
events is considered jointly in an unsorted fashion [153], split by model segment [142] or simulated
entity [179, 5, 98], split according to a fixed policy [124, 166], or split randomly [108]. To determine
the events that can be executed without violating the simulation correctness, a parallel reduction is
performed to determine the minimum timestamp among the events.
Baudis et al. [13] evaluate the performance of PQs on a GPU implemented as a single parallel
heap or as a set of ring buffers, implicit binary heaps, and splay trees [146] in the context of DES
and path finding on grids. Their results indicate that for up to about 500 elements per PQ, ring
buffers achieve the highest performance. At larger element counts, implicit heaps outperform the
other approaches in their study. Their results suggest that higher performance is achieved by relying
on multiple PQs, one for each agent or set of agents, compared to a single PQ holding all events.
4.4 Maximisation of Parallelism
The limited predictability of how the state of a simulated system evolves over time translates not
only to scattered memory accesses, but also to an irregular control flow, which can negatively affect
performance in two ways: first, variations in the computational intensity among the model segments
may leave some processing elements idle. Second, the single-instruction multiple-thread execution
17
model of GPUs requires divergent operations within a warp to be serialised.
The existing techniques to maximise the parallelism of ABS using accelerators can be roughly
categorized as follows:
1. Multiple replications in parallel: full utilisation of a massively parallel accelerator requires
large numbers of computations that are independent and can thus be executed in parallel. If
a simulation involves a sequence of mostly dependent computations, the overheads for commu-
nication may outweigh the gains from parallelisation. Thus, techniques have been proposed to
perform computations from multiple simulation runs in parallel.
2. Window-based event execution: in simulations involving a discrete-event mechanism, only
a proper subset of the simulated entities may require an update at a certain point in simulation
time. Multiple authors have proposed gathering events across a window in simulated time, and
executing these events in parallel. In effect, this approach forces a discrete-event approach into
a time-stepped execution. A key difference among the techniques lies in whether the simulation
correctness is strictly maintained.
3. Speculative execution: as in general optimistic parallel and distributed simulation [47],
computations may be performed speculatively to improve hardware utilisation. A rollback
mechanism is required to revert to a correct simulation state after erroneous computations.
4. Computation sorting: when assigning neighbouring threads of a GPU to individual agents
or events, divergence occurs if required computations are inhomogeneous. Some authors have
proposed sorting of computations to minimise divergence.
4.4.1 Multiple Replications in Parallel
If an individual simulation run does not provide sufficient parallelism to fully utilise the available
hardware, a Multiple Replications in Parallel (MRIP) approach [125] can be applied, as shown by Shen
et al. [143]: in their approach, multiple replications of a traffic simulation [164] are executed in parallel
on a GPU. Thus, both the parallelism among agents as well as the parallelism across replications
can be exploited. Laville et al. [93] implement a multi-agent simulation of microorganisms in soil for
CPU/GPU in OpenCL. Each GPU thread manages one agent and each block is responsible for one
simulation instance so that multiple simulation instances can run concurrently on one graphics card.
The idea is applied to discrete-event simulations by Kunz et al. [89], focusing on executing parameter
studies comprised of multiple replications on a GPU.
In addition to exploiting the parallelism across replications, Li et al. [97] aim to avoid unnecessary
redundant computations common to multiple replications. They propose a cloning mechanism for
ABS on the GPU: in an ensemble simulation run comprised of multiple simulation instances, the
computations that are common to multiple instances are only performed once. When the behaviour
of an agent diverges between two simulation instances, a clone of the agent is created. Since the
agent may affect other agents, cloning is performed according to the propagation of the effects of the
original change in agent behaviour. Across cloned simulation instances, neighbour detection can be
aggregated to improve the utilisation of the GPU resources. The benefit of cloning is limited when
simulation runs diverge strongly, e.g., across multiple runs of a stochastic simulation using different
seeds for random number generation. Recently, the cloning approach has been applied to large-scale
cellular simulations on GPU clusters [175].
4.4.2 Window-based event execution
On a GPU, all threads in a warp execute the same sequence of instructions on different elements of
data. If no input data is available for some of the threads within a warp, the hardware utilisation is
reduced. In ABS, this issue is particularly obvious when time advancement is performed in a discrete-
event fashion to accommodate varying state update intervals among the agents. Then, events are
scattered along the time axis, i.e., the probability that many events share the same timestamp may
be low. Thus, a simple parallelisation across the events at a certain point in model time may be
insufficient. An approach to address this problem is to execute DES models in a time-stepped fashion:
all events within a certain time interval are executed in parallel. The lower bound of this time interval
18
is usually referred to as Lower Bound on Time Stamp (LBTS), which is similar to Global Virtual
Time in optimistically synchronised parallel and distributed simulation [81]. With a sufficiently large
time step size, hardware utilisation is increased. However, since dependencies between events are not
considered, the simulation results may differ from a sequential execution.
A study comparing the performance of time advancement mechanisms for simulations on the
CPU and the GPU is presented by Perumalla [126]. They study diffusion simulations running in
a time-stepped, discrete-event, and hybrid fashion. The GPU variant is implemented in the GPU
programming language Brook [24]. While the GPU outperforms the CPU in the time-stepped variant,
it does not perform as well as the discrete-event implementation on the CPU. However, high speedup is
achieved using the hybrid approach, where at each cycle, the minimum gap between two events is used
as a time step. The simulation time then advances according to this time step. Fishwick. [122, 124]
present a method for queuing network simulation that executes a DES model in a time-stepped
fashion. The simulation time advances according to a fixed time step size, but skips periods where
no events occur. All events within the current time step are executed in parallel without considering
their potential dependencies. Although the simulation results may be affected by their approach, the
authors show that for a queueing network simulation, error bounds can be given. Other works assume
a minimum time delta between an event and its creation (lookahead) to guarantee the correctness of
the simulation results [140, 5, 179]. If lookahead is available, a window can be determined within which
events are independent, allowing for parallel execution without affecting the simulation correctness.
The current time window is extended dynamically in work by Tang and Yao [155] to allow more
events to be executed in parallel. After executing all events within the current window, their algorithm
evaluates the first event in the event queue with a timestamp larger than the LBTS that can still
safely be executed according to the lookahead.
4.4.3 Speculative execution
To maintain the correctness of the simulation results when executing in parallel on an accelerator, the
simulator must consider the dependencies between state updates. In some of the approaches described
above, a time window is determined where state updates cannot affect each other. If it is difficult
to determine a time window of sufficient size to extract substantial parallelism, a speculative (also
referred to as optimistic) approach can be employed: state updates are performed without regard for
correctness, and rolled back if errors are detected.
The possibility of speculative execution of simulations on FPGAs has been first demonstrated
by Model and Herbordt [108]. They make use of an event predictor, which predicts the interaction
between two particles and generates new events accordingly. Events may later be cancelled as a
consequence of a false prediction.
Targeting GPUs, Li et al. [95] present an execution model that avoids divergent control flow
by speculative event execution. In an initial step, all events that may occur in the simulation are
created. Subsequently, all events are executed in parallel. A scanning process detects and revokes
causally invalid event executions: if an event leaves the simulation in an incorrect state according to
a model-specific criterion, the erroneous event and all events created by it are revoked recursively.
A more general approach for GPU-based discrete-event simulation is presented by Liu and An-
delfinger [98]. An optimistic execution scheme based on the Time Warp algorithm [81] implemented
in CUDA is shown to be beneficial at low event density in simulated time. To support rollbacks
in case of erroneous computations, the authors show how the default random number generator in
CUDA can be reversed computationally without storing additional data.
4.4.4 Computation sorting
On a GPU, threads within a warp following divergent branches of the control flow are serialised. For
instance, if some threads in a warp execute the body of an if statement, whereas others execute
the body of the corresponding else statement, the two sets of threads perform their actions one
after another. Some approaches attempt to arrange the assignment of computations to the available
threads so that branch divergence is minimised.
In their DES engine on the GPU, Tang and Yao [155] sort events by type before execution, i.e.,
by the code associated with the event.
19
The idea is applied to GPU-based execution of multiple simulation instances at the same time
by Kunz et al. [89] (cf. Section 4.4.1). If the simulation instances do not diverge too strongly, many
events of the same type are available across multiple instances, enabling efficient parallel execution.
Kofler et al. apply computation sorting to their ABS of mosquitoes [85]. In their simulator, a
one-to-one mapping between agents and threads is used. Depending on their current state, agents
may perform different operations, which can result in taking different control flow branches during
the state updates. Thus, to reduce divergence among threads within a warp, agents are sorted by
their current state, so that the state updates of adjacent agents share the same control flow.
4.5 Abstraction from hardware specifics
Compared with model development in CPU-based environments, development for accelerators can be
cumbersome and error-prone. To avoid the need for modellers to gain deep expertise in programming
for specific accelerators, several frameworks have been proposed that enable the specification of parts
of the model structure and behaviour in a hardware-agnostic fashion. The approaches to avoid the
need for modellers to consider low-level aspects of accelerators can be classified as follows:
1. Frameworks to support simulation development: some authors have proposed generating
partial model code to be executed on accelerators from domain-specific languages or the reliance
on a library of pre-defined implementations of common simulation tasks and models. However,
in these approaches, developing a full ABS will typically still require manual implementation
work using a comparatively low-level languages such as CUDA. Further, workload partitioning
and assignment to different hardware devices is currently not considered by these approaches.
2. Unified memory access: since in most cases, the CPU and hardware accelerators involved in
a simulation operate on separate memory, resolving data dependencies may involve cumbersome
explicit data transfers. A number of authors have proposed techniques to transparently access
data in programs executed on heterogeneous hardware.
4.5.1 Frameworks to support simulation development
In the Flexible Large Scale Agent Modelling Environment (FLAME GPU) [135, 137], agent states
are specified using the state machine model X-Machine [41, 74]. Modellers define agent states in
an XML-based format, while state transitions, i.e., the code segments describing the state updates,
have to be manually specified as CUDA code. Generic facilities for exchanging messages between
agents are provided by the framework. Use cases of the FLAME GPU framework can be, e.g., traffic
simulation [71].
Another framework called Many-Core Multi-Agent System (MCMAS) for GPU and other many-
core architectures is introduced in [92]. The framework provides a high-level Java interface to OpenCL
code as well as a set of pre-defined data structures and functions called plugins. To implement agent
models, users either rely on plugins or define their own plugins as OpenCL code that can be called
from Java code. The authors state that unlike FLAME GPU, in which models are targeted exclusively
to the framework, the models defined in MCMAS can be reused by other agent-based simulators.
While FLAME and MCMAS both reduce the implementation work required to develop agent-
based simulations targeting accelerators, these frameworks do not provide guidance or automation
in distributing the simulation workload to the available hardware. Thus, manual experimentation is
required to determine a suitable hardware mapping.
4.5.2 Unified memory access
GPGPU frameworks such as OpenCL or CUDA require the user to either explicitly trigger data trans-
fers between host and device memory, to explicitly select certain variables or memory regions for access
from both CPU and GPU code [117], or to annotate the program to manage data transfers [173, 94].
These manual steps complicate the development of agent-based simulations in heterogeneous envi-
ronments. Some works aim to improve on this situation by transparently transferring required data
between host and graphics memory. However, in languages based on C or C++, static alias analysis,
i.e., determining which pointers refer to the same memory regions, is known to be undecidable [78].
20
Jablin et al. [78, 77] presented the first fully automated data management system based on com-
pilation steps and a runtime library. The developer formulates his program and GPU code as if
all data resides in host memory and can be accessed both from the CPU and GPU. The proposed
approach instruments the code to track accesses to different memory regions using code instrumen-
tation and trapping of system calls. To avoid the need for static pointer analysis, memory accesses
through pointers are tracked by the runtime library.
In addition to transparently handling data
transfers, CPU-GPU communication is optimised during compile time by re-ordering the program
flow to reduce the alternation between computations and data transfers. Unnecessary data transfers
are avoided by leaving data in the GPU memory until it is accessed from the host.
While the work of Jablin et al. could be applied to automate data transfers in heterogeneous
ABS, the detection of parallelism is not covered. In Section 5, we sketch research directions towards
automation in porting ABS to accelerators.
5 Towards an automated offloading procedure
From the observations in the previous section, we can state that there is a vast range of techniques
covering the main challenges of high-performance ABS on hardware accelerators. However, there
exist only few ABS frameworks that support such accelerators. Since existing agent-based simulation
and model implementations typically target purely CPU-based environments, there is a clear need
for processes and tools to support the transition to an execution on accelerators. More specifically,
modellers and simulationists should be supported in the parallelisation and hardware mapping as
much as possible. While methodologies have been proposed to systematise the steps of porting a
simulation to a GPU [106, 69], there is still a lack of automated tools to support this process.
The problem of automatic parallelisation of general programs is a broad and active field of re-
search [48]. Substantial successes have been achieved with respect to parallelisation of computation-
ally intensive loops with predictable and mostly static control flow [59], whereas the extraction of par-
allelism across complex and irregular programs is still a largely manual process. Common approaches
include specifying software systems using formalisms that express parallelism explicitly [73, 91, 105]
or annotating programs with parallelisation hints [35]. In essence, these approaches provide the com-
piler or parallelisation middleware with a dependency graph of the statements or code blocks within
the original program.
Fortunately, many agent-based simulators and models roughly follow a common set of properties
that simplify the extraction of parallelism. We identify the following constraints that can be leveraged
to support the parallelisation process:
1. Time-stepped execution: usually, the model time is advanced in fixed increments. At each time
step, all agents update their states.
2. Two states per agent: to decouple the simulation results from ordering agent updates, simulators
commonly support storing each agent's old state at t − 1 and the new state at t separately.
During an update from t− 1 to t, only read accesses are performed to the agents' states and the
environment state at t − 1, and only write accesses to the states at t. Thus, within an update,
there are no read-after-write dependencies across agents.
3. Sense-Think-Act cycle: we assume that agent updates follow the well-known Sense-Think-Act
cycle (cf. Sec. 3.1), with one such cycle per model.
With these constraints, a natural approach to parallelisation is to offload individual stages of a
model's Sense-Think-Act cycle to an accelerator. For instance, in crowd simulations using the social
force model, the Think stage comprised of the computation of the force affecting an agent may be
performed by one thread of a GPU per agent.
In the following, we sketch an envisioned workflow and the required tools to support users in
porting an existing CPU-based ABS to a system equipped with hardware accelerators. For the
targeted simulator architecture, we assume a traditional master-worker scheme, with the host CPU
acting as the master and assigning work to the available accelerators at each time step.
21
Figure 7: Workflow of the envisioned automated offloading procedure.
Coord position;
void executeOnTimeStep():
Algorithm 1 Example for model code annotated with the stages of an agent update.
1: #pragma agent_begin
2: class Agent:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14: #pragma agent_end
#pragma sense_begin
List agents = getNeighbouringAgents(position);
#pragma sense_end
#pragma think_begin
Coord velocity = computeVelocity(agents);
#pragma think_end
#pragma act_begin
position = position + velocity;
#pragma act_end
5.1 Proposed Work Flow
The proposed semi-automated process is visualised in Figure 7. To facilitate the automatic partition-
ing of the simulation source code into segments that can be outsourced to various types of hardware,
we suggest manually annotating the source code according to the Sense-Think-Act paradigm. From
that it follows that the smallest unit that can be offloaded to a hardware accelerator in our proposed
framework is one of these three stages. Each of the stages is profiled in terms of memory and com-
putational requirements. According to the gathered requirements, an optimisation problem is solved
to generate a hardware assignment (rightmost part of Figure 7).
For simplicity, we assume that all data required by the stage fits into one of the accelerator's
memory entirely. Otherwise, agents could be distributed across multiple accelerators or processed in
batches, both implying additional communication costs.
Input
5.1.1
The source code is annotated manually to signify the stages of the Sense-Think-Act cycle, e.g., in
the form #pragma sense_begin, #pragma sense_end, and so forth. A simple example for a crowd
simulation is given in Algorithm 1. In addition to the manual annotations, this clear separation may
require refactoring of the simulation code. By parsing the annotated source code, the framework
obtains a mapping between code and stages that will later be enriched with data from measurements.
The second input is a specification of the available hardware. Each hardware device is charac-
terised by its available memory, computational performance, and host-device data transfer overhead.
The computational performance can be stated in terms of single-threaded performance on CPUs,
many-core CPUs, GPUs, and APUs. We assume that for an FPGA, only model stages for which
implementations already exist are eligible for offloading. Thus, the computational performance of an
FPGA is given with respect to specific model stages.
22
ComputationProfiling OptimisationProblemFPGAGPUCPUAvailable HardwareAnnotatedSOURCECODESenseCPUThinkFPGAActCPU Simulation model NSenseCPUThinkGPUActGPU Simulation model 1...MemoryProfiling 5.1.2 Memory access profiling
Now that the source code is partitioned into offloadable stages and the capabilities of all the hardware
components are known, the data dependencies of each stage are determined. Assuming a node in a
graph represents one stage, then an edge in this graph represents a data dependency between these
stages. The dependency can refer to both agent or environment data. The weight of the edge is the
volume of the data that is accessed in the CPU-based simulator, i.e., that has to be transferred during
offloading. Usually, the Think stage only has a dependency on the Sense stage within the same model
and agent (intra-agent dependency), whereas the Sense stage might depend on the environment and
on other agents' states (inter-agent dependency). Although we assume that an individual stage is not
partitioned across multiple hardware devices, the amount of data gathered during the sense stage may
vary over the course of the simulation. For instance, if agents form clusters in the simulation space,
the number of neighbours per agent may increase over time. Thus, the data dependencies should be
measured with respect to typical scenario conditions. To avoid exceeding the memory capacity of one
of the considered hardware devices, the profiling can be repeated for a worst-case scenario.
Tools exist that are able to ascribe memory accesses performed during a program run to the source
functions, data structures or threads [8]. For instance, the tool PinComm constructs a dynamic data
flow graph from instrumented program executions [66]. The annotations shown in Algorithm 1 allow
us to map function names to the separate agent update stages. Thus, it is possible to obtain the
amount of memory accessed within each stage. Once the graph describing the amount of memory
accesses across stages is created, the implications in terms of memory copying of moving a certain
stage to a hardware device can directly be evaluated. For example, if the Think stage is moved to
the GPU and the Sense and Act stage remains on the host CPU, then the edges entering and leaving
the Think node determine the data transfer overhead. The actual cost of this copy procedure can be
obtained from the device specification or through measurements.
5.1.3 Computational profiling
In addition to the memory requirements of each stage, information about the computational char-
acteristics of each stage is required. The estimated runtime could be inferred from hardware perfor-
mance models [178, 145, 28, 9]. Approaches as those described in Section 4.1.2 can be applied to
estimate the suitability of different agent update stages for execution on a certain accelerator. By
characterising the workload incurred by each stage in terms of instruction mix and memory accesses
as well as the number of agents, the performance of executing the full-scale simulation can be esti-
mated [56, 86, 176, 165]. Alternatively, if the runtime of a stage is dominated by a sub-task that can
easily be ported to an accelerator, measurements with respect to this task can be performed directly
on the accelerator [18].
5.1.4 Optimisation problem
Building on the graph that represents data dependencies, an optimisation problem of assigning stages
to hardware types can be formulated, similar to the approach targeting embedded systems by Zhang
et al. [177]. In essence, constraints are formulated so that each stage is assigned to the host or a device,
resulting in an overall simulation schedule. Importantly, the optimisation problem must reflect the
data location after each stage or time step (e.g., [105]). For instance, to avoid data transfers, it may
be more efficient to execute two subsequent stages on the same accelerator. The objective function
of the optimisation problem is the overall runtime, i.e., the sum of all estimated execution times on
the respective device and the incurred communication costs by distributing nodes of the dependency
graph that are connected by an edge.
5.1.5 Output
The output of the optimisation steps is a recommendation of which stages should be executed on
which hardware device. It is then the task of the user to port the code of each stage so it can be
executed on the assigned device. This might require specific knowledge, e.g., programming in VHDL
or OpenCL and can therefore be an obstacle to some researchers. Given that some established
simulation models are used by many researchers (e.g., a CSMA/CA model in network simulation
23
or different car-following models in traffic simulation), a public repository of common simulation
models could be created, similarly to the plugin approach used in MCMAS [92]. Researchers could
download these crowd-sourced simulation models to enable parts of their own simulations to be run
on heterogeneous hardware environments, and contribute their own model implementations. Such
a repository would also reduce the need to estimate execution times and improve the optimisation
results by allowing direct measurements on the potential target devices. Similarly, after porting a
specific model stage, new measurements may be performed to provide the optimisation process with
more accurate performance data.
5.1.6 Discussion
In our approach, we take a pragmatic perspective: while the envisioned workflow is achievable based
on existing building blocks, our assumptions may leave substantial performance potentials unexplored.
In particular, by assuming that models and their stages are both executed as a series of dependent
steps, we only exploit the inter-agent parallelism within each stage, while any parallelism across stages
is not considered. In the following, we revisit the key challenges of ABS using hardware accelerators
and sketch techniques from the literature that could be applied to maximise the performance benefits
given our assumptions.
The hardware assignment (cf. Section 4.1) is the main focus of the proposed work flow. Above,
we describe a static assignment using a functional decomposition. Still, the optimisation problem
that determines the hardware mapping could be updated according to runtime measurements.
To minimise data transfer overheads that cannot be avoided (cf. Section 4.2), a bulk execution
of multiple simulation runs would be feasible. The optimisation problem could be adapted so that
the computational and memory requirements reflect those of each stage executed within multiple
simulations runs at the same time. The output of the optimisation process would then be a schedule
for an execution in a multiple replications in parallel (MRIP) fashion [125, 89] .
The technique of overlapping computations with data transfers seems challenging in our approach,
since we assume a serialisation of the agent update stages. However, pre-fetching across stages may
be performed by commencing data transfers once some agents have finished a stage.
Scattered memory accesses and the maximisation of parallelism (cf. Sections 4.3 and 4.4)
could be addressed by providing a library of optimised functions and data structures for operations
such as inter-agent communication or neighbour search (e.g., [31, 92]).
A certain degree of abstraction from hardware specifics (cf. Section 4.5) is achieved by the
automated profiling and hardware mapping of our proposed workflow. Since each stage is executed on
a single accelerator, facilities for unified memory access across all devices are not required. Instead,
all agent data is updated locally on the accelerator and transferred automatically according to the
schedule determined in the optimisation process.
Overall, the envisioned workflow is intended to rely on existing tools and techniques to allow
researchers to exploit the hardware at their disposal with reasonable performance gains, while avoiding
the need for costly and time-consuming manual optimisation steps as much as possible.
6 Conclusions
We presented a survey of the literature on agent-based simulation using hardware accelerators. We
categorized existing approaches according to the key challenges of hardware assignment, minimisa-
tion of data transfer overheads, scattered memory accesses, maximisation of parallelism, and the
abstraction from hardware specifics. Our survey provides modellers with an overview of techniques
to execute a certain class of models on the available hardware. Methodology researchers are given a
summary of the existing work, pointing out research gaps where further exploration is required. Our
main observations are two-fold: first, most of the literature in the past years has focused on GPUs.
We expect a significant amount of work exploring agent-based simulations on FPGAs to appear in
the near future. Second, while a vast amount of work has proposed techniques that allow for efficient
execution of agent-based simulations, only a small number of techniques has found their way into a
unified framework. Thus, the burden of developing a simulation that is executable in a heterogeneous
environment is carried by the modeller. Aiming to reduce the need for expertise in the programming
24
for accelerators, we sketched our vision of a framework to perform an automated hardware mapping
and performance optimisation based on building blocks from the literature.
Acknowledgement
This work was financially supported by the Singapore National Research Foundation under its Cam-
pus for Research Excellence And Technological Enterprise (CREATE) programme.
References
[1] Brandon G. Aaby, Kalyan S. Perumalla, and Sudip K. Seal. Efficient Simulation of Agent-Based
Models on Multi-GPU and Multi-Core Clusters. In Proceedings of the International Conference
on Simulation Tools and Techniques (SIMUTools '10), pages 29:1–29:10, Torremolinos, Malaga,
Spain, March 2010. ICST.
[2] Advanced Micro Devices, Inc. Radeon's next-generation Vega architecture. Technical Report
061317_FINAL_V2, Radeon Technologies Group, June 2017.
[3] Spiros N. Agathos, Alexandros Papadogiannakis, and Vassilios V. Dimakopoulos. Targeting the
parallella. In Proceedings of the European Conference on Parallel Processing (Euro-Par '15),
pages 662–674, Vienna, Austria, August 2015. Springer.
[4] Gary An, Qi Mi, Joyeeta Dutta-Moscato, and Yoram Vodovotz. Agent-Based Models in Trans-
lational Systems Biology. Wiley Interdisciplinary Reviews: Systems Biology and Medicine,
1(2):159–171, September 2009.
[5] Philipp Andelfinger and Hannes Hartenstein. Exploiting the Parallelism of Large-Scale
Application-Layer Networks by Adaptive GPU-Based Simulation. In Proceedings of the Winter
Simulation Conference (WSC '14), pages 3471–3482, Savannah, GA, USA, December 2014.
IEEE.
[6] Philipp Andelfinger, Jens Mittag, and Hannes Hartenstein. GPU-Based Architectures and Their
Benefit for Accurate and Efficient Wireless Network Simulations. In Proceedings of the Interna-
tional Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication
Systems (MASCOTS '11), pages 421–424, Singapore, July 2011. IEEE.
[7] Philipp Andelfinger, Yadong Xu, David Eckhoff, Wentong Cai, and Alois Knoll. Fast-
Forwarding Agent States to Accelerate Microscopic Traffic Simulations. In Proceedings of the
ACM SIGSIM Conference on Principles of Advanced Discrete Simulation (PADS '18), pages
113–124, Rome, Italy, May 2018. ACM.
[8] Imran Ashraf, Mottaqiallah Taouil, and Koen Bertels. Memory Profiling for Intra-Application
Data-Communication Quantification: A Survey. In Proceedings of the International Design &
Test Symposium (IDT '15), pages 32–37, Dead Sea, Jordan, December 2015. IEEE.
[9] Sara S. Baghsorkhi, Matthieu Delahaye, Sanjay J. Patel, William D. Gropp, and Wen-mei W.
Hwu. An adaptive performance modeling tool for GPU architectures. In Proceedings of the
ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (PPoPP '10),
pages 105–114, Bangalore, India, January 2010. ACM.
[10] Geoffrey H. Ball and David J. Hall. ISODATA, a Novel Method of Data Analysis and Pattern
Classification. Technical Report AD0699616, Stanford Research Institute, Menlo Park, CA,
April 1965.
[11] Masako Bando, Katsuya Hasebe, Ken Nakanishi, Akihiro Nakayama, Akihiro Shibata, and Y¯uki
Sugiyama. Phenomenological Study of Dynamical Model of Traffic Flow. Journal de Physique
I, 5(11):1389–1399, November 1995.
25
[12] Taylor Barnes, Brandon Cook, Jack Deslippe, Douglas Doerfler, Brian Friesen, Yun (Helen)
He, Thorsten Kurth, Tuomas Koskela, Mathieu Lobet, Tareq Malas, Leonid Oliker, Andrey
Ovsyannikov, Abhinav Sarje, Jean-Luc Vay, Henri Vincenti, Samuel Williams, Pierre Carrier,
Nathan Wichmann, Marcus Wagner, Paul Kent, Christopher Kerr, and John Dennis. Evaluating
and Optimizing the NERSC Workload on Knights Landing. In Proceedings of the International
Workshop on Performance Modeling, Benchmarking and Simulation of High Performance Com-
puting Systems (PMBS '16), pages 43–53, Salt Lake City, UT, USA, November 2016. IEEE.
[13] Nikolai Baudis, Florian Jacob, and Philipp Andelfinger. Performance Evaluation of Priority
Queues for Fine-Grained Parallel Tasks on GPUs. In Proceedings of the International Sym-
posium on Modeling, Analysis, and Simulation of Computer and Telecommunication Systems
(MASCOTS '17), pages 1–11, Banff, Canada, September 2017. IEEE.
[14] David W. Bauer, Matthew McMahon, and Ernest H. Page. An Approach for the Effective
Utilization of GP-GPUs in Parallel Combined Simulation. In Proceedings of the Conference on
Winter Simulation (WSC '08 ), pages 695–702, Miami, FL, USA, December 2008. IEEE.
[15] Michael Bauer, Henry Cook, and Brucek Khailany. CudaDMA: Optimizing GPU Memory
Bandwidth via Warp Specialization. In Proceedings of the International Conference for High
Performance Computing, Networking, Storage and Analysis (SC '11), pages 12:1–12:11, Seattle,
DC, USA, November 2011. ACM.
[16] Michael Bauer, Sean Treichler, and Alex Aiken. Singe: Leveraging Warp Specialization for High
Performance on GPUs. In Proceedings of the ACM SIGPLAN Symposium on Principles and
Practice of Parallel Programming (PPoPP '14), pages 119–130, Orlando, FL, USA, February
2014. ACM.
[17] Nathan Bell and Jared Hoberock. Thrust: A Productivity-Oriented Library for CUDA.
GPU Computing Gems Jade Edition, pages 359–371. Elsevier, 2011.
In
[18] Mehmet E. Belviranli, Laxmi N. Bhuyan, and Rajiv Gupta. A Dynamic Self-scheduling Scheme
for Heterogeneous Multiprocessor Architectures. ACM Transactions on Architecture and Code
Optimization - Special Issue on High-Performance Embedded Architectures and Compilers,
9(4):57:1–57:20, January 2013.
[19] Jacob L. Berlin. Design of a Parallel Discrete Event Simulation Coprocessor. Master's thesis,
December 1993.
[20] Ranjita Bhagwan and Bill Lin. Fast and Scalable Priority Queue Architecture for High-Speed
Network Switches. In Proceedings of the IEEE INFOCOM Conference on Computer Commu-
nications (INFOCOM '00), pages 538–547, Tel Aviv, Israel, March 2000. IEEE.
[21] Ben Romdhanne Bilel, Nikaein Navid, and Mohamed Said Mosli Bouksiaa. Hybrid CPU-GPU
Distributed Framework for Large Scale Mobile Networks Simulation.
In Proceedings of the
IEEE/ACM International Symposium on Distributed Simulation and Real Time Applications
(DS-RT '12), pages 44–53, Dublin, Ireland, October 2012. IEEE.
[22] André R. Brodtkorb, Trond R. Hagen, and Martin L. Saetra. Graphics Processing Unit (GPU)
Programming Strategies and Trends in GPU Computing. Elsevier Journal of Parallel and
Distributed Computing, 73(1):4–13, January 2013.
[23] Ulrich Brüning, Wolfgang K. Giloi, and Wolfgang Schroeder-Preikschat. Latency Hiding in
Message-Passing Architectures. In Proceedings of the International Parallel Processing Sympo-
sium (IPPS '94), pages 704–709, Cancun, Mexico, April 1994. IEEE.
[24] Ian Buck, Tim Foley, Daniel Horn, Jeremy Sugerman, Kayvon Fatahalian, Mike Houston, and
Pat Hanrahan. Brook for GPUs: Stream Computing on Graphics Hardware. In Proceedings
of the International Conference on Computer Graphics and Interactive Techiques (SIGGRAPH
'04), pages 777–786, Los Angeles, CA, USA, August 2004. ACM.
26
[25] Martin Burtscher, Rupesh Nasre, and Keshav Pingali. A Quantitative Study of Irregular Pro-
grams on GPUs. In Proceedings of the IEEE International Symposium on Workload Character-
ization (IISWC '12), pages 141–151, La Jolla, CA, USA, November 2012. IEEE.
[26] Márcio Castro, Emilio Francesquini, Fabrice Dupros, Hideo Aochi, Philippe Navaux, and Jean-
François Mehaut. Seismic Wave Propagation Simulations on Low-Power and Performance-
Centric Manycores. Elsevier Journal of Parallel Computing, 54:108–120, May 2016.
[27] Shuai Che, Jie Li, Jeremy W. Sheaffer, Kevin Skadron, and John Lach. Accelerating Compute-
Intensive Applications With GPUs and FPGAs. In Proceedings of the Symposium on Application
Specific Processors (SASP '08), pages 101–107, Anaheim, CA, USA, June 2008. IEEE.
[28] Xi E. Chen and Tor M. Aamodt. A First-Order Fine-Grained Multithreaded Throughput
Model. In Proceedings of the IEEE International Symposium on High Performance Computer
Architecture (HPCA '09), pages 329–340, Raleigh, NC, USA, February 2009. IEEE.
[29] Thomas M. Cioppa, Thomas W. Lucas, and Susan M. Sanchez. Military Applications of Agent-
In Proceedings of the Winter Simulation Conference (WSC '04), pages
Based Simulations.
171–180, Washington, DC, USA, December 2004. IEEE.
[30] John G. Cleary, Murray Pearson, and Husam Kinawi. The Architecture of an Optimistic CPU:
The WarpEngine. In Proceedings of the Hawaii International Conference on System Sciences
(HICSS '95), pages 163–172, Wailea, HI, USA, January 1995. IEEE.
[31] Simon Coakley, Paul Richmond, Marian Gheorghe, Shawn Chin, David Worth, Mike Holcombe,
and Chris Greenough. Large-scale simulations with flame. In Joanna Kołodziej, Luís Correia,
and José Manuel Molina, editors, Intelligent Agents in Data-intensive Computing, pages 123–
142. Springer International Publishing, 2016.
[32] N.T. Collier and M.J. North. Repast sc++: A platform for large-scale agent-based modeling.
In W. Dubitzky, K. Kurowski, and B Schott, editors, Large-Scale Computing Techniques for
Complex System Simulations. Wiley, 2011.
[33] Biagio Cosenza, Gennaro Cordasco, Rosario De Chiara, and Vittorio Scarano. Distributed Load
Balancing for Parallel Agent-Based Simulations. In Proceedings of the International Euromicro
Conference on Parallel, Distributed and Network-Based Processing (PDP '11), pages 62–69,
Ayia Napa, Cyprus, February 2011. IEEE.
[34] Lintao Cui, Jing Chen, Yu Hu, Jinjun Xiong, Zhe Feng, and Lei He. Acceleration of Multi-Agent
Simulation on FPGAs. In Proceedings of the International Conference on Field Programmable
Logic and Applications (FPL '11), pages 470–473, Chania, Greece, September 2011. IEEE.
[35] Leonardo Dagum and Ramesh Menon. OpenMP: An Industry Standard API for Shared-Memory
Programming. IEEE Journal of Computational Science and Engineering, 5(1):46–55, January
1998.
[36] B. D. de Dinechin, R. Ayrignac, P. E. Beaucamps, P. Couvert, B. Ganne, P. G. de Massas,
F. Jacquet, S. Jones, N. M. Chaisemartin, F. Riss, and T. Strudel. A Clustered Manycore
Processor Architecture for Embedded and Accelerated Applications.
In Proceedings of the
IEEE High Performance Extreme Computing Conference (HPEC '13), pages 1–6, Waltham,
MA, USA, September 2013. IEEE.
[37] Narsingh Deo and Sushil Prasad. Parallel Heap: An Optimal Parallel Priority Queue. Springer
Journal of Supercomputing, 6(1):87–98, March 1992.
[38] Chris Ding and Yun He. A Ghost Cell Expansion Method for Reducing Communications in
Solving PDE Problems. In Proceedings of the ACM/IEEE Conference on Supercomputing (SC
'01), pages 55–55, Denver, CO, USA, November 2001. IEEE.
[39] Arnaud Doniec, René Mandiau, Sylvain Piechowiak, and Stéphane Espié. A Behavioral Multi-
Agent Model for Road Traffic Simulation. Elsevier Journal of Engineering Applications of
Artificial Intelligence, 21(8):1443–1454, December 2008.
27
[40] Roshan M. D'Souza, Mikola Lysenko, Simeone Marino, and Denise Kirschner. Data-Parallel
Algorithms for Agent-Based Model Simulation of Tuberculosis on Graphics Processing Units.
In Proceedings of the Spring Simulation Multiconference (SpringSim '09), pages 21:1–21:12, San
Diego, CA, USA, March 2009. SCSI.
[41] Samuel Eilenberg. Automata, Languages, and Machines. Academic Press, 1974.
[42] Joshua M. Epstein. Agent-Based Computational Models and Generative Social Science. Com-
plexity, 4(5):41–60, May 1999.
[43] Fernando A. Escobar, Xin Chang, and Carlos Valderrama. Suitability Analysis of FPGAs for
IEEE Transactions on Parallel and Distributed Systems,
Heterogeneous Platforms in HPC.
27(2):600–612, February 2016.
[44] Babak Falsafi, Bill Dally, Desh Singh, Derek Chiou, J Yi Joshua, and Resit Sendag. FPGAs
Versus GPUs in Data Centers. IEEE Micro, 37(1):60–72, January 2017.
[45] Naznin Fauzia, Louis-Noël Pouchet, and P. Sadayappan. Characterizing and Enhancing Global
Memory Data Coalescing on GPUs. In Proceedings of the Annual IEEE/ACM International
Symposium on Code Generation and Optimization (CGO '15), pages 12–22, San Francisco, CA,
USA, February 2015. IEEE.
[46] Richard M. Fujimoto. Performance of Time Warp Under Synthetic Workloads. In Proceedings of
the Distributed Simulation Conference (DSC '90), pages 23–28, San Diego, CA, USA, January
1990. SCS.
[47] Richard M. Fujimoto. Parallel and Distributed Simulation Systems. Wiley New York, 2000.
[48] Richard M. Fujimoto. Research Challenges in Parallel and Distributed Simulation. ACM
Transactions on Modeling and Computer Simulation (TOMACS), 26(4):22:1–22:29, May 2016.
[49] Richard M. Fujimoto, Conrad Bock, Wei Chen, Ernest Page, and Jitesh H. Panchal. Research
Challenges in Modeling and Simulation for Engineering Complex Systems. Springer, 2017.
[50] Richard M. Fujimoto, Christopher Carothers, Alois Ferscha, David Jefferson, Margaret Loper,
Madhav Marathe, and Simon J.E. Taylor. Computational Challenges in Modeling & Simulation
of Complex Systems. In Proceedings of the Winter Simulation Conference (WSC '17), pages
431–445, Las Vegas, NV, USA, December 2017. IEEE.
[51] Richard M. Fujimoto, Jya-Jang. Tsai, and Ganesh Gopalakrishnan. Design and Performance
of Special Purpose Hardware for Time Warp. In Proceedings of the Annual International Sym-
posium on Computer Architecture (SCA '88), pages 401–409, Honolulu, HI, USA, May 1988.
IEEE.
[52] Martin Gardner. Mathematical Games: The Fantastic Combinations of John Conway's New
Solitaire Game "Life". Scientific American, 223(4):120–123, October 1970.
[53] Ioakeim G. Georgoudas, Panagiotis Kyriakos, G. Ch. Sirakoulis, and I. Th. Andreadis.
An FPGA Implemented Cellular Automaton Crowd Evacuation Model Inspired by the
Electrostatic-Induced Potential Fields. Elsevier Journal of Microprocessors and Microsystems,
34(7):285–300, November 2010.
[54] Jacob Goldenberg, Barak Libai, and Eitan Muller. Talk of the Network: A Complex Systems
Look at the Underlying Process of Word-of-Mouth. Springer Marketing letters, 12(3):211–223,
August 2001.
[55] Mark Granovetter. Threshold Models of Collective Behavior. University of Chicago Press
American Journal of Sociology, 83(6):1420–1443, 1978.
[56] Ivan Grasso, Klaus Kofler, Biagio Cosenza, and Thomas Fahringer. Automatic Problem Size
In Proceedings of the ACM
Sensitive Task Partitioning on Heterogeneous Parallel Systems.
SIGPLAN Symposium on Principles and Practice of Parallel Programming (PPoPP '13), pages
281–282, Shenzhen, China, February 2013. ACM.
28
[57] Simon Green. Particle Simulation Using Cuda. NVIDIA Whitepaper, 6:121–128, 2010.
[58] Tobias Grosser, Armin Groesslinger, and Christian Lengauer. Polly-Performing Polyhedral
Optimizations on a Low-Level Intermediate Representation. World Scientific Parallel Processing
Letters, 22(04):1250010, December 2012.
[59] Tobias Grosser and Torsten Hoefler. Polly-ACC Transparent Compilation to Heterogeneous
Hardware. In Proceedings of the International Conference on Supercomputing (ICS '16), pages
1:1–1:13, Istanbul, Turkey, June 2016. ACM.
[60] Takahiro Harada. Real-Time Rigid Body Simulation on GPUs. NVIDIA GPU Gems, 3:123–148,
December 2007.
[61] Jack Harris and Matthias Scheutz. New Advances in Asynchronous Agent-based Scheduling. In
Proceedings of the International Conference on Parallel and Distributed Processing Techniques
and Applications (PDPTA '12), pages 1–7, Las Vegas, NV, USA, July 2012.
[62] Mark J. Harris, Greg Coombe, Thorsten Scheuermann, and Anselmo Lastra. Physically-
Based Visual Simulation on Graphics Hardware.
the ACM SIG-
GRAPH/EUROGRAPHICS Conference on Graphics Hardware (HWWS '02), pages 109–118,
Saarbrücken, Germany, September 2002. ACM.
In Proceedings of
[63] Scott Hauck and Andre DeHon. Reconfigurable Computing: The Theory and Practice of FPGA-
Based Computation. Morgan Kaufmann, 2010.
[64] Xi He, Deborah Agarwal, and Sushil K. Prasad. Design and Implementation of a Parallel
Priority Queue on Many-Core Architectures. In Proceedings of the International Conference on
High Performance Computing (HiPC '12), pages 1–10, Pune, India, December 2012. IEEE.
[65] Ulrich Heinkel, Martin Padeffke, Werner Haas, Thomas Buerner, Herbert Braisz, Thomas Gen-
tner, and Alexander Grassmann. The VHDL Reference: A Practical Guide to Computer-Aided
Integrated Circuit Design (Including VHDL-AMS). John Wiley & Sons, Inc., 2000.
[66] Wim Heirman, Dirk Stroobandt, Narasinga Rao Miniskar, Roel Wuyts, and Francky Catthoor.
PinComm: Characterizing Intra-Application Communication for the Many-Core Era. In Pro-
ceedings of the IEEE International Conference on Parallel and Distributed Systems (ICPADS
'10), pages 500–507, Shanghai, China, December 2010. IEEE.
[67] Emmanuel Hermellin and Fabien Michel. GPU Environmental Delegation of Agent Perceptions:
Application to Reynolds's Boids. In Proceedings of the International Workshop on Multi-Agent
Systems and Agent-Based Simulation (MABS '15), pages 71–86, Istanbul, Turkey, May 2015.
Springer.
[68] Emmanuel Hermellin and Fabien Michel. Defining a Methodology Based on GPU Delegation
for Developing MABS Using GPGPU. In Proceedings of the International Workshop on Multi-
Agent Systems and Agent-Based Simulation (MABS '16), pages 24–41, Singapore, May 2016.
Springer.
[69] Emmanuel Hermellin and Fabien Michel. GPU Delegation: Toward a Generic Approach for
Developping MABS using GPU Programming. In Proceedings of the International Conference
on Autonomous Agents & Multiagent Systems (AAMAS '16), pages 1249–1258, Singapore, May
2016. IFAAMAS.
[70] James D. Hess, Jacqueline J. Kacen, and Junyong Kim. Mood-Management Dynamics: The
Interrelationship Between Moods and Behaviours. Wiley Online Library British Journal of
Mathematical and Statistical Psychology, 59(2):347–378, November 2006.
[71] Peter Heywood, Paul Richmond, and Steve Maddock. Road Network Simulation Using FLAME
GPU. In Proceedings of the European Conference on Parallel Processing (Euro-Par '15), pages
430–441, Vienna, Austria, August 2015. Springer.
29
[72] Manato Hirabayashi, Shinpei Kato, Masato Edahiro, and Yuki Sugiyama. Toward GPU-
Accelerated Traffic Simulation and Its Real-Time Challenge.
In Proceedings of the Interna-
tional Workshop on Real-time and Distributed Computing in Emerging Applications (REAC-
TION '12), pages 45–50, San Juan, Puerto Rico, December 2012. Universidad Carlos III de
Madrid.
[73] Charles Antony Richard Hoare. Communicating Sequential Processes. Communications of the
ACM, 21(8):666–677, August 1978.
[74] Mike Holcombe. X-Machines as a Basis for Dynamic System Specification.
Engineering Journal, 3(2):69–76, March 1988.
IET Software
[75] Intel Corporation. Intel R(cid:13) FPGA SDK for OpenCL – Programming Guide. Technical Report
UG-OCL002, December 2017.
[76] Intel Corporation. Intel R(cid:13) Architecture Instruction Set Extensions and Future Features – Pro-
gramming Reference, January 2018.
[77] Thomas B. Jablin, James A. Jablin, Prakash Prabhu, Feng Liu, and David I. August. Dy-
namically Managed Data for CPU-GPU Architectures.
In Proceedings of the International
Symposium on Code Generation and Optimization (CGO '12), pages 165–174, San Jose, CA,
USA, March 2012. ACM.
[78] Thomas B. Jablin, Prakash Prabhu, James A. Jablin, Nick P. Johnson, Stephen R. Beard, and
David I. August. Automatic CPU-GPU Communication Management and Optimization. In
Proceedings of the ACM SIGPLAN Conference on Programming Language Design and Imple-
mentation (PLDI '11), pages 142–151, San Jose, CA, USA, June 2011. ACM.
[79] Deepak Jagtap, Ketan Bahulkar, Dmitry Ponomarev, and Nael Abu-Ghazaleh. Charac-
terizing and Understanding PDES Behavior on Tilera Architecture.
In Proceedings of the
ACM/IEEE/SCS Workshop on Principles of Advanced and Distributed Simulation (PADS '12),
pages 53–62, Zhangjiajie, China, July 2012. IEEE.
[80] Myeong-Wuk Jang and Gul Agha. Agent Framework Services to Reduce Agent Communication
Overhead in Large-Scale Agent-Based Simulations. Elsevier Journal of Simulation Modelling
Practice and Theory, 14(6):679–694, August 2006.
[81] David R. Jefferson. Virtual Time. ACM Transactions on Programming Languages and Systems,
7(3):404–425, July 1985.
[82] Y. Jiao, H. Lin, P. Balaji, and W. Feng. Power and Performance Characterization of Com-
putational Kernels on the GPU. In Proceedings of the IEEE/ACM International Conference
on Green Computing and Communications & International Conference on Cyber, Physical and
Social Computing (GREENCOM-CPSCOM '10), pages 221–228, Hangzhou, China, December
2010. IEEE.
[83] Jiangming Jin, Stephen John Turner, Bu-Sung Lee, Jianlong Zhong, and Bingsheng He. HPC
Simulations of Information Propagation Over Social Networks. In Proceedings of the Interna-
tional Conference on Computational Science (ICCS '12), pages 292–301, Omaha, NE, USA,
June 2012. Elsevier.
[84] Jiangming Jin, Stephen John Turner, Bu-Sung Lee, Jianlong Zhong, and Bingsheng He. Simula-
tion of Information Propagation Over Complex Networks: Performance Studies on Multi-GPU.
In Proceedings of the IEEE/ACM International Symposium on Distributed Simulation and Real
Time Applications (DS-RT '13), pages 179–188, Delft, Netherlands, October 2013. IEEE.
[85] Klaus Kofler, Gregory Davis, and Sandra Gesing. Sampo: An Agent-Based Mosquito Point
Model in OpenCL. In Proceedings of the Symposium on Agent Directed Simulation (ADS '14),
pages 5:1–5:10, Tampa, FL, USA, April 2014. SCSI.
30
[86] Klaus Kofler, Ivan Grasso, Biagio Cosenza, and Thomas Fahringer. An Automatic Input-
Sensitive Approach for Heterogeneous Task Partitioning. In Proceedings of the International
ACM Conference on International Conference on Supercomputing (ICS '13), pages 149–160,
Eugene, OR, USA, June 2013. ACM.
[87] Andreas Kolb, Lars John, et al. Volumetric Model Repair for Virtual Reality Applications. In
EUROGRAPHICS Short Presentation (2001), pages 249–256, Manchester, England, September
2001. The Eurographics Association.
[88] Andreas Kolb, Lutz Latta, and Christof Rezk-Salama.
Hardware-Based Simulation
and Collision Detection for Large Particle Systems.
In Proceedings of the ACM SIG-
GRAPH/EUROGRAPHICS Conference on Graphics Hardware (HWWS '04), pages 123–131,
Grenoble, France, August 2004. ACM.
[89] Georg Kunz, Daniel Schemmel, James Gross, and Klaus Wehrle. Multi-Level Parallelism for
Time- and Cost-Efficient Parallel Discrete Event Simulation on GPUs. In Proceedings of the
ACM/IEEE/SCS Workshop on Principles of Advanced and Distributed Simulation (PADS '12),
pages 23–32, Zhangjiajie, China, July 2012. IEEE.
[90] Chenggang Lai, Miaoqing Huang, Xuan Shi, and Haihang You. Accelerating Geospatial Appli-
cations on Hybrid Architectures. In Proceedings of the IEEE International Conference on High
Performance Computing and Communications, IEEE International Conference on Embedded
and Ubiquitous Computing (HPCC & EUC '13), pages 1545–1552, Singapore, November 2013.
IEEE.
[91] Leslie Lamport. Specifying Systems: The TLA+ Language and Tools for Hardware and Software
Engineers. Addison-Wesley Longman Publishing Co., Inc., 2002.
[92] Guillaume Laville, Kamel Mazouzi, Christophe Lang, Nicolas Marilleau, Bénédicte Herrmann,
and Laurent Philippe. MCMAS: A Toolkit to Benefit From Many-Core Architecure in Agent-
Based Simulation. In Proceedings of the European Conference on Parallel Processing (Euro-Par
'13), pages 544–554, Aachen, Germany, August 2013. Springer.
[93] Guillaume Laville, Kamel Mazouzi, Christophe Lang, Laurent Philipppe, and Nicolas Marilleau.
Using GPU for Multi-Agent Soil Simulation.
In Proceedings of the Euromicro International
Conference on Parallel, Distributed and Network-Based Processing (PDP '13), pages 392–399,
Belfast, UK, February 2013. IEEE.
[94] Seyong Lee, Seung-Jai Min, and Rudolf Eigenmann. OpenMP to GPGPU: A Compiler Frame-
In Proceedings of the ACM SIGPLAN
work for Automatic Translation and Optimization.
Symposium on Principles and Practice of Parallel Programming (PPoPP '09), pages 101–110,
Raleigh, NC, USA, February 2009. ACM.
[95] Xiaosong Li, Wentong Cai, and Stephen John Turner. GPU Accelerated Three-Stage Execu-
tion Model for Event-Parallel Simulation. In Proceedings of the ACM SIGSIM Conference on
Principles of Advanced Discrete Simulation (PADS '13), pages 57–66, Montreal, Canada, May
2013. ACM.
[96] Xiaosong Li, Wentong Cai, and Stephen John Turner. Efficient Neighbor Searching for Agent-
Based Simulation on GPU. In Proceedings of the IEEE/ACM International Symposium on Dis-
tributed Simulation and Real Time Applications (DS-RT '14), pages 87–96, Toulouse, France,
October 2014. IEEE.
[97] Xiaosong Li, Wentong Cai, and Stephen John Turner. Cloning Agent-based Simulation on GPU.
In Proceedings of the ACM SIGSIM Conference on Principles of Advanced Discrete Simulation
(PADS '15), pages 173–182, London, UK, June 2015. ACM.
[98] Xinhu Liu and Philipp Andelfinger. Time Warp on the GPU: Design and Assessment.
In
Proceedings of the ACM SIGSIM Conference on Principles of Advanced Discrete Simulation
(PADS '17), pages 109–120, Singapore, May 2017. ACM.
31
[99] Xu Liu, Langshi Chen, Jesun S Firoz, Judy Qiu, and Lei Jiang. Performance Characterization
of Multi-Threaded Graph Processing Applications on Intel Many-Integrated-Core Architecture.
August 2017.
[100] Qingqi Long, Jie Lin, and Zhixun Sun. Agent Scheduling Model for Adaptive Dynamic Load
Balancing in Agent-Based Distributed Simulations. Elsevier Journal of Simulation Modelling
Practice and Theory, 19(4):1021–1034, April 2011.
[101] Sean Luke, Claudio Cioffi-Revilla, Liviu Panait, Keith Sullivan, and Gabriel Balan. MASON:
A Multiagent Simulation Environment. SCSI Journal of Simulation, 81(7):517–527, July 2005.
[102] Elizabeth Whitaker Lynch. Hardware Acceleration for Conservative Parallel Discrete Event
Simulation on Multi-Core Systems. PhD thesis, School of Electrical and Computer Engineering,
Georgia Institute of Technology, February 2011.
[103] Mikola Lysenko and Roshan M. D'Souza. A Framework for Megascale Agent Based Model
Simulations on Graphics Processing Units. JASSS Journal of Artificial Societies and Social
Simulation, 11(4):10, October 2008.
[104] Charles M. Macal and Michael J. North. Tutorial on Agent-Based Modeling and Simulation.
Springer Journal of Simulation, 4(3):151–162, September 2010.
[105] Deepak Majeti, Kuldeep S. Meel, Rajkishore Barik, and Vivek Sarkar. Automatic Data Layout
Generation and Kernel Mapping for CPU+GPU Architectures. In Proceedings of the Interna-
tional Conference on Compiler Construction (CC '16), pages 240–250, Barcelona, Spain, March
2016. ACM.
[106] Fabien Michel. Translating Agent Perception Computations Into Environmental Processes in
Multi-Agent-Based Simulations: A Means for Integrating Graphics Processing Unit Program-
ming Within Usual Agent-Based Simulation Platforms. Wiley Online Library Journal of Sys-
tems Research and Behavioral Science, 30(6):703–715, November 2013.
[107] Sparsh Mittal and Jeffrey S. Vetter. A Survey of CPU-GPU Heterogeneous Computing Tech-
niques. ACM Computing Surveys, 47(4):69:1–69:35, July 2015.
[108] Josh Model and Martin C. Herbordt. Discrete Event Simulation of Molecular Dynamics With
Configurable Logic. In Proceedings of the International Conference on Field Programmable Logic
and Applications (FPL '07), pages 151–158, Amsterdam, Netherlands, August 2007. IEEE.
[109] Kai Nagel and Marcus Rickert. Parallel Implementation of the TRANSIMS Micro-Simulation.
Elsevier Journal of Parallel Computing, 27(12):1611–1639, November 2001.
[110] Razvan Nane, Vlad-Mihai Sima, Christian Pilato, Jongsok Choi, Blair Fort, Andrew Canis,
Yu Ting Chen, Hsuan Hsiao, Stephen Brown, Fabrizio Ferrandi, Jason Anderson, and Koen
Bertels. A Survey and Evaluation of FPGA High-Level Synthesis Tools. IEEE Transactions on
Computer-Aided Design of Integrated Circuits and Systems, 35(10):1591–1604, October 2016.
[111] Roland Neumann and Fritz Strack.
'Mood Contagion": The Automatic Transfer of Mood
Between Persons. American Psychological Association Journal of Personality and Social Psy-
chology, 79(2):211, August 2000.
[112] Michael J. North and Charles M. Macal. Managing Business Complexity: Discovering Strategic
Solutions With Agent-Based Modeling and Simulation. Oxford University Press, 2007.
[113] NVIDIA Corporation. CUDA Toolkit 4.2 – CUBLAS Library. Technical Report PG-05326-
041_v01, Santa Clara, CA, USA, February 2012.
[114] NVIDIA Corporation. Whitepaper – NVIDIA GeForce GTX 1080 – Gaming Perfected, 2016.
[115] NVIDIA Corporation. Whitepaper – NVIDIA Tesla P100 – The Most Advanced Datacenter
Accelerator Ever Built Featuring Pascal GP100, the World's Fastest GPU. Technical Report
WP-08019-001_v01.1, 2016.
32
[116] NVIDIA Corporation. NVIDIA TESLA V100 GPU Architecture – The World's Most Advanced
Data Center GPU. Technical Report WP-08608-001_v1.1, 2017.
[117] NVIDIA Corporation. NVIDIA CUDA C Programming Guide. Version 9.1.85, NVIDIA Cor-
poration, 2018.
[118] Christopher Oat, Joshua Barczak, and Jeremy Shopf. Efficient Spatial Binning on the GPU.
Technical report, Advanced Micro Devices, Inc, February 2009.
[119] Andreas Olofsson. Epiphany-V: A 1024 processor 64-bit RISC System-On-Chip. October 2016.
[120] OpenACC Working Group. The OpenACC Application Programming Interface. Technical
Report Version 2.5, October 2015.
[121] Samir Palnitkar. Verilog R(cid:13)Hdl: A Guide to Digital Design and Synthesis, Second Edition.
Prentice Hall Press, 2003.
[122] Hyungwook Park and Paul A. Fishwick. A Fast Hybrid Time-Synchronous/Event Approach
In Proceedings of the Winter
to Parallel Discrete Event Simulation of Queuing Networks.
Simulation Conference (WSC '08), pages 795–803, Miami, FL, USA, December 2008. IEEE.
[123] Hyungwook Park and Paul A. Fishwick. A GPU-Based Application Framework Supporting
Fast Discrete-Event Simulation. SCSI Simulation, 86(10):613–628, October 2010.
[124] Hyungwook Park and Paul A. Fishwick. An Analysis of Queuing Network Simulation Using
GPU-Based Hardware Acceleration. ACM Transactions on Modeling and Computer Simulation,
21(3):18:1–18:22, March 2011.
[125] Roman Pavlov and Jörg P. Müller. Multi-Agent Systems Meet GPU: Deploying Agent-Based
Architectures on Graphics Processors. In Proceedings of the Doctoral Conference on Computing,
Electrical and Industrial Systems (DoCEIS '13), pages 115–122, Costa de Caparica, Portugal,
April 2013. Springer.
[126] Kalyan S. Perumalla. Discrete-Event Execution Alternatives on General Purpose Graphical
Processing Units (GPGPUs). In Proceedings of the Workshop on Principles of Advanced and
Distributed Simulation (PADS '06), pages 74–81, Singapore, May 2006. IEEE.
[127] Kalyan S. Perumalla. Efficient Execution on GPUs of Field-Based Vehicular Mobility Models.
In Proceedings of the Workshop on Principles of Advanced and Distributed Simulation (PADS
'08), pages 154–154, Roma, Italy, June 2008. IEEE.
[128] Kalyan S. Perumalla and Brandon G. Aaby. Data Parallel Execution Challenges and Run-
In Proceedings of the Spring Simulation
time Performance of Agent Simulations on GPUs.
Multiconference (SpringSim '08), pages 116–123, Ottawa, Canada, April 2008. SCSI.
[129] Kalyan S. Perumalla, Brandon G. Aaby, Srikanth B. Yoginath, and Sudip K. Seal. GPU-Based
Real-Time Execution of Vehicular Mobility Models in Large-Scale Road Network Scenarios.
In Proceedings of the ACM/IEEE/SCS Workshop on Principles of Advanced and Distributed
Simulation (PADS '09), pages 95–103, Lake Placid, NY, USA, June 2009. IEEE.
[130] Louis-Noël Pouchet. Polybench: the Polyhedral Benchmark suite. http://web.cs.ucla.edu/
~pouchet/software/polybench/, 2012.
[131] Sebastian Raase and Tomas Nordstrm. On the Use of a Many-Core Processor for Computational
Fluid Dynamics Simulations. pages 1403–1412, Reykjavík, Iceland, June 2015. Elsevier.
[132] Shafiur Rahman, Nael Abu-Ghazaleh, and Walid Najjar. PDES-A: A Parallel Discrete Event
Simulation Accelerator for FPGAs. In Proceedings of the ACM SIGSIM Conference on Princi-
ples of Advanced Discrete Simulation (PADS '17), pages 133–144, Singapore, May 2017. ACM.
[133] Ashu Rege. An Introduction to Modern GPU Architecture, 2008.
33
[134] Paul F. Reynolds, Carmen M. Pancerella, and Sudhir Srinivasan. Design and Performance Anal-
ysis of Hardware Support for Parallel Simulations. Elsevier Journal of Parallel and Distributed
Computing, 18(4):435–453, August 1993.
[135] Paul Richmond, Simon Coakley, and Daniela Romano. Cellular Level Agent Based Modelling
on the Graphics Processing Unit. In Proceedings of the International Workshop on High Perfor-
mance Computational Systems Biology (HIBI '09), pages 43–50, Trento, Italy, October 2009.
IEEE.
[136] Paul Richmond and Daniela Romano. Agent Based GPU, a Real-time 3D Simulation and
Interactive Visualisation Framework for Massive Agent Based Modelling on the GPU. In Pro-
ceedings of the International Workshop on Supervisualisation (IWSV '08), Island of Kos, Aegean
Sea,Greece, June 2008.
[137] Paul Richmond and Daniela Romano. Template-Driven Agent-Based Modeling and Simula-
tion With CUDA. Elsevier GPU Computing Gems Emerald Edition, Applications of GPU
Computing Series, pages 313–324, February 2011.
[138] Patrick F. Riley and George F. Riley. Next Generation Modeling III-Agents: Spades-a Dis-
tributed Agent Simulation Environment With Software-in-the-Loop Execution. In Proceedings
of the Conference on Winter Simulation (WSC '03), pages 817–825, New Orleans, LA, USA,
December 2003. IEEE.
[139] Robert Rönngren and Rassul Ayani. A Comparative Study of Parallel and Sequential Priority
Queue Algorithms. ACM Transactions on Modeling and Computer Simulation, 7(2):157–209,
April 1997.
[140] Janche Sang, Che-Rung Lee, Vernon Rego, and Chung-Ta King. A Fast Implementation of
Parallel Discrete-Event Simulation on GPGPU. In Proceedings of the International Conference
on Parallel and Distributed Processing Techniques and Applications (PDPTA '13), page 501,
Las Vegas, NV, US, July 2013.
[141] Thomas C. Schelling. Micromotives and Macrobehavior. WW Norton & Company, 2006.
[142] Moon Gi Seok and Tag Gon Kim. Parallel Discrete Event Simulation for DEVS Cellular Models
Using a GPU. In Proceedings of the Symposium on High Performance Computing (HPC '12),
pages 11:1–11:7, Orlando, FL, USA, March 2012. SCSI.
[143] Zhen Shen, Kai Wang, and Fenghua Zhu. Agent-Based Traffic Simulation and Traffic Signal
Timing Optimization With GPU.
In Proceedings of the International IEEE Conference on
Intelligent Transportation Systems (ITSC '11), pages 145–150, Washington, DC, USA, October
2011. IEEE.
[144] Xuan Shi and Fei Ye. Kriging Interpolation Over Heterogeneous Computer Architectures and
Systems. Taylor & Francis Journal of GIScience & Remote Sensing, 50(2):196–211, 2013.
[145] Jaewoong Sim, Aniruddha Dasgupta, Hyesoon Kim, and Richard Vuduc. A Performance Anal-
ysis Framework for Identifying Potential Benefits in GPGPU Applications. In Proceedings of
the ACM SIGPLAN symposium on Principles and Practice of Parallel Programming (PPoPP
'12), pages 11–22, New Orleans, LA, USA, February 2012. ACM.
[146] Daniel Dominic Sleator and Robert Endre Tarjan. Self-Adjusting Binary Search Trees. Journal
of the ACM, 32(3):652–686, July 1985.
[147] Avinash Sodani, Roger Gramunt, Jesus Corbal, Ho-Seop Kim, Krishna Vinod, Sundaram
Chinthamani, Steven Hutsell, Rajat Agarwal, and Yen-Chen Liu. Knights Landing: Second-
Generation Intel Xeon Phi Product. IEEE Micro, 36(2):34–46, April 2016.
[148] Xiao Song, Ziping Xie, Yan Xu, Gary Tan, Wenjie Tang, Jing Bi, and Xiaosong Li. Support-
ing Real-World Network-Oriented Mesoscopic Traffic Simulation on GPU. Elsevier Journal of
Simulation Modelling Practice and Theory, 74:46–63, May 2017.
34
[149] Russell K. Standish and Richard Leow. EcoLab: Agent based modeling for C++ programmers.
January 2004.
[150] John E. Stone, David Gohara, and Guochun Shi. OpenCL: A Parallel Programming Standard
for Heterogeneous Computing Systems. IEEE Computing in Science & Engineering, 12(3):66–
73, May 2010.
[151] David Strippgen and Kai Nagel. Using Common Graphics Hardware for Multi-Agent Traffic
Simulation With CUDA. In Proceedings of the International Conference on Simulation Tools
and Techniques (Simutools '09), pages 62:1–62:8, Rome, Italy, March 2009. ICST.
[152] Herb Sutter. The Free Lunch Is Over: A Fundamental Turn Toward Concurrency in Software.
Dr. Dobb's Journal, 30(3):202–210, March 2005.
[153] Brian Paul Swenson. Techniques to Improve the Performance of Large-Scale Discrete-Event
Simulation. Dissertation, Georgia Institute of Technology, May 2015.
[154] G. Szemes, L. Gulyás, G. Kampis, and W. de Back. GridABM-templates for Distributed Agent
Based Simulation. In Open Grid Forum, volume 28, Munich , Germany, 2010.
[155] Wenjie Tang and Yiping Yao. A GPU-Based Discrete Event Simulation Kernel. SCSI Journal
of Simulation, 89(11):1335–1354, October 2013.
[156] Kardi Teknomo, Yasushi Takeyama, and Hajime Inamura. Review on Microscopic Pedestrian
Simulation Model. September 2016.
[157] Leigh Tesfatsion. Agent-Based Computational Economics: A Constructive Approach to Eco-
nomic Theory. Elsevier Handbook of Computational Economics, 2:831–880, May 2006.
[158] Y. Torres, A. Gonzalez-Escribano, and D.R. Llanos. Understanding the Impact of CUDA Tuning
In Proceedings of the International Conference on High Performance
Techniques for Fermi.
Computing and Simulation (HPCS '11), pages 631–639, Istanbul, Turkey, July 2011. IEEE.
[159] Justin L. Tripp, Henning S. Mortveit, Anders A. Hansson, and Maya Gokhale. Metropolitan
Road Traffic Simulation on FPGAs. In Proceedings of the Annual IEEE Symposium on Field-
Programmable Custom Computing Machines (FCCM '05), pages 117–126, Napa, CA, USA,
April 2005. IEEE.
[160] Mario Vestias and Horácio Neto. TRENDS OF CPU, GPU AND FPGA FOR HIGH-
PERFORMANCE COMPUTING.
In Proceedings of the International Conference on Field
Programmable Logic and Applications (FPL '14), pages 1–6, Munich, Germany, September
2014. IEEE.
[161] Guillermo Vigueras, Juan M. Orduña, Miguel Lozano, José M. Cecilia, and José M. García. Ac-
celerating Collision Detection for Large-Scale Crowd Simulation on Multi-Core and Many-Core
Architectures. Sage Publications The International Journal of High Performance Computing
Applications, 28(1):33–49, February 2014.
[162] Ioannis Vourkas and Georgios Ch. Sirakoulis. FPGA Based Cellular Automata for Environ-
mental Modeling. In Proceedings of the International Conference on Electronics, Circuits and
Systems (ICECS '12), pages 93–96, Seville, Spain, December 2012. IEEE.
[163] Jin Wang, Norman Rubin, Haicheng Wu, and Sudhakar Yalamanchili. Accelerating Simulation
of Agent-Based Models on Heterogeneous Architectures. In Proceedings of the Workshop on
General Purpose Processor Using Graphics Processing Units (GPGPU '06), pages 108–119,
Houston, TX, USA, March 2013. ACM.
[164] Kai Wang and Zhen Shen. A GPU Based Trafficparallel Simulation Module of Artificial Trans-
portation Systems. In Proceedings of the IEEE International Conference on Service Operations
and Logistics, and Informatics (SOLI '12), pages 160–165, Suzhou, China, July 2012. IEEE.
35
[165] Yuan Wen, Zheng Wang, and Michael F.P. O'Boyle. Smart Multi-Task Scheduling for OpenCL
Programs on CPU/GPU Heterogeneous Platforms. In Proceedings of the International Confer-
ence on High Performance Computing (HiPC '14), pages 1–10, Dona Paula, India, December
2014. IEEE.
[166] Tang Wenjie, Yao Yiping, and Zhu Feng. An Expansion-Aided Synchronous Conservative Time
Management Algorithm on GPU. In Proceedings of the ACM SIGSIM Conference on Principles
of Advanced Discrete Simulation (PADS '13), pages 367–372, Montreal, Canada, May 2013.
ACM.
[167] Barry Williams, Dmitry Ponomarev, Nael Abu-Ghazaleh, and Philip Wilsey. Performance
Characterization of Parallel Discrete Event Simulation on Knights Landing Processor. In Pro-
ceedings of the ACM SIGSIM Conference on Principles of Advanced Discrete Simulation (PADS
'17), pages 121–132, Singapore, Singapore, May 2017. ACM.
[168] Wąs, Jarosław and Mróz, Hubert and Topa, Paweł. GPGPU Computing for Microscopic
Simulations of Crowd Dynamics. Slovak Academy of Sciences Computing and Informatics,
34(6):1418–1434, February 2016.
[169] Fulong Wu. Calibration of Stochastic Cellular Automata: The Application to Rural-Urban
Land Conversions. Taylor & Francis International Journal of Geographical Information Science,
16(8):795–818, 2002.
[170] Xilinx. XilinxVirtex UltraScale+ Product Table.
https://www.xilinx.com/products/
silicon-devices/fpga/virtex-ultrascale-plus.html#productTable, December 2017.
[171] Yadong Xu, Wentong Cai, David Eckhoff, Suraj Nair, and Alois Knoll. A Graph Partition-
ing Algorithm for Parallel Agent-Based Road Traffic Simulation. In Proceedings of the ACM
SIGSIM Conference on Principles of Advanced Discrete Simulation (PADS '17), pages 209–219,
Singapore, May 2017. ACM.
[172] Yan Xu, Gary Tan, Xiaosong Li, and Xiao Song. Mesoscopic traffic simulation on CPU/GPU.
In Proceedings of the ACM SIGSIM Conference on Principles of Advanced Discrete Simulation
(PADS '14), pages 39–50, Denver, CO, USA, May 2014. ACM.
[173] Yonghong Yan, Max Grossman, and Vivek Sarkar. JCUDA: A Programmer-Friendly Inter-
face for Accelerating Java Programs With CUDA. In Proceedings of the European Conference
on Parallel Processing (Euro-Par '09), pages 887–899, Delft, The Netherlands, August 2009.
Springer.
[174] Mingyu Yang, Philipp Andelfinger, Wentong Cai, and Alois Knoll. Evaluation of Conflict
Resolution Methods for Agent-Based Simulation on the GPU.
In Proceedings of the ACM
SIGSIM Conference on Principles of Advanced Discrete Simulation (PADS '18), pages 129–
132, Rome, Italy, May 2018. ACM.
[175] Srikanth B. Yoginath and Kalyan S. Perumalla. Scalable Cloning on Large-Scale GPU Platforms
With Application to Time-Stepped Simulations on Grids. ACM Transactions on Modeling and
Computer Simulation, 28(1):5:11–5:26, January 2018.
[176] Feng Zhang, Jidong Zhai, Bingsheng He, Shuhao Zhang, and Wenguang Chen. Understanding
Co-Running Behaviors on Integrated CPU/GPU Architectures. IEEE Transactions on Parallel
and Distributed Systems, 28(3):905–918, March 2017.
[177] Tao Zhang, Xin Zhao, Xinqi An, Haojun Quan, and Zhichun Lei. Using Blind Optimization
Algorithm for Hardware/Software Partitioning. IEEE Access, 5:1353–1362, February 2017.
[178] Yao Zhang and John D Owens. A Quantitative Performance Analysis Model for GPU Ar-
chitectures. In Proceedings of the 2011 IEEE International Symposium on High Performance
Computer Architecture (HPCA '11), pages 382–393, San Antonio, TX, USA, February 2011.
IEEE.
36
[179] Li Zhen, Qiuxiao Gang, Guo Gang, and Chen Bin. A GPU-Based Simulation Kernel within
IACSIT Press
Heterogeneous Collaborative Computation on Large-Scale Artificial Society.
International Journal of Modeling and Optimization, 4(3):205–210, Jun 2014.
[180] Hamid Reza Zohouri, Naoya Maruyama, Aaron Smith, Motohiko Matsuda, and Satoshi Mat-
suoka. Evaluating and Optimizing OpenCL Kernels for High Performance Computing With
FPGAs. In Proceedings of the International Conference for High Performance Computing, Net-
working, Storage and Analysis (SC '16), pages 35:1–35:12, Salt Lake City, UT, USA, November
2016. IEEE.
[181] Peng Zou, Ya-shuai Lü, Li-li Chen, and Yi-ping Yao. Epidemic Simulation of Large-Scale Social
Contact Network on GPU Clusters. Sage Publications Journal of Simulation, 89(10):1154–1172,
2013.
37
|
1606.07525 | 1 | 1606 | 2016-06-24T00:32:41 | Relating Knowledge and Coordinated Action: The Knowledge of Preconditions Principle | [
"cs.MA",
"cs.AI",
"cs.DC",
"cs.LO"
] | The Knowledge of Preconditions principle (KoP) is proposed as a widely applicable connection between knowledge and action in multi-agent systems. Roughly speaking, it asserts that if some condition is a necessary condition for performing a given action A, then knowing that this condition holds is also a necessary condition for performing A. Since the specifications of tasks often involve necessary conditions for actions, the KoP principle shows that such specifications induce knowledge preconditions for the actions. Distributed protocols or multi-agent plans that satisfy the specifications must ensure that this knowledge be attained, and that it is detected by the agents as a condition for action. The knowledge of preconditions principle is formalised in the runs and systems framework, and is proven to hold in a wide class of settings. Well-known connections between knowledge and coordinated action are extended and shown to derive directly from the KoP principle: a "common knowledge of preconditions" principle is established showing that common knowledge is a necessary condition for performing simultaneous actions, and a "nested knowledge of preconditions" principle is proven, showing that coordinating actions to be performed in linear temporal order requires a corresponding form of nested knowledge.
| cs.MA | cs | Relating Knowledge and Coordinated Action:
The Knowledge of Preconditions Principle
Yoram Moses∗
Technion -- Israel Institute of Technology
[email protected]
The Knowledge of Preconditions principle (KoP) is proposed as a widely applicable connection
between knowledge and action in multi-agent systems. Roughly speaking, it asserts that if some
condition ϕ is a necessary condition for performing a given action α, then knowing ϕ is also a neces-
sary condition for performing α. Since the specifications of tasks often involve necessary conditions
for actions, the KoP principle shows that such specifications induce knowledge preconditions for
the actions. Distributed protocols or multi-agent plans that satisfy the specifications must ensure
that this knowledge be attained, and that it is detected by the agents as a condition for action. The
knowledge of preconditions principle is formalised in the runs and systems framework, and is proven
to hold in a wide class of settings. Well-known connections between knowledge and coordinated
action are extended and shown to derive directly from the KoP principle: a common knowledge
of preconditions principle is established showing that common knowledge is a necessary condition
for performing simultaneous actions, and a nested knowledge of preconditions principle is proven,
showing that coordinating actions to be performed in linear temporal order requires a corresponding
form of nested knowledge.
Keywords: Knowledge, multi-agent systems, common knowledge, nested knowledge, coordinated
action, knowledge of preconditions principle.
1 Introduction
While epistemology, the study of knowledge, has been a topic of interest in philosophical circles for cen-
turies and perhaps even millennia, in the last half century it has seen a flurry of activity and applications
in other fields such as AI [19], game theory [2] and distributed computing [13]. At least in the latter two
fields a particular, information-based, notion of knowledge plays a prominent and useful role.
This paper proposes an essential connection between knowledge and action in such a setting. Using
doesi(α) to denote "Agent i is performing action α" and Kiϕ to denote that Agent i knows the fact ϕ,
the connection can intuitively be formulated as follows:
The KNOWLEDGE OF PRECONDITIONS Principle (KoP):
ϕ is a necessary condition for doesi(α)
If
then Kiϕ is a necessary condition for doesi(α)
This statement appears deceptively simple. In fact, many successful applications of knowledge to the
design and analysis of distributed protocols over the last three decades are rooted in the KoP. Moreover,
some of the deeper insights obtained by knowledge theory in this field can be derived in a fairly direct
∗The Israel Pollak academic chair at Technion. This work was supported in part by ISF grant 1520/11.
R. Ramanujam (Ed.): TARK 2015
EPTCS 215, 2016, pp. 231 -- 245, doi:10.4204/EPTCS.215.17
c(cid:13) Yoram Moses
This work is licensed under the
Creative Commons Attribution License.
232
The Knowledge of Preconditions Principle
Figure 1: A simple four-agent system
fashion from the KoP. We will argue and demonstrate that this principle lies at the heart of coordination
in many distributed and multi-agent systems.
This paper is structured as follows. Section 1.1 illustrates the central role of knowledge in a natural
distributed systems application. Section 1.2 provides a high-level discussion of the knowledge of pre-
conditions principle and its connection to coordinating actions. In Section 2 we review and discuss the
modelling of knowledge in the runs and systems model of distributed systems based on [10]. A formal
statement and proof of the KoP are presented in Section 3. Then, in Section 4, the KoP is used to es-
tablish a common knowledge of preconditions principle. It states that in order to perform simultaneously
coordinated actions, agents must first attain common knowledge of any of the actions' preconditions. An
example of its use is provided in Section 4.1. Section 5 present an additional use of the KoP, and shows
that coordinating a sequence of actions to occur in a prescribed temporal order requires attaining nested
knowledge of their preconditions. Finally, Section 6 discusses additional applications, extensions and
future directions.
1.1 The Case for Knowledge in Distributed Systems
Why should knowledge play a central role in distributed computing? As pointed out in [13], most every-
one who designs or even just tries to study the workings of a distributed protocol is quickly found talking
in terms of knowledge, making statements such as "once the process receives an acknowledgement, it
knows that the other process is ready. . . ". An essential aspect of distributed systems is the fact that an
agent chooses which action to perform based on the local information available to it, which typically
provides only a partial view of the overall state of the system. To get a sense of the role of knowledge in
distributed systems, consider the following example.
Example 1. Given is a distributed network modeled by a graph, with agents located at the nodes, and the
edges standing for communication channels (see Figure 1). In the problem we shall call Computing the
Max (or CTM for short), each agent i starts out with a natural number vi ∈ N as an initial value. The goal
is to have Agent 1 print the maximum of all of the initial values (we denote this value by Max), and print
nothing else. In the instance depicted in Figure 1, the maximal value happens to be 100. Initially, Agent 1
clearly can't print its own initial value of 75. Suppose that Agent 1 receives a message µ (cid:44) "v2 = 100"
from Agent 2 reporting that its value is 100. At this point, Agent 1 has access to the maximum, and
printing 100 would satisfy the problem specification. Compare this with a setting that is the same in all
respects, except that Agent 3's value is v3 = 150. In this case, of course, Max (cid:54)= 100 and so printing 100 is
forbidden. But if Agent 1 can receive the same message µ under similar circumstances in both scenarios,
7510080100 Yoram Moses
233
then it is unable to distinguish whether or not Max = 100 upon receiving µ. Intuitively, even in the first
scenario, the agent does not know that Max = 100.
What information does Agent 1 need, then, in order to be able to print the maximum? Notice that
it is not necessary, in general, to collect all of the initial values in order to print the maximum. For
example, suppose that the agents follow a bottom-up protocol in which values are sent from right to left,
starting from Agent 4, and every agent passes to the left the larger of its own value and the value it
received from its neighbor on the right (if such a neighbor exists). In this protocol, Agent 1 can clearly
print the maximum after receiving the message "v2 = 100", and seeing just one value besides its own.
Interestingly, even collecting all of the values is not a sufficient condition for printing the Max. Imagine
a setting in which the network is as in Figure 1, but Agent 1 considers it possible that there are more than
four nodes in the network. In this case, even if Agent 1 receives all (four) values, it may still need to wait
(cid:3)
for proof that there is no additional, larger, value in the system.
CTM is a simplified example in the spirit of many distributed systems applications. In fact, a central
problem called Leader Election, for example, is often solved by computing a node with maximal ID
[1, 17]. The solution to such a problem is typically in the form of a set of short computer programs
(jointly constituting a distributed protocol), each executed at one of the nodes. When the nodes follow
such a protocol, the resulting execution should satisfy the problem specification. Of course, the programs
are written in a standard programming language, without any reference to knowledge or possibility. In
the vast majority of cases, the programs in question do not enumerate and/or explore possible states or
scenarios. Indeed, the program designer is typically unfamiliar with formal notions of knowledge. This
being the case, what sense does it make to talk of Agent 1 in Example 1 "knowing" or "not knowing"
that Max = c? Can it make sense to say that the Agent "considers it possible that there may be more than
four nodes in the system"? After all, we may be talking about a 10-line program. It has no soul. Does it
have thoughts, doubts and mental states?
Since agents act based on their local information, a protocol designer must ensure that agents ob-
tain the necessary information for a given task, and that this information is applied correctly. Using the
information-based notion of knowledge, the designer can ascribe knowledge to an agent without requir-
ing it to have a soul, feelings, and self-awareness. As seen in the CTM example, it is natural to think in
terms of whether or not Agent 1 knows Max = c at any given point in a run of a protocol solving CTM.
(A formal definition of knowledge will be provided in Section 2.) Suppose that a protocol is designed
to solve CTM in networks that may have a variety of sizes. If Agent 1 does not start out with local
information ensuring that there are no more than four nodes in the system, then from the point of view of
an outside observer the agent can be thought of as "considering it possible" that there may be more than
four nodes.
Even in a simple network as in Figure 1, the CTM problem can be posed in different models, which
can differ in essential aspects. A solution to CTM in one model might not solve the problem in another
model. Indeed, the rationale behind distinct solutions, as well as their details, may vary considerably.
Are there common features shared by all solutions to CTM?
Interestingly, all solutions to CTM, in all models, share one property: Agent 1 must know that
Max = c in order to print the value c. Indeed, the ability to print the answer in a protocol for CTM
reduces to detecting when the Max value is known. Of course, once Agent 1 knows that Max = c it
can safely print c. Hence, knowing that Max = c is not just necessary, but also a sufficient condition
for printing c. The CTM problem shows that knowledge and attaining knowledge can be a central and
crucial aspect of a standard distributed application.
The need to know Max = c in solving CTM suggests that we consider a natural question: When
does Agent 1 know that Max = c? The answer is less straightforward than we might initially expect.
234
The Knowledge of Preconditions Principle
What is known depends in a crucial way on the protocol that the agents are following. Thus, in the
setting of Example 1, if the agents follow the bottom-up protocol, then Agent 1 knows the maximum
once it receives a single message from Agent 2. Knowledge is also significantly affected by features of
the model. In CTM, if there is an upper bound (say 100) on the possible initial values, then an agent
that sees this value knows the maximum. Knowledge about the network topology and properties of
communication play a role as well. For example, consider a model in which Agent 1 has a clock, and
a single clock cycle suffices for a message to be delivered, digested, and acted upon. Suppose that the
protocol is such that all agents start simultaneously at time 0 and an agent forwards a value towards
Agent 1 only if this value is larger than any value it has previously sent. Then in the network of Figure 1
Agent 1 will receive a message with value 100 from Agent 2 at time 1, and no further messages. If
Agent 1 knows that the diameter of the network is 3, it will not know the maximum upon receiving
this message. However, without receiving any further messages, at time 3 Agent 1 will know that the
maximum is 100; no larger value can be lurking in the system.
1.2 KoP and Coordination
The fact that Max = c is a necessary condition for printing c is an essential feature of the CTM problem.
We have argued that, in fact, K1(Max = c) is also a necessary condition for printing c, as the KoP would
suggest. But this is just one instance. Let us briefly consider another example.
Example 2. Consider a bank whose ATMs are designed in such a way that an ATM will dispense cash
only to a customer whose account shows a sufficiently large positive balance. Along comes Alice, who
has a large positive balance, and tries to obtain a modest sum from the ATM. On this day, however,
the ATM is unable to communicate with the rest of the bank and it declines to pay Alice. Thus, despite
the fact that Alice has good credit, the ATM frustrates her and denies her request. Apparently, given
its specification, the ATM has no choice. Intuitively, in order to satisfy the credit restriction, the ATM
needs to know that a customer has good credit before dispensing cash. If the ATM may pay a customer
that is not known to have good credit, there will be possible scenarios in which the ATM will violate its
specification, and pay a customer that does not have credit. Notice, however, that the specification said
nothing about the ATM's knowledge. It only imposed a restriction on the ATM's action, based on the
(cid:3)
state of Alice's account.
Both the CTM problem and the ATM example are instances in which the KoP clearly applies. The
intuitive argument for why the KoP should apply very broadly is straightforward. If ϕ is a necessary
condition for performing α, and agent i ever performs α without knowing ϕ, then there should be a
possible scenario that is indistinguishable to agent i, in which ϕ does not hold. Since the two scenarios
are indistinguishable, the agent can perform α in the second scenario, and violate the requirement that ϕ
is a necessary condition. A formal statement and proof requires a definition of necessary conditions,
knowledge, as well as capturing a sense in which an action at one point implies the same action at any
other, indistinguishable, point. This will be done in Section 3.
Most tasks in distributed systems are described by way of a specification. Such specifications typ-
ically impose a variety of necessary conditions for actions. The KoP implies that even though such
specifications often do not explicitly discuss the agents' knowledge, they do in fact impose knowledge
preconditions. Observe that the KoP applies to a task regardless of the means that are used to implement
it. Any engineer implementing a particular task will have to ensure that preconditions are known when
actions are taken. This is true whether or not the engineer reasons explicitly in terms of knowledge, and
it is true even if the engineer is not even aware of the knowledge terminology. (Normally, neither may be
Yoram Moses
235
the case, of course.) The need to satisfy the KoP suggests that the design of distributed implementations
must involve at least two steps. One is to make sure that the required knowledge is made available to
an agent who needs to performed a prescribed action, and the other is ensuring that the agent detect
that it knows the required preconditions. This is quite different from common practice in engineering
distributed implementations [28].
We remark that the KoP can be expected to hold in a variety of multi-agent settings well beyond the
realm of distributed systems. Thus, for example, suppose that a jellyfish is naturally designed so that it
will never sting its own flesh. By the KoP, the cell activating the sting at a given point needs to know
that it is not stinging the jellyfish's body when it "fires" its sting. The jellyfish is thus designed with
some form of a "friend or foe" mechanism that is used in the course of activating the sting. Various
biological activities can similarly be considered in light of the KoP: How does the organism know that
certain preconditions are met? Our last example will come from the social science arena. Suppose that
a society designs a legal system, that is required to satisfy the constraint that only people who are guilty
of a particular crime are ever put in jail for committing this crime. By the KoP, the judge (or jury) must
know that the person committed this crime in order to send him to jail.
As discussed above, specifications impose preconditions. Typically, these conditions relate an action
to facts about the world (e.g., the maximal value, or the customer's good credit). In many cases, however,
actions of different agents need to be coordinated. Consider a variant of CTM in which in addition to
Agent 1 printing the maximum, Agent 4 needs to perform an action (say print the same value or print the
minimal value), but not before Agent 1 does. Then Agent 1 performing her action is a condition for 4's
action. In particular, Agent 4 would need to know that Agent 1 has already come to know Max = c for
some c before 4 acts. In some cases, the identity of actions performed needs to be coordinated.
For a final example, suppose that Alice should perform an action αA only if Bob performs an ac-
tion αB at least 5 time steps earlier. Then she needs to know that Bob acted at least 5 steps before
when she acts. Indeed, if ψ is a necessary condition for αB, then Alice must know that "Bob knew ψ at
least 5 time steps ago" when she acts, since knowing ψ is a necessary condition for Bob's performing αB
(see [4, 5]). As these examples illustrate, given KoP, coordination can give rise to nested knowledge.
Simple instances of the KoP are often quite straightforward. Ensuring and detecting K1(Max = c)
is often fairly intuitive, and it not justify the overhead involved in developing a theory of knowledge for
multi-agent systems. However, satisfying statements involving nested knowledge in particular models
of computation can quickly become nontrivial. For this, it is best to have a clear mathematical model of
knowledge in multi-agent systems. The next section reviews the runs and systems model.
2 Modeling Knowledge Using Runs and Systems
We now review the runs and systems model of knowledge of [10, 13]. The interested reader should
consult [10] for more details. A global state is an "instantaneous snapshot" of the system at a given time.
Let G denote a set of global states. Time will be identified with the natural numbers N = {0,1,2, . . .} for
ease of exposition. A run is a function r: N → G associating a global state with each instant of time.
Thus, r(0) is the run's initial state, r(1) is the next global state, and so on. A system is a set R of runs.
The same global state can appear in different runs, and in some systems may even appear more than once
in the same run.
A central notion in our framework is that of an agent's local state, whose role is to capture the agent's
local information at a given point. The precise details of the local state depend on the application. It could
be the complete contents of an agent's memory at the given instant, or the complete sequence of events
236
The Knowledge of Preconditions Principle
that it has observed so far. for example. The rule of thumb is that the local state should consist of the
local information that the agent may use when deciding which actions to take. Thus, for example, if
agents are finite-state machines, it is often natural to identify an agent's local state with the automaton
state that it is in. Formally, we assume that every global state determines a unique local state for each
agent. We denote agent i's local state in the global state r(t) by ri(t). Moreover, a global state with n
agents A = {1, . . . ,n} will have the form r(t) = (cid:104)re(t),r1(t), . . . ,rn(t)(cid:105), where re(t) is called the local
state of the environment, and will serve to represent all aspects of the global state that are not included in
the agents' local states. For example, it could represent messages in transit, the current topology of the
network including what links may be down, etc.
2.1 Syntax and Semantics
We are interested in a propositional logic of knowledge, in which propositional facts and epistemic facts
can be expressed. Facts will be considered to be true or false at a point (r,t), with respect to a system R.
More formally, given a set Φ of primitive propositions and a set P ={1, . . . ,n} of the agents in the system,
n (Φ) by closing Φ under negation '¬' and conjunction '∧', as well
we define a propositional language L K
as under knowledge operators Ki for all i ∈ P (see [14]). Thus, for example, if p,q ∈ Φ are primitive
propositions and i, j ∈ P are agents, then ¬Ki p∧ KjKi¬Kjq is a formula in L K
n (Φ). We typically omit
the set Φ and call L K
n the language for knowledge with n agents.
In a multi-agent system facts about the world, as well as the knowledge that agents have, can change
dynamically from one time point to the next. We thus consider the truth of formulas of L K
n at points of a
system R, where a point is a pair (r,t) ∈ R×N, and it is used to refer to time t in the run r. We denote the
set of points of a system R by Pts(R) (cid:44) R× N. Points will play the role of states of a Kripke structure.
The set Φ of primitive propositions used in the analysis of any given multi-agent system R will
depend on the application. Their truth at the points of the system needs to be explicitly defined. This is
done by an interpretation π : Φ× Pts(R) → {T,F}, where π(cid:0)q, (r,t)(cid:1) = T means that the proposition q
holds at (r,t). Formally, an interpreted system w.r.t. a set Φ of primitive propositions is a pair (R,π)
consisting of the system R and interpretation π for Φ over Pts(R). Just as we typically omit explicit
reference to Φ, we shall omit π as well, when this is unambiguous.
We assume from here on that the environment's state re(t) in a global state r(t) contains a "history"
component h that records all actions taken by all agents at times 0,1,. . . ,t − 1. Formally, we take h to be
a set of triples (cid:104)α,i,t(cid:48)(cid:105), which grows monotonically in time. An action α is considered to be performed
by i at the point (r,t) if and only if the triple (cid:104)α,i,t(cid:105), denoting that action α was performed by agent i at
time t, appears in the history component h of re(t(cid:48)) for all times t(cid:48) > t.1 For the analysis in this paper,
we will also assume that Φ includes propositions of the form doesi(α) and didi(α) for agents i ∈ P
and actions α. With this assumption, what actions are performed at any given point (r,t) is uniquely
determined by the run r.
π(cid:0)doesi(α), (r,t)(cid:1) = T
π(cid:0)didi(α), (r,t)(cid:1) = T
We will consider interpretations π that, on these propositions, are defined by
iff
agent i performs α at (r,t)
iff π(cid:0)doesi(α), (r,t(cid:48))(cid:1) = T holds for some t(cid:48) ≤ t
We allow t(cid:48) = t in the definition of didi(α) for technical convenience; it simplifies our later analysis
slightly.
1Our definition does not imply or assume that the actions are observed, observable or recorded by any of the agents. Whether
that is the case depends on the application.
Yoram Moses
237
Our model of knowledge will follow the standard Kripke-style possible worlds approach. The pos-
sibility relations that we use are induced directly from the system R being analyzed; two points are
considered indistinguishable to an agent if its local states at the two points are the same. More formally:
Definition 2.1. If ri(t) = r(cid:48)
i(t(cid:48)), then (r,t) and (r(cid:48),t(cid:48)) are called indistinguishable to i, denoted by
(r,t) ≈i (r(cid:48),t(cid:48)).
Formulae of L K
n are interpreted at a point (r,t) of an interpreted system (R,π) by means of the
satisfaction relation '=', which is defined inductively by:
(R,r,t) = p iff (r,t) ∈ π(p);
(R,r,t) = ¬ϕ iff (R,r,t) (cid:54)= ϕ;
(R,r,t) = ϕ ∧ ψ iff both (R,r,t) = ϕ and (R,r,t) = ψ;
(R,r,t) = Kiϕ iff (R,r(cid:48),t(cid:48)) = ϕ for all (r(cid:48),t(cid:48)) ∈ Pts(R) such that (r(cid:48),t(cid:48)) ≈i (r,t).
We say that ϕ is valid in the system R, and write R = ϕ, if (R,r,t) = ϕ for all points (r,t) ∈ Pts(R).
We say that ϕ validly implies ψ in R if ϕ ⇒ ψ is valid in R. Since, by Definition 2.1, the ≈i relations
are equivalence relations, each knowledge operator Ki satisfies the S5 axiom system [14]. In particular,
it satisfies the knowledge property (or Knowledge Axiom) that Kiϕ ⇒ ϕ is valid in all systems.
It is instructive to relate our modeling using runs and systems to standard multi-agent Kripke struc-
tures. For every system R there is a corresponding Kripke structure MR = (SR,π,∼1, . . . ,∼n) for n agents
such that SR = Pts(R) and '∼i' = '≈i' for every i. They correspond in that (R,r,t) = ϕ iff MR, (r,t) = ϕ
is guaranteed for all (r,t) ∈ Pts(R) = SR and ϕ ∈ L K
n (Φ) (for details, see [10]).
The system R will determine the space of possible runs and possible points, which play a crucial
role in determining the truth of facts involving knowledge. For example, consider a run r in which Alice
sends Bob a message at time 1, and Bob receives it at time 2. If R is a system in which messages may
be lost, or may take longer than one time step to be delivered, then Alice would not know at time 2(cid:0)i.e.,
w.r.t. (R,r,2)(cid:1) that her message has been delivered, because there is another run r(cid:48) ∈ R that she cannot
tell apart from r at time 2, in which her message is not (or not yet) delivered by that time. The same run r
also belongs to another system R(cid:48) in which messages are always reliably delivered in exactly one round.
With respect to (R(cid:48),r,2), however, Alice would know at time 2 that her message has been delivered.
Our definition of knowledge is rather flexible and widely applicable. The set R of the possible runs
immediately induces what the agents know. Observe that the definition of knowledge is completely
external. It ascribes knowledge to agents in the system even if the protocol they follow, as well as the
actions that they perform, do not involve the knowledge terminology in any way. Moreover, the agents
do not need to be complex or sophisticated for the definition to apply. Indeed, in a model of a very simple
system consisting of a bed lamp and its electric cable, a switch in the OFF state can be said to know that
the lamp is not lit; what the same switch would know in the ON state would depend on the system R under
consideration, which determines the runs considered possible. E.g., if R contains a run in which the lamp
is burnt out, then in the ON state the switch would not know that the lamp is shining light. On the other
hand, if the lamp can never burn out, and the cord, plug and switch are in proper working order in all runs
of R, then in the ON state the switch would know that the lamp is shining light. As this example shows,
knowledge under this definition does not require the "knower" to compute what it knows. Indeed, this
definition of knowledge is not sensitive to the computational complexity of determining what is known.
In most cases, of course, we will ascribe knowledge to agents or components that can perform actions,
which is not the case in the light switch example. And agents might need to explicitly establish whether
they know relevant facts. We now provide a statement and proof of the knowledge of preconditions
principle KoP.
238
The Knowledge of Preconditions Principle
3 Formalizing the Knowledge of Preconditions Principle
Intuitively, the KoP states that if a particular fact ψ is a necessary condition for an agent to perform
an action α, then the agent must in fact know ψ in order to act. In other words, knowing ψ is also a
necessary condition for performing the action. We formalize the claim and prove it as follows. We say
that ψ is a necessary condition for doesi(α) in R if (R,r,t) = doesi(α) holds only if (R,r,t) = ψ, for
all (r,t) ∈ Pts(R). Clearly, the customer's good credit is a necessary condition for the ATM dispensing
cash. That is, suppose that a bank makes use of a correct implementation of an ATM protocol, which
satisfies the credit requirement. Then, in the system R consisting of the set of all possible histories (runs)
of the bank's (and the ATM's) transactions, good credit is a necessary condition for receiving cash from
the ATM.
It is often of interest to consider facts whose truth depends only on a given agent's loca state. Such,
for example, may be the receipt of a message, or the observation of a signal, by the agent. Whether x = 0
for a local variable x, for example, would be a natural local fact. Moreover, if an agent has perfect recall,
then any events that it has observed in the past will give rise to local facts. Finally, since knowledge is
defined based on an agent's local state, then a fact of the form Kiϕ constitutes a local fact. Indeed, there
is a simple way to define the local facts above using knowledge. Namely, we say that ϕ is i-local in R if
R = (ϕ ⇒ Kiϕ).
The formalism of [10] defines protocols as explicit objects, and defines contexts that describe the
possible initial states and the model of computation. This provides a convenient and modular way of
constructing systems. Namely, given a protocol P and a context γ, the system R = R(P,γ) is defined to
be the set of all runs of protocol P in γ. The runs of this system embody all of the properties of the
context, as they arise in runs of P. This includes, for example, any timing assumptions, possible values
encountered, possible topologies of the network, etc. They also embody the relevant properties of the
protocol, because in all runs considered possible the agents follow P.
In this paper, we do not define protocols and contexts. Rather, we treat the KoP in a slightly simpler
and more abstract setting. We say that an action α is a conscious action for i in R if i's local state
completely determines whether i performs α. If its local state at two points (r,t) and (r(cid:48),t(cid:48)) of R is
the same, then (R,r,t) = doesi(α) iff (R,r(cid:48),t(cid:48)) = doesi(α). Conscious actions are quite prevalent in
many systems of interest. For example, suppose that agent i follows a deterministic protocol, so that its
action at any given point is a function of its local state. If, in addition, agent i is allowed to move at
every time step, then all of its actions are conscious actions. We remark that, since conscious actions
depend on an agent's local state, then if α is conscious for i in R then (R,r,t) = doesi(α) holds iff
(R,r,t) = Kidoesi(α) does, for all (r,t) ∈ Pts(R).
We are now ready to prove a formal version of the KoP:
Theorem 3.1 (The KoP Theorem). Let α be a conscious action for i in R. If ψ is a necessary condition
for doesi(α) in R, then Kiψ is also a necessary condition for doesi(α) in R.
Proof. We will show the contrapositive. Let α be a conscious action for i in R, and assume that Kiψ
is not a necessary condition for doesi(α) in R. Namely, there exists a point (r,t) ∈ Pts(R) such that
both (R,r,t) = doesi(α) and (R,r,t) (cid:54)= Kiψ. Given the latter, we have by the definition of '=' for Ki
that there exists a point (r(cid:48),t(cid:48)) ∈ Pts(R) such that both (r(cid:48),t(cid:48)) ≈i (r,t) and (R,r(cid:48),t(cid:48)) (cid:54)= ψ. Since α is a
conscious action for i in R and (R,r,t) = doesi(α) we have that (R,r,t) = Kidoesi(α). It follows from
(r(cid:48),t(cid:48)) ≈i (r,t) by the definition of '=' for Ki that (R,r(cid:48),t(cid:48)) = doesi(α) holds. But since (R,r(cid:48),t(cid:48)) (cid:54)= ψ,
we conclude that ψ is not a necessary condition for doesi(α) in R, establishing the countrapositive
claim.
Yoram Moses
239
Theorem 3.1 applies to all multi-agent systems. It immediately implies, for example, that a necessary
condition for the ATM to dispense cash is Katm(good credit). The theorem is model independent; it
does not depend on timing assumptions, on the topology of the system (even on whether agents commu-
nicate by message passing or via reading and writing to registers in a shared memory), or on the nature
of the activity that is carried out. For every necessary condition for a conscious action, knowing that the
condition holds is also a necessary condition.
4 Coordinating Simultaneous Actions
Recall that the language L K
n contains formulas in which knowledge operators can be nested to arbitrary
finite depth. It is sometimes useful to consider a state of knowledge called common knowledge that goes
beyond any particular nested formula. Intuitively, a fact ψ is common knowledge if everyone knowing
that everyone knows . . . , that everyone knows the fact ψ, to every finite depth. Common knowledge has
a number of equivalent definitions, one of which is as follows:
Definition 4.1 (Common Knowledge). Fix a set of agents G and a fact ψ. We denote by CGψ the fact
that ψ is common knowledge to G. Its truth at points of a system R is defined by:
(R,r,t) = Ki1Ki2 ···Kimψ holds for all (cid:104)i1,i2, . . . ,im(cid:105) ∈ Gm and all m ≥ 1.
(R,r,t) = CGψ iff
Common knowledge, a term coined by Lewis in [18], plays an important role in the analysis of
games [2], distributed systems [13], and many other multi-agent settings. Clearly, common knowledge
is much stronger than "plain" knowledge. Indeed, CGψ validly implies Kjψ, for all agents j ∈ G. Since
common knowledge requires infinitely many facts to hold, it is not a priori obvious that CGϕ can be
attained at a reasonable cost, or even whether it can ever be attained at all, in settings of interest (see [7,
10, 13]). We will now show that there are natural applications for which attaining common knowledge is
essential.
Intuitively, distinct actions are simultaneous in R if they can only be performed together; whenever
one is performed, all of them are performed simultaneously. It is possible to define simultaneous coordi-
nation formally in terms of necessary conditions:
Definition 4.2 (Simultaneous Actions). Let G be a set of agents. We say that a set of actions A = {αi}i∈G
is (necessarily) simultaneous in R if doesi(αi) is a necessary condition for does j(α j) in R, for all
i, j ∈ G.
Suppose that the actions in A are simultaneous in R in the above sense. Then the KoP immediately
implies (by Theorem 3.1) that a necessary condition for performing an action in A is knowing that the
other actions are also (currently) being performed. In fact, however, much more must be true. We now
present a strong variant of the KoP, which shows that in order to perform simultaneous actions agents
must attain common knowledge of their necessary conditions. Notice that in order to allow a set of
actions by the agents in G to be simultaneous, the system R must be sufficiently deterministic to ensure
that if i, j ∈ G are distinct agents and (R,r,t) = doesi(α) holds, then j will be scheduled to perform an
action at (r,t). For otherwise, there would be no way to ensure simultaneous execution of the actions by
the agents in G. Conscious actions fit this setting well in this case. We proceed as follows.
Theorem 4.3 (C-K of Preconditions). Let G be a set of agents and let A = {αi}i∈G be a set of necessarily
simultaneous actions in the system R. Moreover, suppose that each action αi ∈ A is a conscious action
for its agent i in R. If ψ is a necessary condition for doesi(αi) for some i ∈ G, then CGψ is a necessary
condition for does j(α j), for all j ∈ G.
240
The Knowledge of Preconditions Principle
Proof. Assume that A is a set of necessarily simultaneous actions for G in R. It is straightforward to
show the following claim.
Observation 1. Let αi,α j ∈ A be the actions for agents i and j, respectively. If a fact ϕ is a necessary
condition for doesi(αi) in R then ϕ is also a necessary condition for does j(α j) in R.
To prove this observation notice that, by assumption, both (a) R = does j(α j) ⇒ doesi(αi) and
(b) R = doesi(αi) ⇒ ϕ hold. For all (r,t) ∈ Pts(R), if (R,r,t) = does j(α j) then (R,r,t) = doesi(αi)
holds by (a) and so (R,r,t) = ϕ by (b). Thus, ϕ is a necessary condition for does j(α j) in R.
Assume that ψ is a necessary condition for doesi(αi), for some i ∈ G. We shall prove by induc-
tion on m ≥ 0 that Ki1Ki2 ···Kimψ is a necessary condition for does j(α j) in R, for every j ∈ G and all
sequences (cid:104)i1, . . . ,im(cid:105) ∈ Gm (of m agent names from G). This will establish that (R,r,t) = does j(α j)
implies (R,r,t) = CGψ for all (r,t) ∈ Pts(R), and thus CGψ is a necessary condition for does j(α j) for
all j ∈ G, as claimed.
• Base case: Let m = 0. The claim in this case is that if ψ is a necessary condition for doesi(αi)
then ψ is also a necessary condition for does j(α j). This is precisely Observation 1, with ϕ (cid:44) ψ.
• Inductive step: Let m ≥ 1, and assume that the claim holds for all j(cid:48) ∈ G and all sequences in Gm−1.
Fix j ∈ G and a sequence (cid:104)i1,i2, . . . ,im(cid:105) ∈ Gm. Its suffix (cid:104)i2, . . . ,im(cid:105) is a sequence in Gm−1. Thus,
Ki2 ···Kimψ is a necessary condition for doesi1(αi1) by the inductive hypothesis for m − 1 (ap-
plied to Gm−1 and agent j(cid:48) = i1 ∈ G). Given that αi1 is a conscious action by i1, we can apply
Theorem 3.1 to the necessary condition Ki2 ···Kimψ and obtain that Ki1Ki2 ···Kimψ is a necessary
condition for doesi1(αi1). By Observation 1 we have that Ki1Ki2 ···Kimψ is also a necessary con-
dition for does j(α j) in R, and we are done.
4.1 Common Knowledge and the Firing Squad Problem
As an illustration of the applicability of Theorem 4.3 to a concrete application, consider a simple version
of the Firing Squad problem. In this instance, the set of agents G in the system must simultaneously
perform an action (say each agent i ∈ G should perform the action firei) in response to the receipt, by
any agent in G, of a particular external input called a 'go' message. The firei action can stand for a
simultaneous change in shared copies of a database, a public announcement at different sites of the
system, or any other actions that need to take place simultaneously. Moreover, firei actions are allowed
only if they are preceded by such a go message. For simplicity, we consider a case in which none of the
agents in G may fail, and they all must satisfy the specification.
Let ψgo be a proposition that is true at (r,t) ∈ Pts(R) if a go message is received by any of the agents
in G at a point (r,t(cid:48)) of r at a time t(cid:48) ≤ t. According to the specification of the Firing Squad problem, ψgo
is a necessary condition for the firei actions. An immediate consequence of Theorem 4.3 is:
Corollary 4.4. CGψgo is a necessary condition for all firei actions in the Firing Squad problem.
Given Corollary 4.4, any solution to the firing squad problem must first attain common knowledge
that a go message has been received. It is well-known (see [11, 13]) that common knowledge of a fact
is observed simultaneously at all agents it involves. Suppose that every i ∈ G performs firei when CGψgo
first holds. Since all agents in G will come to know that CGψgo immediately, they will fire simultaneously,
as required by the problem specification. Indeed, Theorem 4.3 shows that this is the first time at which
they can perform according to a correct protocol. Implementing simultaneous tasks such as the Firing
Yoram Moses
241
Squad therefore inherently involves, and often reduces to, ensuring and detecting CGψgo. Recall that
depending on the properties of the system, attaining such common knowledge might be impossible in
some cases, or it might incur a substantial cost in others. Just as in the case of the KoP, this necessity is
not due to our formalism. It is only exposed by our analysis. In every protocol that implements such a
task correctly, the firing actions cannot be performed unless CGψgo is attained.
There is an extensive literature on using common knowledge to obtain optimal protocols for simul-
taneous tasks [8, 9, 16, 20, 21, 23, 24, 26, 25]. Typically, they involve an explicit proof that common
knowledge of a particular fact is a necessary condition for performing a set A of necessarily simultaneous
actions. Theorem 4.3 or a variant of it suited for fault-tolerant systems can be used to establish this result
in all of these cases. Moreover, one of the main insights from the analysis of [13] and of [11] is that
when simultaneous actions are performed, the participating agents have common knowledge that they
are being performed. Theorem 4.3 is a strict generalization of this fact.
5 Temporally Ordering Actions
So far, we have seen two essential connections between knowledge and coordinated action: performing
actions requires knowledge of their necessary conditions, and performing simultaneous actions requires
common knowledge of their necessary conditions. We now further extend the connection between states
of knowledge and coordination, by showing that temporally ordering actions depends on attaining nested
knowledge of necessary conditions. Following [5], we define temporally ordered actions:
Definition 5.1 (Ben Zvi and Moses). A sequence of actions (cid:104)α1, . . . ,αk(cid:105) (for agents 1, . . . ,k, respectively)
is (linearly) ordered in R if did j−1(α j−1) is a necessary condition for does j(α j) in R.
Observe that this definition does not force an action α j to occur in a run in which α j−1 occurs.
Rather, if an action α j is performed in a given run, then it must be preceded by all actions α1, . . . ,α j−1.
Moreover, if we denote the time at which an action αi is performed in a run r by ti, then we require that
t j−1 ≤ t j for every action α j performed in r.
Claim 1. Assume that the sequence (cid:104)α1, . . . ,αk(cid:105) is ordered in R. Then R =(cid:0)did j(α j) ⇒ did j−1(α j−1)(cid:1)
for all 2 ≤ j ≤ k.
Proof. Assume that (R,r,t) = did j(α j). Then, by definition of did j(α j), we have (R,r, t ) = does j(α j)
for some t ≤ t. The fact that (cid:104)α1, . . . ,αk(cid:105) is ordered in R implies that did j−1(α j−1) is a necessary
condition for does j(α j) in the system R, and so (R,r, t ) = did j−1(α j−1). Since did j−1(α j−1) is a
stable fact and t ≥ t, we obtain that (R,r,t) = did j−1(α j−1). The claim follows.
We say that a fact ϕ is stable in R if once true, ϕ remains true. Formally, if (R,r,t) = ϕ and t(cid:48) > t
then (R,r,t(cid:48)) = ϕ, for all r ∈ R and t,t(cid:48) ≥ 0. Notice that while doesi(α) is, in general, not a stable fact,
didi(α) is always stable.
Definition 5.2. We say that agent i recalls ψ in R if the fact Kiψ is stable in R.
The notion of perfect recall, capturing the assumption that agents remember all events that they take
part in, is popular in the analysis of games and multi-agent systems [10, 29]. While perfect recall is
a nontrivial assumption often requiring significant storage costs, selective recall of single facts such as
does j(α j) is a much weaker assumption, that can be assumed of a system R essentially without loss of
generality. By adding a single bit to Agent j's local state, whose value is 0 as long as j has not performed
α j and 1 once the action has been performed, we can obtain a system R(cid:48) that is isomorphic to R, in which
Agent j recalls does j(α j).
242
The Knowledge of Preconditions Principle
Claim 2. Assume that α j is a conscious action for j in R, and that j recalls did j(α j) in R. Then
did j(α j) is a j-local fact in R.
Proof. Suppose that (R,r,t) = did j(α j). Then, by definition of did j(α j), we have (R,r, t ) = does j(α j)
for some t ≤ t. Choose an arbitrary (r(cid:48),t(cid:48)) ∈ Pts(R) satisfying that (r(cid:48),t(cid:48)) ≈ j (r, t ).
It follows that
(R,r(cid:48),t(cid:48)) = does j(α j) since α j is a conscious action for j in R. By definition of did j(α j) it follows
that (R,r(cid:48),t(cid:48)) = did j(α j). Now, by definition of = for Kj we have that (R,r, t ) = Kjdid j(α j). By
assumption, j recalls did j(α j) in R, and so Kjdid j(α j) is stable in R. Thus, since t ≥ t, we obtain that
(R,r,t) = Kjdid j(α j), as claimed.
We can now show:
Theorem 5.3 (Ordering and Nested Knowledge). Assume that
• the actions (cid:104)α1, . . . ,αk(cid:105) are ordered in R,
• each agent j = 1, . . . ,k recalls did j(α j) in R,
• α j is a conscious action for j in R, for all j = 1, . . . ,k, and
• ψ is a stable necessary condition for the first action does1(α1) in R
Then KjKj−1···K1ψ is a necessary condition for the j th action does j(α j) in R, for all j ≤ k.
Proof. Assuming the conditions of the theorem, we will prove by induction on j ≤ k that did j(α j)
validly implies KjKj−1···K1ψ in R. Since does j(α j) validly implies did j(α j) by definition of did j(α j),
this will yield that KjKj−1···K1ψ is a necessary condition for does j(α j) in R, as claimed. We proceed
with the inductive argument.
• Base case j = 1: Assume that (R,r,t) = did1(α1). Claim 2 implies that (R,r,t) = K1did1(α1).
Let (r(cid:48),t(cid:48)) ∈ Pts(R) be an arbitrary point satisfying that (r(cid:48),t(cid:48)) ≈1 (r,t). Then (R,r(cid:48),t(cid:48)) = did1(α1)
by the knowledge property. Thus, (R,r(cid:48), t ) = does1(α1) holds for some t ≤ t(cid:48), and because ψ is a
necessary condition for does1(α1) in R, we obtain that (R,r, t ) = ψ. Since ψ is stable and t(cid:48) ≥ t,
we have that (R,r(cid:48),t(cid:48)) = ψ. By choice of (r(cid:48),t(cid:48)) we have that (R,r,t) = K1ψ, as claimed.
• Inductive step: Let j > 1 and assume that Kj−1···K1ψ is a necessary condition for did j−1(α j−1)
in R. Moreover, let (R,r,t) = did j(α j). Since α j is a conscious action for j, Claim 2 implies
that (R,r,t) = Kjdid j(α j). Choose an arbitrary (r(cid:48),t(cid:48)) ∈ Pts(R) satisfying that (r(cid:48),t(cid:48)) ≈ j (r,t). By
definition of Kj, it follows that (R,r(cid:48),t(cid:48)) = did j(α j). By Claim 1, since the sequence (cid:104)α1, . . . ,αk(cid:105)
is ordered in R and j > 1 we have that (R,r(cid:48),t(cid:48)) = did j−1(α j−1). We now apply the inductive
hypothesis to obtain that (R,r(cid:48),t(cid:48))= Kj−1···K1ψ. Finally, we obtain that (R,r,t)= KjKj−1···K1ψ
by choice of (r(cid:48),t(cid:48)) and the definition of '=' for Kj. The claim now follows.
A slightly more restricted version of Theorem 5.3 was proved in [5]. Rather than consider an arbitrary
necessary condition for α1, they proved a version for the case in which the first action α1 is triggered by
an external input to agent 1. Technically, the proofs are quite similar.
Theorem 5.3 provides a necessary, but possibly not sufficient, condition for ordering actions in dis-
tributed systems. If agent j acts strictly later than when KjKj−1···K1ψ first holds, then it may be inappro-
priate for agent j + 1 to act when it knows that the fact KjKj−1···K1ψ holds (i.e., when Kj+1Kj ···K1ψ
first holds). Nevertheless, Theorem 5.3 is often very useful because it can be used as a guide for effi-
ciently, and sometimes even optimally, performing a sequence of ordered actions. Intuitively, suppose
Yoram Moses
243
that we have a protocol whose goal is to perform (cid:104)α1, . . . ,αk(cid:105) in response to an externally generated
trigger ψ (such as the 'go' message in Firing Squad). In particular, assume that ψ is a necessary condi-
tion for α1. Keeping the communication aspects of this protocol fixed, an optimally fast solution would
be for each agent j ≤ k to perform α j when KjKj−1···K1ψ first holds. Let R be the set of runs of
such a protocol with r ∈ R, and let t j and t j−1 be the earliest times at which (R,r,t j) = KjKj−1···K1ψ
and (R,r,t j−1) = Kj−1···K1ψ hold in a run r, respectively. The knowledge property guarantees that
KjKj−1···K1ψ validly implies that Kj−1···K1ψ in R, and so t j ≥ t j−1. Since, by assumption, α j is
performed at time t j and α j−1 at t j−1, we have that agents perform actions in linear temporal order, as
required by Definition 5.1. Clearly, none of the actions can be performed any earlier, as Theorem 5.3
shows. We conclude that in time-efficient protocols, the nested knowledge formula presented by the
theorem can be both necessary and sufficient. In this sense, Theorem 5.3 suggests a recipe for obtaining
time-efficient solutions for ordering actions.
Just as Theorem 4.3 implies that common knowledge is a necessary condition for simultaneous ac-
tions, we now have by Theorem 5.3 that nested knowledge is a necessary condition for performing actions
in linear temporal order. And just as there is an established literature on when common knowledge is
and is not attainable and on how it may arise, there are results concerning the communication structure
that underlies attaining nested knowledge. Indeed, in a seminal paper [7], Chandy and Misra showed that
in asynchronous systems R, if (R,r,t) = ¬ϕ and at a time t(cid:48) > t (R,r,t(cid:48)) = KjKj−1···K1ϕ, then there
must be a message chain in the run r between times t and t(cid:48), passing through the agents 1,2,. . . , j in this
order (possibly involving additional agents as well). Given Theorem 5.3, this implies that the only way
to coordinate actions in a linear temporal order in an asynchronous setting is by way of such message
chains.2
More recently, Ben Zvi and Moses extended Chandy and Misra's work to systems in which commu-
nication is not asynchronous, but rather agents may have access to clocks and the transmission time for
each of the channels is bounded [5]. They show that a communication structure called a centipede must
be constructed in order to obtain nested knowledge of spontaneous facts such as the arrival of an external
input. They prove a slightly more restricted instance of Theorem 5.3 (without using KoP directly), and
use it to show that ordering actions in their setting requires the construction of the appropriate centipedes.
Finally, Parikh and Krasucki analyze the ability to create levels of knowledge consisting of collections of
nested knowledge formulas in [27]. Theorem 5.3 relates levels of knowledge to coordination.
6 Discussion
This paper formulated the knowledge of preconditions principle and presented three theorems relating
knowledge and coordinated action: the first is the KoP itself -- necessary conditions for an action must be
known to hold when the action is performed. Next, we showed that necessary conditions for simultaneous
actions must be commonly known when the actions are taken. Finally, nested knowledge is a necessary
condition for coordinating linearly ordered actions. The latter two are fairly direct consequences of the
KoP. We discussed some of the uses of the latter two results in Sections 4 and 5. Indeed the KoP has
many further implications.
In recent years, several works that make use of KoP have appeared, citing the unpublished [22].
For example, Castaneda, Gonczarowski and Moses used the KoP to analyze the consensus problem [6],
2Theorems 3.1 and 5.3 depend on conscious actions and therefore do not apply to asynchronous systems. Nevertheless,
variants of these theorems can be presented that do apply to asynchronous systems and nondeterministic protocols. Details will
appear in [22].
244
The Knowledge of Preconditions Principle
in which agents need to agree on a binary value in a fault-prone system. They designed a protocol in
two steps -- applying the KoP once to derive a rule by which, roughly, agents decide on 0 when they
know of an initial value of 0. Then, they applied the KoP again assuming that this is the rule used for
making decisions on 0, and obtained a rule involving nested knowledge (roughly, a statement of the
form "knowing that nobody knows 0") for deciding on the value 1. The result of their analysis was a
very efficient solution to consensus that is optimal in a strong sense: It is the first unbeatable consensus
protocol. No protocol can strictly dominate it, by having processes always decide at least as fast, and
sometimes strictly faster, than this protocol does. The work of [6] complements an earlier work by
Halpern, Moses and Waarts [15], in which a fixed point analysis of optimal consensus was obtained. The
latter, too, is closely related to the KoP.
Gonczarowski and Moses used the KoP to analyze the epistemic requirements of more general forms
of coordination [12]. Namely, they considered a setting in which k agents need to perform actions, and
there are time bounds on the relative times at which the actions of any pair of agents is performed. The
simple instance in which all bounds are 0 is precisely that of the simultaneous actions considered in Sec-
tion 4. They show that such coordination requires vectorial fixed points of knowledge conditions, which
are naturally related to fixed points and equilibria. The papers [3, 4, 5, 12] together can all be viewed
as making use of the KoP to provide insights into the interaction between time and communication for
coordinating actions in a distributed and multi-agent system. Describing them is beyond the scope of the
current paper.
The most significant aspect of the KoP, in our view, is the fact that it places a new emphasis on
the epistemic aspects of problem solving in a multi-agent system. Simple necessary conditions induce
epistemic conditions. Thus, in order to act correctly, one needs a mechanism ensuring that the agents
obtain the necessary knowledge, and that they discover that they have this knowledge. Most problems and
solutions are not posed or described in this fashion. We believe that the KoP encapsulates an important
connection between knowledge, action and coordination that will find many applications in the future.
References
[1] Hagit Attiya & Jennifer Welch (2004): Distributed Computing: Fundamentals, Simulations and Advanced
Topics. John Wiley & Sons, doi:10.1002/0471478210.
Agreeing to disagree.
J. Aumann (1976):
[2] R.
doi:10.1214/aos/1176343654.
Annals of Statistics 4(6), pp. 1236 -- 1239,
[3] Ido Ben-Zvi & Yoram Moses (2013): Agent-Time Epistemics and Coordination. In: Proceedings of ICLA,
pp. 97 -- 108, doi:10.1007/978-3-642-36039-8 9.
[4] Ido Ben-Zvi & Yoram Moses (2013): The Shape of Reactive Coordination Tasks. In: Proceedings of TARK,
TARK XIV, pp. 29 -- 38.
[5] Ido Ben-Zvi & Yoram Moses (2014): Beyond Lamport's Happened-before: On Time Bounds and the Order-
ing of Events in Distributed Systems. J. ACM 61(2), p. 13, doi:10.1145/2542181.
[6] Armando Castaneda, Yannai A Gonczarowski & Yoram Moses (2014): Unbeatable Consensus. In: Proceed-
ings of DISC, Springer, pp. 91 -- 106, doi:10.1007/978-3-662-45174-8 7.
[7] K. M. Chandy & J. Misra (1986): How processes learn. Distributed Computing 1(1), pp. 40 -- 52,
doi:10.1007/BF01843569.
[8] Danny Dolev, Ezra N Hoch & Yoram Moses (2012): An optimal self-stabilizing firing squad. SIAM Journal
on Computing 41(2), pp. 415 -- 435, doi:10.1137/090776512.
[9] C. Dwork & Y. Moses (1990): Knowledge and common knowledge in a Byzantine environment: crash fail-
ures. Information and Computation 88(2), pp. 156 -- 186, doi:10.1016/0890-5401(90)90014-9.
Yoram Moses
245
[10] R. Fagin, J. Y. Halpern, Y. Moses & M. Y. Vardi (2003): Reasoning about Knowledge. MIT Press, Cambridge,
Mass.
[11] Ronald Fagin, Joseph Y. Halpern, Yoram Moses & Vardi. Moshe Y. (1999): Common knowledge revisited.
Annals of Pure and Applied Logic 96(13), doi:10.1016/S0168-0072(98)00033-5.
[12] Y Gonczarowski & Y Moses (2013): Timely common knowledge: Characterising asymmetric distributed
coordination via vectorial fixed points. In: Proceedings of TARK XIV.
[13] J. Y. Halpern & Y. Moses (1990): Knowledge and Common Knowledge in a Distributed Environment. Journal
of the ACM 37(3), pp. 549 -- 587, doi:10.1145/800222.806735. A preliminary version appeared in Proc. 3rd
ACM PODC, 1984.
[14] J. Y. Halpern & Y. Moses (1992): A guide to completeness and complexity for modal logics of knowledge and
belief. Artificial Intelligence 54, pp. 319 -- 379, doi:10.1016/0004-3702(92)90049-4.
[15] Joseph Y. Halpern, Yoram Moses & Orli Waarts (2001): A Characterization of Eventual Byzantine Agree-
ment. SIAM J. Comput. 31(3), pp. 838 -- 865, doi:10.1137/S0097539798340217.
[16] Maurice P Herlihy, Yoram Moses & Mark R Tuttle (2011): Transforming worst-case optimal solutions for si-
multaneous tasks into all-case optimal solutions. In: Proceedings of the 30th annual ACM SIGACT-SIGOPS
symposium on Principles of distributed computing, ACM, pp. 231 -- 238, doi:10.1145/1993806.1993849.
[17] G´erard Le Lann (1977): Distributed Systems-Towards a Formal Approach. In: IFIP Congress, 7, Toronto,
pp. 155 -- 160.
[18] D. Lewis (1969): Convention, A Philosophical Study. Harvard University Press, Cambridge, Mass.
[19] J. McCarthy & P. J. Hayes (1969): Some Philosophical Problems From the Standpoint of Artificial Intelli-
gence. In: Machine Intelligence 4, Edinburgh University Press, pp. 463 -- 502, doi:10.1016/b978-0-934613-
03-3.50033-7.
[20] Tal Mizrahi & Yoram Moses (2008): Continuous consensus via common knowledge. Distributed Computing
20(5), pp. 305 -- 321, doi:10.1007/s00446-007-0049-6.
[21] Tal Mizrahi & Yoram Moses (2008): Continuous consensus with ambiguous failures. In: Distributed Com-
puting and Networking, Springer, pp. 73 -- 85, doi:10.1016/j.tcs.2010.04.025.
[22] Y Moses (2016): Knowledge and Coordinated Action, to appear.
[23] Y. Moses & M. R. Tuttle (1988): Programming simultaneous actions using common knowledge. Algorithmica
3, pp. 121 -- 169, doi:10.1007/BF01762112.
[24] G. Neiger (1990): Consistent coordination and continual common knowledge. Manuscript.
[25] G. Neiger & M. R. Tuttle (1993): Common knowledge and consistent simultaneous coordination. Distributed
Computing 6(3), pp. 334 -- 352, doi:10.1007/BF02242706.
[26] Gil Neiger & Rida A Bazzi (1999): Using knowledge to optimally achieve coordination in distributed sys-
tems. Theoretical computer science 220(1), pp. 31 -- 65, doi:10.1016/S0304-3975(98)00236-9.
[27] R. Parikh & P. Krasucki (1992): Levels of knowledge in distributed computing. S¯adhan¯a 17(1), pp. 167 -- 191,
doi:10.1007/bf02811342.
[28] Fred B Schneider (1990): Implementing fault-tolerant services using the state machine approach: A tutorial.
ACM Computing Surveys (CSUR) 22(4), pp. 299 -- 319, doi:10.1145/98163.98167.
[29] R. Selten (1975): Reexamination of the perfectness concept for equilibrium points in extensive games. Inter-
national Journal of Game Theory 4, pp. 25 -- 55, doi:10.1145/2542181.
|
cs/0308030 | 1 | 0308 | 2003-08-19T15:45:16 | Learning in Multiagent Systems: An Introduction from a Game-Theoretic Perspective | [
"cs.MA",
"cs.AI"
] | We introduce the topic of learning in multiagent systems. We first provide a quick introduction to the field of game theory, focusing on the equilibrium concepts of iterated dominance, and Nash equilibrium. We show some of the most relevant findings in the theory of learning in games, including theorems on fictitious play, replicator dynamics, and evolutionary stable strategies. The CLRI theory and n-level learning agents are introduced as attempts to apply some of these findings to the problem of engineering multiagent systems with learning agents. Finally, we summarize some of the remaining challenges in the field of learning in multiagent systems. | cs.MA | cs |
In Eduardo Alonso, editor, Adaptive Agents: LNAI 2636. Springer Verlag, 2003.
Learning in Multiagent Systems: An
Introduction from a Game-Theoretic Perspective
Jos´e M. Vidal
University of South Carolina, Computer Science and Engineering,
Columbia, SC 29208
[email protected]
Abstract. We introduce the topic of learning in multiagent systems. We
first provide a quick introduction to the field of game theory, focusing
on the equilibrium concepts of iterated dominance, and Nash equilib-
rium. We show some of the most relevant findings in the theory of learn-
ing in games, including theorems on fictitious play, replicator dynamics,
and evolutionary stable strategies. The CLRI theory and n-level learning
agents are introduced as attempts to apply some of these findings to the
problem of engineering multiagent systems with learning agents. Finally,
we summarize some of the remaining challenges in the field of learning
in multiagent systems.
1 Introduction
The engineering of multiagent systems composed of learning agents brings to-
gether techniques from machine learning, game theory, utility theory, and com-
plex systems. A designer must choose carefully which machine-learning algorithm
to use since otherwise the system’s behavior will be unpredictable and often un-
desirable. Fortunately, we can use the tools from these areas in an effort to
predict the expected system behaviors. In this article we introduce these tech-
niques and explain how they are used in the engineering of learning multiagent
systems.
The goal of machine learning research is the development of algorithms that
increase the ability of an agent to match a set of inputs to their corresponding
outputs [7]. That is, we assume the existence of a large, sometimes infinite, set
of examples E . Each example e ∈ E is a pair e = {a, b} where a ∈ A represents
the input the agent receives and b ∈ B is the output the agent should produce
when receiving this input. The agent must find a function f which maps A → B
for as many examples of A as possible. In a controlled test the set E is usually
first divided into a training set which is used for training the agent, and a testing
set which is used for testing the performance of the agent. In some scenarios it
is impossible to first train the agent and then test it. In these cases the training
and testing examples are interleaved. The agent’s performance is assessed on an
ongoing manner.
When a learning agent is placed in a multiagent scenario these fundamental
assumptions of machine learning are violated. The agent is no longer learning
c(cid:13) Springer Verlag, 2003.
to extrapolate from the examples it has seen of fixed set E , instead it’s target
concept keeps changing, leading to a moving target function problem [10]. In
general, however, the target concept does not change randomly; it changes based
on the learning dynamics of the other agents in the system. Since these agents
also learn using machine learning algorithms we are left with some hope that
we might someday be able to understand the complex dynamics of these type of
systems.
Learning agents are most often selfish utility maximizers. These agents often
face each other in encounters where the simultaneous actions of a set of agents
leads to different utility payoffs for all the participants. For example, in a market-
based setting a set of agents might submit their bids to a first-price sealed-bid
auction. The outcome of this auction will result in a utility gain or loss for all
the agents. In a robotic setting two agents headed in a collision course towards
each other have to decide whether to stay the course or to swerve. The results
of their combined actions have direct results in the utilities the agents receive
from their actions. We are solely concerned with learning agents that maximize
their own utility. We believe that systems where agents share partial results or
otherwise help each other can be considered extension on traditional machine
learning research.
2 Game Theory
Game theory provides us with the mathematical tools to understand the possible
strategies that utility-maximizing agents might use when making a choice. It
is mostly concerned with modeling the decision process of rational humans, a
fact that should be kept in mind as we consider its applicability to multiagent
systems.
The simplest type of game considered in game theory is the single-shot
simultaneous-move game. In this game all agents must take one action. All ac-
tions are effectively simultaneous. Each agent receives a utility that is a function
of the combined set of actions. In an extended-form game the players take turns
and receive a payoff at the end of a series of actions. A single-shot game is a
good model for the types of situations often faced by agents in a multiagent
system where the encounters mostly require coordination. The extended-form
games are best suited to modeling more complex scenarios where each succes-
sive move places the agents in a different state. Many scenarios that first appear
like they would need an extended-form game can actually be described by a se-
ries of single-shot games. In fact, that is the approach taken by many multiagent
systems researchers.
In the one-shot simultaneous-move game we say that each agent i chooses a
strategy si ∈ Si , where Si is the set of all strategies for agent i. These strategies
represent the actions the agent can take. When we say that i chooses strategy
si we mean that it chooses to take action si . The set of all strategies chosen by
all the agents is the strategy profile for that game and it is denoted by s ∈ S ≡
×I
i=iSi . Once all the agents make their choices and form the strategy profile s
A B
A 1,2 3,4
B 3,2 2,1
Fig. 1. Sample two-player game matrix. Agent 1 chooses from the rows and
agent 2 chooses from the columns.
then each agent i receives a utility which is given by the function ui (s). Notice
that a player’s utility depends on the choices made by all the agents.
Two player games involve only two players, i and j . They are often repre-
sented using a game matrix such as the one shown in Figure 1. In that matrix
we see that if agent 1 (the one who chooses from the rows) chooses action A and
agent 2 chooses action B then agent 1 will receive a utility of 3 while agent 2
receives a utility of 4. Using our notation for strategies we would say that if the
strategy profile is (s1 , s2 ) then the payoff vector is
(u1 (s1 , s2 ), u2 (s1 , s2 ))
It is possible that a player will choose randomly between its action choices,
using different prior probabilities for each choice. These types of strategies are
called mixed strategies and they are a probability distribution over an agent’s
actions. We say that a mixed strategy for agent i is σi ∈ Σi ≡ P (Si ) where
P (Si ) is the set of all probability distributions over the set of pure strategies
Si . Although a real agent can not take a “mixed action”, mixed strategies are
useful abstractions since they allow us to model agents who might use some
randomization subroutine to choose their action.
3 Solution Concepts
Much of the work in game theory has concentrated in the definition of plausible
solution concepts. A solution concept tries to define the set of actions that a
set of rational agents will choose when faced with a game. The most common
assumptions are that the agents are rational, have common knowledge1 of the
payoffs in the game matrix, and that they are intelligent enough to re-create
the thought process that the mathematician went through to come up with the
solution concept. As such, most solution concepts are geared towards an under-
standing of how smart, well-informed people would act. They are not necessarily
meant to explain the behavior of machine-learning agents. Still, the fact that
they provide the “best” solution makes them a useful tool.
3.1 Iterated Dominance
The iterated dominance approach is to successively eliminate from consideration
those actions that are worst than some other action, no matter what the other
1 Common knowledge about p means that everybody knows that everybody knows,
and so on to infinity, about p.
A B
A 8,2 9,4
B 1,2 3,1
Fig. 2. A game where agent 1’s action B is dominated by A.
player does. For example, in Figure 2 we see a game where agent 1’s action B
is dominate by A. That is, no matter what agent 2 does, agent 1 should choose
action A. Then, if agent 1 chooses action A then agent 2 should choose action
B. Therefore, the solution strategy profile for this game is (A, B ).
Formally, we say that a strategy σi is strictly dominated for agent i if there
is some other strategy σi ∈ Σi for which ui (σi , σ−i ) > ui (σi , σ−i ) for all σ−i ,
where σ−i is a set of strategies for all agents except i. Notice that the inequality
sign is a greater-than. If we change that sign to a greater-than-or-equal then we
have the definition for a weakly dominated strategy.
There is no reason for a rational agent to choose a strictly dominated strategy.
That is, there is no reason for an agent to choose σi when there exists a σi which
will give it a better utility no matter what the other agents do. Similarly, there
is no reason for the agent to choose a weakly dominated strategy. Of course,
this reasoning relies on the assumption that the agent can indeed determine the
existence of a σi . This assumption can be hard to justify in cases where the
better strategy is a mixed strategy where the agent has an infinite number of
possible strategies to verify, or in cases where the number of actions and agents
is too large to handle.
The iterated dominance algorithm consists of calculating all the strategies
that are dominated for all the players, eliminating those strategies from consid-
eration, and repeating the process until no more strategies are dominated. At
that point it might be the case that only one strategy profile is left available. In
this case that profile is the one all agents should play. However, in many cases
the algorithm still leaves us with a sizable game matrix with a large number of
possible strategy profiles. The algorithm then serves only to reduce the size of
the problem.
3.2 Nash Equilibrium
The Nash equilibrium solution concept is popular because it provides a solution
where other solution concepts fail. The Nash equilibrium strategy profile is de-
fined as σ such that for all agents i it is true that there is no strategy better than
σi given that all the other agents take the actions prescribed by σ−i . Formally,
we say that σ is a Nash equilibrium strategy profile if for all i it is true that
σi ∈ BRi ( σ−i ), where BRi (s−i ) is the best response for i to s−i . That is, given
that everyone else plays the strategy given by the Nash equilibrium the best
strategy for any agent is the one given by the Nash equilibrium. A strict Nash
equilibrium states that σi is strictly (i.e., greater than) better than any other
alternative.
It has been shown that every game has at least one Nash equilibrium, as
long as mixed strategies are allowed. The Nash equilibrium has the advantage
of being stable under single agent desertions. That is, if the system is in a Nash
equilibrium then no agent, working by itself, will be tempted to take a different
action. However, it is possible for two or more agents to conspire together and find
a set of actions which are better for them. This means that the Nash equilibrium
is not stable if we allow the formation of coalitions.
Another problem we face when using the Nash equilibrium is the fact that
a game can have multiple Nash equilibria. In these cases we do not know which
one will be chosen, if any. The Nash equilibrium could also be a mixed strategy
for some agent while in the real world the agent has only discrete actions avail-
able. In both of these cases the Nash equilibrium is not sufficient to identify a
unique strategy profile that rational agents are expected to play. As such, fur-
ther studies of the dynamics of the system must be carried out in order to refine
the Nash equilibrium solution. The theory of learning in games—a branch of
game theory—has studied how simple learning mechanisms lead to equilibrium
strategies.
4 Learning in Games
The theory of learning in games studies the equilibrium concepts dictated by var-
ious simple learning mechanisms. That is, while the Nash equilibrium is based
on the assumption of perfectly rational players, in learning in games the as-
sumption is that the agents use some kind of algorithm. The theory determines
the equilibrium strategy that will be arrived at by the various learning mecha-
nisms and maps these equilibria to the standard solution concepts, if possible.
Many learning mechanisms have been studied. The most common of them are
explained in the next few sub-sections.
4.1 Fictitious Play
A widely studied model of learning in games is the process of fictitious play. In it
agents assume that their opponents are playing a fixed strategy. The agents use
their past experiences to build a model of the opponent’s strategy and use this
model to choose their own action. Mathematicians have studied these types of
games in order to determine when and whether the system converges to a stable
strategy.
Fictitious play uses a simple form of learning where an agent remembers
everything the other agents have done and uses this information to build a
probability distribution for the other agents’ expected strategy. Formally, for the
two agent (i and j ) case we say that i maintains a weight function ki : Sj → R+ .
The weight function changes over time as the agent learns. The weight function
at time t is represented by k t
i which keeps a count of how many times each
strategy has been played. When at time t − 1 opponent j plays strategy st−1
j
A B
A 0,0 1,1
B 1,1 0,0
Fig. 3. A game matrix with an infinite cycle.
then i updates its weight function with
(sj ) + (cid:26) 1 if st−1
j = sj ,
0 if st−1
6= sj .
j
Using this weight function, agent i can now assign a probability to j playing
any of its sj ∈ Sj strategies with
i (sj ) = k t−1
k t
i
(1)
.
(2)
Prt
i [sj ] =
k t
i (sj )
k t
Psj ∈Sj
i (sj )
Player i then determines the strategy that will give it the highest expected
utility given that j will play each of its sj ∈ Sj with probability Prt
i [sj ]. That
is, i determines its best response to a probability distribution over j ’s possible
strategies. This amounts to i assuming that j ’s strategy at each time is taken
from some fixed but unknown probability distribution.
Several interesting results have been derived by researches in this area. These
results assume that all players are using fictitious play. In [3] it was shown that
the following two propositions hold.
Proposition 1. If s is a strict Nash equilibrium and it is played at time t then
it wil l be played at al l times greater than t.
Intuitively we can see that if the fictitious play algorithm leads to all players
to play the same Nash equilibrium then, afterward, they will increase the prob-
ability that all others are playing the equilibrium. Since, by definition, the best
response of a player when everyone else is playing a strict Nash equilibrium is to
play the same equilibrium, all players will play the same strategy and the next
time. The same holds true for every time after that.
Proposition 2. If fictitious play converges to a pure strategy then that strategy
must be a Nash equilibrium.
We can show this by contradiction. If fictitious play converges to a strategy
that is not a Nash equilibrium then this means that the best response for at
least one of the players is not the same as the convergent strategy. Therefore,
that player will take that action at the next time, taking the system away from
the strategy profile it was supposed to have converged to.
An obvious problem with the solutions provided by fictitious play can be seen
in the existence of infinite cycles of behaviors. An example is illustrated by the
game matrix in Figure 3. If the players start with initial weights of k0
1 (A) = 1,
k0
1 (B ) = 1.5, k0
2 (A) = 1, and k0
2 (B ) = 1.5 they will both believe that the other
will play B and will, therefore, play A. The weights will then be updated to
k1
1 (A) = 2, k1
1 (B ) = 1.5, k1
2 (A) = 2, and k1
2 (B ) = 1.5. Next time, both agents
will believe that the other will play A so both will play B . The agents will engage
in an endless cycle where they alternatively play (A, A) and (B , B ). The agents
end up receiving the worst possible payoff.
This example illustrates the type of problems we encounter when adding
learning to multiagent systems. While we would hope that the machine learning
algorithm we use will be able to discern this simple pattern and exploit it, most
learning algorithms can easily fall into cycles that are not much complicated
than this one. One common strategy for avoiding this problem is the use of
randomness. Agents will sometimes take a random action in an effort to exit
possible loops and to explore the search space. It is interesting to note that, as
in the example from Figure 3, it is often the case that the loops the agents fall
in often reflect one of the mixed strategy Nash equilibria for the game. That
is, (.5, .5) is a Nash equilibrium for this game. Unfortunately, if the agents are
synchronized, as in this case, the implementation of a mixed strategy could lead
to a lower payoff.
Games with more than two players require that we decide whether the agent
should learn individual models of each of the other agents independently or a
joint probability distribution over their combined strategies. Individual mod-
els assume that each agent operates independently while the joint distributions
capture the possibility that the others agents’ strategies are correlated. Unfor-
tunately, for any interesting system the set of all possible strategy profiles is too
large to explore—it grows exponentially with the number of agents. Therefore,
most learning systems assume that all agents operate independently so they need
to maintain only one model per agent.
4.2 Replicator Dynamics
Another widely studied model is replicator dynamics. This model assumes that
the percentage of agents playing a particular strategy will grow in proportion to
how well that strategy performs in the population. A homogeneous population of
agents is assumed. The agents are randomly paired in order to play a symmetric
game, that is, a game where both agents have the same set of possible strategies
and receive the same payoffs for the same actions. The replicator dynamics model
is meant to capture situations where agents reproduce in proportion to how well
they are doing.
Formally, we let φt (s) be the number of agents using strategy s at time t.
We can then define
θt (s) =
φt (s)
Ps′ ∈S φt (s′ )
(3)
to be the fraction of agents playing s at time t. The expected utility for an
agent playing strategy s at time t is defined as
ut (s) ≡ Xs′ ∈S
θt (s′ )u(s, s′ ),
(4)
where u(s, s′ ) is the utility than an agent playing s receives against an agent
playing s′ . Notice that this expected utility assumes that the agents face each
other in pairs and choose their opponents randomly. In the replicator dynamics
the reproduction rate for each agent is proportional to how well it did on the
previous step, that is,
φt+1 (s) = φt (s)(1 + ut (s)).
(5)
Notice that the number of agents playing a particular strategy will continue
to increase as long as the expected utility for that strategy is greater than zero.
Only strategies whose expected utility is negative will decrease in population. It
is also true that under these dynamics the size of a population will constantly
fluctuate. However, when studying replicator dynamics we ignore the absolute
size of the population and focus on the fraction of the population playing a par-
ticular strategy, i.e., θt (s), as time goes on. We are also interested in determining
if the system’s dynamics will converge to some strategy and, if so, which one.
In order to study these systems using the standard solution concepts we view
the fraction of agents playing each strategy as a mixed strategy for the game.
Since the game is symmetric we can use that strategy as the strategy for both
players, so it becomes a strategy profile. We say that the system is in a Nash
equilibrium if the fraction of players playing each strategy is the same as the
probability that the strategy will be played on a Nash equilibrium. In the case
of a pure strategy Nash equilibrium this means that all players are playing the
same strategy.
An examination of these systems quickly leads to the conclusion that every
Nash equilibrium is a steady state for the replicator dynamics. In the Nash
equilibrium all the strategies have the same average payoff since the fraction of
other players playing each strategy matches the Nash equilibrium. This fact can
be easily proven by contradiction. If an agent had a pure strategy that would
return a higher utility than any other strategy then this strategy would be a
best response to the Nash equilibrium. If this strategy was different from the
Nash equilibrium then we would have a best response to the equilibrium which
is not the equilibrium, so the system could not be at a Nash equilibrium.
It has also been shown [4] that a stable steady state of the replicator dynam-
ics is a Nash equilibrium. A stable steady state is one that, after suffering from
a small perturbation, is pushed back to the same steady state by the system’s
dynamics. These states are necessarily Nash equilibria because if they were not
then there would exist some particular small perturbation which would take the
system away from the steady state. This correspondence was further refined by
Bomze [1] who showed that an asymptotically stable steady state corresponds to
a Nash equilibrium that is trembling-hand perfect and isolated. That is, the sta-
ble steady states are a refinement on Nash equilibria—only a few Nash equilibria
can qualify. On the other hand, it is also possible that a replicator dynamics sys-
tem will never converge. In fact, there are many examples of simple games with
no asymptotically stable steady states.
While replicator dynamics reflect some of the most troublesome aspects of
learning in multiagent systems some differences are evident. These differences
are mainly due to the replication assumption. Agents are not usually expected
to replicate, instead they acquire the strategies of others. For example, in a
real multiagent system all the agents might choose to play the strategy that
performed best in the last round instead of choosing their next strategy in pro-
portion to how well it did last time. As such, we cannot directly apply the results
from replicator dynamics to multiagent systems. However, the convergence of the
systems’ dynamics to a Nash equilibrium does illustrate the importance of this
solution concept as an attractor of learning agent’s dynamics.
4.3 Evolutionary Stable Strategies
An Evolutionary Stable Strategy (ESS) is an equilibrium concept applied to
dynamic systems such as the replicator dynamics system of the previous section.
An ESS is an equilibrium strategy that can overcome the presence of a small
number of invaders. That is, if the equilibrium strategy profile is ω and small
number ǫ of invaders start playing ω ′ then ESS states that the existing population
should get a higher payoff against the new mixture (ǫω ′ + (1 − ǫ)ω ) than the
invaders.
It has been shown [9] that an ESS is an asymptotically stable steady state
of the replicator dynamics. However, the converse need not be true—a stable
state in the replicator dynamics does not need to be an ESS. This means that
ESS is a further refinement of the solution concept provided by the replicator
dynamics. ESS can be used when we need a very stable equilibrium concept.
5 Learning Agents
The theory of learning in games provides the designer of multiagent systems with
many useful tools for determining the possible equilibrium points of a system.
Unfortunately, most multiagent systems with learning agents do not converge
to an equilibrium. Designers use learning agents because they do not know, at
design time, the specific circumstances that the agents will face at run time. If
a designer knew the best strategy, that is, the Nash equilibrium strategy, for his
agent then he would simply implement this strategy and avoid the complexities
of implementing a learning algorithm. Therefore, the only times we will see a
multiagent system with learning agents are when the designer cannot predict
that an equilibrium solution will emerge.
The two main reasons for this inability to predict the equilibrium solution
of a system are the existence of unpredictable environmental changes that affect
the agents’ payoffs and the fact that on many systems an agent only has access
to its own set of payoffs—it does not know the payoffs of other agents. These two
reasons make it impossible for a designer to predict which equilibria, if any, the
system would converge to. However, the agents in the system are still playing a
game for which an equilibrium exists, even if the designer cannot predict it at
design-time. But, since the actual payoffs keep changing it is often the case that
the agents are constantly changing their strategy in order to accommodate the
new payoffs.
Learning agents in a multiagent system are faced with a moving target func-
tion problem [10]. That is, as the agents change their behavior in an effort to max-
imize their utility their payoffs for those actions change, changing the expected
utility of their behavior. The system will likely have non-stationary dynamics—
always changing in order to match the new goal. While game theory tells us
where the equilibrium points are, given that the payoffs stay fixed, multiagent
systems often never get to those points. A system designer needs to know how
changes in the design of the system and learning algorithms will affect the time to
convergence. This type of information can be determined by using CLRI theory.
5.1 CLRI Theory
The CLRI theory [12] provides a formal method for analyzing a system composed
of learning agents and determining how an agent’s learning is expected to affect
the learning of other agents in the system. It assumes a system where each agent
has a decision function that governs its behavior as well as a target function that
describes the agent’s best possible behavior. The target function is unknown to
the agent. The goal of the agent’s learning is to have its decision function be
an exact duplicate of its target function. Of course, the target function keeps
changing as a result of other agents’ learning.
Formally, CLRI theory assumes that there are N agents in the system. The
world has a set of discrete states w ∈ W which are presented to the agent with
a probability dictated by the probability distribution D(W ). Each agent i ∈ N
has a set of possible actions Ai where Ai ≥ 2. Time is discrete and indexed by
a variable t. At each time t all agents are presented with a new w ∈ D(W ), take
a simultaneous action, and receive some payoff. The scenario is similar to the
one assumed by fictitious play except for the addition of w.
Each agent i’s behavior is defined by a decision function δ t
i (w) : W → A.
When i learns at time t that it is in state w it will take action δ t
i (w). At any
time there is an optimal function for i given by its target function ∆t
i (w). Agent
i’s learning algorithm will try to reduce the discrepancy between δi and ∆i by
using the payoffs it receives for each action as clues since it does not have direct
access to ∆i . The probability that an agent will take a wrong action is given
by its error e(δ t
i ) = Pr[δ t
i (w) 6= ∆t
i (w) w ∈ D(W )]. As other agents learn and
change their decision function, i’s target function will also change, leading to the
moving target function problem, as depicted in Figure 4.
An agent’s error is based on a fixed probability distribution over world states
and a boolean matching between the decision and target functions. These con-
8q
q
Learn
q
q
q
q
q
q
q
q
q
q
q
δ t
i
δ t+1
i
e(δt+1
i
)
e(δt
i )
/ ∆t+1
∆t
i Move
i
Fig. 4. The moving target function problem.
straints are often too restrictive to properly model many multiagent systems.
However, even if the system being modeled does not completely obey these two
constraints, the use of the CLRI theory in these cases still gives the designer
valuable insight into how changes in the design will affect the dynamics of the
system. This practice is akin to the use of Q-learning in non-Markovian games—
while Q-learning is only guaranteed to converge if the environment is Markovian,
it can still perform well on other domains.
The CLRI theory allows a designer to understand the expected dynamics
of the system, regardless of what learning algorithm is used, by modeling the
system using four parameters: Change rate, Learning rate, Retention rate, and
Impact (CLRI). A designer can determine values for these parameters and then
use the CLRI difference equation to determine the expected behavior of the
system.
The change rate (c) is the probability that an agent will change at least one
of its incorrect mappings in δ t (w) for the new δ t+1 (w). It captures the rate at
which the agent’s learning algorithm tries to change its erroneous mappings. The
learning rate (l) is the probability that the agent changes an incorrect mapping
to the correct one. That is, the probability that δ t+1 (w) = ∆t (w), for all w. By
definition, the learning rate must be less than or equal to the change rate, i.e.
l ≤ c. The retention rate (r) represents the probability that the agent will retain
its correct mapping. That is, the probability that δ t+1 (w) = δ t (w) given that
δ t (w) = ∆t (w).
CLRI defines a volatility term (v) to be the probability that the target func-
tion ∆ changes from time t to t + 1. That is, the probability that ∆t (w) 6=
∆t+1 (w). As one would expect, volatility captures the amount of change that
the agent must deal with. It can also be viewed as the speed of the target func-
tion in the moving target function problem, with the learning and retention rates
representing the speed of the decision function. Since the volatility is a dynamic
property of the system (usually it can only be calculated by running the sys-
tem) CLRI provides an impact (Iij ) measure. Iij represents the impact that i’s
learning has on j ’s target function. Specifically, it is the probability that ∆t
j (w)
will change given that δ t+1
(w) 6= δ t
i (w).
i
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
8
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
Someone trying to build a multiagent system with learning agents would
determine the appropriate values for c, l, r, and either v or I and then use
E [e(δ t+1
i
Ai − 1 (cid:19)
)] = 1 − ri + vi (cid:18) Ai ri − 1
i ) (cid:18)ri − li + vi (cid:18) Ai (li − ri ) + li − ci
(cid:19)(cid:19) (6)
+ e(δ t
Ai − 1
in order to determine the successive expected errors for a typical agent i. This
equation relies on a definition of volatility in terms of impact given by
(7)
(w) 6= δ t
j (w)]),
i = Pr[∆t+1
(w) 6= ∆t
∀w∈W v t
i (w)]
i
= 1 − Yj∈N−i
(1 − Ij iPr[δ t+1
j
which makes the simplifying assumption that changes in agents’ decision func-
tions will not cancel each other out when calculating their impact on other
agents. The difference equation (6) cannot, under most circumstances, be col-
lapsed into a function of t so it must still be iterated over. On the other hand,
a careful study of the function and the reasoning behind the choice of the CLRI
parameter leads to an intuitive understanding of how changes in these parame-
ters will be reflected in the function and, therefore, the system. A knowledgeable
designer can simply use this added understanding to determine the expected be-
havior of his system under various assumptions. An example of this approach is
shown in [2].
For example, it is easy to see that an agent’s learning rate and the system’s
volatility together help to determine how fast, if ever, the agent will reach its
target function. A large learning rate means that an agent will change its decision
function to almost match the target function. Meanwhile, a low volatility means
that the target function will not move much, so it will be easy for the agent to
match it. Of course, this type of simple analysis ignores the common situation
where the agent’s high learning rate is coupled with a high impact on other
agents’ target function making their volatility much higher. These agents might
then have to increase their learning rate and thereby increase the original agent’s
volatility. Equation (6) is most helpful in these type of feedback situations.
5.2 N-Level Agents
Another issue that arises when building learning agents is the choice of a model-
ing level. A designer must decide whether his agent will learn to correlate actions
with rewards, or will try to learn to predict the expected actions of others and
use these predictions along with knowledge of the problem domain to determine
its actions, or will try to learn how other agents build models of other agents,
etc. These choices are usually referred to as n-level modeling agents—an idea
first presented in the recursive modeling method [5] [6].
A 0-level agent is one that does not recognize the existence of other agents
in the world. It learns which action to take in each possible state of the world
because it receives a reward after its actions. The state is usually defined as a
static snapshot of the observable aspects of the agent’s environment. A 1-level
agent recognizes that there are other agents in the world whose actions affect
its payoff. It also has some knowledge that tells it the utility it will receive
given any set of joint actions. This knowledge usually takes the form of a game
matrix that only has utility values for the agent. The 1-level agent observes the
other agents’ actions and builds probabilistic models of the other agents. It then
uses these models to predict their action probability distribution and uses these
distributions to determine its best possible action. A 2-level agent believes that
all other agents are 1-level agents. It, therefore, builds models of their models of
other agents based on the actions it thinks they have seen others take. In essence,
the 2-level agent applies the 1-level algorithm to all other agents in an effort
to predict their action probability distribution and uses these distributions to
determine its best possible actions. A 3-level agent believes that all other agents
are 2-level, an so on. Using these guidelines we can determine that fictitious play
(Section 4.1) uses 1-level agents while the replicator dynamics (Section 4.2) uses
0-level agents.
These categorizations help us to determine the relative computational costs
of each approach and the machine-learning algorithms that are best suited for
that learning problem. 0-level is usually the easiest to implement since it only
requires the learning of one function and no additional knowledge. 1-level learn-
ing requires us to build a model of every agent and can only be implemented
if the agent has the knowledge that tells it which action to take given the set
of actions that others have taken. This knowledge must be integrated into the
agents. However, recent studies in layered learning [8] have shown how some
knowledge could be learned in a “training” situation and then fixed into the
agent so that other knowledge that uses the first one can be learned, either at
runtime or in another training situation. In general, a change in the level that an
agent operates on implies a change on the learning problem and the knowledge
built into the agent.
Studies with n-level agents have shown [11] that an n-level agent will always
perform better in a society full of (n-1)-level agents, and that the computational
costs of increasing a level grow exponentially. Meanwhile, the utility gains to
the agent grow smaller as the agents in the system increase their level, within
an economic scenario. The reason is that an n-level agent is able to exploit the
non-equilibrium dynamics of a system composed of (n-1)-level agents. However,
as the agents increase their level the system reaches equilibrium faster so the
advantages of strategic thinking are reduced—it is best to play the equilibrium
strategy and not worry about what others might do. On the other hand, if all
agents stopped learning then it would be very easy for a new learning agent to
take advantage of them. As such, the research concludes that some of the agents
should do some learning some of the time in order to preserve the robustness of
the system, even if this learning does not have any direct results.
6 Conclusion
We have seen how game theory and the theory of learning in games provide us
with various equilibrium solution concepts and often tell us when some of them
will be reached by simple learning models. On the other hand, we have argued
that the reason learning is used in a multiagent system is often because there
is no known equilibrium or the equilibrium point keeps changing due to outside
forces. We have also shown how the CLRI theory and n-level agents are attempts
to characterize and predict, to a limited degree, the dynamics of a system given
some basic learning parameters.
We conclude that the problems faced by a designer of a learning multiagent
systems cannot be solved solely with the tools of game theory. Game theory
tells us about possible equilibrium points. However, learning agents are rarely
at equilibrium, either because they are not sophisticated enough, because they
lack information, or by design. There is a need to explore non-equilibirium sys-
tems and to develop more predictive theories which, like CLRI, can tell us how
changing either the parameters on the agents’ learning algorithms or the rules
of the game will affect the expected emergent behavior.
References
1. Bomze, I.: Noncoopertive two-person games in biology: A classification. Interna-
tional Journal of Game Theory 15 (1986) 31–37
2. Brooks, C.H., Durfee, E.H.: Congregation formation in multiagent systems. Journal
of Autonomous Agents and Multi-agent Systems (2002) to appear.
3. Fudenberg, D., Kreps, D.: Lectures on learning and equilibrium in strategic-form
games. Technical report, CORE Lecture Series (1990)
4. Fudenberg, D., Levine, D.K.: The Theory of Learning in Games. MIT Press (1998)
5. Gmytrasiewicz, P.J., Durfee., E.H.: A rigorous, operational formalization of re-
cursive modeling. In: Proceedings of the First International Conference on Multi-
Agent Systems. (1995) 125–132
6. Gmytrasiewicz, P.J., Durfee., E.H.: Rational communication in multi-agent sys-
tems. Autonomous Agents and Multi-Agent Systems Journal 4 (2001) 233–272
7. Mitchell, T.M.: Machine Learning. McGraw Hill (1997)
8. Stone, P.: Layered Learning in Multiagent Systems. MIT Press (2000)
9. Taylor, P., Jonker, L.: Evolutionary stable strategies and game dynamics. Mathe-
matical Biosciences 16 (1978) 76–83
10. Vidal, J.M., Durfee, E.H.: The moving target function problem in multi-agent
learning. In: Proceedings of the Third International Conference on Multi-Agent
Systems. (1998)
11. Vidal, J.M., Durfee, E.H.: Learning nested models in an information economy.
Journal of Experimental and Theoretical Artificial Intelligence 10 (1998) 291–308
12. Vidal, J.M., Durfee, E.H.: Predicting the expected behavior of agents that learn
about agents: the CLRI framework. Autonomous Agents and Multi-Agent Systems
(2002)
|
1703.02399 | 1 | 1703 | 2017-03-07T14:33:00 | On time and consistency in multi-level agent-based simulations | [
"cs.MA"
] | The integration of multiple viewpoints became an increasingly popular approach to deal with agent-based simulations. Despite their disparities, recent approaches successfully manage to run such multi-level simulations. Yet, are they doing it appropriately?
This paper tries to answer that question, with an analysis based on a generic model of the temporal dynamics of multi-level simulations. This generic model is then used to build an orthogonal approach to multi-level simulation called SIMILAR. In this approach, most time-related issues are explicitly modeled, owing to an implementation-oriented approach based on the influence/reaction principle. | cs.MA | cs |
On time and consistency in multi-level
agent-based simulations
Gildas MORVAN, Yoann KUBERA
http://www.lgi2a.univ-artois.fr/~morvan/
[email protected]
Laboratoire de Génie Informatique et d'Automatique de l'Artois (LGI2A)
Univ. Artois, EA 3926,
Béthune, France
Abstract
The integration of multiple viewpoints became an increasingly popular ap-
proach to deal with agent-based simulations. Despite their disparities, recent ap-
proaches successfully manage to run such multi-level simulations. Yet, are they
doing it appropriately?
This paper tries to answer that question, with an analysis based on a generic
model of the temporal dynamics of multi-level simulations. This generic model is
then used to build an orthogonal approach to multi-level simulation called SIMI-
LAR. In this approach, most time-related issues are explicitly modeled, owing to
an implementation-oriented approach based on the influence/reaction principle.
Keywords: multi-level agent-based modeling, large-scale simulation
Introduction
Simulating complex systems often requires the integration of knowledge coming from
different viewpoints (e.g. different application fields, different focus points) to obtain
relevant results. Yet, the representations of the agents, the environment and the tem-
poral dynamics in regular multi-agent based simulation meta-models are designed to
support a single viewpoint. Therefore, they lack the structure to manage the integra-
tion of such systems, called multi-level simulations.
1
Managing multiple viewpoints on the same phenomenon induces the use of het-
erogeneous time models, thus raising issues related to time and consistency. Multi-
Level Agent-based Modeling (ML-ABM) is a recent approach that aims at extending
the classical single-viewpoint agent-based modeling paradigm to cope with these issues
and create multiple-viewpoints based simulations [1, 2, 4, 6, 7, 9, 11, 12]. Considering
the disparities between the various ML-ABM approaches, a natural question comes
to mind: is there a "right" way to do ML-ABM?
In this context, the aim of this paper is double. We first aim at eliciting the is-
sues and simulation choices underlying such simulations, with an analysis based on
a generic model of the temporal dynamics of a multi-level simulation. Then, we
present SIMILAR, a ML-ABM approach using the influence/reaction principle to
manage explicitly the issues related to the simultaneous actions of agents in multiple
levels [3, 8, 10].
1 Temporal dynamics in Multi-level simulations
In this section some issues related to multi-level simulation are emphasized using a
generic model describing the temporal dynamics of a multi-level simulation.
1.1 General case
From a coarse grain viewpoint, simulation is a process transforming the data about a
phenomenon from initial values into a sequence of intermediate values, until a final
state is reached. This evolution is characterized by: 1) A dynamic state δ(t)∈ ∆ mod-
eling the data of the simulation at time t; 2) A time model (cid:84) representing the moments
when each state of the discrete evolution was obtained; 3) A behavior model describ-
ing the evolution process of the dynamic state between two consecutive moments of
the time model.
The exact content of the time model, dynamic state, as well as the behavior model
of a simulation depends on the simulation approach being used. Yet, despite their
disparities, many common points can be identified among them.
First, since real time can be seen as a continuous value, most simulation assume
that (cid:84) ⊂ (cid:82). Moreover, we can assume that a simulation eventually ends. Thus, (cid:84)
contains an ordered, finite and discrete set of time values t ∈ (cid:84). Then, the dynamic
state contains data related to the agents and the environment1.
1.2 Multi-level case
In multi-level simulations, each level embodies a specific viewpoint on the studied
phenomenon. Since these viewpoints can evolve using very different time scales, each
level l ∈ (cid:76) (where (cid:76) is the set of all levels) has to define its own time model (cid:84)
l .
1In this paper, we use a simplistic definition of these concepts: an agent is an entity that can perceive data
about itself, the environment and the other agents, possibly memorize some of them and decide to perform
actions
2
The interaction of the levels is possible only by defining when and under which
circumstances interaction is possible. For this purpose, we introduce a multi-level
specific terminology to the temporal dynamics.
1.2.1 Local information
We consider that agents can lie in more than one level at a time. (cid:65) (t, l) ∈ (cid:65)
l de-
notes the set of agents of the level l ∈ (cid:76) at time t ∈ (cid:84)
l . Since levels can have very
different temporal dynamics, this point has various implications on the structure of
a,l in each level l ∈(cid:76) where
(t, l)∈ Φ
the simulation: 1) Agents have a local state2 φ
they lie; 2) Agents perform decisions differently depending on the level from which
the decisions originates; 3) A level l can only trigger the local behavior of the agents
lying in l.
Similarly, the environment has a local state φω(t, l) ∈ Φω in each level of the
simulation. Yet, contrary to the agents, the environment is present in each level of
the simulation. Each local state embodies any agent-independent information like a
topology or a state (e.g. an ambient temperature).
a
1.2.2 Global information
The coherence of agent behaviors in each level can require information like cross-level
plans or any other level-independent information. Therefore, we consider that agents
have a global state3 µ
a, which is independent from any level.
(t)∈ (cid:77)
a
1.2.3 Content of a dynamic state
Owing to the abovementioned information, the dynamic state δ(t) ∈ ∆ of a multi-
level simulation at time t can be defined as the sum of the local dynamic state δ(t, l)∈
l of each level l ∈ (cid:76) and the global dynamic state δ
∆
G containing the global
×(cid:89)
state of the agents.
∀δ(t)∈ ∆, δ(t) =
(t),δ(t, l)
∧∀l ∈ (cid:76), δ(t, l)∈ ∆
(t)∈ ∆
(t)∈ ∆
∆ = ∆
G
l∈(cid:76)
∆
l
(1)
l (2)
G
δ
G
l∈(cid:76)
with δ
G
G
1.2.4 Time model of the simulation
The interaction between levels is possible only if their time models are somehow cor-
related. Since the time model of each level is a discrete ordered set, it is possible to
build an order between their elements.
l∈(cid:76)(cid:84)
models of all the levels: (cid:84) =(cid:83)
The time model of a multi-level simulation is defined as the union of the time
l . For consistency reasons, the time models (cid:84)
and (cid:84)
l must have the same bounds. Since (cid:84) and (cid:84)
l are ordered, we also define s + d t
(resp. s + d tl ) as the successor of s ∈ (cid:84) (resp. s ∈ (cid:84)
l ).
2Also called "physical state" or "face" in the literature [3, 8, 10, 11]
3Also called "mind", "memory state" or "core" in the literature [3, 8, 10, 11]
3
1.2.5 Consistent and transitory states
l , the level l ∈ (cid:76) is in a transitory state. No guarantee can
In the case where t (cid:54)∈ (cid:84)
be provided on such a state, since it corresponds to a temporary value used by l to
compute its future consistent dynamic state. On the opposite, the data contained in
the dynamic state of a level l ∈ (cid:76) can be safely read or perceived at times in (cid:84)
l , where
this state is considered as consistent.
Figure 1: Illustration of a time model, for a simulation using two levels "A" and
"B".The first line represents real time.The second and third line represent the time
model of the levels "A" and "B". The last line represents the time model of the whole
simulation. At t1, the state of the simulation is consistent. At t4, it is half-consistent.
From a global viewpoint, the dynamic state of the simulation is consistent (resp.
transitory) if all of its levels are consistent (resp.
It can also be in an
intermediate situation called half-consistent state, if a level is in a consistent and another
level is in a transitory state. These concepts are illustrated in Figure 1.
To clarify our speech, we write δ(t) the consistent (or half-consistent) dynamic state
of the simulation at time t ∈ (cid:84) and δ(]t, t(cid:48)[) the transitory dynamic state of the sim-
ulation between the times t ∈ (cid:84) and t(cid:48) ∈ (cid:84).
transitory).
1.3 Multi-level inherent issues
When a simulation is in a transitory phase, each level performs operations in parallel
to determine the next consistent value of their dynamic state. The transitory periods
of the levels are not necessarily in sync. Therefore, each level can be at a different step
of its transitory operations when an interaction occurs. This point raises the follow-
ing time-related issues: 1) Determine on which dynamic states is based the decision
in a level to interact with another level; 2) Determine when to take into considera-
tion the modifications in a level resulting from an action initiated in another level; 3)
Determine how to preserve the consistency of the global state of the agents despite
having its update occurring after the level-dependent perceptions.
4
This section illustrates these issues on an example containing two levels "A" and
"B", presented in Figure 2
Figure 2: Illustration of a subset of a time model containing two levels "A" and "B".
1.3.1 Level interaction through perception
The first issue is related to the perception of the dynamic state of the other levels. It
happens for instance at the time t2 (see fig. 2), when an agent from the level "B" tries
to read information from the level "A". Indeed, since "A" is in a transitory state at
that time, the data being read by the agent might have arbitrary values. Therefore, a
heuristic has to be used to disambiguate that value. For instance using the last consis-
tent dynamic state of the level (in this case the dynamic state of "A" at the time t1),
using the arbitrary values from the transitory state at t2 or anticipating the modifica-
tions that might have occurred in "A" between t1 and t2.
1.3.2 Level interaction through actions
The second issue is related to the side effects of an interaction between two levels. It
happens for instance during the transitory period ]t1, t3
[ (see fig. 2) of the level "A",
if an agent from "A" tries to interact with the level "B". Indeed, since both levels have
a different time scale, it is difficult to determine when the actions of "A" have to be
taken for account into the computation of the dynamic state of "B". It can be during
the transitory phases ]t1, t2
A generic answer to that problem might be "the next time both levels are in sync"
(t3 in this case). Yet, this leads to aberrations like taking into consideration these
actions at the end of the simulation (for instance in fig 1, if an agent from the level
"A" interacts with "B" during the transitory period ]t3, t5
[).
[, ]t2, t3
[ or a later one.
1.3.3 Global state update
The third issue is related to the read and write access of the global state of an agent and
the update of that value. Indeed, during the transitory phase of each level, the agent
has to read and possibly update the value of the global states, to take into account the
information that were perceived. Yet, since the perception is relative to each level, the
global state is the subject of the same issues than the interactions between levels.
5
For instance, in figure 2, the period ]t1, t2
[ of the simulation is a transitory period
for the level "B" and a subset of the transitory period for the level "A". The latter
raises the question of whether if the global state of the agent at the time t2 has to take
into consideration the data being perceived by the agent from "A" or not. Indeed,
perception might not be complete at that time in "A".
1.4 Differences between multi-level approaches
There is no universal answer to the issues presented in this section, since the coherence
between heterogeneous time scales is itself an ill-defined notion. The main differences
between existing ML-ABM approaches are the way these issues are handled, through
the answer of the following questions about the operations performed during a tran-
sitory phase ]t, t + d t[: 1) Which agents can perform a decision during a transitory
period of a level? 2) How many actions can be performed by the decision process of
an agent? 3) How are committed the results of the action to the future dynamic state
of a level? 4) When are performed these operations during the transitory state? 5)
Which dynamic state of a level k is read by a level l initiating an interaction with k?
A consistent one? A transitory one? Which ones ? 6) When is taken into account the
interaction initiated by a level l with a level k? 7) How is managed the consistency of
the global state of agents?
In the next section, we present an agent-based approach called SIMILAR, that aims
at addressing these issues.
2 SIMILAR
Many meta-models and simulation engines dedicated to ML-ABM have been proposed
in the literature such as IRM4MLS [10], PADAWAN [11], GAMA [5] or NETLOGO
LEVELSPACE [6]. All these approaches provide a different and yet valid answer to the
multi-level simulation issues. In this paper, we do not aim at detailing precisely their
differences: a comprehensive survey of the different approaches can be found in [9].
Existing approaches like GAMA or PADAWAN (Pattern for Accurate Design of
Agent Worlds in Agent Nests) are complete approaches providing various interesting
features respectively including the agentification of emerging structures or the elicita-
tion of interactions between agents. However, these approaches rely on a time model
where the management of the potentially simultaneous actions is strongly constrained
by the sequential execution of agent actions.
In this paper, we investigate another approach where agent actions are separated
from their consequences in the dynamic state of the simulation, using the influence/reaction
principle [3]. The resulting approach, called SIMILAR (SImulations with MultI-Level
Agents and Reactions), is deeply inspired by IRM4MLS [10], a multi-level extension
of IRM4S [8]. The main differences between SIMILAR and IRM4MLS are the more
precise and less misleading terminology and simulation algorithms, as well as a more
precise and implementation-oriented model for the reaction phase (the latter is not
described in this paper due to the lack of space).
6
2.1 Core concepts
SIMILAR revolves around five core concepts: 1) Levels, modeling different view-
points on the simulated phenomenon; 2) Agents lying in one or more levels. From
each level where they lie, they perceive the state of one or more levels to decide how
they wish to influence the evolution of the system; 3) the Environment modeling the
topology, the local information (e.g. temperature) and the natural evolution4 of each
level; 4) Influences modeling actions which effect has yet to be committed to the state
of the simulation; 5) Reactions modeling how the changes depicted by the influences
are committed to the state of the simulation.
We note (cid:76) the levels defined for a simulation, (cid:73) the domain space of all the possible
influences of the simulation and (cid:65) all possible agents of the simulation.
2.2 Heuristics
SIMILAR relies on the following heuristics and choices to manage the issues raised in
[ of a level l ∈ (cid:76), the agents
the section 1.4. 1) During the transitory period ]t, t +d tl
from (cid:65) (t, l) decide once in parallel; 2) The number of influences produced by each
decision is not constrained; 3) The result of the actions is committed to the future
dynamic state of a level using a reaction mechanism [3] ; 4) During the transitory
[ of a level l ∈ (cid:76), the behavior of the agents is triggered slightly after
period ]t, t +d tl
t and the reaction occurs slightly before t + d tl ; 5) The dynamic state being read by
the behavior of an agent (or of the environment) is always the most recent consistent
state of the level5; 6) The actions emitted by an agent from a level l to a level k during
a transitory period ]t, t +d tl
[ are taken into account in the next reaction of k after the
time t (i.e. the reaction occurring during the transitory period containing or starting
with the time t); 7) The consistency of agent global states is attained by: i) Computing
the revised global state of the agents at the beginning of the transitory period of a level
(i.e. before any reaction); ii) Computing the revised global state of an agent once for
all the levels starting a new transitory period at the same time; iii) Use this revised
global state as the global state of the agent for the next half- consistent state of the
simulation. This approach is summarized in Figure 3
2.3 Dynamic state
In SIMILAR, we consider that each point of view on a phenomenon has to be em-
bodied in a level l ∈ (cid:76). As a consequence, the dynamic state δ(t) of the simulation
is divided in level-specific dynamic states δ(t, l). Two kind of data can be obtained
from the dynamic state of a level l ∈ (cid:76): a state valuation σ(t, l), defining a valuation
of the level-related properties of the agents (e.g. their location or their temperature) or
the environment (e.g. an ambient temperature) and the state dynamics γ(t, l), defining
the actions that were still being performed6 in that level during the observation.
4i.e. without the intervention of the behavior of an agent
5Default heuristic of SIMILAR. SIMILAR also allows the definition of user-defined disambiguation
heuristics
6Actions that started before the time t and that will end after the time t
7
Figure 3: Illustration of the operation performed independently in each level (squares)
and joint for all the levels (rectangles) during a simulation. The letters describe the
type of the operation: Perception (P), Global state revision (M), Decision (D) and
Reaction (R). Any arrow starting on a consistent dynamic state at a time t points to
the revised global state used as the new global state of the agents in the half-consistent
dynamic state of the time t. This example focuses on the operations performed in a
simulation containing one agent lying in three levels.
∀l ∈ (cid:76),∀t ∈ (cid:84)
l , δ(t, l) =< σ(t, l), γ(t, l) >
(3)
2.3.1 State valuation
The state valuation σ(t, l) of a level l ∈ (cid:76) is the union of the local state of the en-
vironment φω(t, l) ∈ Φω, containing agent-unrelated information and a local state7
(t, l)∈ Φ
a,l for each agent a ∈ (cid:65) contained in the level.
φ
a
2.3.2 State dynamics
SIMILAR relies on the influence/reaction principle to model the actions resulting
from the decision of the agents, from the natural evolution of the environment and
the actions still being performed at time t. Therefore, the state dynamics γ(t, l) of a
level l ∈ (cid:76) contains a set of influences.
∀l ∈ (cid:76),∀t ∈ (cid:84)
l , γ(t, l)⊆ (cid:73)
(4)
Since the data contained in an influence are mostly domain-dependent, no specific
model is attached to them. They usually contain the subjects of the action (e.g. the
physical state of one or more agents) as well as parameters (e.g. an amount of money
to exchange).
2.4 General behavior model
The dynamic state of a simulation models a "photograph" of the simulation at time
t. Motion is attained owing to the behavior of the agents, the natural action of the
7This replaces the term "physical state" from IRM4MLS, which was misleading, since that state also
contains mental information like a desired speed.
8
environment and the reaction of each level to influences.
2.4.1 Behavior of the agents
The behavior of an agent in a level l ∈ (cid:76) has three phases: 1) Perception: extract
information from the dynamic state of the levels that can be perceived from l; 2)
Global state revision: use the newly perceived data to revise the content of the global
state of the agent; 3) Decision: use the perceived data and the revised global state to
create and send influences to the levels that can be influenced by l. Each influence
models a modification request of the dynamic state of a level.
2.4.2 Natural action of the environment
The natural action of the environment is simpler than the behavior of agents: it only
has one phase, where the dynamic state of the levels that can be perceived from l are
used to create influences sent to one or more levels that can be influenced by l.
2.4.3 Reaction to the influences
As in IRM4MLS, in SIMILAR the reaction of a level l ∈ (cid:76) is a process computing
the new consistent dynamic state of l. The reaction phase occurs at the end of a
transitory period ]t, t + d tl
[ of a level, and is computed using the value of the most
recent consistent dynamic state of l and the influences that were sent to l during the
transitory period ]t, t + d tl
Yet, contrary to IRM4MLS, SIMILAR provides an explicit model to the generic
influences that can be found in any simulation, like the addition/removal of an agent
from the simulation/a level. Such influences are called system influences, in opposition
to regular influences, which are user-defined. A model is also provided to their generic
reaction. These points are not detailed in this paper.
[;
2.5 Formal notations and simulation algorithm
Not all levels are able to interact. Therefore, the interactions between levels are con-
strained by two digraphs: A perception relation graph (cid:71)
influence relation
graph (cid:71)
I ) defines which levels can be perceived (resp. influenced) during the behav-
ior of the agent/environment in a specific level.
P (resp.
l1, l2
∈ (cid:71)
P (resp. (cid:71)
I ) ⇐⇒ An agent from l1 can perceive (resp.
influence) the dynamic state of l2
The out neighborhood (cid:78) +
(5)
(l)) of a level l ∈ (cid:76) in the perception
(resp. influence) relation graph defines the levels that can be perceived (resp. influ-
enced) by l.
(l) (resp. (cid:78) +
P
I
2.5.1 Agent behavior
Since the content of the dynamic state is not trustworthy during transitory periods,
the natural action of the environment and the perception of the agents are based on
9
}
{u ≤ tu ∈ (cid:84)
the last consistent dynamic state of the perceptible levels. This time is identified by
(]t, t(cid:48)[), which models the last time when the dynamic state of a
the notation f l oorl
level l was consistent for a perception occurring during a transitory period ]t, t(cid:48)[.
(]t, t
∀t, t
(cid:48)[) = max
(cid:48) ∈ (cid:84)2,∀l ∈ (cid:76), f l oorl
(6)
Based on these information, the perception phase of an agent a ∈(cid:65) (t, l) from a level
l ∈ (cid:76) for the transitory period ]t, t+d tl
This application reads the last consistent dynamic state of each perceptible level to
produce the perceived data:
[), k
(cid:81)
∀l ∈ (cid:76),∀t ∈ (cid:84)
δf l oork
l
k∈(cid:78) +
pe r ce pt iona,]t,t+d tl
−→ (cid:80)
(cid:55)−→ pa,l
[ is defined as an application pe r ce pt iona,]t,t+d tl
)},∀a ∈(cid:65) (t, l),
\{max((cid:84)
(l) ∆
k
(]t, t + d tl
(]t, t + d tl
[,l :
(7)
[)
a,l
P
l
l
[,l .
k∈(cid:78) +
P
(l)
In this notation, (cid:80)
a models the domain space of the data that can be perceived by the
agent a, from the perspective of the level l. It can contain raw data from the dynamic
states, or an interpretation of these data. For instance, in a road traffic simulation, the
drivers do not need to put the absolute position of the leading vehicle (i.e. raw data
from the dynamic state) in their perceived data: the distance between the two vehicles
is sufficient.
The revision of the global state of an agent a ∈ (cid:65) (t, l) for a transitory period
t) of the agent a and the perceived
starting at the time t is defined as an application g l obal Reva,]t,t+d t[. This applica-
tion reads the most recent consistent global state µ
data pa,l
the time t, in order to determine the value of the revised global state µ
of the agent during the transitory period.
∀t ∈ (cid:84)\{max((cid:84))},∀a ∈(cid:65) (t),
l of all the levels that started a transitory phase at
(]t, t + d t[)
(]t, t + d tl
a
a
[),∀l ∈ (cid:76)t ∈ (cid:84)
×(cid:81)
(t),pa,l
l∈(cid:76)t∈(cid:84)
a
l
(cid:77)
a
µ
(cid:80)
a,l
g l obal Reva,]t,t+d t[ :
−→ (cid:77)
(cid:55)−→ µ
[)
a
(]t, t + d t[)
(]t, t + d tl
l
a
a
l∈(cid:76)t∈(cid:84)
(]t, t+d t[) of a and the perceived data pa,l
[ is defined as an application d eci siona,]t,t+d tl
(8)
Finally, the decision of an agent a ∈ (cid:65) (t, l) from a level l ∈ (cid:76) for the transitory
period ]t, t + d tl
[,l . This application
[)
(]t, t+d tl
reads the revised global state µ
computed for the level l to create the influences that will modify levels during their
respective next reaction.
∀l ∈ (cid:76),∀t ∈ (cid:84)
\{max((cid:84)
×(cid:80)
(]t, t + d t[), pa,l
(9)
If we note l eve l(i) the level at which the influence i ∈ (cid:73) is aimed, then the influence
relation graph imposes the following constraint to the decision:
)},∀a ∈(cid:65) (t, l), d eci siona,]t,t+d tl
[) (cid:55)−→ (cid:73)
−→ 2(cid:73)
(]t, t + d tl
(]t, t + d tl
(cid:77)
a
µ
[,l :
[)
a,l
a,l
a
l
l
∀l ∈ (cid:76),∀t ∈ (cid:84)
\{max((cid:84)
)}, i ∈(cid:73)
l
l
[)⇒ l eve l(i)∈(cid:78) +
I
(l)
(10)
(]t, t + d tl
a,l
10
As a result to this phase, each created influence i ∈ (cid:73)
transitory state dynamics of k = l eve l(i), for the transitory period ]f l oork
d tl
(]t, t + d tl
(]t, t + d tl
[), f l oork
[) + d tk
[) is added to the
(]t, t +
[.
a,l
2.5.2 Natural action of the environment
The natural action of the environment from a level l ∈ (cid:76) for the transitory period
]t, t + d tl
[,l . This application reads the
last consistent dynamic state of each perceptible level to create the influences that will
modify the dynamic state of the influenceable levels (during their reaction).
[ is defined as an application nat u ral]t,t+d tl
(cid:81)
∀l ∈ (cid:76),∀t ∈ (cid:84)
δf l oork
l
k∈(cid:78) +
\{max((cid:84)
(l) ∆
P
l
k
(]t, t + d tl
)}, nat u ral]t,t+d tl
[), k
k∈(cid:78) +
P
(l)
[,l :
−→ 2(cid:73)
(cid:55)−→ (cid:73)ω,l
(]t, t + d tl
[)
(11)
The resulting influences are managed with the same process than the ones coming
from the decisions of the agents.
[ where t ∈ (cid:84)
l . It is defined as an application r eac t ionl,]t,t+d tl
2.5.3 Reaction of a level
The reaction of a level l ∈ (cid:76) is computed at the end of each transitory period ]t, t +
[ reading the tran-
d tl
[, l) of the level to determine the next consistent
sitory dynamic state δ(]t, t + d tl
value of the dynamic state δ(t + d tl , l).
∀l ∈ (cid:76),∀t ∈ (cid:84)
\{max((cid:84)
r eac t ion]t,t+d tl
−→ ∆
[, l (cid:55)−→ δt + d tl , l
δ]t, t + d tl
)},
(12)
[,l :
∆
l
l
l
l
The reaction has the following responsibilities: 1) Take into consideration the influ-
[, l) to update the local state of the agents, update the local state
ences of γ(]t, t + d tl
of the environment, create/delete agents from the simulation or add/remove agents
from the level; 2) Determine if the influences of γ(]t, t +d tl
[, l) persist in γ(t +d tl , l)
(if they model something that has not finished at the time t +d tl ); 3) Manage the col-
liding influences of γ(]t, t + d tl
[, l).
2.5.4 Simulation algorithm
The simulation algorithm of SIMILAR is presented in Figure 4. It relies on the pre-
sented concepts and complies with the time constraints defined in [10].
Conclusion and perspectives
In this paper, we elicited several issues about time and consistency raised in multi-level
simulations. There is no clear solution to theme since the notion of time consistency
11
t ←− min((cid:84));
D ←− Map<level,dynamic state> //Most recent consistent dynamic state
M ←− Map<agent,memory state> //Most recent consistent global state
T ←− Map<level,time interval> //Next transitory state time interval
←− Map<level,dynamic state> //Next transitory state
DT
←− Map<agent,memory state> //Most recent revised global state
MT
for l ∈ (cid:76) do
D[l]←− δ(t, l);
T [l]←−]t, t + d tl
DT
end
while t (cid:54)= max((cid:84)) do
for l ∈ (cid:76) do
[l]←− copy of D[l];
[;
if t = end(T [l]) then
D[l]←− r eac t ionT [l],l
[l]←− copy of D[l];
DT
T [l] =]t, t + d tl
[;
end
for a ∈(cid:65) (t) do
(DT
[l]);
end
[a];
M[a] = copy of MT
end
Li ←−{l ∈ (cid:76)s t ar t(T [l]) = t};
p ←− Map<agent,Map< level, perceived data >>;
A←−(cid:59);
for l ∈ Li do
for a ∈(cid:65) (t, l) do
p[a][l]←− pe r ce pt iona,T [l],l
A←− A∪{a}
(D[k])
;
k∈(cid:78) +
P
(l)
end
end
for a ∈ A do
MT
end
I ←−(cid:59);
for l ∈ Li do
end
end
for i ∈ I do
[a] = g l obal Reva,]t,t+d t[(M[a], p[a]);
I ←− I ∪ nat u ralT [l],l
for a ∈(cid:65) (t, l) do
(D[k])
I ←− I ∪ d eci siona,T [l],l
;
k∈(cid:78) +
p (l)
(MT
[a], p[a]);
Add i to the state dynamics of DT
r i g ht(T [l]), l ∈ (cid:76)
end
t ←− min
[l eve l(i)];
//t becomes equal to t + d t
end
Figure 4: The simulation algorithm used in SIMILAR
among heterogeneous time models is itself ill-defined. Therefore, rather than distin-
guishing the "right" or "wrong" approaches, we defined a theoretical frame giving a
better understanding of the choices underlying each approach. Then, it is up to mod-
elers and domain specialists to tell if these choices are appropriate or not for the study
of a given phenomenon.
To cope with the multi-level related issues, we introduced a meta-model named
SIMILAR based on the influence/reaction principle. This model is designed to reify as
much as possible the concepts involved in the abovementionned issues, thus providing
a better support to the definition of explicit solutions to them. SIMILAR includes a
generic and modular formal model, a methodology and a simulation API preserving
the structure of the formal model. Thus, the design of simulations is in addition
more robust to model revisions and relies on a structure fit to represent the intrinsic
complexity of the simulated multi-level phenomena.
SIMILAR has been implemented in Java and is available under the CeCILL-B li-
cense.
12
It is available at http://www.lgi2a.univ-artois.fr/~morvan/similar.html.
References
[1] B. Camus, C. Bourjot, and V. Chevrier. Considering a multi-level model as
a society of interacting models: Application to a collective motion example.
Journal of Artificial Societies & Social Simulation, 18(3), 2015.
[2] D. David and R. Courdier. See emergence as a metaknowledge. a way to reify
emergent phenomena in multiagent simulations? In Proceedings of ICAART'09,
pages 564 -- 569, Porto, Portugal, 2009.
[3] J. Ferber and J-P. Müller. Influences and reaction: a model of situated multia-
gent systems. In 2nd International Conference on Multi-agent systems (ICMAS'96),
pages 72 -- 79, 1996.
[4] J. Gil-Quijano, T. Louail, and G. Hutzler. From biological to urban cells:
In Principles and Practice
Lessons from three multilevel agent-based models.
of Multi-Agent Systems, volume 7057 of LNCS, pages 620 -- 635. Springer, 2012.
[5] A. Grignard, P. Taillandier, B. Gaudou, D. A. Vo, N. Q. Huynh, and A. Dro-
goul. GAMA 1.6: Advancing the art of complex agent-based modeling and sim-
ulation. In PRIMA 2013: Principles and Practice of Multi-Agent Systems, volume
8291 of LNCS, pages 117 -- 131. Springer, 2013.
[6] A. Hjorth, B. Head, and U. Wilensky. Levelspace netlogo extension. http:
//ccl.northwestern.edu/rp/levelspace/index.shtml, 2015.
[7] T. Huraux, N. Sabouret, and Y. Haradji. A multi-level model for multi-agent
based simulation. In Proc. of 6th Int. Conf. on Agents and Artificial Intelligence
(ICAART), 2014.
[8] F. Michel. The IRM4S model: the influence/reaction principle for multiagent
based simulation. In Proc. of 6th Int. Conf. on Autonomous Agents and Multiagent
Systems (AAMAS), pages 1 -- 3, 2007.
[9] G. Morvan. Multi-level agent-based modeling - a literature survey. CoRR,
abs/1205.0561, 2013.
[10] G. Morvan, A. Veremme, and D. Dupont. IRM4MLS: the influence reaction
model for multi-level simulation. In Multi-Agent-Based Simulation XI, volume
6532 of LNCS, pages 16 -- 27. Springer, 2011.
[11] S. Picault and P. Mathieu. An interaction-oriented model for multi-scale simula-
tion. In Proc of the 22nd Int. Joint Conf. on Artificial Intelligence, pages 332 -- 337.
AAAI Press, 2011.
[12] D-A. Vo. An operational architecture to handle multiple levels of representation in
agent-based models. PhD thesis, Université Paris VI, 2012.
13
|
1204.0634 | 1 | 1204 | 2012-04-03T09:27:26 | Multi-level agent-based modeling with the Influence Reaction principle | [
"cs.MA"
] | This paper deals with the specification and the implementation of multi-level agent-based models, using a formal model, IRM4MLS (an Influence Reaction Model for Multi-Level Simulation), based on the Influence Reaction principle. Proposed examples illustrate forms of top-down control in (multi-level) multi-agent based-simulations. | cs.MA | cs |
Multi-level agent-based modeling
with the Influence Reaction principle
Gildas Morvan, Daniel Jolly
Univ. Lille Nord de France
1bis rue Georges Lef`evre 59044 Lille cedex, France
LGI2A, U. Artois, Technoparc Futura 62400 B´ethune, France.
email: first [email protected]
Abstract
This paper deals with the specification and the implementation of
multi-level agent-based models, using a formal model, IRM4MLS (an In-
fluence Reaction Model for Multi-Level Simulation), based on the Influ-
ence Reaction principle. Proposed examples illustrate forms of top-down
control in (multi-level) multi-agent based-simulations.
Keywords: multi-level simulation, influence reaction model, micro-macro
link, cellular automata.
1
Introduction
Four main theoretical issues emerge in the literature on multi-level1 agent-based
modeling: the conception of a meta-model allowing a non ambiguous charac-
terization of a multi-level model at the conceptual level [18, 23, 29, 33], the
introduction of a dynamic level of detail [25, 31], the detection and reification of
emergent phenomena [2, 3, 5, 21, 30, 34] and the representation of aggregated
entities [27]. This paper focuses on the first one, with respect to the Influence
Reaction (IR) principle, shortly, action as a two step process: (1) agents pro-
duce "influences", i.e., individual decisions, according to their internal state and
1The term multi-scale may also be found. Intuitively, a level and scale are similar concepts
In the following we
that both mean viewpoint. However, this notion should be clarified.
assume that two agents are not at the same scale iff they represent processes that have
different spatial and/or temporal extents. Two agents are not in the same level iff they cannot
interact directly, i.e., with a single interaction function. It should follow from the previous
definitions that it exists multi-scale models that are mono-level and conversely.
1
perceptions (2) the system "reacts", i.e., computes the consequences of influ-
ences, according to the state of the world [10]. This model has been extended in
several ways, and notably for multi-agent based-simulation (MABS), by adding
an explicit representation of time [19]. An IR-based meta-model, IRM4MLS
(an Influence Reaction Model for Multi-Level Simulation), and its Java imple-
mentation are introduced in the section 2. Using, and then extending a simple
example: the Conway's game of life, two toy-models of increasing complexity are
presented in the section 3, illustrating forms of top-down control in (multi-level)
MABS. Results are discussed in the section 4.
2
IRM4MLS: an Influence Reaction Model for
Multi-Level Simulation
In this section, IRM4MLS, an Influence Reaction Model for Multi-Level Sim-
ulation, is introduced [23]. It extends IRM4S (an Influence Reaction Model for
Simulation) in order to deal with multi-level models [19]. From a technical perspec-
tive, levels can be viewed as interacting IRM4S simulations2.
A multi-level model is defined by a set of levels L and a specification of the relations
between levels. Two types of relations are specified: an influence relation (agents in a
level l are able to produce influences in a level l(cid:48) (cid:54)= l) and a perception relation (agents
in a level l are able to perceive the dynamic state of a level l(cid:48) (cid:54)= l), represented by
directed graphs denoted respectively < L, EI > and < L, EP >, where EI and EP
are two sets of edges, i.e., ordered pairs of elements of L. Influence and perception
relations in a level are systematic and thus not specified in EI and EP (cf. eq. 1 and
2).
The in and out neighborhood in < L, EI > (respectively < L, EP >) are denoted
−
I and N +
N
I (resp. N
∀l ∈ L, N
−
P and N +
−
I (l) (resp. N
P ) and are defined as follows:
P (l)) = {l} ∪ {l
−
(cid:48) ∈ L : l
(cid:48)
l ∈ EI (resp. EP )},
(1)
∀l ∈ L, N +
P (l)) = {l} ∪ {l
(cid:48) ∈ L : ll
(cid:48) ∈ EI (resp. EP )},
I (l) (resp. N +
(2)
The set of agents in the system at time t is denoted A(t). ∀l ∈ L, the set of agents
belonging to l at t is denoted Al(t) ⊆ A(t). An agent belongs to a level iff a subset of
its physical state φa belongs to the state of the level. Thus, an agent belongs to zero,
one, or more levels. An environment models the natural dynamics of level properties
and can be shared by different levels (fig. 1).
is a tu-
ple < σl(t), γl(t) >, where σl(t) ∈ Σl and γl(t) ∈ Γl are the sets of environmen-
tal properties and influences of l. The behavior of an agent a ∈ Al is defined as
Decisionl
The dynamic state of a level l ∈ L at time t, denoted δl(t) ∈ ∆l,
a ◦ M emorizationa ◦ P erceptionl
a, with
(cid:89)
∆lP (cid:55)→ (cid:89)
P erceptionl
a :
P lP
a ,
(3)
lP ∈N +
P (l)
lP ∈N +
P (l)
2Therefore, each level has a microscopic side: the agent behaviors, and a macroscopic side:
the reaction function. This aspect can also be found in holonic multi-agent systems [4].
2
(cid:89)
(cid:89)
a : Sa (cid:55)→ (cid:89)
lP ∈N +
l∈La∈Al
lI∈N +
I (l)
P (l)
a × Sa (cid:55)→ Sa,
P lP
ΓlI (cid:48)
.
(4)
(5)
M emorizationa :
Decisionl
N aturall
(6)
(7)
There is no memorization function specific to a level to preserve the coherence of the
internal state of the agents. The environment ω of a level l produces influences through
a function:
ω : ∆l (cid:55)→ (cid:89)
ΓlI (cid:48)
.
The reaction function computes next level state and time advance:
lI∈N +
I (l)
Reactionl : Σl × Γl(cid:48) (cid:55)→ ∆l × Tl.
cation) [35]. T = (cid:83)
The time representation is inspired by DEVS (Discrete EVent System specifi-
l∈L{Tl} denotes the time vector of the simulation, such as
∀l ∈ L, Tl =< tl, dtl >, where tl represents when the current event (or step, de-
pending on the simulation model) time and dtl its lifespan. The final simulation time
is denoted tf . The algorithm 1 ensures the scheduling of these different functions with
respect to temporal constraints of perception and memorization, influence production
and reaction [23].
The implementation of IRM4MLS is based on the idea of micro kernel, taken from
MadKit3 [13]. Thus, in this approach, a technical agent, e.g., an observer or a message
broker, would be scheduled with respect to the IR principle (cf. algo. 1), the concept
of level ensuring a clear separation between system and simulation agents.
The API is minimal (seven high-level abstractions) and specifies only the methods
needed to schedule a model (fig. 2). Most methods are generic and then are imple-
mented at an abstract level. Basically, to implement a model one only has to override
perception, memorization, influence production, reaction and initialization functions.
Agent and behaviors (such as environment and natural dynamics) are represented
by different entities to clearly distinguish the core side of an agent, its state and mem-
orization function, and the level sides (perception and influence production functions)
that can change according to simulations.
3http://www.madkit.org
0..n
agent
0..n
0..n
environment
1
1..n
0..n
level
Figure 1: Main concepts of IRM4MLS (as a simplified class diagram)
3
Algorithm 1: simulation model of IRM4MLS
Input: < L, EI , EP >, A, δ, T, tf
Output: δ(tf )
1 while ∃tl ≤ tf do
foreach a ∈ A(t) do
foreach l ∈ L : a ∈ Al(t) ∧ ∀lP ∈ N +
a(< δlP (tlP ) : lP ∈ N +
P (l), tl ≥ tlP do
P (l) >);
pa(tl) = P erceptionl
end
sa(tl + dtl) = M emorizationa(pa);
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
foreach lI ∈ N +
end
foreach l ∈ L : ∀lI ∈ N +
(cid:48)(tlI ) = N aturall
γlI
foreach a ∈ Al(t) do
ω
ω(δl(tl)) ;
I (l), tl ≤ tlI ∨ tl + dtl < tlI + dtlI do
I (l) : tl ≤ tlI ∧ tl + dtl > tlI do
(cid:48)(tlI ) = Decisionl
γlI
a
a(sa(tl + dtl));
end
end
γl(cid:48)(t) = {γl(t)(cid:83)
end
foreach l ∈ L : tl + dtl ∈ min(t + dt) do
a∈AlI
(cid:48)(t)};
<δl(tl + dtl), Tl >= Reactionl(σl(tl), γl(cid:48)(tl));
(cid:48)(t)(cid:83)
I (l) γlI
−
lI∈N
γlI
a
ω
end
19
20 end
3
IRM4MLS in practice: multi-level games of
life
3.1 Introducing a macroscopic parameter (top-down con-
trol)
In this section, a toy IRM4S (or 1-level IRM4MLS) model, is presented: a modified
agent-based version of the Conway's game of life (or simply Life). This simple example
illustrates where a macroscopic parameter, should (but should not) be introduced in
an agent-based model that relies on the influence reaction principle:
in the reaction
function (but not in the behavioral functions of agents). Therefore, this parameter
has a non-ambiguous semantics that does not depend on the updating scheme of the
simulations, even in the case of strong interaction 4.
Each agent represents a cell that can be dead or alive and that has eight neighbors
4Strong interaction implies that agents agree on the outcome of the interaction [20]. Thus,
such model should not be simulated with a STRIPS-like meta-model, i.e., that views an action
as a change of the state of the system. It would lead to problems of result replication [1], but
also of parameter and result interpretation [8, 20].
4
kernel
<<interface>>
Simulation
+void runSimulation(String logFile)
+void init()
+List<Agent> getAgents()
+List<Level> getLevels()
+Time getInitialTime()
+Time getFinalTime()
<<interface>>
Natural
+Map<Level,Map> natural()
1..*
1..*
<<interface>>
Environment
0..*
0..*
<<interface>>
Agent
0..*
1..*
+Map<Level, Natural> getLevels()
1
1..*
<<interface>>
Level
+Map<Level, Behavior> getLevels()
+void memorize()
0..*
0..*
1..*
1..*
<<interface>>
Behavior
+List<Agent> getAgents()
+Environment getEnvironment()
+List<Level> getPerceptibleLevels()
+List<Level> getInfluenceableLevels()
+DynamicState getDynamicState()
+void reaction()
1
1
<<interface>>
DynamicState
+void perceive()
+Map<Level,Map> decision(List<Level> li)
+Map getTemporaryInfluences()
+void addTemporaryInfluences(Map influences)
+Time getTime()
+Time getNextTime()
Figure 2: Java API of IRM4MLS
in a toroidal grid. The set of environmental properties is then:
{neighbors(a), alive(a)}.
∀t, σ(t) =
(cid:91)
a∈A
(8)
Cells evolve in parallel: the reaction function can then simply be defined as "ap-
plying agent influences" (algo. 2). If a cell is dead and has three living neighbors or is
alive and has two or three living neighbors, it will be alive at the next step; in other
cases, it will be dead. Let specify the behavior of the agents:
1. they perceive the number of living cells in their neighborhood (algo. 3),
2. memorize their internal state, i.e., their next state (algo. 4),
3. and then, decide whether or not they will be alive at the next step according to
their internal state (algo. 5).
The environment is static: there is no natural dynamics and thus, N aturalω returns
∅.
One cruel and ironic aspect of Life is that a cell has generally little chance to
remain alive in the long run (fig. 3(a)). Moreover, what you get most of the time,
5
Algorithm 2: Reaction
Input: σ(t), γ(cid:48)(t)
Output: δ(t + 1)
alive(a) = γ(cid:48)
1 foreach a ∈ A(t) do
a(t) ;
2
3 end
Algorithm 4: M emorizationa
Input: pa(t), sa(t)
Output: sa(t + dt)
1 if alive(a) ∧ pa(t) ∈ {2, 3}∨
2 ¬alive(a) ∧ pa(t) ∈ {3} then
sa(t + dt) = 1 ;
sa(t + dt) = 0 ;
3
4 else
5
6 end
Algorithm 3: P erceptiona
Input: δ(t)
Output: pa(t)
(cid:88)
1 pa(t) =
n∈neighbors(a)
alive(n) ;
Algorithm 5: Decisiona
Input: sa(t + dt)
Output: γ(cid:48)
a(t) = sa(t + dt) ;
a(t)
1 γ(cid:48)
is a board composed of small still lifes and 1-period oscillators. This behavior is
predictable knowing the λ parameter of the game of life5. λ is a complexity measure
of cellular automata introduced by [15] that depends on the number of cell states K,
the neighborhood N and the number n of transitions to a quiescent state Sq in the
transition function such as
with K = 2, N = 9 and n =(cid:0)8
(cid:1) + 2 ·(cid:80)
λ = 1 − n
K N ,
i=0,1,4−8
(cid:0)8
(cid:1) for Life.
i
2
To improve λlif e to a value λ+, one needs to change the rules. In this example,
λ+ is regarded as a macroscopic parameter, explicitly introduced in the model and
independent from cell behaviors6: influences of dying cells are not taken into account
by the reaction function with a probability p such as
n =(cid:0)8
(cid:1) + (2 − p) · (cid:88)
2
(cid:0)8
(cid:1).
i
i=0,1,4−8
(9)
(10)
Thus, their is a simple linear relation between p and λ+: p = (λ+ − λlif e)/0.3359.
For λ+ ∈ [0.48, 0.6], the number of dying cells tends to decrease in time and large
structures of vertical or horizontal rows eventually emerge, shaped by moving groups
of switching state cells that seem to work at their boundaries, and eventually vanish
when the board becomes a dense still life of density ≈ 0.5 (fig. 4).
5λlif e = 0.2734375. For λ ≈ 0.25, "structures of period 1 appear. Thus, there are now three
different possible outcomes for the ultimate dynamics of the system, depending on the initial
state. The dynamics may reach a homogeneous fixed point consisting entirely of state Sq, or
it may reach a heterogeneous fixed point consisting mostly of cells in state Sq with a sprinkling
of cells stuck in one of the other states, or it may settle down to periodic behavior" [15, p.
17].
6An other macroscopic parameter, the asynchrony, has been previously introduced in Life
is such way [9].
6
2
/
ρ
σ
±
¯ρ
0.5
0.4
0.3
0.2
0.1
0
2
/
ρ
σ
±
¯ρ
0.5
0.4
0.3
0.2
0.1
0
0
200 400 600 800 1000
step
0
200 400 600 800 1000
step
(a) Macroscopic level
(b) Mesoscopic level
Figure 3: Dynamics of the density of living cells, ρ, starting from a random grid
with 100 replications (a) Macroscopic dynamics: expected density and variabil-
ity of living cells in the whole grid. (b) Mesoscopic dynamics: expected density
and variability of living cells in 10 × 10 cell clusters.
3.2 Top-down feedback control
In the previous example, while the macroscopic parameter λ+ has an influence on
agents, it is not related to the state of the system and therefore, there is no need to
observe it. The reaction function can then be viewed as an open-loop controller. In
the next example, a top-down feedback control is introduced.
The goal of the multi-scale model presented in this section is to keep Life boards at
the desired density (ρ+), by controlling the proportion of dying cells at the mesoscopic
level, to account for the natural variability of density between regions of the grid
(fig. 3(b)). Moreover, the control should affect as less as possible simulations at the
microscopic level and, to keep it simple, should be tuned by a single linear parameter
such as λ+ in the previous model.
Two levels are considered: the cell (or microscopic) level, lm and the cell region (or
mesoscopic) level, lM . At the mesoscopic level, the model properties are the expected
density and the cells in each region:
∀t, σlM (t) = {ρ+} (cid:91)
{cells(alM )}.
(11)
alM ∈AlM
The cells behave according to the game of life rules (algo. 3, 4 and 5). However,
Reactionlm depends on mesoscopic influences (algorithm 6). EP and EI are equal to
{lM lm}. Mesoscopic agents have a proportional control behavior. They
1. perceive the density of living cells in a region,
2. memorize their internal state, i.e., the difference between expected and actual
densities,
3. and then decide the influence sent to agents of lm:
∀alM ∈ AlM∀alm ∈ cells(alM ), command(alm ) = kP · .
(12)
7
s
p
e
t
s
f
o
r
e
b
m
u
n
20000
18000
16000
14000
12000
10000
8000
0.45 0.5 0.55 0.6 0.65
λ+
(a) Convergence conditions
(b) Simulation example
Figure 4: (a) Mean number of steps needed to converge to a steady state starting
from a random grid (100 replications), simulations are stopped after 2 · 104
steps (simulations converge in the dark area). (b) Example of still life found for
λ+ = 0.5.
Algorithm 6: Reactionlm
Input: σlm(t), γlm(cid:48)(t)
Output: δlm(t + 1)
1 foreach alm ∈ alm do
rand ∈ [0, 1[ from pseudorandom uniform distribution ;
if command(alm) > rand ∧ alive(alm) ∧ ¬γ(cid:48)
alm (tlm) then
2
3
4
5
6
alive(alm) = (cid:62) ;
alive(alm) = γ(cid:48)
else
alm (tlm) ;
end
7
8 end
The kP parameter has to be carefully tuned to run realistic simulations: too small
simulations do not achieve the desired solution (¯ρ = ρ+), too big the board density
tends to oscillate around ρ+ and the number of micro influences not taken into account
by Reactionlm becomes too important. However, for appropriate kP values, this simple
linear controller achieves good results and allows to find a good compromise between
conflicting micro and meso knowledge (fig. 5).
4 Conclusion
4.1 Discussion
An important issue of multi-level agent-based modeling, only briefly discussed here,
is to define the adequate methodology. Indeed, the traditional MABS methodology is
purely bottom-up: microscopic knowledge is used to construct models while macro-
8
0.2
0.15
0.1
0.05
0
r (%)
ρ+
ρ
0
200 400 600 800 1000
step
(a) ρ+ = 0.07
0.2
0.15
0.1
0.05
0
r (%)
ρ+
ρ
0
200 400 600 800 1000
step
(b) ρ+ = 0.09
Figure 5: Simulation examples with kP = 10·ρ+ and initial density ρ(t0) = 2·ρ+.
r represents the rate of microscopic influences not taken into account (in %) by
Reactionlm.
scopic knowledge is used to validate models [6]. Thus, it seems irrelevant in a multi-
level context. Three general conclusions can be drawn from the previous examples:
• a parameter should be introduced at its observation level in the model; therefore,
each observed level should be explicitly represented in the model,
• an intuitive way to model an external control on a level l is to modify the reaction
function of l, i.e., to modify the way influences of agents of l are taken into
account (but not agent behavioral functions), according to external influences;
an external feedback control implies both observation and influence relations:
N
• such a controller can be viewed as a technical tool that aims to find a compro-
mise between conflicting knowledge from the different studied levels and achieve
realistic simulations from different studied points of view [16].
−
I (l) = N
−
P (l),
Pattern oriented modeling (POM) consists in "the multi-criteria design, selection
and calibration of models of complex systems" [12]. Many aspects of this methodology,
developed in the context of ecology, seem particularly relevant for multi-level agent-
based models since "patterns" are generally observed at different levels of organization
in complex systems. However, the problem is far from being solved. For instance,
the introduction of a dynamic level of detail raises several questions regarding, e.g.,
the validation of the model or the representation of composite agents [25, 27]. A
case study of three real world multi-level agent-based models reveals other interesting
methodological issues [11].
4.2 Related works
At least three works could be related to this one:
• ML-DEVS is an extension of DEVS that allows the simulation of multi-level
models (and not only coupled models in which the behavior of a model is deter-
mined by the behaviors of its sub-models) [33]. Two types of relation between
9
levels are defined: information propagation and event activation which are quite
similar to those defined in IRM4MLS. However, ML-DEVS supports only pure
hierarchies of models, i.e., interaction graphs are viewed as trees [18]. DEVS, as
a generic event-based simulation framework, has also been extended to support
agent-based models [24]. A major design difference between IR and DEVS based
approaches is the technical orientation of the latter leading to an important gap
between conceptual and computational models.
• PADAWAN (Pattern for Accurate Design of Agent Worlds in Agent Nests)
is a multi-scale agent-based meta-model based on a compact matricial repre-
sentation of interactions: IODA (Interaction-Oriented Design of Agent simula-
tions) [14, 29]. Moreover, authors analyze the structure of what is called here
interaction graphs in multi-scale models (a relevant issue for IRM4MLS as well),
and conclude they should be viewed as upper semilattices and not simply trees
as suggested elsewhere. A major design difference between IR and IODA based
approaches is that the latter constraints the definition of interactions, leading
to a simple but restrictive simulation framework.
• GAMA7 is a MABS platform with a dedicated modeling language, GAML, that
offers multi-level capabilities [32]. Moreover, it includes a framework (a set
of predefined GAML commands) to agentify emerging structures [34].
It is
certainly the most advanced platform, from an end-user point of view, that
integrates a multi-level approach.
4.3 Perspectives
The main perspectives of this work concern the implementation of existing works with
IRM4MLS:
• the concept of PolyAgent [26, 28],
• multi-level organizational models widely used in engineering sciences such as
holonic multi-agent systems (cf. footnote 2), system of systems and heterarchical
control [4, 17, 22],
• multi-scale tools: generic scaling operators and emergence detection and reifica-
tion algorithms [3, 25].
Moreover, the first model presented in this paper could be used to explore the
relations between computational capabilities of a cellular automaton (λ+), noise (a
function of λ+ − λ) and entropy. Moreover, finding the conditions under which cells
arrange themselves in a steady state could be an interesting way to solve heuristically
large instances of the maximum density still life problem [7].
The second model illustrates a form of simple proportional top-down feedback
control. Such approach could be generalized to model more complex of cross-level
feedback control, using integrations and derivates of observed variables.
References
[1] Bigbee, A., Cioffi-Revilla, C., Luke, S.: Replication of sugarscape using mason.
In: Terano, T., Kita, H., Deguchi, H., Kijima, K. (eds.) Agent-Based Approaches
7http://code.google.com/p/gama-platform/
10
in Economic and Social Complex Systems IV, Agent-Based Social Systems, vol. 3,
pp. 183 -- 190. Springer (2007)
[2] Caillou, P., Gil-Quijano, J.: Simanalyzer : Automated description of groups dy-
namics in agent-based simulations. In: Proc. of 11th Int. Conf. on Autonomous
Agents and Multiagent Systems (AAMAS 2012) (2012)
[3] Chen, C.C., Clack, C., Nagl, S.: Identifying multi-level emergent behaviors in
agent-directed simulations using complex event type specifications. Simulation
86(1), 41 -- 51 (2010)
[4] Cossentino, M., Gaud, N., Hilaire, V., Galland, S., Koukam, A.: Aspecs: an
agent-oriented software process for engineering complex systems. Autonomous
Agents and Multi-Agent Systems 20(2), 260 -- 304 (2010)
[5] David, D., Courdier, R.: See emergence as a metaknowledge. a way to reify
emergent phenomena in multiagent simulations? In: Proceedings of ICAART'09.
pp. 564 -- 569. Porto, Portugal (2009)
[6] Drogoul, A., Vanbergue, D., Meurisse, T.: Multi-agent based simulation: Where
are the agents? In: Sichman, J., Bousquet, F., Davidsson, P. (eds.) Multi-Agent-
Based Simulation II, Lecture Notes in Computer Science, vol. 2581, pp. 1 -- 15.
Springer (2003)
[7] Elkies, N.: The still-life density problem and its generalizations. Voronoi's Impact
on Modern Science, Book I. Proceedings of the Institute of Mathematics of the
National Academy of Sciences of Ukraine 21, 228 -- 253 (1998)
[8] Fat`es, N., Chevrier, V.: How important are updating schemes in multi-agent
systems? an illustration on a multi-turmite model. In: Proc. of 9th Int. Conf. on
Autonomous Agents and Multiagent Systems (AAMAS 2010). vol. 1, pp. 533 -- 540
(2010)
[9] Fat`es, N.: Does "life" resist asynchrony? In: Adamatzky, A. (ed.) Game of Life
Cellular Automata, pp. 257 -- 274. Springer (2010)
[10] Ferber, J., Muller, J.P.: Influences and reaction: a model of situated multiagent
systems. In: 2nd International Conference on Multi-agent systems (ICMAS'96).
pp. 72 -- 79 (1996)
[11] Gil-Quijano, J., Louail, T., Hutzler, G.: From biological to urban cells: Lessons
from three multilevel agent-based models. In: Desai, N., Liu, A., Winikoff, M.
(eds.) Principles and Practice of Multi-Agent Systems, Lecture Notes in Computer
Science, vol. 7057, pp. 620 -- 635. Springer (2012)
[12] Grimm, V., Railsback, S.: Pattern-oriented modelling: a 'multi-scope' for predic-
tive systems ecology. Philosophical Transactions of the Royal Society B: Biological
Sciences 367, 298 -- 310 (2012)
[13] Gutknecht, O., Ferber, J.: The madkit agent platform architecture. In: Agents
Workshop on Infrastructure for Multi-Agent Systems. pp. 48 -- 55 (2000)
[14] Kubera, Y., Mathieu, P., Picault, S.:
Interaction-oriented agent simulations:
From theory to implementation. In: Proc. of the18th European Conf. on Artificial
Intelligence (ECAI'08). pp. 383 -- 387 (2008)
[15] Langton, C.: Computation at the edge of chaos: Phase transitions and emergent
computation. Physica D: Nonlinear Phenomena 42(1-3), 12 -- 37 (1990)
11
[16] Li, J., Ge, W., Zhang, J., Kwauk, M.: Multi-scale compromise and multi-level
correlation in complex systems. Chemical Engineering Research and Design 83(6),
574 -- 582 (2005)
[17] Maier, M.: Architecting principles for system of systems. Systems Engineering
1(4), 267 -- 284 (1998)
[18] Maus, C., John, M., Rohl, M., Uhrmacher, A.: Hierarchical modeling for com-
putational biology. In: Bernardo, M., Degano, P., Zavattaro, G. (eds.) Formal
Methods for Computational Systems Biology, Lecture Notes in Computer Sci-
ence, vol. 5016, pp. 81 -- 124. Springer (2008)
[19] Michel, F.: The IRM4S model: the influence/reaction principle for multiagent
based simulation. In: Proc. of 6th Int. Conf. on Autonomous Agents and Multi-
agent Systems (AAMAS 2007). pp. 1 -- 3 (2007)
[20] Michel, F., Gouaıch, A., Ferber, J.: Weak interaction and strong interaction in
agent based simulations. In: Multi-Agent-Based Simulation III, Lecture Notes in
Computer Science, vol. 2927, pp. 43 -- 56. Springer (2003)
[21] Moncion, T., Amar, P., Hutzler, G.: Automatic characterization of emergent
phenomena in complex systems. Journal of Biological Physics and Chemistry 10,
16 -- 23 (2010)
[22] Morvan, G., Dupont, D., Soyez, J.B., Merzouki, R.: Engineering hierarchical
complex systems: an agent-based approach -- the case of flexible manufactur-
ing systems. In: Service Orientation in Holonic and Multi Agent Manufacturing
Control, Studies in Computational Intelligence, vol. 402. Springer (2012)
[23] Morvan, G., Veremme, A., Dupont, D.: IRM4MLS: the influence reaction model
for multi-level simulation. In: Bosse, T., Geller, A., Jonker, C. (eds.) Multi-Agent-
Based Simulation XI, Lecture Notes in Artificial Intelligence, vol. 6532, pp. 16 -- 27.
Springer (2011)
[24] Muller, J.P.: Towards a formal semantics of event-based multi-agent simulations.
In: David, N., Sichman, J. (eds.) Multi-Agent-Based Simulation IX, Lecture Notes
in Computer Science, vol. 5269, pp. 110 -- 126. Springer Berlin / Heidelberg (2009)
[25] Navarro, L., Flacher, F., Corruble, V.: Dynamic level of detail for large scale
agent-based urban simulations. In: Proc. of 10th Int. Conf. on Autonomous
Agents and Multiagent Systems (AAMAS 2011). pp. 701 -- 708 (2011)
[26] Parunak, H.: Pheromones, probabilities and multiple futures. In: Multi-Agent-
Based Simulation XI, Lecture Notes in Artificial Intelligence, vol. 6532, pp. 44 -- 60.
Springer (2011)
[27] Parunak, H.: Between agents and mean fields. In: Multi-Agent-Based Simulation
XII, Lecture Notes in Artificial Intelligence, vol. 7124. Springer (2012)
[28] Parunak, H., Brueckner, S.: Concurrent modeling of alternative worlds with
polyagents. In: Multi-Agent-Based Simulation VII, pp. 128 -- 141. Lecture Notes
in Computer Science, Springer (2007)
[29] Picault, S., Mathieu, P.: An interaction-oriented model for multi-scale simulation.
In: Twenty-Second International Joint Conference on Artificial Intelligence (2011)
[30] Pr´evost, G., Bertelle, C.: Detection and reification of emerging dynamical ecosys-
tems from interaction networks. In: Complex Systems and Self-organization Mod-
elling, Understanding Complex Systems, vol. 39, pp. 139 -- 161. Springer (2009)
12
[31] Soyez, J.B., Morvan, G., Merzouki, R., Dupont, D., Kubiak, P.: Multi-agent
multi-level modeling -- a methodology to simulate complex systems. In: Proceed-
ings of the 23rd European Modeling & Simulation Symposium (2011)
[32] Taillandier, P., Vo, D.A., Amouroux, E., Drogoul, A.: GAMA: A simulation
platform that integrates geographical information data, agent-based modeling
and multi-scale control. In: Desai, N., Liu, A., Winikoff, M. (eds.) Principles and
Practice of Multi-Agent Systems, Lecture Notes in Computer Science, vol. 7057,
pp. 242 -- 258. Springer (2012)
[33] Uhrmacher, A.M., Ewald, R., John, M., Maus, C., Jeschke, M., Biermann, S.:
Combining micro and macro-modeling in devs for computational biology. In: Pro-
ceedings of the 39th conference on Winter simulation. pp. 871 -- 880 (2007)
[34] Vo, D.A., Drogoul, A., Zucker, J.D., Ho, T.V.: A modelling language to represent
and specify emerging structures in agent-based model. In: Desai, N., Liu, A.,
Winikoff, M. (eds.) Principles and Practice of Multi-Agent Systems, Lecture Notes
in Computer Science, vol. 7057, pp. 212 -- 227. Springer (2012)
[35] Zeigler, B., Kim, T., Praehofer, H.: Theory of Modeling and Simulation. Aca-
demic Press, 2nd edn. (2000)
13
|
0912.2826 | 1 | 0912 | 2009-12-15T09:28:33 | An iterative approach for generating statistically realistic populations of households | [
"cs.MA",
"cs.CY"
] | Background: Many different simulation frameworks, in different topics, need to treat realistic datasets to initialize and calibrate the system. A precise reproduction of initial states is extremely important to obtain reliable forecast from the model. Methodology/Principal Findings: This paper proposes an algorithm to create an artificial population where individuals are described by their age, and are gathered in households respecting a variety of statistical constraints (distribution of household types, sizes, age of household head, difference of age between partners and among parents and children). Such a population is often the initial state of microsimulation or (agent) individual-based models. To get a realistic distribution of households is often very important, because this distribution has an impact on the demographic evolution. Usual techniques from microsimulation approach cross different sources of aggregated data for generating individuals. In our case the number of combinations of different households (types, sizes, age of participants) makes it computationally difficult to use directly such methods. Hence we developed a specific algorithm to make the problem more easily tractable. Conclusions/Significance: We generate the populations of two pilot municipalities in Auvergne region (France), to illustrate the approach. The generated populations show a good agreement with the available statistical datasets (not used for the generation) and are obtained in a reasonable computational time. | cs.MA | cs | An iterative approach for generating statistically realistic populations of
households
Floriana Gargiulo1,*, Sônia Ternes1,2, Sylvie Huet1, Guillaume Deffuant1
1 LISC, Cemagref, Clermont Ferrand,France
2 Embrapa Agricultural Informatics, Campinas - SP, Brazil
*[email protected]
Abstract
Background
Many different simulation frameworks, in different topics, need to treat realistic datasets to
initialize and calibrate the system. A precise reproduction of initial states is extremely important to
obtain reliable forecast from the model.
Methodology/Principal Findings
This paper proposes an algorithm to create an artificial population where individuals are described
by their age, and are gathered in households respecting a variety of statistical constraints
(distribution of household types, sizes, age of household head, difference of age between partners
and among parents and children). Such a population is often the initial state of microsimulation or
(agent) individual-based models. To get a realistic distribution of households is often very
important, because this distribution has an impact on the demographic evolution. Usual techniques
from microsimulation approach cross different sources of aggregated data for generating
individuals. In our case the number of combinations of different households (types, sizes, age of
participants) makes it computationally difficult to use directly such methods. Hence we developed a
specific algorithm to make the problem more easily tractable.
Conclusions/Significance
We generate the populations of two pilot municipalities in Auvergne region (France), to illustrate
the approach. The generated populations show a good agreement with the available statistical
datasets (not used for the generation) and are obtained in a reasonable computational time.
1 Introduction
With the increasing computing power, researchers tend to develop models which include more and
more diversity and details. A considerable effort has been made, both in academic and corporate
research, to generate modelling frameworks simulating policy impacts on complex dynamics: from
traffic studies [1] to epidemic diffusion [2,3,4,5], to policy impact studies [6,7,8,9]. These
approaches require using various sources of data, detailed at local level to test scenarios with
different policies (for instance mitigation strategies) and analyse their impact. For instance, an
increasing research effort targets the simulation of epidemic evolution: starting from SARS [3], to
the new virus of Influenza A (H1N1) [10]. Many different simulations, at global level or at local
level aim at providing precise forecast on the number of infected, with the actuation of different
containment strategies. One can expect that such tools become more and more commonly used to
support political decisions.
Many models consider populations with an explicit representation of each individual or of the
household structures. These individuals are characterised by some state variables (e.g. age,
profession, marital status), and often a spatial position. Two main types of modelling approaches
can be identified in the literature - Microsimulation and dynamical Individual Based Models (IBMs)
1
1: The microsimulation approach defines individual economic and social trajectories through a set
of events which occur with given probabilities, generally neglecting interactions between
individuals. It provides a mechanism to analyse the effects of policy changes at the level of the
decision making units as individuals and households. Individual based models (IBMs) consider the
same type of population but generally include more elaborated models of decisions and actions,
where individuals take into account the interactions with their environment and other individuals.
In both cases, the dynamics of the whole system is given by the aggregation of all individual
behaviours. Hence these modelling approaches are often used to explore the link between the micro
and macro dynamics. For instance models of evolving human populations yield demographic
patterns in geographical space, which can be compared with census-based data [11].
In both approaches, the first step for the simulation is to initialize the system with a realistic
population: the state variables defining the agents or the individuals, must replicate, as closely as
possible, the statistical properties of the targeted population. In particular, the demographic
evolution must take into account the structure of the distribution of households. Indeed, for the
same age structure of the population, different household structures evolve differently.
If individual data were available about the household structure, the problem would be solved
quickly by creating a one to one correspondence between the agents and the real persons. However,
such a situation rarely occurs, because the institutes managing statistics usually provide aggregated
datasets, describing the global properties of the households and individuals. Therefore we must use
these aggregate data to generate the artificial sets of individuals and households.
This paper focuses on the specific case of generating a population distributed in households to
initialise a dynamical microsimulation model for the PRIMA project. PRIMA – Prototypical Policy
Impacts on Multifunctional Activities in Rural Municipalities – is a European project (FP7-ENV-
2007) which aims to model the impact of European policies on land use at municipality level in a
set of case study regions. Hence in this project, the microsimulation process represents a population
of individuals at municipality level, living in households of different types. Once generated, the
initial synthetic population evolves through different processes such as birth, death, marriage,
divorce, leaving parental house, getting a job and retirement. The quality of the final results
depends heavily on the accuracy with which the initial synthetic population represents the available
real data.
According to the literature, there are two approaches commonly used to create a synthetic
population. In the first approach, some data at individual level are used to create the synthetic
population. For instance in the SVERIGE model [9], the whole population of Sweden in 1990 is the
starting population, and large longitudinal data sets are used for estimation of many equations for
the demographic process. In a similar way, DYNAMOD [12] is a dynamic model designed to
project population characteristics over a 50-year period, using a 1% sample. A second approach
uses the Iterative Proportional Fitting [13] to estimate the joint probability of characteristics
belonging to different sets of aggregated data. This approach is used in the SMILE model [6] where
the synthetic population is generated from Census of Small Area Population Statistics (SAPS) in
1996 in Ireland, considering characteristics as gender, age, employment status and industry, for a
given group of the population in a specific location. IPF can be applied when the Census data,
describing the aggregate properties of individuals and households, are integrated with individual
1 Sometimes they are also called "agent based models", because the individuals represent economic or social agents. But
there is an ambiguity with a different research trend of "agent based models", more related to computer science, which
investigates computer agents that cooperate for achieving some tasks, for instance foraging on the internet. To avoid
this ambiguity, we prefer to use the expression "Individual Based Models", which originally comes from modelling in
ecology.
2
data, extracted by surveys on samples that can be bigger or smaller than the size of the desired
artificial population. Thus, the initialization process consists in finding the good weight to attribute
to each sub-element of the analyzed sample to make it representative of the objective population.
Some methods to solve the up-scaling or downscaling initialization problem, with stochastic and
deterministic approaches, are described in [14,15,16,17,18]
In our problem, individual data to cross with the aggregate properties are not available. This
situation does not allow us to apply the IPF method. Moreover computing the joint probability of
characteristics of households, including size, type and age of members, implies heavy
computations. In this paper, we propose an iterative semi-stochastic algorithm, involving a
sequence of stochastic extractions, which considerably decreases the computational cost of the
population generation. This algorithm uses only aggregated datasets from Census, and the missing
crossings between the data are obtained through testing procedures.
The algorithm is adjusted for data from the Auvergne region (France), but the general concept can
be easily adapted to different uses. The next section describes the details of the problem to solve.
Section 3 describes the available data in Auvergne region, as well the attributes of the synthetic
population to be generated. The iterative algorithm is described in detail in section 4. Sections 5 and
6 present the results and conclusions.
2 Materials and Methods
2.1 General Formulation of the problem
The classical generation approach only considers one micro level (individuals or households). The
specificity of this work is that we need to respect statistical constraints on the distribution of the
individual ages, the distribution of household size and the distribution of individual ages within
households.
More precisely our problem is to generate a set of households comprising individuals taken in a
distribution of age of the population, and which respect all the constraints we found in the data
about the distributions of:
-
sizes and types of households,
-
ages of the head of the household,
- differences of age between partners,
-
ages of children according to mother's age.
Let us call:
•
t the type of household, the values of t can be: 'single', 'couple', 'single-parent', 'complex';
• s the size of the household, the values of s can be: 1, 2, 3, 4, 5, >5;
• ar the age of the head of the household;
• a1, …, as-1 the age of the children of single-parent households;
• ar' , the age of the head's partner, and a1, …, as-2, the age of the children for couple
households
(ai ) generally represents the list of the ages of the household members.
•
In a first approach we would suppose that we are able to compute a good approximation of the
probability of a given household P(t, s, (ai)) (a possible method to compute these probabilities is
described in section 2.3). Then, a straightforward way to proceed is described in algorithm 1.
Algorithm 1:
3
1. Generate all possible households, considering all possible combinations of types, sizes and
ages of members;
2. Associate with each of these households, defined by the values of (t, s, (ai)), the probability
P(t, s, (ai));
3. Generate a void list H. Repeat, until the size of H reaches the expected number of
households:
a. Pick a household generated in step 1 according to its probability associated in step 2;
b. Add the household to list H.
4. Return H.
This algorithm shows a significant drawback. Although the average on a large runs of this
algorithms of the distribution of age will be close to the data, one can expect significant differences
between the age distribution of a specific run and the data. Since the data about the distribution of
ages are reliable in our problem, we would like to keep it as precise as possible in our approach.
This leads to algorithm 2, where we use the list of ages of individuals directly taken from the data,
and a probability of household P'(t, s, (ai)), independently from the distribution of ages in the
population:
Algorithm 2:
1. Generate a population of individuals following the age structure of the population. Let us
call it the list I = { aj} (to each element of the list an age is associated);
2. Generate all possible households, considering all possible combinations of types, sizes and
ages of members;
3. Associate with each of these households, defined by the values of (t, s, (ai)), the probability
P'(t, s, (ai)) of the household, independently from the age distribution of the population;
4. Generate a void list H. Repeat, until list I is void or a number N of iterations is reached:
a. Pick a household h generated in step 2 according to its probability P'(t, s, (ai));
b. If ages (ai) are included in I then remove them from I and copy household h in H.
5. Return H.
With algorithm 2, we guarantee to keep the final distribution of individual ages close to the data.
Generating the list of individuals following the age structure of the population is straightforward.
The Census data of 1990 [19], the starting point at which we initialize the model for the Auvergne
region, chosen in the PRIMA project as a pilot region to be studied, provides the age distribution for
the population at municipality level. Two municipalities are chosen to test the algorithm: Abrest,
which was composed by 964 households with a total population of 2545 individuals, and Bellerive-
sur-Allier, composed by 8530 individuals organized in 3520 households. The choice of these
municipalities was made arbitrary, considering the difference of sizes, for testing the algorithm.
These data, displayed in Figure 1, allow us to generate directly the list I of individuals following the
age structure of the population. Simply, going through all the age brackets, and for each one, we
add to the list the corresponding number of individuals.
However, the other steps of the algorithms involve several difficulties:
• To evaluate the probability of a given household. This will be addressed in section 2.2.
• To manage the complexity of the set of all possible households. This will be addressed in
section 2.3.
In general, the algorithm leaves some individual ages unused at the end, and generates a
smaller number of households than expected (because of the impossibility to find the
necessary individuals to fit the drawn households). This is also addressed in section 2.3
•
4
2.2 Calculating the probability of a household
Census data, [19], provide also some information about households: the size distribution, the age
distribution for people living alone (single households) and the age distribution of the head of the
household. Figures 2 to 4 show those available data for the two municipalities. From those data, we
can calculate the probability of each household.
Data of figure 2 provide us with P(s), the probability of having a household of size s.
Data of figure 3 provide us with P(ar s=1), the probability of age range of the head for households
of size 1 (single).
Data of figure 4 together with data of figure 3, provide with P(ar s>1), the probability of age range
of the head for households of size superior to 1.
Data of figure 5 provide us with P(t ar = α), the probability of a household type given the age of
the head equals α, and the probability P(child a = α) for a individual of age α to live in a
household without being the head or the partner (this means, as a "child"2). Involving this constraint
is very important to avoid to get households with very old parents (e.g. 90 years) and old children
(around 70).
Clearly these data at local level are not sufficient to characterize a household. We lack constraints
on the distribution of ages inside a given type of household. Hence we used some data at national
level about the age structure inside the households regarding the ages of parents on one hand, and
the ages of children on the other hand. Figures 6 and 7 show the national level data that we use to
calculate the probability of the structure of ages, [20,21].
Data of figure 6 provide us with P(ar' ar = α), the probability of the age of the head's partner, given
the age of the head.
From data of figure 7, we can derive P((ai) am = α, s = σ) the probabilities of children ages
knowing the number of children and that the age of the mother is α. We consider that in couple
households, the mother is the partner, and in single-parent households, the head is the mother.
We can now use these partial probabilities to evaluate the probability of a given household P(t, s,
(ai)). We must distinguish cases 'single', 'single-parent', 'couple':
= P(s=1)* P(ar s=1)
P'(t='single', s , ar)
P'(t='singlep', s, ar, (ai))
= P(s= σ)* P(ar s>1)*P('singlep' ar)* Π P(ai ar) P(child ai)
P'(t='couple', s, ar, ar', (ai)) =
P(s= σ))* P(ar s>1)*P('couple' ar)*P(ar'' ar) * Π P(ai ar') P(child ai)
This evaluation theoretically allows us to apply the approach of algorithm 2. However, to generate
all the combinations of households and picking one according to these probabilities is
computationally expensive. In the next section, we propose an iterative algorithm which is more
efficient computationally.
2 That is the definition of "child" for the French Census managed by INSEE
5
2.3 An iterative algorithm avoiding to generate all possible households
The principle of the algorithm is to build progressively the household, by picking its member(s)
according to the previously described probabilities, and, for each new member, to test if there is an
individual of this age in the list of individuals I. If not, we stop the process for this household and
begin to build another one.
The flux diagram describing the process is represented in Figure 8.
The algorithm consists of five main steps (see algorithm 3).
Algorithm 3
1. Pick the size of the household according to P(s);
2. Pick the age range of the head according to P(ars). If there is no individual in I of the age range,
the process is stopped and a new attempt for building a household is launched. Otherwise an
individual of the chosen age range is added to the household, and removed from list I;
3. If s > 1, pick a household type ('couple' or 'single-parent') according to P(tar). “Complex”
households are not considered at this stage.
4. If t = 'couple', pick the age of the partner according to P(ar''ar). Again, if there is no individual
in I of the chosen age range, then the household is abandoned, the head is put back to list I, and
a new attempt to build a household is launched. Otherwise an individual of the chosen age range
is added to the household and remove from list I;
5. We pick the age of children with probability P(ai ar)*P(child ai) for single-parent and P(ai
ar')*P(child ai) for couples. Again, for each child, if there is no individual in I of the chosen
age range, then the household is abandoned, its members put back to list I, and a new attempt to
build a household is launched. Otherwise an individual of the chosen age range is added to the
household and removed from list I.
This process is equivalent to pick one household according to its evaluated probability, and keeping
it if all the ages of its members are present in list I. Indeed, the process of picking the different
members of the household leads to the same overall probability to pick a household, and since the
attempt is cancelled as soon as one age is lacking in list I, it changes nothing to make these tests
iteratively.
Moreover, we can constrain even more the process by considering the list of household sizes which
is directly derived from the data. The rest of the process remains the same. Then we are sure to have
the right number of households, even though when algorithm 3 stops, some void households remain
in the list.
Indeed, the described algorithm should a priori be repeated until all the households of the list are
filled with all the individuals of the availability vector. However, this situation is never reached and
after the creation of almost all the households, the program reaches a point where no more
households can be achieved given the remaining individuals. For this reason, when this situation is
reached, the algorithm is stopped. The remaining households can be considered as “complex
structures”, namely all the housing solutions that cannot be placed into the usual categorization of
household type (single, couple, single-parent). A complex household can be, for example, a group
of students occupying the same dwelling or two familiar groups sharing the same location.
Therefore, since we do not have any information about these structures from the data sets, to
conclude the generation of the artificial population, the complex households are filled randomly
with the remaining individuals in the availability list.
6
3 Results
We tested the algorithm for two different municipalities in Auvergne: Abrest and Bellerive-sur-
Allier. The first one had a population of 2545 inhabitants in 1990, while the second one had 8530.
In the following we compare the statistical properties of the artificial population with the real
Census data. We use for the comparison both the data implicitly used in the building algorithm and
other national and municipality level data, which were not used in the generation process. We
calculate the distributions both for one single realization of the system and for a sequence of 100
realizations (the random nature of the algorithm leads to some variations from one run to the other).
By construction of the algorithm, the age distribution and the size distribution of the household are
directly derived from the data for the two villages. In Figure 9 we show the distributions of the age
of head for real data and the artificial population. The distribution of the age of head was used
inside the generation process, but the stochastic extractions from this distribution were spaced out
from various tests; for this reason we can expect some discrepancy between the real data and the
generated population.
As we can observe in Figure 9, the artificial population respects quite well the real distribution.
In Figure 10, we compare the obtained artificial population with the real distribution of number of
children in households. This particular data set was not used in the generation, so the comparison
can give an idea of the accuracy of the algorithm; this data set is reported in Table 1. Also in this
case we can observe a good agreement between the real data and the simulations.
The final comparison (Figure 11) regards the household typology. For this comparison we will not
use directly the data that we have used in the generation (the probability for a person to be in a
certain type of household) but another dataset containing the direct proportions of household types
at national level. This dataset is reported in Table 2.
In this case the differences from the real data, for both municipalities, are quite significant. It could
be expected: the data we are using in this case for the comparison are at national data, and therefore
keep into account of the population of metropolitan areas. The discrepancy between our results and
the national data, therefore, do not highlight an error in the generating process, but show the
behavioural difference between metropolitan area and rural villages, with small population.
Moreover, it is noticeable that the data reported in the previous graph provide relevant information
about the complex households. We lack completely this information at village level and therefore
we cannot use any constraint on complex households in the building procedure. In the proposed
algorithm, complex households are created randomly, grouping together the individuals that the
generating procedure cannot assign to a household according to the selection/test mechanism.
Nevertheless, we observe that the proportion of complex households that we obtain is close to the
data at national level.
Finally, we need to stress that this kind of algorithm is strictly correlated to the data structure we
have: for Auvergne region such as for France and most of occidental countries the main household
structures are based on the concept of “nuclear family”: a couple of parents and a certain number of
children, or a subset of this structure. In some other cultures the basic household can have
completely different structure (for example many generations sharing the same dwelling), and
therefore this kind of approach can give rise to potential bias without any additional information
about the structure of complex households.
7
4 Conclusion
In this paper we proposed an algorithm for the generation of a synthetic population organized in
households that can be applied in various modeling contexts. This method gives good results
without using a set of prototypical households that, in many cases, are not available. This is an
advantage compared with existing methods such as IFT. Moreover it allows one to reproduce
exactly some features of the real population that are particularly important for the subsequent
analysis.
This algorithm is a practical implementation of a general approach where the households are picked
according to their probability, among all the possible household structures. The method builds the
households iteratively. It tests the availability of the age of its members at each step, and backtracks
as soon as an age is lacking. This saves a lot of computations.
We presented the example of the PRIMA project, where the artificial population is needed as
initialization of a dynamical microsimulation model at municipality level. We showed that the
algorithm yields a good agreement between the statistics of the artificial population and the real
one. Clearly, the approach can be adapted to other cases where it is necessary to generate a
population organized in households. During the project, we shall have to adapt it to other sets of
data that can be found in different case study regions.
The algorithm can deal with other properties of the individuals and of the households. For instance,
we could add a gender variable to describe the individuals of our example. We would need to split
the list I of individuals of different ages into two lists, one for males and one for females.
Moreover, we would need to include the percentage of household where the head is a male and
about the percentage of heterosexual couples. Then the principle of the method remains the same.
The only difference is that to build the households, we pick either in the list of males or in the list of
females.
More generally, after the set-up of the demographical structure, other characteristics can be
assigned to each individual, through stochastic extractions or deterministic associations: the level of
instruction, the professional activity, the favorite recreational activities, the commuting pattern, etc.
According to the available datasets, these properties can be assigned to each individual
independently from the household in which it is embedded, or some correlations can be considered
inside the same household.
Acknowledgment
This publication has been funded under the PRIMA (Prototypical policy impacts on
multifunctional activities in rural municipalities) collaborative project, EU 7th Framework
Programme (ENV 2007-1), contract no. 212345.
References
1. Nagel K, Beckman RL, Barrett CL. TRANSIMS for transportation planning; 1999.
2. Colizza V, Barrat A, Barthelemy M, Valleron A, Vespignani A (2007) Modeling the worldwide
spread of pandemic influenza: Baseline case and containment interventions. PLoS Medicine
4: 95.
3. Colizza V, Barrat A, Barthélemy M, Vespignani A (2007) Predictability and epidemic pathways
in global outbreaks of infectious diseases: the SARS case study. BMC medicine 5: 34.
4. degli Atti MLC, Merler S, Rizzo C, Ajelli M, Massari M, et al. (2008) Mitigation measures for
8
pandemic influenza in Italy: an individual based model considering different scenarios.
PLoS One 3.
5. Eubank S, Guclu H, Anil Kumar VS, Marathe MV, Srinivasan A, et al. (2004) Modelling disease
outbreaks in realistic urban social networks. Nature 429: 180-184.
6. Ballas D, Clarke GP, Wiemers E (2005) Building a dynamic spatial microsimulation model for
Ireland. Population, Space and Place 11: 157-172.
7. Ballas D, Clarke GP, Wiemers E (2006) Spatial microsimulation for rural policy analysis in
Ireland: The implications of CAP reforms for the national spatial strategy. Journal of Rural
Studies 22: 367-378.
8. Gotts NM, Polhill JG, Law ANR (2003) Aspiration levels in a land use simulation. Cybernetics
and Systems 34: 663-683.
9. Holm E, Holme K, Mäkilä K, Mattson-Kauppi M, Mörtvik G (2004) The microsimulation model
SVERIGE; content, validation and applications. SMC, Kiruna, Sweden (www sms kiruna
se).
10. Balcan D, Hu H, Goncalves B, Bajardi P, Poletto C, et al. (2009) Seasonal transmission
potential and activity peaks of the new influenza A (H 1 N 1): a Monte Carlo likelihood
analysis based on human mobility. BMC Medicine 7: 45.
11. Mahdavi B, O'Sullivan D, Davis P. An agent-based microsimulation framework for
investigating residential segregation using census data. In: Oxley LaK, D., editor; 2007. pp.
365-371.
12. King A, Baekgaard H, Robinson M (1999) DYNAMOD-2: An overview. Technical Paper 19.
13. Norman P (1999) Putting iterative proportional fitting on the researcher's desk. WORKING
PAPER-SCHOOL OF GEOGRAPHY UNIVERSITY OF LEEDS.
14. Ballas D, Clarke G, Dorling D, Rossiter D (2007) Using SimBritain to model the geographical
impact of national government policies. Geographical Analysis 39: 44.
15. Ballas D, Kingston R, Stillwell J, Jin J (2007) Building a spatial microsimulation-based
planning support system for local policy making. Environment and Planning A 39: 2482-
2499.
16. Birkin M, Turner A, Wu B. A synthetic demographic model of the UK population: methods,
progress and problems; 2006.
17. Smith DM, Clarke GP, Harland K (2009) Improving the synthetic data generation process in
spatial microsimulation models. Environment and Planning A 41: 1251–1268.
18. Williamson P, Birkin M, Rees PH (1998) The estimation of population microdata by using data
from small area statistics and samples of anonymised records. Environment and Planning A
30: 785-816.
19. INSEE (1990) French Population Census of 1990. Centre Maurice Halbwarchs, 48 Boulevard
Jourdan, Paris.
20. Eurostat (1999) Demography, fecondity data for 1999.
21. INSEE (1990) Enquête sur l'étude de l'histoire familiale de 1999. Centre Maurice Halbwarchs,
48 Boulevard Jourdan, Paris.
9
Figures
Figure 1: Histogram of the number of individuals according to various age ranges of 5 years each
for Abrest and Bellerive-sur-Allier. Source: INSEE, French Census data, 1990.
Figure 2: Histogram of the number of households according to their size (number of individuals in
the household) in Abrest and Bellerive-sur-Allier. Source: INSEE, French Census data, 1990..
10
Figure 3: Histogram of the number of households according to the age ranges of person living
alone in Abrest and Bellerive-sur-Allier. Source: INSEE, French Census data, 1990..
Figure 4: Histogram of the number of households according to the age ranges of the head in Abrest
and Bellerive-sur-Allier. Source: INSEE, French Census data, 1990.
11
Figure 5: Histogram of the number of individuals of age > 15 according to different age ranges,
from the top to the bottom, living as partners in couple, as head in single-parent households or
living with parent(s) in Abrest and Bellerive-sur-Allier. Source: INSEE, French Census data, 1990
12
Figure 6: Histogram of the number of couples according to their difference of ages in France in
1999. Source: INSEE, “Enquête sur l'étude de l'histoire familiale de 1999”.
Figure 7: distribution of live births by birth order and mother's age range in France. Source:
Eurostat Data 1999.
13
Figure 8: Flux diagram describing the algorithm for the generation of an artificial population for
PRIMA project
Figure 9: Histograms for age of head distribution for the municipality of Abrest (left plot) and of
Bellerive-sur-Allier (right plot). The light purple bars represents the real data, the dark purple bars
the average for 100 realizations for the artificial population. The error is the standard deviation on
the 100 replicas.
14
Figure 10: Histograms for age number of children distribution for the village of Abrest (left plot)
and of Bellerive-sur-Allier (right plot). The light purple bars represents the real data, the dark purple
bars the average for 100 realizations for the artificial population. The error is the standard deviation
on the 100 replica.
Figure 11: Histograms for the household type distribution for the village of Abrest (left plot) and of
Bellerive-sur-Allier (right plot). The light purple bars represents the real data, the dark purple bars
the average for 100 realizations for the artificial population. The error is the standard deviation on
the 100 replica.
Tables
15
Type
Household without child
Household with one child
Household with two children
Household with three children
Household with four or more children
ABREST BELLERIVE
1316
360
192
580
444
156
120
48
16
44
Table 1: Distribution of households according to the number of children for the municipalities
Abrest and Bellerive sur Allier. Source: INSEE, French Census data, 1990.
Type
Table 2: Distribution of households according to the type in France. Source: INSEE, 1990.
Single
Single-parent
Couple
Couple with Children
Complex
Proportion
0,2720
0,0660
0,2370
0,3640
0,0610
16
|
1705.09838 | 1 | 1705 | 2017-05-27T16:38:35 | Applying Artificial Intelligence and Internet Techniques in Rural Tourism Domain | [
"cs.MA"
] | Society has become more dependent on automated intelligent systems, at the same time, these systems have become more and more complicated. Society's expectation regarding the capabilities and intelligence of such systems has also grown. We have become a more complicated society with more complicated problems. As the expectation of intelligent systems rises, we discover many more applications for artificial intelligence. Additionally, as the difficulty level and computational requirements of such problems rise, there is a need to distribute the problem solving. Although the field of multiagent systems (MAS) and distributed artificial intelligence (DAI) is relatively young, the importance and applicability of this technology for solving today's problems continue to grow. In multiagent systems, the main goal is to provide fruitful cooperation among agents in order to enrich the support given to all user activities. This paper deals with the development of a multiagent system aimed at solving the reservation problems encountered in rural tourism. Due to their benefits over the last few years, online travel agencies have become a very useful instrument in planning vacations. A MAS concept (which is based on the Internet exploitation) can improve this activity and provide clients with a new, rapid and efficient way of making accommodation arrangements. | cs.MA | cs | APPLYING ARTIFICIAL INTELLIGENCE AND INTERNET
TECHNIQUES IN RURAL TOURISM DOMAIN
Cristina TURCU, Cornel TURCU
"Stefan cel Mare" University of Suceava
13, University Street, Suceava, 720225, Romania, tel.: +40-230-522978
e-mails: {cristina, cturcu}@eed.usv.ro
Abstract
Society has become more dependent on automated intelligent systems; at the same time, these systems have become
more and more complicated. Society's expectation regarding the capabilities and intelligence of such systems has also
grown. We have become a more complicated society with more complicated problems. As the expectation of intelligent
systems rises, we discover many more applications for artificial intelligence. Additionally, as the difficulty level and
computational requirements of such problems rise, there is a need to distribute the problem solving. Although the field of
multiagent systems (MAS) and distributed artificial intelligence (DAI) is relatively young, the importance and applicability of
this technology for solving today's problems continue to grow. In multiagent systems the main goal is to provide fruitful
cooperation among agents in order to enrich the support given to all user activities. This paper deals with the development of
a multiagent system aimed at solving the reservation problems encountered in rural tourism. Due to their benefits over the
last few years, online travel agencies have become a very useful instrument in planning vacations. A MAS concept (which is
based on the Internet exploitation) can improve this activity and provide clients with a new, rapid and efficient way of
making accommodation arrangements.
Introduction
Tourism is sometimes believed to be a remedy for the development of rural areas [1]. Especially in the
northern part of Romania where forestry and agriculture has lost their importance with respect to the number of
people employed tourism based on natural resources, rural tourism is seen as the solution or part of the solution
for creating jobs and a viable socio-economic situation.
Over the last few years, travel agencies have taken great pride in providing both individuals &
companies with the highest quality service and cost effective savings when planning business or vacation travel
arrangements. Constantly changing airfares and schedules, thousands of available vacation packages, and a vast
amount of travel information on the Internet can make travel planning frustrating and time-consuming. To sort
out the many travel options, tourists and business people often turn to travel agents, who assess their needs and
help them make the best possible travel arrangements. In general, travel agents give advice on destinations and
make arrangements for transportation, hotel accommodations, car rentals, tours, and recreation. For international
travel, agents also provide information on customs regulations, required papers (passports, visas, etc), and
currency exchange rates. Travel agents consult a variety of published and computer-based sources for
information on departure and arrival times, fares, and hotel ratings and accommodations.
So a travel agency has become a full service organization that handles travel plans from start to finish.
But in the Internet ages the personal computers can be used as private travel agencies. An online travel agency is
more convenient than a traditional one because there are many advantages using it: many hotel searching and
reservation services are on the web, which can offer online information about accommodation conditions and
prices; also time schedules for air, sea, railway and ground transportation are available (all) over the world; no
appointments and face to face meetings with travel agent are necessary. That's what a person gets when using
the online services of the travel agency Web site. With a click of a mouse, one can access many travel agency
benefits and services 24 hours a day, 365 days a year.
One important aspect in the rural tourism domain is related to booking. Traditional booking enforces
the use of telephone or fax. Due to the several attempts to make a contact the first method is unsatisfactory. Also
the fax method is characterized by large time delay. Using web based applications the booking process may be
substantially improved. But the access to information is efficient when one knows where to find the information
s/he is looking for. However, this task becomes complex when the information source is unknown. The
complexity considerably increases when one does not know the type of information useful for his or her purpose
and has only limited knowledge of the information sources put at his or her disposal without mentioning the fact
that the arrangement and use of these sources can considerably vary from one to another. Thus, it becomes
necessary either to ease the task of information retrieval, or to provide the users with some automation
mechanisms [6].
Multiagent systems
An agent is a software entity that applies Artificial Intelligence techniques to choose the best set of
actions to perform in order to reach a goal specified by the user. It should react in a flexible, proactive, dynamic,
autonomous and intelligent way to the changes produced in its environment. A multiagent system [8, 9] may be
defined as a collection of autonomous agents that communicate between themselves to coordinate their activities
in order to be able to solve collectively a problem that could not be tackled by any agent individually.
In the last years it has been argued that multiagent systems may be considered as the latest software
engineering paradigm [2, 7]. This kind of systems may be used in domains with the following features [5]:
Knowledge is distributed in different locations.
Several entities, while keeping their autonomous behaviour, have to join their problem-solving
abilities to be able to solve a complex problem.
The problems in the domain may be decomposed in different sub-problems, even if they have some
kind of inter-dependencies.
Due to the similarities with rural tourism domain the conclusion is that the multiagent systems seem to
be adequate to be used in booking process.
A multiagent system in rural tourism domain
During a booking process the user access a lot of distributed information sources that are now available
on electronic supports. So it is necessary to create some tools to be used in retrieval information. To be efficient,
such tools must be able to assess the context of an information request. The questions to be answered are the
following: Who is the user? For what purpose is the research being undertaken? What are the relevance and
quality of the retrieved information [3, 4]?
Obviously the implementation of an intelligent system can improve the whole booking process.
Automated computer reservation systems may provide clients with instant access to thousands of guesthouse
offers. Thus, clients can choose the best price and the fastest way of reaching their destination. Furthermore, one
of the most important functions of this system would be the obtainment of best offers from guesthouses. Due to
the rapid distribution of all travel information and the complexity of the planning process, intelligent systems can
reduce the allocated time to a client and increase service quality.
In the last few years many guesthouses were built in the Romanian area. A lot of them are included into
a national network named ANTREC1. Due to geographical distribution of the information one of the major
problems occurres during the booking process. The development of an intelligent system requires the assessment
of the agency's most important function: the client's request, alternatives to accommodation, notification on
accommodation alternatives, reservation and payment. The temporal distribution of the reservation steps is
presented in figure 1.
Figure 1. Reservation – temporal distribution
agents:
The system will be able to provide the following services:
The user may ask for a reservation satisfying certain preferences (e.g. those guesthouses that are
located in a particular area, price, facilities etc.).
The user is able to access his/her historical record.
The proposed multiagent architecture is presented in figure 2. It consists in the following types of
Personal agent: Each personal agent represents a human user. The personal agent trustfully keeps
information such as user profile and reservations made and is responsible for asking for possible
bookings, in compliance with the users' preferences. It also controls the access of the user interface
agent to the services provided by MAS, by avoiding direct communication with other agents.
User interface agent: The user interface agent provides a graphical interface of the MAS with the
user. This interface is used to introduce the requirements of a search or to show the results of a
query to the user. It is necessary to define personalized interfaces that permit an easy, flexible and
customizable access to the information the user need.
1 ANTREC is a non-profit association that identifies, develops and promotes Romanian rural hospitality and
tourism.
Alternatives to acommodation Notification on alternatives Reservation & payment Client Request Registration Registration National agent: it receives the users' preferences and sends a request to the zonal agents; also it
receives all possible reservations and sends a classification to the personal agent.
Zonal agent: For each zone there is one zonal agent and several guesthouse agents. The ZA has the
general information of the guesthouses (e.g. its address, telephone number etc.).
Guesthouse agent: It is one guesthouse agent for each guesthouse of a system. This agent could be
executing in the desktop computer of each guesthouse. A guesthouse agent can encapsulate a web
site, which provides reservation services (it access a database). It would also have a graphical
interface that allows the personnel to update the information and request specific information from
the database etc.
Figure 2. Multiagent system architecture (PA-Personal Agent, UIA-User Interface Agent, NA-National Agent)
Usually the system receives a user's request parameters set. These parameters refer to location wanted,
number of persons, arrival & departure days, number of rooms wanted (single/double/triple), maximum price
and other required facilities. If a guesthouse matchs the user's preferences then the correspondent guesthouse
agent will send an accepting message to ZA. The facilities of the guesthouse are presented in figure 3.
Figure 3. Guesthouse facilities
A possible reservation (from a zonal point of view) determination is presented in figure 4. In the
represented situation the zonal agent ZA1 receives the user's preferences and a request id (which is used as a
parameter in all ask messages) from national agent. It will ask the guesthouse agents (GA1_1, GA1_2,…) for
possible reservations. GA1_3 does not match the user's preferences and send a sorry message to ZA1. The
GA1_2 agent does match the preferences and will send a tell message to ZA1 about a possible reservation. The
GA1_1 agent does not match the preferences but it's possible to ask GA1_2 and GA1_3 for collaboration (for
example, GA1_1 cannot make a reservation for all the period requested by the user). GA1_2 already made a
reservation proposal (it checks id request) and will send a sorry message to GA1_1. If GA1_3 can complete the
period and match the user's preferences it will send a tell message to ZA1. So ZA1 has two possible reservations:
one from GA1_2 and another one from GA1_1 in collaboration with GA1_3.
The NA agent will receive all reservation proposals that meet the user's preferences from all zonal
agents. After that it will reason about all of them and make a classification depending on one or more criteria
(for example the price). The user will chose the best proposal and will initiate the booking process sending a
Guesthouse agent (GA) Zonal agent (ZA) NA PA UIA message to the proper GA (through the MAS chain).
If the user specifies a requested zone, the entire reservation process will be made without NA
participation. This implies the occurrence of direct communication between PA and ZA agents.
Figure 4. The reservation determination process
The system guarantees a secure access to information records, so that only authenticated users are
allowed to update the information. There is a two-level security mechanism:
Using a security plug-in of JADE (Java Agent Development Environment), JADE-S, all the
messages between the agents in the system are encoded using SSL.
The JADE-S module also allows the definition of different types of agents, and the assignment of
different permissions to each kind of agent. For example, a personal agent cannot access directly
any of the other agents in the system; it only has permissions to send and receive messages from
national agent or zonal agent.
To ask information users may use diverse information technology. So, this system can be developed by
considering the following easy and flexible communication ways: SMS messages through mobile phones,
standard e-mail messages from a PC etc. In this case, it is necessary to add a new agent, named Communication
Manager Agent. This agent is responsible of dealing with different types of information technologies. For
example, the user might carry Personal Digital Assistants (PDAs) when they are on the way. We have also
considered the possibility of attaching an intelligent agent to the PDA, who could request the information and
send it wirelessly to the communication manager agent. The communication manager agent, after receiving the
answer to the query could send this information to the agent permanently running in the PDA of a user.
Conclusions
Due to the distribution of information the booking process in rural tourism domain is actually quite
difficult. A solution is represented by a multiagent system approach. The multiagent system model seems to be
the adequate framework for dealing with the design and development of an application, which is flexible,
adaptable to the environment, versatile and robust enough to supply the booking process with efficiency and
reliability.
The benefits of the system would reach two kinds of users:
Citizens that need tourism information, who could obtain any kind of information related to the
guesthouses of a given zone and make a reservation.
Guesthouses personnel, who could access and update the guesthouse's information.
References
[1]. Hall, C.M. & J. M. Jenkins, "The policy dimensions of rural tourism and recreation", In R. Butler, C.M.
Hall and J. Jenkins (eds.) Tourism and Recreation in Rural Areas (pp. 19-42). Chichester: Wiley, 1998
[2]. Jennings, N., "On agent-based software engineering", Artificial Intelligence 117, 2000
[3]. King, J.A., "Intelligent Retrieval", AI Expert, 10.1, January, 15–17, 1995
[4]. King, J.A., "Intelligent Agents: Bringing Good Things to Life", AI Expert, February, 17–19, 1995
[5]. Moreno, A., "Agents applied in health care", Guest editorial, AI Commun 16(3):135-137, 2003
[6]. Pelletier, S., Pierre, S., Hoang, H., "Modeling a Multiagent System for Retrieving Information from
Distributed Sources", Journal of Computing and Information Technology - CIT 11, 1, 1–10, 2003
[7]. Petrie, C., "Agent-based software engineering", in: Agent-Oriented Software Engineering, 2001
[8]. Weiss, G., "Multiagent systems. A modern approach to Distributed AI", M.I.T. Press, 1999
[9]. Wooldridge, M., "An introduction to Multi Agent Systems", Wiley Ed., 2002
ask(user's preferences) NA ZA1 Cases ask(user's preferences) sorry(cause) tell(GA1_2 match) ask(collaboration) SUSPENDED Requested accommodation GA1_1 GA1_2 GA1_3 [sorry(cause)] tell(GA1_3 match) tell(GA1_1 match collab) tell(GA1_2, GA1_1 collab) NA |
cs/9903017 | 1 | 9903 | 1999-03-28T20:46:49 | SIMMUNE, a tool for simulating and analyzing immune system behavior | [
"cs.MA"
] | We present a new approach to the simulation and analysis of immune system behavior. The simulations that can be done with our software package called SIMMUNE are based on immunological data that describe the behavior of immune system agents (cells, molecules) on a microscopial (i.e. agent-agent interaction) scale by defining cellular stimulus-response mechanisms. Since the behavior of the agents in SIMMUNE can be very flexibly configured, its application is not limited to immune system simulations. We outline the principles of SIMMUNE's multiscale analysis of emergent structure within the simulated immune system that allow the identification of immunological contexts using minimal a priori assumptions about the higher level organization of the immune system. | cs.MA | cs |
SIMMUNE,
a tool for simulating and analyzing
Immune System behavior ∗
M. Meier-Schellersheim, G. Mack
II.Institut fur Theoretische Physik
Universitat Hamburg†
August 6, 2018
Abstract
We present a new approach to the simulation and analysis of immune system
(IS) behavior. The simulations that can be performed with our software package
called SIMMUNE are based on immunological data that describe the behavior
of IS agents (cells, molecules) and the IS's challengers (bacteria, viruses) on a
microscopical (i.e. agent-agent interaction) scale by defining cellular stimulus
response mechanisms. All processes within the simulated IS are based on these
mechanisms. Since the behavior of the agents in SIMMUNE can be very flexibly
configured, its application is not limited to IS simulations.
We outline the principles of SIMMUNE's multiscale analysis of emergent struc-
ture within the simulated IS that allow the identification of immunological con-
texts using minimal a priori assumptions about the higher-level organization of
the IS.
Keywords:
analysis.
immune system simulation, locally interacting agents, multiscale
1 Introduction
For quite a long time immunological research limited itself mainly to the inves-
tigation of molecular details of cellular mechanisms within the immune system
(IS). The structure of immune responses was believed to be rather simple: Upon
infection of the organism IS agents had to 'recognize' the foreign material that
caused the infection, the pathogen, with highly specific receptors and then re-
move everything within the organism that would bind to these receptors. To
∗Work supported by Deutsche Forschungsgemeinschaft and by the German Israeli Foun-
dation for Scientific Research and Development (GIF)
†email [email protected], [email protected]
1
make this principle work, the IS only had to ensure that it did not produce
receptors complementary to components of host organism.
From this point of view, developing a vaccine against some kind of disease thus
simply meant searching for the right (harmless) fragment of the pathogen that
could be presented to the IS as an antigen (a substance that provokes an im-
mune response). Having recognized this fragment once, the IS would memorize
this knowledge through maintaining a large enough number of the right recep-
tors and then upon contact with the real pathogen roll out its arsenal of defense
quickly enough to avoid a spreading of the pathogen within the organism.
In many cases vaccine development is more complicated. The IS normally looks
for more than just one signal (the antigen) before it produces a full response.
Given that the IS can hardly rule out every receptor that fits to material of the
own organism without crippling its own functionality because of the degeneracy
of receptors 1 it is understandable that it should ask for more information.
(For a mathematical discussion of the problem of immune receptor specificity
see for example [1].)
Providing the right adjuvans, the biochemical context within which the antigen
is seen by the IS, is important. Adjuvants often contain fragments of classical
pathogens, i.e. pathogens the organism and its ancestors have been accustomed
to for a long time. Being confronted to these pathogen fragments the IS more
readily switches into 'defense mode'.
Unfortunately, within certain contexts the IS readily accepts the organism's own
material as antigenic. Autoimmune diseases like diabetes are examples where
the IS, once it finds itself in a certain context, attacks its own host.
Cohen [2] formulated the cognitive paradigm postulating that cognitive abilities
- enabling the IS to select its response according to the context of the presenta-
tion of antigen - were an indispensable ingredient of IS behavior.
Grossman [4] explained why we should investigate the IS's context recognition
and proposed some models for appropriate cell behavior.
Segel and Bar-Or [5] investigated the question how the IS's cells might be able
to optimize their contribution to immune responses with the help of feedback
mechanisms and compared the IS to other systems of decentralized organizing
agents. Segel [6] also presented the idea of a diffuse informational network of
cytokines encoding the molecular context the IS finds itself in. Atlan and Cohen
[7] pointed out that the IS must achieve the ability to extract -- in a cognitive
process -- meaning from the wealth of information its cells gather via their re-
ceptors.
Often, the term 'recognition' in an immunological sense refers to the ability of
the IS to provide receptors that can bind to the surface of foreign material with
high enough affinity to direct an effective immune response against this mate-
rial. Recognition of this kind comes down to the question whether molecular
1IS receptors usually fit to more than just one single molecular structure. The neccessity
to be able to provide receptors to virtually any (foreign) molecular structure the IS might get
confronted to, prevents the IS from destroying every receptor that fits to the material of its
host organism.
2
shapes are complementary while the ability to act context dependent requires
the IS's agents to mutually coordinate the processing of the molecular signals
which they receive from their milieu.
How could the IS's information processing work to achieve context recognition?
What is the nature of the context, that make the IS respond in a protective (or
harmful) manner? How is the IS provided with the specific context that would
activate an immune response to some threatening disease?
To be able to start answering these questions, we need to investigate how the
agents of the IS exchange information, how they influence each other's states,
how their reactions to combined signals differ from their reactions to these sig-
nals when occurring at different times. Further, what is the spatial scope of
specific signals, i.e., do they influence only their direct neighborhood or do they
spread over larger areas of the organism?
The software package SIMMUNE which we want to introduce here was designed
to facilitate simulations dealing with these questions.
SIMMUNE simulates the IS on the agent level, i.e. on the level of interactions
between cells and molecules. The analysis that is performed on the simulations,
however, operates on multiple scales; this is described in section 3.4. Application
of the approach of identifying immunological contexts by multiscale analysis to
more elaborate simulations will be described in a forthcoming publication [15].
In section 2 we will briefly describe some of the classical methods of IS modelling
to be able to point out the differences and similarities between them and the
approach presented here. In section 3 our software package SIMMUNE will be
introduced. We will give an overview over its structure and present some simple
examples of SIMMUNE applications, as well as some of the methods to analyze
the simulations. Section 3 concludes with some remarks on the limitations of
the current version of SIMMUNE.
In section 3.3.2 we will describe some immunological mechanisms as far as they
are implemented in the example simulations. Descriptions of immunological
mechanisms refer to the way they are implemented in the simulations which
we present here. In those cases where they differ fundamentally from their real
biological counterparts we will mention this.
2 Modelling IS Behavior
The possibilities of tracing directly the complex sequences of interactions in real
ISs of living organisms are very limited. However, Jenkins [9] has presented a
method of monitoring the activities of selected cell clones in situ, i.e.
in the
living organism. (A definition of 'cell clones' as they are implemented in our
simulations will be given in chapter 3.3 .)
Various techniques of IS modelling have been developed that allow to investigate
theoretically different aspects of immunology. Perelson and Weisbuch [1] have
provided a comprehensive survey of this area.
3
2.1 Reaction-Kinetics Models
The great complexity of IS behavior is due to the large number of different
types of IS agents (cells, molecules) that can interact with each other in various
ways. Deriving the system's behavior from the interactions between its many
constituents is one of the goals of theoretical immunology.
Abstracting from the specific details of interaction between the IS's agents one
can formulate systems of coupled differential equations describing how the time
development of an agent's concentration in the modelled organism depends on
the concentrations of other types of agents. We call these models reaction-
kinetics models as they bear resemblence to models of reaction kinetics in
chemistry. A very simple example for such a system is the following.
dI
dt = pinf ectIC − pkillIK − dI I
dK
dt = prespIK − dK K
dC
dt = s − pinf ectIC − dC C
Using as a shortcut notation I, K and C for agent names as well as for
their concentrations, the equations above describe a situation that may be in-
terpreted as follows: Infectious agents of type I transform agents of type C into
new agents of type I upon contact with a rate pinf ect. Agents of type I get
removed (killed) upon contact with agents of type K with a rate pkill. Agents
K proliferate upon contact with I with a rate presp. Agents of all three types
die naturally at their specific rates dI , dK, dC . C type agents are produced
at a constant rate s. K could be considered to be an immune system cell type
I
K
C
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
n
o
i
t
a
r
t
n
e
c
n
o
c
t
n
e
g
a
0
0
100
200
300
400
500
time
600
700
800
900
1000
Figure 1: time development of agent concentration in the simple IS model
being produced as a response to the appearance of the infectious I. C may be
4
presenting any possible type of target cell for the pathogen I.
Integrating this system of equations yields different kinds of time development
for I, K and C depending on the parameters pinf ect, pkill, presp, s, the death
rates dx and of course the initial values of I, K and C.
Fig. 1 shows the time development of the agents for one set of parameters (see
appendix). We see how at first the number of infected agents I grows while
the number of 'healthy' agents C decreases. Then, as the response from the IS
agents K grows, the number of infected agents declines while C recovers. Finally
the system ends up in a steady state that may be interpreted as a chronic in-
fection: Infectious agents persist even though the IS constantly prepares agents
of type K to suppress the infection.
Other sets of parameters may lead to stronger oszillations of the agents con-
centrations. Fig. 2 shows in a half-logarithmic plot a system that settles down
into a steady 'chronic infection' state after having gone through states where
the concentration of I is very low.
n
o
i
t
a
r
t
n
e
c
n
o
c
t
n
e
g
a
1
0.1
0.01
0.001
0.0001
1e-05
1e-06
1e-07
I
K
C
0
200
400
600
800
1000
time
1200
1400
1600
1800
2000
Figure 2: The model IS settles down into 'cronic infection' after oscillations
with very low concentrations of I.
Fig. 2 illustrates one of the limits of applicability of simulation results of the
reaction-kinetics approach: In a real system of interacting entities at a given
time all agents of a certain type may have dissappeared due to destructive in-
teraction with other agents (in our example, such an interaction is for example
the suppression of I by K ). In the reaction-kinetics approach there are no in-
dividual agents. Even agent types with unrealistically low concentrations may
experience a 'comeback'. Of course, this problem can be avoided by appropriate
setup of the model, but it points out one of the weaknesses of the approach:
Information is global in reaction-kinetics models. In contrast, information in
nature is something local.
In physics, this principle of locality is one of the major foundations of modern
5
theories. It lies at the heart of field theory and can be used as a starting point
for entering Einstein's General Theory of Relativity.
In immunology, the locality of information processing by the IS's agents may
prove to play an important role too.
2.2 Automata
Automata models of the IS neglect the microscopic details of IS behavior. They
identify a set of characteristic states of the IS (like 'in rest' or 'with infection')
together with transition rules that define how the automaton may switch from
one state to another.
At discrete timesteps the state of the automaton is evaluated and the rules are
applied to define the automaton's state for the next timestep.
Atlan and Cohen [10] investigated the effects of suppressor T cells using a neural
network automaton.
2.3 Cellular Automata
Cellular automata (CA) were invented to investigate how simple building blocks
could locally cooperate to produce aggregates with interesting behavior. The
building blocks are automata living on a grid. Their rules define how the change
in state of a single automaton at the next time step depends on his own current
state and the states of his direct neighbor automata.
Clearly, the most fascinating aspect of CA modelling is the fact that even rather
simple transition rules together with strictly local interactions can lead to very
complex behavior of automata aggregates. An overview and classification of
different types of CAs can be found in [11].
In immunology we encounter a similar situation -- all IS activities are based on
the actions of cells reacting to their direct neighbor cells and molecules. There
is no central supervision of immune responses. Nevertheless the IS manages to
coordinate the actions of its constituents over larger spans of space and time.
Celada and Seiden [12] developed a CA model of the IS where the cells are simu-
lated by automata that may (as a modification of the usual CA concept of static
correspondence between information and position) carry their state information
with them while they are moving on a 2-dimensional grid. Depending on the
agents (cells, molecules) they encounter they may change their state e.g. from
naive to activated.
With their simulation program called IMMSIM Celada and Seiden were able
to investigate a number of IS phenomena, for example Affinity maturation and
hypermutation [. . . ] of the humoral immune response [13].
6
3 SIMMUNE
SIMMUNE is an attempt to derive IS behavior from immunological data that
describe the behavior of the cells of the IS on a microscopical level by defining
cellular stimulus response mechanisms.
A cellular stimulus response mechanism (cellular mechanism for short) consists
of a description of a set of stimuli a cell needs to experience before it performs
certain actions, and a description of those actions.
Metaphorically, a mechanism thus may be considered to be a set of conditional
actions: The cell checks whether certain conditions are fulfilled and if they are,
it performs certain actions.
In SIMMUNE, the condition part of a mechanism can consist of an arbitrary
number of conditions that can be combined through logical AND or AND NOT.
The action part may consist of one or more actions. Fig. 3 illustrates the concept
of a cellular mechanism.
condition 1
condition 2
...
condition n
cell action 1
cell action 2
...
cell action k
then
mechanism
Figure 3: cellular mechanism
Fig. 4 presents an example of a cellular mechanism. There, the stimuli
consist of the B:C complex and the D molecule on the cell's surface.
3.1 Components of SIMMUNE
3.1.1 Celltypes
Like in cellular automata models, in SIMMUNE too cells are individual entities
that interact with each other only locally. The cells live on a 3-dimensional
grid. Each cell in a SIMMUNE model belongs to a certain cell type that is
defined by the set of mechanisms according to which cells of a this type act.
SIMMUNE cell types need not be equivalent to cell types as they are identified
7
B
then
B
C
C
D
D
cell of
type X
cell of
type X
A
A
A
molecule complex
B:C
on any of the sides?
AND
then
secrete
molecules
of type A
molecule D
on same side?
mechanism of celltype X
Figure 4: an example of a cellular mechanism
in experimental immunology.2
It is important to notice that mechanisms in SIMMUNE do not define cell
states (that would be described by a fixed set of attributes, like certain types
of molecules on the cell's surface) but cell behavior, while the usual approach
in CA immune system models is to define a set of cell states and the rules how
cells may switch from one state to another (cellular actions simply being the
process of changing the cell's state).
Depending on the stimuli they receive from their environment, cells in SIM-
MUNE may change their attributes in various ways according to the mecha-
nisms of their type. They may express receptor molecules on their surface,
incorporate material from the extracellular milieu, secrete certain messenger
substances, kill neighbor cells or move into a certain direction. They may also
have part of their mechanisms modified (for example as a consequence of viral
attack) without completely changing their type.3
In SIMMUNE several conditions, each checking a simple attribute of the
cell's state may be combined in the condition part of a mechanism. A condition
for a given cell action may be fulfilled in many different ways -- all of which may
bear specific information for other mechanisms.
Fig. 5 illustrates how the mechanism based approach of SIMMUNE is able to
provide more flexibility in describing cell behavior than IS models that com-
2Interestingly enough the categorization of IS cells into rigidly distinct cell types has often
been a controversial subject in immunology. Currently this is the case for T-helper1 vs.
T-helper2 cells.
3 Here, the analogy between sets of cellular mechanisms in SIMMUNE and the genetic
encoding of cellular behavior in real biological systems is obvious.
8
pletely define cell states and transitions between them.
Besides its mechanisms, cell type properties in SIMMUNE include typical (mean)
lifetime and the size of the cells.
molecule D
on surface?
molecule
complex B:C
on surface?
secrete
molecules
of type A
CB
D
CB
D
CB
D
D
E
E
CB
D
CB
D
CB
D
A
A
A
A
A
A
D
E
E
Figure 5: different ways of satisfying common conditions
3.1.2 Molecules
In SIMMUNE cells are the basic units of signal- and information processing. The
signals they receive consist of molecules from the extracellular milieu that bind
to the receptor molecules which the cells display on their surfaces. (Molecules
also perform signal- and information processing simply by forming aggregates
according to their mutual binding possibilities.
In real ISs certain kinds of
molecular aggregates are responsible of attacking cellular pathogens as a first
line of defense before the IS has build up its complete arsenal of effector cells.
The death of a cellular pathogen as a reaction to contact with the above men-
tioned molecular aggregates would, however, be encoded as a cellular action in
SIMMUNE.) The significant properties of a molecule hence are described by the
set of its possibilities to bind to other molecules.
A part of a molecule that is visible to the milieu and hence can be used to bind
9
the molecule to receptors is called an epitope. Even though one usually only
refers to the binding sites of 'normal' molecules as epitopes, we treat receptor
binding sites just as 'normal' molecule binding sites and call them epitopes too.
The binding possibilities of any molecule are thus defined by the binding possi-
bilities of its epitopes.
Other molecule properties in SIMMUNE include the typical lifetime of the
molecule before it desintegrates into its fragment molecules and the types of
fragments that are produced upon desintegration of the molecule.
3.1.3 Compartments
The effects of many interactions in the IS are local by their nature. An example
is the cell-cell communication via contact receptors. But moving molecules or
cells may influence all those parts of the organism which they can access.
To provide the right specific milieus for the different tasks it has to fulfill, the
IS uses different compartments. For example, as mentioned above, the IS needs
to avoid producing too many cells with receptors with high affinity for material
of the own organism. It achieves this by establishing a central 'school' com-
partment, the thymus, where cells of a certain very important IS celltype, the
T-cells, are tested to be useful and to be not too self-reactive before they are
allowed to start their immune activities.
Another wellknown type of IS compartment are the lymph nodes. They are used
to bring together the different cell types of the IS for information exchange.
Every type of IS cell may be considered to see a different aspect of an antigen.
Hence, in order to exploit all the information available about the antigen, the
IS needs to provide the lymph nodes as meeting points for the different cells
carrying different pieces of information. From a simplified and abstract point
of view, compartments simply gather certain kinds of agents while excluding
others. A simulation with SIMMUNE may comprise different compartments.
3.2 Operation of SIMMUNE
Besides the properties of cell and molecule types, SIMMUNE lets its user define
the properties of the compartments within the simulated IS. Dimensions of the
compartments, diffusion rates of molecules and cells within the compartments
as well as initial concentrations of the different types of agents can be given.
Furthermore, the exchange of agents between the different compartments can
be regulated: which kinds of agents are allowed to pass from one compartment
to another and at which rate.
SIMMUNE offers a graphical user interface that can be used to watch agents'
concentrations and the spatial distribution of cells or manipulate the running
simulation by injecting new cells or molecules.
3.3 Example Applications
10
3.3.1 Local vs. Global Interaction between Agents
This example is meant to demonstrate with a very simple model differences
between a reaction-kinetics simulation and a simulation with locally interacting
agents. Five types of cells interact as illustrated in Fig. 6.
ID0 cells proliferate upon contact with cells of type OC. Upon contact with
certain cytokines (signal molecules) they transform into cells of type ID1 or
ID2, depending on the kind of cytokine they register. ID1 and ID2 themselves
produce the cytokines C1 and C2 that make their precursor ID0 switch into
their own state respectively. AID cells kill all ID type cells upon contact.
AID
OC
suppress
ID1
suppress
suppress
enhance
own prod.
differentiate
stimulate
proliferation
ID0
enhance
own prod.
ID2
differentiate
Figure 6: Network of cell interactions in the 'Local vs. Global Interaction'
simulation
Translation of this interaction network into a system of coupled differential
equations leads to a situation with an unstable equilibrium between the two cell
types ID1 and ID2. Because of the positive feedback between ID1/ID2 and ID0
small deviations from equal concentrations of both types may push the system
into nearly exclusive production of one type.
Simulation of this system with locally interacting agents yields a different be-
havior. The cytokines that are secreted by ID1 and ID2 do not instantly spread
all over the compartment. If the cytokines' lifetimes are short and diffusion is
not too strong, ID1 and ID2 act with their cytokines only on ID0 cells that are
located within a small neighborhood of themselves.
The parameters of the simulation are given in the appendix.
11
Fig. 10 in the appendix shows a cut through the compartment after the simu-
lation has reached a state of dynamic equilibrium. Neither ID1 nor ID2 have
achieved global dominance. Instead, the two competitors cluster in areas with
a diameter of typically 5-10 cells. Another interesting effect is that the AID
cells do not appear inside of these clusters. Although they perform free ran-
dom moves (controlled only by the availability of free space) they remain on the
clusters' surfaces. One is reminded of the situation of a water drop on a hot
stove: while the cells on the surface of the cluster get killed, those cells inside
the cluster manage to survive.
3.3.2 A simple Immune System
Here we demonstrate a simple IS model simulation. The simulation includes
five types of cells, three of which may go through differentiations during their
lifetimes. Furthermore one type of virus is part of the model.
Most of the cellular mechanisms in this model are strong simplifications of the
processes in real ISs. The reason why we present this simulation is that all of its
behavior results from direct cell-cell respectively cell-molecule interactions and
that it suffices as an example simulation that allows to describe how we analyze
the behavior of the simulations and to explain the notion of an IS context. We
will return to this in section 3.4 .
The first cell type OC is an organism cell not belonging to the IS. Cells of this
type divide at a certain (low) rate. This reproduction is controlled by contact
inhibition: OC cells have receptors on their membranes that are able to bind
to receptors of the same type on the membranes of neighbor cells. If such a
cell finds too many of these receptor complexes on its surface, it refrains from
dividing.
OC cells are the target of the virus V. Upon contact with V the OC type cells
get infected and start producing new viruses that are kept inside the cells. After
some time however the cells burst and release their virus content to the extra-
cellular milieu. Even though infected cells do not constitute a completely new
cell type in immunological terminology they will be called IC type cells here to
facilitate the discussion of the model and to identify them in the diagrams.
IC cells besides producing new viruses present virus epitopes on MHC1 receptor
molecules on their surface. They do not proliferate and have a shorter lifespan
than OC cells.
T-Cells appear in three states of activity. As naive T-cells NT and activated
T-killer (TK ) or T-helper (TH ) cells. NT cells possess two mechanisms. The
first makes them express receptor molecules TCR on their surfaces, the second
mechanism induces their transformation into TK or TH cells and proliferation
(reproduction). Using a rather simplified T-cell activation scheme, we define
the stimulus for this transformation to be the presence of a TCR being bound
to a receptor molecule of type MHC1 that besides its own epitopes presents an
additional antigen epitope.
12
A receptor molecule TCR possesses a binding site that uses two epitopes. Both
of these are selected at random. This means their binding possibilities (i.e.
which epitopes they can bind to) are arbitrarily chosen. Each NT cell once in
its lifetime selects its own (random-) epitopes. All TCR molecules produced by
this cell will possess these two epitopes. When the cell transforms into a TK or
TH cell it keeps this epitope choice; if such a cell divides, both daughtercells also
will use these two epitopes for their receptors. Thus all TK cells stemming from
a common NT cell will bind with their receptors to the same MHC1 /antigen
complexes. These cells are said to constitute a clone.
A TCR to be able to bind to a MHC must be able to bind to the MHC 's own
epitopes as well as to the antigen epitope which the MHC presents.
Naive T-cells (NT ) are stimulated to change their type and become (activated)
TK or TH cells upon registering a TCR:MHC complex. TK cells react to such
complexes by killing the cell that presents the MHC receptor. This is the way
the IS tries to remove infected cells (IC ) before they are able to release their
virus load to the milieu.
The TCRs of TH cells bind not to MHC1 but to similar membrane molecules
called MHC2 that are used by B-cells (see below). Similar to the contact in-
hibition of OC cells, TK and TH cells have a contact mediated mechanism
that controlls their proliferation. 4 They have receptors (called FAS ) and their
counterparts (FAS-ligand ) on their surface. An activated T-cell that finds a
FAS:FAS-ligand complex on its surface commits suicide.
Besides the T-cell response that needs direct cell-cell contact between IS effector
cells and infected cells, the IS uses cells that -- after having been activated --
secrete molecules which may bind to the virus particles to mark them for later
destruction. These cells are called B-cells and the marker molecules they pro-
duce are called antibodies (AB ). Similar to the T-cells, the B-cells exist in our
simulation as a pre-activation type NB (naive B-cell) and as activated B-cells
of type B.
The NB cells express on their surface B-cell receptors called BCR. The epitopes
of these receptors are randomly determined -- analogously to the random epitope
selection of NT cells. Some of the NB cells may possess BCR molecules on their
surface that are able to bind to the virus V while others may possess receptors
that bind to molecules that are used by the cells of the IS. The NB cells possess
mechanisms that make them present the epitopes of anything which got bound
by their BCR to the extracelular milieu with the help of membrane molecules
called MHC2. If a naive B-cell encounters a T-helper cell TH that has TCR
which bind to the MHC2 of the B-cell (including the additional epitopes the
MHC2 presents) the B-cell gets activated. Activated B-cells start secreting an-
tibodies AB that have the same random epitopes that were used by the BCRs
of the naive cell. These antibodies are hence capable of binding to the virus that
activated the T-helpers. The viruses that are marked for destruction by AB s
4Knowledge about the mechanisms that are used by real ISs to controll T-cell proliferation
seems to be still rather limited.
13
are removed by cells of type macrophage. These cells have receptors that fit to
a non-random binding site of antibodies. In this way they are able to destroy
the viruses that are bound to antibodies.
Fig. 7 shows the behavior of this simplified IS. The system starts with a cer-
tain concentration of OC that quickly enters a plateau concentration. Then the
virus V is injected at a high dose and infection spreads: the concentration of
infected cells IC grows. After a while, infected cells meet NT cells that possess
the appropriate receptors to react to the MHC presenting the virus-epitope.
T-helpers and -killers appear. T-cell proliferation stops at a certain concentra-
tion of T-cells as the encounters between T-cells and thus the FAS:FAS-ligand
induced T-cell death get frequent.
OC
TK
TH
IC
AB
V
100
10
1
0.1
OC
0.01
0.001
0.0001
n
o
i
t
a
r
t
n
e
c
n
o
c
t
n
e
g
a
AB
V
IC
TH
TK
1e-05
0
500
1000
timesteps
1500
2000
Figure 7: time dependency of the concentration of organism cells (OC), T-killer
cells (TK), T-helper cells (TH), infected cells (IC), antibodies (AB) and virus
(V) in the simple IS model.
B-cells, once being activated by T-helpers, start producing AB. As the T-
cell concentration grows, infected cells are effectively removed from the system.
High antibody concentration allows the removal of so many viruses that the rate
of infection of OC finally gets low enough to let the organism experience recon-
valescence - the OC concentration returns to its (contact-inhibition controlled)
plateau. As in reality, the antibodies in our simulation not only mark viruses
for destruction, they also may block the binding sites of the viruses and prevent
them from attaching themselves to their target cells IC.
14
3.3.3 B-Cell Activation
B-cells in real ISs may be activated in two different ways. One of them depends
on the assistance of T-helper cells, but B-cells are also capable of responding to
certain kinds of antigens directly, i.e. without the assistance of T-cells. These
antigens need to possess several indentical epitopes at the right distance from
each other as to allow a simultaneous binding of more than one B-cell antigen
receptor (BCR). Polysaccharides as they appear on the membranes of bacteria
have this property. The aggregation of several antigen receptors on the mem-
brane of a B-cell triggers a cascade of intracellular mechanisms of the B-cell that
leads to its activation. The cell will differentiate into an antibody secreting cell
and will proliferate.
This T-cell independent B-cell activation features an interesting dose-response
curve. Instead of inducing a stronger response, very large concentrations of an
antigen lead to an attenuated B-cell activation. The reason for this is that if
the antigen is present in abundance, the probability is rather high that each of
the receptors gets bound by an antigen molecule of its own. The probability
of B-cell receptor aggregation decreases. (For a mathematical discussion of this
effect as well as for a list of references to work on real ISs see [1].)
In our computer-experiment we inject a large dose of antigen (AG) concentrated
at one of the walls of the compartment. While the antigen diffuses we inves-
tigate the B-cell activation by looking at the concentration of the molecules A
that are produced by the activated B-cells in slices of the compartment parallel
to the gradient of the antigen concentration.
cell of
type SimpleB
A
G
BCR
BCR
then
A
A
A
cell of
type SimpleB
BCR
BCR
A
G
molecule complex
BCR:AG:BCR
on any of the sides?
then
secrete
molecules
of type A
mechanism of SimpleB
Figure 8: mechanism of simple B-cell: cross linking of two B-cell receptors makes
the B-cell secrete A.
In the simulation simple B-cells are created with the ability to distinguish
signals from six different sides. They possess one important mechanism which
15
is illustrated in Fig. 8.
A
f
o
n
o
i
t
a
r
t
n
e
c
n
o
c
G
A
f
o
n
o
i
t
a
r
t
n
e
c
n
o
c
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1000
100
10
1
0.1
0.01
0.001
0.0001
0
A
2
4
6
8
10
12
14
distance from AG injection side
AG
2
4
6
8
10
12
14
distance from AG injection side
Figure 9: Spatial variation of the concentration of molecules A produced by
activated B-cells and of the concentration of activating antigen AG.
The binding of the AG-epitopes by the receptors is reversible (with equal
reaction constants for binding and release of AG) and the A molecules desinte-
grate after a certain mean lifetime. The AG molecules are stable. The resulting
concentration of A thus indicates how successfully receptor-antigen-receptor ag-
gregates are produced. Fig. 9 shows the number of AG and A molecules per cell
as a function of the distance from the wall of the compartment where the high
antigen concentration was initially injected. The maximum of the concentration
of A is located around the area where 10 AG molecules per cell can be found.
This corresponds -- as expected -- to just below 2 AG molecules per side of the
cell.
16
3.4 Simulation Analysis
The main purpose of SIMMUNE is to provide a tool to investigate how context
adaptive behavior of the IS might emerge from local cell-cell and cell-molecule
interactions.
The most obvious kind of context dependent behavior of the IS can be found
in the ability of single cells to react differently to a stimulus when it comes in
combination with other stimuli as opposed to an isolated event. Thus, mecha-
nisms with several conditons already encode context dependent behavior.
More difficult is the analysis of context recognition that emerges from coopera-
tions between several cells. Cooperations between cells manifest themselves as
correlations between the actions of these cells. A very primitive example from
the simulation of section 3.3.2 would be the correlation between the 'kill' -action
of T-killer cells and the 'die' -action of infected cells.
Whether a cell-cell interaction is of just local importance or part of a network
of interdependent cell actions cannot be decided on the microscopical level of
single cells. It requires to analyze on multiple scales the correlations between
cell actions.
Analyzing the behavior of a system on multiple scales means identifying on each
scale collections of objects that may be regarded as single composite objects on
a coarser scale. This identification of composite objects is based on the recog-
nition of common attributes or coherent behavior of the component objects. In
order to allow an analysis that involves a series of scales that emerge from each
other, the rules for the composition of objects must be applicable not only to
the elementary objects of the original, finest scale but also to the composite
objects of coarser scales.
In our case, the elementary objects are cell actions. Collections of them are
identified to constitute composite objects if they are spatially and temporally
correlated.
Iterating the process by applying the correlation analysis to the
objects of the first coarse scale may yield a further scale that is coarser than
the first in two aspects: The components of its objects are themselves compos-
ite objects and the correlations between the constituents of its objects involve
larger spatial and temporal distances.
Starting from single-cell behavior we enter the scale of cell-cell cooperations by
looking for correlations between cell actions. Looking for correlations between
cell cooperations takes us to the next scale. Repeated application of this process
might lead to a scale that directly describes the macroscopical behavior of the
IS.
The network of interactions of our example simulation from section 3.3.2 leads
to two (yet simple) coarse scales: On the first we register the correlations be-
tween the actions of IS effector cells like T-killer cells and non-IS cells like
infected cells. The second coarse scale contains for example the correlation be-
tween T-cell activation and the appearance of activated B-cells.
Having explained how we proceed in order to look for context dependent behav-
ior of the IS we are now able to give a definition of what we call a 'recognized
immunological context': An immunological context that is recognized by the
17
IS5 leads to a response of the IS that can be traced on multiple scales as de-
scribed above. It can be distinguished from other contexts that are descibed by
different patterns of correlation.
Recently, the principle of identifying biological contexts on different scales
by iterated clustering of data has successfully been applied to the infrared ana-
lysis of human blood serum by Werner et al. [16].
Usually, models about higher level cooperative effects within the IS (like
automata models that describe the behavior of the IS in terms of predefined
states) must be based on assumptions that try to bridge the scale gap between
the scale of direct cellular interactions and overall behavior of the IS.
We do not need to make such assumptions. Cooperative phenomena are detected
using methods of statistical analysis.
Of course we pay a price for this: Since SIMMUNE is entirely based on direct
cell-cell or cell-molecule interactions we need to make assumptions about cellular
mechanisms whenever we make use of aspects of single-cell behavior that are
not yet understood in detail.
3.5 Limitations of SIMMUNE
Due to the faithful modelling of cell behavior, computer power requirements of
SIMMUNE are quite high. With currently available single processor computer
power and memory the maximum number of cells (per computer processor) that
can participate in a simulation is practically not larger than 500.000.
Recently Smith et al. [14] presented a technique called lazy evaluation to sim-
ulate quite realistic clone numbers and sizes. Lazy evaluation makes use of the
fact that out of the large number of clones only a small fraction is actually able
to interact with a given antigen and the presenting cells of the IS. Only this frac-
tion will be activated to proliferate and play an important role in the immune
response. By simulating only this fraction while limiting its overall concentra-
tion to the value it would have with the full cell repertoire being existent, major
memory and cpu-time savings can be achieved. To accelerate the simulation of
those models that do not need to consider idiotypic networks (see [17] or e.g.
[1]) future versions of SIMMUNE may adopt this technique.
Another limitation of SIMMUNE will not be circumvented by increasing com-
puter power or techniques as lazy evaluation: As mentioned in the foregoing
section, simulations with SIMMUNE sometimes have to involve mechanisms
and parameters that have not yet been established by experiment. This often
makes quantitative predictions difficult. On the other hand, the need for de-
tailed descriptions of cellular mechanisms may point out gaps in immunological
knowledge that have not yet been sufficiently investigated.
5Context dependent behavior of single cells as we mentioned above would be encoded in
cellular mechanisms. We want to distinguish between such context recognition and the context
recognition that emerges from multi-cellular cooperation.
18
3.6 Beyond Immunology
Since all celltypes are user-defined and the basic cellular actions of SIMMUNE
are typically as general as 'expression of (user-defined) membrane receptors' or
'movement along gradients of concentrations of molecules', the software may be
used to simulate populations of cells of any kind. It has already been used to
simulate systems of neurons with some of the molecular structure of synaptic
gaps between them.
4 Conclusion
Besides providing its user with the possibility to define in detail the behavior
and properties of cells and molecules in simulations of the IS, the mechanism
based approach of SIMMUNE yields the fundament for a new kind of simulation
analysis in immunology. Multiscale correlation analysis of cellular actions may
allow the automated classification of immunological contexts that have not yet
been considered until now.
Institutes interested in working with SIMMUNE may contact us to receive a
copy of the program together with a manual.
Acknowledgements
Irun Cohen for helpful discussions and for
The authors wish to thank Prof.
providing copies of his forthcoming book.
Dr. Ute Kerres (Univ. of Hamburg) has contributed to this work during many
valuable discussions. We thank Dr. Andrew Yates (Institute of Child Health,
London) for reading the manuscript.
Support by Deutsche Forschungsgemeinschaft (GRK 135/3-98) and GIF travel
grants are gratefully acknowledged.
19
A Simulation Parameters and Diagrams
Parameters for reaction-kinetics simulations from 2.1:
figure
Fig. 1
Fig. 2
pinf ect
0.3
0.3
pkill
0.5
1.0
presp
0.1
0.8
s
0.01
0.01
dI
0.01
0.01
dK
0.01
0.01
dC
0.01
0.01
I0 K0 C0
1.0
0.1
0.1
1.0
0.1
0.1
Parameters for the simulation from 3.3.1:
OC ID0
initial concentration
0.01
0.05
av. lifespan (timesteps) ∞
∞
ID1
0.005
∞
∞
ID2 AID C1
0
0.005
0.01
∞ 100
C2
0
100
molecular diffusion rate: 0.0001
cellular diffusion rate: 0.01
(The diffusion rate defines the per timestep probability for agents to get trans-
ported to a neighbor grid point. In this simulation cellular diffusion replaces
active cell movement. The product of molecular diffusion rate and molecular
lifespan defines the mean range of a molecular signal emitted by a cell.)
compartment dimensions: 80 x 80 x 10
20
References
[1] A.S. Perelson and G. Weissbuch, Immunology for Physicists, Rev. Mod.
Phys. 69, No. 4, 1997.
[2] I.R. Cohen, The Cognitive Principle Challenges Clonal Selection, Immunol-
ogy Today 13, No.11, 441-444, 1992.
[3] I.R. Cohen, Tending Adams Garden: Evolving the Cognitive Immune Self,
Academic Press, to appear in 1999.
[4] Z. Grossman, Contextual Discrimination of Antigens by the Immune Sys-
tem: Towards a Unifying Hypothesis, in Theoretical and Experimental In-
sights into Immunology, edited by A.S. Perelson and G. Weisbuch, Springer
Berlin, 1992
[5] L. Segel and Ruth Lev Bar-Or, Immunology Viewed as the Study of an
Autonomous Decentralized System, in Artificial Immune Systems and Their
Applications, edited by D. Dasgupta, Springer Berlin, 1998.
[6] L. Segel, How Can Perception of Context Improve the Immune Response,
in L. Steinman (Ed.) Festschrift for Yaron Cohen, in preparation.
[7] H. Atlan, I.R. Cohen, Immune Information, Self-organization and Meaning,
Int. Immunol. 10, No. 6, 711-717, 1998.
[8] C.A. Janeway and P. Travers, Immunobiology, 1994 Current Biology
Ltd./Garland Publishing Inc..
[9] M. Jenkins, Tracking the Interactions Between Antigen-Specific T and B
Cells in situ, Talk given at the IMA Workshop 'IS Modeling and Cell Sig-
nalling', October 12-16, 1998.
[10] H. Atlan and I.R. Cohen, Paradoxical Effects of Suppressor T Cells in
Adjuvant Arthritis: A Neural Network Analysis, in Theoretical and Experi-
mental Insights into Immunology, edited by A.S. Perelson and G. Weisbuch,
Springer Berlin, 1992.
[11] S. Wolfram (Ed.), Theory and Applications of Cellular Automata, World
Scientific (Singapore), 1986.
[12] F. Celada, P.E. Seiden, A computer model of cellular interactions in the
immune system, Immunol. Today 13, 56-62, 1992.
[13] F. Celada, P.E. Seiden, Affinity maturation and hypermutation in a simu-
lation of the humoral immune response, Eur. J. Immunol. 26, 1350-1358,
1996.
[14] D.J. Smith et al., Using Lazy Evaluation to Simulate Realistic-Size Reper-
toires in Models of the Immune System, Bull. Math. Biol., 60, 647-658,
1998.
21
[15] M. Meier-Schellersheim, Immunological Contexts -- a Multiscale Approach,
in preparation.
[16] G. Werner et al., in H. Mansch, M. Jackson (Eds.), Infrared Spectroscopy:A
New Tool in Medicine, Proc. of SPIE, 3257, 1998.
[17] N.K. Jerne, Towards a network theory of the immune system, Ann. Im-
munol. (Inst. Pasteur), 125C, 373-389, 1974.
22
o o *
o + #
o + o + + + o + +
o o
+
+ + + + * + o
*
+ + *
+ *
# + o
* * * + + *
+ *
* + + # +
o
*
* * *
# *
# o o
* *
* * # # * o o
# # o o
* * * *
# # # *
# # o *
* * * *
*
# * * * o * o o o
* # * * *
o o * * * o o o
# *
o o o # o o
+ + *
o o o o
+
#
o o
+ + # o
o
o o o o
# o o o
+ +
+
* + + o *
o
* + *
+ + + +
* * * * * * * * *
* * * + + # * *
+ * * + + # * * + * * * *
* *
*
*
o # * * *
o o o
o #
o
o *
o
* * * *
# o o o o + *
o # + o + o * *
* o o o o o o * #
o o o o o * o * * *
o o # o #
o #
# o * * #
o o * * # o * *
o o *
o o + * o *
+ o + o o
o + + + + # o
o o + + + o
o o + + + + + +
o o o
* # + o o o o
* * o o o
* * o o o o o o o o
* o *
o + * o # # + o # * * * * * *
+ +
*
o # # + + + + + + + + +
*
* * *
o o o * * * *
o * * * * *
* * * * * * * * * * + + + * o + + o + + + + + +
* + + * * * o o o o o + + + + + +
*
o o *
* * * * * * * + + *
+ + + +
*
+ + * +
o o + + + + + + +
+ o o * *
* * * *
# + + + +
+ + + + +
* # o * * o
# * *
+ + +
* * o + +
o
o + + + o
# o o # o o o o o o
# + o o + +
+ + # +
+ + + o
# + o + + + o o o * *
o o o o
o o +
+ + + + + + + + o o + *
o # o o o o o
+ + + +
o o
+ + + + + # #
o + + + o
+ + +
+ + + + + # # # # + + + o o
+ + + + o o o o
o o o o o o o o
+ + + + + + o o
+
o o o #
*
+ + + * o +
* * *
* * *
o * + + +
o o + + * + + + + + +
* o o o
* o o # o
* * o
o o
# o o o o o o
o
o + + + o
o
+ + + + + + +
+ + + * +
+ + + + +
o
+ + +
* * * * *
* *
+
o
#
#
#
#
*
#
+
o o
o o
o o o o
o + o +
o
o
o
o +
+ + + + + + + +
o o o
o *
o o * o
o o
# o o
o o
o
o o + # o
*
o o o o o * o
+ + + + + +
o *
* * * *
+
* * *
o o
+ + + + + + + + +
o * * * * * *
+
* * *
+ * * * * * * *
* * *
o o o + o + + + * * # # *
o
o * o o #
o o o
o o + o o o o o
o o o
o o o o o
o o o o
o o o o + +
o * o o o o + o o
o *
+ + + + + * * + +
o o * * *
+ + * * *
* * + + + + +
+ + + + + # + + o + + + +
+ + + + + + # + + + + + o o o o + + + * * *
+ + + + + + + + +
+ + + + + + +
* + o
o + + o o o o o
*
+ + + + # # o o # o o o
*
* *
# + + + # *
o o o
*
o o *
o o o o * *
o o o o * *
o o o
o o o #
o o o o
* o
o o o o
o o o
# o o o o o + o o o
o o o o o o + o +
* * * * o o o o o o o + + o
* * * * * * # o
* * * * # o o o
* + * + + o #
o * * * * + + + +
+ o o + *
# # * *
# * + + + + + * * # + o o o o * o
+ * + + + +
+ + + + + + + + # + o * * * * * #
o o o o #
o
o o o
o o o o
o o o o o o
+ + + + o o o o # o o o # o
o o o o
# # o o o o * * # o * o
o #
+ + + o o +
+ + + +
* * * + + + + + + o + + +
+ o o o o o
+ + + + o
o *
o o o
+ o o o
+ + + + #
* + + +
o o o
o o o
o
o
o o
* * *
* # *
*
o o * * * *
o * o o * * *
# +
#
#
* * + # * *
*
# # # # * * o
# * * * * * * * *
* # * * *
o * * * * * * *
o * * * * *
# # o * * *
o # # # # *
o o o
o o o o o o # o o o
o o o + o o o # #
o o o o
o o o o o o
o o o o
o o
#
+
+ + +
o
o o o
o + + + + o o o o
o o + + o
o o + + o o
o o + + + o
* * * * o o o o
o
*
o o o
#
#
o
o * * o
*
*
o
o
+
+
+
+
+ *
* * *
* o o
o o
+ #
*
* o
* o
o o + + +
* * * *
* * *
#
o
o
o + *
o + * *
+
* *
+ + + # # #
* * *
+ + + + + + # #
o o o + * + o + + + + + + + + +
* * * * * * * * o o o o o + o * + o + o + + + + + o
+ o + + o
# o + o o
o o o
# * * # #
* * * o *
*
* * *
o
* * * o * *
o #
* * *
o # o
* * # o + +
o
o + + * * * *
# # # # o *
# # # *
*
* * * *
* * + * # # # # o
* * * *
* * * * * *
*
o o * * * * o * + +
* *
+ + + + o * *
* + * + + + +
# * * * + # *
+ o
o
#
+ o + o
+ + o o
# o o o o o o o
#
o * * *
*
+ * * o +
o o o o o o
o + o
o o + + + + + + #
# # + + + + + + + + + + *
*
* * * + + + + + + + + + * + *
#
* * +
# + + + + + + + + + + +
o o o + o + + + + + + + + + * *
+ + + + + + + + * + * *
* + * *
+ + + + o # # * * + + + + + * + + + + + *
#
+ + *
+ + * * * *
* +
# * * * * * * *
* * * * *
+ + + + + o # * * * * * o * *
+ * * o o # + * + * *
*
o * * * * * + o o o o o # # # * * * * * * * *
* * * * * * * *
o + + + + + * *
o o o + + + + + + *
#
* * * * * #
* * *
# # + + + + + * + * *
* + # + + + + # # # *
*
* *
* * * *
o + o + + + +
o # #
+ + + + + +
* * * # # *
# +
+ + + + + + # *
+
+ + + + # o
o
o + + + + #
o o o + o o
o o o
o
#
*
o
* *
* * * * * *
o o o o + # *
o *
o o o o * # * * * * * *
o o o
+ * *
o o o
o o o o o o
* *
*
* * * *
o o o o *
+ #
* + *
o o o
o o o
# + + #
* * #
o o
#
#
#
o
o
o o * o o #
# o * * + + + + + + o o + + +
o o * *
#
o
o
+
o o * *
+ + + + o o
o + o
* *
# o o * * * * * *
+ # + o * * * * * +
+ # * o * * * *
* * *
o *
o o #
o # + + # +
+ + + * + * *
# # # + + + * + * + + + + # + # #
+ + * * + * * + + # + +
+ + + * * * * + + + # + + #
o * * + * + *
# + + # # # +
* * * *
# # + + + + + + o
* #
o #
+
#
o # o o o
# o o
o o o o + o * o o
* * o
* * *
* + # # + + # # #
# # # o + + +
# # + + + +
o o o * o o
o o o o o
+ + o o + + + +
o #
o #
# #
#
+
#
o
+ o o + + o # #
+
o o o o
o o o
o
o
+ +
#
#
* o *
+ + * + * * * *
# o o o # # # o
o o
+ * * * * * * * o o
+ o * *
+ +
+ + +
* * * * + +
+ + + + +
* * *
* *
+ * * o
*
o *
* o o o
* *
* * * * # # + *
* * * *
* * # # *
# # * * *
# * * * * * * * * * # * *
+ + *
+ + * * # # # * * * * * *
* *
# + * + # o # # o + + + * +
* *
* * # * # * * * * * * * * * *
* * + + * * * *
o
o o o
+ + + + o o o o
+ o o + + + + + + o o
+ o * * * * * o + o o o o
+ o o o
o
o o o * * + *
o o * o # o
o o o
o o
o
o
# * * * * * * * *
o o o
o * o o + + * + + o # o o o *
* * * + * + + # * o *
* + +
* o o o o
o o * # # # + + + + + + + o o o * # o o #
* * *
o o o o + + + + # + + o o *
o o o * * *
o o o o + + + + # + +
o o o
*
# #
+ + + # +
+
+ * *
* #
o * *
* * * *
*
# * * * *
o
o
#
o + # * * *
o # o # # + + + o + * * o
#
o o + + + + o # * * * * o
*
* # * + + + * * * *
#
* *
* * * * * * + + +
* * * * * * * * + + +
# * * * + o o + + + # # o * *
* * * * o o * * * + + + + + + +
o
o o + + +
o o +
o o o o #
# o
* * # # + #
+ + + o * *
+ + + + o + + o * o o * * # # +
+ o + o o + # # o * * * * * + + o
+
o o o o + +
o o o o o + + + o
+ * o *
# # o # # # o o
+ o o
o o
o
#
+ o
o +
* + o
* *
* + + # # o * *
* o # # # * * *
* + + + o
* * o o
# * + + + +
o o o
#
+ # + * o o
o * * * *
*
*
* o
* *
o
o
*
*
o o *
+ *
+ * +
# + + + * + + + +
+ # + + + + + + * +
o + + + + + + + + +
o
o o o
+ + + + + + +
o o o o o o o + +
+ +
o o
+ + + + + + + + +
o * o o o o + + # # o # + + + + + + + +
*
o o * *
o + + o * o o # + + + + +
* o # + + + + + +
o + * * * * *
o o o o * *
o o o o o + * * * *
* o o o
o o o # + # #
o o o # # + + + * *
o o
o
+ + + + + o * * * o o o # #
+ +
+ +
# #
o o
+ o
o + + + + + + * o o o + +
o o
o + + + # + * * *
+ + + + *
o
# #
# # # # # #
+ # +
# * * o o o + + *
o * * * o # + + + * * *
*
o o o o o + * + + + # + *
o o o o o o *
* o o + + # # + * *
o o o o o o o + * + * * + + * *
* * * *
# * * * * *
+ + + o o o o o * * * *
o o o o o o o
#
# #
# #
o # #
* #
# * o #
#
o o o o o o o o o o *
o o o o o o o o
o o o o o
,
∗
1
D
I
,
o
0
D
I
:
s
l
o
b
m
y
s
;
1
.
3
.
3
m
o
r
f
n
o
i
t
a
l
u
m
i
s
k
c
a
b
d
e
e
f
n
i
s
r
e
t
s
u
l
c
:
0
1
e
r
u
g
i
F
#
D
I
A
+
2
D
I
,
3
2
|
1909.09397 | 1 | 1909 | 2019-09-20T09:54:30 | Simulating Crowds in Real Time with Agent-Based Modelling and a Particle Filter | [
"cs.MA"
] | Agent-based modelling is a valuable approach for systems whose behaviour is driven by the interactions between distinct entities. They have shown particular promise as a means of modelling crowds of people in streets, public transport terminals, stadiums, etc. However, the methodology faces a fundamental difficulty: there are no established mechanisms for dynamically incorporating real-time data into models. This limits simulations that are inherently dynamic, such as pedestrian movements, to scenario testing of, for example, the potential impacts of new architectural configurations on movements. This paper begins to address this fundamental gap by demonstrating how a particle filter could be used to incorporate real data into an agent-based model of pedestrian movements at run time. The experiments show that it is indeed possible to use a particle filter to perform online (real time) model optimisation. However, as the number of agents increases, the number of individual particles (and hence the computational complexity) required increases exponentially. By laying the groundwork for the real-time simulation of crowd movements, this paper has implications for the management of complex environments (both nationally and internationally) such as transportation hubs, hospitals, shopping centres, etc. | cs.MA | cs |
SIMULATING CROWDS IN REAL TIME WITH AGENT-BASED
MODELLING AND A PARTICLE FILTER ∗
A PREPRINT
Nick Malleson†
School of Geography
University of Leeds
Leeds, LS2 9JT, UK
[email protected]
Jonathan A. Ward
School of Mathematics
University of Leeds
Leeds, LS2 9JT, UK
Leeds Institute for Data Analytics
Leeds Institute for Data Analytics
Le-Minh Kieu
University of Leeds
Leeds, LS2 9JT, UK
Kevin Minors
University of Leeds
Leeds, LS2 9JT, UK
Alison Heppenstall
Alan Turing Institute
British Library
London NW1 2DB
Andrew A. West
School of Geography
University of Leeds
Leeds, LS2 9JT, UK
September 23, 2019
ABSTRACT
Agent-based modelling is a valuable approach for systems whose behaviour is driven by the in-
teractions between distinct entities. They have shown particular promise as a means of modelling
crowds of people in streets, public transport terminals, stadiums, etc. However, the methodology
faces a fundamental difficulty: there are no established mechanisms for dynamically incorporating
real-time data into models. This limits simulations that are inherently dynamic, such as pedestrian
movements, to scenario testing of, for example, the potential impacts of new architectural configu-
rations on movements. This paper begins to address this fundamental gap by demonstrating how a
particle filter could be used to incorporate real data into an agent-based model of pedestrian move-
ments at run time. The experiments show that it is indeed possible to use a particle filter to perform
online (real time) model optimisation. However, as the number of agents increases, the number
of individual particles (and hence the computational complexity) required increases exponentially.
By laying the groundwork for the real-time simulation of crowd movements, this paper has impli-
cations for the management of complex environments (both nationally and internationally) such as
transportation hubs, hospitals, shopping centres, etc.
Keywords Agent-based modelling · Particle Filter · Data assimilation · Crowd simulation
1
Introduction
Agent-based modelling is a form of computer simulation that is well suited to modelling human systems [1, 2]. In
recent years it has emerged as an important tool for decision makers who need to base their decisions on the behaviour
of crowds of people [3]. Such models, that simulate the behaviour of synthetic individual people ('agents'), have
been proven to be useful as tools to experiment with strategies for humanitarian assistance [4], emergency evacuations
[5, 6], religious festivals [7], crowd stampedes [8] etc. Although many agent-based crowd simulations have been
developed, there is a fundamental methodological difficulty that modellers face: there are no established mechanisms
∗This project has received funding from the European Research Council (ERC) under the European Union's Horizon 2020
research and innovation programme (grant agreement No. 757455), through a UK Economic and Social Research Council (ESRC)
Future Research Leaders grant [number ES/L009900/1], and through an internship funded by the UK Leeds Institute for Data
Analytics (LIDA).
†Corresponding author
A PREPRINT - SEPTEMBER 23, 2019
for incorporating real-time data into simulations [9 -- 11]. Models are typically calibrated once, using historical data,
and then projected forward in time to make a prediction independently of any new data that might arise. Although this
makes them effective at analysing scenarios to create information that can be useful in the design of crowd management
policies, it means that they cannot currently be used to simulate real crowd systems in real time. Without knowledge of
the current state of a system it is difficult to decide on the most appropriate management plan for emerging situations.
Fortunately, methods do exist to reliably incorporate emerging data into models. Data assimilation (DA) is a technique
that has been widely used in fields such as meteorology, hydrology and oceanography, and is one of the main reasons
that weather forecasts have improved so substantially in recent decades [12]. Broadly, DA refers to a suite of techniques
that allow observational data from the real world to be incorporated into models [13]. This makes it possible to more
accurately represent the current state of the system, and therefore reduce the uncertainty in future predictions.
It is important to note the differences between the data assimilation approach used here and that of typical agent-
based parameter estimation / calibration. The field of optimisation -- finding suitable estimates for the parameters of
algorithms -- has (and continues to be) an extremely well-researched field that agent-based modellers often draw on.
For example, agent-based models regularly make use of sampling methods, such as Latin Hypercube sampling [14]
or evolutionary / heuristic optimisation algorithms such as simulated annealing [15], genetic algorithms, [16], and
approximate Bayesian computation [17]. There are also new software tools becoming available to support advanced
parameter exploration [18]. It is, however, worth noting that in most cases agent-based models are not calibrated
to quantitative data [14].
In the cases where parameter estimation does take place, it is typically performed as a
single calibration step in a waterfall-style development process -- e.g. design, implementation, calibration, validation.
Although there are some more recent studies that do attempt to re-calibrate model parameters dynamically during
runtime [e.g. 19, 20] there is another, more fundamental, difference to typical parameter optimisation (be it static or
dynamic) and the data assimilation approach.
Even if optimal model parameters have been found, there will usually be a degree of uncertainty in the model. With
crowd simulations, for example, it is impossible to know exactly how individuals will behave in a given situation -- will
someone turn left or right given two competing options? -- nor can individual parameters such as walking speed ever
be known exactly. Data assimilation algorithms use 'state estimation' to calculate the difference between the model
and the 'true' state of the underlying system at runtime. They are then able to adjust the current model state in order
to constrain a model's continued evolution against the real world [11]. Although it is possible to re-calibrate models
dynamically during runtime (e.g. [19]), this would not reduce the natural uncertainty that arises as stochastic models
evolve.
This paper is part of a wider programme of work3 whose main aim is to develop data assimilation methods that
can be used in agent-based modelling. The software codes that underpin the work discussed here are available in
full from the project code repository; see Appendix A. The work here focuses on one particular system -- that of
pedestrian movements -- and one particular method -- the particle filter. A particle filter is a brute force Bayesian state
estimation method whose goal is to estimate the 'true' state of a system, obtained by combining a model estimate with
observational data, using an ensemble of model instances called particles. When observational data become available,
the algorithm identifies those model instances (particles) whose state is closest to that of the observational data, and
then re-samples particles based on this distance. It is worth noting that once an accurate estimate of the current state
of the system has been calculated, predictions of future states should be much more reliable -- c.f.
the substantial
improvements in weather forecasting that have come about as a result of modern data assimilation methods [12].
Predicting future system states is beyond the scope of this paper however.
The overall aim of the paper is to:
quantify the conditions under which a typical particle filter is able to reliably estimate the 'true' state
of an underlying pedestrian system through the combination of a modelled state estimate, produced
using an agent-based model, and observational data.
This will be achieved through a number of experiments following an 'identical twin' approach [10]. The agent-based
model is first executed to produce hypothetical real data -- also known as pseudo-truth [17] data -- and these data are
assumed to be drawn from the real world. During data assimilation, observations are derived from the pseudo-truth
data. This approach has the advantage that the 'true' system state can be known precisely, and so the accuracy of the
particle filter can be calculated. In reality, the true system state can never be known.
The agent-based model under study is designed to represent a very simple pedestrian system. It has been kept inten-
tionally simple because the aim here is to experiment with the particle filter, not to accurately simulate a pedestrian
system. Were the model more complicated it would become more difficult to understand the internal uncertainties,
3http://dust.leeds.ac.uk/
2
A PREPRINT - SEPTEMBER 23, 2019
which would in turn make it more difficult to understand how well the particle filter was able to handle these uncer-
tainties. The model is sufficiently complex to allow the emergence of crowding, so its dynamics could not be easily
replicated by a simpler mathematical model -- as per [9] and [11] -- but is otherwise as simple as possible. Crowding
occurs because the agents have a variable maximum speed, therefore slower agents hold up faster ones who are behind
them. The only uncertainty in the model, which the particle filter is tasked with managing, occurs when a faster agent
must make a random choice whether to move round a slower agent to the left or right. Without that uncertain behaviour
the model would be deterministic. A more realistic crowd simulation [e.g. 8] would exhibit much more complicated
behavioural dynamics.
The paper is outlined as follows: Section 2 reviews the relevant literature; Section 3 outlines the methods, including
a description of the agent based model and particle filter; Section 4 outlines the experiments that are conducted and
their results; and Section 5 draws conclusions and outlines opportunities for future work.
2 Background
2.1 Agent-Based Crowd Modelling
One of the core strengths of agent-based modelling is its ability to replicate hetereogeneous individuals and their
behaviours. These individuals can be readily placed within an accurate representation of their physical environment,
thus creating a potentially powerful tool for managing complex environments such as urban spaces. Understanding
how pedestrians are likely to move around different urban environments is an important policy area. Recent work
using agent-based models have shown value in simulating applications from the design of railway stations [22, 23],
the organisations of queues [24], the dangers of crowding [8] and the management of emergency evacuations [25].
However, a drawback with all agent-based crowd simulations -- aside from a handful of exceptions discussed below
-- is that they are essentially models of the past. Historical data are used to estimate suitable model parameters and
models are evolved forward in time regardless of any new data that might arise. Whilst this approach has value for
exploring the dynamics of crowd behaviour, or for general policy or infrastructure planning, it means that models
cannot be used to simulate crowd dynamics in real time. The drivers of these systems are complex, hence the models
are necessarily probabilistic. This means that a collection of models (or an ensemble) will rapidly diverge from each
other and from the real world even under identical starting conditions. This issue has fostered an emerging interest in
the means of better associating models with empirical data from the target system [see, e.g., 10, 11, 26].
2.2 Data-Driven Agent-Based Modelling
The concept of Data-Driven Agent-Based Modelling (DDABM) emerged from broader work in data-driven application
systems [27] that aims to enhance or refine a model at runtime using new data. One of the most well developed
models in this domain is the 'WIPER' system [6]. This uses streaming data from mobile telephone towers to detect
crisis events and model pedestrian behaviour. When an event is detected, an ensemble of agent-based models are
instantiated from streaming data and then validated in order to estimate which ensemble model most closely captured
the particular crisis scenario [28]. Although innovative in its use of streaming data, the approach is otherwise consistent
with traditional methods for model validation based on historical data [19]. Similar attempts have been made to model
solar panel adoption [29], rail travel [30], crime [9], bird flocking [20] and aggregate pedestrian behaviour [11], but
whilst promising, these models contain their own limitations. For example, they either assume that agent behaviours
can be proxied by simple regression models [29] (which will make it impossible to use more advanced behavioural
frameworks to encapsulate the more interesting features of agent behaviour), are calibrated manually [30] (which is
infeasible in most cases), optimse model parameters dynamically but not the underlying model state (which might
have diverged substantially from reality) [20], or are simple enough to be approximated by an aggregate mathematical
model [9, 11], neglecting the importance of using agent-based modelling in the first place.
There are two studies of direct relevance. The first is that of [31], who developed an agent-based model of an evac-
uation coupled with a particle filter to conduct real-time data assimilation. This paper presents a novel mathematical
approach that can map observations from a simple to a complex domain. In the authors' example, simulated data from
hypothetical population counters are mapped to the complex agent-based model, which represents the heterogeneous
locations of the individual agents. Although it is beyond the scope of the study here, the proposed mapping method
will be useful for future work that experiments with variations in the complexity of the observations that are drawn
from the real world (e.g. examining how well the particle filter can perform when presented with aggregate population
counts rather than individual agent locations).
The second study is that of [10] who investigated the viability of simulating individual movements in an indoor
environment using streams of real-time sensor data to perform dynamic state estimation. As with this paper, the
3
A PREPRINT - SEPTEMBER 23, 2019
authors used an 'identical twin' experimental framework; the agent-based model is used to create pseudo-realistic
data rather than using those from real-world sensors. Next, an ensemble of models were developed to represent the
target system with a particle filter used to constrain the models to the hypothetical reality. A new particle resampling
method ('component set resampling') was also proposed that is shown to mitigate the particle deprivation problem
(see Section 2.3 for more details). The research presented within this paper builds on [10] by: (i) attempting to apply
data assimilation to a system that exhibits emergence; and (ii) performing more rigorous experiments to assess the
conditions under which a particle filter is appropriate for assimilating data into agent-based crowd models.
2.3 Data Assimilation and the Particle Filter
'Data assimilation' refers to a suite of mathematical approaches that are capable of using up-to-date data to adjust the
state of a running model, allowing it to more accurately represent the current state of the target system. They have
been successfully used in fields such as meteorology, where in recent years 7-day weather forecasts have become more
accurate than 5-day forecasts were in the 1990s [32]; partly due to improvements in data assimilation techniques [12].
The need for data assimilation was initially born out of data scarcity; numerical weather prediction models typically
have two orders of magnitude more degrees of freedom than they do observation data.
Initially the problem was
addressed using interpolation [e.g. 33, 34], but this soon proved insufficient [12]. The eventual solution was to use a
combination of observational data and the predictions of short-range forecasts (i.e. a model) to create the full initial
conditions (the 'first guess') for a model of the current system state that could then make forecasts. In effect, the basic
premise is that by combining a detailed but uncertain model of the system with sparse but less uncertain data, "all the
available information" can be used to estimate the true state of the target system [35].
A particle filter is only one of many different methods that have been developed to perform data assimilation. Others
include the Successive Corrections Method, Optimal Interpolation, 3D-Var, 4D-Var, and various variants of Kalman
Filtering [36], but it is beyond the scope of the paper to review all of these methods in detail. The particle filter
method is chosen here because they are non-parametric methods and are better suited to performing data assimilation
in systems that have non-linear and non-Gaussian behaviour [26], such as agent-based models. In fact, agent-based
models are typically formulated as computer programs rather than in the mathematical forms required of many data
assimilation methods, such as the Kalman filter and its variants [10].
The particle filter is a brute force Bayesian state estimation method. The goal is to estimate a posterior distribution (i.e.
the probability that the system is in a particular state conditioned on the observed data) using an ensemble of model
instances, called particles. Each particle has an associated weight, normalised to sum to one, that are drawn from a
prior distribution. In the data assimilation step (discussed shortly) each particle is confronted with observations from
the (pseudo) real system and weights are adjusted depending on the likelihood that a particle could have produced the
observations. Unlike most other data assimilation methods, a particle filter does not actually update the internal states
of its particles. Instead, the worst performing particles -- those that are least likely to have generated the observations
-- are removed from subsequent iterations, whereas better performing particles are duplicated. This has the advantage
that, when performing data assimilation on an agent-based model, it is not necessary to derive a means of updating
unstructured variables. For example, it is not clear how data assimilation approaches that have been designed for
mathematical models that consist of purely numerical values will update models that contain agents with categorical
variables.
A particle filter (PF) Pt at time t with N particles is the set
Pt =(cid:8)(xi
t) : i ∈ {1, . . . , N}(cid:9) ,
t, wi
t, which are subject to the condition(cid:80)N
(1)
i=1 wi
t is the state of the i-th particle with associated weight wi
where xi
t = 1
for all t. The general method of the particle filter is to use Pt and new information in the form of observations to
determine Pt+1.
There are many different PF methods. The standard form is the sequential importance sampling (SIS) PF which
selects the weights using importance sampling [37, 38]. The particles are sampled from an importance density [39].
One pitfall of the SIS PF is particle degeneracy. This occurs when the weights of all the particles tend to zero except for
one particle which has a weight very close to one. This results in the population of particles being a very poor estimate
for the posterior distribution. One method to prevent particle degeneracy is to resample the particles, duplicating
particles with a high weight and discarding particles with a low weight. The probability of a particle being resampled
is proportional to its weight; known as the sequential importance resampling (SIR) PF. The particles can be resampled
in a variety of different ways, including multinomial, stratified, residual, systematic, and component set resampling
[10, 40, 41]. Although resampling helps to increase the spread of particle weights, it is often not sufficient [36]. In
a similar manner to particle degeneration in the SIS PF, particle collapse can occur in the SIR PF. This occurs when
only one particle is resampled so every particle has the same state. One of the main drawbacks with PF methods, as
4
A PREPRINT - SEPTEMBER 23, 2019
many studies have found [e.g. 36, 42], is that the number of particles required to prevent particle degeneracy or collapse
grows exponentially with the dimensionality of the model. This is an ongoing problem that will be revisited throughout
the paper. It is worth nothing that there are many other PFs including the auxiliary SIR PF and the regularised PF [43],
the merging PF [44], and the resample-move PF [45]. In Section 3.3, we will consider a SIR PF with systematic
resampling because it ranks higher in resampling quality and computational simplicity compared to other resampling
methods [41, 46].
3 Method
3.1 The Agent-Based Model: StationSim
StationSim is a simple agent-based model that has been designed to very loosely represent the behaviour of a crowd of
people moving from an entrance on one side of a rectangular environment to an exit on the other side. This is analogous
to a train arriving at a train station and passengers moving across the concourse to leave. A number of agents, N, which
varies in the later experiments, are created when the model starts. They are able to enter the environment (leave their
train) at a uniform rate through one of three entrances. They move across the 'concourse' and then leave by one of
the two exits. The entrances and exits have a set size, such that only a limited number of agents can pass through
them in any given iteration. Once all agents have entered the environment and passed through the concourse then the
simulation ends. The model environment is illustrated in Figure 1, with the trajectories of two interacting agents for
illustration.
Figure 1: StationSim environment with 3 entrance and 2 exit doors
.
The model has deliberately been designed to be extremely simple and does not attempt to match the behavioural
realism offered by more developed crowd models [8, 22, 23, 25]. The reason for this simplicity is so that: (1) the
model can execute relatively quickly; (2) the probabilistic elements in the model are limited (we know precisely from
where probabilistic behaviour arises); (3) the model can be described fully using a relatively simple state vector, as
discussed in Section 3.3.1. Importantly, the model is able to capture the emergence of crowding. This results because
each agent has a different maximum speed that they can travel at. Therefore, when a fast agent approaches a slower
one, they attempt to get past by making a random binary choice to move left or right around them. Depending on the
agents in the vicinity, this behaviour can start to lead to the formation of crowds. To illustrate this, Figure 2 shows
the paths of the agents (2a) and the total agent density (2b) during an example simulation. The degree and location
of crowding depends on the random allocation of maximum speeds to agents and their random of direction taken to
avoid slower agents; these cannot be estimated a priori. Unlike in previous work where the models did not necessarily
meet the common criteria that define agent-based models [9, 11, e.g.] this model respects three of the most important
characteristics:
• individual heterogeneity -- agents have different maximum travel speeds;
• agent interactions -- agents are not allowed to occupy the same space and try to move around slower agents
who are blocking their path;
• emergence -- crowding is an emergent property of the system that arises as a result of the choice of exit that
each agent is heading to and their maximum speed.
The model code is relatively short and easy to understand. It is written in Python, and is available in its entirety at in
the project repository [21].
5
A PREPRINT - SEPTEMBER 23, 2019
(a) Individual trails showing the paths taken by agents
(b) The total crowd density over the simulation run.
Figure 2: An example of crowding in the StationSim model
3.2 Data Assimilation - Introduction and Definitions
DA methods are built on the following assumptions:
1. Although they have low uncertainty, observational data are often spatio-temporally sparse. Therefore there
are typically insufficient amounts of data to to describe the system in sufficient detail and a data-driven
approach would not work.
2. Models are not sparse; they can represent the target system in great detail and hence fill in the spatio-temporal
gaps in observational data by propagating data from observed to unobserved areas [36]. For example, some
parts of a building might be more heavily observed than others, so a model that assimilated data from the
observed areas might be able to estimate the state of the unobserved areas. However, if the underlying
systems are complex, a model will rapidly diverge from the real system in the absence of up to date data [11].
3. The combination a model and up-to-date observational data allow "all the available information" to be used
to determine the state of the system as accurately as possible [35].
DA algorithms work by running a model forward in time up to the point that some new observational data become
available. This is typically called the predict step. At this point, the algorithm has an estimate of the current system
state and its uncertainty (the prior). The next step, update, involves using the new observations, and their uncertainties,
to update the current state estimate to create a posterior estimate of the state. As the posterior has combined the best
guess of the state from the model and the best guess of the state from the observations, it should be a closer estimate
of the true system state than that which could be estimated from the observations or the model in isolation.
3.3 The Particle Filter
There are many different ways to perform data assimilation, as discussed in Section 2.3. Here, a potentially appropriate
solution to the data assimilation problem for agent-based models is the particle filter -- also known as a Bayesian
bootstrap filter or a sequential Monte Carlo method -- which represents the posterior state using a collection of model
samples, called particles [10, 36, 47, 48]. Figure 3 illustrates the process of running a particle filter. Note that the
'pseudo-truth model' is a single instance of the agent-based model that is used as a proxy for the real system as per
the identical twin experimental framework that we have adopted.
The data assimilation 'window' determines how often new observations are assimilated into the particle filter. The
size of the window is an important factor -- larger windows result in the particles deviating further from the real system
state -- but here we fix the window at 100 iterations. The simulation terminates when all agents have left the system.
3.3.1 The State Vector and Transition Function
Here, the state vector, at a time t, contains all the information that a transition function needs to iterate the model
forward by one step, including all of the agent (i = {0, 1, . . . , N}) parameters (−→pi ) and variables (−→vi ) as well as
global model parameters
−→
P :
(2)
(cid:104) −→p0
St =
(cid:105)
−→v0
−→p1
−→v1
. . . −→pN
−→vN
−→
P
6
A PREPRINT - SEPTEMBER 23, 2019
Figure 3: Flowchart of data assimilation process using a particle filter.
A similar structure, the observation vector, contains all of the observations made from the 'real world' (in this case the
pseudo-truth model) at a time t. Here, the particle filter is only used to estimate the state of the models variables (−→vi ),
−→
not any of the parameters (−→pi and
P ) (although it is worth noting that parameter estimation is technically feasible
and will be experimented with in future work). Also, the current speed of an agent can be calculated from its current
location and the locations of the agents surrounding it, so in effect the observation vector only needs to include the
positions of the agents with the addition of some Gaussian noise, :
Ot = [ x0
y0 x1
y1
. . .
xn
yn ]
(3)
Therefore in the experiments conducted here, all model parameters are fixed. Hence a further vector is required to
map the observations to the state vector that the particles can actually manipulate. We define the partial state vector S(cid:48)
to match the shape of O, i.e.:
S(cid:48)
t = [ x0
y0 x1
y1
. . .
xn
yn ]
(4)
This has the effect of 'pairing' agents in the particles to those in the pseudo-truth data, in a similar approach to that
taken by [10]. It is worth noting that, because the particle filter will not be tasked with parameter estimation, then the
data assimilation is somewhat simpler than it would be in a real application. For example, one of the agent parameters
is used to store the location of the exit out of the environment that the agent is moving towards. As this parameter is
set a priori for each agent, then the particle filter does not need to estimate where the agents are ultimately going, only
where they currently are.
3.3.2 Observations from the pseudo-truth data
In a real application, the particle filter would be assimilating data in real time from sensors of the real world. This is
not the case here so instead, we take 'observations' from the pseudo-truth data (which are, as it happens, generated
by the StationSim model). Each particle evolves in time according to the StationSim dynamics and receives new
observations at regular intervals. Measurement error (i.e. noise, ) is added to the observations (in the real world,
sensors observations will be noisy). Here, observations take the form of the (x, y) locations of all agents. This is
analogous to tracking all individuals in a crowd and providing snapshots of their locations at discrete points in time
to the particle filter. This 'synthetic observation' is probably more detailed than reality, so future work will vary the
amount of detail provided to the algorithm. It will firstly reduce the number of agents who are observed (e.g. tracking
only some people) and then provide only aggregate population counts (which is analogous to using a camera or other
sensor to count the number of people at a certain point).
3.3.3 Particle Weights
Each particle in the particle filter has a weight associated with it that quantifies how similar a particle is to an ob-
servation. The weights are calculated at the end of each data assimilation window (i.e. when observations become
available). At the start of the following window the particles then evolve independently from each other [49]. The
weights are, in effect, the average distance between agents in that particle and the corresponding agents in StationSim
(recall that there is a one-to-one mapping between agents in the particles and agents in the truth model). Formally, let
xn(i, t) be the location of the i-th agent at time t in the n-th particle for n ∈ {1, . . . , N} and let x(i, t) be the location
7
Pseudo-Truth ModelParticlesObservations (pseudo-truth)ReweightResampleNew set of ParticlesData AssimilationTime (model iterations)t01. . . . . 2tt+1t+2. . . . . . . t-1 A PREPRINT - SEPTEMBER 23, 2019
of the i-th agent in StationSim for i ∈ {1, . . . , I}. The error of the n-th particle n(t) at time t is then given by
I(cid:88)
i=1
n(t) =
1
I
x(i, t) − xn(i, t),
and the particle filter error ν(t) at time t is given by
N(cid:88)
N(cid:88)
n=1
n(t),
I(cid:88)
n=1
i=1
ν(t) =
1
N
=
1
N I
x(i, t) − xn(i, t).
(5)
(6)
(7)
(9)
It is worth noting that, because the agent locations are the only data stored in the partial state vector, the particle error
is equivalent to the Euclidean distance (l2-norm) between the particle partial state vector S(cid:48)
t and the observation vector
Ot,
ν(t) = S(cid:48)
t − Ot2.
Particles with relatively large error are likely to be removed during the sampling procedure (discussed in the following
section), whereas those with low error are likely to be duplicated. In addition for their use in resampling, the population
of particle weights can be used to gain insight into how well the particle filter is able to represent the 'true' system
state overall.
3.3.4 Sampling Procedure
Here, a bootstrap filter is implemented which uses systematic resampling [10, 26, 36, 41, 50]. This begins by taking
a random sample U from the uniform distribution on the interval [0, 1/N ] and then selecting N points U i for i ∈
{1, . . . , N} on the interval [0, 1] such that
(8)
Let the particles currently have locations xi. Using the inversion method, we calculate the cumulative sum of the
normalised particle weights wi and define the inverse function D of this cumulative sum, that is:
U i = (i − 1)/N + U.
i−1(cid:88)
i(cid:88)
wj,
wj
.
D(u) = i for u ∈
j=1
j=1
Finally, the new locations of the resampled particles are given by xD(U i).
As discussed in Section 2, a well-studied issue that particle filters face is that of particle deprivation [42], which refers
to the problem of particles converging to a single point such that all particles, but one, vanish [51]. Here, the problem
is addressed in two ways. Firstly by simply using large numbers of particles relative to the size of the state space
and, secondly, by diversifying the particles [52] -- also known as roughening, jittering, and diffusing [53 -- 55]. In each
iteration of the model we add Gaussian white noise to the particles' state vector to increase their variance, which
increases particle diversity. This encourages a greater variety of particles to be resampled and therefore makes the
algorithm more likely to represent the state of the underlying model. This method is a special case of the resample-
move method presented in [45]. The amount of noise to add is a key hyper-parameter -- too little and it has no effect,
too much and the state of the particles moves far away from the true state -- as discussed in the following section.
4 Experimental Results
4.1 Outline of the Experiments
Recall that the aim of the paper is to quantify the conditions under which a typical particle filter is able to reliably
estimate the 'true' state of an underlying pedestrian system through the combination of a modelled state estimate,
produced using an agent-based model, and observational data. By examining the error (i.e. the difference between
the particle filter estimate of the system state and the pseudo-truth data) under different conditions, the paper will
present some preliminary estimates of the potential for the use of particle filtering in real-time crowd simulation. In
particular, the paper will estimate the minimum number of particles (Np) that are needed to model a system that has a
given number of agents (Na). Although the dynamics of the model used here are simpler than those of more advanced
8
A PREPRINT - SEPTEMBER 23, 2019
crowd models, and hence the particle filter is tasked with an easier problem than it would have to solve in a real-world
scenario, these experiments provide valuable insight into the general potential of the method for crowd simulation.
It is well known that one of the main drawbacks with a particle filter is that the number of particles required can explode
as the complexity of the system increases [e.g. 36, 42]. In this system, the complexity of the model is determined by
the number of agents. As outlined in Section 3, randomness is introduced through agent interactions. Therefore the
fewer the number of agents, the smaller the chance of interactions occurring and the lower the complexity. As the
number of agents increases,we find that the number of interactions increases exponentially (Figure 5, discussed in the
following section, will illustrate this). The paper will also experiment with the amount of particle noise that needs to
be included to reduce the problem of particle deprivation (σp), but this is not a focus of the experiments.
To quantify the 'success' of each parameter configuration -- i.e. the number of particles Np and amount of particle
noise σp that allow the particle filter to reliably represent a system of Na agents -- an estimate of the overall error
associated with a particular particle filter configuration is required. There are a number of different ways that this error
could be calculated. Here, we calculate the mean weight of each particle after resampling in every data assimilation
window (the weights vary during an experiment because in early windows there are few agents and hence very little
stochasticity so particle errors are low). Then we take the mean of these individual particle errors. In addition, because
the results from an individual experiment can vary slightly, each particle configuration is executed a number of times,
M, and the median error across experiments is calculated.
Formally if ν(t, j) is the error of particle j in a data assimilation window t, and the total number of windows is T ,
then the total error of that particle filter configuration across M experiments is:
(cid:80)T
Np(cid:88)
j=1
t=1 ν(t)
T
ENp,σp,Na = medianM
(10)
where medianM () calculates the median over M experiments.
In summary, the Table 1 outlines the parameters that are used in the experiments. There are a number of other model
and particle filter parameters that can vary, but ultimately these do not influence the results outlined here and are not
experimented with. The parameters, and code, are available in full from the project repository [21].
Table 1: Main parameters used in the experiments
Symbol Value / Range
Parameter
Number of agents
Na
Number of particles
Np
Particle noise
σp
Measurement noise
σm
Number of experiments (repetitions)
M
Model iterations in each data assimilation window -
[2, 40]
[1, 10000]
[0.25, 0.5]
1.0
20
100
4.2 Results
4.2.1 Overall Error
Figure 4 plots the median of the mean error over all experiments to show how it varies with differing numbers of
agents and particles. Due to the computational difficulty in running experiments with large numbers of particles and
the need for an exponential increase in the number of particles with the number of agents (discussed in detail below),
there are fewer experiments with large numbers of particles and hence the the experiments are not distributed evenly
across the agents/particles space. Thus the error is presented in the form of an interpolated heat map. Broadly there is a
reduction in error from the bottom-right corner (few particles, many agents) to the top left corner (many particles, few
agents). The results illustrate that, as expected, there is a larger error with increasing numbers of agents but this can
be mediated with larger numbers of particles. Note the logarithmic scale used on the vertical axis; an exponentially
greater number of particles is required for each additional agent included in the simulation.
There are two reasons for this exponential increase in the number of particles required. Firstly, as the number of agents
increases so does the dimensionality of the state space. Also, and perhaps more importantly, with additional agents the
chances of collisions, and hence stochastic behaviour, increases exponentially. This is illustrated by Figure 5 which
presents the total number of collisions that occur across a number of simulations with a given number of agents.
9
A PREPRINT - SEPTEMBER 23, 2019
Figure 4: Median of the mean errors after resampling
Figure 5: The total number of collisions that occur given the number of agents. The blue line represents a polynomial
regression model of order 2 with 99% confidence intervals.
On its own, the overall particle filter error (Figure 4) reveals little information about which configurations would be
judged 'sufficiently reliable' to be used in practice. Therefore it is illuminating to visualise some of the results of
individual particle filter runs to see how the estimates of the individual agents' locations vary, and what might be
considered an 'acceptable' estimate of the true state. Figure 6 illustrates the state of a particle filter with 10 particles
and 10 agents at the end of its first data assimilation window (after resampling). With only ten agents in the system
there are few, if any, collisions and hence very little stochasticity; all particles are able to represent the locations of the
agents accurately.
As the number of agents increases, collisions become much more likely and the chance of a particle diverging from the
pseudo-truth state increases considerably. It becomes more common that no single particle will correctly capture the
behaviours of all agents. Therefore even after resampling there are some agents whose locations the particle filter is
10
A PREPRINT - SEPTEMBER 23, 2019
Figure 6: State of a particle filter with 10 agents and 10 particles after 100 and 200 iterations (after resampling at
the end of the first and second data assimilation windows). Black squares illustrate the (pseudo) real locations of
the agents and red circles represent the locations of those agents as predicted by individual particles (in this case all
particles predict the agents' locations accurately and hence the red circles overlap).
not able to accurately represent. Figure 7 illustrates a filter running with 40 agents and still only 10 particles. The long
black lines show the locations of pseudo-real agents and the corresponding agents in the individual particles; it is clear
that for some agents none of the particles have captured their pseudo-real locations. This problem can be mediated by
increasing the number of particles. Figure 4 showed that with approximately 10,000 particles the error for simulations
with 40 agents drops to levels that are comparable to the simulations of 10 agents. Hence a rule of thumb is that any
particle filter with an overall error that is comparable to a clearly successful filter (i.e. Figure 6) are reliably estimating
the state of the system.
Figure 7: State of a particle filter with 50 agents and 10 particles. Lines connecting black squares (the pseudo-real
agent locations) to red circles (particle estimate of the agent location) show that some particle estimates are a long way
from the true agent locations.
4.2.2 Impact of Resampling
Resampling is the process of weighting all particles according to how well they represent the pseudo-truth data; those
with higher weights are more likely to be sampled and used in the following data assimilation window. This is
important to analyse because it is the means by which the particle filter improves the overall quality of its estimates.
Figure 8 illustrates the impact of resampling on the error of the particle filter. With fewer than 10 agents in the system
resampling is unnecessary because all particles are able to successfully model the state of the system. With more than
10 agents, however, it becomes clear that the population of particles will rapidly diverge from the pseudo-truth system
state and resampling is essential for the filter manage to limit the overall error.
11
0255075100125150175200X position020406080100Y position10 agents, 10 particles, 100 iterations0255075100125150175200X position020406080100Y position10 agents, 10 particles, 200 iterations0255075100125150175200X position020406080100Y position40 agents, 10 particles, 100 iterations0255075100125150175200X position020406080100Y position40 agents, 10 particles, 200 iterations A PREPRINT - SEPTEMBER 23, 2019
Figure 8: Median of the mean errors before and after resampling
4.2.3 Particle Variance
As discussed in Section 3.3.4, the variance of the population of particles can be a reliable measure for estimating
whether particle deprivation is occurring. If there is very little variance in the particles, then it is likely that they have
converged close to a single point in the space of possible states. This needs to be avoided because, in such situations,
it is extremely unlikely that any particles will reliably represent the state of the real system. Figure 9 illustrates this by
visualising the mean error of all particles, ν(t) (defined in Equation 6), in each data assimilation window, t, and their
variance under different numbers of agents and particles. Note that each agent/particle configuration is executed 10
times and the results are visualised as boxplots. Also, simulations with larger numbers of agents are likely to run for a
larger number of iterations, but the long-running models usually have very few agents in them in later iterations (most
have left, leaving only a few slow agents). Hence only errors up to 600 iterations, where most of the agents have left
the environment in most of the simulations, are shown. The graphs in Figure 9 can be interpreted as follows:
• With 5 agents and 10 particles there is very low error and low variance (so low that the box plots are difficult
to make out on Figure 9). This suggests that particle deprivation is occurring, but there are so few agents that
the simulation is largely deterministic so the particles are likely to simulate the pseudo-truth observations
accurately regardless.
• When the number of agents is increased to 30 agents and 10 particles, the errors are much larger. The
increased non-linearity introduced by the greater number of agents (and hence greater number of collisions)
means that the population of particles, as a whole, is not able to simulate the pseudo-truth data. Although
particle variance can be high, none of the particles are successfully simulating the target.
• Finally, with 30 agents and 10,000 particles, the errors are relatively low in comparison, especially after the
first few data assimilation windows.
4.2.4 Particle Noise
The number of particles (Np) is the most important hyper-parameter, but the amount of noise added to the particles
(σp) is also important as this is the means by which particle deprivation is prevented. However, the addition of too
much noise will push a particle a long way away from the true underlying state. Under these circumstances, although
particle deprivation is unlikely none of the particles will be close to the true state. To illustrate this, Figure 10 presents
the median error with a greater amount of noise (σp = 0.5). The overall pattern is similar to the equivalent graph
produced with σp = 0.25 in Figure 4 -- as the number of agents in the simulation (Na) increases so does the number of
particles required to maintain low error (Np) -- but the overall error in each Na/Np combination is substantially larger
for the experiments with additional noise. Future work could explore the optimal amount of noise in more detail;
σ = 0.25 was found to be the most reliable through trial and error.
12
A PREPRINT - SEPTEMBER 23, 2019
Figure 9: Mean error and variance of particles under different combinations of the number of particles and number of
agents
5 Discussion and Future Work
5.1 Discussion
This paper has experimented with the use of a sequential importance resampling (SIR) particle filter (PF) as a means
of dynamically incorporating data into a simple agent-based model of a pedestrian system. The results demonstrate
that it is possible to use a particle filter to perform dynamic adjustment of the model. However, they also show that
(as expected [36, 56]) as the dimensionality of the system increases, the number of particles required to maintain an
acceptable approximation error grows exponentially. The reason for this is because, as the dimensionality increases,
it becomes less likely that an individual particle will have the 'correct combination' of values [26]. In this work, the
dimensionality is proportional to the number of agents. At most 10,000 particles were used, which was sufficient for
a simulation with 30-40 agents. However, for a more complex and realistic model containing hundreds or thousands
of agents, the number of particles required would most likely number in the millions. The particle filter used in this
study was provided with more information than would normally be available. For example, information was supplied
as fixed parameters on when each agent entered the simulation, their maximum speeds, and their chosen destinations.
Therefore the only information that the particle filter was lacking was the actual locations of the agents and whether
13
A PREPRINT - SEPTEMBER 23, 2019
Figure 10: Median of the mean errors after resampling with additional noise (σ = 0.5).
they would chose to move left or right to prevents a collision with another agent. It is entirely possible to include
these agent-level parameters in the state vector, but this would further increase the size of the state space and hence
the number of particles required. This is an important caveat as in a real-world situation it is very unlikely that such
detail would be available. Future work should begin to experiment with the number of particles that would be required
when observational data that are more akin to those available in the real world are used.
There are a number of possible improvements that could be made to the basic SIR particle filter to reduce the number
of particles required. For example, [10] propose component set resampling -- details below -- but exploring these
further is beyond the scope of this paper. Overall, these results reflect those of other studies which demonstrate that
particle filtering has value for simulations with relatively few agents and interactions [e.g. 10, 31, 57] but that the large
dimensionality of a pedestrian system poses problems for the standard (unmodified) bootstrap filter.
5.2
Improvements to the particle filter
There are a number of adaptions to the main particle filtering algorithm that might make the method more amenable to
use with complex agent-based models. The aforementioned Component Set Resampling [10] approach proposes that
individual components of particles are sampled, rather than whole particles in their entirety. A more commonly used
approach is to reduce the dimensionality of the problem in the first place. With spatial agent-based models, such as the
one used here, spatial aggregation provides such an opportunity. In the data assimilation stage, the state vector could
be converted to an aggregate form, such as a population density surface, and particle filtering could be conducted on
the aggregate surface rather than on the individual locations of the agents. After assimilation, the particles could be
disaggregated and then run as normal. This will, of course, introduce error because the exact positions of agents in the
particles will not be known when disaggregating, but that additional uncertainty might be outweighed by the benefits
of a more successful particle filtering overall. In addition, the observations that could be presented to an aggregate
particle filter might be much more in line with those that are available in practice (as will be discussed shortly). If
the aim of real-time model calibration is to give decision makers a general idea about how the system is behaving in
real time, then this additional uncertainty might not be problematic. A similar approach, proposed by [56], could be to
divide up the space into smaller regions and then apply the algorithm locally to these regions (i.e. a divide-and-conquer
approach). Although it is not clear how well divide-and-conquer would work in an agent-based model -- for example,
[26] developed the method for a discrete cellular automata model -- it would be an interesting avenue to explore.
5.3 Real-World Implications and Future Work
It is important to note that, unlike other data assimilation approaches, the particle filter does not dynamically alter
the state of the running model. This could be advantageous because, with agent-based modelling, it is not clear
that the state of an agent should be manipulated by an external process. Agents typically have goals and a history,
and behavioural rules that rely on those features, so artificially altering an agent's internal state might disrupt their
behaviour making it, at worst, nonsensical. Experiments with alternative (potentially more efficient) algorithms such
as 4DVar or a the Ensemble Kalman Filter should be conducted to test this.
14
A PREPRINT - SEPTEMBER 23, 2019
Ultimately the aim of this work is to develop methods that will allow simulations of human pedestrian systems to be
optimised in real time. Not only will such methods provide decision makers with more accurate information about the
present, but they could also allow for better predictions of the near future [c.f. 57]. One assumption made throughout
the paper, which limits its direct real-world use, is that the locations of pseudo-real individuals are known, albeit with
some uncertainty. Not only is this assumption unrealistic -- it is rare for individuals to be tracked to this degree in
public places -- but we would also argue that the privacy implications of tracking individual people are not outweighed
by the benefits offered by better understanding the system. Therefore, immediate future work will test how well a
data assimilation algorithm would fare were it supplied only aggregate information such as the number of people
who pass through a barrier, or the number of people recorded by a CCTV camera within a particular area. Both of
these measures can be used in aggregate form and would be entirely anonymous. It is unclear whether such aggregate
data would be sufficient to identify a 'correct' agent-based model, so experiments should explore the spatio-temporal
resolution of the aggregate data are required. Also, identifiability/equifinality analysis might help initially as a means
of estimating whether the available data are sufficient to identify a seemingly 'correct' model in the first place. In the
end, such research might help to provide evidence to policy makers for the number, and characteristics, of the sensors
that would need to be installed to accurately simulate the target system, and how these could be used to maintain the
privacy of the people who they are recording.
References
[1] E Bonabeau. Agent based modeling: Methods and techniques for simulating human systems. Proceedings of the
National Academy of Sciences, 99(90003):7280 -- 7287, 2002.
[2] J. Doyne Farmer and Duncan Foley. The economy needs agent-based modelling. Nature, 460(7256):685 -- 686,
August 2009.
[3] Colin M. Henein and Tony White. Agent-Based Modelling of Forces in Crowds. In David Hutchison, Takeo
Kanade, Josef Kittler, Jon M. Kleinberg, Friedemann Mattern, John C. Mitchell, Moni Naor, Oscar Nierstrasz,
C. Pandu Rangan, Bernhard Steffen, Madhu Sudan, Demetri Terzopoulos, Dough Tygar, Moshe Y. Vardi, Ger-
hard Weikum, Paul Davidsson, Brian Logan, and Keiki Takadama, editors, Multi-Agent and Multi-Agent-Based
Simulation, volume 3415, pages 173 -- 184. Springer Berlin Heidelberg, Berlin, Heidelberg, 2005.
[4] Andrew T. Crooks and Sarah Wise. GIS and agent-based models for humanitarian assistance. Computers,
Environment and Urban Systems, 41:100 -- 111, 2013.
[5] Chuanjun Ren, Chenghui Yang, and Shiyao Jin. Agent-Based Modeling and Simulation on Emergency Evacu-
ation. In Ozgur Akan, Paolo Bellavista, Jiannong Cao, Falko Dressler, Domenico Ferrari, Mario Gerla, Hisashi
Kobayashi, Sergio Palazzo, Sartaj Sahni, Xuemin (Sherman) Shen, Mircea Stan, Jia Xiaohua, Albert Zomaya,
Geoffrey Coulson, and Jie Zhou, editors, Complex Sciences, volume 5, pages 1451 -- 1461. Springer Berlin Hei-
delberg, Berlin, Heidelberg, 2009.
[6] Timothy Schoenharl and Greg Madey. Design and Implementation of An Agent-Based Simulation for Emergency
Response and Crisis Management. Journal of Algorithms & Computational Technology, 5(4):601 -- 622, 2011.
[7] Zarita Zainuddin, Kumatha Thinakaran, and Ibtesam Abu-Sulyman. Simulating the Circumambulation of the
Ka'aba using SimWalk. European Journal of Scientific Research, 38:454 -- 464, 2009.
[8] Dirk Helbing, Illés Farkas, and Tamás Vicsek. Simulating dynamical features of escape panic. Nature, 407:487
EP -- , September 2000.
[9] David J. B. Lloyd, Naratip Santitissadeekorn, and Martin B. Short. Exploring data assimilation and forecasting
issues for an urban crime model. European Journal of Applied Mathematics, 27(Special Issue 03):451 -- 478,
2016.
[10] Minghao Wang and Xiaolin Hu. Data assimilation in agent based simulation of smart environments using particle
filters. Simulation Modelling Practice and Theory, 56:36 -- 54, 2015.
[11] Jonathan A. Ward, Andrew J. Evans, and Nicolas S. Malleson. Dynamic calibration of agent-based models using
data assimilation. Royal Society Open Science, 3(4), 2016.
[12] Eugenia Kalnay. Atmospheric Modeling, Data Assimilation and Predictability. Cambridge University Press,
2003.
[13] John M. Lewis, S. Lakshmivarahan, and Sudarshan Dhall. Dynamic Data Assimilation: A Least Squares Ap-
proach. Cambridge University Press, Cambridge, 2006.
[14] Jan C. Thiele, Winfried Kurth, and Volker Grimm. Facilitating Parameter Estimation and Sensitivity Analysis of
Agent-Based Models: A Cookbook Using NetLogo and 'R'. Journal of Artificial Societies and Social Simulation,
17(3), 2014.
15
A PREPRINT - SEPTEMBER 23, 2019
[15] M. Pennisi, R. Catanuto, F. Pappalardo, and S. Motta. Optimal vaccination schedules using simulated annealing.
Bioinformatics, 24(15):1740 -- 1742, August 2008.
[16] A. J. Heppenstall, A.J. Evans, and M H Birkin. Genetic algorithm optimisation of an agent-based model for
simulating a retail market. Environment and Planning B: Planning and Design, 34:1051 -- 1070, 2007. Cited by
0018.
[17] Jakob Grazzini, Matteo G. Richiardi, and Mike Tsionas. Bayesian estimation of agent-based models. Journal of
Economic Dynamics and Control, 77:26 -- 47, April 2017.
[18] Jonathan Ozik, Nicholson T. Collier, Justin M. Wozniak, Charles M. Macal, and Gary An. Extreme-Scale
Dynamic Exploration of a Distributed Agent-Based Model With the EMEWS Framework. IEEE Transactions
on Computational Social Systems, 5(3):884 -- 895, September 2018.
[19] Francis Oloo and Gudrun Wallentin. An Adaptive Agent-Based Model of Homing Pigeons: A Genetic Algorithm
Approach. ISPRS International Journal of Geo-Information, 6(1):27, January 2017.
[20] Francis Oloo, Kamran Safi, and Jagannath Aryal. Predicting Migratory Corridors of White Storks, Ciconia
ciconia, to Enhance Sustainable Wind Energy Planning: A Data-Driven Agent-Based Model. Sustainability,
10(5):1470, 2018.
[21] Dust project code repository, including stationsim. XXXXgithuburl. Accessed: XX/XX/2019.
[22] Franziska Klügl and Guido Rindsfüser. Large-Scale Agent-Based Pedestrian Simulation. In Paolo Petta, Jörg P.
Müller, Matthias Klusch, and Michael Georgeff, editors, Multiagent System Technologies, volume 4687, pages
145 -- 156. Springer Berlin Heidelberg, Berlin, Heidelberg, 2007.
[23] Xu Chen, Haiying Li, Jianrui Miao, Shixiong Jiang, and Xi Jiang. A multiagent-based model for pedestrian
simulation in subway stations. Simulation Modelling Practice and Theory, 71:134 -- 148, February 2017.
[24] Inhi Kim, Ronald Galiza, and Luis Ferreira. Modeling pedestrian queuing using micro-simulation. Transporta-
tion Research Part A: Policy and Practice, 49:232 -- 240, March 2013.
[25] C. Natalie van der Wal, Daniel Formolo, Mark A. Robinson, Michael Minkov, and Tibor Bosse. Simulating
In Jacek Mercik, editor, Trans-
Crowd Evacuation with Socio-Cultural, Cognitive, and Emotional Elements.
actions on Computational Collective Intelligence XXVII, volume 10480, pages 139 -- 177. Springer International
Publishing, Cham, 2017.
[26] Yuan Long and Xiaolin Hu. Spatial Partition-Based Particle Filtering for Data Assimilation in Wildfire Spread
Simulation. ACM Transactions on Spatial Algorithms and Systems, 3(2):1 -- 33, August 2017.
[27] Frederica Darema. Dynamic data driven applications systems: A new paradigm for application simulations and
measurements. In Computational Science-ICCS 2004, pages 662 -- 669. Springer, 2004.
[28] Timothy W. Schoenharl and Greg Madey. Evaluation of Measurement Techniques for the Validation of Agent-
Based Simulations Against Streaming Data. In Marian Bubak, Geert Dick van Albada, Jack Dongarra, and Peter
M. A. Sloot, editors, Computational Science -- ICCS 2008, number 5103 in Lecture Notes in Computer Science,
pages 6 -- 15. Springer Berlin Heidelberg, 2008.
[29] Haifeng Zhang, Yevgeniy Vorobeychik, Joshua Letchford, and Kiran Lakkaraju. Data-Driven Agent-Based Mod-
In Proceedings of the 2015 International Conference on
eling, with Application to Rooftop Solar Adoption.
Autonomous Agents and Multiagent Systems, AAMAS '15, pages 513 -- 521, Richland, SC, 2015. International
Foundation for Autonomous Agents and Multiagent Systems.
[30] Nasri Bin Othman, Erika Fille Legara, Vicknesh Selvam, and Christopher Monterola. A data-driven agent-based
model of congestion and scaling dynamics of rapid transit systems. Journal of Computational Science, 10:338 --
350, 2015.
[31] Jordan Lueck, Jason H. Rife, Samarth Swarup, and Nassin Uddin. Who Goes There? Using an Agent-based
Simulation for Tracking Population Movement. In Winter Simulation Conference, Dec 8 - 11, 2019., National
Harbor, MD, USA, 2019.
[32] Peter Bauer, Alan Thorpe, and Gilbert Brunet. The quiet revolution of numerical weather prediction. Nature,
525(7567):47 -- 55, 2015.
[33] R. A. Panofsky. Objective Weather-Map Analysis. Journal of Meteorology, 6(6):386 -- 392, 1949.
[34] J. G. Charney. Dynamic Forecasting by Numerical Process.
In Thomas F. Malone, editor, Compendium of
Meteorology, pages 470 -- 482. American Meteorological Society, Boston, MA, 1951.
[35] Oliver Talagrand. The Use of Adjoint Equations in Numerical Modelling of the Atmospheric Circulation. In An-
dreas Griewank and George F. Corliss, editors, Automatic Differentiation of Algorithms: Theory, Implementation,
and Application, pages 169 -- 180. SIAM, Philadelphia, PA, 1991.
16
A PREPRINT - SEPTEMBER 23, 2019
[36] Alberto Carrassi, Marc Bocquet, Laurent Bertino, and Geir Evensen. Data assimilation in the geosciences: An
overview of methods, issues, and perspectives. Wiley Interdisciplinary Reviews: Climate Change, 9(5):e535,
September 2018.
[37] Niclas Bergman. Recursive Bayesian Estimation: Navigation and Tracking Applications. Univ., Linköping,
1999. OCLC: 186143552.
[38] Arnaud Doucet, Simon Godsill, and Christophe Andrieu. On sequential Monte Carlo sampling methods for
Bayesian filtering. Statistics and Computing, 10(3):197 -- 208, July 2000.
[39] K. Uosaki, Y. Kimura, and T. Hatanaka. Nonlinear state estimation by evolution strategies based particle filters. In
The 2003 Congress on Evolutionary Computation, 2003. CEC '03., volume 3, pages 2102 -- 2109 Vol.3, December
2003.
[40] Jun S. Liu and Rong Chen. Sequential Monte Carlo Methods for Dynamic Systems. Journal of the American
Statistical Association, 93(443):1032 -- 1044, September 1998.
[41] Randal Douc, Olivier Cappé, and Eric Moulines. Comparison of Resampling Schemes for Particle Filtering.
arXiv:cs/0507025, July 2005.
[42] Chris Snyder, Thomas Bengtsson, Peter Bickel, and Jeff Anderson. Obstacles to High-Dimensional Particle
Filtering. Monthly Weather Review, 136(12):4629 -- 4640, December 2008.
[43] M. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp. A tutorial on particle filters for online nonlinear/non-
Gaussian Bayesian tracking. IEEE Transactions on Signal Processing, 50(2):174 -- 188, February 2002.
[44] S. Nakano, G. Ueno, and T. Higuchi. Merging particle filter for sequential data assimilation. Nonlinear Processes
in Geophysics, 14(4):395 -- 408, July 2007.
[45] Walter R. Gilks and Carlo Berzuini. Following a moving target-Monte Carlo inference for dynamic Bayesian
models. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 63(1):127 -- 146, February
2001.
[46] J. D. Hol, T. B. Schon, and F. Gustafsson. On Resampling Algorithms for Particle Filters.
Nonlinear Statistical Signal Processing Workshop, pages 79 -- 82, September 2006.
In 2006 IEEE
[47] N.J. Gordon, D.J. Salmond, and A.F.M. Smith. Novel approach to nonlinear/non-Gaussian Bayesian state esti-
mation. IEE Proceedings F Radar and Signal Processing, 140(2):107, 1993.
[48] J. Carpenter, P. Clifford, and P. Fearnhead. Improved particle filter for nonlinear problems. IEE Proceedings -
Radar, Sonar and Navigation, 146(1):2 -- 7, February 1999.
[49] Paul Fearnhead and Hans R. Künsch. Particle Filters and Data Assimilation. Annual Review of Statistics and Its
Application, 5(1):421 -- 449, 2018.
[50] Arnaud Doucet, Nando Freitas, and Neil Gordon. An Introduction to Sequential Monte Carlo Methods.
In
Arnaud Doucet, Nando Freitas, and Neil Gordon, editors, Sequential Monte Carlo Methods in Practice, pages
3 -- 14. Springer New York, New York, NY, 2001.
[51] Augustine Kong, Jun S. Liu, and Wing Hung Wong. Sequential Imputations and Bayesian Missing Data Prob-
lems. Journal of the American Statistical Association, 89(425):278 -- 288, March 1994.
[52] P. Vadakkepat and L. Jing. Improved Particle Filter in Sensor Fusion for Tracking Randomly Moving Object.
IEEE Transactions on Instrumentation and Measurement, 55(5):1823 -- 1832, October 2006.
[53] Tiancheng Li, Shudong Sun, Tariq Pervez Sattar, and Juan Manuel Corchado. Fight sample degeneracy and
impoverishment in particle filters: A review of intelligent approaches. Expert Systems with Applications,
41(8):3944 -- 3954, June 2014.
[54] Neil Shephard and Thomas Flury. Learning and filtering via simulation: Smoothly jittered particle filters. Tech-
nical Report 469, University of Oxford, Department of Economics, December 2009.
[55] Juan Jose Pantrigo, Angel Sanchez, Kostas Gianikellis, and Antonio S. Montemayor. Combining Particle Filter
and Population-based Metaheuristics for Visual Articulated Motion Tracking. ELCVIA Electronic Letters on
Computer Vision and Image Analysis, 5(3):68 -- 83, November 2005.
[56] Patrick Rebeschini and Ramon van Handel. Can local particle filters beat the curse of dimensionality? The
Annals of Applied Probability, 25(5):2809 -- 2866, October 2015.
[57] Le-Minh Kieu, Nicolas Malleson, and Alison Heppenstall. Dealing with uncertainty in agent-based models for
short-term predictions. arXiv:1908.08288 [cs], 2019.
17
A PREPRINT - SEPTEMBER 23, 2019
A Instructions for Running the Source Code
The source code to run the StationSim model and the particle filter experiments can be found in the main 'Data
Assimilation for Agent-Based Modelling' (DUST) project repository:
github.com/urban-analytics/dust
Specifically, scripts and instructions to run the experiments are available at:
github.com/Urban-Analytics/dust/tree/master/Projects/ABM_DA/experiments/pf_experiments
18
|
1710.10013 | 1 | 1710 | 2017-10-27T07:39:49 | Declarative vs Rule-based Control for Flocking Dynamics | [
"cs.MA",
"eess.SY"
] | The popularity of rule-based flocking models, such as Reynolds' classic flocking model, raises the question of whether more declarative flocking models are possible. This question is motivated by the observation that declarative models are generally simpler and easier to design, understand, and analyze than operational models. We introduce a very simple control law for flocking based on a cost function capturing cohesion (agents want to stay together) and separation (agents do not want to get too close). We refer to it as {\textit declarative flocking} (DF). We use model-predictive control (MPC) to define controllers for DF in centralized and distributed settings. A thorough performance comparison of our declarative flocking with Reynolds' model, and with more recent flocking models that use MPC with a cost function based on lattice structures, demonstrate that DF-MPC yields the best cohesion and least fragmentation, and maintains a surprisingly good level of geometric regularity while still producing natural flock shapes similar to those produced by Reynolds' model. We also show that DF-MPC has high resilience to sensor noise. | cs.MA | cs |
Declarative vs Rule-based Control for Flocking Dynamics
Usama Mehmood
Department of Computer Science,
Stony Brook University, USA
Nicola Paoletti
Department of Computer Science,
Stony Brook University, USA
Radu Grosu
Cyber-Physical Systems Group,
Technische Universitat Wien, Austria
Ashish Tiwari
SRI International, USA
Shan Lin
University, USA
Junxing Yang
Department of Electrical and
Computer Engineering, Stony Brook
Department of Computer Science,
Stony Brook University, USA
Dung Phan
Department of Computer Science,
Stony Brook University, USA
Scott D. Stoller
Department of Computer Science,
Stony Brook University, USA
Scott A. Smolka
Department of Computer Science,
Stony Brook University, USA
ABSTRACT
The popularity of rule-based flocking models, such as Reynolds'
classic flocking model, raises the question of whether more declar-
ative flocking models are possible. This question is motivated by
the observation that declarative models are generally simpler and
easier to design, understand, and analyze than operational models.
We introduce a very simple control law for flocking based on a
cost function capturing cohesion (agents want to stay together) and
separation (agents do not want to get too close). We refer to it as
declarative flocking (DF). We use model-predictive control (MPC)
to define controllers for DF in centralized and distributed settings. A
thorough performance comparison of our declarative flocking with
Reynolds' classic flocking model, and with more recent flocking
models that use MPC with a cost function based on lattice struc-
tures, demonstrate that DF-MPC yields the best cohesion and least
fragmentation, and maintains a surprisingly good level of geomet-
ric regularity while still producing natural flock shapes similar to
those produced by Reynolds' model. We also show that DF-MPC
has high resilience to sensor noise.
ACM Reference Format:
Usama Mehmood, Nicola Paoletti, Dung Phan, Radu Grosu, Shan Lin, Scott
D. Stoller, Ashish Tiwari, Junxing Yang, and Scott A. Smolka. 2018. Declara-
tive vs Rule-based Control for Flocking Dynamics. In Proceedings of ACM/SIGAPP
Symposium On Applied Computing (SAC 2018). ACM, New York, NY, USA,
8 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn
1 INTRODUCTION
Flocking is a collective behavior exhibited by a large number of
interacting agents possessing a common group objective [7]. The
term is most commonly associated with birds, and more recently,
drones. Examples include foraging for food, executing a predator-
avoidance maneuver, and engaging in migratory behavior.
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
SAC 2018, April 9-13, 2018, Pau, France
© 2018 Copyright held by the owner/author(s).
ACM ISBN 978-x-xxxx-xxxx-x/YY/MM.
https://doi.org/10.1145/nnnnnnn.nnnnnnn
With the introduction of Reynolds' model [12, 13], rule-based
control became the norm in the flocking community. Specifically,
in this model, at each time-step, each agent executes a control law
given in terms of the weighted sum of three competing forces to
determine its next acceleration. Each of these forces has its own rule:
separation (keep a safe distance away from your neighbors), cohesion
(move towards the centroid of your neighbors), and alignment (steer
toward the average heading of your neighbors). As the descriptions
suggest, these rules are executed by each agent in a distributed
environment with limited-range sensing and no communication.
The popularity of Reynolds' model and its many variants raises
the question: Is there a more abstract declarative form of control for
flocking? This question is important because declarative models
are generally simpler and easier to design, understand, and analyze
than operational models. This is analogous to declarative programs
(e.g., functional programs and logic programs) being easier to write
and verify than imperative programs.
We show that the answer to this question is indeed positive by
providing a very simple control law for flocking based on a cost
function comprising two main terms: cohesion (the average squared
distance between all pairs of agents) and separation (a sum of in-
verse squared distances, except this time between pairs of agents
within each other's sensing range). That is it. For example, no term
representing velocity alignment is needed. The cost function speci-
fies what we want as the goal, and is hence declarative. In contrast,
the update rules in Reynolds' model aim to achieve an implicit goal
and hence are operational. Executing declarative control amounts
to finding the right balance between attracting and repelling forces
between agents. We refer to this approach as Declarative Flocking
(DF). We use MPC (model-predictive control) to define controllers
for DF, and refer to this approach as DF-MPC. We define a cen-
tralized version of DF-MPC, which requires communication, and a
distributed version, which does not.
Previous MPCs for flocking exist, e.g., [16–18]. Most of these
MPCs are designed to conform to the α-lattice model of flocking
proposed in [7]. α-lattices impose a highly regular structure on
flocks: all neighboring agents are distance d apart, for a specified
constant d. This kind of structure is seen in some settings, such as
beehives, but is not expected in many other natural and engineered
settings, and it is not imposed by Reynolds' model.
SAC 2018, April 9-13, 2018, Pau, France
U. Mehmood et al.
In this paper, we show, via a thorough performance evaluation,
how centralized and distributed DF-MPC compare to Reynolds' rule-
based approach [12, 13], Olfati-Saber's potential-based approach
[7], a variant of Zhan and Li's centralized lattice-based MPC ap-
proach [15, 16], and Zhang et al.'s distributed lattice-based MPC
approach [17]. We consider performance measures that capture mul-
tiple dimensions of flocking behavior: number of sub-flocks (flock
fragmentation), maximum sub-flock diameter (cohesion), velocity
convergence, and a new parameter-free measure of the geometric
regularity of the formation.
Our experimental results demonstrate that DF-MPC yields the
best cohesion and least fragmentation, and produces natural flock
shapes like those produced by Reynolds' model. Also, distributed
DF-MPC maintains a surprisingly good level of geometric regularity.
We also analyze the resiliency of DF-MPC and the lattice-based
MPC approaches by considering the impact of sensor noise. Our
results demonstrate a remarkably high level of resiliency on the
part of DF-MPC in comparison with these other approaches.
The rest of the paper is organized as follows. Section 2 presents
the rule-based, potential-based, and lattice-based MPC approaches
mentioned above. Section 3 defines our declarative flocking ap-
proach. Section 4 defines our performance measures for flocking
models. Section 5 presents our experimental results and perfor-
mance evaluation. Section 6 discusses related work. Finally, Sec-
tion 7 offers our concluding remarks and directions for future work.
xi(k + 1) = xi(k) + dT · vi(k), vi(k) ∈ V
vi(k + 1) = vi(k) + dT · ai(k), ai(k) ∈ A,
2 MODELS OF FLOCKING BEHAVIOR
We consider a set of dynamic agents B = {1, . . . , n} that move
according to the following discrete-time equation of motion:
(1)
(2)
where xi(k), vi(k), ai(k) ∈ Rm are respectively position, velocity
and acceleration of agent i ∈ B in the m-dimensional space at
step k, and dT ∈ R+ is the time step. We consider physical con-
straints on velocities and accelerations, described by the sets V
v ≤ ¯v} and
and A, respectively, which are defined by V = {v
a ≤ ¯a}, where ¯v and ¯a limit the allowed magnitude of
A = {a
the velocity and acceleration vectors, respectively.
In most flocking models, agents update their motion by changing
their acceleration. In this sense, ai(k) represents the control input
for agent i.
The configuration of all agents is described by the vector x(k) =
[xT
1 (k) . . . xT
n (k)]T ∈
n(k)]T ∈ Rm·n. Then the equation
Rm·n, and a(k) = [aT
of motion for all agents can be expressed as
(3)
(4)
The local neighborhood of agent i is defined by the set of other
agents, called neighbors, within a given distance from i, mimicking
the agent's visibility sphere. For an interaction radius r > 0 and
configuration x, the set of spatial neighbors of agent i, Ni(x) ⊆ B,
is given by:
(5)
1 (k) . . . aT
x(k + 1) = x(k) + dT · v(k),
v(k + 1) = v(k) + dT · a(k),
n (k)]T ∈ Rm·n. Let v(k) = [vT
Ni(x) =(cid:8)j ∈ B j , i ∧ ∥xi − xj ∥ < r(cid:9) ,
where ∥ · ∥ denotes the Euclidean norm.
1 (k) . . . vT
Figure 1: Examples of α-lattice a) and quasi α-lattice b). Solid
lines connect agents in the same neighborhood that have dis-
tance d. Dashed lines connect those with have distance d ± ϵ
for ϵ ≤ δ (the tolerance).
E(x) =(cid:8)(i, j) ∈ B × B ∥xi − xj ∥ < r , i , j(cid:9) ,
For configuration x ∈ Rm·n, we define the associated proximity
net G(x) = (B, E(x)) as the graph that connects agents within their
interaction radius:
(6)
To capture the regular geometry of flocks, Olfati-Saber intro-
duced the notions of α-lattices, i.e. configurations where each agent
is equally distant from its neighbors, and quasi α-lattices, i.e. con-
figurations that are α-lattices modulo a small error in the distances
[7]. The scale parameter d defines the ideal inter-agent distance.
Definition 2.1 (α-lattice [7]). A configuration x ∈ Rm·n is called
α-lattice if for all i ∈ B and all j ∈ Ni(x), ∥xi − xj ∥ = d, where d ∈
R+ is the scale of the α-lattice. For tolerance δ ∈ R+, a configuration
x ∈ Rm·n is called a quasi α-lattice if for all i ∈ B and all j ∈ Ni(x),
∥xi − xj ∥ − d ≤ δ.
2.1 Sensing noise
We extend the classical equations of motion, Eqs. (1)–(2), with sens-
ing noise affecting how each agent perceives positions and velocities
of its neighbors. Existing work has put little focus on flocking dy-
namics subject to noise, which is unfortunately unavoidable in
realistic natural and engineered flocks.
For actual positions x(k) and velocities v(k) at step k, let x(k)
and v(k) denote their noisy counterparts sensed by a generic agent,
defined by:
x(k) = x(k) + nx(k) and v(k) = v(k) + nv(k),
(7)
where nx(k) and nv(k) in Rm·n are vectors of independent and
identically distributed (i.i.d.) random variables. The position noise
nx(k) and velocity noise nv(k) are distributed according to Gauss-
ian distributions with mean 0 and standard deviation σx and σv,
respectively. We stress the dependency on k because noise variables
are independent across time steps.
In centralized flocking algorithms, where agent decisions are
computed by a single controller with information about the whole
population, we use Eq. 7 to define noisy measurements. In dis-
tributed algorithms, sensing noise is independent for each agent.
We denote the noisy measurement of agent i by x▷i(k) and v▷i(k),
where positions and velocities are noisy for all agents except agent
i
x▷i(k) =[ xT
v▷i(k) =[ vT
1 (k) . . . xT
1 (k) . . . vT
i (k) . . . xT
i (k) . . . vT
n (k)]T and
n (k)]T ,
(8)
(9)
a)
b)
Declarative vs Rule-based Control for Flocking Dynamics
SAC 2018, April 9-13, 2018, Pau, France
with x1(k), . . . , xn(k) and v1(k), . . . , vn(k) defined as per (7); implic-
itly, for each agent i and each other agent j, the noise distribution is
j (k) component of x▷i(k).
sampled independently to compute the xT
2.2 Reynolds' rule-based model
In Reynolds' rule-based distributed model [12, 13], agents follow
simple rules to compute their accelerations from the positions and
velocities of their neighbors. The rules are illustrated in Figure 2.
They do not explicitly specify the desired flocking formation as an
objective; rather, flocking emerges from the interaction rules.
Specifically, each agent i ∈ B updates its acceleration ai(k) at
step k by considering the following three components (adapted to
include sensing noise):
(1) Alignment: agents match their velocities with the average
velocity of nearby agents.
aal
i (k) = wal ·'›«'›«
i (k) = wc ·'›«'›«
ac
as
i (k) = ws ·
(2) Cohesion: agents move towards the centroid of the agents in
the local neighborhood.
(3) Separation: agents move away from nearby neighbors.
1
Ni(x▷i(k)) · j∈Ni(x▷i(k))
1
Ni(x▷i(k)) · j∈Ni(x▷i(k))
j∈Ni(x▷i(k))
Ni(x▷i(k)) ·'›«
1
(10)
(11)
(12)
vj(k)"fi‹ − vi(k)"fi‹
xj(k)"fi‹ − xi(k)"fi‹
∥xi(k) − xj(k)∥2"fi‹
xi(k) − xj(k)
The cohesion and alignment rules help form and maintain a closely
packed, flock-like formation. The separation rule prevents agents
from coming too close to each other, thus reducing crowding and
collisions.
Non-negative constants wal , wc and ws are the weights for each
acceleration component. Typically, a smaller interaction radius
(hence a smaller neighborhood) is used for the separation rule,
because it is significant only when agents are very close to each
other. The overall acceleration in Reynolds' model is given by:
ai(k) = aal
i (k) + ac
i (k) + as
i (k).
(13)
2.3 Olfati-Saber's potential-based model
In potential-based flocking models, the interaction between a pair
of agents is modeled by a potential field. It is assumed that an agent
is a point source, and it has a potential field around it, which ex-
erts a force, equal to its gradient, on other agents in its range of
influence. The potential field has circular symmetry and hence is a
function of distance from the source. In the work of Olfati-Saber [7],
the potential function ψα for a pair of agents has its minimum at
the desired inter-agent distance d of the desired α-lattice. Outside
the interaction radius r, the potential function is constant, so the
potential field exerts no force. The exact definition of ψα is com-
plicated: it is the definite integral of an "action function" ϕα that
is the product of a "bump function" ρh and an uneven sigmoidal
function ϕ. The control law computes an agent's acceleration based
on the sum of the forces from all other agents in its neighborhood
and a velocity alignment term.
2.4 MPC-based models
Model predictive control (MPC) [2] is a well-established control
technique that works as follows: at each time step k, it computes
the optimal control sequence (agents' accelerations in our case)
that minimizes a given cost function with respect to a predictive
model of the controlled system and a finite prediction horizon of
length T , i.e., from step k + 1 to k + T . Then, the first control input
of the optimal sequence is applied (the remainder of the sequence
is unused), and the algorithm proceeds with a new iteration.
Two main kinds of MPC-based flocking models exist, centralized
and distributed. Centralized models assume that information about
positions and velocities of all agents is available to compute their
optimal accelerations. Formally, at each time step k, it solves the
following optimization problem:
min
a(k k), ...,a(k +T−1k)∈A
J(k) + λ ·
T−1t =0 ∥a(k + t k)∥2
(14)
where a(k + t k) is the control input (accelerations) for all agents
at predicted time step k + t starting from step k. The first term J(k)
is the primary model-specific cost function that the controller seeks
to optimize within the prediction horizon; it is implicitly a function
of the predicted configurations during the prediction horizon for
time step k. The second term is standard for MPC problems and
penalizes large control inputs, with weight λ > 0.
In distributed flocking models, each agent computes its optimal
acceleration based only on information about its neighbors. Each
agent i solves an optimization problem of the form:
min
ai(k k), ...,ai(k +T−1k)∈A
Ji(k) + λ ·
T−1t =0 ∥ai(k + t k)∥2
(15)
where ai(k + t k) is the acceleration for agent i at predicted time
step k + t starting from step k, and Ji(k) is the model-specific cost
function for agent i. In distributed MPC, an agent has no way to
know current or future control decisions of its neighbors, which
are needed to make accurate predictions about their behavior. To
address this problem, some approaches allow agents to communi-
cate their local control decisions or future positions (e.g. [16, 18]),
or assume that neighbors follow some default motion law, e.g.,
they move with constant velocities. We adopt the second strategy,
because it does not require any communication.
The majority of existing MPC-based approaches to flocking are
designed to optimize the regularity of the flock, by penalizing con-
figurations where neighboring agents are not exactly distance d
apart, i.e., configurations that differ from an α-lattice [15–18]. We
call these approaches lattice-based MPC. Next we describe repre-
sentative centralized and distributed lattice-based MPC flocking
models, which we extend to account for sensing noise. The cen-
tralized model is a variant of a model by Zhan and Li [15, 16]. The
distributed model is by Zhang et al. [17].
2.4.1 Centralized lattice-based MPC flocking. The centralized
lattice-based MPC problem is defined as:
min
a(k k), ...,a(k +T−1k)∈A
T
t =1 ∥д (x(k + t k)) ∥2+λ·
T−1t =0 ∥a(k+t k)∥2
(16)
SAC 2018, April 9-13, 2018, Pau, France
U. Mehmood et al.
(a) Alignment
(b) Cohesion
(c) Separation
Figure 2: Interaction rules for flocking behavior in Reynolds Model.
where x(k + t k) is the configuration of the system at predicted
time step k + t starting from step k, following the dynamics:
xi(k k) = xi(k) vi(k k) = vi(k)
xi(k + t + 1 k) = xi(k + t k) + dT · vi(k + t k)
vi(k + t + 1 k) = vi(k + t k) + dT · ai(k + t k),
where the initial state of the prediction window is given by noisy
measurements. For configuration x, д(x) captures the α-lattice ir-
regularity as the total deviation between agent distances and d:
∥д(x)∥2 = (i, j)∈E(x)(cid:13)(cid:13)(cid:13)(cid:13)xji −
d · xji
∥xji ∥(cid:13)(cid:13)(cid:13)(cid:13)
2
, with xji = xj − xi .
(17)
This model is inspired by [15] and [16] but differs from both:
in [15] the cost function also contains a velocity alignment term,
which the same authors removed in their subsequent work, while
in [16], "impulsive MPC" is used, which means that agents directly
control their velocities (instead of accelerations), an abstraction
that allows physically unrealizable accelerations.
2.4.2 Distributed lattice-based MPC flocking. In the distributed
MPC flocking model of Zhang et al. [17], each agent i controls its
acceleration based on position and velocity measurements of the
neighbors and assumes they have constant velocity (zero accelera-
tion) during the prediction horizon. Similarly, the set of neighbors
of i is assumed invariant during the prediction horizon, and we
denote it by Ni(k) = Ni(x(k)). The control law for agent i is:
3 DECLARATIVE FLOCKING
This section introduces centralized and distributed versions of our
Declarative Flocking (DF) model, and presents a flocking algorithm
based on MPC. Our formulation is declarative in that it consists of
just two simple terms: (1) a cohesion term based on the average
squared distance between pairs of agents, to keep the flock together,
and (2) a separation term based on the inverse squared distances
between pairs of agents, to avoid crowding. These two terms repre-
sent opposing forces on agents, causing agents to move towards
positions in which these forces are balanced. Unlike the majority
of existing MPC-based approaches that are designed to optimize
conformance to an α-lattice, our design does not impose a specific
geometric structure.
3.1 Centralized DF model
The cost function J for our centralized DF model contains the two
terms described above, with the cohesion term considering all pairs
of agents, and the separation term considering only pairs of agents
that are neighbors. The weight ω of the separation term provides
control over the density of the flock.
JC (x) =
The control law is Eq. (14) with J(k) equal toT
B · (B − 1) ·i∈B j∈B,i <j
1
∥xij ∥2
t =1 JC (x(k + t k)).
∥xij ∥2 + ω · (i, j)∈E(x)
3.2 Distributed DF model
The cost function J for our distributed DF model is similar to the
centralized one, except that both terms are limited to consider pairs
of agents that are neighbors.
2
1
Ni(k) · j∈Ni(k)
∥xij ∥2 + ω · j∈Ni(k)
1
∥xij ∥2
(20)
T
i (x(k + t k)).
The control law for agent i is Eq. (15) with Ji(k) equal to
t =1 JD
4 MEASURES OF FLOCKING PERFORMANCE
We introduce four key measures of flocking performance. A single
measure is insufficient, because flocking is indeed characterized by
multiple desirable properties, such as aligned velocities and cohe-
sion. Olfati-Saber introduces four main properties for flocking [7],
informally described as:
(1) the group of agents stays connected in a unique flock, i.e., no
sub-flocks and fragmentation should emerge;
(2) the group remains cohesive, in a close-knit formation;
min
ai(k k), ...,ai(k +T−1k)∈A
T
t =1 ∥дi (x(k + t k)) ∥2+
T−1t =0 ∥ai(k + t k)∥2.
λ ·
(18)
JD
i (x) =
where the predicted future dynamics of i is determined by:
xi(k k) = xi(k) vi(k k) = vi(k)
xi(k + t + 1 k) = xi(k + t k) + dT · vi(k + t k)
xi(k + t + 1 k) = xi(k + t k) + dT · ai(k + t k),
while i's neighbors j ∈ Ni(k) have constant velocity:
xj(k k) = xj(k) xj(k + t + 1 k) = xj(k + t k) + dT · vj(k).
For configuration x, дi(x) is defined in a similar way to Eq. (17)
and quantifies how much i's neighborhood Ni(k) deviates from an
α-lattice:
(19)
d · xji
2
.
∥дi(x)∥2 = j∈Ni(k)(cid:13)(cid:13)(cid:13)(cid:13)xji −
∥xji ∥(cid:13)(cid:13)(cid:13)(cid:13)
Declarative vs Rule-based Control for Flocking Dynamics
SAC 2018, April 9-13, 2018, Pau, France
(3) the group moves in a coherent way as if it was a unique body,
i.e., agents' velocities are aligned; and
(4) the group maintains a regular geometry (in the α-lattice
sense).
We introduce the following four measures to capture these four
requirements. An important concept in these definitions is a sub-
flock, which is a set of interacting agents that is too far apart from
other agents to interact with them. Formally, a sub-flock in a config-
uration x corresponds to a connected component of the proximity
net G(x). Let CC(x) ⊆ 2B be the set of connected components of
the proximity net G(x).
(1) The number of connected components of the proximity net
quantifies connectedness-or, equivalently, fragmentation-of the
flock. There is no fragmentation when CC(x) = 1. Fragmentation
exists when CC(x) > 1. Fragmentation may be temporary or, if
sub-flocks move in different directions, permanent.
(2) The maximum component diameter, denoted D(x), quantifies
cohesion. It is defined by
(21)
(22)
where D(x, B′) is the diameter of connected component B′:
D(x) = max
B′∈CC(x)
D(x, B′)
D(x, B′) = max
(i, j)∈B′×B′
i,j
∥xij ∥.
Note that when all agents are isolated, i.e., CC(x) =i∈B {{i}},
D(x) = −∞ because the domain of the max function in Equation 22
is empty when B′ is a singleton. Note that we consider the maxi-
mum diameter of a sub-flock in order to make this measure more
independent of connectedness. If we instead considered the overall
diameter of the entire (possibly fragmented) flock, any flocking
model that did poorly on connectedness would also do very poorly
on this measure.
(3) The velocity convergence measure, adopted from [17], quanti-
fies the average discrepancy between each agent's velocity and the
average velocity of the flock. In particular, we extend the measure
of [17] to average velocity convergence values across sub-flocks:
VC(x, v) =B′∈CC(x)(cid:13)(cid:13)(cid:13)i∈B′ vi −(cid:16)j∈B′ vj
CC(x)
B′
(cid:17)(cid:13)(cid:13)(cid:13)2(cid:30)B′
(23)
(4) To measure the regularity of the geometric structure of a flock,
as reflected in the inter-agent spacing, we introduce a parameter-
free and model-independent irregularity measure I(x). For a con-
nected component (sub-flock) B′, it is defined as the sample stan-
dard deviation of the distances between each agent in B′ and its
closest neighbor. Thus, the measure penalizes configurations where
there is dispersion in inter-agent distances, while not imposing any
fixed distance between them (unlike α-lattices).
Let CC′(x) = CC(x) \i∈B {{i}} be the set of connected com-
ponents where isolated agents are excluded. For CC′(x) = 0 (or
equivalently, CC(x) = B), i.e., all agents are isolated, we set the
irregularity I(x) = 0, which is the optimal value. This reflects the
fact that a single point is a regular structure on its own. Moreover,
such a configuration is already highly penalized by CC(x) and
VC(v). For CC′(x) > 0, the measure is defined by:
I(x) =B′∈CC′ σ(cid:0)i∈B′ minj,i ∥xij ∥(cid:1)
CC′
.
(24)
and is the sum operator (or disjoint union) for multisets.
where σ(S) is the standard deviation of the multiset of samples S
An α-lattice (see Def. 2.1) has the optimal value of I(x), i.e.,
I(x) = 0, since all neighboring agents are located at the same
distance d from each other, leading to zero standard deviation for the
term σ ({d, d, . . . , d}). This shows that I(x) captures the regularity
underlying the concept of α-lattice.
We introduce this measure because previous measures of regu-
larity or irregularity, such as those in [7, 16, 17], measure deviations
from an α-lattice with a specified inter-agent distance d and are
therefore inapplicable to flocking models, such as Reynolds' model
and our DF models, that are not based on α-lattices and do not
have a specified target inter-agent distance. Also, our irregularity
measure is more flexible than those based on α-lattices, because it
gives an optimal score to some configurations that are geometrically
regular but not α-lattices. For example, consider a configuration x
in which the agents are on the vertices of a grid with edge length e,
and the interaction radius is equal to the length of the diagonal of
a box in the grid. This configuration has an optimal value for our
irregularity measure, i.e., I(x) = 0, because the distance from every
agent to its nearest neighbor is e. This configuration is not an α-
lattice and hence does not nave an optimal value for the irregularity
measures used in prior work.
5 PERFORMANCE EVALUATION
We compare the performance of the models of Section 2 with the
newly introduced DF flocking models in the 2-dimensional setting.
In the first set of experiments (Section 5.1), we evaluate the per-
formance measures illustrated in Section 4. In the second set of
experiments (Section 5.2), we analyze the resilience of the algo-
rithms to sensor noise.
For consistency with the experimental settings of [17], the lattice-
based MPC problems are solved using the interior point method im-
plemented in MATLAB's fmincon function. Our DF-MPC problems
are solved using gradient descent optimization. Unless otherwise
specified, the population size is n = 30, the simulation length is
100, dT = 0.3, ¯v = 8, ¯a = 1, r = 8.4, d = 7, T = 3, and λ = 1.
These parameter values are the same ones reported in [17]. Follow-
ing the settings in the OpenSteer project [11], the parameters for
Reynolds' model are rc = 9, rs = 5, ral = 7.5, wc = 8, ws = 12,
and wal = 8. The weight of the separation term in our centralized
and distributed DF-MPC is ω = 50. As in [17], initial positions and
initial velocities of agents are uniformly sampled from [−15, 15]2
and [0, 2]2, respectively.
5.1 Performance Comparison of Flocking
Algorithms
Fig. 3 shows examples of final formations for all flocking models.
In particular, we chose configurations where fragmentation did
not occur. We observe that the formations for lattice-based MPC
algorithms have spread-out, rigid structures, consistent with the
SAC 2018, April 9-13, 2018, Pau, France
U. Mehmood et al.
design objective of maximizing the α-lattice regularity. On the other
hand, Reynolds and our DF MPC models result in more natural
flock shapes.
In Fig. 4, we compare the performance measures averaged over
100 runs for each flocking model. Regarding the number of con-
nected components (sub-flocks), our centralized DF-MPC registers
the best behavior, rapidly stabilizing to an average of 1 component
(see plot a). Our distributed DF-MPC and Reynolds' model have
comparable performance, reaching an average number of sub-flocks
below 1.4. The lattice-based MPCs and Olfati-Saber instead lead
to constant fragmentation, with more than 2 sub-flocks for the
distributed lattice-based MPC, 6 for the centralized lattice-based
MPC, and more than 8 for Olfati-Saber's model.
This ranking is confirmed by the diameter measure (plot b),
where our centralized and distributed DF-MPC and Reynolds' model
show the best cohesion, outperforming the lattice-based approaches.
Recall that this measure indicates the maximum diameter over all
sub-flocks, not the diameter of the entire population. As a conse-
quence, fragmentation tends to improve diameter values since it
produces sub-flocks with fewer individuals. This explains why our
distributed DF-MPC performs better on this measure than the cen-
tralized version, and similarly why Olfati-Saber's model has smaller
diameter measure than centralized lattice-based MPC, which in turn
has smaller diameter measure than the distributed variant.
As expected, Olfati-Saber's model and the lattice-based MPCs
have very good performance for irregularity (plot c), since they are
designed to achieve the regular geometric formation of α-lattice.
Surprisingly, our distributed DF-MPC performs almost as well as
them on this measure. Centralized DF-MPC and Reynolds' model
have the least regular formations.
For velocity convergence (plot d), we find that all models perform
comparably well and are able to achieve flocks with consistent
velocities fairly quickly after an initial spike.
5.2 Robustness to Sensing Noise
To evaluate the resiliency of the models to sensor noise, we per-
formed 20 runs for each model at 10 noise levels. The noise levels
are numbered from 1 to 10, and noise level i has σx = 0.2i and
σv = 0.1i. For each performance metric, we averaged its final val-
ues over 20 runs for each noise level. The results are plotted in Fig. 5.
Of the six models, Olfati-Saber's model is the most vulnerable to
sensing noise: the number of sub-flocks CC in Olfati-Saber's model
quickly increases to nearly 30, rendering other metrics irrelevant.
The lattice-based MPC models also exhibit high fragmentation, lead-
ing to nominally good but largely irrelevant values for the other
performance metrics. Our distributed DF-MPC and Reynolds' model
have the best resiliency to sensing noise, with both models exhibit-
ing similar profiles in all metrics. While the irregularity and velocity
convergence measures increase with noise level, as expected, both
models remarkably maintain almost a single connected component
with a nearly constant component diameter for all 10 noise levels,
with DF-MPC achieving a smaller diameter than Reynolds' model.
6 RELATED WORK
Reynolds [12] introduced the first rule-based approach for simula-
tion of flocking behavior. With only three simple rules, his model is
able to capture complex flocking behaviors of animals. Additional
rules can be added to the model to simulate specific behaviors, such
as leader following and predator avoidance. Pearce et al. [8] present
a rule-based strategy for flocking, where agents move to maximize
their view out of the flock. Cucker and Dong [3] present a rule-
based flocking approach with proofs of convergence and collision
avoidance.
Cucker and Smale [4] introduced another popular rule-based
flocking model. The Cucker-Smale model is parameterized by a
constant β such that if β < 1/2, velocity convergence is guaranteed.
If β ≥ 1/2, velocity convergence can also be achieved under some
conditions on the initial positions and initial velocities of the agents.
Ahn and Ha [1] investigated the effects of multiplicative noise on
the long term dynamics of the Cucker-Smale model. Erban et al. [5]
extend the Cucker-Smale model to take into account stochasticity
(imperfections) of agent behavior and delay in agents' responses to
changes in their environment.
Flocking models based on potential fields have been proposed
in several papers. Tanner et al. [14] propose a potential function
Uij, given in Equation 25, where rij2 is the distance between
agents i and j. For distances greater than R, the potential is set to a
constant value, C, indicating a zero force. In their control law, the
acceleration of agent i is based on the sum over all neighbors j of
the gradient of the potential function Uij.
ri j 2 + loдrij2,
(25)
Uij =(
A similar potential function is also proposed by [10]. Furthermore,
potential-based solutions have been extended with additional behav-
iors such as obstacle avoidance and leader following. For example,
Ogren et.al. [9] use the motion of the leader to guide the motion of
the flock; the leader's motion is independent, i.e., is not influenced
by other agents.
La and Sheng [6] propose an extension of Olfati-Saber's model
designed for noisy environments. In addition to the terms found in
Olfati-Saber's model, their control law contains feedback terms for
position and velocity, to make agents tend to stay close to the cen-
troid of their neighborhood and minimizing the velocity mismatch
with their neighbors. They show that adding these feedback terms
to the control law helps bound the error dynamics of the system.
1
C,
rij2 < R
rij2 ≥ R
7 CONCLUSIONS
This paper presents an abstract declarative form of control for
flocking behavior and the results of a thorough comparison of
centralized and distributed versions of our MPC-based declarative
flocking with four other flocking models. Our simulation results
demonstrate that DF-MPC yields the best cohesion and least frag-
mentation, and produces natural flock shapes like those produced
by Reynolds' rule-based model. Our resiliency analysis shows that
the distributed version of our DF-MPC is highly robust to sensor
noise.
As future work, we plan to study resilience of the flocking models
with respect to additional noisy scenarios such as actuation noise
(i.e., noise affecting acceleration) and faulty agents with deviant
behavior. We also plan to investigate smoothing techniques to
increase resilience to sensor noise.
Declarative vs Rule-based Control for Flocking Dynamics
SAC 2018, April 9-13, 2018, Pau, France
(a) Reynolds
(b) Lattice-based central-
ized MPC
Lattice-based
(c)
tributed MPC
dis-
(d) DF centralized MPC
(e) DF distributed MPC
(f) Olfati-Saber
Figure 3: Examples of final formations for different flocking models. The red dots are the agent positions. The blue lines
denote the agent velocities; the line lengths are proportional to the speeds.
(a) Number of connected components
CC
(b) Max component diameter D
(c) Irregularity I
(d) Velocity convergence V C
Figure 4: Comparison of performance measures obtained with 100 runs for each flocking algorithm.
(a) Number of connected components
CC
Figure 5: Comparison of the final values of the performance measures obtained with 20 runs for each flocking algorithm and
for each noise level.
(b) Max component diameter D
(d) Velocity convergence V C
(c) Irregularity I
REFERENCES
[1] Shin Mi Ahn and Seung-Yeal Ha. 2010.
ics of the CuckerâĂŞSmale model with multiplicative white noises.
Math. Phys. 51, 10 (2010), 103301.
arXiv:http://dx.doi.org/10.1063/1.3496895
Stochastic flocking dynam-
J.
https://doi.org/10.1063/1.3496895
[2] E.F Camacho and C. Bordons. 2007. Model predictive control. Springer.
[3] Felipe Cucker and Jiu-Gang Dong. 2011. A general collision-avoiding flocking
framework. IEEE Trans. Automat. Control 56, 5 (2011), 1124–1129.
[4] F. Cucker and S. Smale. 2007. Emergent Behavior in Flocks. IEEE Trans. Automat.
Control 52, 5 (May 2007), 852–862. https://doi.org/10.1109/TAC.2007.895842
[5] Radek Erban, Jan Hasźkovec, and Yongzheng Sun. 2016. A Cucker–Smale Model
with Noise and Delay. SIAM J. Appl. Math. 76, 4 (July 2016), 1535–1557. https:
//doi.org/10.1137/15M1030467
[6] H. M. La and W. Sheng. 2010. Flocking control of multiple agents in noisy
environments. In 2010 IEEE International Conference on Robotics and Automation.
4964–4969. https://doi.org/10.1109/ROBOT.2010.5509668
[7] Reza Olfati-Saber. 2006. Flocking for multi-agent dynamic systems: Algorithms
and theory. IEEE Transactions on automatic control 51, 3 (2006), 401–420.
[8] Daniel J. G. Pearce, Adam M. Miller, George Rowlands, and Matthew S. Turner.
2014. Role of projection in the control of bird flocks. Proceedings of the National
Academy of Sciences 111, 29 (2014), 10422–10426. https://doi.org/10.1073/pnas.
1402202111 arXiv:http://www.pnas.org/content/111/29/10422.full.pdf
[9] Naomi Ehrich Leonard Peter Ogren. 2004. Cooperative control of mobile sen-
sor networks:Adaptive gradient climbing in a distributed environment. IEEE
transactions on Automatic Control 49, 8 (2004).
[10] John H. Reif and Hongyan Wang. 1999. Social potential fields: A distributed
behavioral control for autonomous robots. Robotics and Autonomous Systems 27,
3 (1999), 171 – 194. https://doi.org/10.1016/S0921-8890(99)00004-4
[11] Craig Reynolds. 2004. OpenSteer, steering behaviors for autonomous characters.
(2004). http://opensteer.sourceforge.net/
[12] Craig W. Reynolds. 1987. Flocks, Herds and Schools: A Distributed Behavioral
Model. SIGGRAPH Comput. Graph. 21, 4 (Aug. 1987), 25–34. https://doi.org/10.
1145/37402.37406
[13] Craig W. Reynolds. 1999. Steering Behaviors For Autonomous Characters. In
Proceedings of Game Developers Conference 1999. 763–782.
75
70
65
60
55
50
190
200
210
80
60
40
20
0
-20
0
20
40
60
140
130
120
110
100
90
100
120
140
135
130
125
120
115
110
90
100
110
125
120
115
110
105
100
80
90
100
60
40
20
0
-20
0
20
40
60
80
Lattice Distributed
Reynolds
DF Centralized
DF Distributed
Lattice Centralized
O-S
0
10
20
30
40
50
Time
60
70
80
90
10
02468
CC
20
40
60
80
Time
10
5
CC
0
0
D
60
50
40
30
20
10
0
20
40
60
80
Time
2
1.5
I
1
0.5
0
0
20
40
60
80
Time
0
20
40
60
80
Time
4
3
2
1
0
VC
Lattice Distributed
Reynolds
DF Centralized
DF Distributed
Lattice Centralized
O-S
0
10
20
30
40
50
Time
60
70
80
90
10
02468
CC
2
6
4
Noise Level
8
10
30
20
CC
10
0
D
50
40
30
20
10
0
2
4
6
8
10
Noise Level
I
1.5
1
0.5
0
2
4
6
8
10
Noise Level
2
4
6
8
10
Noise Level
1.5
1
0.5
0
VC
SAC 2018, April 9-13, 2018, Pau, France
U. Mehmood et al.
[14] H. G. Tanner, A. Jadbabaie, and G. J. Pappas. 2003. Stable flocking of mobile
agents part I: dynamic topology. In 42nd IEEE International Conference on Decision
and Control (IEEE Cat. No.03CH37475), Vol. 2. 2016–2021 Vol.2.
[15] Jingyuan Zhan and Xiang Li. 2011. Flocking of discrete-time multi-agent systems
with predictive mechanisms. IFAC Proceedings Volumes 44, 1 (2011), 5669–5674.
[16] Jingyuan Zhan and Xiang Li. 2013. Flocking of multi-agent systems via model
predictive control based on position-only measurements. IEEE Transactions on
Industrial Informatics 9, 1 (2013), 377–385.
[17] Hai-Tao Zhang, Zhaomeng Cheng, Guanrong Chen, and Chunguang Li. 2015.
Model predictive flocking control for second-order multi-agent systems with
input constraints. IEEE Transactions on Circuits and Systems I: Regular Papers 62,
6 (2015), 1599–1606.
[18] Lifeng Zhou and Shaoyuan Li. 2017. Distributed model predictive control for
multi-agent flocking via neighbor screening optimization. International Journal
of Robust and Nonlinear Control 27, 9 (2017), 1690–1705.
|
1702.00785 | 3 | 1702 | 2017-03-28T03:28:28 | Evaluation of Automated Vehicles Encountering Pedestrians at Unsignalized Crossings | [
"cs.MA",
"cs.RO",
"eess.SY"
] | Interactions between vehicles and pedestrians have always been a major problem in traffic safety. Experienced human drivers are able to analyze the environment and choose driving strategies that will help them avoid crashes. What is not yet clear, however, is how automated vehicles will interact with pedestrians. This paper proposes a new method for evaluating the safety and feasibility of the driving strategy of automated vehicles when encountering unsignalized crossings. MobilEye sensors installed on buses in Ann Arbor, Michigan, collected data on 2,973 valid crossing events. A stochastic interaction model was then created using a multivariate Gaussian mixture model. This model allowed us to simulate the movements of pedestrians reacting to an oncoming vehicle when approaching unsignalized crossings, and to evaluate the passing strategies of automated vehicles. A simulation was then conducted to demonstrate the evaluation procedure. | cs.MA | cs | Evaluation of Automated Vehicles Encountering Pedestrians at
Unsignalized Crossings
Baiming Chen1, Ding Zhao2, Huei Peng3
7
1
0
2
r
a
M
8
2
]
A
M
.
s
c
[
3
v
5
8
7
0
0
.
2
0
7
1
:
v
i
X
r
a
Abstract- Interactions between vehicles and pedestrians
have always been a major problem in traffic safety. Experienced
human drivers are able to analyze the environment and choose
driving strategies that will help them avoid crashes. What is not
yet clear, however, is how automated vehicles will interact with
pedestrians. This paper proposes a new method for evaluating
the safety and feasibility of the driving strategy of automated
vehicles when encountering unsignalized crossings. MobilEye®
sensors installed on buses in Ann Arbor, Michigan, collected
data on 2,973 valid crossing events. A stochastic interaction
model was then created using a multivariate Gaussian mixture
model. This model allowed us to simulate the movements of
pedestrians reacting to an oncoming vehicle when approaching
unsignalized crossings, and to evaluate the passing strategies
of automated vehicles. A simulation was then conducted to
demonstrate the evaluation procedure.
I. INTRODUCTION
Traffic safety has become an issue of growing concern in
the U.S. According to the NHTSA, the country lost 35,092
people in crashes on U.S. roadways during 2015, up 7.2%
from 32,744 in 2014, the largest increase in nearly 50 years
[1]. One approach to decreasing traffic deaths is to use
automated vehicles, which do not suffer from challenges such
as fatigue, distraction and drunk driving that humans might
face. Researches have been conducted on the interaction
between vehicles and vehicles [2], [3], [4]. What is not
yet known, however, is exactly how automated vehicles will
interact with pedestrians. Ways to test automated vehicles
must be found before the vehicles can be put into use. The
most critical places with a high concentration of interac-
tions are intersections, particularly unsignalized crossings,
where the right of way is not clear. As it is difficult for
automated vehicles to decide the appropriate strategy for
passing through an unsignalized intersection with pedestrians
crossing the street, this scenario is a good one to consider.
The first stage of the evaluation requires an analysis of
pedestrian behaviors. Early studies on pedestrian movement
focused on pedestrians that did not interact with vehicles.
Three typical kinds of pedestrian models were developed:
discrete cellular automata models [5], continuous force-based
models [6], and macroscopic pedestrian stream models [7].
* This work is funded by the Mobility Transformation Center Denso
Tailor Project at the University of Michigan with grant No. N020210.
1B. Chen is with the Department of Automotive Engineering, Tsinghua
University, Beijing, China, 100084, and is currently a visiting scholar at the
University of Michigan, Ann Arbor, MI 48109, USA.
2D. Zhao is with the University of Michigan Transportation Re-
search Institute, Ann Arbor, MI 48109, USA (correspondig author:
[email protected]).
3H. Peng is with the Department of Mechanical Engineering, University
of Michigan, Ann Arbor, MI 48109, USA.
Additional attributes of pedestrian behavior were studied
when vehicles were introduced into the picture. For example,
the walking speed of pedestrians might vary depending
on several factors, one of which being where pedestrians
crossed the street. Thus, Bennett et al. [8] found that the
average speed of those crossing at mid-block was slower
than those crossing at a signalized intersection. Additional
factors significantly contributing to pedestrian walking speed
(i.e., age, gender, group size and street width) were found
by Tarawneh [9] in Jordan. Pedestrians 21-30 years of age
were found to be the fastest; those over 65, the slowest. The
author also found that pedestrians tended to walk faster after
longer wait times. Whether pedestrians were in a group or
alone also had an impact. Yagil [10] found, for instance, that
pedestrians were more likely to wait at an intersection if a
group of pedestrians were already waiting there.
One approach to analyzing pedestrian behaviors is to
apply the concept of gap acceptance. When arriving at an
unsignalized intersection, pedestrians must decide whether
the gap in the traffic stream large enough for them to cross
the street safely. To describe the gap acceptance behavior
of pedestrians, probit and binary models were developed
by Sun et al. [11]. Regression analysis found the important
factors for pedestrian gap acceptance to be gap size, number
of pedestrians waiting, and pedestrian age. Schroeder [12]
improved gap acceptance models for unsignalized crossings
by incorporating vehicle dynamics, pedestrian characteris-
tics and concurrent events at the crosswalk. This model is
difficult to use for the evaluation of automated vehicles,
however, because information about the pedestrians (e.g.,
age and gender) cannot as yet be detected by the vehicles
approaching the intersection.
The driving strategy of human drivers when encounter-
ing unsignalized crossings was studied. To describe driver
yielding and pedestrian gap acceptance behaviors, empirical
logit models were developed by Schroeder et al. [13]. The
author also proposed a driver dynamic model when faced
with one crossing pedestrian. This model is introduced and
used as the driving strategy of automated vehicles in the
simulation section of this paper. A key component of the
driving strategy is the braking behavior of drivers when a
pedestrian comes out from the sidewalk to the road. For
example, Suzuki et al. [14] concluded that the timing of
the braking operation is approximately the same in terms of
TTC (Time to Collision), though the velocity of the subject
vehicles and passing pedestrians are different.
Before automated vehicles are put into use, they must be
evaluated for their potential interaction with pedestrians. The
participating in the SPMD (Safety Pilot Model Deployment)
program [16] were equipped with MobilEye® which can
distinguish pedestrians and provide their relative positions to
the instrumented vehicle. In addition, u-Box GPS/IMU unit
installed on the university buses can provide the geograph-
ical location. Vehicle velocity and yaw rate were recorded
from CAN bus. The routes of the instrumented buses and
pedestrians detected by MobilEye® are shown in Fig. 2.
Fig. 2. Routes of the instrumented buses and detected pedestrians in Ann
Arbor, Michigan. The Central Campus Transit Center (CCTC) has a high
pedestrian density. The figure is created based on Google Maps.
The unsignalized crossing chosen for collecting data is
in Ann Arbor, Michigan, near the University of Michigan's
Central Campus Transit Center (CCTC), where the pedes-
trian density is high (Fig. 2). The geographic location is
42.278415◦, −83.735580◦. A picture of this crossing path
is shown in Fig. 3, where 2,973 valid passing events were
collected.
driving strategy of automated vehicles should be neither too
aggressive (endangering traffic safety) nor too conservative
(leading to a waste of time and traffic congestion). Current
experimental approaches are limited, however. The European
New Car Assessment Programme (Euro NCAP) established
a test procedure for AEB VRU systems,
the automatic
emergency braking systems that specifically look for and
react to pedestrians [15]. The purpose of that test differs
from the one in this paper, however. Our goal is to simulate
the movement of multiple pedestrians and test the automated
vehicles' passing strategy, while the procedure proposed
by Euro NCAP was designed to test the effectiveness of
the braking system with the sudden appearance of only
one pedestrian. More importantly, if an automated vehicle's
driving strategy is too conservative, it may still pass the
NCAP test, but leads to traffic congestion and road rage in
reality.
The procedures used in this paper are shown in Fig. 1.
A stochastic interaction model was first developed based on
naturalistic traffic data. Then a method for evaluating the
automated vehicles was proposed. Finally, the test procedure
was demonstrated in simulation.
Fig. 1. Procedures for designing the evaluation method
The main contributions of this paper are:
• Naturalistic data of 2,973 passing events encountering
pedestrians at unsignalized crossings were recorded by
in-car sensors. To the best of our knowledge, this is
the largest data set describing vehicle-pedestrian inter-
actions that has ever been collected.
• A stochastic interaction model was created by fitting a
bounded multivariate Gaussian mixture model to traffic
data.
• A new method for evaluating automated vehicles at
unsignalized crossings was proposed.
II. DATA COLLECTION
Fig. 3. Crossing path environment (based on Google Maps)
The first step in being able to simulate the movements
of pedestrians and vehicles was to collect naturalistic data.
In this study, data on 2,973 passing events encountering
pedestrians at unsignalized crossings were collected by in-
car sensors. The devices used to collect pedestrian behavior
were MobilEye® installed on university buses in Ann Arbor,
Michigan. The university's 12 bus routes carry approximately
7.2 million passengers per year. Three of the university buses
The GPS system recorded the latitude and longitude of
the instrument vehicles, while the MobilEye® recorded the
relative positions of pedestrians. Using the data collected,
each passing event was drawn in a Cartesian Coordinate
system, so that the trajectories of the vehicles and pedestrians
could be observed. One example is shown in Fig. 4. The color
bar indicates the time of the samples.
instrumented vehicle directly, T −1
and vp can be calculated by the lateral distance (L) data.
B. Algorithms
Adv can be calculated by (1),
The probability density function (p.d.f.) of a Gaussian
mixture model is
K(cid:88)
f (y; Θ) =
πkfk(y; θk),
(2)
Fig. 4. Sample of passing events (based on Google Maps)
III. INTERACTION MODEL
A multivariate Gaussian mixture model is created based
on naturalistic data to simulate the movement of vehicles and
pedestrians when encountering unsignalized crossings. This
stochastic model demonstrates how vehicles and pedestrians
will interact with each other. This section shows the proce-
dures used to develop this model and how to use it for the
simulation.
A. Variables
Appropriate variables must first be selected to develop
a multivariate Gaussian mixture model. We simplified the
passing scenario by assuming that the trajectories of the
vehicle and the pedestrian are two perpendicular straight
lines. Then, longitudinal distance (R) and lateral distance
(L) are used to describe the relative position of a pedestrian
(Fig. 5).
Fig. 5. Definition of R (longitudinal distance) and L (lateral distance)
Four variables are required to define the state of the
passing scenario: the reciprocal of the longitudinal distance
(R−1), vehicle speed (v), pedestrian walking speed (vp)
and the reciprocal of the Time Advantage (T −1
Adv). Time
Advantage is an indicator used to describe situations where
two road users pass a common spatial zone, but at different
times, thus avoiding a collision [17]. The definition of Time
Advantage is the time between the first road user leaving
the common spatial zone and the second arriving. The
mathematical definition is:
TAdv = T T C − L/vp.
(1)
The reciprocals are used for statistical convenience. Among
the four variables, R−1 and v can be obtained by the
k=1
where πk are weights of components, fk are density func-
tions of each component, θk are parameters to decide each
component. The distribution of each component is a multi-
variate Gaussian distribution with mean µk and covariance
Σk.
Algorithms are required to fit Gaussian mixture models
to the traffic data. Here, since all the variables (i.e., R−1,
v, vp, T −1
Adv) are positive and have boundaries, a truncated
Gaussian mixture model will better fit the data. Lee et al. [18]
developed for fitting multivariate Gaussian mixture models
to data that is truncated. By applying these algorithms to the
data, parameters πk, µk and Σk of the multivariate Gaussian
mixture model can be calculated.
The number of components (K) of the multivariate Gaus-
sian mixture model will influence how well the data fit,
with a higher K helping to improve the precision of the
modeling. However, it might also lead to greater computing
time, as well as statistical errors such as singular covariance
matrices which endanger the stability of the model. Thus,
an appropriate K must be selected for the Gaussian mixture
model. The Bayesian information criterion (BIC) provides a
means for model selection. It can help estimate the quality
of the statistical models, with a model with a lower BIC
preferred. In the interaction model, this criterion was used
to determine the appropriate value of K, the BIC of the
Gaussian mixture models with different K can be seen in
Fig. 6. The BIC continues to decrease with an increase
in number of components, with little change in rate (less
than 10%) when the number of components is larger than
10. Considering the computing time and the complexity
of the models, K = 10 was selected when generating the
multivariate Gaussian mixture model.
(a) Change of BIC
(b) Change Rate of BIC
Fig. 6. Bayesian Information Criterion test of the interaction models with
different number of components (K). The BIC continues to decrease with
an increase in number of components, with little change in rate (less than
10%) when the number of components is larger than 10.
An interaction model was then created by fitting a 10-
component truncated multivariate Gaussian mixture model
to the collected naturalistic data using the algorithms above.
C. Model Display and Utilization
Since there were four variables (i.e., {R−1, v, vp, T −1
Adv})
the distribution of the
in the Gaussian mixture model,
interaction model was a 4-D function, which cannot be
shown effectively on a flat piece of paper. We projected the
function to three 2-D functions to illustrate the probability
distribution, and then compared them with the raw data
collected. The results are shown in Fig. 7.
calculated. Examples of distribution functions under different
conditions are shown in Fig. 8. We can generate a random
vp from this distribution and take it as the walking speed of
a pedestrian under the given conditions for the simulation
and evaluation.
(a) Raw data
(b) Model distribution
(c) Raw data
(d) Model distribution
(e) Raw data
(f) Model distribution
Fig. 7. Comparison of raw data and Gaussian mixture model
Fig. 8.
different conditions
Probability density functions of pedestrian walking speed under
The movement of vehicles can also be simulated using the
same method by calculating the conditional distribution. To
do that, we set ym = v and yo = {R−1, vp, T −1
Adv}, and then
the distribution function of v can be calculated. Examples of
distribution functions under different conditions are shown
in Fig. 9. It is clear that, in this model, the vehicle is likely
to have a lower speed when the distance is shorter, which
makes sense considering traffic safety. The speed with the
highest probability might be a good choice for setting the
desired speed of the vehicle under given conditions when
designing driving strategy.
The stochastic interaction model developed here can help
to simulate the walking speed of pedestrians. When encoun-
tering an unsignalized crossing, the pedestrian will decide the
appropriate walking speed to cross the street, depending on
the speed and position of the oncoming vehicle; that is, the
distribution of pedestrian walking speed (vp) is a conditional
distribution of the multivariate Gaussian mixture model. By
calculating this conditional distribution, the walking speed
of pedestrians (vp) can be simulated.
For a Gaussian distribution, if y = (yT
o )T is with
mean µ and covariance Σ, then the conditional distribution of
ymyo is also normally distributed with mean and covariance:
(3)
m, yT
µmo = µm + Σm,oΣ−1
Σmo = Σm,m − Σm,oΣ−1
o,o(yo − µo),
o,oΣo,m.
(4)
To calculate the conditional distribution of a mixture Gaus-
sian model, first calculate the conditional distribution of
each component, then normalize all components and set new
weights.
Adv}, then the
distribution function of vp under given conditions can be
If we set ym = vp and yo = {R−1, v, T −1
Fig. 9.
conditions
Probability density functions of vehicle speed under different
IV. SIMULATION
A simulation is used to show how to evaluate the driving
strategy of automated vehicles. The automated vehicle is
evaluated against the human drivers for better performance
when encountering unsignalized crossings in the simulation.
We expect that the driving strategy of automated vehicles
be neither too aggressive (endangering traffic safety) nor too
conservative (leading to traffic congestion). With an increase
in aggressiveness of the automated vehicle,
less time is
then needed to pass through the intersection, resulting in
a possible rise in the crash rate. As shown in Fig. 10, the
aggressiveness needs to be within the allowed interval when
the automated vehicle has a passing time less of than µ0 and
a denied crash rate less than κ0.
Based on this distribution, an arriving time example was
generated and is shown in Fig. 12. Five pedestrians are
scheduled to arrive at this crossing within 60 seconds.
Fig. 10.
Influence of the vehicle's aggressiveness
A. Procedures for the Evaluation Experiment
PreScan® is used for the simulation. The experiment takes
place at an unsignalized crossing (Fig. 11). At first, the vehi-
cle comes from a long distance at a constant speed v0. When
it is close, at a certain distance R0, pedestrians will start to be
generated to go across the street, and the vehicle will interact
with pedestrians and try to pass through the intersection
without any crashes. Each experiment is done twice. The
first time, the automated vehicle is tested to interact with
the pedestrians; the second time, the human driver passing
strategy is applied to interact with the same pedestrian which
is recorded as a reference for the evaluation. In this study,
the simulation is run 50 times. The passing strategies of
pedestrians, automated vehicles and human drivers simulated
in this section are described below separately. The results are
analyzed at the end of this section.
Fig. 11. Environment of the simulation. When the vehicle is at a certain
distance (R = R0), pedestrians will start to arrive at the unsignalized
crossing.
B. Pedestrian Crossing Strategy
Pedestrians are made to cross the street from both sides.
When the vehicle is at a certain distance (R = R0),
pedestrians will start to arrive at the unsignalized crossing.
Each pedestrian has an arriving time when they will be put
on the side of the street and start to cross. The arriving time
of the pedestrians obeys the Poisson process:
e(−λt).
P{N (t) = n} =
(5)
(λt)n
n!
Parameter λ can be set to different values depending on the
density of pedestrian flow, which varies greatly depending
on time and location. For reference, pedestrian flow in the
campus area during peak hours is around 250 peds/hr [13].
Fig. 12. Arriving times of pedestrians within 60 seconds
Though the pedestrians' arriving times are random, their
walking speeds are usually not. Each pedestrian has a
constant walking speed that is calculated by the interaction
model depending on the distance and speed of the oncoming
vehicle when the pedestrian arrives, which is similar to real
world process, in which most pedestrians will look at the
oncoming vehicle and decide what to do before moving.
C. Automated Vehicle's Passing Strategy
The driving strategy evaluated in the simulation is the
Soft-Yield model proposed by Schroeder et al. [13] which
provides vehicle trajectories when facing one pedestrian.
The model was developed based on GPS data from the
instrumented vehicle study. The generalized vehicle distance-
speed and time-speed models are shown in Fig. 13. When
approaching the unsignalized crossing,
the vehicle has a
constant deceleration for time length T1 and then starts to
coast to the crosswalk.
Fig. 13. Driving strategy of the automated vehicle when encountering an
unsignalized crossing (developed by Schroeder et al. [13])
Based on a regression analysis, vehicle acceleration a is
set to
a = p1 + p2v + p3R,
(6)
where p1 = 0.0169, p2 = −0.13986, p3 = 0.010115. v and
R are the vehicle's information when the vehicle perceives
the pedestrian and makes its decision.
Deceleration time T1 is calculated as follows:
L0
vp
,
tL =
(cid:114)
L − 2(R − v ∗ tL)
t2
a
(7)
(8)
,
T1 = tL −
where L0 is the length of the crossing path, and tL is the
time needed for the pedestrian to complete the crossing.
D. Human Driver's Passing Strategy
A human driver's passing strategy was developed by
studying the interaction model based on the naturalistic
traffic data. When encountering an unsignalized crossing,
the strategy of the simulated human driver is set to adjust
the vehicle's speed if a pedestrian is detected. The desired
speed vd for adjusting is calculated based on the position
and velocity of the pedestrian in the current state. Thus, the
desired acceleration of the vehicle is
vd − v
∆t
ad =
(9)
,
where ∆t is the time interval between two updates of the
desired speed. In this study, we set ∆t = 1 s.
Each vehicle, of course, has a maximum limit of acceler-
ation am. Thus, if ad > am, the acceleration of the vehicle
a will be set to a = am. Otherwise, the acceleration will be
set to just a = ad.
When no pedestrian is present within the radar's detection
range, or the pedestrians have left the crossing, the vehicle
will accelerate to v0 at a constant speed of a0 = 1 m/s.
E. Simulation Results
All the above procedures are simulated using PreScan®.
The parameters set for the simulation is shown in Table I.
One pedestrian is generated in each experiment.
TABLE I
PARAMETERS IN THE SIMULATION
symbol
R0
L0
v0
unit
m
m
m/s
value
30
9
5
Each experiment is done twice. The first time, the au-
tomated vehicle's strategy is tested to interact with the
pedestrians generated based on the Poisson process;
the
second time, the human driver's strategy is tested to interact
with the same pedestrian. The time used and the safety of
the pedestrian is recorded. We evaluate the aggressiveness of
the automated vehicle by analyzing its results compared to
those of the human driver strategy.
a and ti
Assume ti
h are the time used by the automated vehi-
cle and the human driver for passing through the unsignalized
crossing in experiment i, respectively. Then, τi = ti
a/ti
h
indicates the efficiency of the automated vehicle compared
to the human driver. After N = 50 experiments, the mean
value and distribution of {τi} is shown in Fig. 14. As n
increases, the mean value of {τi} approaches a limit value,
which means that the time used by the automated vehicle to
pass through the intersection is stable.
Fig. 14. Mean value and distribution of {τi} (the ratio of time used in
experiments i). As the number of simulated experiment increases, the mean
value of {τi} approaches a limit value.
µ (The mean value of {τi}), cv (the coefficient of variation
of {τi}) and κ (the crash rate) are parameters to indicate the
passing strategy's efficiency, stability and safety, respectively.
They are calculated to evaluate the automated vehicle based
on the experimental results:
(cid:80) τi
(cid:80)N
N
i=1(τi − µ)2
µ
= 0.7044,
κ =
ncrash
N
= 0.
µ =
(cid:113)
1
N
cv =
σ
µ
=
= 4.54%
(10)
(11)
(12)
It turns out that the passing strategy of the automated
vehicle is more efficient than that of a human driver, taking
only about 70% of the time a human driver uses to pass
the unsignalized crossing. It is also quite stable since cv
is under 5%. More importantly, no crash occurred during
the simulation, meaning the strategy is also safe under ideal
conditions.
V. CONCLUSION AND DISCUSSION
In this paper, we proposes a method to evaluate the passing
strategy of automated vehicles at unsignalized crossings. A
stochastic interaction model is developed to predict pedes-
trian movement. A simulation is conducted to demonstrate
the evaluation procedure. The automated vehicle is evaluated
against the human drivers for better performance in terms
of the time used and the crash rate. One Soft-Yield driver
model is evaluated as an example. The simulation results
indicate that this driving strategy is more efficient than that of
a human driver that is modeled by naturalistic data collected
in Ann Arbor.
The pedestrians in our evaluation experiments comes from
both sides of the crossing. Their arriving times obey the Pois-
son process; the density of pedestrian flow can be adjusted
by setting different values of parameters. The pedestrians
in the simulation will have a look at the oncoming vehicle
and then decide on their walking speed before moving,
[17] A. Laureshyn, s. Svensson, and C. Hyd'en, "Evaluation of traffic
safety, based on micro-level behavioural data Theoretical framework
and first implementation," Accident Analysis and Prevention, vol. 42,
no. 6, pp. 1637–1646, 2010.
[18] G. Lee and C. Scott, "EM algorithms for multivariate Gaussian mixture
models with truncated and censored data," Computational Statistics
and Data Analysis, vol. 56, no. 9, pp. 2816–2829, 2012.
similar to the process in the real world. However, a more
detailed pedestrian model remains to be developed, and more
behaviors and features can be taken into account in future
work.
DISCLAIMERS
This work was funded in part by the University of Michi-
gan Mobility Transformation Center Denso Tailor project.
The findings and conclusions in the report are those of the
authors and do not necessarily represent the views of the
MTC or Denso.
REFERENCES
[1] N. Highway Traffic Safety Administration and U. Department of
"Research Note 2015 Motor Vehicle Crashes
Transportation,
Overview," no. August, 2016.
[2] D. Zhao, H. Lam, H. Peng, S. Bao, D. J. Leblanc, and C. S.
Pan, "Accelerated Evaluation of Automated Vehicles Safety in Lane
Change Scenarios based on Importance Sampling Techniques," IEEE
Transactions on Intelligent Transportation Systems, 2016.
[3] D. Zhao, X. Huang, H. Peng, H. Lam, and D. J. LeBlanc, "Accelerated
Evaluation of Automated Vehicles in Car-Following Maneuvers,"
submitted to IEEE Transactions on Intelligent Transportation Systems,
7 2017. [Online]. Available: http://arxiv.org/abs/1607.02687
[4] Z. Huang, D. Zhao, H. Lam, and D. J. LeBlanc, "Accelerated
Evaluation of Automated Vehicles Using Piecewise Mixture Models,"
submitted to IEEE Transactions on Intelligent Transportation Systems,
7 2017. [Online]. Available: http://arxiv.org/abs/1701.08915
[5] C. Burstedde, K. Klauck, A. Schadschneider, and J. Zittartz, "Simula-
tion of pedestrian dynamics using a two-dimensional cellular automa-
ton," Physica A Statistical Mechanics and its Applications, vol. 295,
no. 3-4, pp. 507–525, 2001.
[6] M. Chraibi, A. Seyfried, A. Schadschneider, and W. MacKens,
"Quantitative description of pedestrian dynamics with a force-based
model," Proceedings-2009 IEEEWICACM International Conference
on Web Intelligence and Intelligent Agent Technology-Workshops, WI-
IAT Workshops 2009, vol. 3, pp. 583–586, 2009.
[7] T. Xiong, P. Zhang, S. C. Wong, C.-W. Shu, and M.-P. Zhang,
"A Macroscopic Approach to the Lane Formation Phenomenon in
Pedestrian Counterflow," Chinese Physics Letters, vol. 28, p. 108901,
2011.
[8] S. Bennett, A. Felton, and R. Akccelik, "Pedestrian movement char-
acteristics at signalised intersections," 23rd Conference of Australian
Institutes of Transport Research, no. December, pp. 10–12, 2001.
[9] M. S. Tarawneh, "Evaluation of pedestrian speed in Jordan with
investigation of some contributing factors," Journal of Safety Research,
vol. 32, no. 2, pp. 229–236, 2001.
[10] D. Yagil, "Beliefs, motives and situational factors related to pedes-
trians? self-reported behavior at signal-controlled crossings," Trans-
portation Research Part F Traffic Psychology and Behaviour, vol. 3,
pp. 1–13, 2000.
[11] B. Dazhi Sun, S. VSK Ukkusuri, R. F. Benekohal, S. Travis Waller,
D. Sun, R. F. Benekohal, and S. T. Waller, "Modeling of Motorist-
Pedestrian Interaction at Uncontrolled Mid-block Crosswalks," no.
November, 2002.
[12] B. B. J. Schroeder, "A Behavior-Based Methodology for Evaluating
Pedestrian-Vehicle Interaction at Crosswalks," Analysis, p. 332, 2008.
[13] B. Schroeder, N. Rouphail, K. Salamati, E. Hunter, B. Phillips,
L. Elefteriadou, T. Chase, Y. Zheng, V. Sisiopiku P, and S. Mamidi-
palli, "Empirically-Based Performance Assessment and Simulation of
Pedestrian Behavior at Unsignalized Crossings," no. September, 2014.
[14] Keisuke Suzuki, Takuya Kakihara, and Yasutoshi Horii, "Investigation
of Braking Timing of Drivers for Design of Pedestrian Collision
Avoidance System," Journal of Mechanics Engineering and Automa-
tion, vol. 6, no. 3, pp. 118–127, 2016.
Test
- AEB VRU
systems," no. 1.0.1, 2015. [Online]. Available: httpwww.euroncap.
comenfor-engineersprotocolssafety-assist
[15] Euro-NCAP,
"Euro NCAP
Protocol
[16] D. Bezzina and J. Sayer, "Safety pilot model deployment: Test
conductor team report," Report No. DOT HS, vol. 812, p. 171, 2014.
|
1112.1335 | 1 | 1112 | 2011-12-06T16:30:35 | Connectivity and Set Tracking of Multi-agent Systems Guided by Multiple Moving Leaders | [
"cs.MA"
] | In this paper, we investigate distributed multi-agent tracking of a convex set specified by multiple moving leaders with unmeasurable velocities. Various jointly-connected interaction topologies of the follower agents with uncertainties are considered in the study of set tracking. Based on the connectivity of the time-varying multi-agent system, necessary and sufficient conditions are obtained for set input-to-state stability and set integral input-to-state stability for a nonlinear neighbor-based coordination rule with switching directed topologies. Conditions for asymptotic set tracking are also proposed with respect to the polytope spanned by the leaders. | cs.MA | cs |
Connectivity and Set Tracking of Multi-agent
Systems Guided by Multiple Moving Leaders∗
Guodong Shi, Yiguang Hong†and Karl Henrik Johansson ‡
Abstract
In this paper, we investigate distributed multi-agent tracking of a convex set specified by
multiple moving leaders with unmeasurable velocities. Various jointly-connected interaction
topologies of the follower agents with uncertainties are considered in the study of set tracking.
Based on the connectivity of the time-varying multi-agent system, necessary and sufficient
conditions are obtained for set input-to-state stability and set integral input-to-state stability
for a nonlinear neighbor-based coordination rule with switching directed topologies. Condi-
tions for asymptotic set tracking are also proposed with respect to the polytope spanned by
the leaders.
Keywords. Multi-agent systems, multiple leaders, set input-to-state stability (SISS), set
integral input-to-state stability (SiISS), set tracking.
1
Introduction
The last decade has witnessed tremendous interest devoted to the investigation of collective
phenomena in multiple autonomous agents, due to broad applications in various fields of science
ranging from biology to physics, engineering, and ecology, just to name a few [8, 9, 10, 11, 12].
Concerning the issues of multi-agent systems and distributed design, the revolutionary idea
is underpinning a strong interaction of individual dynamics, communication topologies, and
∗This work has been supported in part by the NNSF of China under Grants 60874018, 60736022, and 60821091,
the Knut and Alice Wallenberg Foundation and the Swedish Research Council.
†G. Shi and Y. Hong are with Key Laboratory of Systems and Control, Institute of Systems Science, Chinese
Academy of Sciences, Beijing 100190, China. Email: [email protected], [email protected]
‡K. Johansson is with ACCESS Linnaeus Centre, School of Electrical Engineering, Royal Institute of Technol-
ogy, Stockholm 10044, Sweden. Email: [email protected]
1
distributed controls. The problem is generally very challenging due to the complex dynamics
and hierarchical structures of the systems. However, efforts have been started with relatively
simple problems such as consensus, formation, and rendezvous, and many significant results
have been obtained.
The leader-follower coordination is an important multi-agent control problem, where the
leader may be a real leader (such as a target, an evader, or a predefined position), or a virtual
leader (such as a reference trajectory or a specified path). In most theoretical work, a single
leader with exact measurement is considered on multi-agent systems for each agent to follow.
However, in practical situations, multiple leaders and target sets with unmeasurable variables are
considered to achieve desired collective behaviors. In [8], a simple model was given to simulate
fish foraging and demonstrate the leader effectiveness when the leaders (or informed agents)
guide a school of fish to a particular food region. In [23], a straight-line formation of a group
of agents was discussed, where all the agents converge to the line segment specified by two edge
leaders. A containment control scheme was proposed with fixed undirected interaction in [24],
which aimed at driving a group of agents to a given target location and making their positions
contained in the polytope spanned by multiple stationary or moving leaders during their motion.
Region following formation control was constructed [25], where all the robots are driven and
then stay within a moving target region as a group. Moreover, different dynamic connectivity
conditions were obtained to guarantee that the multiple leaders (or informed agents) aggregate
the whole multi-agent group within a convex target set in [26]. Additionally, control strategies
were demonstrated and analyzed to drive a collection of mobile agents to stationary/moving
leaders with connectivity-maintenance and collision-avoidance with fixed and switching directed
network topologies in [27]. As a matter of fact, multiple leaders are usually assigned to increase
control effectiveness, enhance communication/sensing range, improve reliability, and optimize
energy cost in multi-agent coordination.
Connectivity plays a key role in the coordination of multi-agent networks, which is related to
the influence of agents and controllability of the network. Due to mobility of the agents, inter-
agent topologies usually keep changing in practice. Therefore, the various connectivity conditions
to describe frequently switching topologies in order to deal with multi-agent consensus or flocking
[15, 16, 18, 21]. In fact, the "joint connection" or similar concepts are important in the analysis of
stability and convergence to guarantee multi-agent coordination with time-dependent topology.
Uniformly jointly-connected conditions have been employed for different problems. [28] studied
2
the distributed asynchronous iterations, while [22] proved the consensus of a simplified Vicsek
model. Furthermore, [14] and [6] investigated the jointly-connected coordination for second-order
agent dynamics via different approaches, while [30] worked on nonlinear continuous-time agent
dynamics with jointly-connected interaction graphs. Also, flocking of multi-agent system with
state-dependent topology was studied with non-smooth analysis in [18, 20]. What is more, the
[t,∞) joint connection condition, which is more generalized than the uniformly joint connection
assumption, was discussed by Moreau, in order to achieve the consensus for discrete-time agents
in [31]. This [t,∞) connectivity concept was then extended in the distributed control analysis
for the target set convergence in [26].
It is well known that input-to-state stability (ISS) is an important and very useful tool in the
study of the stability and stabilization of control systems [29, 35]. Its variants such as integral
input-to-state stability (iISS) were discussed in [34]. Then few works on set input-to-state
stability (SISS) were done with respect to fixed sets in [33]. On the other hand, ISS or related
ideas can facilitate the control analysis and synthesis with interconnection conditions like small
gains (referring to [29], for example). ISS has recently been applied to the stability study of a
group of interconnected nonlinear systems [32]. Moreover, an extended concept called leader-
to-formation stability was introduced to investigate the stability of the formation of a group of
agents in light of ISS properties [19]. In fact, ISS application in multi-agent systems is promising.
The contributions of the paper include:
• We propose the generalized set input-to-state stability (SISS) and set integral-input-to-
state stability (SiISS) to handle moving sets with time-varying shapes for switching multi-
agent networks.
• We study the multi-leader coordination from the ISS viewpoint. With the help of SISS and
SiISS, we give explicit expressions to estimate the convergence rate and tracking error of
a group of mobile agents that try to enter the convex hull determined by multiple leaders.
• We show relationships between the connectivity and set tracking of the multi-agent sys-
tem, and find that various jointly-connected conditions usually provide necessary and/or
sufficient conditions for distributed coordination.
• We develop a method to study SISS and SiISS for a moving set and switching topology with
graph theory and non-smooth analysis. In fact, we cannot take the standard approaches
to conventional ISS or iISS using equivalent ISS-Lyapunov functions [34, 35]. In addition,
3
the classic algebraic methods based on Laplacian may fail due to disturbances in nonlinear
agent dynamics, uncertain leader velocities, or moving multi-leader set.
This paper is organized as follows. Section 2 presents the preliminaries and problem formu-
lation, while Section 3 proposes results for the convergence estimation. Section 4 mainly reports
a necessary and sufficient condition for the SISS with respect to the moving multi-leader set
with switching inter-agent topologies, and then presents a set-tracking case based on the SISS.
Correspondingly, Section 5 obtains necessary and sufficient conditions for SiISS and then shows
set-tracking results related to SiISS. Finally, Section 6 gives concluding remarks.
2 Problem Formulation
In this section, we introduce some preliminary knowledge for the following discussion.
First we introduce some basic concepts in graph theory (referring to [13] for details). A
directed graph (digraph) G = (V,E) consists of a finite set V = {1, 2, ..., n} of nodes and an arc
set E, in which an arc is an ordered pair of distinct nodes of V. (i, j) ∈ E describes an arc which
leaves i and enters j. A walk in digraph G is an alternating sequence W : i1e1i2e2 . . . em−1im of
nodes iκ and arcs eκ = (iκ, iκ+1) ∈ E for κ = 1, 2, . . . , m− 1. A walk is called a path if the nodes
of this walk are distinct, and a path from i to j is denoted as [(i, j). Node j is called reachable
from i if there is a path [(i, j). If the nodes i1, . . . , im−1 are distinct and i1 = im, W is called a
(directed) cycle. A digraph without cycles is said to be acyclic.
The union of the two digraphs G1 = (V,E1) and G2 = (V,E2) is defined as G1 ∪ G2 =
(V,E1 ∪ E2) if they have the same node set. Furthermore, a time-varying digraph is defined as
Gσ(t) = (V,Eσ(t)) with σ : t → Q as a piecewise constant function, where Q is the finite set
which consists of all the possible digraphs with node set V. Moreover, the joint digraph of Gσ(t)
in time interval [t1, t2) with t1 < t2 ≤ +∞ is denoted as
G([t1, t2)) = ∪t∈[t1,t2)G(t) = (V,∪t∈[t1,t2)Eσ(t)).
(1)
Next, we recall some notations in convex analysis (see [2]). A set K ⊂ Rd is said to be
convex if (1 − λ)x + λy ∈ K whenever x ∈ K, y ∈ K and 0 ≤ λ ≤ 1. For any set S ⊂ Rd,
the intersection of all convex sets containing S is called the convex hull of S, denoted by co(S).
Particularly, the convex hull of a finite set of points x1, . . . , xn ∈ Rd is a polytope, denoted by
co{x1, . . . , xn}. In fact, we have co{x1, . . . , xn} = {λ1x1 + ··· + λnxnλ1 + ··· + λn = 1, λi ≥ 0}.
4
Let K be a closed convex subset in Rd and denote xK , inf{x − y y ∈ K}, where ·
denotes the Euclidean norm for a vector or the absolute value of a scalar ([35, 34]). Then we
can associate to any x ∈ Rd a unique element PK(x) ∈ K satisfying x − PK (x) = xK , where
the map PK is called the projector onto K and
hPK (x) − x,PK (x) − yi ≤ 0,
∀y ∈ K.
Clearly, x2
K is continuously differentiable at point x, and (see [1])
∇x2
K = 2(x − PK (x)).
The following lemma was obtained in [26], which is useful in what follows.
Lemma 2.1 Suppose K ⊂ Rd is a convex set and xa, xb ∈ Rd. Then
hxa − PK (xa), xb − xai ≤ xaK · xaK − xbK .
Particularly, if xaK > xbK , then
hxa − PK (xa), xb − xai ≤ −xaK · (xaK − xbK ).
(2)
(3)
(4)
(5)
Then we consider the Dini derivative for the following non-smooth analysis. Let a and b (> a)
be two real numbers and consider a function h : (a, b) → R and a point t ∈ (a, b). The upper
Dini derivative of h at t is defined as
D+h(t) = lim sup
s→0+
h(t + s) − h(t)
s
.
It is well known that when h is continuous on (a, b), h is non-increasing on (a, b) if and only if
D+h(t) ≤ 0 for any t ∈ (a, b) (more details can be found in [3]). The next result is given for the
calculation of Dini derivative [4, 30].
Lemma 2.2 Let Vi(t, x) : R × Rd → R (i = 1, . . . , n) be C 1 and V (t, x) = maxi=1,...,n Vi(t, x).
If I(t) = {i ∈ {1, 2, . . . , n} : V (t, x(t)) = Vi(t, x(t))} is the set of indices where the maximum is
reached at t, then D+V (t, x(t)) = maxi∈I(t)
Vi(t, x(t)).
In this paper, we consider the set coordination problems for a multi-agent system consisting
of n follower-agents and k leader-agents (see Fig. 1). The follower set is denoted as VF ,
{v1, . . . , vn}, and the leader set is denoted as VL , {v1, . . . , vk}. In what follows, we will identify
follower vi or leader vi with its index i (namely, agent i or leader i) if there is no confusion.
5
Figure 1: Multiple agents (vi, i = 1, 2, 3, 4) with multiple leaders (vi, i = 1, 2, 3)
Then we describe the communication in the multi-agent network. At time t, if i ∈ VF can
"see" j ∈ VF , there is an arc (j, i) (marking the information flow) from j to i, and then agent
j is said to be a neighbor of agent i. Moreover, if i ∈ VF "sees" j ∈ VL at time t, there is an
arc (j, i) leaving from j and entering i, and then j is said to be a leader of agent i. Let Ni and
Li represent the set of agent i's neighbors and the set of agent i's leaders (that is, the leaders
which are connected to agent i), respectively. Note that, since the leaders are not influenced by
the followers, there is no arc leaving from VF entering VL.
Define V = VF ∪ VL as the whole agent set (including leaders and followers). Denote P as
the set of all possible interconnection topologies, and σ : [0, +∞) → P as a piecewise constant
switching signal function to describe the switchings between the topologies. Thus, the interaction
topology of the considered multi-agent network is described by a time-varying directed graph
Gσ(t) = (V,Eσ(t)). Correspondingly, GF
σ(t) is denoted as the communication graph among the
follower agents. Additionally, let Ni(σ(t)) and Li(σ(t)) represent the set of agent i's neighbors
and the set of its connected leaders in Gσ(t), respectively.
As usual in the literature [22, 30, 26], an assumption is given for the switching signal σ(t).
Assumption 1 (Dwell Time) There is a lower bound τD > 0 between two switching instants.
We give definitions for the connectivity of a multi-agent system with multiple leaders.
Definition 2.1 (i) Gσ(t) is said to be L-connected if, for any i ∈ VF , there exists a leader j ∈ VL
such that there is a path from leader j to agent i in Gσ(t) at time t. Moreover, Gσ(t) is said to be
jointly L-connected in time interval [t1, t2) if the union graph G([t1, t2)) is L-connected;
(ii) Gσ(t) is said to be jointly L-connected (JLC) if the union graph G([t,∞)) is L-connected
6
for any t;
(iii) Gσ(t) is said to be uniformly jointly L-connected (UJLC) if there exists T > 0 such that
the union graph G([t, t + T )) is L-connected for any t ≥ 0.
Remark 2.1 Note that the L-connectedness describes the capacity for the follower agents to get
the information from the moving multi-leader set in the information flow, and an L-connected
graph may not be connected since the graph with leaders as its nodes may not be connected. In
fact, if we consider the group of the leaders as one virtual node in V, then the L-connectedness
becomes the quasi-strong connectedness for a digraph [5, 30].
The state of agent vi ∈ VF , is denoted as xi ∈ Rd (i = 1, . . . , n), and the state of leader
vi ∈ VL, is denoted as yi ∈ Rd (i = 1, . . . , k). Denote x = (x1, . . . , xn)T ∈ Rnd and y =
(y1, . . . , yk)T ∈ Rkd and let the continuous function aij(x, y, t) > 0 be the weight of arc (j, i), if
any, for i, j ∈ VF , and continuous function bij(x, y, t) > 0 be the weight of arc (j, i), if any, for
i ∈ VF ; j ∈ VL.
Then we present the multi-agent model for the active leaders and the (follower) agents
i = 1, . . . , k
yi = ui(y, t),
xi = Pj∈Ni(σ(t))
(6)
where ui(y, t) describes the control inputs of the leader i, i ∈ VL, which is continuous in y for
fixed t and piecewise continuous in t for fixed y, and wi(t) is a continuous function to describe
aij(x, y, t)(xj − xi) + Pj∈Li(σ(t))
bij(x, y, t)(yj − xi) + wi(t),
i = 1, . . . , n
the disturbances in communication links and individual dynamics to follower agent i. Then
another assumption is given on the weight functions aij(x, y, t) and bij(x, y, t).
Assumption 2 (Bounded Weights) There are 0 < a∗ ≤ a∗ and b∗ > 0 such that a∗ ≤
aij(x, y, t) ≤ a∗, b∗ ≤ bij(x, y, t) for any x, y, t.
Remark 2.2 In (6), the weights, aij and bij, may not be constant.
Instead, because of the
complex communication and environment uncertainties, they are dependent on time or space or
relative measurement (see nonlinear models given in [30, 26, 31, 18]). Some models such as those
studied in [30, 26] can be written in the form of (6), while other nonlinear multi-agent models
may be transformed to this class of multi-agent systems in some situations. Here aij(x, y, t) and
bij(x, y, t) are written in a general form simply for convenience, and global information is not
required in our study. For example, aij and bij can depend only on the state of xi, time t and
7
xj (j ∈ Ni), which is certainly a special form of aij(x, y, t) or bij(x, y, t). In other words, the
control laws in specific decentralized forms are still decentralized.
Without loss of generality, we assume the initial time t = 0, and the initial condition x0 =
(x1(0), . . . , xn(0))T ∈ Rnd and y0 = (y1(0), . . . , yk(0))T ∈ Rkd.
Denote the time-varying polytope formed by the k active leaders
L(y(t)) , co{y1(t), . . . , yk(t)},
(7)
and let
x(t)L(y(t))
, max
i∈VF xi(t)L(y(t))
be the maximal distance for the followers away from the moving multi-leader set L(y(t)).
The following definition is to describe the convergence to the moving convex set L(y(t)).
Definition 2.2 The (global) set tracking (ST) with respect to L(y(t)) for system (6) is achieved
if
lim
t→+∞x(t)L(y(t)) = 0
(8)
for any initial condition x0 ∈ Rnd and y0 ∈ Rkd.
For a stationary convex set K, set tracking can be reduced to set stability and attractivity,
and methods to analyze xi(t)K were proposed in some existing works [26]. In fact, [24, 27]
discussed the convergence to the static convex set determined by stationary leaders with well
designed control protocols. Moreover, if we assume that the target set is exactly the polytope
with the positions of the stationary leaders (or informed agents) as its vertices, then the conver-
gence to the polytope, treated as a target set, can be obtained straightforwardly based on the
results and limit-set-based methods given in [26].
Input-to-state stability has been widely used in the stability analysis and set input-to-state
stability (SISS) for a fixed set has been studied in [33]. To study the multi-leader set tracking
in a broad sense, we introduce a generalized SISS with respect to L(y(t)), a moving set with a
time-varying shape, for multi-agent systems with switching interaction topologies. Denote u ,
, {z : R≥0 → R(n+k)m kzk∞ < ∞}
(u1, . . . , uk)T , w , (w1, . . . , wn)T , z , (uT wT )T , and L∞
with kzk∞
, sup{z(t), t ≥ 0} ([35]).
A function γ : R≥0 → R≥0 is said to be a K-class function if it is continuous, strictly
increasing, and γ(0) = 0. Moreover, a function β : R≥0 × R≥0 → R is a KL-class function if
β(·, t) is of class K for each fixed t ≥ 0 and β(s, t) decreases to 0 as t → ∞ for each fixed s ≥ 0.
8
Definition 2.3 System (6) is said to be globally generalized set input-to-state stable (SISS) with
respect to L(y(t)) with input z if there exist a KL-function β and a K-function γ such that
x(t)L(y(t)) ≤ β(x0L(y0), t) + γ(kzk∞)
(9)
for z ∈ L∞ and any initial conditions x0 ∈ Rnd and y0 ∈ Rkd.
Integral-input-to-state stability (iISS) was introduced as an integral variant of ISS, which has
been proved to be strictly weaker than ISS [34]. We also introduce a definition of (generalized)
set integral-input-to-state stability (SiISS) with respect to a time-varying and moving set.
Definition 2.4 System (6) is (globally) generalized set integral-input-to-state stable (SiISS) with
respect to L(y(t)) if there exist a KL-function β and a K-function γ such that
x(t)L(y(t)) ≤ β(x0L(y0), t) +Z t
0
γ(z(s))ds,
(10)
for any initial conditions x0 ∈ Rnd and y0 ∈ Rkd.
The conventional SISS was given for a fixed set K ([33]), while the generalized SISS or SiISS
is proposed with respect to a time-varying set L(y(t)). In the following, we still use SISS or
SiISS instead of generalized SISS or SiISS for simplicity.
Remark 2.3 Similar to the study of conventional ISS, local SISS and SiISS can be defined. In
this paper, we focus on the global SISS and SiISS. In fact, it is rather easy to extend research
ideas of global set tracking to study local cases.
3 Convergence Estimation
For the set tracking with respect to a moving multi-leader set of system (6), we have to deal
with the estimation of xi(t)L(y(t)) when L(y(t)) is a time-varying convex set, where y(t) is a
trajectory of the moving leaders in system (6) with initial condition y0 = y(0). Define
r(t) , max
i∈VL ui(y(t), t);
q(t) , max
i∈VL ui(y(t), t) + max
i∈VF wi(t).
Obviously,
q(t) ≤ u(y(t), t) + w(t) ≤
√2z(t) ≤
√2 max{√n,√k}q(t).
(11)
(12)
The following result is given to estimate the changes of the distance between an agent and
the convex hull spanned by the leaders.
9
Lemma 3.1 For any t, t0 ≥ 0 and i = 1, . . . , n,
xi(t)L(y(t)) − xi(t)L(y(t0)) ≤Z t
t0
r(s)ds.
(13)
Proof: Suppose
PL(y(t0))(xi(t)) =
kXi=1
λiyi(t0) ∈ L(y(t0)),
where λi ≥ 0 for i = 1, . . . , k with
λi = 1. Define y(t) ,
kPi=1
y(t) − y(t0) ≤
kXi=1
λiyi(t) − yi(t0) =
λiZ t
t0
kXi=1
Moreover,
xi(t)L(y(t)) ≤ xi(t) − y(t)
λiyi(t), and then
kPi=1
ui(y(s), s)ds ≤Z t
t0
≤ xi(t) − y(t0) + y(t) − y(t0)
≤ xi(t)L(y(t0)) +Z t
r(s)ds
t0
Also, similar analysis leads to
xi(t)L(y(t0)) ≤ xi(t)L(y(t)) +Z t
t0
r(s)ds
Therefore, (14) and (15) lead to the conclusion.
For simplicity, define ψi(t) , xi(t)2
L(y(t)), i = 1, . . . , n and
Ψ(t) , max
i∈VF
ψi(t),
r(s)ds
(14)
(15)
(cid:3)
which is locally Lipschitz but may not be continuously differentiable. Clearly, xi(t)L(y(t)) =
pψi(t), i = 1, . . . , n and kx(t)kL(y(t)) =pΨ(t).
Then, we get the following lemma to estimate the set convergence.
Lemma 3.2 D+pΨ(t) ≤ q(t).
Proof: It is not hard to see that
dψi(t)
dt
= lim
∆t→0
= lim
∆t→0
ψi(t + ∆t) − ψi(t)
∆t
xi(t + ∆t)2
L(y(t+∆t)) − xi(t + ∆t)2
L(y(t)) − xi(t)2
L(y(t))
∆t
.
∆t
L(y(t))
xi(t + ∆t)2
+ lim
∆t→0
10
(16)
Then, according to (3), we obtain
L(y(t))
xi(t + ∆t)2
L(y(t)) − xi(t)2
=
∆t
lim
∆t→0
d
dsxi(s)2
= h∇xi(s)2
= 2hxi(t) − PL(y(t))(xi(t)), Xj∈Ni(σ(t))
L(y(t))(cid:12)(cid:12)(cid:12)(cid:12)s=t
L(y(t)), xi(s)i(cid:12)(cid:12)(cid:12)s=t
aij(xj(t) − xi(t)) + Xj∈Li(σ(t))
bij(yj(t) − xi(t)) + wi(t)i.
(17)
Furthermore, according to Lemma 3.1,
xi(t + ∆t)L(y(t+∆t)) − xi(t + ∆t)L(y(t))
∆t
lim
∆t→0
and then it is easy to find that
∆t→0R t+∆t
≤ lim
t
∆t
r(s)ds
= r(t),
xi(t + ∆t)2
L(y(t+∆t)) − xi(t + ∆t)2
L(y(t))
∆t
lim
∆t→0
xi(t + ∆t)L(y(t+∆t)) − xi(t + ∆t)L(y(t))
= lim
∆t→0
·(xi(t + ∆t)L(y(t+∆t)) + xi(t + ∆t)L(y(t)))
(18)
∆t
≤ 2r(t)xi(t)L(y(t)).
Therefore,
d
dt
ψi(t) ≤ 2hxi − PL(y(t))(xi), Xj∈Ni(σ(t))
+2r(t)xi(t)L(y(t)).
aij(x)(xj − xi) + Xj∈Li(σ(t))
bij(x)(yj − xi) + wi(t)i
(19)
Moreover, let I(t) denote the set containing all the agents that reach the maximal distance away
from L(y(t)) at time t. Then, for any i ∈ I(t), according to (2), one has
hxi − PL(y(t))(xi), yj − xii ≤ hxi − PL(y(t))(xi), yj − PL(y(t))(xi)i
+hxi − PL(y(t))(xi),PL(y(t))(xi) − xii
≤ hxi − PL(y(t))(xi),PL(y(t))(xi) − xii
= −ψi(t)
(20)
for any j ∈ Li(σ(t)). Furthermore, in light of Lemma 2.1, since i ∈ I(t),
hxi − PL(y(t))(xi), xj − xii ≤ −xiL(y(t))(xiL(y(t)) − xjL(y(t))) ≤ 0
11
for any j ∈ Ni(σ(t)). Therefore, the conclusion follows since
ψi(t)
d
D+Ψ(t) = max
dt
i∈I(t)
[hxi − PL(y(t))(xi), wi(t)i + 2r(t)xi(t)L(y(t))]
≤ 2 max
i∈I(t)
≤ 2(r(t) + max
i∈VF wi(t)) max
= 2q(t)pΨ(t)
i∈I(t) xi(t)L(y(t))
according to Lemma 2.2.
(cid:3)
4 Connectivity and SISS
In this section, we study the SISS with respect to the convex set spanned by the moving leaders
in an important connectivity case, uniformly jointly L-connected (UJLC) topology. Without
loss of generality, we will assume n ≥ 2 in the sequel.
4.1 Main results
Suppose z = (uT , wT )T ∈ L∞ in this section. Then we have the main result on SISS.
Theorem 4.1 System (6) is SISS with respect to L(y(t)) and with z as the input if and only if
Gσ(t) is UJLC.
The main difficulties to obtain the SISS inequalities in the UJLC case are how to estimate the
convergence rate in a time interval by "pasting" time subintervals together and how to estimate
the impact of the input z to the agent motion.
To prove Theorem 4.1, we first present two lemmas to estimate the distance error in the two
standard cases during t ∈ [t0, t0 + T∗] for t0 ≥ 0 and a constant T∗ > τD with τD as the dwell
time of switching.
Lemma 4.1 If there is an arc (j, i) leaving from follower j ∈ VL entering i ∈ VF in Gσ(t) for
all t ∈ [t0, t0 + τD), then there exist a continuous function µ(s) : [0, T∗] 7→ (0, 1] and a constant
γ1 > 0 such that
xi(t)L(y(t)) ≤ µ(t − t0)x(t0)L(y(t0)) + γ1kzk∞, ∀t ∈ [t0, t0 + T∗].
Proof: See Appendix A.1.
12
(21)
(cid:3)
Lemma 4.2 If there is an arc (i, m) leaving from i ∈ VF entering m ∈ VF in Gσ(t) for all
t ∈ [t0, t0 + τD), and
xi(t)L(y(t)) ≤ µ0x(t0)L(y(t0)) + d0, ∀t ∈ [t0, t0 + τD)
(22)
for constants µ0 ∈ (0, 1) and d0 > 0, then there exist a continuous function ξµ0(s) : [0, T∗] 7→
(0, 1] and a positive constant γ2 such that
kxm(t)kL(y(t)) ≤ ξµ0(t − t0)x(t0)L(y(t0)) + γ2kzk∞ + d0, ∀t ∈ [t0, t0 + T∗]
Proof: See Appendix A.2.
(23)
(cid:3)
Remark 4.1 The following properties of µ(s) and ξµ0(s) are quite critical in the study of the
set tracking with jointly L-connected topology (see Fig. 2):
(i) µ(0) = ξµ0(0) = 1.
(ii) µ(s) and ξµ0(s) are strictly decreasing during s ∈ [0, τD].
(iii) µ(s) and ξµ0(s) are strictly increasing during s ∈ [τD, T ∗], and µ(T ∗) < 1, ξµ0 (T ∗) < 1.
Figure 2: µ(s) and ξµ0(s)
Next, we introduce the following lemma to state an important property for UJLC graphs.
Lemma 4.3 If Gσ(t) is UJLC, then, for any t > 0 and i ∈ VF , there is a path d(j, i) from some
leader j ∈ VL to follower i in G([t, t + T0)) with T0 , T + 2τD, and each arc of d(j, i) exists in a
time interval with length τD at least during [t, t + T0).
13
Proof: Denote t1 as the first moment when the interaction topology switches within [t, t + T0)
(suppose there are switchings without loss of generality). If t1 ≥ t + τD, then, for any i ∈ VF ,
there is a path d(j, i) from some leader with index j ∈ VL to agent i in G([t, t + T )), where each
arc stays there for at least the dwell time τD during [t, t + T + τD) due to the definition of τD.
On the other hand, if t1 < t + τD, t1 + T + τD < t + T0. Then, for any i ∈ VF , there is also a
path d(j, i) from some leader j ∈ VL to agent i in G([t1, t1 + T )) in [t1, t1 + T + τD) with each arc
exists for at least τD. This completes the proof.
(cid:3)
Remark 4.2 If there is a convex set Ω such that L(y(t)) ∈ Ω,∀t ≥ 0, that is, Ω is a positively
invariant set for the leaders, then x(t)Ω ≤ x(t)L(y(t)). By Theorem 4.1, system (6) is SISS
with respect to Ω with w as the input if Gσ(t) is UJLC.
Sometimes, the velocities of the moving leaders and uncertainties in agent dynamics (maybe
because of the online estimation) may vanish. To be strict, consider the following condition
limt→+∞ ui(y, t) = 0 uniformly for y; i = 1, . . . , k;
limt→+∞ wj(t) = 0, j = 1, . . . , n.
(24)
Clearly, (24) yields that for any ε > 0, there is Tε > 0 such that kzTεk∞ < ε, where zTε is
the truncated part of z defined on [Tε, +∞). Suppose (24) holds and Gσ(t) is UJLC. Based on
Theorem 4.1, for any ε > 0, there is Tε > 0 such that
x(t)L(y(t)) ≤ β(x(Tε)L(y(Tε)), t) + γ(ε).
Hence, the set tracking for system (6) with respect to set L(y(t)) is achieved easily. On the other
hand, similar to the proof of Theorem 4.1, the necessity of the global set tracking for system (6)
with condition (24) can also be simply proved by counterexamples since z(t) may be large and
the distance error may accumulate to a very large value over a sufficiently long period of time.
Therefore, we have the following result.
Corollary 4.1 The global set tracking with respect to L(y(t)) is achieved for all z(t) satisfying
(24) if and only if Gσ(t) is UJLC.
4.2 Proof of Theorem 4.1
We are now in a position to prove Theorem 4.1: "If" part: Denote T∗ = nT0 with T0 = T + 2τD.
Then we estimate Ψ(t) at subintervals [t∗ + (j − 1)T0, t∗ + jT0] for j = 1, . . . , n.
14
Based on Lemma 4.3, in [t∗, t∗ + T0), there must be an arc (j1, i1) ∈ E([t∗, t∗ + T0)) leaving
from a leader j1 ∈ VL to a follower i1 ∈ VL and this arc remains for at least τD. Suppose
(j1, i1) ∈ Eσ(t) for t ∈ [t1, t1 + τD) ⊂ [t∗, t∗ + T0). According to Lemma 4.1,
xi1(t)L(y(t)) ≤ µ(t − t1)x(t1)L(y(t1)) + γ1kzk∞, t ∈ [t1, t1 + T∗],
(25)
where µ(s) and γ1 were defined in Lemma 4.1. Take η1 = sup{µ(s) s ∈ [T0, T∗]} = µ(T∗). Since
0 < µ1 < 1,
xi1(t)L(y(t)) ≤ η1x(t1)L(y(t1)) + γ1kzk∞, t ∈ [t∗ + T0, t∗ + T∗].
(26)
Furthermore, in [t∗ + T0, t∗ + 2T0), there must be a follower i2 ∈ VF , i2 6= i1, such that there
exists an arc (j2, i2) for some j2 ∈ VL, or an arc (i1, i2) in E([t∗ + T0, t∗ + 2T0)).
There are two cases:
1) If (j2, i2) ∈ Eσ(t) for t ∈ [t2, t2 + τD) ⊂ [t∗ + T0, t∗ + 2T0), one also has
xi2(t)L(y(t)) ≤ η1x(t2)L(y(t2)) + γ1kzk∞, t ∈ [t∗ + 2T0, t∗ + T∗].
(27)
2) If (i1, i2) ∈ Eσ(t) for t ∈ [t2, t2 + τD) ⊂ [t∗ + T0, t∗ + 2T0). According to (12) and Lemma
3.2, one has
x(t1)L(y(t1)) ≤ x(t2)L(y(t2)) + √2kzk∞ · t2 − t1 ≤ x(t2)L(y(t2)) + 2√2kzk∞T0,
Thus, (26) will lead to
xi1(t)L(y(t)) ≤ η1x(t2)L(y(t2)) + (2√2η1T0 + γ1)kzk∞, t ∈ [t∗ + T0, t∗ + T∗].
(28)
Then, by Lemma 4.2, if we take η2 = ξη1((n − 1)T0), then
xi2(t)L(y(t)) ≤ η2x(t2)L(y(t2)) + (γ2 + 2√2η1T0 + γ1)kzk∞, t ∈ [t∗ + 2T0, t∗ + T∗]. (29)
Because η2 > η1,
x(t)L(y(t)) ≤ η2x(t2)L(y(t2)) + (γ2 + 2√2η1T0 + γ1)kzk∞, = i1, i2, t ∈ [t∗ + 2T0, t∗ + T∗]. (30)
Repeating the above procedure yields
ηj = ξηj−1((n − j + 1)T0), j = 3, . . . , n
15
and tj ∈ [t∗ + jT0, t∗ + T∗) such that, there exists ij ∈ VF , j = 3, . . . , n satisfying
x(t)L(y(t)) ≤ ηjx(tj)L(y(tj )) + [(j − 1)γ2 + 2√2
j−1Xl=1
ηlT0 + γ1]kzk∞, = i1, . . . , ij
(31)
for t ∈ [t∗ + jT0, t∗ + T∗]. Moreover, the nodes ij , j = 1, 2, . . . , n are distinct.
Denote η∗ = ηn, and then 0 < η∗ < 1. Thus, (31) leads to
x(t∗ + T∗)L(y(t∗+T∗)) ≤ η∗x(t∗)L(y(t∗)) + [(1 + 2√2)η∗T∗ + (n − 1)γ2 + γ1]kzk∞,
for any ∈ VF , which leads to
x(t∗ + T∗)L(y(t∗+T∗)) ≤ η∗x(t∗)L(y(t∗)) + [(1 + 2√2)η∗T∗ + (n − 1)γ2 + γ1]kzk∞.
Therefore, ∀N = 1, 2, . . . ,
x(N T∗)L(y(N T∗)) ≤ ηN
∗ x0L(y0) +
N−1Xj=0
ηj
∗[(1 + 2√2)η∗T∗ + (n − 1)γ2 + γ1]kzk∞.
Again by Lemma 3.2, one has
x(t)L(y(t)) ≤ β(x0L(y0), t) + γ(kzk∞)
with
(32)
(33)
(34)
(35)
β(x0L(y0), t) , η⌊ t
T ∗ ⌋
∗
T ∗⌋ denotes the largest integer no greater than t
x0L(y0),
γ(s) , [
where ⌊ t
1 − η∗
(1 + 2√2)η∗T∗ + (n − 1)γ2 + γ1
+ T∗]s
T ∗ , which implies the conclusion.
"Only if" part: If Gσ(t) is not UJLC, there is a time sequence 0 < T1 < T2 < . . . such that
G([T2κ−1, T2κ)) is not L-connected for κ = 1, 2, . . . with limκ→∞(T2κ − T2κ−1) = ∞. Taking
xi(0) = (0, . . . , 0)T ∈ Rd,∀i ∈ VF and yi(0) = (1, . . . , 1)T ∈ Rd,∀i ∈ VL with wi(t) ≡ 0,∀i ∈ VF
and ui(y, t) ≡ (1, . . . , 1)T ,∀i ∈ VL, we obtain L(y(t)) = {(1+t, . . . , 1+t)T }. Since G([T2κ−1, T2κ))
is not L-connected, there is i ∈ VF such that agent i is reachable from no leader. Define
V 1
i contains no leader and
there is no arc entering V 1
i } when t ∈ [T2κ−1, T2κ).
Moreover, none of the followers can enter L(y(t)) in finite time. Therefore,
, {j ∈ Vi is reachable from j in graph G([T2κ−1, T2κ))}. Since V 1
leaves co{x(T2κ−1), ∈ V 1
i , no agent in V 1
i
i
lim
κ→∞x(T2κ)L(y(T2κ )) ≥ lim
κ→∞
(T2κ − T2κ−1) = +∞,
∀ ∈ V 1
i .
Thus, the SISS with respect to L(y(t)) cannot be achieved.
(cid:3)
16
5 Connectivity and SiISS
In this section, we aim at the connectivity requirement to ensure the set integral-input-to-state
stability (SiISS) when Gσ(t) is jointly L-connected (JLC).
5.1 Main results
Theorem 4.1 showed an equivalent relationship between SISS and UJLC. However, this is not
true for SiISS. Here, we propose a couple of theorems about SiISS. The proofs of these conclusions
can be found in the following subsection.
First of all, we propose a sufficient condition.
Theorem 5.1 System (6) is SiISS with respect to L(y(t)) if Gσ(t) is UJLC.
Remark 5.1 JLC of Gσ(t) (i.e., G([t,∞)) is L-connected for any t) is necessary for the SiISS,
though it is not sufficient. If G([ T ,∞)) is not L-connected for some T > 0, there is a subset
VF ⊆ VF such that no arcs enter VF in G([ T ,∞)). Hence, the agents in VF may not be SiISS
for some initial conditions since they will not be influenced by the convex leader-set after T .
UJLC, which is a special case of JLC, provides a sufficient condition for SiISS, but UJLC is
not necessary to ensure SiISS. In fact, there are other cases of JLC to make SiISS hold. Here
we consider two important special JLC cases i.e., bidirectional graphs and acyclic graphs.
A digraph G is called a bidirectional graph when i is a neighbor of j if and only if j is a
neighbor of i, but the weight of arc (i, j) may not be equal to that of arc (j, i). The next result
shows a necessary and sufficient condition for the bidirectional case.
Theorem 5.2 Suppose that GF
only if Gσ(t) is JLC.
σ(t) is bidirectional for all t ≥ 0. Then system (6) is SiISS if and
The next lemma shows an important property for an acyclic digraph, that is, a digraph
without cycles.
Lemma 5.1 Assume that GF ([0, +∞)) is acyclic and G([0, +∞)) is L-connected. Then there
is a partition of VF by VF = Sk0
i , k0 ≥ 1 such that in graph G([0, +∞)), all the arcs
entering node set V F
j , j = 2, . . . , k0 are from
VL ∪ (Sj−1
i=1 V F
i ).
1 are from VL ; and all the arcs entering node set V F
i=1 V F
17
1
, {j ∈ VFthere is an arc leaving from VL entering j}.
1 6= ∅. Take i0 ∈ V F
1 exists by contradiction. If V F
1 does not exist, every agent i, i ∈ VF has
Proof: First we prove V F
neighbors within VF in G([0, +∞)). Denote V F
Clearly V F
1 . Then, there is j1 ∈ VF such that (j1, i0) ∈ G([0, +∞)). Moreover,
1 (i1 cannot be i0, of course) such that there is a path \(i1, j1)
we can associate j1 with i1 ∈ V F
1 ). Hence, a path \(i1, i0) in G([0, +∞)) is found. Regarding i1 as
in G([0,∞)) (i1 = j1 if j1 ∈ V F
i0 and repeating the above procedure yields the existence of \(i2, i1) in G([0, +∞)) with i2 ∈ V F
1 .
In this way, we obtain a path \(il+1, il) in G([0, +∞)) with il ∈ V F
1 , l = 2, 3, . . . . Since the nodes
in V F
1 are finite, there has to be il1 = il2 for some l1 > l2 ≥ 0, which lead to a directed cycle in
G([0, +∞)). Therefore, there is V F
Next, by replacing VL with V F
1 to make the conclusion hold.
1 ∪ VL in G([0,∞)), with the same analysis we can find V F
2 to
make the conclusion hold. Repeating this procedure, since the number of all the agents is finite,
there will be a constant k0 ≥ 1 such that VF =Sk0
i=1 V F
Then we have a SiISS result for the acyclic graph case.
i . This completes the proof.
(cid:3)
Theorem 5.3 Assume that GF ([0, +∞)) is acyclic. Then system (6) is SiISS if and only if
Gσ(t) is JLC.
Furthermore, consider the following inequality
Z +∞
0
z(t)dt < ∞.
(36)
It is not hard to obtain the following results based on Theorems 5.1, 5.2, and 5.3. The proofs
are omitted for space limitations.
Corollary 5.1 System (6) achieves the set tracking if (36) holds and Gσ(t) is UJLC.
Corollary 5.2 Suppose (36) holds with either GF
being acyclic. Then system (6) achieves the global set tracking if and only if Gσ(t) is JLC.
σ(t) being bidirectional for all t ≥ 0 or GF ([0, +∞))
Remark 5.2 In general, the condition (24) does not imply and is not implied by the condition
(36). In fact, the considered leaders converge to some points with (36), but the leaders can go to
infinity with (24). However, if z(t) is uniformly continuous in [0, +∞) (which can be guaranteed
once z(t) is bounded for t ∈ [0, +∞)), (24) will then be implied by (36) according to Barbalat's
Lemma.
Remark 5.3 Corollaries 5.1 and 5.2 are consistent with Proposition 6 in [34], where (36)
and integral-ISS together resulted in the state stability. Moreover, the two corollaries are also
18
consistent with Theorems 15 and 17 in [26], respectively, when z ≡ 0. However, different from
the limit-set-based approach given in [26], the proposed method by virtue of (43) and (50) also
provides the estimation of the convergence rate.
Remark 5.4 Theorems 4.1 and 5.1 with Remark 5.1 proved that for system (6), SISS is equiv-
alent to UJLC, which implies SiISS, while JLC is a necessary condition, namely,
SISS ⇐⇒ U J LC =⇒ SiISS =⇒ J LC.
Thus, SISS =⇒ SiISS, which is consistent with Corollary 4 of [34], where ISS implies iISS.
Moreover, Theorems 5.2 and 5.3 show that, in either bidirectional or acyclic case,
SiISS ⇐⇒ J LC.
Remark 5.5 As for set tracking (ST), Corollary 4.1 shows that
U J LC ⇐⇒ ST,∀z(t) satifying (24).
Moreover, Corollaries 5.1 and 5.2 show that as long as (36) holds,
in general directed cases, and
U J LC =⇒ ST
J LC ⇐⇒ ST
in either bidirectional or acyclic case. Usually, SISS goes with (24) and SiISS with (36), con-
sistent with discussions on ISS and iISS [34, 35]. Additionally, it is worth pointing out that the
differences between the statements in Corollaries 4.1 and those in 5.1 result from the fact that
UJLC is necessary for SISS, but not necessary to SiISS.
Although our results are consistent with the results on conventional ISS or iISS, the analysis
methods given in [34, 35] are mainly based on an equivalent ISS-Lyapunov function, which
cannot be applied to our cases with a moving set and switching topologies.
5.2 Proofs
To establish the SiISS in the JLC case, we will analyze the impact of the integral of input z(t) in a
time interval and estimate the convergence rates during this time interval by "pasting" different
time subintervals together within the interval. The following lemmas are given to estimate the
convergence rates in different cases.
19
Lemma 5.2 If there is an arc (j, i) leaving from j ∈ VL entering i ∈ VF in Gσ(t) for t ∈
[t0, t0 + τD), then there exists a strictly decreasing function δ(s) : [0, τD] 7→ (0, 1] with δ(0) = 1
such that
xi(t)L(y(t)) ≤ δ(t − t0)x(t0)L(y(t0)) + 2√2Z t0+τD
t0
z(s)ds, t ∈ [t0, t0 + τD].
(37)
Proof: According to Lemma 3.2, ψj(t) ≤pΨ(t) ≤pΨ(t0) +R t
t > t0 > 0. Since there is an arc (j, i) with j ∈ VL, i ∈ VF in Gσ(t) for t ∈ [t0, t0 + τD),
√2z(s)ds, j = 1, ..., n for any
t0
d
dt
ψi(t) ≤ −2b∗ψi(t) + 2√2z(t)pψi(t) + 2hxi − PL(y(t))(xi), Xj∈Ni(σ(t))
hxi(t) − PL(y(t))(xi(t)), xj(t) − xi(t)i ≤ pψi(t)(pΨ(t) −pψi(t))
Based on Lemma 2.1, when t ∈ [t0, t0 + τD),
aij(x)(xj − xi)i.
≤ pψi(t)(pΨ(t0) +Z t
t0
√2z(s)ds −pψi(t))
Therefore,
d
dt
ψi(t) ≤ −2[b∗ + (n − 1)a∗]ψi(t) + 2[√2z(t) + (n − 1)a∗(pΨ(t0) +Z t
dtpψi(t) ≤ −λpψi(t) + [√2z(t) + (n − 1)a∗(pΨ(t0) +Z t0+τD
d
t0
t0
or equivalently,
√2z(s)ds)]pψi(t),
√2z(s)ds)]
where λ , b∗ + (n − 1)a∗ for t ∈ [t0, t0 + τD). Thus,
b∗ + 2(n − 1)a∗
pψi(t) ≤ δ(t − t0)pΨ(t0) +
λ
Z t0+τD
t0
√2z(s)ds,
t ∈ [t0, t0 + τD)
with δ(s) , b∗e−λs+(n−1)a∗
λ
, s ∈ [0, τD], which implies the conclusion.
(cid:3)
Lemma 5.3 Suppose there is an edge (i, m) leaving from i ∈ VF entering m ∈ VF in Gσ(t) and
xi(t)L(y(t)) ≤ δ0x(t0)L(y(t0)) + c0 with constants δ0 ∈ (0, 1) and c0 > 0 when t ∈ [t0, t0 + τD).
Then there is a strictly decreasing function ϕδ0(s) : [0, τD] 7→ (0, 1] with ϕδ0(0) = 1 such that
xm(t)L(y(t)) ≤ ϕδ0(t − t0)x(t0)L(y(t0)) + c0 + 2√2Z t0+τD
t0
z(s)ds, t ∈ [t0, t0 + τD].
(38)
Lemma 5.4 Given a constant T > 0, if there is t1 ≥ t0 with kxi(t1)kL(y(t1)) ≤ ε0x(t0)L(y(t0)) +
c0 for constants ε0 ∈ (0, 1) and c0 > 0, then there is a strictly increasing function φε0(s) :
[0, T ] 7→ [ε0, 1) with φε0(0) = ε0 such that
xi(t)L(y(t)) ≤ φε0(t − t1)x(t0)L(y(t0)) + c0 + 2√2Z t1+ T
t0
z(s)ds, t ∈ [t1, t1 + T ],
(39)
where φε0(s) = 1 − e−(n−1)a∗ s(1 − ε0).
20
The proofs of Lemmas 5.3 and 5.4 are similar to that of Lemma 5.2, and therefore, omitted.
Lemma 5.5 Suppose V 1
entering V 1
c0, ∀i ∈ V 1
F ⊂ VF is an nonempty subset. If there are no arcs leaving from VF \V 1
F in G([t1, t1 + T )) for a given constant T > 0 and kxi(t1)kL(y(t1)) ≤ ε0x(t0)L(y(t0)) +
F for constants ε0 ∈ (0, 1) and c0 > 0, then
F
xi(t)L(y(t)) ≤ ε0x(t0)L(y(t0)) + c0 + √2Z t1+t
t1
z(s)ds,
(40)
Taking Ψ1(t) = max
i∈V 1
F{ψi(t)} gives D+pΨ1(t) ≤ √2z(t) for t ∈ [t1, t1 + T ] by virtue of the
analysis given for Lemma 3.2. Then Lemma 5.5 can be obtained straightforwardly.
Now we are ready to prove Theorems 5.1, 5.2, and 5.3.
Proof of Theorem 5.1: Denote T∗ = nT0 with T0 = T + 2τD defined in Lemma 4.3.
If
G([t∗, t∗ + T0)) is L-connected, there has to be an arc (j1, i1) ∈ Eσ(t) for t ∈ [t1, t1 + τD) ⊂
[t∗, t∗ + T0) leaving from a leader j1 ∈ VL entering i1 ∈ VL and this arc is kept there for a period
of at least τD. Invoking Lemmas 5.2 and 5.4,
xi1(t)L(y(t)) ≤ c1x(t1)L(y(t1)) + 4√2Z t∗+T∗
t∗
z(s)ds,
t ∈ [t1, t∗ + T∗],
where c1 = φδ(τD )(T∗).
Furthermore, when t ∈ [t∗ + T0, t∗ + 2T0), there must be a follower i2 ∈ VF , i2 6= i1 such
that there exists an arc (j2, i2) for some j2 ∈ VL, or an arc (i1, i2) when t ∈ [t2, t2 + τD) ⊂
[t∗ + T0, t∗ + 2T0)). According to Lemmas 5.3 and 5.4,
xi2(t)L(y(t)) ≤ c2x(t1)L(y(t1)) + 8√2Z t∗+T∗
t∗
z(s)ds, t ∈ [t2, t∗ + T∗],
where c2 = φϕ2(T∗) with ϕ2 = ϕc1(τD).
Repeating the above procedure yields
xiℓ(t)L(y(t)) ≤ cℓx(t1)L(y(t1)) + 4√2ℓZ t∗+T∗
t∗
z(s)ds, t ∈ [t∗ + ℓT0, t∗ + T∗].
for iℓ ∈ VF , ℓ = 3, . . . , n, where
cℓ = φϕℓ−1(T∗), ϕℓ = ϕcℓ−1(τD), [tℓ, tℓ + τD) ⊂ [t∗ + (ℓ − 1)T0, t∗ + ℓT0)
(41)
Moreover, the nodes of iℓ, ℓ = 1, 2, . . . , n are distinct.
Denote c , cn from (41). Then we obtain
x(t∗ + T∗)L(y(t∗+T∗)) ≤ cx(t∗)L(y(t∗)) + (4n + 1)√2Z t∗+T∗
t∗
z(s)ds.
(42)
21
It follows immediately that
x(KT∗)L(y(KT∗)) ≤ cKx0L(y0) + (4n + 1)√2
KXj=1
Z jT∗
(j−1)T∗
cK−jz(s)ds, K = 1, 2, . . .
(43)
Based on Lemma 3.2 and (12), we have
Z jT∗
T ∗ ⌋−jz(s)ds + √2Z t
c⌊ t
(j−1)T∗
c⌊ t
T ∗ ⌋−p(s)z(s)ds
⌊ t
T ∗ ⌋
z(s)ds
(44)
(45)
x(t)L(y(t) ≤ c⌊ t
≤ c⌊ t
where
⌊ t
T ∗ ⌋Xj=1
T ∗ ⌋x0L(y0) + (4n + 1)√2
T ∗ ⌋x0L(y0) + (4n + 1)√2Z t
p(s) =
⌊ t
T ∗⌋,
i,
0
s ∈ [(i − 1)T ∗, iT ∗) f or i = 1, . . . ,⌊ t
T ∗⌋
s ∈ [T ∗ · ⌊ t
T ∗⌋, t)
Hence, (10) holds with γ(s) = (4n + 1)√2s since c⌊ t
T ∗ ⌋−p(s) ≤ 1, which completes the proof. (cid:3)
Proof of Theorem 5.2: The "only if" part is quite obvious, so we focus on the "if" part.
Since Gσ(t) is JLC, there exists a sequence of time instants
0 = T1 < T2 < ··· < Ti < Ti+1 < . . .
such that
Ti , Ti1 < Ti2 < ··· < Tin+1 = Ti+1, i = 1, 2, . . .
(46)
(47)
and G([Tiκ , Tiκ+1)) is L-connected for κ = 1, . . . , n. Moreover, each arc in G([Tik , Tik+1)) will
be kept for at least the dwell time τD during the time interval [Tiκ , Tiκ+1), i = 1, 2, . . . ; κ =
1, 2, . . . , n.
Then we estimate Ψ(t) during t ∈ [Ti, Ti+1]. Since G([Ti1 , Ti2 )) is L-connected, there is a
time interval [t1, t1 + τD) ⊆ [Ti1 , Ti2) such that there is an edge (l, m0) ∈ Eσ(t) between a leader
l ∈ VL and a follower m0 ∈ VF for t ∈ [t1, t1 + τD). Based on Lemma 5.2,
xm0(t1 + τD)L(y(t1+τD)) ≤ δ1x(t1)L(y(t1)) + 2√2Z t1+τD
z(s)ds,
t1
where δ1 , δ(τD).
Furthermore, we define V 1
L
, {vm0} ∪ VL,
t2 , inf
t {t ∈ [t1 + τD, Ti3)there is an edge leaving from V 1
L entering V \ V 1
L in Gσ(t)},
and V 1
F
, { ∈ VF \ m0there is an edge leaving from V 1
L entering when t = t2}.
22
Noting that G([Ti2 , Ti3)) is L-connected, thus, according to Lemma 5.5, one has
xm0 (t2)L(y(t2)) ≤ δ1x(t1)L(y(t1)) + √2Z t2
t1+τD z(s)ds + 2√2Z t1+τD
t1
z(s)ds
Further, by Lemma 5.4,
xm0 (t)L(y(t)) ≤ φ1x(t1)L(y(t1)) + √2[2Z t2+τD
≤ φ1x(t1)L(y(t1)) + 4√2Z t2+τD
t1
t1
+2Z t1+τD
t1
]z(s)ds
+Z t2
z(s)ds
t1+τD
(48)
for t ∈ [t2, t2 + τD], where φ1 , φδ1(τD). Moreover, according to Lemma 5.3,
x(t2 + τD)L(y(t2+τD)) ≤ ϕφ1(τD)x(t1)L(y(t1)) + √2[4Z t2+τD
+2Z t2+τD
≤ ϕφ1(τD)x(t1)L(y(t1)) + 8√2Z t2+τD
z(s)ds
t1
t2
t1
+2Z t2
t1
]z(s)ds
(49)
for ∈ V 1
F . Because φ1 < ϕφ1(τD), (48) and (49) lead to
xi(t2 + τD)L(y(t2+τD)) ≤ δ2x(t1)L(y(t1)) + 8√2Z t2+τD
t1
z(s)ds, ∀i ∈ {vm0} ∪ V 1
F ,
where δ2 , ϕφ1(τD).
Next, define V 2
L
, V 1
L ∪ V 1
F ,
t3 , inf
t {t ∈ [t2 + τD, Ti4)there is an edge leaving from V 2
L and entering V \ V 2
L in Gσ(t)}
and V 2
F
, { ∈ V \ V 2
Lthere is an edge leaving from V 2
L entering when t = t3}.
Similarly, from Lemma 5.5, by φ2 , φδ2(τD), δ3 , ϕφ2(τD), one has
xi(t3 + τD)L(y(t3+τD)) ≤ δ3x(t1)L(y(t1)) + 12√2Z t3+τD
t1
z(s)ds,
Repeating the process gives
φκ , φδκ(τD),
δκ+1 , ϕφκ(τD),
∀i ∈ {vm0} ∪ V 1
F ∪ V 2
F
for κ = 3, 4, . . . , k0 until VF = {vm0} ∪ V 1
F ∪ V 2
F ∪ ··· ∪ V k0
F for some k0 ≤ n such that
xi(tk0 + τD)L(y(tk0 +τD)) ≤ δk0x(t1)L(y(t1)) + 4√2k0Z tk0 +τD
t1
z(s)ds,
∀i ∈ VF
Hence
x(tk0 + τD)L(y(tk0 +τD)) ≤ δk0x(t1)L(y(t1)) + 4√2k0Z tk0 +τD
t1
z(s)ds
23
According to Lemma 5.5, we obtain
x(Ti+1)L(y(Ti+1)) ≤ δk0x(Ti)L(y(Ti)) + (4k0 + 1)√2Z Ti+1
Ti
z(s)ds.
It is obvious to see that k0 ≤ n and 0 < δ1 ≤ δ2 ≤ . . . δn < 1. Therefore, denote δ , δn, then for
K = 1, 2, . . . ,
x(TK+1)L(y(TK+1)) ≤ δKx0L(y0) + (4n + 1)√2
δK−iZ Ti+1
Ti
KXi=1
z(s)ds
Thus, similar to the proof of Theorem 5.1, we also have
x(t)L(y(t) ≤ δΓ(t)x0L(y0) + (4n + 1)√2Z t
0
δΓ(t)−p(s)z(s)ds
where Γ(t) = K0 − 1 when t ∈ [TK0 , TK0+1), K0 = 1, 2, . . . , and
p(s) =
i,
s ∈ [Ti, Ti+1) f or i = 1, . . . , K0 − 1
K0 − 1,
s ∈ [TK0 , t)
Then it is obvious to see that (51) leads to Theorem 5.2 immediately.
(50)
(51)
(52)
(cid:3)
Proof of Theorem 5.3: We also focus on the "if" part since the "only if" part is quite obvious.
Because Gσ(t) is JLC, there is an infinite sequence in the form of (46) with (47) such that
G([Tiκ , Tiκ+1)) is L-connected for κ = 1, . . . , n.
Then, for any ℓ ∈ V1, there is tℓ ∈ [Ti1 , Ti2) such that there is an arc leaving from VF entering
ℓ in Gσ(t). Hence, recalling Lemma 5.2,
xℓ(tℓ + τD)L(y(tℓ+τD)) ≤ d1x(tℓ)L(y(tℓ)) + 2√2Z tℓ+τD
with a constant d1 , δ(τD). According to Lemma 5.1, for any ℓ ∈ V F
tℓ
xℓ(t)L(y(t)) ≤ d1x(Ti)L(y(Ti)) + 2√2Z Ti+1
Ti
Again by Lemmas 5.3 and 5.1, for any ℓ ∈ V F
2 ,
xℓ(t)L(y(t)) ≤ d2x(Ti)L(y(Ti)) + 4√2Z Ti+1
Ti
z(s)ds, ℓ ∈ V F
1
1 , we have
z(s)ds, t ∈ [Ti2 , Ti+1]
z(s)ds, t ∈ [Ti3 , Ti+1],
where d2 = ϕd1(τD). Similarly, with dj = ϕdj−1 (τD), j = 3, . . . , k0,
xℓ(t)L(y(t)) ≤ djx(Ti)L(y(Ti)) + 2√2jZ Ti+1
Ti
z(s)ds, t ∈ [Tij+1 , Ti+1],
for any ℓ ∈ V F
j , j = 3, . . . , k0, which leads to
x(Ti+1)L(y(Ti+1)) ≤ dk0x(Ti)L(y(Ti)) + 2√2k0Z Ti+1
Ti
z(s)ds.
Similar to the proof of Theorem 5.2, SiISS can be obtained.
(cid:3)
24
6 Conclusions
This paper addressed multi-agent set tracking problems with multiple leaders and switching
communication topologies. At first, the equivalence between UJLC and the SISS of a group
of uncertain agents with respect to a moving multi-leader set was shown. Then it was shown
that UJLC is a sufficient condition for SiISS of the multi-agent system with disturbances in
agent dynamics and unmeasurable velocities in the dynamics of the leaders. Moreover, when
communication topologies are either bidirectional or acyclic, JLC is a necessary and sufficient
condition for SiISS. Also, set tracking was achieved in special cases with the help of SISS and
SiISS.
Multiple leaders, in some practical cases, can provide an effective way to overcome the
difficulties and constraints in the distributed design. On the other hand, ISS-based tools were
proved to be very powerful in the control synthesis. Therefore, the study of multiple active
leaders and related ISS tools deserves more attention.
Appendix
A.1 Proof of Lemma 4.1
Due to D+pΨ(t) ≤ √2kzk∞ by Lemma 3.2 and (12), we obtain
qψj(t) ≤pΨ(t) ≤pΨ(t0) + √2kzk∞(t − t0), j = 1, ..., n.
(53)
Since there is an arc (j, i) with j ∈ VL and i ∈ VF in Gσ(t) for t ∈ [t0, t0 + τD), based on (20),
one has
hxi − PL(y(t))(xi), Xj∈Li(σ(t))
bij(x)(yj − xi)i ≤ −b∗ψi(t).
(54)
d
dt
Thus, with (19) and the fact that r(t) + wi(t) ≤ q(t) ≤ √2kzk∞, we obtain
ψi(t) ≤ −2b∗ψi(t) + 2hxi − PL(y(t))(xi), Xj∈Ni(σ(t))
≤ −2b∗ψi(t) + 2hxi − PL(y(t))(xi), Xj∈Ni(σ(t))
aij(xj − xi)i + 2(r(t) + wi(t))pψi(t)
aij(xj − xi)i + 2√2kzk∞pψi(t)
(55)
for t ∈ [t0, t0 + τD).
Then, by Lemma 2.1, if pψj(t) <pψi(t), j ∈ Ni(σ(t)) for t ∈ [t0, t0 + τD), then
hxi(t) − PL(y(t))(xi(t)), xj (t) − xi(t)i ≤ 0.
(56)
25
On the other hand, if pψj(t) ≥pψi(t), j ∈ Ni(σ(t)), from Lemma 2.1 and (53),
hxi(t) − PL(y(t))(xi(t)), xj (t) − xi(t)i ≤ pψi(t)(qψj(t) −pψi(t))
≤ pψi(t)(pΨ(t0) + √2kzk∞(t − t0) −pψi(t))
≤ pψi(t)(pΨ(t0) −pψi(t) + √2kzk∞τD)
(57)
t ∈ [t0, t0 + τD). Therefore, with (55), (56) and (57), it follows that
d
dt
ψi(t) ≤ −2λψi(t) + 2[√2kzk∞(1 + (n − 1)a∗τD) + (n − 1)a∗pΨ(t0)]pψi(t),
where λ , b∗ + (n − 1)a∗, or equivalently,
d
dtpψi(t) ≤ −λpψi(t) + [√2kzk∞(1 + (n − 1)a∗τD) + (n − 1)a∗pΨ(t0))]
for t ∈ [t0, t0 + τD). As a result,
pψi(t) ≤ e−λ(t−t0)pΨ(t0) + (1 − e−λ(t−t0))
≤ µ(t − t0)pΨ(t0) + c0√2kzk∞,
(n − 1)a∗pΨ(t0) + (kuk∞ + kwk∞)(1 + (n − 1)a∗τD)
λ
t ∈ [t0, t0 + τD)
(58)
b∗+(n−1)a∗
where µ(s) , b∗e−λs+(n−1)a∗
, s ∈ [0, τD] and c0 , 1+(n−1)a∗τD
b∗+(n−1)a∗ , because 1 − e−λ(t−t0) < 1.
between the followers and the leaders. Similar analysis gives
Then we evaluate pψi(t) for t ∈ [t0 + τD, t0 + T∗) no matter whether there is any connection
ψi(t) ≤ 2√2kzk∞pψi(t) + 2hxi − PL(y(t))(xi), Xj∈Ni(σ(t))
aij(x)(xj − xi)i
d
dt
≤ 2√2kzk∞pψi(t) + 2(n − 1)a∗pψi(t)(pΨ(t0) −pψi(t) + √2kzk∞T∗)
= −2(n − 1)a∗ψi(t) + 2[√2kzk∞(1 + (n − 1)a∗T∗) + (n − 1)a∗pΨ(t0)]pψi(t),
which is equivalent to
d
dtpψi(t) ≤ −(n − 1)a∗pψi(t) + [√2kzk∞(1 + (n − 1)a∗T∗) + (n − 1)a∗pΨ(t0)].
(59)
Denote µ∗ , µ(τD). From (58), when t ∈ [t0 + τD, t0 + T∗),
pψi(t) ≤ e−(n−1)a∗(t−(t0+τD))pψi(t0 + τD)
+(1 − e−(n−1)a∗(t−(t0+τD)))[pΨ(t0) + √2kzk∞
≤ e−(n−1)a∗(t−(t0+τD))[µ∗pΨ(t0) + c0√2kzk∞]
+(1 − e−(n−1)a∗(t−(t0+τD)))[pΨ(t0) + √2kzk∞ ·
≤ µ(t − t0)pΨ(t0) + γ1kzk∞,
1 + (n − 1)a∗T∗
]
(n − 1)a∗
1 + (n − 1)a∗T∗
(n − 1)a∗
]
(60)
26
where γ1 , √2· 1+(n−1)a∗ T∗
(n−1)a∗ > c0 and µ(s) , 1− e−(n−1)a∗(s−τD))(1− µ∗), s ∈ [τD, T∗]. Therefore,
based on (58) and (60),
pψi(t) ≤ µ(t − t0)pΨ(t0) + γ1kzk∞, µ(s) =
where µ(s) is continuous. Thus, the conclusion follows.
µ(s), s ∈ [0, τD)
µ(s), s ∈ [τD, T∗]
(cid:3)
A.2 Proof of Lemma 4.2
If there is an arc (vi, vm) in Gσ(t) for t ∈ [t0, t0 + τD), then based on (53), Lemmas 2.1 and
3.1, it is easy to see
d
dt
ψm(t) ≤ 2hxm − PL(y(t))(xm), Xj∈Nm(σ(t))
amj(x)(xj − xm) + Xj∈Lm(σ(t))
bmj(x)(yj − xm)i
+2√2kzk∞pψm(t)
≤ 2√2kzk∞pψm(t) + 2hxm − PL(y(t))(xm), Xj∈Nm(σ(t))
= 2√2kzk∞pψm(t) + 2 Xj∈Nm(σ(t))\vi
≤ 2√2kzk∞pψm(t) + 2(n − 2)a∗pψm(t)(pΨ(t0) −pψm(t) + √2kzk∞τD)
+2ami(x)hxm − PL(y(t))(xm), xi − xmi
amj(x)hxm − PL(y(t))(xm), xj − xmi
amj(x)(xj − xm)i
−2a∗pψm(t)(pψm(t) −pψi(t))
for t ∈ [t0, t0 + τD). Then, if (22) holds, as done in the proof of Lemma 4.1, we can obtain
d
dtpψm(t) ≤ −λ1pψm(t) + d1
where λ1 , (n− 2)a∗ + a∗ and d1 , (1 + (n− 2)a∗τD)√2kzk∞ + ((n− 2)a∗ + a∗µ0)pΨ(t0) + a∗d0.
Here are two cases.
• when t ∈ [t0, t0 + τD):
d1
pψm(t) ≤ e−λ1(t−t0)pψm(t0) + (1 − e−λ1(t−t0))
λ1
(n − 2)a∗ + (µ0 + (1 − µ0)e−λ1(t−t0))a∗
√2kzk∞(1 + (n − 2)a∗τD) + a∗d0
(n − 2)a∗ + a∗
≤
·
(n − 2)a∗ + a∗
≤ ξ(t − t0)pΨ(t0) + γ0,
pΨ(t0) + (1 − e−λ1(t−t0))
where ξ(s) , (n−2)a∗+(µ0+(1−µ0)e−λ1s)a∗
(n−2)a∗+a∗
, s ∈ [0, τD] and γ0 , (1+(n−2)a∗τD)√2kzk∞+a∗d0
(n−2)a∗+a∗
27
(61)
.
• when t ∈ [t0 + τD, t0 + T∗): Denote ξ∗ , ξ(τD). By (61), similarly, we have
pψm(t) ≤ e−(n−1)a∗(t−(t0+τD))pψm(t0 + τD)
≤ e−(n−1)a∗(t−(t0+τD))[ξ∗pΨ(t0) + γ0]
≤ ξ(t − t0)pΨ(t0) + γ2kzk∞ + d0,
+(1 − e−(n−1)a∗(t−(t0+τD)))[pΨ(t0) + √2kzk∞
+(1 − e−(n−1)a∗(t−(t0+τD)))[pΨ(t0) + √2kzk∞
1 + (n − 1)a∗T∗
(n − 1)a∗
1 + (n − 1)a∗T∗
(n − 1)a∗
]
]
(62)
where γ2 , √2 · 1+(n−1)a∗T∗
(n−2)a∗+a∗
max{γ0,√2kzk∞
and ξ(s) , 1 − e−(n−1)a∗(s−τD)(1 − ξ∗), s ∈ [τD, T∗], because
1 + (n − 1)a∗T∗
+ d0,
1 + (n − 1)a∗T∗
(n − 2)a∗ + a∗
√2kzk∞
} ≤
(n − 1)a∗
With (61) and (62), we have
pψm(t) ≤ ξµ0(t − t0)pΨ(t0) + γ2kzk∞ + d0,
where ξµ0(s) which is continuous. Thus, the conclusion follows.
ξµ0(s) =
ξ(s), s ∈ [0, τD)
ξ(s), s ∈ [τD, T∗]
(cid:3)
References
[1] J. Aubin and A. Cellina. Differential Inclusions. Berlin: Speringer-Verlag, 1984
[2] R. T. Rockafellar. Convex Analysis. New Jersey: Princeton University Press, 1972.
[3] N. Rouche, P. Habets, and M. Laloy. Stability Theory by Liapunov's Direct Method, New
York: Springer-Verlag, 1977.
[4] J. Danskin. The theory of max-min, with applications, SIAM J. Appl. Math., vol. 14,
641-664, 1966.
[5] C. Berge and A. Ghouila-Houri. Programming, Games, and Transportation Networks, John
Wiley and Sons, New York, 1965.
[6] D. Cheng, J. Wang, and X. Hu, An extension of LaSalle's invariance principle and its
applciation to multi-agents consensus, IEEE Trans. Automatic Control, 53, 1765-1770, 2008.
[7] F. Clarke, Yu.S. Ledyaev, R. Stern, and P. Wolenski, Nonsmooth Analysis and Control
Theory. Speringer-Verlag, 1998
28
[8] I. D. Couzin, J. Krause, N. Franks, and S. Levin. Effective leadership and decision making
in animal groups on the move. Nature, vol. 433, 513-516, 2005.
[9] J. Fang, A. S. Morse, and M. Cao, Multi-agent rendezvousing with a finite set candidate
rendezvous points, Proc. American Control Conference, 765-770, 2008.
[10] S. Martinez, J. Cortes, and F. Bullo. Motion coordination with distributed information,
IEEE Control Systems Magazine, vol. 27, no. 4, 75-88, 2007.
[11] W. Ren and R. Beard, Distributed Consensus in Multi-vehicle Cooperative Control,
Springer-Verlag, London, 2008.
[12] R. Olfati-Saber, Flocking for multi-agent dynamic systems: algorithms and theory, IEEE
Trans. Automatic Control, 51(3): 401-420, 2006.
[13] C. Godsil and G. Royle. Algebraic Graph Theory. New York: Springer-Verlag, 2001.
[14] Y. Hong, L. Gao, D. Cheng, and J. Hu. Lyapuov-based approach to multi-agent systems
with switching jointly connected interconnection. IEEE Trans. Automatic Control, vol. 52,
943-948, 2007.
[15] R. Olfati-Saber and R. Murray. Consensus problems in the networks of agents with switching
topology and time dealys, IEEE Trans. Automatic Control, vol. 49, no. 9, 1520-1533, 2004.
[16] Y. Hong, J. Hu, and L. Gao. Tracking control for multi-agent consensus with an active
leader and variable topology. Automatica, vol. 42, 1177-1182, 2006.
[17] H. G. Tanner, A. Jadbabaie, and G. Pappas, Stable flocking of mobile agents, Part I: fixed
topology, Proc. IEEE Conf. on Decision and Control, Hawaii, 2010-2015, Dec. 2003
[18] H. G. Tanner, A. Jadbabaie, and G. Pappas, Stable flocking of mobile agents, Part II:
dynamic topology, Proc. IEEE Conf. on Decision and Control, Hawaii, 2016-2021, Dec.
2003
[19] H. G. Tanner, G. Pappas and V. Kumar, Leader-to-formation stability, IEEE Transactions
on Robotics and Automation, 20(3), 443-455, 2004
[20] H. G. Tanner, A. Jadbabaie, G. J. Pappas, Flocking in fixed and switching networks, IEEE
Trans. Automatic Control, 52(5): 863-868, 2007.
29
[21] F. Xiao and L. Wang, State consensus for multi-agent systems with swtiching topologies
and time-varying delays, Int. J. Control, 79, 10, 1277-1284, 2006.
[22] A. Jadbabaie, J. Lin, and A. S. Morse. Coordination of groups of mobile agents using
nearest neighbor rules. IEEE Trans. Automatic Control, vol. 48, no. 6, 988-1001, 2003.
[23] Z. Lin, and B. Francis, M. Maggiore. Necessary and sufficient graphical conditions for
formation control of unicycles, IEEE Trans. Automatic Control, 50(1): 121-127, 2005.
[24] M. Ji, G. Ferrari-Trecate, M. Egerstedt, and A. Buffa, Containment control in mobile
networks, IEEE Trans. Automatic Control, 53, no. 8, 1972-1975, 2008.
[25] C. C. Cheah, S. P. Hou, and J. J. E. Slotine, Region following formation control for multi-
robot systems, Proc. of IEEE Int. Conf. Robotics and Automation, pp. 3796-3801, 2008.
[26] G. Shi and Y. Hong, Global target aggregation and state agreement of nonlinear multi-agent
systems with switching topologies, Automatica, vol. 45, 1165-1175, 2009.
[27] Y. Cao and W. Ren, Containment control with multiple stationary or dynamic leaders
under a directed interaction graph, Proc. of Joint 48th IEEE Conf. Decision & Control/28th
Chinese Control Conference, Shanghai, China, Dec. 2009, pp. 3014-3019.
[28] J. Tsitsiklis, D. Bertsekas, and M. Athans. Distributed asynchronous deterministic and
stochastic gradient optimization algorithms, IEEE Trans. Automatic Control, 31, 803-812,
1986.
[29] Z. Jiang, A. Teel, and L. Praly, Small-gain theorem for ISS systems and applications,
Mathematics of Control, Signals, and Systems, 7, 95-120, 1994
[30] Z. Lin, B. Francis, and M. Maggiore. State agreement for continuous-time coupled nonlinear
systems. SIAM J. Control Optim., vol. 46, no. 1, 288-307, 2007.
[31] L. Moreau, Stability of multiagent systems with time-dependent communication links, IEEE
Trans. Automatic Control, 50, 169-182, 2005.
[32] L. Scardovi, M. Arcak, and E. Sontag, Synchronization of interconnected systems with
an input-output approach-I: main results, Proc. of Joint 48th IEEE Conf. Decision &
Control/28th Chinese Control Conference, Shanghai, China, Dec. 2009, pp. 609-614.
30
[33] E. Sontag and Y. Lin, Stabilization with respect to noncompact sets: Lyapunov charac-
terizations and effect of bounded inputs, Proc. Nonlinear Control Systems Design Symp.,
Bordeaus, IFAC Publications, 9-14, June 1992 (M.Fliess, Ed.)
[34] E. Sontag, Comments on integral variants of ISS, Systems & Control Letters, 34, 93-100,
1998
[35] E. Sontag and Y. Wang, On characterizations of the input-to-state stability property, Sys-
tems & Control Letters, 24, 351-359, 1995.
31
|
cs/0306119 | 1 | 0306 | 2003-06-20T15:12:22 | A Method for Solving Distributed Service Allocation Problems | [
"cs.MA"
] | We present a method for solving service allocation problems in which a set of services must be allocated to a set of agents so as to maximize a global utility. The method is completely distributed so it can scale to any number of services without degradation. We first formalize the service allocation problem and then present a simple hill-climbing, a global hill-climbing, and a bidding-protocol algorithm for solving it. We analyze the expected performance of these algorithms as a function of various problem parameters such as the branching factor and the number of agents. Finally, we use the sensor allocation problem, an instance of a service allocation problem, to show the bidding protocol at work. The simulations also show that phase transition on the expected quality of the solution exists as the amount of communication between agents increases. | cs.MA | cs |
Web Intelligence and Agent Systems: An International Journal, 2003.
A Method for Solving Distributed Service
Allocation Problems ∗
Jos´e M. Vidal
Computer Science and Engineering
University of South Carolina
Columbia, SC 29208
[email protected]
April 17, 2003
Abstract
We present a method for solving service allocation problems in which
a set of services must be allocated to a set of agents so as to maximize a
global utility. The method is completely distributed so it can scale to any
number of services without degradation. We first formalize the service
allocation problem and then present a simple hill-climbing, a global hill-
climbing, and a bidding-protocol algorithm for solving it. We analyze the
expected performance of these algorithms as a function of various prob-
lem parameters such as the branching factor and the number of agents.
Finally, we use the sensor allocation problem, an instance of a service al-
location problem, to show the bidding protocol at work. The simulations
also show that phase transition on the expected quality of the solution
exists as the amount of communication between agents increases.
1 Introduction
The problem of dynamically allocating services to a changing set of consumers
arises in many applications. For example, in an e-commerce system, the service
providers are always trying to determine which service to provide to whom,
and at what price [5]; in an automated manufacturing for mass customization
scenario, agents must decide which services will be more popular/profitable [1];
and in a dynamic sensor allocation problem, a set of sensors in a field must
decide which area to cover, if any, while preserving their resources.
While these problems might not seem related, they are instances of a more
general service allocation problem in which a finite set of resources must be
allocated by a set of autonomous agents so as to maximize some global measure
∗This work has been funded by Darpa under contract F30602-99-2-0513.
1
of utility. A general approach to solving these types of problems has been used
in many successful systems , such as [2] [3] [11] [9]. The approach involves three
general steps:
1. Assign each resource that needs to be preserved to an agent responsible
for managing the resource.
2. Assign each goal of the problem domain to an agent responsible for achiev-
ing it. Achieving these goals requires the consumption of resources.
3. Have each agent take actions so as to maximize its own utility, but im-
plement a coordination algorithm that encourages agents to take actions
that also maximize the global utility.
In this paper we formalize this general approach by casting the problem as
a search in a global fitness landscape which is defined as the sum of the agents'
utilities. We show how the choice of a coordination/communication protocol
disseminates information, which in turn "smoothes" the global utility landscape.
This smooth global utility landscape allows the agents to easily find the global
optimum by simply making selfish decisions to maximize their individual utility.
We also present experiments that pinpoint the location of a phase transition
in the time it takes for the agents to find the optimal allocation. The transi-
tion can be seen when the amount of communication allowed among agents is
manipulated. It exists because communication allows the agents to align their
individual landscapes with the global landscape. At some amount of commu-
nication, the alignment between these landscapes is good enough to allow the
agents to find the global optimum, but less communication drives the agents
into a random behavior from which the system cannot recuperate.
1.1 Task Allocation
The service allocation problem we discuss in this paper is a superset of the well
known task allocation problem [10, chapter 5.7]. A task allocation problem is
defined by a set of tasks that must be allocated among a set of agents. Each
agent has a cost associated with each subset of tasks, which represents the cost
the agent would incur if it had to perform those tasks. Coordination proto-
cols are designed to allow agents to trade tasks so that the globally optimal
allocation -- the one that minimizes the sum of all the individual agent costs -- is
reached as soon as possible. It has been shown that this globally optimal allo-
cation can reached if the agents use the contract-net protocol [9] with OCSM
contracts [8]. These OCSM contracts make it possible for the system to transi-
tion from any allocation to any other allocation in one step. As such, a simple
hill-climbing search is guaranteed to eventually reach the global optimum.
In this paper we consider the service allocation problem, which is a superset
of the task allocation because it allows for more than one agent to service a
"task". The service allocation problem we study also has the characteristic that
every allocation cannot be reached from every other allocation in one step.
2
1.2 Service Allocation
In a service allocation problem there are a set of services, offered by service
agents, and a set of consumers who use those services. A server can provide any
one of a number of services and some consumers will benefit from that service
without depleting it. A server agent incurs a cost when providing a service and
can choose not to provide any service.
For example, a server could be an agent that sets up a website with infor-
mation about cats. All the consumer agents with interests in cats will benefit
from this service, but those with other interests will not benefit. Since each
server can provide, at most, one service, the problem is to find the allocation of
services that maximizes the sum of all the agents' utilities, that is, an allocation
that maximizes the global utility.
1.2.1 Sensor Allocation
Another instance of the service allocation problem is the sensor allocation prob-
lem, which we will use as an example throughout this paper.
In the sensor
allocation problem we have a number of sensors placed in fixed positions in a
two-dimensional space. Each sensor has a limited viewing angle and distance
but can point in any one of a number of directions. For example, a sensor might
have a viewing angle of 120 degrees, viewing distance of 3 feet, and be able
to look in three directions, each one 120 degrees apart from the others. That
is, it can "look" in any one of three directions. On each direction it can see
everything that is in the 120 degree and 3 feet long view cone. Each time a
sensor looks in a particular direction is uses energy.
There are also targets that move around in the field. The goal is for the
sensors to detect and track all the targets in the field. However, in order to
determine the location of a target, two or more sensors have to look at it at
the same time. We also wish to minimize the amount of energy spent by the
sensors.
We consider the sensor agents as being able to provide three services, one for
each sector, but only one at a time. We consider the target agents as consuming
the services of the sensors.
2 A Formal Model for Service Allocation
We define a service allocation problem SA as a tuple SA = {C, S} where C is
the set of consumer agents C = {c1, c2, . . . , cC}, and ci has only one possible
state, ci = 0. The set of service agents is S = {s1, s2, . . . , sS} and the
value of si is the value of that service. For the sensor domain in which a
sensor can observe any one of three 120-degree sectors or be turned off, we have
si ∈ {0, 1, 2, off}. An allocation is an assignment of states to the services (since
the consumers have only one possible state we can ignore them). A particular
allocation is denoted by a = {s1, s2, . . . , sS}, where the si have some value
taken from the domain of service states, and a ∈ A, where A is the set of all
3
possible allocations. That is, an allocation tells us the state of all agents (since
consumers have only one state they can be omitted).
Each agent also has a utility function. The utility that an agent receives
depends on the current allocation a, where we let a(s) be the state of service
agent s under a. The agent's utilities will depend on their state and the state of
other agents. For example, in the sensor problem we define the utility of sensor
s as Us(a), where
Us(a) = (cid:26) 0
−K1
if a(s) = off
otherwise.
(1)
That is, a sensor receives no utility when it is off and must pay a penalty of
−K1 when it is running.
The targets are the consumers, and each target's utility is defined as
Uc(a) =
if fc(a) = 0
0
K2
if fc(a) = 1
K2 + n − 2 if fc(a) = n
(2)
where
fc(a) = number of sensors s that see c given their state a(c).
(3)
Finally, given the individual agent utilities, we define the global utility
GU (a) as the sum of the individual agents' utilities:
GU (a) = Xc∈C
Uc(a) +Xs∈S
Us(a).
(4)
The service allocation problem is to find the allocation a that maximizes
GU (a). In the sensor problem, there are 4S possible allocations, which would
make a simple generate-and-test approach take exponential amounts of time.
We wish to find the global optimum much faster than that.
2.1 Search Algorithms
Our goal is to design an interaction protocol whereby an allocation a that max-
imizes the global utility GU (a) is reached in a small number of steps. In each
step of our protocol one of the agents will change its state or send a message to
another agent. The messages might contain the state or utilities of other agents.
We assume that the agents do not have direct access to the other agents' states
or utility values.
The simplest algorithm we can envision involves having each consumer, at
each time, changing the state of a randomly chosen service agent so as to increase
the consumer's own utility. That is, a consumer c will change the current
allocation a into a′ by changing the state of some sensor s such that Uc(a′) >
Uc(a). If the sensor's state cannot be changed so as to increase the utility, then
the consumer does nothing.
In the sensor domain this amounts to a target
4
picking a sensor and changing its state so that the sensor can see the target.
We refer to this algorithm as individual hill-climbing.
The individual hill-climbing algorithm is simple to implement and the only
communication needed is between the consumer and the chose server. This
simple algorithm makes every consumer agent increase its individual utility at
each turn. However, the new allocation a′ might result in a lower global utility,
since a′ might reduce the utility of several other agents. Therefore, it does not
guarantee that an optimal allocation will be eventually reached.
Another approach is for each agent to change state so as to increase the global
utility. We call this a global hill-climbing algorithm. In order to implement
this algorithm, an agent would need to know how the proposed state change
affects the global utility as well as the states of all the other agents. That is, it
would need to be able to determine GU (a′) which requires it to know the state
of all the agents in a′ as well as the utility functions of every other agent, as
per the definition of global utility (4). In order for an agent to know the state
of others, it would need to somehow communicate with all other agents. If the
system implements a global broadcasting method then we would need for each
agent to broadcast its state at each time. If the system uses more specialized
communications such as point-to-point, limited broadcasting, etc., then more
messages will be needed.
Any protocol that implements the global hill-climbing algorithm will reach a
locally optimal allocation in the global utility. This is because it is always true
that, for a new allocation a′ and old allocation a, GU (a′) ≥ GU (a). Whether or
not this local optimum is also a global optimum will depend on the ruggedness
of the global utility landscape. That is, if it consists of one smooth peak then
it is likely that any local optimum is the global optimum. On the other hand,
if the landscape is very rugged then there are likely many local peaks. Studies
in NK landscapes [4] tell us that smoother landscapes result when an agent's
utility depends on the state of smaller number of other agents.
Global hill-climbing is better than individual hill-climbing since it guarantees
that we will find a local optima. However, it requires agents to know each others'
utility function and to constantly communicate their state. Such large amount
of communication is often undesirable in multiagent systems. We need a better
way to find the global optimum.
One way of correlating the individual landscapes to the global utility land-
scape is with the use of a bidding protocol in which each consumer agent
tells each service the marginal utility the consumer would receive if the service
switched its state to so as to maximize the consumer's utility. The service agent
can then choose to provide the service with the highest aggregate demand. Since
the service is picking the value that maximizes the utility of everyone involved
(all the consumers and the service) without decreasing the utility of anyone else
(the other services) this protocol is guaranteed to never decrease the global util-
ity. This bidding protocol is a simplified version of the contract-net [9] protocol
in that it does not require contractors to send requests for bids.
However, in order for a consumer to determine the marginal utility it will
receive from one sensor changing state, it still needs to know the state of all
5
the other sensors. This means that a complete implementation of this protocol
will still require a lot of communication (namely, the same amount as in global
hill-climbing). We can reduce this number of messages by allowing agents to
communicate with only a subset of the other agents and making their decisions
based on only this subset of information. That is, instead of all services telling
each consumer their state, a consumer could receive state information from only
a subset of the services and make its decision based on this (assuming that the
services chosen are representative of the whole). This strategy shows a lot of
promise but its performance can only be evaluated on an instance-by-instance
basis. We explore this strategy experimentally in Section 3 using the sensor
domain.
2.1.1 Theoretical Time Bounds of Global Hill-Climbing
Since we now know that global hill-climbing will always reach a local optimum,
the next questions we must answer are:
1. How many local optima are there?
2. What is the probability that a local optimum is the global optimum?
3. How long does it take, on average, to reach a local optimum?
Let a be the current allocation and a′ be a neighboring allocation. We know
that a is a local optimum if
∀a′∈N (a)GU (a) > GU (a′)
where
N (a) = {x x is a Neighbor of a}.
(5)
(6)
We define a Neighbor allocation as an allocation where one, and only one, agent
has a different state.
The probability that some allocation a is a local optimum is simply the prob-
ability that (5) is true. If the utility of all pairs of neighbors is not correlated,
then this probability is
Pr[∀a′∈N (a)GU (a) > GU (a′)] = Pr[GU (a) > GU (a′)]b,
(7)
where b is the branching factor. In the sensor problem b = 3 · S where S is
the set of all sensors. That is, since each sensor can be in any of four states it will
have three neighbors from each state. In some systems it is safe to assume that
the global utilities of a's neighbors are independent. However, most systems
show some degree of correlation.
Now we need to calculate the Pr[GU (a) > GU (a′)], that is, the probability
that some allocation a has a greater global utility that its neighbor a′, for all a
and a′. This could be calculated via an exhaustive enumeration of all possible
allocations. However, often we can find the expected value of this probability.
6
For example, in the sensor problem each sensor has four possible states. If
a sensor changes its state from sector x to sector y the utility of the target
agents covered by x will decrease while the utility of those in y will increase.
If we assume that, on average, the targets are evenly spaced on the field, then
the global utilities for both of these are expected to be the same. That is, the
expected probability that the global utility of one allocation is bigger than the
other is 1/2.
If, on the other hand, a sensor changes state from "off" to a sector, or
from a sector to "off," the global utility is expected to decrease and increase,
respectively. However, there are an equal number of opportunities to go from
"off" to "on" and vice-versa. Therefore, we can also expect that for these cases
the probability that the global utility of one allocation is bigger than the other
is 1/2.
Based on these approximations, we can declare that for the sensor problem
Pr[∀a′∈N (a)GU (a) > GU (a′)] =
1
2b = λ.
(8)
If we assume an even distribution of local optima, the total number of lo-
cal optima is simply the product of the total number of allocations times the
probability that each one is a local optimum. That is,
Total number of local optima = λA
(9)
For the sensor problem, λ = 1/2b, b = 3 · S and A = bS, so the expected
number of local optima is bS/23S.
Pr[a local optimum is also global] =
1
λA
=
1
2b .
(10)
We can find the expected time the algorithm will take to reach a local opti-
mum by determining the maximum number of steps from every allocation to the
nearest local optimum. This gives us an upper bound on the number of steps
needed to reach the nearest local optimum using global hill-climbing. Notice
that, under either individual hill-climbing or the bidding protocol it is possible
that the local optimum is not reached, or is reached after more steps, since these
algorithms can take steps that lower the global utility.
In order to find the expected number of steps to reach a local optimum, we
start at any one of the local optima and then traverse all possible links at each
depth d until all possible allocations have been visited. This occurs when
Solving for d, and remembering that λ = 1/2b, we get
λ · A · bd > A.
d > b logb 2.
(11)
(12)
The expected worst-case distance from any point to the nearest local opti-
mum is, therefore, b logb 2 (this number only makes sense for b ≥ 2 since smaller
7
number of neighbors do not form a searchable space). That is, the number of
steps to reach the nearest local optima in the sensor domain is proportional to
the branching factor b, which is equal to 3 · S. We can expect search time to
increase linearly with the number of sensors in the field.
3 Simulations
While the theoretical results above give us some bounds on the number of itera-
tions before the system is expected to converge to a local optimum, the bounds
are rather loose and do not tell us much about the dynamics of the executing
system. Also, we cannot show mathematically how changes in the amount of
communication change the search. Therefore, we have implemented a service
allocation simulator to answer these questions. It simulates the sensor allocation
domain described in the introduction.
The simulator is written in Java and the source code is available upon re-
quest.
It gathers and analyzes data from any desired number of runs. The
program can analyze the behavior of any number of target and sensor agents
on a two-dimensional space, and the agents can be given any desired utility
function. The program is limited to static targets. That is, it only consid-
ers the one-shot service allocation problem. Each new allocation is completely
independent of any previous one.
In the tests we performed, each run has seven sensors and seven targets, all
of which are randomly placed on a two-dimensional grid. Each sensor can only
point in one of three directions or sectors. These three sectors are the same for
all sensors (specifically, the first sector is from 0 to 120 degrees, the second one
from 120 to 240, and the third one from 240 to 360). All the sensors use the
same utility function which is given by (1), while the targets use (2). After
a sensor agent receives all the bids it chooses the sector that has the heighest
aggregate demand, as described by the bidding protocol in Section 2.1.
During a run, each of the targets periodically sends a bid to a number of
sensors asking them to turn to the sector that faces the target. We set the bid
amount to a fixed number for these tests. Periodically, the sensors count the
number of bids they have received for each sector and turn their detector (such
as a radar) to face the sector with the highest aggregate demand. We assume
that neither the targets nor the sensors can form coalitions.
We vary the number of sensors to which the targets send their bids in order
to explore the quality of the solution that the system converges upon as the
amount of communication changes. For example, at one extreme if the all the
targets send their bids to all the sensors, then the sensors would always set their
sector to be the one with the most targets. This particular service allocation
should, usually, be the best. However, it might not always be the optimal
solution. For example, if seven targets are clustered together and the eighth
is on another part of the field, it would be better if six sensor agents pointed
towards the cluster of targets while the remaining two sensor agents pointed
towards the stray target rather than having all sensor agents point towards the
8
Results with 1 Neighbor
Results with 3 Neighbors
20 40 60 80 100
time
0
1
0.8
α
0.6
0.4
0.2
30
20
10
0
0
20 40 60 80 100
time
0
1
0.8
α
0.6
0.4
0.2
Results with 5 Neighbors
Results with 7 Neighbors
20 40 60 80 100
time
0
1
0.8
α
0.6
0.4
0.2
30
20
10
0
0
20 40 60 80 100
time
0
1
0.8
α
0.6
0.4
0.2
30
20
10
0
0
30
20
10
0
0
Figure 1: The z-axis on each figure represents the number of runs, out of 100,
which had the particular α ratio at each particular time. α = 1 means the run
is at the global optimum. The optimum is reached more often in the cases with
more communication.
cluster of targets. At the other extreme, if all the targets send their bids to only
one sensor then they will minimize communications but then the sensors will
point to the sector from which they received a message -- an allocation which is
likely to be suboptimal.
These simulations explore the ruggedness of the system's global utility land-
scape and the dynamics of the agents' exploration of this landscape.
If the
agents were to always converge on a local (non-global) optimum then we would
deduce that this problem domain has a very rugged utility landscape. On the
other hand, if they usually manage to reach the global optimum then we could
deduce a smooth utility landscape.
9
Results with 4 Neighbors
30
20
10
0
0
20 40 60 80 100
time
0
1
0.8
α
0.6
0.4
0.2
Figure 2: The transitional case occurs when the target communicates with four
sensors.
10
4 Test Results
In each of our tests we set the number of agents that each target will send its
bid to, that is, the number of neighbors, to a fixed number. Given this fixed
number of neighbors, we then generated 100 random placements of agents on
the field and ran our bidding algorithm 10 times on each of those placements.
Finally, we plotted the average solution quality, over the 10 runs, as a function
of time for each of the 100 different placements. The solution quality is given
by the ratio
Current Utility
Globally Optimal Utility
α =
,
(13)
so if α = 1, then it means that the run has reached the global optimum. Since
the number of agents is small, we were able to calculate the global optimum using
a brute-force method. Specifically, there are 37 = 2187 possible configurations
times 100 random placements leads to 218700 combinations that we had to check
for each run in order to find the global optimum using brute-force. Using more
than 7 sensors made the test take too long. Notice, however, that our algorithm
is much faster than this brute-force search which we perform only to confirm
that our search does find the global optimum.
In our tests there were always seven target agents and seven sensor agents.
We varied the number of neighbors from 1 to 7. If the target can only commu-
nicate with one other sensor, the sensors will likely have very little information
for making their decision, while if all targets communicate with all seven sen-
sors, then each sensor will generally be able to point to the sector with the most
targets. However, because these decisions are made in an asynchronous manner,
it is possible that some sensor will sometimes not receive all the bids before it
has to make a decision. The targets always send their bids to the sensors that
are closest to them.
The results from our experiments are shown in Figure 1 where we can see
that there is a transition in the system's performance as the number of neighbors
goes from three to five. That is, if the targets only send their bids to three
sensors then it is almost certain that the system will stay in a configuration
that has a very low global utility. However, if the targets send their bids to
five sensors, then it is almost guaranteed (98% of the time) that the system will
reach the globally optimal allocation. This is a huge difference in terms of the
performance. We also notice in Figure 2 that for four neighbors there is a fairly
even distribution in the utility of the final allocation.
5 Related Work
There is ongoing work in the field of complexity that attempts to study they
dynamics of complex adaptive systems [4]. Our approach is based on ideas bor-
rowed from the use of NK landscapes for the analysis of co-evolving systems.
As such, we are using some of the results from that field. However, complexity
11
theory is more concerned with explaining the dynamic behavior of existing sys-
tems, while we are more concerned with the engineering of multiagent systems
for distributed service allocation.
The Collective Intelligence (COIN) framework [12] shares many of the same
goals of our research. They start with a global utility function from which they
derive the rewards functions for each agent. The agents are assumed to use some
form of reinforcement learning. They show that the global utility is maximized
when using their prescribed reward functions. They do not, however, consider
how agent communication might affect the individual agent's utility landscape.
The task allocation problem has been studied in [7], but the service allocation
problem we present in this paper has received very little attention. There is also
work being done on the analysis of the dynamics of multiagent systems for other
domains such as e-commerce [5] and automated manufacturing [6]. It is possible
that extensions to our approach will shed some light into the dynamics of these
domains.
6 Conclusions
We have formalized the service allocation problem and examined a general ap-
proach to solving problems of this type. The approach involves the use of
utility-maximizing agents that represent the resources and the services. A sim-
ple form of bidding is used for communication. An analysis of this approach
reveals that it implements a form of distributed hill-climbing, where each agent
climbs its own utility landscape and not the global utility landscape. However,
we showed that increasing the amount of communication among the agents
forces each individual agent's landscape to become increasingly correlated to
the global landscape.
These theoretical results were then verified in our implementation of a sensor
allocation problem -- an instance of a service allocation problem. Furthermore,
the simulations allowed us to determine the location of a phase transition in the
amount of communication needed for the system to consistently arrive at the
globally optimal service allocation.
More generally, we have shown how a service allocation problem can be
viewed as a distributed search by multiple agents over multiple landscapes.
We also showed how the correlation between the global utility landscape and
the individual agent's utility landscape depends on the amount and type of
inter-agent communication. Specifically, we have shown that increased commu-
nications leads to a higher correlation between the global and individual utility
landscapes, which increases the probability that the global optimum will be
reached. Of course, the success of the search still depends on the connectivity
of the search space, which will vary from domain to domain. We expect that
our general approach can be applied to the design of any multiagent systems
whose desired behavior is given by a global utility function but whose agents
must act selfishly.
Our future work includes the study of how the system will behave under
12
perturbations. For example, as the target moves it perturbs the current alloca-
tion and the global optimum might change. We also hope to characterize the
local to global utility function correlation for different service allocation prob-
lems and the expected time to find the global optimum under various amounts
of communication.
References
[1] A.
D.
Baker,
K.
Agents and the internet: Infrastructure for mass customization.
Internet Computing, 3(5):62 -- 69, September-October 1999.
Parunak,
and
H.
V.
Erol.
IEEE
[2] E. H. Durfee, T. Mullen, S. Park, J. M. Vidal, and P. Weinstein.
The dynamics of the UMDL service market society.
In M. Klusch and
G. Weiss, editors, Cooperative Information Agents II, LNAI, pages 55 -- 78.
Springer, 1998.
[3] J. M. Epstein and R. L. Axtell. Growing Artificial Societies : Social Science from the Bottom Up.
Brookings Institute, 1996.
[4] S. Kauffman. The Origins of Order: Self-Organization and Selection in Evolution.
Oxford University Pres, 1993.
[5] J. O. Kephart,
J.
E. Hanson,
Dynamic pricing by software agents.
752, 2000.
and A. R. Greenwald.
Computer Networks, 32(6):731 --
[6] H. V. D. Parunak. "go to the ant": Engineering principles from natural agent systems.
Annals of Operation Research, 75:69 -- 101, 1997.
[7] J. S. Rosenschein and G. Zlotkin. Rules of Encounter. The MIT Press,
Cambridge, MA, 1994.
[8] T. W. Sandholm. Necessary and sufficient contract types for optimal task
allocation. In Proceedings of the Fourteenth International Joint Conference
on Artificial Intelligence, 1997.
[9] R. G. Smith. The contract net protocol: High-level communication and
control in a distributed problem solver. IEEE Transactions on Computers,
C-29(12):1104 -- 1113, 1981.
[10] G. Weiss, editor. Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence.
MIT Press, 1999.
[11] M. P. Wellman. Market-oriented programming: Some early lessons.
In
S. Clearwater, editor, Market-Based Control: A Paradigm for Distributed
Resource Allocation. World Scientific, 1996.
13
[12] D. H. Wolpert and K. Tumer. An introduction to collective intelligence.
1999.
report, ACM Computing Research Repository,
Technical
cs.LG/9908014.
14
|
cs/9809108 | 1 | 9809 | 1998-09-26T17:43:36 | Learning Nested Agent Models in an Information Economy | [
"cs.MA",
"cs.AI"
] | We present our approach to the problem of how an agent, within an economic Multi-Agent System, can determine when it should behave strategically (i.e. learn and use models of other agents), and when it should act as a simple price-taker. We provide a framework for the incremental implementation of modeling capabilities in agents, and a description of the forms of knowledge required. The agents were implemented and different populations simulated in order to learn more about their behavior and the merits of using and learning agent models. Our results show, among other lessons, how savvy buyers can avoid being ``cheated'' by sellers, how price volatility can be used to quantitatively predict the benefits of deeper models, and how specific types of agent populations influence system behavior. | cs.MA | cs | Learning Nested Agent Models in an
Information Economy
Jos´e M. Vidal and Edmund H. Durfee
Artificial Intelligence Laboratory
University of Michigan
1101 Beal Avenue
Ann Arbor, MI 48109-2110
[email protected], [email protected]
Abstract
We present our approach to the problem of how an agent, within an
economic Multi-Agent System, can determine when it should behave
strategically (i.e. learn and use models of other agents), and when it
should act as a simple price-taker. We provide a framework for the
incremental implementation of modeling capabilities in agents, and
a description of the forms of knowledge required. The agents were
implemented and different populations simulated in order to learn
more about their behavior and the merits of using and learning agent
models. Our results show, among other lessons, how savvy buyers can
avoid being “cheated” by sellers, how price volatility can be used to
quantitatively predict the benefits of deeper models, and how specific
types of agent populations influence system behavior.
1 Introduction
In open, multi-agent systems, agents can come and go without any central
control or guidance, and thus how and which agents interact with each other
will change dynamically. Agents might try to manipulate the interactions to
their individual benefits, at the cost of the global efficiency. To avoid this, the
protocols and mechanisms that the agents engage in might be constructed
1
to make manipulation irrational (Rosenschein and Zlotkin, 1994), but unfor-
tunately this strategy is only applicable in restricted domains. By situating
agents in an economic society, as we do in the University of Michigan Digital
Library (UMDL), we can make each agent responsible for making its own
decisions about when to buy/sell and who to do business with (Atkins et al.,
1996). A market-based infrastructure, built around computational auction
agents, serves to discourage agents from engaging in strategic reasoning to
manipulate the system by keeping the competitive pressures high. However,
since many instances can arise where imperfections in competition could be
exploited, agents might benefit from strategic reasoning, either by manipu-
lating the system or not allowing others to manipulate them. But strategic
reasoning requires effort. An agent in an information economy like the UMDL
must therefore be capable of strategic reasoning and of determining when it
is worthwhile to invest in strategic reasoning rather than letting its welfare
rest in the hands of the market mechanism.
In this paper, we present our approach to the problem of how an agent,
within an economic MAS, can determine when it should behave strategically,
and when it should act as a simple price-taker. More specifically, we let the
agent’s strategy consist of learning nested models of the other agents, so the
decision it must make refers to which of the models will give it greater gains.
We show how, in some circumstances, agents benefit by learning and using
models of others, while at other times the extra effort is wasted. Our results
point to metrics that can be used to make quantitative predictions as to the
benefits obtained from learning and using deeper models.
1.1 Related work
Different research communities have run across the problems that arise from
having agents learning in societies of learning agents. The studies of (Shoham
and Tennenholtz, 1992), and (Glance, 1993) focus on very simple but numer-
ous agents and emphasize their emergent behavior. (Hu and Wellman, 1996)
show that learning agents in an economic domain sometimes converge to
globally sub-optimal equilibria. The work on agent-based modeling (Hubler
and Pines, 1994; Axelrod, 1984; Epstein and Axtell, 1996) of complex sys-
tems studies slightly more complicated agents that are meant as stand-ins
for real world agents (e.g. insects, communities, corporations, people).
All these researchers used agents whose learning abilities consist of choos-
ing from among a set of fixed strategies. They do not explicitly consider the
2
fact that the agents inhabit communities of learning agents. That is, their
agents do not try to model other agents. We address this issue and try to
determine when and which models an agent should keep.
Within the MAS community, some work (Sen, 1996) has focused on
how artificial AI-based learning agents would fare in communities of simi-
lar agents. For example, (Nadella and Sen, 1997) and (Terabe et al., 1997)
show how agents can learn the capabilities of others via repeated interactions,
but these agents do not learn to predict what actions other might take. Most
of the work in MAS also fails to recognize the possible gains from using ex-
plicit agent models to predict agent actions. (Tambe and Rosenbloom, 1996)
is an exception and gives another approach for using nested agent models.
However, they do not go so far as to try to quantify the advantages of their
nested models or show how these could be learned via observations. We
believe that our research will bring to the foreground some of the common
observations seen in these research areas and help to clarify the implications
and utility of learning and using nested agent models.
2 Description of the UMDL
The UMDL pro ject is a large-scale, multidisciplinary effort to design and
build a flexible, scalable infrastructure for rendering library services in a
digital networked environment.
In order to meet these goals, we chose to
implement the library as a collection of interacting agents, each specialized
to perform a particular task. These agents buy and sell goods/services from
each other, within an artificial economy, in an effort to make a profit. Since
the UMDL is an open system, which will allow third parties to build and
integrate their own agents into the architecture, we treat all agents as purely
selfish.
2.1
Implications of the information economy.
Information goods/services, like those provided in the UMDL, are very hard
to compartmentalize into equivalence classes that all agents can agree on.
For example, if a web search engine service is defined as a good, then all
agents providing web search services can be considered as selling the same
good.
It is likely, however, that a buyer of this good might decide that
seller s1 provides better answers than seller s2. We cannot possibly hope to
3
enumerate the set of reasons an agent might have for preferring one set of
answers (and thus one search agent) over another, and we should not try to
do so. It should be up to the individual buyers to decide what items belong
to the same good category, each buyer clustering items in possibly different
ways.
This situation is even more evident when we consider an information
economy rooted in some information delivery infrastructure (e.g. the Inter-
net). There are two main characteristics that set this economy apart from a
traditional economy.
• There is virtually no cost of reproduction. Once the information is
created it can be duplicated virtually for free.
• All agents have virtually direct and free access to all other agents.
If these two characteristics are present in an economy, it is useless to
talk about supply and demand, since supply is practically infinite for any
particular good and available everywhere. The only way agents can survive
in such an economy is by providing value-added services that are tailored to
meet their customers’ needs. Each provider will try to differentiate his goods
from everyone else’s while each buyer will try to find those suppliers that
best meet her value function. We propose to build agents that can achieve
these goals by learning models of other agents and making strategic decisions
based on these models. These techniques can also be applied, with variable
levels of efficacy, to traditional economies.
3 A Simplified Model of the UMDL
In order to capture the main characteristics of the UMDL, and to facilitate
the development and testing of agents, we have defined an “abstract” eco-
nomic model. We define an economic society of agents as one where each
agent is either a buyer or a sel ler of some particular good. The set of buyers
is B and the set of sellers is S . These agents exchange goods by paying some
price p ∈ P , where P is a finite set. The buyers are capable of assessing the
quality of a good received and giving it some value q ∈ Q, where Q is also a
finite set.
The exchange protocol, seen in Figure 1, works as follows: When a buyer
b ∈ B wants to buy a good g , she will advertise this fact. Each seller s ∈ S
4
that sells that good will give his bid in the form of a price pg
s . The buyer
will pick one of these and will pay the seller. All agents will be made aware
of this choice along with the prices offered by all the sellers.. The winning
seller will then return1 the specified good. Note that there is no law that
forces the seller to return a good of any particular quality. For example, an
agent that sells web search services returns a set of hits as its good. Each
buyer of this good might determine its quality based on the time it took for
the response to arrive, the number of hits, the relevance of the hits, or any
combination of these and/or other features. Therefore, it would usually be
impossible to enforce a quality measure that all buyers can agree with.
It is thus up to the buyer to assess the quality q of the good received.
Each buyer b also has a value function V g
b (p, q ) for each good g ∈ G that she
might wish to buy. The function returns a number that represents the value
that b assigns to that particular good at that particular price and quality.
Each seller s ∈ S , on the other hand, has a cost cg
s associated with each good
he can produce. Since we assume that costs and payments are expressed
in the same units (i.e. money) then, if seller s gets paid p for good g , his
profit will be Profit(p, cg
s ) = p − cg
s . The buyers, therefore, have the goal of
maximizing the value they get for their transactions, while the sellers have
the goal of maximizing their profits.
4 Learning recursive models
Agents placed in the economic society we just described will have to learn,
typically via trial and error, what actions give them the highest expected
reward and under which circumstances. In this section we will present tech-
niques that these agents might use to maximize their rewards.
An important question we wish to answer is: when do agents benefit
from having deeper nested models of other agents? It seems intuitive that,
ignoring computational costs, the agents with more complete models of others
will always do better. We find this to be usually true; however, while there
are instances when it is significantly better to have deeper models, there are
also instances when the difference is barely noticeable and, instances when
its better to ignore deeper models if they are imperfect or provide no useful
1 In the case of agent/link failure, each agent is free to set its own timeouts and assess
the quality of the never-received good accordingly. Bids that are not received in time will,
of course, not be considered.
5
information. These instances are defined in part by the set of other agents
present, their capabilities and preferences, and the dynamics of the system.
In order to precisely determine what these instances are, and in the hopes
of providing a more general framework for studying the effects of increased
agent-modeling capabilities within our economic model, we have defined a
set of techniques that our agents can use for learning and using models.
We divide the agents into classes that correspond to their modeling ca-
pabilities. The hierarchy we present is inspired by the Recursive Modeling
Method (Gmytrasiewicz, 1996), but is function-based rather than matrix-
based, and includes learning. We will first describe our agents at the knowl-
edge level, stating only the type of knowledge the agents are either trying to
acquire through learning, or already have (i.e. knowledge that was directly
implemented by the designers of the agents), and will then explain the details
of how this knowledge was implemented.
At the most abstract level, we can say that every agent i is trying to learn
the oracle decision function ∆i : w → ai , which maps the state w of the world
into the action ai that the agent should take in that state. This function will
not be fixed throughout the lifetime of the agent because the other agents
are also engaged in some kind of learning themselves. The agents that try
to directly learn ∆i (w) we refer to as 0-level agents, because they have no
explicit models of other agents. In fact, they are not aware that there are
other agents in the world. Any such agent i will learn a decision function
δi : w → ai where w is what agent i knows about its external world and ai
is its rational action in that state. For example, a web search agent might
look at the going price for web searches, in order to determine how much to
charge for its service.
Agents with 1-level models of other agents, on the other hand, are aware
that there are other agents out there but have no idea what the “interior”
of these agents looks like. They have two kinds of knowledge— a set of
functions δij : w → aj which capture agent i’s model of each of the other
agents (j ), and δi : (w , ~a−i ) → ai which captures i’s knowledge of what action
to take given w and the collective actions ~a−i the others will take. We define
~a−i = {a1 · · · ai−1 , ai+1 · · · an }, where n is the number of agents. An agent’s
model of others might not be correct; therefore, it is not always true that
δj (w) = δij (w). The δij (w) knowledge for all j 6= i turns out to be easier to
learn than the joint action δi (w , ~a−i) because the set of possible hypotheses
is smaller.
Agents with 2-level models are assumed to have deeper knowledge about
6
the other agents; that is, they have knowledge of the form δij : (w , ~a−j ) → aj .
This knowledge tells them how others determine which action to take. They
also know what actions others think others are going to take, i.e. δij k : w →
ak , and (like 1-level modelers) what action they should take given others’
actions, δi : (w , ~a−i ) → ai . Again, the δij k (w) is easier to learn that the other
two, as long as all agents use the same features to discriminate among the
different worlds (i.e. share the same w).
4.1 Populating the knowledge
If the different level agents had to learn all the knowledge then, since the 0-
level agents have a lot less knowledge to learn, they would learn it much faster.
However, in the economic domain, it is likely that the designer has additional
knowledge which could be incorporated into the agents. The agents we built
incorporated extra knowledge along these lines.
We decided that 0-level agents would learn all their knowledge by tracking
their actions and the rewards they got. These agents, therefore, receive
no extra domain knowledge from the designers and learn everything from
experience. 1-level agents, on the other hand, have a priori knowledge of
what action they should take given the actions that others will take. That
is, while they try to learn knowledge of the form δij (w) by observing the
actions others take (i.e.
in a form of supervised learning where the other
agents act as tutors), they already have knowledge of the form δi (w , ~a−i ).
In our economic domain, it is reasonable to assume that agents have this
knowledge since, in fact, this type of knowledge can be easily generated.
That is, if I know what all the other sellers are going to bid, and the prices
that the buyer is willing to pay, then it is easy for me to determine which
price to bid. We must also point out that in this domain, the δi (w , ~a−i )
knowledge cannot be used by a 0-level agent. If this knowledge had said,
for instance, that from some state w agent i will only ever take one of a
few possible actions, then this knowledge could have been used to eliminate
impossibilities from the δi (w) knowledge of a 0-level agent. However, this
situation never arises in our domain because, as we shall see in the following
Sections, the states used by the agents permit the set of reasonable actions
to always be equal to the set of all possible actions.
The 2-level agents learn their δij k (w) knowledge from observations of
others’ actions, under the already stated assumption that there is common
knowledge of the fact that all agents see the actions taken by all. The rest of
7
the knowledge, i.e. δij (w , ~a−j ) and δi (w , ~a−i), is built into the 2-level agents
a priori. As with 1-level agents, we cannot use the δij (w , ~a−j ) knowledge to
add δij (w) knowledge to a 1-level modeler, because other agents are also free
to take any one of the possible actions in any state of the world. There are
many reasonable ways to explain how the 2-level agents came to possess the
δij (w , ~a−j ) knowledge. It could be, for instance, that the designer assumed
that the other designers would build 1-level agents with the same knowledge
we just described. This type of recursive thinking (i.e. “they will do just
as I did, so I must do one better”), along with the obvious expansion of
the knowledge structure, could be used to generate n-level agents, but so far
we have concentrated only on the first three levels. The different forms of
knowledge, and their form of acquisition, are summarized in Table 1. In the
following sections, we talk about each one of these agents in more detail and
give some specifics on their implementation. Our current model emphasizes
transactions over a single good, so each agent is only a buyer or a seller, but
cannot be both.
4.2 Agents with 0-level models
Agents with 0-level models must learn everything they know from observa-
tions they make about the environment, and from any rewards they get. In
our economic society this means that buyers see the bids they receive and
the good received after striking a contract, while sellers see the request for
bids and the profit they made (if any). In general, these agents get some
input, take an action, then receive some reward. This framework is the same
framework used in reinforcement learning, which is why we decided to use a
form of reinforcement learning (Sutton, 1988) (Watkins and Dayan, 1992),
for implementing learning in our agents.
Both buyers and sellers will use the equations in the next few sections for
determining what actions to take. But, with a small probability ǫ they will
choose to explore, instead of exploit, and will pick their actions at random
(except for the fact that sellers never bid below cost). The value of ǫ is
initially 1 but decreases with time to some empirically chosen, fixed minimum
value ǫmin . That is,
ǫt+1 = ( γ ǫt
if γ ǫt > ǫmin
ǫmin otherwise
where 0 < γ < 1 is some annealing factor.
8
Time 1
request-bid
Time 2
p1
S1
S2
S1
B
B
Pick-bid
p2
S2
Time 3
payment
S3
B
S3
Time 4
p3
S3
B
Assess-quality
good
S3
Figure 1: View of the protocol. We show only one buyer B and three sellers
S 1, S 2, and S 3. At time 1 the buyer requests bids for some good. At time
2 the sellers send their prices for that good. At time 3 the buyer picks one
of the bids, pays the seller the amount and then, at time 4, she receives the
good.
Level
0-level
1-level
2-level
Form of Knowledge Method of Acquisition
δi (w) Reinforcement Learning
δi (w , ~a−i) Previously known
δij (w) Learn from observation
δi (w , ~a−i) Previously known
δij (w , ~a−j ) Previously known
δij k (w) Learn from observation.
Table 1: Summary of the forms of knowledge that the different agents have
or are trying to learn.
9
4.2.1 Buyers with 0-level models.
A buyer b will start by requesting bids for a good g . She will then receive all
bids for good g and will pick the seller:
s∗ = args∈S max f g (pg
s )
(1)
This function implements the buyer’s δb(w) which, in this case, can be
rephrased as δb(p1 . . . pS ). The function f g (p) returns the value the buyer
expects to get if she buys good g at a price of p. It is learned using a simple
form of reinforcement learning, namely:
f g
t+1 (p) = (1 − α)f g
t (p) + α · V g
b (p, q )
Here α is the learning rate, p is the price b pays for the good, and q is
the quality she ascribes to it. The learning rate is initially set to 1 and, like
ǫ, is decreased until it reaches some fixed minimum value αmin .
(2)
4.2.2 Sellers with 0-level models.
When asked for a bid, the seller s will provide one whose price is greater
than or equal2 to the cost for producing it (i.e. pg
s ≥ cg
s ). From these prices,
he will chose the one with the highest expected profit:
p∗
s = argp∈P max hg
s (p)
(3)
Again, this function encompasses the seller’s δs (g ) knowledge, where we now
have that the states are the goods being sold w = g , and the actions are
prices offered a = p. The function hg
s (p) returns the profit s expects to get if
he offers good g at a price p. It is also learned using reinforcement learning,
as follows:
hg
t+1 (p) = (1 − α)hg
t (p) + α · Profitg
s (p)
(4)
where
s (p) = ( p − cg
Profitg
s
0
2We could just as easily have said that the price must be strictly greater than the cost.
if his bid is chosen
otherwise
(5)
10
4.3 Agents with One-level Models
The next step is for an agent to keep one-level models of the other agents.
This means that it has no idea of what the interior (i.e. “mental”) processes of
the other agents are, but it recognizes the fact that there are other agents out
there whose behaviors influence its rewards. The agent, therefore, can only
model others by looking at their past behavior and trying to predict, from it,
their future actions. The agent also has knowledge, implemented as functions,
that tells it what action to take, given a probability distribution over the set
of actions that other agents can take. In the actual implementation, as shown
below, the δi (w , ~a−i) knowledge takes into account the fact that the δij (w)
knowledge is constantly being learned and, therefore, is not correct with
perfect certainty.
4.3.1 Buyers with one-level models.
A buyer with one-level models can now keep a history of the qualities she
ascribes to the goods returned by each seller. She can, in fact, remember
the last N qualities returned by some seller s for some good g , and define
a probability density function q g
s (x) over the qualities x returned by s (i.e.
q g
s (x) returns the probability that s returns an instance of good g that has
quality x). This function provides the δbs (g ) knowledge. She can use the
expected value of this function to calculate which seller she expects will give
her the highest expected value.
= args∈S max
s∗ = args∈S max E (V g
s , q g
b (pg
s (x)))
1
s (x) · V g
b (pg
q g
Q Xx∈Q
s , x)
The δb (g , q1 · · · qS ) is given by the previous function which simply tries
to maximize the value the buyer expects to get. The buyer does not need to
model other buyers since they do not affect the value she gets.
(6)
4.3.2 Sellers with one-level models.
Each seller will try to predict what bid the other sellers will submit (based
solely on what they have bid in the past), and what bid the buyer will likely
pick. A complete implementation would require the seller to remember past
combinations of buyers, bids and results (i.e. who was buying, who bid what,
11
and who won). However, it is unrealistic to expect a seller to remember all
this since there are at least P S · B possible combinations.
However, the seller’s one-level behavior can be approximated by having
him remember the last N prices accepted by each buyer b for each good g ,
and form a probability density function mg
b (x), which returns the probability
that b will accept (pick) price p for good g . The expected value of this
function provides the δsb (g ) knowledge. Similarly, the seller remembers other
sellers’ last N bids for good g and forms ng
s (y ), which gives the probability
that s will bid y for good g . The expected value of this function provides
the δs (g ) knowledge. The seller s can now determine which bid maximizes
his expected profits.
(7)
if mg
b (p′ ) ≤ mg
b (p)
otherwise
s ) · Ys′∈{S−s} Xp′∈P ( ng
s′ (p′ )
p∗ = argp∈P max(p − cg
0
Note that this function also does a small amount of approximation by
assuming that s wins whenever there is a tie3 . The function calculates the
best bid by determining, for each possible bid, the product of the profit and
the probability that the agent will get that profit. Since the profit for lost bids
is 0, we only need to consider the cases where s wins. The probability that
s will win can then be found by calculating the product of the probabilities
that his bid will beat the bids of each of the other sellers. The function
approximates the δs (g , pb , p1 · · · pS ) knowledge.
4.4 Agents with Two-level Models
The intentional models we use correspond to the functions used by agents
that use one-level models. The agents’ δi (w , ~a−i) knowledge has again been
expanded to take into account the fact that the deeper knowledge is learned
and might not be correct. The δij k (w) knowledge is learned from observation,
under the assumption that there is common knowledge of the fact that all
agents see the bids given by all agents.
3The complete solution would have to consider the probabilities that s ties with 1, 2,
3,. . . other agents. In order to do this we would need to consider all P S subsets.
12
4.4.1 Buyers with two-level models.
Since the buyer receives bids from the sellers, there is no need for her to try to
out-guess or predict what the sellers will bid. She is also not concerned with
what the other buyers are doing since, in our model, there is an effectively
infinite supply of goods. The buyers are, therefore, not competing with each
other and do not need to keep deeper models of others.
4.4.2 Sellers with two-level models.
A seller will model other sellers as if they were using the one-level models.
That is, he thinks they will model others using policy models and make
their decisions using the equations presented in Section 4.3.2. He will try to
predict their bids and then try to find a bid for himself that the buyer will
prefer more than all the bids of the other sellers. His model of the buyer will
also be an intentional model. He will model the buyers as though they were
implemented as explained in Section 4.3.1. A seller, therefore, assumes that
it has the correct intentional models of other agents.
The algorithm he follows is to first use his models of the sellers to predict
what bids pi they will submit. He has a model of the buyer C (s1 · · · sn , p1 · · · pn ) →
si , that tells him which seller she might choose given the set of bids pi sub-
mitted by each seller si . The seller sj uses this model to determine which of
his bids will bring him higher profit, by first finding the set of bids he can
make that will win:
P ′ = {pj pj ∈ P , sj = C (s1 · · · sj · · · sn , p1 · · · pj · · · pn )}
And from these finding the one with the highest profit:
p∗ = argp∈P ′ max(p − cg
s )
These functions provide the δs (g , pb , p1 · · · pS ) knowledge.
(8)
(9)
5 Tests
Since there is no obvious way to analytically determine how different pop-
ulations of agents would interact and, of greater interest to us, how much
better (or worse) the agents with deeper models would fare, we decided to
implement a society of the agents described above and ran it to test our
13
hypotheses. In all tests, we had 5 buyers and 8 sellers. The buyers had the
same value function Vb(p, q ) = 3q − p, which means that if p = q then the
buyers will prefer the seller that offers the higher quality. The quality that
they perceived was the same only on average, i.e. any particular good might
be thought to have quality that is slightly higher or lower than expected. All
sellers had costs equal to the quality they returned in order to support the
common sense assumption that quality goods cost more to produce. A set
of these buyers and sellers is what we call a population. We tried various
populations; within each population we kept constant the agents’ modeling
levels, the value assessment functions and the qualities returned. The tests
involved a series of such populations, each one with agents of different mod-
eling levels, and/or sellers with different quality/costs. We also set αmin = .1,
ǫmin = .05, and γ = .99. There were 100 runs done for each population of
agents, each run consisting of 10000 auctions (i.e. iterations of the protocol).
The lessons presented in the next section are based on the averages of these
100 runs.
6 Lessons
From our tests we were able to discern several lessons about the dynamics
of different populations of agents. Some of these lead to methods that can
be used to make quantitative predictions about agents’ performance, while
others make qualitative assessments about the type of behaviors we might
expect. We detail these in the next subsections, and summarize them in
Table 2.
6.1 Micro versus macro behaviors.
In all tests, we found the behavior for any particular run does not necessarily
reflect the average behavior of the system. The prices have a tendency to
sometimes reach temporary stable points. These conjectural equilibria, as
described in (Hu and Wellman, 1996), are instances when all of the agents’
models are correctly predicting the others’ behavior and, therefore, the agents
do not need to change their models or their actions. These conjectural equi-
libria points are seldom global optima for the agents. If one of our agents
finds itself at one of these equilibrium points, since the agent is always explor-
ing with probability ǫ, it will in time discover that this point is only a local
14
optima (i.e.
it can get more profit selling/buying at a different price) and
will change its actions accordingly. Only when the price is an equilibrium
price4 do we find that the agents continue to forever take the same actions,
leaving the price at its equilibrium point.
In order to understand the more significant macro-level behaviors of the
system, we present results that are based on the averages from many runs.
While these averages seem very stable, and a good first step in learning
to understand these systems, in the future we will need to address some
of the micro-level issues. We do notice from our data that the micro-level
behaviors (e.g. temporary conjectural equilibria, price fluctuations) are much
more closely tied, usually in intuitive ways, to the agents’ learning rate α and
exploration rate ǫ. That is, higher rates for both of these lead to more price
fluctuations and shorter temporary equilibria.
6.2
0-level buyers and sellers.
This type of population is equivalent to a “blind” auction, where the agents
only see the price and the good, but are prevented from seeing who the
seller (or buyer) was. As expected, we found that an equilibrium is reached
as long as all the sellers are providing the same quality. This is the case
for population 1 in Figure 2. Otherwise, if the sellers offer different quality
goods, the price fluctuates as the buyers try to find the price that on the
whole returns the best quality, and the sellers try to find the price5 the
buyers favor.
In these populations, the sellers offering the higher quality,
at a higher cost, lose money. Meanwhile, sellers offering lower quality, at
a lower cost, earn some extra income by selling their low quality goods to
buyers that expect, and are paying for, higher quality. As more sellers start to
offer lower quality, we find that the mean price actually increases, evidently
because price acts as a signal for quality and the added uncertainty makes
the higher prices more likely to give the buyer a higher value. We see this
in Figure 2, where population 1 has all sellers returning the same quality
while in each successive population more agents offer lower quality. The
price distribution for population 1 is concentrated on 9, but for populations
2 through 6 it flattens and shifts to the right, increasing the mean price. It is
4That is, p is an equilibrium price if every seller that can sell at that price (i.e. those
whose cost is less than p) does.
5Remember, the sellers are constrained to return a fixed quality. They can only change
the price they charge.
15
y
t
i
l
i
b
a
b
o
r
P
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Price Distributions
Pop. 1
Pop. 2
Pop. 3
Pop. 4
Pop. 5
Pop. 6
Pop. 7
Pop. 8
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Prices
Figure 2: Price distributions for populations of 0-level buyers and sellers.
The prices are 0 · · · 19. The columns represent the percentage of time the
good was sold at each price, in each population. In p1 sellers return qualities
{8, 8, 8, 8, 8, 8, 8, 8}, in p2 its {8, 8, 8, 8, 8, 8, 7, 8}, and so on such that by p8
its {1, 2, 3, 4, 5, 6, 7, 8}. The highest peak in all populations corresponds to
price 9.
16
only by population 7 when it starts to shift back to the left, thus reducing the
mean price, as seen in Figure 3. That is, it is only after a significant number
of sellers start to offer lower quality that we see the mean price decrease.
6.3
0-level buyers and sellers, plus one 1-level seller.
In these population sets we explored the advantages that a 1-level seller has
over identical 0-level sellers. The advantage was non-existent when all sellers
returned the same quality (i.e. when the prices reached an equilibrium as
shown in population 1 in Figure 4), but increased as the sellers started to
diverge in the quality they returned. In order to make these findings useful
when building agents, we needed a way to make quantitative predictions as to
the benefits of keeping 1-level models. It turns out that these benefits can be
predicted, not by the population type as we had first guessed, but by the price
volatility. We define volatility as the number of times the price changes from
one auction to the next, divided by the total number of auctions. Figure 5
shows the linear relation between volatility and the percentage of times the
1-level seller wins. The two lines correspond to two “types” of volatility. The
first line includes populations 1 through 5 (p1-p5). It reflects the case where
the buyers’ second-favorite (and possibly, the third, fourth, etc.) equilibrium
price is greater than her most preferred price.
In these cases the buyers
and sellers fight among the two most preferred prices, the sellers pulling
towards the higher equilibrium price and the buyers towards the lower one,
as shown by the two peaks in populations 4 and 5 in Figure 4. The other line,
which includes populations 6 and 7, corresponds to cases where the buyers’
preferred equilibrium price is greater than the runner-ups.
In these cases
there is no contest between two equilibria. We observe only one peak in the
price distribution for these populations.
The slope of these lines can be easily calculated and the resulting function
can be used by a seller agent for making a quantitative prediction as to
how much he would benefit by switching to 1-level models. That is, he
could measure price volatility, multiply it by the appropriate slope, and the
resulting number would be the percentage of times he would win. However,
for this to work the agent needs to know that all eight buyers and five sellers
are 0-level modelers because different types of populations lead to different
slopes. Also, slight changes in our learning parameters (.02 ≤ ǫmin ≤ .08 and
.05 ≤ αmin ≤ .2) lead to slight changes in the slopes so these would have to
be taken into account if the agent is actively changing its parameters.
17
e
c
i
r
P
n
a
e
M
10.4
10.2
10
9.8
9.6
9.4
9.2
1
Mean Price for each Population
Price
2
3
5
4
Population
6
7
8
Figure 3: Mean price for the populations from Figure 2
y
t
i
l
i
b
a
b
o
r
P
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Price Distributions
Pop. 1
Pop. 2
Pop. 3
Pop. 4
Pop. 5
Pop. 6
Pop. 7
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Prices
Figure 4: Price distributions for populations of 0-level buyers and 0-level
sellers plus one 1-level seller.
In population 1 sellers return qualities
{2, 2, 2, 2, 2, 2, 2, 2}, in pop. 2 its {2, 3, 2, 2, 2, 2, 2, 2}, and so on such that
by pop. 7 it’s {2, 3, 4, 5, 6, 7, 8, 2}. The 1-level seller is the last on the list. It
always returns quality 2.
18
We also want to make clear a small caveat, which is that the volatility
that is correlated to the usefulness of keeping 1-level models is the volatility
of the system with the agent already doing 1-level modeling. Fortunately,
our experiments show that having one agent change from 0-level to 1-level
does not have a great effect on the volatility as long as there are enough (i.e.
more than five or so) other sellers.
The reason volatility is such a good predictor is that it serves as an accu-
rate assessment of how dynamic the system is and, in turn, of the complexity
of the learning problem faced by the agents. It turns out that the learning
problem faced by 1-level agents is “simpler” than the one faced by 0-level
modelers. Our 0-level agents use reinforcement learning to learn a good
match between world states and the actions they should take. The 1-level
agents, on the other hand, can see the actions other agents take and do not
need to learn their models through indirect reinforcements. They instead
use a form of supervised learning to learn the models of others. Since 1-level
agents need fewer interactions to learn a correct model, their models will,
in general, be better than those of 0-level agents in direct proportion to the
speed with which the target function changes. That is, in a slow-changing
world both of them will have time enough to arrive at approximately cor-
rect models, while in a fast-changing world only the 1-level agents will have
time to arrive at an approximately correct model. This explains why high
price volatility is correlated to an increase in the 1-level agent’s performance.
However, as we saw, the relative advantages for different volatilities (i.e. the
slope in Figure 5) will also depend on the shape of the price distribution and
the particular population of agents.
Finally, in all populations where the buyers are 0-level, we saw that it
really pays for the sellers to have low costs because this allows them to lower
their prices to fit almost any demand. Since the buyers have 0-level models,
the sellers with low quality and cost can raise their prices when appropriate,
in effect “pretending” to be the high-quality sellers, and make an even more
substantial profit. This extra profit comes at the cost of a reduction in the
average value that the buyers receive.
In other words, the buyers get less
value because they are only 0-level agents and are less able to detect the
sellers’ deception. In the next Section we will see how this is not true for
1-level buyers.
Of course, the 1-level sellers were more successful at this deception strat-
egy than the 0-level sellers. Figure 6 shows the profit of several agents in a
population as a function of their cost. We can see how the 0-level agents’
19
l
e
v
e
l
-
1
r
o
f
s
n
i
W
f
o
e
g
a
t
n
e
c
r
e
P
2200
2100
2000
1900
1800
1700
1600
1500
1400
1300
1200
Correlation of Volatility and Wins
p5
p6
p3
p4
p7
Wins
p2
p1
0.1
0
0.2
0.3
Volatility
0.4
0.5
Figure 5: Scatter plot of volatility versus the percentage of time that the
1-level seller wins (w). The populations are the same as in Figure 4.
t
n
e
g
A
e
h
t
r
o
f
t
i
f
o
r
P
l
a
t
o
T
f
o
e
g
a
t
n
e
c
r
e
P
0.3
0.25
0.2
0.15
0.1
0.05
1
Average Profit for Agents
1-level
1-level
Profit
0-level
0-level
2
3
4
5
Agent’s Cost
6
7
8
9
Figure 6: Agents’ profit in population 7 from Figure 4. The sellers return
qualities of {2, 3, 4, 5, 6, 7, 8, 2}, respectively, where the 1-level seller is the
first on the list.
20
profit decreases with increasing costs, and how the 1-level agent’s profit is
much higher than the 0-level with the same costs. We also notice that, since
the 0-level agents are not as successful as the 1-level at taking advantage of
their low costs, the first 0-level seller (that returns quality 2) has lower profit
than the rest as some of his profit was taken away by the 1-level seller (that
returns the same quality).
6.4
1-level buyers and 0 and 1-level sellers.
In these populations the buyers have the upper hand. They quickly identify
those sellers that provide the highest quality goods and buy exclusively from
them. The sellers do not benefit from having deeper models; in fact, Figure 7
shows how the 1-level seller’s profit is less than that of a similar 0-level seller
because the 1-level seller tries to charge higher prices than the 0-level seller.
The 1-level buyers do not fall for this trick— they know what quality to
expect, and buy more from the lower-priced 0-level seller(s). We have here
a case of erroneous models— 1-level sellers assume that buyers are 0-level,
and since this is not true, their erroneous deductions lead them to make bad
decisions. To stay a step ahead, sellers would need to be 2-level in this case.
In Figure 7, the first population has all sellers returning a quality of 8
while by population 7 they are returning qualities of {8, 2, 3, 4, 5, 6, 7, 8}, re-
spectively, with the 1-level always returning quality of 8. We notice that the
difference in profits between the 0-level and the 1-level increases with succes-
sive populations. This is explained by the fact that in the first population
all seven 0-level sellers are returning the same quality, while by population 7
only the 0-level pictured (i.e. the first one) is still returning quality 8. This
means that his competition, in the form of other 0-level sellers returning the
same quality, decreases for successive populations. Meanwhile, in all pop-
ulations there is only one 1-level seller who has no competition from other
1-level sellers. To summarize, the 0-level seller’s profit is always higher than
the similar 1-level seller’s, and the difference increases as there are fewer
other competing 0-level sellers who offer the same quality.
6.5
1-level buyers and several 1-level sellers.
We have shown how 1-level sellers do better, on average, than 0-level sellers
when faced with 0-level buyers, but this is not true anymore if too many
0-level sellers decide to become 1-level. Figure 8 shows how the profits of a
21
t
n
e
g
A
e
h
t
r
o
f
t
i
f
o
r
P
l
a
t
o
T
f
o
e
g
a
t
n
e
c
r
e
P
0.5
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
1
Average Profit for Agents
0-level agent
1-level agent
2
3
4
Population
5
6
7
Figure 7: Profit of a 1-level seller and a 0-level seller, both with the same
costs and quality, in a population with 1-level buyers.
t
n
e
g
A
e
h
t
r
o
f
t
i
f
o
r
P
l
a
t
o
T
f
o
e
g
a
t
n
e
c
r
e
P
0.22
0.21
0.2
0.19
0.18
0.17
0.16
0.15
0.14
0.13
0.12
0.11
Average Profit for Agents
0-level agent
1-level agent
0
1
2
4
3
Number of 1-level Agents
5
6
Figure 8: Profit of 1-level seller and similar 0-level seller as a function of the
number of 1-level sellers in the population. There are a total of 8 agents in
all populations.
22
1-level seller decrease as he is joined by other 1-level sellers. In this Figure
the sellers are returning qualities of {2, 2, 2, 2, 2, 3, 4}. Initially they are all
0-level, then one of the sellers with quality 2 becomes 1-level (he is the seller
shown in the Figure), then another one and so on. . . until there is only one
0-level seller with quality two. Then the seller with quality three becomes
1-level and, finally the seller with quality four becomes 1-level. At this point
we have six 1-level sellers and one 0-level seller. We can see that with more
than four 1-level sellers the 0-level seller is actually making more profit than
the similar 1-level seller. The 1-level seller’s profit decreases because, as
more sellers change from 0 to 1-level, they are competing directly with him
since they are offering the same quality and are the same level. Notice that
the 1-level seller’s curve flattens after four 1-level sellers are present in the
population. The reason is that the next sellers to change over to 1-level
return qualities of 3 and 4, respectively, so that they do not compete directly
with the seller pictured. His profits, therefore, do not keep decreasing.
For this test, and other similar tests, we had to use a population of sellers
that produce different qualities because, as explained in Section 6.3, if they
had returned the same quality then an equilibrium would have been reached
which would prevent the 1-level sellers from making a significantly greater
profit than the 0-level sellers.
6.6
1-level buyers and 1 and 2-level sellers.
Assuming that the 2-level seller has perfect models of the other agents, we
find that he wins an overwhelming percentage of the time. This is true,
surprisingly enough, even when some of the 1-level sellers offer slightly higher
quality goods. However, when the quality difference becomes too great (i.e.
greater than 1), the buyers finally start to buy from the high quality 1-
level sellers. This case is very similar to the ones with 0-level buyers and 0
and 1-level sellers and we can start to discern a recurring pattern. In this
case, however, it is much more computationally expensive to maintain 2-level
models. On the other hand, since these 2-level models are perfect, they are
better predictors than the 1-level, which explains why the 2-level seller wins
much more than the 1-level seller from Section 6.3.
23
Buyers Sellers
0-level
0-level
0-level Any
0-level
0-level and
one 1-level
0-level
1-level
0-level and
many 1-level
0-level and
one 1-level
1-level
1-level and
one 2-level
Lessons
Equilibrium reached only when all sellers offer
the same quality. Otherwise, we get oscillations.
Mean price increases when quality offered decreases.
Sellers have big incentives to lower quality/cost.
1-level seller beats others.
Quantitative advantage of being 1-level predicted by
volatility and price distribution.
1-level sellers do better, as long as there
are not too many of them.
Buyers have upper hand. They buy from the most
preferred seller.
1-level sellers are usually at a disadvantage.
Since 2-level has perfect models, it wins an
overwhelming percentage of time, except
when it offers a rather lower quality.
Table 2: Summary of lessons. In all cases the buyers had identical value
and quality assessment functions. Sellers were constrained to always return
the same quality.
24
7 Conclusions
We have presented a framework for the development of agents with incremen-
tal modeling/learning capabilities, in an economic society of agents. These
agents were built, and the execution of different agent populations leads us to
the discovery of the lessons summarized in Table 2. The discovery of volatil-
ity and price distributions as predictors of the benefits of deeper models will
be very useful as guides for deciding how much modeling capability to build
into an agent. This decision could either be done prior to development or,
given enough information, it could be done at runtime. We are also encour-
aged by the fact that increasing the agents’ capabilities changes the system
in ways that we can recognize from our everyday economic experience.
Some of the agent structures shown in this paper are already being im-
plemented into the UMDL (Atkins et al., 1996). We have a basic economic
infrastructure that allows agents to engage in commerce, and the agents
use customizable heuristics for determining their strategic behavior. We are
working on incorporating the more advanced modeling capabilities into our
agents in order to enable more interesting strategic behaviors.
Our results showed how sellers with deeper models fare better, in gen-
eral, even when they produce less valuable goods. This means that we should
expect those types of agents to, eventually, be added into the UMDL6 . For-
tunately, this advantage is diminished by having buyers keep deeper models.
We expect that there will be a level at which the gains and costs associ-
ated with keeping deeper models balance out for each agent. Our hope is
to provide a mechanism for agents to dynamically determine this cutoff and
constantly adjust their behavior to maximize their expected profits given the
current system behavior. The lessons in this paper are a significant step in
this direction. We have seen that one needs to look at price volatility and at
the modeling levels of the other agents to determine what modeling level will
give the highest profits. We have also learned how buyers and sellers of dif-
ferent levels and offering different qualities lead to different system dynamics
which, in turn, dictate whether the learning of nested models is useful or not.
We are considering the expansion of the model with the possible additions
of agents that can both buy and sell, and sellers that can return different
quality goods. Allowing sellers to change the quality returned to fit the
buyer will make them more competitive against 1-level buyers. We are also
6 If not by us, then by a profit-conscious third party.
25
continuing tests on many different types of agent populations in the hopes
of getting a better understanding of how well different agents fare in the
different populations.
In the long run, another offshoot of this research could be a better charac-
terization of the types of environments and how they allow/inhibit “cheating”
behavior in different agent populations. That is, we saw how, in our economic
model, agents are sometimes rewarded for behavior that does not seem to be
good for the community as a whole (e.g. when some of the sellers raised their
price while lowering the quality they offered). The rewards, we are finding,
start to diminish as the other agents become “smarter”. We can intuit that
the agents in these systems will eventually settle on some level of nesting that
balances their costs of keeping nested models with their gains from taking
better actions (Kauffman, 1994). It would be very useful to characterize the
environments, agent populations, and types of “equilibria” that these might
lead to, especially as interest in multi-agent systems grows.
References
Akerlof, G. A. (1970). The market for ‘lemons’: Quality uncertainty and the
market mechanism. The Quaterly Journal of Economics, pages 488–500.
Atkins, D. E., Birmingham, W. P., Durfee, E. H., Glover, E. J., Mullen, T.,
Rundensteiner, E. A., Soloway, E., Vidal, J. M., Wallace, R., and Well-
man, M. P. (1996). Toward inquiry-based education through interacting
software agents. IEEE Computer.
Axelrod, R. M. (1984). The Evolution of Cooperation. Basic Books.
Durfee, E. H., Gmytrasiewicz, P. J., and Rosenschein, J. S. (1994). The
utility of embedded communications and the emergence of protocols. In
Proceedings of the 13th International Distributed Artificial Intel ligence
Workshop.
Epstein, J. M. and Axtell, R. L. (1996). Growing Artifical Societies: Social
Science from the Bottom Up. Brookings Institution.
Glance, N. S. (1993). Dynamics with Expectations. PhD thesis, Stanford
University.
26
Gmytrasiewicz, P. J. (1996). On reasoning about other agents. In Wooldridge,
M., Muller, J. P., and Tambe, M., editors, Intel ligent Agents Volume II,
Lecture Notes in Artificial Intelligence, pages 143–155. Springer-Verlag.
Hu, J. and Wellman, M. P. (1996). Self-fulfilling bias in multiagent learning.
In Proceedings of the Second International Conference on Multi-Agent
Systems, pages 118–125.
Hubler, A. and Pines, D. (1994). Complexity: Methaphors, Models and Re-
ality, chapter Prediction and Adaptation in an Evolving Chaotic Envi-
ronment, pages 343–379. Addison Wesley.
Kauffman, S. A. (1994). Complexity: Models, Metaphors and Reality, chapter
Whispers from Carnot: The Origins of Order and Principles of Adapta-
tion in Complex Nonequilibrium Systems, pages 83–136. Addison Wes-
ley.
Mullen, T. and Wellman, M. P. (1996). Some issues in the design of market-
oriented agents.
In Wooldridge, M., Muller, J. P., and Tambe, M.,
editors, Intel ligent Agents Volume II, Lecture Notes in Artificial Intelli-
gence, pages 283–298. Springer-Verlag.
Nadella, R. and Sen, S. (1997). Correlating internal parameters and exter-
nal performance. In Weiss, G., editor, Distributed Artificial Intel ligence
Meets Machine Learning, pages 137–150. Springer.
Rosenschein, J. S. and Zlotkin, G. (1994). Rules of Encounter. The MIT
Press, Cambridge, Massachusetts.
Russell, S. (1995). Rationality and intelligence. In Proceedings of the 14th
International Joint Conference on Artificial Intel ligence, pages 950–957.
Sen, S., editor (1996). Working Notes for the AAAI Symposium on Adapta-
tion, Co-evolution and Learning in Multiagent Systems.
Shoham, Y. and Tennenholtz, M. (1992). Emergent conventions in multi-
agent systems. In Proceedings of Know ledge Representation.
Sutton, R. S. (1988). Learning to predict by the methods of temporal differ-
ences. Machine Learning, 3:9–44.
27
Tambe, M. and Rosenbloom, P. S. (1996). Architectures for agents that
track other agents in multi-agent worlds.
In Wooldridge, M., Muller,
J. P., and Tambe, M., editors, Intel ligent Agents Volume II, Lecture
Notes in Artificial Intelligence, pages 156–170. Springer-Verlag.
Terabe, M., Wasio, T., Katai, O., and Sawaragi, T. (1997). A study of
organizational learning in multi-agent systems.
In Weiss, G., editor,
Distributed Artificial Intel ligence Meets Machine Learning, pages 168–
179. Springer.
Vidal, J. M. and Durfee, E. H. (1996a). The impact of nested agent models
in an information economy. In Proceedings of the Second International
Conference on Multi-Agent Systems, pages 377–384. http://jmvidal.
ece.sc.edu/papers/amumdl/.
Vidal, J. M. and Durfee, E. H. (1996b). Using recursive agent models
effectively.
In Wooldridge, M., Muller, J. P., and Tambe, M., edi-
tors, Intel ligent Agents Volume II, Lecture Notes in Artificial Intelli-
gence, pages 171–196. Springer-Verlag. http://jmvidal.ece.sc.edu/
papers/lr-rmm2/.
Watkins, C. J. and Dayan, P. (1992). Q-learning. Machine Learning, 8:279–
292.
28
|
1910.13880 | 1 | 1910 | 2019-10-30T14:26:30 | Path Planning Games | [
"cs.MA",
"cs.AI",
"cs.GT",
"cs.RO"
] | Path planning is a fundamental and extensively explored problem in robotic control. We present a novel economic perspective on path planning. Specifically, we investigate strategic interactions among path planning agents using a game theoretic path planning framework. Our focus is on economic tension between two important objectives: efficiency in the agents' achieving their goals, and safety in navigating towards these. We begin by developing a novel mathematical formulation for path planning that trades off these objectives, when behavior of other agents is fixed. We then use this formulation for approximating Nash equilibria in path planning games, as well as to develop a multi-agent cooperative path planning formulation. Through several case studies, we show that in a path planning game, safety is often significantly compromised compared to a cooperative solution. | cs.MA | cs |
Path Planning Games
Yi Li and Yevgeniy Vorobeychik
Electrical Engineering and Computer Science, Vanderbilt University, Nashville, TN
Abstract. Path planning is a fundamental and extensively explored problem in
robotic control. We present a novel economic perspective on path planning. Specif-
ically, we investigate strategic interactions among path planning agents using a
game theoretic path planning framework. Our focus is on economic tension be-
tween two important objectives: efficiency in the agents' achieving their goals,
and safety in navigating towards these. We begin by developing a novel mathe-
matical formulation for path planning that trades off these objectives, when be-
havior of other agents is fixed. We then use this formulation for approximating
Nash equilibria in path planning games, as well as to develop a multi-agent co-
operative path planning formulation. Through several case studies, we show that
in a path planning game, safety is often significantly compromised compared to
a cooperative solution.
Keywords: Multi-agent system · Path planning
1
Introduction
Path planning is a fundamental technical problem in autonomous robotic control. Decades
of development have led to significant theoretical and algorithmic progress, with au-
tonomous vehicles (including autonomous cars and UAVs) increasingly finding their
way to urban roads and skies.
In much of the research on path planning, including mobile robot navigation [10, 1,
11], a fundamental task is to find a collision-free motion from a starting position to the
goal position given a collection of known obstacles. Variations on this theme, such as
dealing with stochastic and moving obstacles, have received recent attention with the
emergence of numerous novel unmanned robotic systems and aerial vehicles [16, 2, 9].
As interactions among autonomous vehicles, be it on our roads or in the skies, be-
comes more routine, we can expect a certain amount of conflict to emerge, as the au-
tonomous agents, designed in service of their individual goals, must occasionally find
these goals dependent on other autonomous agents nearby. However, remarkably lit-
tle research has been devoted to the question of what autonomous vehicle ecosystem
would thereby emerge, when many autonomous agents attempt to achieve their individ-
ual goals, but must necessarily interact with one another in doing so.
To investigate the consequences of such strategic interactions among multiple path
planners, we propose a study of path planning games. An important feature of such
games is that a collection of self-interested path planners each trade off two objectives:
efficiency, or speed with which their goals are achieved, and safety, or probability that
they crash before reaching their goals. Moreover, they trade these off in individual,
2
Yi Li and Yevgeniy Vorobeychik
potentially diverse, ways. Consequently, in order to study path planning games we must
take an economic, rather than a purely algorithmic, perspective on path planning.
To this end, we first develop a novel mathematical programming method for com-
puting a single-agent path plan, accounting for these two objectives, given fixed dy-
namic behavior (i.e., path plans) of all other agents, as well stochastic disturbances
in the environment. Next, we propose a simple iterative algorithm, best response dy-
namics, for approximately computing Nash equilibria of path planning games, given
the best response mathematical programs. Finally, we develop a novel mathematical
program for computing a cooperative multi-agent path plan which optimally trades off
efficiency and safety among all agents -- that is, again, taking the economic perspective
on the multi-agent path planning problem.
We numerically investigate path planning games through several case studies in-
volving two and three agents. Our central observation is that as safety becomes more
important to agents, a large gap opens up between safety achieved by a socially optimal
and Nash equilibrium outcomes; in other words, Nash equilibria exhibit significantly
more collisions than desirable by all agents. The main reason for this is that while each
agent is concerned with safety, they only account in their objective for the impact of
collisions on themselves, and not on other agents who crash along with them.
Our observation about safety consequences of path planning games raises a con-
cern as we look towards the future of autonomous vehicles interacting in populated
environments, particularly as they tend to be designed primarily in service of their in-
dividual ends, rather than those of the entire autonomous and non-autonomous vehicle
ecosystem.
2 Related Work
One common paradigm for studying multi-agent path planning problems is by consid-
ering cooperative path planning involving multiple agents. For example, Shen et al. [22]
studied cooperative path planning in UAV control system, while LaValle [15] presented
an algorithm for applying path planning with stochastic optimal control.
Game theoretic problems related to path planning have been considered from sev-
eral perspectives. Closest to traditional path planning are zero-sum models of games
against nature in which agents are designed to be robust against adversarial uncertainty
in the environment [8, 7]. Classic approaches consider rules of interaction and nego-
tiation among self-interested agents, including planning agents [19, 13, 14]. Loosely
related also is the extensive literature on multi-agent learning, in which multiple agents
repeatedly interact in strategic scenarios in which rewards and dynamics depend on all
agents (often modeled as stochastic games) [23].
Another important class of game theoretic models related to path planning are rout-
ing games. The routing games, as a framework for modeling routing traffic in a large
communication network, were first informally discussed by Pigou [18]. This model was
first formally defined by Wardrop [24] based on a flow network under the non-atomicity
assumption. Therefore, equilibrium flows in non-atomic selfish routing games are often
called Wardrop equilibria. Since then, a number of fundamental results for the non-
atomic routing games have been proved by various researchers, such as the existence
Path Planning Games
3
and uniqueness of equilibrium flows [3], first-order conditions for convex program-
ming problem [4], and the theory of general non-cooperative non-atomic games [21].
The seminal work by Roughgarden and Tardos [20] first characterized the gap between
centralized and decentralized control in multi-agent routing problems, formalized as
the price of anarchy, or ratio of socially optimal to worst-case equilibrium outcomes.
Their work explained the principles behind a broad class of counter-intuitive phenom-
ena, such as Braess's Paradox [6].
Both routing games and path planning games investigate the competition among
agents during their navigation tasks (e.g. passing through bottlenecks). However, in
routing games, the state space is a graph-based structure, and the cost of competition
is modeled by a set of latency functions without considering the agents' dynamics,
while path planning games consider the problem at higher fidelity, with a continuous
state space where the latency is caused by the interaction among agents. Moreover, our
model of path planning games allows us to explicitly study the tradeoff agents make
between performance and safety, an issue not considered in routing games.
3 Model
We describe the problem by first introducing the model of agents' motions, and then
formulating the path planning game.
Consider a state space X = Rn. We represent an agent i by a polyhedron described
ijx ≤ bij, j ∈ {0, ..., Mi}}. Each agent
by a collection of Mi hyperplanes: Pi = {aT
polyhedron Pi contains a point ri ∈ X called the reference which rigidly attaches to
the polyhedron such that the state of an agent can be determined by the position of its
reference. We assume that agents move in discrete time, and a control input uit ∈ Uit ⊂
Rm applied to the ith agent at time t moves the agent from state ri,t ∈ X at time t to
state ri,t+1 ∈ X at time t + 1 according to a linear stochastic dynamic model
ri,t+1 = Airit + Biuit + ωi,
(1)
where Ai ∈ Rn×n, Bi ∈ Rn×m, and ωi ∼ N (0, Σi) is the process noise for ith agent
at time t following an n-dimension zero-mean Gaussian distribution with a covariance
matrix Σi.
For each agent we are given its initial placement r0 ∈ X (i.e., where the agent starts)
and a goal rgoal ∈ X which the agent needs to reach. Let ri,0:T =< ri0, ..., riT > be a
state sequence of the (reference point of the) ith agent from time 0 to T and ui,0:T =<
ui0, ..., uiT > be a corresponding control sequence. However, once the agent reaches
its goal, it remains there deterministically, and has no effect on other agents. We aim to
find the optimal control sequence for the ith agent in this stochastic motion model, with
the following criteria in mind:
1. After applying the resulting control sequence, the expected terminal position of the
3. the agent reaches the goal in as few time steps as possible.
2. the upper bound of the probability that the ith agent collides with other agents
ith agent is ri,goal,
should be minimized, and
4
Yi Li and Yevgeniy Vorobeychik
For the moment, we allow no feedback from observed state to control; we relax this
restriction below.
Path Planning Game: Given these models of individual agents, we define a path plan-
ning game by a collection of N agents, with each agent i's action space comprised
Uit. In this game, each agent aims to compute
of all possible control sequences,
an optimal control sequence, given the behavior of others, trading off two objectives:
efficiency, or the number of times steps it takes to reach the goal, and safety, or the
probability of collision. To formalize, let Ti be the expected number of times steps to
reach the goal (if no collision occurs), and Gi the safety margin, related to the upper
bound on the probability of collision as discussed below. An agent i's objective is then
T(cid:81)
t=0
Ji(ui,0:Tmax, u−i,0:Tmax ) = λTi + (1 − λ)Gi,
(2)
What makes this a game is that the safety Gi of an agent i depends on the paths taken
by all agents, rather than i alone. For example, if two agents are moving towards one
another, and directly towards their respective goals, the only way for one of them to
avoid collision is to circumnavigate the other, taking a longer path towards the goal.
Next, we describe how to define and compute Ti and Gi, and compute a best response
for a given agent i, fixing behavior of all others.
4 Computing an Agent's Best Response
An important subproblem of computing a Nash equilibrium of a path planning game is
to compute a best response of an arbitrary agent i when we fix the control policies of all
others. We show that calculating agents' best responses in path planning games amounts
to a single-agent path planning problem with motion uncertainty. Blackmore et al. [5]
previously developed a probabilistic approach for computing a robust optimal path for
a robot in the environment with a static obstacle and motion uncertainty via mathemat-
ical programming. However, in our context, where an agent trades off efficiency and
safety, with stochastic moving obstacles (representing other agents), this prior approach
is inadequate. In this section we develop a novel method for solving such problems.
4.1 Best Response for a Point-Like Agent
First, consider a simple path planning problem illustrated in Figure 1. In this problem,
there is a set of static obstacles and an agent, represented by a point, aiming to find a
collision-free minimum-time path from its initial placement to its goal position under
motion uncertainty. Assume each obstacle has a given collision volume which can be
represented by a polyhedron. To create a mathematical program for solving this prob-
lem, two factors need to be taken into account: goal position constraints and collision
avoidance constraints.
Formally, let rt denote the position of an agent at time t with its initial placement r0
and the goal position rgoal. Suppose that the motion dynamics of the agent follows (1)
(from which, we remove the index i, since there is only one agent). Assume there are K
Path Planning Games
5
(a)
(b)
Fig. 1: Single agent path planning with a point-like agent.
obstacles represented by polyhedra Pn, n = 1, ..., K, with Pn = {xaT
npx ≤ bnp, p =
1, . . . , En}, where En is the number of hyperplanes representing the polyhedron Pn.
As before, let T denote the planning horizon (so that the goal must be reached by time
T ; we assume the horizon is long enough that the goal can be successfully reached even
with the obstacles).
Efficiency and Reachability: Let {d0, ..., dT} denote a collection of binary indicators
which indicate whether the agent has reached its goal, i.e., dt = 1 iff rt = rgoal. Then,
with a large positive number M, the constraints
∀t,rt − rgoal ≤ M (1 − dt)
T(cid:88)
t=0
dt = 1
(3)
(4)
make sure that the agent will reach to its goal position sooner or later (and we assume
that there exists a feasible solution). Moreover, the number of time steps to reach its
goal position can be represented by
T(cid:88)
t=0
T =
t · dt
(5)
which is one of our objectives (corresponding to Ti, for an agent i above). Since rt
is a random variable, this constrain is stochastic. We approximate it by a determinis-
tic constraint, replacing the position of the agent rt with its expected position rt in
Constraint (3).
Collision Avoidance: Let A denote the event that the agent has a collision, and let
A(n, t), n ∈ {1, ..., K} denote the event that the agent collides with the nth obstacle at
time step t. We wish to minimize the probability of a collision, P r(A), or minimize G
such that
(6)
The agent has a collision if the agent collides with any of obstacles at any time steps,
which is the event
A =
A(n, t)
(7)
P r(A) ≤ G.
T(cid:95)
K(cid:95)
t=0
n=1
6
Yi Li and Yevgeniy Vorobeychik
Then, by the union bound
P r(A) = Pr
(cid:32) T(cid:95)
K(cid:95)
t=0
n=1
(cid:33)
A(n, t)
Pr(A(n, t)) ≤ G
K(cid:88)
≤ T(cid:88)
T(cid:88)
K(cid:88)
n=1
t=0
t=0
n=1
⇐ [∀n, t, P r(A(n, t)) ≤ g(n, t)] ∧ [
g(n, t) = G],
where g(·) is risk allocation which indicates how the risks are distributed among obsta-
cles and time steps. Next, we consider the event that the agent collides with an obstacle
at time step t, which means that the position of the agent is inside the corresponding
polyhedron. Thus, collision with the nth obstacle can be described by
A(n, t) :
np · rt ≤ bnp
aT
(10)
Since the condition (10) including rt is also stochastic, to convert it into a deterministic
one, we consider its probabilistic measure, Pr{A(n, t)}. Following (6), our constraints
then become
(cid:41)
np · rt ≤ bnp
aT
≤ g(n, t).
Since a polyhedron is convex, a sufficient condition is,
Pr(aT
np · rt ≤ bnp) ≤ g(n, t).
(8)
(9)
(11)
(12)
En(cid:94)
p=1
Pr
(cid:40) En(cid:94)
En(cid:95)
p=1
p=1
Based on the approach by Blackmore et al. [5], expression (11) can be further simplified
using the linear approximation of the upper bound on the probability of collision. First,
consider rt, the position of agent at time step t given its initial placement r0 and the
control sequence u0:t, which is a random variable following a Gaussian distribution,
rt ∼ N (rt, Σt), where
and
rt =
Σt =
At−k−1Buk + Atr0
At−k−1Σ(AT )t−k−1.
(13)
(14)
For a single Gaussian random variable X ∼ N (µ, σ2), we can take the inverse Gaussian
distribution function at both sides of Pr(X < 0) ≤ δ and get u ≥ √
2σerf−1(1 − 2δ).
nprt−bnp) ∼ N (aT
Similarly, from rt ∼ N (rt, Σt), we can get (aT
npΣtanp).
Then, we take the inverse Gaussian distribution function at both sides of (12), and
nprt−bnp, aT
nprt − bnp ≥ e(n, t)
aT
(15)
t−1(cid:88)
t−1(cid:88)
k=0
k=0
En(cid:95)
p=1
(cid:113)
2aT
npΣtanp · erf
−1(1 − 2g(n, t)) and erf (z) = 2√
dt. We
where e(n, t) =
call this the safety margin, because it expands the margin of obstacles and shrinks the
feasible planning domain in order to consider motion uncertainty. Because the motion
dk to
of the agent after it reaches its goal has no further effect, we add the term M
these constraints where M is a large positive number.
−1(1 − 2g(n, t)). Since erf
−1 is strictly monotonically in-
Define s(n, t) = erf
k=0
π
creasing, we can minimize
g(n, t) by minimizing
Path Planning Games
7
(cid:82) z
0 e−t2
t(cid:80)
T(cid:80)
K(cid:80)
G = − T(cid:88)
K(cid:88)
t=0
n=1
s(n, t).
(16)
t=0
n=1
This is the safety portion of an agent's objective (Gi for an agent i above).
A Path Planning Mathematical Program: Our goal is to minimize J = λT + (1 −
α)G, balancing efficiency and safety using an exogenously specified parameter λ. Com-
bining this objective with the goal and collision avoidance constraints described above,
we obtain the following mathematical program for single-agent path planning:
MP1:
λT (d) + (1 − λ)G(s)
min
u,s,d
s.t.
∀t, ut ∈ Ut
∀t, rt =
At−k−1Buk + Atr0
∀t,rt − rgoal1 ≤ M · (1 − dt)
∀t, dt ∈ {0, 1}
t−1(cid:88)
k=0
T(cid:88)
t=0
∀t∀n,
dt = 1
En(cid:95)
p=1
(cid:113)
n,prt > bnp + e(n, t) − M
aT
t−1(cid:88)
At−k−1Σ(AT )t−k−1
aT
npΣtanp
∀t, e(n, t) = s(n, t)
∀t, Σt =
∀t∀n, 0 ≤ s(n, t) ≤ M(cid:48)
k=0
(17)
(18)
(19)
(20)
(21)
(22)
(23)
(24)
(25)
(26)
t(cid:88)
k=0
dk
One residual concern is that if an agent cannot possibly collide with an nth obstacle at
time step t (i.e., if g(n, t) = 0), s(n, t) can become unbounded. To address this, we add
Constraint (26) which imposes an upper bound M(cid:48) on s(·), where M(cid:48) is an appropriate
positive number so that erf (M(cid:48)) (cid:39) 1.
8
Yi Li and Yevgeniy Vorobeychik
Since MP1 is a disjunctive linear program which can be solved by an off-the-shelf
linear programming solver. A solution < u, s(·), d > found by MP1 with dT0 = 1
means that the agent can reach to its goal position in T0 time steps with the probability
of collision at most
by applying the control sequence u0:T0.
1−erf (s(n,t))
T0(cid:80)
K(cid:80)
2
t=0
n=1
4.2 Generalization: Feedback Control
Above we considered open loop path planning where the control sequence is determin-
istic and fixed a priori. We now extend our approach to closed loop (feedback) control,
following the ideas in Geibel and Wysotzki [12] and Oldewurtel et al. [17].
Assume we have a nominal control sequence u0:T . Then, the feedback control se-
quence can be obtained by integrating the nominal control sequence and the feedback
gain:
(27)
where xt is the observed and xt the predicted position, and K is an exogenous pa-
rameter which determines the importance of the error feedback term (xt − xt). In this
approach, ut is computed using the MP1 offline, and the actual control sequence is then
generated at runtime by applying (27). As a consequence, the Constraints (25) above
become
ut = ut + K(xt − xt),
t−1(cid:88)
Σt =
(A + BK)t−k−1Σ[(A + BK)T ]t−k−1.
(28)
k=0
Notice that when there is no error feedback (K = 0) this becomes equivalent to open
loop control.
4.3 Collision Avoidance for Polyhedral Agents
Having considered the problem for point-like agents, and then generalizing the ap-
proach to consider error feedback, we now generalize the collision avoidance con-
straints to polyhedral agents.
Consider states of the agent and the nth obstacle, both represented by polyhedra
Pt and Pn, respectively. The position of the agents' reference is rt. Since the reference
point rigidly attaches to the agent, let C = {x−rtx ∈ Pt} denote the relative region of
the agent to its time-dependent reference. When the agent collides with the nth obstacle
at time t, we know that ∃x ∈ Pt ∩ Pn (i.e., the intersection of these time-dependent
polyhedra is non-empty). Thus, from the point view of the agent, the set of positions of
its reference causing collision with the nth obstacle can be represented by Kn = {x −
cx ∈ Pn, c ∈ C} = −C ⊕ Pn, where ⊕ is the Minkowski addition. Since both C and
Pn are polyhedra, Kn is a polyhedron and can be represented by a set of hyperplanes:
Kn = {xaT
npx ≤ bnp, p = 0, ..., En}, where En the number of hyperplanes of Kn.
The agent collides with the nth obstacle at time step t if the position of its reference is
in Kn, that is, when
nprt ≤ bnp.
aT
(29)
rt ∈ Kn ⇔ En(cid:94)
p=1
Comparing (29) with (10), we can see that the problem with polyhedral agents can also
be solved via the mathematical program above, if we treat the agent as its reference
point, and assign the collision volume Kn to each obstacle.
Path Planning Games
9
4.4 Best Response Solver
Our final challenge is to consider the actual best response problem of an arbitrary agent
in the path planning game, where all other agents are moving (rather than static) ob-
stacles with known stochastic motion policies. We now address this problem, obtaining
the final mathematical program for computing a single-agent best response.
Let i denote the agent for whom we are computing a best response, with −i =
{1, ..., i − 1, i + 1, ..., N} the set of all others. Let i be represented by a polyhedron
Pit with reference rit and let j ∈ −i be represented by Pjt with reference rjt. Let Ci
denotes the relative region of i to its reference, while Cj denotes the relative region of
j ∈ −i to its reference. Suppose that j reaches its goal position by time step Tj with
the corresponding known control sequences uj,0:Tj . Then, for each j and t, Kijt =
−Ci ⊕ Pjt is a polyhedron with Kijt = {xaT
ijpx ≤ bijtp, p ∈ {0, ..., Eij}} where Eij
is the number of hyperplanes related to the shapes of Ci and Cj.
Now we formalize how the control sequence uj,0:Tj of each agent j affects Pjt so
that we can determine Kijt. From motion dynamics of i and j,
rit =
∀j, rjt =
At−k−1
i
Biuik + At
iri0 + ωit
At−k−1
j
Bjujk + At
jrj0 + ωjt
(30)
(31)
t−1(cid:88)
t−1(cid:88)
k=0
k=0
t−1(cid:88)
t−1(cid:88)
k=0
k=0
t−1(cid:88)
From the perspective of agent i, the motion of agent j can be treated as deterministic if
we "migrate" motion uncertainty from j to i so that
∀j, r(cid:48)
ijt =
∀j, r(cid:48)
jt =
At−k−1
i
Biuik + At
iri0 + ωit − ωjt
At−k−1
j
Bjujk + At
jrj0.
(32)
For each j, let ω(cid:48)
uncertainty of i to j at time t. Let
ijt = (ωit − ωjt) ∼ N (0, Σit + Σjt) denote the relative motion
∀j, ∆r(cid:48)
jt =
At−k−1
j
Bjujk + At
jrj0 − rj0
(33)
k=0
denote the position shift of agent j at time step t determined by its control sequence
uj,0:Tj . Then, we obtain the position of Kijt by shifting Kij0 by ∆rjt. Since Kijt =
{xaT
ijpx ≤ bijtp}, we obtain
bijtp = bij0p + aT
ijp · ∆r(cid:48)
jt.
(34)
10
Yi Li and Yevgeniy Vorobeychik
Consequently, we obtain the following mathematical program for i's best response:
MP2:
Ji = λTi + (1 − λ)Gi
min
u,si(·),d
s.t.
∀t, uit ∈ Uit
∀t, rit =
iri0
∀t,rit − ri,goal1 ≤ M · (1 − dit)
∀t, dit ∈ {0, 1}
Biuik + At
t−1(cid:88)
At−k−1
k=0
i
dit = 1
∀j∀t = 0, ..., Tj,
T(cid:88)
Eij(cid:95)
t=0
p=1
− M
dik
k=0
∀i∀t, Σit =
i,j,prit > bij0p + aijp · ∆r(cid:48)
aT
t(cid:88)
jt + eijt
k=0
t−1(cid:88)
t−1(cid:88)
ijp(Σit + Σjt)aijp · si(j, t)
aT
Bjujk + At
At−k−1
k=0
j
jrj0 − rj0
∀t∀j, ∆r(cid:48)
jt =
(cid:113)
∀j, eijt =
∀t∀n, 0 ≤ si(n, t) ≤ M(cid:48)
(Ai + KiBi)t−k−1Σi[(Ai + KiBi)T ]t−k−1
(35)
(36)
(37)
(38)
(39)
(40)
(41)
(42)
(43)
(44)
(45)
Notice that the constraints (41) are effective only for t = 0, ..., Ti, and i is not affected
by any j who reached its goal.
5 Finding Equilibria in Path Planning Games
Armed with the best response solvers for each agent i in a path planning game, our goal
is to approximate a Nash equilibrium in the resulting game. We do so by applying best
response dynamics which, if it converges (which it does in our experiments), yields a
Nash equilibrium.
Best response dynamics is an asynchronous iterative algorithm in which a single
agent i is chosen in each iteration, and we maximize i's utility (i.e., compute its best
response) fixing control strategies for all other agents. Best response of an agent i can
be calculated as discussed above.
6 Optimal Multi-Agent Path Planning
Path Planning Games
11
jectives, i.e., the new objective is J = (cid:80)
We now extend the single-agent best response problem to compute an optimal multi-
agent path plan. In this case, the control sequences ui,0:Ti of all agents are unknown
a priori (as they are being computed jointly). Compared to calculating an agents' best
response, we replace the objective of the current agent with the sum of all agents' ob-
i Ji, where Ji is the objective of agent i.
Moreover, we add constraints analogous to MP2 to make sure that the collision avoid-
ance conditions hold from the perspective of every agent simultaneously. We thus obtain
the following mathematical program:
MP3:
N(cid:88)
i=1
Ji
min
u,s,d
J =
At−k−1
i
t−1(cid:88)
s.t.
∀i, t, uit ∈ Uit
∀i, t, rit =
iri0
∀i, t,rit − ri,goal1 ≤ M · (1 − dit)
∀i, t, dit ∈ {0, 1}
∀i
Tmax(cid:88)
Biuik + At
dit = 1
k=0
t=0
∀i, t,−i,
Ei,−i(cid:95)
p=1
(46)
(47)
(48)
(49)
(50)
(51)
(52)
(53)
(54)
(55)
i,−i,prit > bi,−i,0,p + ai,−i,p · ∆r(cid:48)
aT
−i,t
+ ei,−i,t − M
(dik + d−i,k)
t(cid:88)
k=0
t−1(cid:88)
t−1(cid:88)
k=0
k=0
(cid:113)
∀i∀t, Σit =
(Ai + KiBi)t−k−1Σi[(Ai + KiBi)T ]t−k−1
∀i, t, ∆r(cid:48)
it =
At−k−1
i
Biui,k + At
iri0 − ri0
i,−i,p(Σit + Σ−i,t)ai,−i,p · si(t, j)
∀i, t∀ − i, ei,−i,t =
aT
∀i∀t∀n, 0 ≤ si(n, t) ≤ M(cid:48)
t(cid:80)
The term −M
(dik + d−i,k) in Constraints (52) means that an agent will not be
affected by other agents who have reached their goal position by time step t, and, con-
versely, it will not affect the final solution once it reaches its goal position.
k=0
12
Yi Li and Yevgeniy Vorobeychik
7 Experiments
(a)
(b)
(c)
(d)
Fig. 2: Experiment scenarios.
Armed with the techniques for computing both Nash equilibria in path planning
games, as well as a socially optimal solution of the corresponding "cooperative" multi-
agent planning scenario, we now consider several case studies to understand the impact
of self-interested behavior. Specifically, we consider the following 2D scenarios:
-- 2 agents with opposing goal positions (Figure 2(a)): the goal position of each
agent is behind the initial placement of the other. In this scenario, the first agent
moves from starting coordinate position (10, 50) to goal at position (95, 50), and
the second agent moves from (90, 50) to (5, 10).
-- 2 agents moving in parallel (Figure 2(b)): the initial and goal positions of both
agents are near one another. In this scenario, the first agent moves from (10, 70) to
(95, 70) and the second agent moves from (10, 35) to (95, 35).
-- Intersection with 2 agents (Figure 2(c)): one agent moves from the bottom to the
top of the 2D grid, and the other moves from left to right. In this scenario the first
agent moves from (10, 50) to (90, 50) and the second agent moves from (50, 10) to
(50, 90).
-- Intersection with 3 agents (Figure 2(d)): one agent starts at the top of a 2D grid
and moves down, while the other two start at southeast and southwest, and move
northwest and southeast, respectively. In this scenario the first agent moves from
(50, 90) to (50, 5), the second agent moves from (85, 30) to (11, 73), and the third
agent moves from (14, 29) to (90, 73).
In each experiment, each agent is represented by a square with each side of length 15
and parallel to either the x or the y axis. The control inputs are 2D velocity vectors and
the maximum velocity of agents in both x and y direction is 10 (thus, A = B = I in
agents' motion dynamic). Agents' motion is distorted by a Gaussian distribution with
the covariance matrix 1.9I. For each scenario we consider solutions with and without
feedback control, where the feedback gain for the latter was chosen to be K = 0.5.
Throughout, we assume that all players are equally concerned about safety vs. effi-
ciency; formally, all players share the same parameter λ.
The results are shown in Figures 3-10. In each figure, the horizontal axis is the λ
value which represents the importance of safety for both agents, where lower values of
λ imply that safety is more important. The left plots show the objective value, where
lower is better. The middle plots give the time to goal, where lower is, again, better.
The right plots show safety margin, where again lower is better. We present average
Path Planning Games
13
quantities over all agents; the qualitative observations are similar if we consider these
at individual agent level.
The first observation is that the difference between socially optimal and equilibrium
objective values appears small ((a) plots in Figures 3-10). It is therefore tempting to
conclude that equilibrium behavior is similar to socially optimal, but it turns out that
this is not the case: in particular, it turns out that the trade-off between efficiency and
safety made by the agents in equilibrium is very different from optimal.
Considering next the (b) and (c) columns of the figures, we can observe that sys-
tematically performance improves, while safety is often significantly compromised, in
equilibrium as compared to a social optimum. The difference is particularly dramatic in
the first two scenarios, when the agents are in direct conflict in their quest to reach their
respective goals. The gap between optimal and equilibrium safety in the other scenarios
tends to be larger for relatively high values of λ.
Another general observation we can make is that often the solutions with a feedback
controller are closer to optimal, particularly from the perspective of safety. The excep-
tions involve the intersection scenarios, where the gap is larger for higher values of λ
in the feedback controller solution than with the open-loop controller. However, even in
these scenarios, the feedback controller yields solutions closer to socially optimal for
most values of λ. This is not surprising: since all agents are concerned about safety,
they are more able to dynamically adjust to avoid collisions when some feedback about
state is available.
To understand why safety is systematically compromised, consider a single agent's
incentive. Even though an agent is interested in reaching the goal safely, it does not
account for the fact that being involved in a crash also crashes the other agent. Thus, in
equilibrium safety is compromised relative to social optimum, as agents fail to capture
the externalities associated with crashes.
(a)
(b)
(c)
Fig. 3: Opposing goal positions without the feedback gain(K = 0).
(a)
(b)
(c)
Fig. 4: Opposing goal positions with the feedback gain(K = 0.5).
14
Yi Li and Yevgeniy Vorobeychik
(a)
(b)
(c)
Fig. 5: Moving in parallel without the feedback gain(K = 0).
(d)
(e)
(f)
Fig. 6: Moving in parallel with the feedback gain(K = 0.5).
(a)
(b)
(c)
Fig. 7: Intersection without the feedback gain(K = 0, 2 agents).
(a)
(b)
(c)
Fig. 8: Intersection with the feedback gain(K = 0.5, 2 agents).
(a)
(b)
(c)
Fig. 9: Intersection without the feedback gain(K = 0, 3 players).
(a)
(b)
(c)
Fig. 10: Intersection with the feedback gain(K = 0.5, 3 players).
Bibliography
[1] Arkin, R.C.: Motor Schema-Based Mobile Robot Navigation (1989)
[2] Auat Cheein, F.A., Carelli, R.: Agricultural robotics: Unmanned robotic service units in
agricultural tasks. IEEE Industrial Electronics Magazine 7(3), 48 -- 58 (2013)
[3] Beckmann, M., McGuire, C., Winsten, C.: Studies in the Economics of Transportation
(1956)
[4] Bertsekas, D.: Nonlinear programming (1999)
[5] Blackmore, L., Li, H., Williams, B.: A probabilistic approach to optimal robust path plan-
ning with obstacles. In: American Control Conference (2006)
[6] Braess, D.: Uber ein Paradoxon aus der Verkehrsplanung. Unternehmensforschung Opera-
tions Research - Recherche Op´erationnelle 12(1), 258 -- 268 (dec 1968)
[7] Chen, M., Zhou, Z., Tomlin, C.J.: Multiplayer reach-avoid games via low dimensional so-
lutions and maximum matching. In: American Control Conferenc. pp. 1444 -- 1449 (2014)
[8] Chen, M., Zhou, Z., Tomlin, C.J.: A path defense approach to the multiplayer reach-avoid
game. In: Annual Conference on Decision and Control. pp. 2420 -- 2426 (2014)
[9] Craighead, J., Murphy, R., Burke, J., Goldiez, B.: A survey of commercial & open source
unmanned vehicle simulators. In: Proceedings - IEEE International Conference on Robotics
and Automation. pp. 852 -- 857 (2007)
[10] DeSouza, G.N., Kak, A.C.: Vision for mobile robot navigation: A survey. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence 24(2), 237 -- 267 (2002)
[11] Elfes, A.: Using occupancy grids for mobile robot perception and navigation. Computer
22(6), 46 -- 57 (1989)
[12] Geibel, P., Wysotzki, F.: Risk-sensitive reinforcement learning applied to control under con-
straints. J. Artif. Intell. Res. 24, 81 -- 108 (2005)
[13] Jonsson, A., Rovatsos, M.: Scaling up multiagent planning: A best-response approach. In:
ICAPS (2011)
[14] Jord´an, J., Torreno, A., de Weerdt, M., Onaindia, E.: A better-response strategy for self-
interested planning agents. Applied Intelligence pp. 1 -- 21 (2017)
[15] LaValle, S.M.: Robot motion planning: A game-theoretic foundation. Algorithmica 26(3-4),
430 -- 465 (2000)
[16] Mahony, R., Kumar, V.: Aerial robotics and the quadrotor. IEEE Robotics and Automation
Magazine 19(3), 19 (2012)
[17] Oldewurtel, F., Jones, C.N., Morari, M.: A tractable approximation of chance constrained
stochastic mpc based on affine disturbance feedback. In: IEEE Conference on Decision and
Control. pp. 4731 -- 4736 (2008)
[18] Pigou, A.: The economics of welfare, 1920. McMillan&Co., London (1932)
[19] Rosenschein, J.S., Zlotkin, G.: Rules of encounter: designing conventions for automated
negotiation among computers. MIT press (1994)
[20] Roughgarden, T., Tardos, E.: How bad is selfish routing? Proceedings 41st Annual Sympo-
sium on Foundations of Computer Science 49(2), 1 -- 26 (2000)
[21] Schmeidler, D.: Equilibrium points of nonatomic games. Journal of Statistical Physics 7(4),
295 -- 300 (1973)
[22] Shen, D., Chen, G., Cruz, J.B., Blasch, E.: A game theoretic data fusion aided path planning
approach for cooperative UAV ISR. In: Aerospace Conference. pp. 1 -- 9 (2008)
[23] Stone, P., Veloso, M.: Multiagent systems: A survey from a machine learning perspective.
Autonomous Robots 8(3), 345 -- 383 (2000)
[24] Wardrop, J.G.: Road paper. some theoretical aspects of road traffic research. Proceedings
of the institution of civil engineers 1(3), 325 -- 362 (1952)
|
1903.00714 | 1 | 1903 | 2019-03-02T14:55:40 | A Cooperative Multi-Agent Reinforcement Learning Framework for Resource Balancing in Complex Logistics Network | [
"cs.MA",
"cs.AI"
] | Resource balancing within complex transportation networks is one of the most important problems in real logistics domain. Traditional solutions on these problems leverage combinatorial optimization with demand and supply forecasting. However, the high complexity of transportation routes, severe uncertainty of future demand and supply, together with non-convex business constraints make it extremely challenging in the traditional resource management field. In this paper, we propose a novel sophisticated multi-agent reinforcement learning approach to address these challenges. In particular, inspired by the externalities especially the interactions among resource agents, we introduce an innovative cooperative mechanism for state and reward design resulting in more effective and efficient transportation. Extensive experiments on a simulated ocean transportation service demonstrate that our new approach can stimulate cooperation among agents and lead to much better performance. Compared with traditional solutions based on combinatorial optimization, our approach can give rise to a significant improvement in terms of both performance and stability. | cs.MA | cs | A Cooperative Multi-Agent Reinforcement Learning Framework
for Resource Balancing in Complex Logistics Network
Xihan Li
Key Laboratory of Machine
Perception, Peking University
Beijing, China
[email protected]
Jia Zhang
Microsoft Research Asia
Beijing, China
[email protected]
Jiang Bian
Microsoft Research Asia
Beijing, China
[email protected]
9
1
0
2
r
a
M
2
]
A
M
.
s
c
[
1
v
4
1
7
0
0
.
3
0
9
1
:
v
i
X
r
a
Yunhai Tong
Key Laboratory of Machine
Perception, Peking University
Beijing, China
[email protected]
Tie-Yan Liu
Microsoft Research Asia
Beijing, China
[email protected]
ABSTRACT
Resource balancing within complex transportation networks is one
of the most important problems in real logistics domain. Traditional
solutions on these problems leverage combinatorial optimization
with demand and supply forecasting. However, the high complex-
ity of transportation routes, severe uncertainty of future demand
and supply, together with non-convex business constraints make
it extremely challenging in the traditional resource management
field. In this paper, we propose a novel sophisticated multi-agent
reinforcement learning approach to address these challenges. In
particular, inspired by the externalities especially the interactions
among resource agents, we introduce an innovative cooperative
mechanism for state and reward design resulting in more effective
and efficient transportation. Extensive experiments on a simulated
ocean transportation service demonstrate that our new approach
can stimulate cooperation among agents and lead to much better
performance. Compared with traditional solutions based on combi-
natorial optimization, our approach can give rise to a significant
improvement in terms of both performance and stability.
KEYWORDS
multi-agent; reinforcement learning, resource balancing, logistics
network
ACM Reference Format:
Xihan Li, Jia Zhang, Jiang Bian, Yunhai Tong, and Tie-Yan Liu. 2019. A
Cooperative Multi-Agent Reinforcement Learning Framework for Resource
Balancing in Complex Logistics Network. In Proc. of the 18th International
Conference on Autonomous Agents and Multiagent Systems (AAMAS 2019),
Montreal, Canada, May 13 -- 17, 2019, IFAAMAS, 14 pages.
1 INTRODUCTION
With the rapid growth of logistics industry, the imbalance between
the resource's supply and demand (SnD) has become one of the
most important problems in many real logistics scenarios. For ex-
ample, in the domain of ocean transportation, the SnD of empty
Proc. of the 18th International Conference on Autonomous Agents and Multiagent Systems
(AAMAS 2019), N. Agmon, M. E. Taylor, E. Elkind, M. Veloso (eds.), May 13 -- 17, 2019,
Montreal, Canada. © 2019 International Foundation for Autonomous Agents and
Multiagent Systems (www.ifaamas.org). All rights reserved.
containers are very unequal due to the world trade imbalance [20];
in the domain of express delivery, there exists severe emerging
unevenness of the SnD of carriers within local areas; in the fast-
growing car-sharing and bike-sharing areas, the unbalanced SnD of
shared taxis and bikes are also explicit due to various temporal and
spatial factors [10, 17]. Henceforth, efficient resource balancing has
risen to be the critical approach to solve the resource imbalance in
the logistics industry. The failure of that will cause large amounts
of unfulfilled resource demand, further resulting in reduction of
customer satisfaction, increasing resource shortage cost and de-
clining revenue. Persistent unsolved SnD imbalance can give rise
to accumulated resource shortage and, even worse, a stalemate of
SnD [17] with unexpected amplified price.
Traditional solutions for resource balancing leverage operational
research (OR) based methods [20], which are typically multistage:
they first use forecasting techniques to estimate the future SnD of
each resource agent; then, the combinatorial optimization approach
is employed to find each resource agent's optimal action to mini-
mize a pre-defined objective, which is usually formed as the total
cost caused by resource shortage; finally, the feasible execution
plan is generated by tailoring the raw solution obtained by OR-
based models. Nevertheless, the drastic uncertainty of future SnD,
complex business constraints in the non-convex form, as well as
the high complexity of transportation networks make it extremely
challenging to generate satisfying action plans by using traditional
OR solutions.
More concretely, the first crucial challenge, i.e., the uncertainty
of future SnD, is mainly caused by multiple external highly dynamic
factors, either temporal or spatial, such as special days/events,
emerging market changes, unstable policies [20], etc. Moreover,
such uncertainty can be even aggravated due to the inherent mu-
tual dependency between the OR-based model and future SnD.
Particularly, the future SnD can be dramatically deviated by action
plans generated by the OR model, which in turn heavily relies on
the future SnD. Henceforth, the uncertainty of future SnD, as drasti-
cally increasing the difficulty of accurate SnD forecasting, tends to
fail the effectiveness of the traditional multistage OR-based method.
The second major challenge is reflected by many important but
complex business rules in real logistics services. On the one hand,
they are hard to be formulated in constraints of linear or convex
forms, which, therefore, makes it quite hard to model and solve
the problem precisely using traditional OR-based method such as
linear programming and convex optimization. On the other hand,
ignoring these necessary constraints is unacceptable since it will
cause a big gap between the model and the real world, leading to
significant performance drop and even unfeasible solutions.
Furthermore, since the transportation networks in real logistics
services are usually very complex, consisting of various types of
terminals and complex connecting routes, the consequential com-
plicated dependencies among terminals rise another vital challenge
when building effective OR-based model. Specifically, those com-
plicated dependencies make it quite difficult to create acceptable
number of constraints and variables to balance between the indi-
vidual and the collective objectives in the OR-based model.
To address these challenges, in this paper, we formally formulate
the resource balancing problem in complex logistics networks as a
stochastic game and then propose a novel cooperative multi-agent
reinforcement learning (MARL) framework. With the dedicated de-
sign of the agent set, joint action space, state set, reward functions,
transition probability functions, and discount factor, respectively,
our multi-agent reinforcement learning framework provides an
end-to-end and high-capability solution, which can not only com-
pensate the imperfect forecasting results to avoid further error
propagation in multistage OR methods, but also enable to optimize
the obtained action plans towards complicated constraints based on
real business rules. Moreover, in contrast to applying MARL under
some easier logistics scenarios, a blind employment of reinforce-
ment learning approach may not produce satisfactory results in
complex logistics networks, because of its incapability of enhancing
cooperation among highly dependent resource agents. To tackle
this challenge, we further introduce three levels of cooperative
metrics and, accordingly, improve the state and reward design to
better promote the cooperation in the complex logistics networks.
To demonstrate the superiority of the MARL framework, we
implement our approach under an empty container repositioning
(ECR) task in a complex ocean transportation network. In fact, such
maritime transportation is essential to the world's economy as 80%
of global trade is carried by sea [21]. By far, maritime transporta-
tion is the most cost-effective way to move bulk commodity and
raw materials around the world. Extensive experiments show that
our method can achieve nearly optimal resource balancing results,
which yields a significant improvement over the traditional OR
baseline.
transportation network as a stochastic game.
Our major contributions can be summarized as follows:
• Formulating the resource balancing problem in a complex
• Introducing a cooperative multi-agent reinforcement learn-
ing framework as an end-to-end and high-capability solution
to the resource balancing problem, as it is not only more
robust to the imperfect SnD forecasting but yields higher
capability and flexibility compared with the traditional mul-
tistage OR-based methods.
• Proposing three levels of cooperative metrics to provide guid-
ance to improve state and reward design, in order to better
promote the cooperation in the complex logistics network.
• Conducting extensive experiments on the empty container
repositioning task in the scenario of real-world ocean logis-
tics industry.
2 RELATED WORKS
Resource balancing in transportation network, which can be re-
garded as a branch of scheduling problem, is comprehensively stud-
ied in the field of OR [2, 5, 9, 18]. Among them, Epstein et al. [5]
studied the ECR problem, and developed a logistics optimization
system to manage the imbalance with a multicommodity network
flow model based on demand forecasting and safety stock con-
trol. For more works about ECR, Song and Dong [20] provides an
in-depth review of the OR-based literature.
With the prosperity of deep learning, deep reinforcement learn-
ing (RL) methods like DQN [15] has achieved great success in
modeling and solving many intellectual challenging problems, such
as video games [15] and go [19]. However, they are not widely
applied to complicated real-world applications, especially for those
who have high-dimensional action spaces and need cooperation
between lots of agents.
In recent years, motivated by the great success of deep RL, some
methods have been proposed based on RL to address resource bal-
ancing problem, especially rebalancing homogeneous, flexible ve-
hicles. Pan et al. [17] proposed a deep reinforcement learning al-
gorithm to tackle the rebalance problem for shared bikes, which
learns a pricing strategy to incentivize users to rebalance the sys-
tem. Lin et al. [10] proposed a contextual multi-agent reinforcement
learning framework to tackle the rebalance problem for online ride-
sharing platforms, in which every taxi is treated as an agent that
learns its action to move to its neighboring grids. Xu et al. [23] pro-
posed a learning and planning approach in on-demand ride-hailing
platforms, which combines RL for learning and combinatorial op-
timizing algorithm for planning. These works have successfully
modeled and handled large-scale and real-world traffic scenarios.
However, compared with resource balancing in complicate logis-
tics network, the environments in their scenarios are much looser,
and the dependency of agents is simple and straightforward. Thus
their methods can hardly be applied to solve the resource balancing
problem.
To apply MARL in resource balancing, one of the main obstacles
is to deal with collaboration of agents with complicated depen-
dency. This dependency is mainly caused by complicated logistics
network structures. In the area of traditional multi-agent system,
fruitful works are done by dealing with collaboration of multi-
agents. Among them, FF-Q [11], Nash-Q [7] and Correlated-Q [6]
are famous methods achieving convergence and optimum. How-
ever, all of them adopt the joint action approach, which is hardly
applied in real-world multi-agent system with lots of agents, due
to the extremely large joint action space. Similar limitation occurs
in other joint action or best response based methods [8, 22]. Some
other works [3, 4, 13] managed to apply potential based reward
shaping in MARL to stimulate cooperation. Methods in these works
achieve performance improvement in their own scenarios. How-
ever, in resource balancing scenarios, where agents' actions have a
long-term and immeasurable effect on the ultimate results, more
efforts should be put to understand the problem and design rewards.
0
i
, Dt
i
, and we use Ct
i
3 PROBLEM STATEMENT
In this section, we will formally define the resource balancing prob-
lem in a complex logistic network.
A typical logistic network can be defined as G = (P, R, V), in
which P, R and V stand for the set of terminals, routes, and vehicles,
respectively. More specifically,
• Each terminal Pi ∈ P represents a place that can store re-
sources and generate corresponding SnD. We denote the
initial resources in stock at Pi as C
, and
(t = 1 · · ·T ) to represent the numbers of stocks, resource
St
i
demands, and resource supplies at different time, respec-
tively.
• Each route Ri ∈ R is a cycle in the logistic network, consist-
ing of a sequence of consecutive terminals {Pi1 , Pi2 , · · · , PiRi },
where Ri is the number of stops on Ri and the next desti-
nation of PiRi is Pi1. Each route can intersect with others
in the network.
• On each route Ri, there is a fixed set of vehicles VRi ⊆ V ,
each of which, Vj ∈ VRi
, yields an initial position, a duration
function dj(Pu , Pv) : P × P → N + (mapping from an origin
terminal Pu and a destination one Pv into the transit time),
a capacity Capt
(the maximum number of resources it can
j
convey). When a vehicle arrives at a terminal, it can either
load resources from or discharge its resources to the terminal.
The objective of resource balancing is to minimize the resource
shortage among all terminals. At a specific time t, the terminal can
only use the stock in the last day, i.e, Ct−1
, to fulfill the current
. 1 Once the stock is not enough, the shortage happens.
demand Dt
i
Thus, we denote the number of shortage as Lt
i
Accordingly, the objective of resource balancing is to minimize the
i − Ct−1
Dt
total resource shortage: L =
i −V
After the current demand is processed, new resource supplies and
those discharged from the vehicle will be added to the stock, thus we
can compute the new stock amount as Ct
+
i
j ∈ N denotes the number of resources
St
loaded onto vehicle Vj at time t. xt
can be negative to denote the
j
discharged amount of resources from the vehicle, and I(i, j, t) is a
indicator variable defined as
(cid:26)1, Vj arrives at Pi at time slot t
j=1 I(i, j, t)xt
= max(cid:16)
= max(cid:16)
, 0(cid:17)
i , 0(cid:17)
Pi ∈P,t ∈T Lt
i
, where xt
Ct−1
i
− Dt
i
.
i
.
j
I(i, j, t) =
We further define Ct
time slot t, and clearly, Ct
V , j
0, otherwise.
as the amount of resources on vehicle Vj at
V , j
= Ct−1
V , j
+ xt
j
.
4 COOPERATIVE MULTI-AGENT
REINFORCEMENT LEARNING
FRAMEWORK
As aforementioned, traditional solutions for resource balancing
employ combinatorial optimization with SnD forecasting. However,
1This is because new supplies and discharged resources at time t are usually un-
available temporarily for realistic reasons, such as inner terminal transportation and
maintenance. This logic can change with specific application scenarios, and will not
affect our framework.
it suffers from failures in front of uncertainty of SnD, complex busi-
ness constraints, and high complexity of transportation networks.
To address these challenges, in this section, we first model the re-
source balancing in complex logistic network as a stochastic game
and then propose a novel cooperative multi-agent reinforcement
learning (MARL) framework to solve it.
4.1 Resource Balancing as a Stochastic Game
The resource balancing problem can be formally modeled as a
stochastic game G = (N , A,S, R, P, γ), where N is the agent set, A
is the joint action space, S is the state set, R is the reward function,
P is the transition probability function, and γ is the discount factor.
More formally definitions are shown below:
Agent set N . We define each vehicle as an agent, which yields
two major advantages: (1) As each vehicle agent continuously sails
circularly along the certain route, it can be aware of the larger
scope of information within the whole route such that optimizing
towards maximizing its own reward, i.e., minimizing the shortage,
can benefit the total reward of the entire route. (2) Since multiple
vehicle agents navigating along the same route usually share the
similar environment, it is natural for them to share the same policy
so as to significantly reduce the model complexity in MARL and
boost the learning process.
Joint action space A. We define the action of a vehicle agent Vj
as loading or discharging resources when it arrives at a terminal
Pi. Similar to Menda et al. [14], we apply the idea of event-driven
reinforcement learning. To be more concrete, we treat agents' each
arrival at a terminal as a trigger event, and an agent only needs to
take action once a trigger event happens. Under this event-driven
to denote the action taken by agent Nj ∈ N at t-th
setting, we use at
j
arrival event. For agent Nj, we define its action space as Aj = [−1, 1],
resources
where at
from the vehicle, at
resources
= 0 means no loading or discharging. Then,
onto the vehicle, and at
j
the joint action space is A = A1 × A2 × · · · × AN , where N is
the number of agents. The total amount of resources that can be
discharged or loaded at t is usually restrictively determined by the
dynamic values of Ct
as well as some other external
i
factors, which are controlled by domain-specific business logics.
State set S. The state S is a finite set that stands for all possible
situations of the whole logistics network. Note that, from a practical
point of view, it is not necessary for the agents to take action based
on the whole state information, due to the extremely large state
space and the potential noise introduced by unrelated information.
We will elaborate more on the practical state design later in this
section.
Rewards function R. The objective of the resource balancing
problem is to minimize the accumulated shortage for all terminals.
With respect to each individual action, i.e., loading or discharg-
ing some resources at a terminal, the impact can be spread to its
follow-up periods. To model such delayed reward, it usually lever-
ages rewards shaping to guide the learning process [16], a typical
specification of which is to measure the difference of the ultimate
accumulated shortage between with and without this action. How-
ever, this reward is very hard to compute in practice. Thus, we find
j ∈ [−1, 0) means discharging a portion of at
j ∈ (0, 1] means loading a portion of at
, Capt
i
, Ct
V , j
j
j
(cid:16)
C
i , · · ·
1
• Current available resources C
• Historical information of available resources ϕ
tk
i
.
(cid:17) and shortages ψ
i , · · · , L
1
L
tk−1
i
tk−1
i
, C
(cid:17).
(cid:16)
tk
berth length, etc.
• Other domain-specific information, such as terminal ID,
where ϕ(·) and ψ(·) denote some statistical function (Mean, Me-
dian, etc.) or more advanced sequential data processing models
(CNN, RNN, etc.). Specific implementation should depend on the
application scenario.
(cid:104)
tk
V , j
j − C
cle type, etc.
for vehicle Vj can be comprised of:
.
tk
j
(cid:105) for self awareness agents. The self awareness agents
State s
• Current available resources onboard C
• Available space Cap
• Other domain-specific information, such as vehicle ID, vehi-
Concatenating the above information, we get the state sI =
tk
P,i , s
only concerns if shortage happens between tk and t′
k ≥ tk
stands for the time of next vehicle's arrival at Pi. Besides, inspired
by the idea of safety stock in traditional methods, we add a small
positive reward if no shortage happens. This reward is calculated
according to a function f : N → R that has diminishing marginal
gain2. The purpose is to encourage the agents to put some safety
stock with upper limit on terminals. In summary, the reward can
be written as follows:
where t′
tk
V , j
.
tk
j
k
s
where д : N → R is the loss defined on the total shortage.
4.2.2 Territorial Awareness. According to the problem defini-
tion, a vehicle needs to navigate along with the certain route and
is obliged to balance the SnD within its own territory, i.e., the ter-
minals in its route. Apparently, each agent with self awareness,
with no consideration on other terminals and vehicles in its route,
cannot balance the resources SnD within its route. Thus, we intro-
duce territorial awareness agent to minimize the total shortage of
all terminals in the route. Specifically, for an agent Vj on route Rq,
we hope the agent to get the accurate information of neighboring
environment on the route, which is more likely to influence the
current decision. We add extra successive information as follows:
• Information about n successive terminals {s
where Sci, j(n) is the set of n terminals to which vehicle Vj
will travel after terminal Pi.
• Information about m future vehicles {s
where Fui, j(m) stands for the set of m vehicles that will
arrive at Pi just after Vj's arrival.
As we can see, the larger n and m are, the more information can be
used for decision. However, in practice, we usually set small values
for n and m to control the model complexity and noise introduced by
unimportant information. To compensate the potential information
loss, we introduce the overall statistical territory information s
tk
R,q
for route Rq:
T ,i′Pi′ ∈ Sci, j(n)(cid:9)
V , j′Vj′ ∈ Fui, j(m)(cid:9)
tk
tk
2For example, f (x) =x
i =0 β i for 0 < β < 1.
(cid:18)
t′
k
i
C
(cid:19)
k
− д(cid:169)(cid:173)(cid:173)(cid:171) t′
t =tk
(cid:170)(cid:174)(cid:174)(cid:172) ,
Lt
i
rI = f
(1)
Figure 1: Illustration of three levels of cooperative metrics.
(a) Self awareness agent V only consider information of
(P, V) to make decision. (b) Territorial agent V1 will make
decision based on information within its territory. It could
load more resources at arrival port P2 with the awareness
that port P4 on its route R has low stock. (c) Agent V1 with
diplomatic awareness can look far beyond its route. It could
load more resources at current port P2 and discharge them
at transshipment port P3 or P4 later with the awareness that
port P6 on its neighboring route R2 needs support.
other more realistic rewards shaping methods, which will also be
discussed later in this section.
Transition probability function P. It is defined as a mapping
S × A × S → [0, 1], which can be specified by the definition of
S, R, V and the distribution behind SnD within particular logistics
networks.
4.2 Cooperative Metrics for State and Reward
Design
After formulating the resource balancing problem as a stochastic
game, applying MARL approach to the real world, however, requires
a dedicated design on the game state and the action's reward to pro-
mote cooperation and improve performance. Based on the scope of
agents' awareness of cooperation, we identify three levels of coop-
erative metrics: self awareness, territorial awareness, and diplomatic
awareness. In general, agents with self awareness are fully selfish
and shortsighted and only consider immediate information and
interests; agents with territorial awareness have a broader vision
and make decision based on information belonging to their terri-
tories, i.e., routes in this problem. At last, agents with diplomatic
awareness even overlook beyond their own routes and conduct
resource balancing, in a diplomatic way, by cooperating with in-
tersecting routes so that resources can flow from fertile routes to
barren routes.
4.2.1
Self Awareness. When agent Vj arrives at terminal Pi, it
is natural that Vj makes decisions just based on the information of
itself and Pi. Regarding the reward of this action, a straightforward
metric is to consider whether shortages will happen before next
vehicle's arrival at Pi. Obviously, this is a very shortsighted agent.
Suppose the time of k-th arrival event of a vehicle agent Vj is tk
for terminal Pi can be
and the arrival terminal is Pi. The state s
formed up by:
tk
P,i
PVP2V1P3P4RP2V1P3P4R1V3P5R2P1V2(a) Self Awareness(b) Territorial Awareness(c) Diplomatic AwarenessP1V2PVRTerminalVehicleResourceStockRouteCurrent EventP6 • Information of available resources in all the terminals in the
• Information of shortage in all the terminals in the route
(cid:16)(cid:110)
route Φ
C
tk
i
(cid:16)(cid:110)
(cid:16)
Ψ
ψ
Pi ∈ Rq
tk−1
i
i , · · · , L
1
L
(cid:111)(cid:17)
(cid:17) Pi ∈ Rq
(cid:111)(cid:17)
Similar as ϕ(·) and ψ(·), Φ(·) and Ψ(·) are statistical functions or
models based on series data.
We concatenate all information above with sI to get the territorial
state sT . Territorial awareness agents will make decision based on
the state sT .
4.2.3 Diplomatic Awareness. In real logistics networks, imbal-
ance can also happen among different routes: there may be a large
amount of supplies but very few demands on some routes, while
some other routes may be opposite, with a large amount of de-
mands that cannot be satisfied with limited supplies. In this case, it
is infructuous to attempt balancing SnD within the territory of sin-
gle route. To solve this problem substantially, agents should learn
the diplomacy: solving imbalance collaboratively with agents in
intersecting routes.
s
tk
(cid:16)(cid:110)
(cid:16)(cid:110)
(cid:111)(cid:17)
(cid:111)(cid:17) where Rti is the
To this end, more information about neighboring routes should
be considered. Assume an event (Pi, Vj, Rq), and denote Crq as the
crossing routes having common terminal(s) with route Rq. First,
R,pRp ∈ Crq
statistic information for all neighboring routes Φr
should be involved to represent the general status of crossing routes.
Moreover, we add additional information when agents arrive at
transfer terminals, that is Φn
set of routes that pass through terminal Pi. We concatenate all
information above with sT as the diplomatic state sD.
To encourage cooperation, we extend the reward by considering
cross routes shortage. For an agent Vj on a route Rq, its action not
only influences the reward on its own route, but also influences
the reward of agents in the neighboring routes in Crq, especially
on the transfer terminals where routes are intersecting. To take
neighboring routes into consideration, we use rD = αrI +(1− α)rC,
where α is a soft hyper-parameter and
R,pRp ∈ Rti
tk
s
(cid:18)
(cid:18)(cid:26)
− д(cid:0)ξ2(cid:0)(cid:8)Lt
t′
k
i
ξ1
C
rC =f
(cid:27)(cid:19)(cid:19)
Pi ∈ Rp, Rp ∈ Crq
i tk ≤ t ≤ t′
k, Pi ∈ Rp, Rp ∈ Crq
(cid:9)(cid:1)(cid:1) ,
for statistical functions or advanced models ξ1(·) and ξ2(·).
The three levels of cooperative metrics are illustrated in Fig-
ure 1. The whole cooperative MARL framework for resource bal-
ancing is shown in Algorithm 1. From Line 4 to 13, the agents
interact with environment by function calls, and collect transition
experiences. It should be emphasized that GetState(cid:16)
triggered. GetDelayedReward(cid:16)
Sj,k , Pi , Vj
refers to the process of constructing state based on current event
(Pi , Vj) and global environment snapshot Sj,k. This snapshot con-
tains complete information of the environment when the event is
(cid:17)
(cid:17) refers to the process
to calculate the delayed reward based on shortage happens between
these two snapshots. The detail implementation of GetState(·) and
GetDelayedReward(·) will be determined based on the adopted
level of cooperative metric.
Sj,k−1, Sj,k
each agent Vj
Algorithm 1 Cooperative MARL Framework
1: Initialize replay memory Dj to capacity M for each agent Vj
2: Initialize action-value function Qj with random weights θj for
3: for episode ← 1 to MAX do
4:
5: while environment is not terminated do
6:
7:
8:
ResetEnvironment()
// k means the k-th event of agent Vj
(Pi , Vj , k) ← WaitingEvent()
Sj,k ← GetEnvironmentSnapshot()
(cid:17)
sk ← GetState(cid:16)
rk−1 ← GetDelayedReward(cid:16)
(cid:17)
StoreExperience(cid:0)Dj ,(sk−1, ak−1, rk−1, sk)(cid:1)
ak ← ϵ-Greedy(cid:0)arg maxa Qj(sk , a)(cid:1)
Sj,k−1, Sj,k
Sj,k , Pi , Vj
Execute(Pi , Vj , ak)
end while
for l ← 1 to MAX-TRAIN do
for each Vj in V do
Sample a batch of data (s, a, r , s′) from Dj
Compute target y ← r + γ maxa′ Qj(s′, a′; θj)
Update Q-network for agent Vj as
θj ← θj − ∇θj(y − Qj(s, a; θj))2
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
end for
20:
21:
22: end for
end for
5 EXPERIMENTS
To evaluate the effectiveness of our proposed approach, we conduct
experiments on resource balancing in the scenario of ocean con-
tainer transportation. In this task, the resource balancing mainly
corresponds to Empty Container Repositioning (ECR). In the fol-
lowing of this section, we will first introduce the background of
ECR, then we will show the experimental results on a part of real
ocean logistics network.
5.1 The ECR Problem
As containers are the most important asset in ocean logistics in-
dustry, the resource balancing in this scenario corresponds to ECR,
which is quite necessary since the SnD of empty containers are very
unequal due to the world trade imbalance [20]. In particular, the
goal of ECR is to reposition empty containers by container vessels
sailing on pre-determined routes within ocean logistics networks
to fulfill the dynamic transportation demand of ports. According to
Asariotis et al. [1], the estimated cost of seaborne empty container
repositioning was about 20 billion dollars in 2009, with 50 million
empty containers movement, which has demonstrated the necessity
to optimize ECR in ocean logistics industry. More formally, ports,
container vessels, and predetermined routes for vessels correspond
to terminals P, vehicles V , and routes R, respectively. External de-
mands and supplies of empty containers for port Pi at time slot t
correspond to Dt
i
, respectively.
and St
i
5.1.1 Domain-specific Features of the ECR Problem. Nonethe-
less, there are several domain-specific feature for the ECR problem.
Figure 2: The container transportation chain in ECR problem. Blue lines indicate laden container flows and green lines indicate
empty container flows. All flows are under the control of specific business logics in real logistics scenarios.
and St
i
In ECR problem, the external demands and supplies Dt
are
i
determined by transportation orders O, which are also external and
dynamic. An order o ∈ O is a tuple (Pu , Pv , n, to), which denotes
departure port, destination port, amount of needed containers and
order time. The container transportation chain for orders can be
described as follows, also illustrated in Figure 2: when an order
(Pu , Pv , n, to) is placed at time slot to, the external demand of depar-
ture port D
u will be added by n, which means Pu need to provide
to
n empty containers to fulfill the order at time slot to. If the order is
fulfilled, cargoes will be loaded into these empty containers, and
they are transformed to laden containers waiting for vessels to
transport them to destination port Pv. Laden containers will be
lifted on the arriving vessel Vj on route Rk if Pv ∈ Rk.3 When the
laden containers are discharged to the destination port Pv at time
slot t′
o, the cargoes in laden containers will be unloaded and these
containers will be returned to Pv as empty containers at time slot
t′
o + tret, in which tret is a constant. Therefore, the external supply
of destination port S
will be added by n. To summarize, the
specification of ECR problem is concluded as follows:
• Empty containers are reusable, which will circulate between
• Laden containers and empty containers share the same ves-
sel. i.e., the space for empty containers Capt
for vessel Vj
j
will change dynamically depending on the amount of laden
containers on the vessel;
• The whole order will fail if not enough empty containers
can be served from departure port when the order is placed.
The resource shortage Lo for a single order o is defined as
Lo = n, when n > C
u , and Lo = 0 for otherwise.
to
ports as receptacles for cargoes;
t′
o +tret
v
corresponds to the state change of containers, i.e. from empty to
laden and vice versa. To build constraints fully representing the
SnD balance, the OR-based method must consider the state changes
of containers. However, this is quite difficult in real world, because
these state changes are completely controlled by business operators
and yield quite different rules according to different customers and
regions. As a blackbox in the ECR problem, such business logic thus
cannot be exactly modeled by traditional OR-based methods. In
the real world of container transportation, there are more business
logics, e.g., regional policy regulation, which are in fact hard to be
modeled by OR-based methods.
To leverage OR-based methods, we have to relax corresponding
constraints and take an approximation approach in the baseline
algorithms, including:
• The transportation of empty and laden are decoupled. The
state changes of containers are pre-determined rather than
dynamically decided by business logic (nonlinear and even
black-boxed in real scene), leading to simplified SnD predic-
tion in OR model by decomposing future order information.
• The atomicity of one order is not preserved. In our running
example, the whole order will fail if the amount of remaining
empty containers is not sufficient, even if the gap is very
small. In OR models, this property cannot be guaranteed
since orders are decomposed into SnD.
vessels.
5.2 Experimental Setting
In the following experiments, we extract a main ocean transporta-
tion network among Asia, North America and Europe based on the
real world service loops of a commercial company. This network
consists of 4 route, 17 ports and 31 vessels. The routes are listed as
follows and illustrated in Figure 3:
• R1: Pacific Atlantic route, 94 days with 14 vessels.
• R2: Central Asia to Southeast Asia route, 60 days with 9
• R3: Japan to America route, 33 days with 5 vessels.
• R4: Japan-China-Singapore route, 19 days with 3 vessels.
The vessels are uniformly distributed with a interval around
one week in their routes. Initially, there are 3000 empty contain-
ers distributed in the 17 ports based on historical statistic from
a commercial ocean logistics company, and all vessels are empty
without any laden or empty containers. The distribution of SnD
of all 17 ports in the simulated environment is shown in Figure 4
based on information provided by the same company. Every vessel
has a capacity of 200 containers. i.e., the total amount of laden and
empty containers cannot exceed 200 for every vessel. To assist the
training of our cooperative MARL approach, we build a simulated
5.1.2 Difficulties of the ECR Problem with OR-based Methods.
The first difficulty of OR-based methods for ECR is brought by
the uncertainty of future SnD forecasts. As aforementioned, such
uncertainty is mainly caused by multiple external dynamic factors,
such as market changes, and will be even aggravated by the inherent
mutual dependency between the OR-based model and future SnD
forecast. Since typical OR-based methods generate action plans
based on future SnD forecasts for a long time span, the severe
uncertainty of long-term forecasts and ignorance of the inherent
mutual dependency between OR and SnD forecast will lead to poor
performance of OR-based methods.
The second major difficulty is caused by the certain business
logic in the container transportation chain. A typical and impor-
tant business logic that is hard to model by OR-based methods
3Without loss of generality, we only deal with non-transshipment order, that is we
suppose Pu and Pv are always within one route. A transshipment order can be viewed
as multiple separated non-transshipment orders.
Port BPort BPort APort ACargoes load into containersLaden containers load to vesselLaden containers discharge from vesselCargoes unload from containersEmpty containersEmpty containersConsumeTransportReturnEmpty Container RepositioningOrder placementCargoes receivedConsignerDepotDepotConsignee (cid:3) by loading or discharging containers. For-
i , F e
i
mally, suppose xt
is the number containers loading from Pi
i, j
(negative value means discharging to this port), it satisfies
i − F e
i − C t
i , Capt
i , C t
j − C t
V , j),
min(C t
− min(F s
0,
V , j, C t
i ), C t
i > F e
i ,
C t
i < F s
i ,
otherwise.
the range(cid:2)F s
x t
i, j =
Figure 3: The extracted ocean transportation network
among Asia, North America and Europe.
Figure 4: The distribution of demand and supply of all 17
ports in the environment.
ECR environment based on real historical data from the commercial
ocean logistics company.
To measure the performance of our approach, we use the metric
of fulfillment ratio, which is defined as the ratio of total successfully
fulfilled containers compared to all containers requested in one
episode (400 time steps, where one time step corresponds to one
day). In real-world, there are many other types of cost for container
repositioning, including loading/discharging cost, storage cost, etc.
Among all of them, however, the cost of shortage, measured by
the fulfillment ratio, is the dominant one since it will directly af-
fect the booking acceptance and consequently the transportation
company's reputation. Therefore, we focus on minimizing the cost
of shortage in this work. Indeed, other types of cost can also be
naturally captured by MARL through rewards shaping and specific
action space design, which will be one of our future targets.
5.3 Compared Methods
In the following experiments, we compare the following methods
on the ECR problem:
• No Reposition: Empty containers are never repositioned.
The flow of containers will only depends on the laden con-
tainer transportation.
• Rule-Based Inventory Control (IC). With the idea in in-
ventory management theory, this method sets two inven-
and excess threshold F e
tory thresholds, safety threshold F s
i
i
i ≤ F e
(F s
), for each port Pi based on the historical informa-
tion of SnD respectively. When a vessel Vj arrives at Pi at
time slot t, it will try to maintain the stock Ct
located in
i
i
• Online Linear Programming (LP). With some approxi-
mation approaches mentioned above, ECR problem can be
modeled in linear programming (LP) by adopting the mathe-
matical definitions in problem statement section. However,
it is hard to apply the solution directly due to the gap caused
by simplified model. Here, we apply rolling horizon policy
described in Long et al. [12] to solve the problem: empty repo-
sition plan are generated for a long period on the planning
horizon based on LP model with forecasting information for
this period, but only partial planning at the beginning are
executed. Repeat this procedure until termination. This is the
so called online LP method. Note that, our proposed end-to-
end MARL method directly interacts with the simulator with
no explicit forecasting stage, therefore, for the purpose of
appropriate comparison, we use exact future order informa-
tion to replace the forecasted future demand in the LP model
so as to eliminate the effects of external factors leading to
uncertain forecasts, which can be seen as a relatively ideal
condition. More details about the online LP can be found in
the appendix.
• Online LP with Inventory Control. In this baseline, we
adopt the idea from Epstein et al. [5] which combines LP
model with inventory control. This method sets a safety
for each port Pi based on the historical informa-
threshold F s
i
− F s
tion of SnD, and then constrains Lt
i
• Self Awareness MARL (SA-MARL). This is the MARL
model described in the previous section with self aware-
, ϕ(·) is an average
ness agents. For terminal (port) state s
function while ψ(·) is a sum function. For vehicle (vessel)
state s
, we add amount of laden containers onboard as ad-
ditional domain-specific information. As for reward, we set
f (x) = 1 − 0.5x and д(y) = 5y, where x and y are calculated
as in Equation (1).
• Territorial Awareness MARL (TA-MARL). This is the
MARL model with territorial awareness agents. For succes-
sive terminal information, both m and n are set to 1. Φ(·) and
Ψ(·) in s
• Diplomatic Awareness MARL (DA-MARL). This is the
MARL model described in previous session with diplomatic
awareness agents. Φr(·) and Φn(·) are set to be average func-
tions with α = 0.5. Both ξ1(·) and ξ2(·) are 2-layer average
are set to be average functions.
= max(cid:16)
i − (Ct−1
Dt
functions Avg{Avg{t′
• Offline Optimal LP (Upper Bound). In this case, the short-
age will be directly calculated as objective by LP model men-
tioned above, which has the knowledge of all orders in ad-
vance, without implementation in simulated environment.
This can be seen as an upper bound for the problem. i.e., it
i Pi ∈ Rp}Rp ∈ Crq}.
Lt
k
t =tk
tk
R,q
tk
V ,i
tk
P,i
i ), 0(cid:17).
i
ShanghaiNingboYantianShekouThailandSingaporeArabHong KongTaiwanTokyoKobeYokohamaOaklandLos AngelesSavannahNew YorkEUR1 (Pacific Atlantic route)R2 (Central Asia to Southeast Asia route)R3 (Japan to America route)R4 (Japan-China-Singapore route)Mainland ChinaMainland ChinaJapanJapanAmericaAmericaOthersOthers0200040006000800010000Total Containers in Order (Demand)Imported Laden Containers (Supply) Table 1: Performance comparison with different baselines.
Method
Fulfillment Ratio (%)
80% Container 100% Container 150% Container
29.87 ± 0.85
26.58 ± 0.90
38.25 ± 1.07
No Reposition
61.07 ± 0.98
68.63 ± 0.98
58.30 ± 0.93
IC
85.75 ± 1.34
94.48 ± 1.00
76.28 ± 1.54
Online LP
88.99 ± 0.89
Online LP with IC 81.09 ± 1.21
96.30 ± 0.80
72.04 ± 1.57
65.39 ± 1.20
84.21 ± 1.45
SA-MARL
93.75 ± 0.69
75.25 ± 1.38
83.48 ± 0.94
TA-MARL
97.70 ± 0.98
82.04 ± 1.69 95.97 ± 0.63
DA-MARL
Offline LP
98.32 ± 0.60
98.95 ± 0.31
99.42 ± 0.25
(Upper Bound)
Table 2: Performance comparison with different delay pa-
rameter k in DA-MARL
k
1
5
10
15
Fulfillment Ratio
95.87 ± 0.65
95.76 ± 0.67
95.49 ± 0.65
94.71 ± 0.93
k
20
30
40
50
Fulfillment Ratio
94.52 ± 0.89
93.23 ± 1.76
90.39 ± 2.50
85.87 ± 3.23
is not likely for any methods to achieve better performance
than this.
All MARL methods are trained 10000 episodes with ϵ-greedy
exploration. The ϵ is annealed linearly from 0.5 to 0.01 across the
first 8000 episodes, and fixed at 0.01 in the rest episodes. We use
Adam Optimizer with a learning rate of 10−4. Batch size is fixed
to 32. All agents in the same route share the same Q-network, and
each Q-network is parameterized by a 2-layer MLP with node size
of 16 and 16, activated by ReLU. Since DQN works on discrete
action space, we discretize the continuous action space Ai = [−1, 1]
uniformly by 21 actions, that is A′
i
= {−1,−0.9, · · · , 0.9, 1}.
5.4 Results Analysis
To compare all the methods aforementioned, we run our trained
models and baseline methods in 100 randomly initialized environ-
ments. For baseline methods, we run grid search to find suitable
parameters. To test the robustness of the learned policy in our
framework, we also evaluate the model trained under 100% (3000)
empty containers setting by changing the total amount of con-
tainers to 80% (2400 containers) and 150% (4500 containers). The
results are summarized in Table 1, in which we report the mean
and standard deviations of the fulfillment ratios. As we can see, DA-
MARL method achieves the best performance in all initial container
settings. Even TA-MARL method is comparable with traditional
online LP method. The SA-MARL achieves the poorest performance
among our MARL methods, while it is still better than rule-based
inventory control. The testing of robustness shows that agents
have learned efficient policies to deal with dramatic environment
changes. The trained DA-MARL model still performs better than
the online LP and its IC version, which in fact are built on changed
environments.
The convergence comparison of MARL methods are shown in
Figure 5a. Each MARL method is trained for 10 times, and we report
the mean and standard deviation of performance during training. As
we can see, all MARL methods converge very quickly at first 1000
episodes. After that, DA-MARL will get a much larger improvement
than the others.
In Diplomatic Awareness MARL, α is an important parameter to
control the proportion between territorial reward and diplomatic
reward. We train the model with different α and the results are
shown in Figure 5b. Every model is trained for 5 times due to time
limitation, and every trained model is tested for 100 times. The
result shows that neither rI alone (α = 1) nor rC alone (α = 0)
performs well alone, and a combination (α = 0.4 in our case) of
them is essential to achieve better performance.
Communication is a crucial part to build up cooperation in MAS,
and in our Diplomatic Awareness MARL design, shared information
Φr(·) and Φn(·) about neighboring routes and transshipment routes
is required to achieve high performance. However, it is possible
that these information cannot be transferred in real-time in realistic
scenario, i.e., agents can only have access to an outdated version
of these information. Table 2 shows the fulfillment ratio when
all agents can only access these information of k days ago. The
result shows that our proposed method performs robustly without
significant loss when the delay is in a reasonable range, i.e., k ≤ 20.
(a)
(b)
Figure 5: (a) Convergence comparison of MARL methods.
The X-axis is number of episodes. (b) Performance compar-
ison with different α in Diplomatic Awareness MARL. The
X-axis is α. The Y-axis is fulfillment ratio in both figures.
5.5 Cooperation Ability Analysis
The major objective of ECR is to balance the SnD so that the short-
age costs of deficit ports are minimized. Figure 6a shows the amount
of imported empty containers of Shekou and Thailand, two major
ports that are deficient of empty containers, by different methods.
From Figure 3, Thailand is the next ports of a surplus port Singapore
on route R2, which means it is not hard to obtain empty containers
without complicated cooperative mechanism. For Shekou, the sitia-
tion is much more severe as it need more containers than Thailand
(shown in Figure 4) while the only supply port, Singapore, in route
R4 doesn't have enough containers to supply Shekou. The only way
that demand of Shekou can be sufficiently fulfilled is to use Tokyo
and Kobe as transshipment ports to transport empty containers
%7,33 584/08
:12039#,94$
#%
#
#
0.20.30.40.50.60.70.80.91.0
0.50.60.70.80.91.0 :12039#,94 from America regions, which requires strong ability of cooperation
between regions. Figure 6a shows that all the three MARL methods
performs well on Thailand, while Diplomatic Awareness MARL out-
performs all other methods on Shekou, indicating that our design is
capable to fulfill the demand that requires inter-route cooperation.
For inter-route cooperation, the amount of exported empty con-
tainers at transshipment port is essential, since it is the source
from which deficient ports such as Shekou obtain empty containers.
Figure 6b shows the amount of exported empty containers of Singa-
pore, Tokyo and Kobe, which are three major transshipment ports
between different routes in our setting. It shows that the amount
of exported empty containers at transshipment ports significantly
increases with more cooperative awareness of MARL agent, which
indicates that our cooperative design is effective. Online LP method
with its IC version can also perform well on transshipment ports
since they are globally optimized. However, the gap between LP
models and environment confines their overall performance.
(a)
(b)
Figure 6: (a) Imported empty containers of Shekou and Thai-
land, two major ports that are deficient of empty containers,
by different methods. (b) Exported empty containers of Sin-
gapore, Tokyo and Kobe, three major transshipment ports
between different routes, by different methods. "No Reposi-
tion" method is omitted since it won't import or export any
empty containers.
6 CONCLUSION
In this paper, we first formulate the resource balancing problem
in logistics networks as a stochastic game. Given this setting, we
propose a cooperative multi-agent reinforcement learning frame-
work, in which three levels of cooperative metrics are identified
based on the scope of agents' awareness of cooperation, which
promote efficient and cost-effective transportation. Extensive ex-
periments on a simulated ocean transportation service demonstrate
that our new approach can stimulate the cooperation among agents
and give rise to a significant improvement in terms of both perfor-
mance and stability. In future, we will integrate more types of cost,
such as transport cost and inventory cost in real logistic scenarios,
into a unified objective to optimize. Moreover, we will investigate
more advanced RL techniques to achieve a more precise control of
actions.
ACKNOWLEDGEMENT
We sincerely appreciate Ryan Ho, Johnson Lui, Karab Sze, Jeffrey
Ko, Simon Choi, Tony Y Li, Apple Ng, Terry Tam and Wyatt Lei
from Orient Overseas Container Line for their great support on
this work.
A APPENDIX
A.1 Linear Programming Model for the ECR
Problem
The linear programming is given by:
min
Pi ∈P,t ∈Event(Pi)
Subject to
Ct
P,i
= C
i −
i + St
prev(P,i,t)
P,i
− Dt
prev(P,i,k)
P,i
,
i − C
i ≥ Dt
Lt
i ≥ 0,
Lt
for Pi ∈ P, t ∈ Event(Pi);
Ct
+ xt
j ,
V , j
0 ≤ Ct
for Vj ∈ V , t ∈ Event(Vj);
prev(V , j,t)
= C
V , j
V , j ≤ Capt
j ,
Lt
i
V
j=1
I(i, j, t)xt
j ,
(2)
(3)
(4)
(5)
(6)
(7)
and supply St
i
Event(·) denotes the set of time slot that an event is predicted
to be triggered for the argument, which can be inferred by V , R and
duration function dj(·, ·). Indicator function I(i, j, t) can be inferred
by similar manner. prev(P, i, t) (prev(V , j, t)) denotes the previous
time slot that an event is triggered on a port Pi (vessel Vj). External
for each port Pi ∈ P are provided by
demand Dt
i
external forecast model. In ECR problem, Capt
will dynamically
j
change according to the amount of laden containers in Vj at time
slot t. For order-based forecast model, i.e., the model forecasts future
order set O first and calculates predicted Dt
based on O,
i
can be also computed based on O with the assumption that
Capt
j
all the external demand Dt
can be fulfilled (so that the amount of
i
laden containers for each vessel at each time slot can be estimated).
The LR model is solved by GNU Linear Programming Kit (GLPK)
as integer programming.
and St
i
A.2 Details of Simulated ECR Environment
A.2.1 Route Schedule. The schedule of each route is shown in
Table 3, Table 4, Table 5 and Table 6 respectively based on infor-
mation provided by the same commercial company mentioned in
experiment section. All the routes are cycled. To achieve uniform
distribution of vessels in each route, vessels are not required to
berth in certain port when the environment is initialized.
A.2.2 Business Logic of Port-Vessel Interaction. When an event
(Pi , Vj) is triggered, i.e., a vessel Vj (on route Rk) arrives at a port
Pi, our simulated environment follows a 4-stage business logic to
execute action a ∈ [−1, 1]:
(1) Laden container discharge: all laden containers on Vj with
destination port Pi are discharged from the vessel. Notices
due to the decrease of laden
that Capt
j
containers in the vessel;
V , j] empty
(2) (if a < 0) Empty container discharge: [−a ∗ Ct
containers on Vj are discharged from the vessel;
will increase to Cap′t
j
02000400060008000ShekouThailandImported Empty ContainersICOnline LPOnline LP with ICSA-MARLTA-MARLDA-MARL0200040006000SingaporeTokyoKobeExported Empty ContainersICOnline LPOnline LP with ICSA-MARLTA-MARLDA-MARL Table 3: Route schedule of R1
Table 6: Route schedule of R4
Region/City
Port
STN Europe Union
NYC New York
SAV
Sawannah
LAS
Los Angeles
OAK Oakland
YOK
Yokohama
SHA Shanghai
KOY
TKY
OAK Oakland
Los Angeles
LAS
SAV
Sawannah
NYC New York
STN Europe Union
Kobe
Tokyo
Transit day
-
15
18
31
32
44
47
51
52
67
68
82
85
94
Table 4: Route schedule of R2
Region/City Transit day
Arab
Singapore
Thailand
Yantian
Los Angeles
Port
JEB
SIN
LCB
YAT
LAS
OAK Oakland
SHA Shanghai
NIN
YAT
SIN
JEB
Ningbo
Yantian
Singapore
Arab
-
3
6
9
26
28
43
44
46
51
60
-
3
17
18
31
33
Table 5: Route schedule of R3
Region/City Transit day
Kobe
Tokyo
Los Angeles
Port
KOY
TKY
LAS
OAK Oakland
TKY
KOY
Tokyo
Kobe
(3) Laden container loading: laden containers in Pi with des-
tination port in Rk are loaded into the vessel as much as
possible with the order of received date. Laden containers
in the same order can be separately transported. Similarly,
Cap′t
due to the increase of laden
j
containers in the vessel;
j −
(4) (if a > 0) Empty container loading: [a ∗ min(Cap′′t
P,i)] empty containers are loaded into the vessel.
will decrease to Cap′′t
j
Ct
V , j , Ct
Region/City Transit day
Tokyo
Kobe
Port
TKY
KOY
KHH Taiwan
HKG Hong Kong
SKZ
SIN
SKZ
HKG Hong Kong
KHH Taiwan
TKY
Tokyo
Shekou
Singapore
Shekou
-
2
5
6
7
11
14
15
16
19
Here [·] denotes the nearest integer function. In this business
logic, laden container transportation has priority over empty con-
tainer repositioning, which conforms real-world scenario in ocean
container transport logistics.
A.3 Regional Statistics
The regional statistics of seven methods in experiment part are
listed in Table 7, 8, 9, 10, 11, 12, 13 respectively. All methods are
tested 100 times and we report the average in the tables.
REFERENCES
[1] Regina Asariotis, Hassiba Benamara, Hannes Finkenbrink, Jan Hoffmann, Jennifer
Lavelle, Maria Misovicova, Vincent Valentine, and Frida Youssef. 2011. Review of
Maritime Transport, 2011. Technical Report.
[2] Teodor Gabriel Crainic and Gilbert Laporte. 1997. Planning models for freight
transportation. European journal of operational research 97, 3 (1997), 409 -- 438.
[3] Sam Devlin and Daniel Kudenko. 2011. Theoretical Considerations of Potential-
based Reward Shaping for Multi-agent Systems. In The 10th International Con-
ference on Autonomous Agents and Multiagent Systems - Volume 1 (AAMAS '11).
International Foundation for Autonomous Agents and Multiagent Systems, Rich-
land, SC, 225 -- 232. http://dl.acm.org/citation.cfm?id=2030470.2030503
[4] Sam Devlin, Logan Yliniemi, Daniel Kudenko, and Kagan Tumer. 2014. Potential-
based Difference Rewards for Multiagent Reinforcement Learning. In Proceed-
ings of the 2014 International Conference on Autonomous Agents and Multi-agent
Systems (AAMAS '14). International Foundation for Autonomous Agents and
Multiagent Systems, Richland, SC, 165 -- 172. http://dl.acm.org/citation.cfm?id=
2615731.2615761
[5] Rafael Epstein, Andres Neely, Andres Weintraub, Fernando Valenzuela, Sergio
Hurtado, Guillermo Gonzalez, Alex Beiza, Mauricio Naveas, Florencio Infante,
Fernando Alarcon, Gustavo Angulo, Cristian Berner, Jaime Catalan, Cristian
Gonzalez, and Daniel Yung. 2012. A Strategic Empty Container Logistics Op-
timization in a Major Shipping Company.
Interfaces 42, 1 (Feb. 2012), 5 -- 16.
https://doi.org/10.1287/inte.1110.0611
[6] Amy Greenwald and Keith Hall. 2003. Correlated-Q Learning. In Proceedings
of the Twentieth International Conference on Machine Learning (ICML'03). AAAI
Press, 242 -- 249. http://dl.acm.org/citation.cfm?id=3041838.3041869
[7] Junling Hu and Michael P. Wellman. 2003. Nash Q-learning for General-sum
Stochastic Games. J. Mach. Learn. Res. 4 (Dec. 2003), 1039 -- 1069. http://dl.acm.
org/citation.cfm?id=945365.964288
[8] Marc Lanctot, Vinicius Zambaldi, Audr¯unas Gruslys, Angeliki Lazaridou, Karl
Tuyls, Julien Pérolat, David Silver, and Thore Graepel. 2017. A Unified Game-
theoretic Approach to Multiagent Reinforcement Learning. In Proceedings of the
31st International Conference on Neural Information Processing Systems (NIPS'17).
Curran Associates Inc., USA, 4193 -- 4206. http://dl.acm.org/citation.cfm?id=
3294996.3295174
[9] Jing-An Li, Stephen CH Leung, Yue Wu, and Ke Liu. 2007. Allocation of empty
containers between multi-ports. European Journal of Operational Research 182, 1
(2007), 400 -- 412.
[10] Kaixiang Lin, Renyu Zhao, Zhe Xu, and Jiayu Zhou. 2018. Efficient Large-Scale
Fleet Management via Multi-Agent Deep Reinforcement Learning. ACM Press,
1774 -- 1783. https://doi.org/10.1145/3219819.3219993
[11] Michael L. Littman. 2001. Friend-or-Foe Q-learning in General-Sum Games.
In Proceedings of the Eighteenth International Conference on Machine Learning
Table 7: Regional Statistic of No Reposition Method
Total
Containers
400.26
0
0
601.08
8010.75
5008.44
797.06
3991.25
1403.97
2200.88
3010.1
0
199.28
1403.94
1002.21
200.86
1003.48
29 233.56
Failed
Containers
0
0
405.24
7124.58
4719.1
0
0
2501.22
944.97
1038.99
1881.84
0
6.12
630.41
534.1
0
505.85
20 292.42
Imported
Laden
Containers
343.61
136.56
134.25
785.03
109.05
860.58
0
1129.67
296.11
866.4
980.57
289.72
805.74
762.26
457
445.1
279.56
8681.21
Imported
Empty
Containers
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Exported
Laden
Containers
400.26
0
195.84
886.17
289.34
797.06
0
1490.03
459
1161.89
1128.26
0
193.16
773.53
468.11
200.86
497.63
8941.14
Exported
Empty
Containers
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Fulfillment
Ratio
1
/
/
/
0.325 814
0.110 623
0.057 77
1
0.373 324
0.326 93
0.527 921
0.374 825
0.969 289
0.550 971
0.467 078
1
0.495 904
0.299 626
Table 8: Regional Statistic of Inventory Control Method
Total
Containers
397.12
597.89
8007.37
4994.01
801.94
3998.32
1402.76
2177.95
2969.29
0
0
0
199.29
1397.14
998.67
200.87
994.69
29 137.31
Failed
Containers
6.81
0
9.01
6068.76
2303.97
1.28
0
1817.13
671.1
4.2
6.64
0
5.74
41.29
7.9
4.01
48.39
10 996.23
Imported
Laden
Containers
600.44
208.77
946.32
1321.94
108.97
1602.52
0
2051.84
650.05
1369.37
1526.76
463.89
2225.95
2071.77
883.59
862.88
459.47
17 354.53
Imported
Empty
Containers
52.13
0
109.82
536.18
2470.64
108.31
0
6.09
47.26
1006.98
1526.89
0
42.91
127.65
275.88
43.47
481.44
6835.65
Exported
Laden
Containers
390.31
588.88
1938.61
2690.04
800.66
2181.19
731.66
2173.75
2962.65
0
0
0
193.55
1355.85
990.77
196.86
946.3
18 141.08
Exported
Empty
Containers
589.51
352.35
507.6
0
28.3
1314.41
269
231.21
111.99
431.89
152.97
535.82
2067.45
804.18
131.33
745.36
177.7
8451.07
Fulfillment
Ratio
0.982 852
/
/
/
0.984 93
0.242 103
0.538 653
0.998 404
0.545 527
0.521 586
0.998 072
0.997 764
0.971 198
0.970 447
0.992 089
0.980 037
0.951 352
0.612 14
Region/City
Shanghai
Ningbo
Yantian
Shekou
Thailand
Singapore
Arab
Hong Kong
Taiwan
Tokyo
Kobe
Yokohama
Oakland
Los Angeles
Sawannah
New York
EU
Total
Region/City
Shanghai
Ningbo
Yantian
Shekou
Thailand
Singapore
Arab
Hong Kong
Taiwan
Tokyo
Kobe
Yokohama
Oakland
Los Angeles
Sawannah
New York
EU
Total
(ICML '01). Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 322 -- 328.
http://dl.acm.org/citation.cfm?id=645530.655661
[12] Yin Long, Loo Hay Lee, and Ek Peng Chew. 2012. The sample average approxi-
mation method for empty container repositioning with uncertainties. European
Journal of Operational Research 222, 1 (Oct. 2012), 65 -- 75. https://doi.org/10.1016/
j.ejor.2012.04.018
[13] Patrick Mannion, Jim Duggan, and Enda Howley. 2016. Generating multi-agent
potential functions using counterfactual estimates. Proceedings of Learning,
Inference and Control of Multi-Agent Systems (at NIPS 2016) (2016).
[14] Kunal Menda, Yi-Chun Chen, Justin Grana, James W. Bono, Brendan D. Tracey,
Mykel J. Kochenderfer, and David Wolpert. 2017. Deep Reinforcement Learning
for Event-Driven Multi-Agent Decision Processes. arXiv:1709.06656 [cs] (Sept.
2017). http://arxiv.org/abs/1709.06656
[15] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness,
Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg
Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen
King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. 2015.
Human-level control through deep reinforcement learning. Nature 518, 7540
(Feb. 2015), 529 -- 533. https://doi.org/10.1038/nature14236
[16] Andrew Y. Ng, Daishi Harada, and Stuart J. Russell. 1999. Policy Invariance
Under Reward Transformations: Theory and Application to Reward Shaping.
In Proceedings of the Sixteenth International Conference on Machine Learning
(ICML '99). Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 278 -- 287.
http://dl.acm.org/citation.cfm?id=645528.657613
[17] Ling Pan, Qingpeng Cai, Zhixuan Fang, Pingzhong Tang, and Longbo Huang.
2018. Rebalancing Dockless Bike Sharing Systems. arXiv:1802.04592 [cs] (Feb.
2018). http://arxiv.org/abs/1802.04592 arXiv: 1802.04592.
[18] Warren B Powell. 1996. Toward a Unified Modeling Framework for Real-Time
Logistics Control. Military Operations Research (1996), 69 -- 79.
Table 9: Regional Statistic of Online LP Method
Failed
Containers
14.88
0
158.21
1194.66
1760.6
82.85
0
156.27
101.41
112.21
158.49
0
5.09
49.25
11.4
1.94
143.08
3950.34
Imported
Laden
Containers
597.56
204.94
1110.7
1663.6
100.3
3172.18
0
3664.9
614.38
2892.28
3196.83
470.02
2231.34
2129.32
840.71
816.29
455.15
24 160.5
Imported
Empty
Containers
511.27
132.56
362.69
5213.06
3051.04
1992.17
60.54
1011.68
1108.52
1727.29
2769.94
122.06
979.83
1035.11
534.48
152.33
341.7
21 106.27
Exported
Laden
Containers
382.43
0
440.5
6768.18
3210.48
720.9
0
3826.72
1301.99
2062.85
2816.82
0
194.08
1342.87
989.3
198.9
854.59
25 110.61
Exported
Empty
Containers
1023.93
470.35
1071.43
142.81
66.81
4842.45
329.54
1164.39
536.47
2765.18
3194.36
624.55
2844.75
1760.18
319.59
759.67
146.09
22 062.55
Fulfillment
Ratio
0.962 548
/
/
/
0.735 749
0.849 971
0.645 831
0.896 921
0.960 766
0.927 74
0.948 411
0.946 732
0.974 444
0.964 622
0.988 608
0.990 341
0.856 586
0.859 485
Total
Containers
397.31
598.71
7962.84
4971.08
803.75
3982.99
1403.4
2175.06
2975.31
0
0
0
199.17
1392.12
1000.7
200.84
997.67
29 060.95
Table 10: Regional Statistic of Online LP with IC Method
Total
Containers
399.33
598.65
7972.27
4936.96
796.67
3981.14
1407.85
2194.04
2969.91
0
0
0
199.22
1396.98
997.87
200.77
989.84
29 041.5
Failed
Containers
21.13
0
87.58
677.09
1927.75
87.21
0
1.53
25
21.85
111.72
0
5.01
42.27
8.76
5.5
64.39
3086.79
Imported
Laden
Containers
606.51
207.5
1029.14
1732.56
99.03
3331.11
0
3896.77
628.13
3093.01
3325.12
465.3
2221.68
2138.31
871.68
852.01
455.31
24 953.17
Imported
Empty
Containers
569.89
134.47
326.27
5657.99
2832.29
1853.54
26.51
1140.93
1284.43
1740.66
3043.36
109.34
936.29
1026.7
598.89
163.15
408.18
21 852.89
Exported
Laden
Containers
378.2
0
511.07
7295.18
3009.21
709.46
3979.61
1382.85
2172.19
2858.19
0
0
194.21
1354.71
989.11
195.27
925.45
25 954.71
Exported
Empty
Containers
1105.43
481.19
877.82
74.89
17.63
4859.85
295.51
1248.01
647.28
2786.07
3536.31
638.74
2814.07
1702.91
421.21
834.53
128.4
22 469.85
Fulfillment
Ratio
0.947 086
/
/
/
0.853 704
0.915 069
0.609 527
0.890 532
0.999 616
0.982 242
0.990 041
0.962 383
0.974 852
0.969 742
0.991 221
0.972 605
0.934 949
0.889 923
Region/City
Shanghai
Ningbo
Yantian
Shekou
Thailand
Singapore
Arab
Hong Kong
Taiwan
Tokyo
Kobe
Yokohama
Oakland
Los Angeles
Sawannah
New York
EU
Total
Region/City
Shanghai
Ningbo
Yantian
Shekou
Thailand
Singapore
Arab
Hongkong
Taiwan
Tokyo
Kobe
Yokohama
Oakland
Los Angeles
Sawannah
New York
EU
Total
[19] David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George
van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershel-
vam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalch-
brenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu,
Thore Graepel, and Demis Hassabis. 2016. Mastering the game of Go with
deep neural networks and tree search. Nature 529, 7587 (Jan. 2016), 484 -- 489.
https://doi.org/10.1038/nature16961
[20] Dong-Ping Song and Jing-Xin Dong. 2015. Empty Container Repositioning.
In Handbook of Ocean Container Transport Logistics. Springer, Cham, 163 -- 208.
https://doi.org/10.1007/978-3-319-11891-8_6
[21] UNCTAD. 2017. Review of maritime transport 2017. OCLC: 1022725798.
[22] Xiaofeng Wang and Tuomas Sandholm. 2002. Reinforcement Learning to Play
an Optimal Nash Equilibrium in Team Markov Games. In Proceedings of the
15th International Conference on Neural Information Processing Systems (NIPS'02).
MIT Press, Cambridge, MA, USA, 1603 -- 1610. http://dl.acm.org/citation.cfm?id=
2968618.2968817
[23] Zhe Xu, Zhixin Li, Qingwen Guan, Dingshui Zhang, Qiang Li, Junxiao Nan,
Chunyang Liu, Wei Bian, and Jieping Ye. 2018. Large-Scale Order Dispatch in
On-Demand Ride-Hailing Platforms: A Learning and Planning Approach. In
Proceedings of the 24th ACM SIGKDD International Conference on Knowledge
Discovery & Data Mining (KDD '18). ACM, New York, NY, USA, 905 -- 913. https:
//doi.org/10.1145/3219819.3219824
Table 11: Regional Statistic of Self Awareness MARL Method
Total
Containers
396.75
595.06
8010.53
4940.34
799.44
3992.99
1405.77
2192.7
2970.62
0
0
0
199.28
1396.48
1006.08
200.75
994.14
29 100.93
Failed
Containers
15.46
0
35.19
5979.48
254.41
89.14
0
1133.04
430.5
100.74
180.91
0
7.37
67.75
36.43
1.81
61.25
8393.48
Imported
Laden
Containers
588.56
204.44
1588.92
1467.91
96.13
1743.24
0
2252.99
604.48
1515.5
1710.44
456.66
2718.49
2731.53
839.19
831.19
449.12
19 798.79
Imported
Empty
Containers
327.65
25.86
145.66
1101.63
4639.55
1250.79
224.56
710.83
453.21
1663.18
1887.26
10.66
188.76
906.05
450.62
69.71
573.6
14 629.58
Exported
Laden
Containers
381.29
0
559.87
2031.05
4685.93
710.3
0
2859.95
975.27
2091.96
2789.71
0
191.91
1328.73
969.65
198.94
932.89
20 707.45
Exported
Empty
Containers
871.11
366.75
1201.87
646.8
87.87
2679.4
491.8
412.74
200.81
1317.72
883.26
531.22
2693.95
2263.86
289.04
731.6
272.27
15 942.07
Fulfillment
Ratio
0.961 033
/
/
/
0.940 863
0.253 548
0.948 504
0.888 497
0.716 243
0.693 762
0.954 057
0.9391
0.963 017
0.951 485
0.963 79
0.990 984
0.938 389
0.702 295
Table 12: Regional Statistic of Territorial Awareness MARL Method
Total
Containers
400.81
0
597.05
7967.28
4939.66
805.05
0
3997.4
1402.58
2181.59
2969.32
0
199.3
1393.16
998.18
200.76
993.28
29 045.42
Failed
Containers
20.7
0
47.7
2248.07
135.07
178.25
0
1045.12
425.2
169.94
234.94
0
5.12
54.74
32.36
0.04
51.05
4648.3
Imported
Laden
Containers
587.02
206.01
1633.54
1447.38
85.8
2630.17
0
3129.24
605.39
2362.92
2671.21
465.79
2729.45
2726.46
827.99
813.79
452.57
23 374.73
Imported
Empty
Containers
337.98
30.21
195.44
4352.67
4730.73
659.64
152.09
350.29
506.39
1617.55
1985.45
16.68
226.27
870.76
558.79
124.39
505.77
17 221.1
Exported
Laden
Containers
380.11
0
549.35
5719.21
4804.59
626.8
0
2952.28
977.38
2011.65
2734.38
0
194.18
1338.42
965.82
200.72
942.23
24 397.12
Exported
Empty
Containers
885.46
377.1
1306.06
98.23
25.76
3079.51
420.43
868.87
270.44
2208.95
1986.71
542.53
2744.53
2228.29
404.74
772.29
202.39
18 422.29
Fulfillment
Ratio
0.948 355
/
/
/
0.920 107
0.717 837
0.972 656
0.778 585
0.738 55
0.696 844
0.922 103
0.920 878
0.974 31
0.960 708
0.967 581
0.999 801
0.948 605
0.834 133
Region/City
Shanghai
Ningbo
Yantian
Shekou
Thailand
Singapore
Arab
Hong Kong
Taiwan
Tokyo
Kobe
Yokohama
Oakland
Los Angeles
Sawannah
New York
EU
Total
Region/City
Shanghai
Ningbo
Yantian
Shekou
Thailand
Singapore
Arab
Hong Kong
Taiwan
Tokyo
Kobe
Yokohama
Oakland
Los Angeles
Sawannah
New York
EU
Total
Table 13: Regional Statistic of Diplomatic Awareness MARL Method
Region/City
Shanghai
Ningbo
Yantian
Shekou
Thailand
Singapore
Arab
Hong Kong
Taiwan
Tokyo
Kobe
Yokohama
Oakland
Los Angeles
Sawannah
New York
EU
Total
Total
Containers
398.2
0
596.74
7963.94
4937.34
800.5
0
3994.47
1401.06
2184.47
2975.98
0
199.32
1399.43
1003.12
200.8
995.32
29 050.69
Failed
Containers
2.46
0
20.2
383.02
205.8
68.58
0
117.05
47.4
64.91
61.25
0
7.29
44.17
35.03
9.73
67.04
1133.93
Imported
Laden
Containers
591.15
205.53
1614.78
1704.92
100.41
3395.41
0
3919.51
619.78
3109.54
3380.17
462.15
2766.72
2763.45
887.31
853.98
447.44
26 822.25
Imported
Empty
Containers
487.31
122.52
240.68
5951.09
4670.81
601.86
216.75
543.61
1025.34
2088.9
2273.1
21.55
449.23
661.21
419.99
67.35
561.59
20 402.89
Exported
Laden
Containers
395.74
576.54
7580.92
4731.54
731.92
3877.42
1353.66
2119.56
2914.73
0
0
0
192.03
1355.26
968.09
191.07
928.28
27 916.76
Exported
Empty
Containers
1001.32
469.3
1298.11
11.96
57.05
3674.31
484.96
881.82
405.71
3245.84
2754.09
541.47
2955.57
2028.79
196.33
711.75
202.88
20 921.26
Fulfillment
Ratio
0.993 822
/
/
/
0.966 149
0.951 906
0.958 318
0.914 329
0.970 697
0.966 168
0.970 286
0.979 419
0.963 426
0.968 437
0.965 079
0.951 544
0.932 645
0.959 447
|
1909.00991 | 1 | 1909 | 2019-09-03T08:07:27 | Modelling Bushfire Evacuation Behaviours | [
"cs.MA",
"cs.AI"
] | Bushfires pose a significant threat to Australia's regional areas. To minimise risk and increase resilience, communities need robust evacuation strategies that account for people's likely behaviour both before and during a bushfire. Agent-based modelling (ABM) offers a practical way to simulate a range of bushfire evacuation scenarios. However, the ABM should reflect the diversity of possible human responses in a given community. The Belief-Desire-Intention (BDI) cognitive model captures behaviour in a compact representation that is understandable by domain experts. Within a BDI-ABM simulation, individual BDI agents can be assigned profiles that determine their likely behaviour. Over a population of agents their collective behaviour will characterise the community response. These profiles are drawn from existing human behaviour research and consultation with emergency services personnel and capture the expected behaviours of identified groups in the population, both prior to and during an evacuation. A realistic representation of each community can then be formed, and evacuation scenarios within the simulation can be used to explore the possible impact of population structure on outcomes. It is hoped that this will give an improved understanding of the risks associated with evacuation, and lead to tailored evacuation plans for each community to help them prepare for and respond to bushfire. | cs.MA | cs |
Modelling bushfire evacuation
behaviours
J. Robertson
Supervisor: D.Singh
November 2018
Abstract
Bushfires pose a significant threat to Australia's regional areas. To
minimise risk and increase resilience, communities need robust evacu-
ation strategies that account for people's likely behaviour both before
and during a bushfire. Agent-based modelling (ABM) offers a practi-
cal way to simulate a range of bushfire evacuation scenarios. However,
the ABM should reflect the diversity of possible human responses in a
given community. The Belief-Desire-Intention (BDI) cognitive model
captures behaviour in a compact representation that is understand-
able by domain experts. Within a BDI-ABM simulation, individual
BDI agents can be assigned profiles that determine their likely be-
haviour. Over a population of agents their collective behaviour will
characterise the community response. These profiles are drawn from
existing human behaviour research and consultation with emergency
services personnel and capture the expected behaviours of identified
groups in the population, both prior to and during an evacuation. A
realistic representation of each community can then be formed, and
evacuation scenarios within the simulation can be used to explore the
possible impact of population structure on outcomes. It is hoped that
this will give an improved understanding of the risks associated with
evacuation, and lead to tailored evacuation plans for each community
to help them prepare for and respond to bushfire.
1
Acknowledgements
I would like to thank my supervisor, Dr. Dhirendra Singh, for his help and
advice. His guidance has been invaluable and I have enjoyed our collabora-
tion immensely. Thanks also to Dr. Vincent Lemiale, Dr. Leorey Marquez
and Rajesh Subramanian from Data61, who I have worked closely with at
times during this year. Without Data61 I would not have been introduced to
bushfire evacuation modelling, and without their support this thesis would
not have been possible.
2
Contents
1 Introduction
2 Background
2.1.1 Bushfire Behaviours
2.2.1 Belief-Desire-Intention Model
2.2 Representing Behaviour in Computational Models
2.1 Behavioural Research . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . .
. . . . . . . . . . . . . .
2.3 Social Simulation . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 Agent-based Modelling . . . . . . . . . . . . . . . . . .
2.3.2
Synthetic Populations . . . . . . . . . . . . . . . . . .
2.4 Existing Evacuation Models . . . . . . . . . . . . . . . . . . .
2.5 The Emergency Evacuation Simulator . . . . . . . . . . . . .
3 Population Generation Algorithm
3.1
3.2.1
3.2.2
Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1
Subgroups . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.2 Activity Distributions . . . . . . . . . . . . . . . . . .
3.1.3 Location Mapping . . . . . . . . . . . . . . . . . . . .
3.2 Activity Assignment . . . . . . . . . . . . . . . . . . . . . . .
Selecting a Feasible Activity Set
. . . . . . . . . . . .
Sequencing the Daily Plan . . . . . . . . . . . . . . . .
3.3 Assigning Locations to Activities . . . . . . . . . . . . . . . .
3.3.1 Localities . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.2 Allocation Considerations . . . . . . . . . . . . . . . .
3.3.3 Home Selection . . . . . . . . . . . . . . . . . . . . . .
3.4 Adding Bushfire Behaviours . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
3.4.1 Threshold Model
3.4.2 Possible Bushfire Behaviours
. . . . . . . . . . . . . .
3.4.3 Assigning Attributes . . . . . . . . . . . . . . . . . . .
4 Applications
Subgroups in Surf Coast Shire
4.1 Algorithm Inputs in Surf Coast Shire . . . . . . . . . . . . . .
4.1.1
. . . . . . . . . . . . .
4.1.2 Activities in Surf Coast Shire . . . . . . . . . . . . . .
4.1.3 Distributions in Surf Coast Shire . . . . . . . . . . . .
4.1.4 Locations in Surf Coast Shire . . . . . . . . . . . . . .
4.2 Results of the Algorithm . . . . . . . . . . . . . . . . . . . . .
4.2.1 Plan Times . . . . . . . . . . . . . . . . . . . . . . . .
3
4
6
6
7
10
11
13
13
15
16
17
21
22
23
23
24
25
26
28
29
30
34
35
36
37
39
41
44
44
44
45
47
48
49
50
4.3 Adding BDI Attributes
4.2.2 Plan Locations . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
4.3.1 Alerts in the Surf Coast Shire . . . . . . . . . . . . . .
4.3.2 Attribute Parameters
. . . . . . . . . . . . . . . . . .
4.4 SCS Application . . . . . . . . . . . . . . . . . . . . . . . . .
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5 Conclusion
5.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
63
63
64
66
70
71
72
1
Introduction
A key question posed in the aftermath of the 2009 Black Saturday bushfires
was whether the official Prepare to Stay and Defend or Leave Early (PSDLE)
policy was an effective way to protect under-threat communities (CFA &
Sweeney, 2009). Whilst it was intended to minimise the number of residents
attempting a dangerous late evacuation, there are concerns that the 'Stay
or Go' approach only leads to confusion over the level of preparation and
planning required to stay and defend, and does not properly define how late
is 'too late' to leave. Subsequent policies emphasising early evacuation as the
only safe option have shown that blanket advisory policy has little effect on
people's behaviour (Reid & Beilin, 2014). This highlights a need for localised
bushfire evacuation responses that are co-developed with communities and
minimise the risks specific to their townships and localities.
Realising this goal requires an understanding of the dynamics of an evac-
uation as well as an insight into how the actions of evacuees and responders
influence the process. One way to achieve this is to simulate the traffic condi-
tions of a community evacuation using an agent-based model (ABM). ABMs
are particularly suited to modelling discrete real-world phenomena, where
agents represent individuals in the population that can interact with each
other and the environment (Bonabeau, 2002). This method allows emer-
gent or collective behaviours of the community to be captured whilst still
preserving the details of the underlying system (Chen & Zhan, 2008). The
Belief-Desire-Intention (BDI) architecture (Rao & Georgeff, 1991) provides
a cognitive framework that captures human behaviour within an intuitive
structure that is both programmatically efficient and understandable by do-
main experts (Norling et al., 2000). BDI can be used alongside an ABM
to add a layer of human reasoning to the actions of agents, allowing them
to pro-actively reason about and respond to changing circumstances within
4
the simulation. In this work we will use the BDI-ABM integration outlined
by Singh et al. (2016).
Any evacuation model should take into consideration the expected re-
sponse of people in the affected community, both to the fire itself and to any
evacuation instructions (Pel et al., 2011). One issue is that this behaviour
data may be scarce. Behavioural science research post Black Saturday has
focussed on PSDLE decisions (Whittaker et al., 2013), but less is known
about evacuation travel decisions which also heavily influence an individ-
ual's safety. To extract these behaviours, we will use existing data from
interviews, surveys and consultation with domain experts to form repre-
sentative behaviour profiles. These profiles will be motivated by bushfire
response research that focuses on archetypes in the community (Strahan
et al., 2018), as well as parameter based non-bushfire evacuation response
models (Lee et al., 2010).
In addition to this, our work will focus on what these behaviour pro-
files mean for the population prior to the evacuation scenario. The types
of activities that people are likely to be engaged in and the dynamics of
the background traffic in the region will have an influence on an evacuation
should a bushfire threat arise. We account for these by linking sets of activi-
ties to each behaviour profile. The profile of an agent then determines when
it is likely to begin a given activity, and where in the region that activity
will occur. A plan for each agent is generated with details of its activities
during the day. In a given scenario, these plans are interrupted by a specific
bushfire threat. An agent's response depends on its current location and
activity as well as its parameterised behaviour profile.
Evaluation focuses on communities in regional Victoria. Behavioural
profiles are informally validated by local emergency service personnel and
bushfire behaviour experts before being applied to specific communities. The
simulation is built using the BDI-ABM framework1, combining the JILL
BDI engine2 with the MATSim traffic ABM3. A range of evacuation scenar-
ios are constructed to experimentally evaluate how sensitive the evacuation
outcomes are to the demographic make-up of different communities.
In this thesis we:
• Present a review of the literature related to bushfire behaviour, the
representation of behaviour in computational models, and the design
of social simulations that consider these behaviours.
1https://github.com/agentsoz/bdi-abm-integration.
2https://github.com/agentsoz/jill.
3https://github.com/matsim-org/matsim.
5
• Outline the design of an algorithm that takes a set of generalised be-
haviour profiles with associated parameters, and generates a list of
agent plans suitable to be run in MATSim.
• Describe how these behaviour profiles can be enriched by adding at-
tributes to the plans that dictate how they will respond to a bushfire
threat in the BDI-ABM simulation.
• Apply the algorithm by generating plans for regions in the Surf Coast
Shire of Victoria, describing tangible behaviour profiles drawn from
interaction with the local community and its emergency services.
• Quantify the bushfire behavioural model through consultation with
domain experts and append the BDI attributes to the generated plans.
• Run instances of these plans in the BDI-ABM model to illustrate the
benefit of conducting bushfire evacuation simulation with the dynam-
ics of the community in mind.
The remainder of this document is structured as follows: we present a
review of the related literature in Section 2; the design of the algorithm and
the BDI behavioural model is given in Section 3; in Section 4 the evacuation
simulation application for the Surf Coast Shire is described, and we analyse
the algorithm output and the resulting simulation outcomes; lastly, in Sec-
tion 5 we conclude with a discussion on the current model and directions
for future work.
2 Background
In this section we review related research into evacuation behaviour, with
a particular focus on bushfire and community specific responses. Following
that, we examine social simulation, and specifically, agent-based modelling
applied in that context. Finally, existing models that implement some as-
pects of our research scope are described.
2.1 Behavioural Research
We proceed by first exploring how behavioural science has understood evac-
uations in general, and more specifically bushfire behaviours. Any evacua-
tion simulation needs to take into account the expected human response to
environmental cues and received warnings (Pel et al., 2011). Predicting be-
haviour in an emergency is difficult, and one of the main motivating factors
6
for modelling bushfire evacuations is to gain better insight into expected
response to direction and coordination (Dash & Gladwin, 2007). The model
can only provide this if the underlying human behaviours that the agents
are given are realistic.
The first aspect to consider is that one should expect systematic de-
viations from rational/optimal behaviour in emergency situations (Dash &
Gladwin, 2007). Because the situation is likely to be unfamiliar and unusual,
people generally do not rely on prior experience and knowledge and instead
focus on the information directly available to them, resulting in 'myopic'
behaviour (Pel et al., 2011). This is exacerbated by the potential for unex-
pected behaviour in stressful or mentally challenging emergency conditions
(Knoop, 2009). People may also fail to comply with directions that they
are given, especially if their prepared plan contradicts the instructed action
(Pel et al., 2011). Conversely, spontaneous evacuation is known to happen
even in places that have not been told to evacuate, so instructed behaviour
should not be taken as actual behaviour (Lindell & Prater, 2007).
2.1.1 Bushfire Behaviours
Whilst there is a body of research into traffic behaviours in evacuations
and accidents (Lindell & Prater, 2007; Tu et al., 2010; Pel et al., 2011), it
is less available in the context of bushfire disasters. Behavioural research
into bushfires, particularly in an Australian context, has instead tended to
focus on the decisions people make between either evacuating or staying
and defending property (Johnson et al., 2011; Haynes et al., 2010). This is
largely a result of the Prepare to Stay and Defend or Leave Early (PSDLE)
policy -- colloquially known as 'Stay or Go' -- which has been official policy in
Australia since the 1990s and explicitly puts the evacuation decision in the
hands of individuals rather than the state (Reynolds, 2017). In other places
where bushfires (wildfires) are common, such as Canada and parts of USA,
the norm is that mandatory mass evacuation can and will be enforced by
authorities (Moritz et al., 2014). However, related research in North America
still tends to focus on the issue of evacuation timing and compliance, and
often questions local evacuation policies (Paveglio et al., 2008; McCaffrey
et al., 2015). This predominant focus on behaviour prior to an evacuation
reflects the often rapid and unpredictable nature of a bushfire as compared
to other natural disasters, which may allow more calculated and decisive
evacuation plans (Adam et al., 2016). In a bushfire, the decision between
staying or going is most likely to be significant in terms of survival, and so
it has become the main focus of behavioural research.
7
Accordingly, we should expect that 'Stay or Go' decisions play an im-
portant part in an evacuation on a macroscopic level, and that a model of
evacuation traffic behaviour will need to account for the circumstances of
a person's departure. In their analysis of community preparedness for the
2009 Victorian 'Black Saturday' bushfires, Whittaker et al. (2013) highlight
the dangers of late evacuation. People who leave late are usually triggered
by the sight of heavy smoke or flames, which means that
'By this time it is likely that driving a vehicle will have become
very difficult, with flames, smoke, strong winds, fallen trees, traf-
fic and the urgency of the situation increasing the likelihood of
accidents (Tibbits & Whittaker, 2007).'
Late evacuations are associated with a lack of planning and preparedness
among those who initially choose to stay and defend, but are not fully com-
mitted to that decision.
Interviews with Black Saturday survivors estab-
lish that there can be a disconnect between intended action and actual be-
haviour; for instance, 63% of people who adopted a 'wait and see' approach
ended up leaving once the fire was visible.
A difficulty here is whether to define waiting and seeing as an intended
action; as McLennan et al. (2012) note, to 'wait and see' is more of an
observation that describes a person's mindset rather than a fully formed plan
that represents their intentions. One new approach is to move away from
viewing the decision as a binary question of leaving or staying. Strahan et al.
(2018) form seven self-evacuation archetypes which accommodate a diverse
range of responses to a bushfire threat. With foundations in Jung's work on
the collective unconscious (Jung, 1969), archetypes represent 'fundamental
characteristics of humanity'. There is also a basis for their use in Australian
public policy, where they are generated in a similar way using cluster and
discriminant function analysis.
The results in Strahan et al. (2018) were derived from interviews with 452
participants who had recently experienced bushfire. The questions focussed
on factors like experience, intended and actual responses, self-responsibility,
access to information and demographics. The seven archetypes they found
are shown in Table 1.
Reid & Beilin (2014) conducted similar interviews following 11 January
2010, a day where a 'Catastrophic' fire warning was issued to residents of
Halls Gap. This warning category was introduced in the aftermath of Black
Saturday to indicate that evacuation is highly recommended; despite this,
the majority of residents did not leave. One factor attributed to this re-
sponse was the nature of risk perception in the community and that 'aware-
8
Table 1: Self-evacuation archetypes according to Strahan et al. (2018).
Archetype
Responsibility
Denier
Dependent
Evacuator
Considered
Evacuator
Community
Guided
Worried
Waverer
Threat
Denier
Experienced
Independent
Key characteristics
Believe they are not responsi-
ble for their personal safety or
for their property.
Expect
the emergency ser-
vices to protect them and
their property because they
are incapable of taking re-
sponsibility for themselves.
Having carefully considered
evacuation, are committed to
it as soon as they are aware of
a bushfire threat.
Seek guidance from neigh-
bours, media and members
of the community who they
see as knowledgeable, well in-
formed and providing reliable
advice.
Prepare and equip their prop-
erty and train to defend it
but worry they lack practi-
cal experience to fight bushfire
putting their personal safety
at risk.
Do not believe that their per-
sonal safety or property is
threatened by bushfire.
Are highly knowledge, compe-
tent and experienced and are
responsible and self-reliant
fighting bushfire.
Evacuate or Remain
Highly committed evac-
uators but expect oth-
ers to direct and assist.
Highly committed evac-
uators but expect oth-
ers to direct and assist.
Committed
directed evacuation.
to
self-
Committed to evacua-
tion on community ad-
vice.
Wavering between evac-
uating and remaining.
Committed to remain
as perceived lack of
threat makes
evacua-
tion unnecessary.
Highly committed to re-
maining because they
are highly experienced
and well prepared.
ness does not always result in a realistic understanding of how to respond
to risk'. Residents may believe that they have a better understanding of
local weather and geographic conditions than a state-wide warning system,
9
and prior experience of similar bushfire events will also strongly influence
people's decisions (McCaffrey et al., 2011). Further to this, trust in warning
systems can be eroded if they are perceived as a political initiative, or even
as an overly 'scientific assessment' of risk. Community response resources
that incorporate local knowledge are likely to have more impact on people's
behaviour.
Horse ownership is a good example of a community specific behaviour
which might be better understood from a local perspective. Thompson
et al. (2018) highlights that although pet ownership is generally a factor
in people's evacuation plans, horse owners (who are relatively common in
rural areas) are presented with unique challenges during a bushfire. Due
to the difficulty of transporting horses pre-emptive or early evacuation is
advised by the CFS as a priority action. However, Thompson et al. find
that this blanket strategy may be detrimental if it comes at the expense of
other preparatory planning, and could prevent horse owners from forming
contingency plans that improve the likelihood of horse survival in the event
they are left in place.
2.2 Representing Behaviour in Computational Models
The complex nature of human behaviour presents challenges for modellers in
many different domains. Kulash (1990) highlights a particular issue in urban
planning, where traffic engineering that fails to take human behaviour into
account can result in contradictory or unsafe policies. Designing hierarchi-
cal traffic systems that rigidly place different roads in different classes (often
inspired by natural phenomena like river systems) is not effective because
it considers humans merely as particles in a flow, rather than autonomous
agents who can decide to use a road for reasons separate from its assigned
purpose. Moreover, urban planners cannot assume random distribution or
optimal dispersal across a network. Song et al. (2010) explore patterns
of human mobility, and note that traditional continuous-time random-walk
models of movement do not reflect the tendency of humans to restrict their
exploration to locations near them, nor their propensity to return to fre-
quently visited locations.
This also has relevance in epidemiological modelling, where the dis-
tinction between density and frequency-based transmission is important
(Antonovics et al., 1995; Lloyd-Smith et al., 2004). The ubiquitous model in
epidemiology is the Susceptible-Infected-Recovered (SIR) set of differential
equations, and indeed most macroscopic representations of disease are exten-
sions of Kermack & McKendrick (1927)'s model. Funk et al. (2010) review
10
the presence of behavioural considerations in disease models, and note that
whilst human behaviour is intrinsically linked to disease spread, it is diffi-
cult to completely synthesise the pathogen dynamics that the SIR model
captures and the complex interplay of 'attitudes, belief systems, opinions
and awareness of a disease... both in an individual and in the population on
the whole'. One alternative is to approach disease modelling from a contact
network perspective, viewing the population as a complex system (Galea et
al., 2010; Christakis, 2007). Considering individuals within a larger social
network and using a bottom-up method allows various different behaviours
to be examined; for instance, clustered interaction networks permit the use
of percolation thresholds to characterise epidemics (Moore & Newman, 2000;
Davis et al., 2008). This is also of interest in economics, where the classi-
cal notion of the rational agent is often poorly represented by cumbersome
mathematical machinery (Farmer & Foley, 2009). It may not be appropriate
to assume that a market will reach a long term equilibrium when complex in-
dividual behaviours within the population give rise to sub-optimal emergent
behaviours in the collective.
Whilst most behavioural research uses surveys and interviews to analyse
bushfire responses, there is usually the caveat that both the sample size and
the qualitative nature of the data limit the predictive power of results. The
acknowledgement that intended and actual actions may differ (Whittaker
& Handmer, 2010) means that the interview approach has limited appli-
cation in capturing behaviour, and actual modelling of bushfire evacuation
behaviour has tended to use 'the judgement of evacuation planners or ex-
perts familiar with the area under study' (Beloglazov et al., 2016). This is
further enforced by the lack of behavioural research into on-road evacua-
tion behaviours in bushfires. Kennedy (2012) is critical of the tendency to
model behaviour using uniform random variables as it does not leave room
for biases and illogical decisions, ignores any notion of memory/learning,
and makes it difficult to identify behavioural preferences between different
groups of people. Instead, threshold-based rules are recommended as a sim-
ple way of approximating consistent behaviours within an individual agent.
2.2.1 Belief-Desire-Intention Model
Any computational model that seeks to use human behaviour to observe
emergent outcomes will require a cognitive architecture to represent the hu-
man decision making process (Gilbert, 2006). There are several approaches
to forming a cognitive model. A connectionist approach seeks to mirror the
design (but not necessarily the complexity) of the brain's neural networks,
11
giving a natural representation of the biological process (McClelland, 1988).
One shortcoming of a neural network model is that the reasoning process
that leads to a particular behaviour is not readily explainable (Medsker,
1994). Other models form a more abstract picture of brain function, with
a focus on the storage and deployment of knowledge -- ACT-R and Soar are
two well-known examples (J. R. Anderson, 1996; Laird, 2007). Whilst these
models are firmly based on experimental evidence and psychological theory
(Gilbert, 2006), they focus heavily on the individual level of cognition. The
Belief-Desire-Intention (BDI) framework provides a way to represent ratio-
nal decision making that lies more between the individual and social context
(Bratman et al., 1988; Jarvis et al., 2008). Less tied to cognitive science,
BDI relies on a 'folk psychology' approach that captures how people think
about how people think (Norling et al., 2000).
The BDI framework is formally described by Rao & Georgeff (1991).
Based on a branching-time possible-worlds model, BDI seeks to model ratio-
nal agents that can vary their actions based on changing beliefs, desires/goals
and intentions. Beliefs represent an agent's knowledge of the environment,
and are not necessarily perfect pictures of the current state of the system
(Norling et al., 2000). Combinations of beliefs will leave an agent with cer-
tain desires that they wish to fulfil. Desires may be inconsistent with one
another, but goals are a subset of desires that are consistent and achievable
according to the agent's beliefs. To achieve these goals, an agent will form
intentions, or plans of action to execute.
Rao & Georgeff (1995) acknowledge that the abstract formalism of their
BDI theory does not lend itself to practical computing, and instead pro-
pose a simplified architecture which is more efficient (Ingrand et al., 1992).
Many modern BDI programming languages are implementations of Rao and
Georgeff's abstract interpreter, and some have maintained a link to the for-
mal underpinnings of BDI; the AgentSpeak(L) language (Rao, 1996) and its
descendent interpreter Jason (Bordini et al., 2005) are useful in this sense
because they allow properties of an empirical BDI implementation to be
proven methodically. GOAL (de Boer et al., 2002) extends this capacity
to allow programmable declarative goals as part of the propositional logic.
This means that goals, like beliefs, are separate from any required action
and can form part of an agent's abstract reasoning. Other languages in-
clude the C++ based dMARS (D'Inverno et al., 2004) and the Java based
JADE (Bellifemine et al., 2001), JACK (Busetta et al., 1999) and Jill. Jill
is a relatively new Java-based BDI platform that is lightweight, scalable
and geared towards integration with large-scale simulations.
It has been
12
developed specifically to handle a large number of agents.4
2.3 Social Simulation
The concept of applying computational techniques to social problems has
emerged in the last 30 years as a new way to model the complex and often
non-linear systems that occur in human societies. Axelrod (1997) proposes
in a somewhat extrapolatory manner that social simulation is a 'third way of
doing science', and that it allows both adaptive (non-optimal) and rational
behaviour to be analysed by modellers. Miller & Page (2007) consider the
relatively new computational approach as complementary to the more tradi-
tional methodologies in social science. Another major benefit of employing
simulation is that very simple and explicit assumptions can be used to gen-
erate complex and unpredictable behaviour which would be very difficult to
understand analytically (Gilbert & Troitzsch, 2011).
Drawing ideas from computer science, mathematics, biology, and social
science, social simulation provides a forum for cross-discipline interaction,
and development has come from a number of domains (Axelrod, 1997). The
UrbanSim model (Waddell, 2002) is a polished web-based program that ap-
plies simulation to urban planning problems and allows metropolitan plan-
ning organisations to forecast the effects of growth in a city's population.
In economics, simulation allows researchers to observe aggregate rational-
ity emerging from irrational individual actions (Gode & Sunder, 1993) and
compare the effectiveness of different trading strategies (Rust et al., 1994).
Queuing models are one of the more established applications of social simu-
lation, where discrete events dictate the evolution of a system, and the times
between each event are given by some probability distribution (Gilbert &
Troitzsch, 2011). Schedules generated in this manner generally apply in
customer service settings, so such simulations can be used to measure the
expected average waiting times and idle times associated with different sys-
tem states.
2.3.1 Agent-based Modelling
One commonly used method in social simulation is agent-based modelling
(ABM). Through the actions and interactions of autonomous agents within
the simulated environment, ABM allows modellers to capture the emergent
behaviours of complex systems and has been applied in a range of fields, in-
cluding public health, urban planning, disaster management and economics
4See https://github.com/agentsoz/jill/ for details.
13
(Dawson et al., 2011; Auchincloss & Garcia, 2015). Macal & North (2005),
maintain that 'modelling human social behaviour and individual decision-
making' is the main purpose of ABM. They see the possibilities of ABM
creating new challenges in social science research, including two central ques-
tions:
• How much do we know about credibly modelling people's behaviour?
• How much do we know about modelling human social interaction?
ABMs are particularly suited to modelling discrete real-world phenomena,
where agents represent individuals in the population that can interact with
each other and the environment (Bonabeau, 2002). The scope of ABM
ranges from small-scale academic demonstrations which focus on the key fea-
tures of the system, to large-scale, validated models with millions of agents
that can be used to support policy decisions (Macal & North, 2005). Agents
are required to be discrete, autonomous components of a system that are
able interact with the system's environment (including other agents) and
independently adapt behaviour based on these interactions (Chen & Zhan,
2008).
The ABM approach allows modellers to:
• Expose emergent behaviours: These are defined as characteristics
of the system that are best observed via the interactions between the
individual agents (Chan et al., 2010).
• Naturally describe systems: ABM also allows a natural synthesis
of a complex system into a descriptive model (Bonabeau, 2002).
• Apply stochasticity: Complex behaviours are represented in terms
of discrete agent activities, which allows randomness to be applied in
a targeted and realistic manner (Crooks & Heppenstall, 2012).
• Seek expert validation: Because agent behaviour is explicitly de-
fined by these activities, domain experts are able to more easily un-
derstand, calibrate and validate the model (Bonabeau, 2002).
ABM is considered by Bonabeau (2002) to be a way of thinking about
and analysing systems rather than a specific technological tool.
It pro-
vides a microscopic, bottom-up approach to modelling wherein a system
is represented by its constituent parts. The contrast between ABM and
'macroscopic' modelling provides a new perspective on the scientific pro-
cess. For instance, Epstein (2009) argues that simulating a pandemic with
14
individual agents interacting via social network structures is more effective
than classical disease modelling, as it easily allows for heterogeneous fac-
tors and non-random mixing (Grimm & Railsback, 2005). Bazzan et al.
(1999) consider the application of microscopic simulation to transportation
science. They argue that the concept of rationality is very relevant in a
traffic simulation, and that maximisation or optimisation should not be the
goal.
Instead it should feature the emotional, non-logical components of
decision making, like impulsive lane changing, trying a different route due
to impatience, or how a person's internal mental state might be expressed in
their driving performance. BDI is cited as an appropriate tool for introduc-
ing these individual actions within the social framework of the road network.
Gilbert (2006) makes a similar conclusion, holding that the micro cognitive
level should both influence and be influenced by the macro social level in
models of human behaviour. This is a crucial element of social simulation
and illustrates that an agent-based evacuation model should not focus too
heavily on one level of detail. The interaction between the agents and their
surrounding environment will reveal as much as analysis of standalone social
or individual outcomes.
2.3.2 Synthetic Populations
Social simulation often requires that the modelled agents to reflect attributes
observed in real world data. A synthetic population is one that is derived
from aggregate statistical data to represent the actual population. This is of-
ten necessary because the micro-data required to accurately model the pop-
ulation at an individual level is either unavailable or inaccessible (Moeckel
et al., 2003). There are several established techniques for creating synthetic
population datasets, and often these use actual but anonymous individual
sample sets attained from census data to generalise incrementally to create
a full population (Harland et al., 2012). These methods, including deter-
ministic re-weighting (Ballas et al., 2005) and the conditional probabilities
model (Birkin & Clarke, 1988), aim to match the distribution of persons
and households to demographic data (Jain et al., 2015). However, these
processes require a stable and realistic set of actual sample data (Wickra-
masinghe et al., 2017). In regions that feature a large transient or fluctuat-
ing non-resident population then this information may not exist in a reliable
form, as any existing individual or household data may not generalise to the
period that the simulation is aiming to model.
15
2.4 Existing Evacuation Models
In this section, several models related to the aims of this thesis are de-
scribed. They help guide strategies in the development of our own model,
as well as reveal approaches that might be counter-productive. Shahparvari
et al. (2015) approach the problem of late evacuation in a bushfire from an
optimisation perspective with a focus on limited resources and constraining
time windows, road disruptions and shelter capacities. This process is useful
in determining which factors have the most impact on an evacuation but has
little potential to accurately predict the outcome of an evacuation as human
behaviour in an unpredictable situation is likely to be sub-optimal, although
there have been attempts to validate optimised plans via agent-based simu-
lation (Pillac et al., 2016).
One way that other models have sought to capture suboptimal behaviour
has been to emulate anxiety and human interaction using sensory param-
eters. This has been a focus in crowd-based evacuation simulation, where
agents will speed up due to an awareness of danger or follow another agent
if they are related to them (i.e. parent-child pairs) (Okaya & Takahashi,
2011). Lee et al. (2010) expand on this idea with their leader-follower dy-
namic, which they define via a 'confidence index'. Agents who perceive
sounds and smoke from a bomb blast will update their confidence index and
then, based on some threshold, transition through different states that affect
their decision options (using the BDI framework). If their confidence index
is low, they are likely to follow other agents rather than assess other cues for
exit points. Agents also have a predefined type -- 'Novice' or 'Commuter' --
which changes their knowledge of escape routes and their initial likelihood
to be a follower or leader.
This concept of following or leading can be translated to traffic evacua-
tions and on road behaviour. Yuan et al. (2017) propose a modular driving
decision framework that allows for herding, congestion avoidance, panic-
affected driving and variable shelter selection. They incorporate a number
of different routing behaviour options -- e.g. following shortest path, follow-
ing a leading car or using a GPS -- and then use a BDI cognitive layer to
determine the course of action to achieve their goal. One important aspect
of their model is that it allows for heterogeneity of route choices rather than
focussing on the optimal solution, highlighting Sadri Arif Mohaimin et al.
(2014) findings that 'evacuees may not follow a recommended route but may
take a usual or familiar one instead'. Further, they claim that the route and
destination choices of drivers have the most affect on clearance time in a
large scale evacuation.
16
Scerri et al. (2010) use this same modular approach to explore bushfire
response strategies. The decision on when and where to leave uses a cell-
based ABM and focuses on a number of attributes like age, gender, panic
level and known information. A separate traffic model uses the lightweight
coordinate-based Repast program, and has agents choose a route that takes
the shortest path they believe to be safe. This belief is updated dynamically
as the traffic simulation progresses. The separation of the cognitive and
traffic models, with a parameterised trigger leading to the dynamic traffic
simulation, allows the outcomes of each decision model to be more easily
analysed.
A somewhat unique bushfire evacuation model is provided by Adam &
Gaudou (2017), who directly use survivor testimonies to formulate agent
behaviour. They argue that this allows them to simulate actual behaviours,
rather than behaviours prescribed by experts. A tension between perceived
and actual danger was highlighted in their review of post Black-Saturday
interviews, where people would quickly shift their perception from passive to
hyper-alert once the risk became undeniable. Using a finite-state machine,
survivor responses were translated into several stage states that an agent
may pass through -- 'Unaware', 'Aware indecisive','Preparing to escape', 'Es-
caping' and 'Preparing to defend'. These stages are linked by various triggers
that allow them to move into a new state. Triggers are changes in person
attributes in response to changes in the environment, which is limited to a
simplistic grid model lacking geospatial attributes. However, their focus on
reproducing realistic behaviours allows them to closely match their results
with observed death causes in a fire, validating their intensive, qualitative
approach.
2.5 The Emergency Evacuation Simulator
As we have already seen, capturing irrational behaviours is especially im-
portant in an evacuation simulation (Barrett et al., 2000). Time pressures
and limited route choice, along with competing and overwhelming new lay-
ers of information (much of which may be unknown to an individual agent)
should result in a wide array of responses which diverge from a hypothetical
optimal solution. Padgham et al. (2014) provide one such model which em-
ploys BDI to augment an ABM in the context of bushfire evacuation. They
describe a method where an BDI agent acts as the cognitive 'brain' to the
active 'body' of a corresponding ABM agent. The ABM will act based upon
instructions from the BDI agent in addition to responding directly to inter-
actions within the ABM system, and the BDI agent will reason and make
17
decisions based on information the ABM agent encounters. These coupled
agents can communicate via
• percepts: where an ABM agent notifies the BDI agent of an event in
the system.
• queries: where a BDI agent requests information from the ABM
agent.
• actions: where the BDI agent, having reached a decision, informs the
ABM agent of what action to take.
• action states: where an agent conveys to its counterpart that the
status of an action has been updated.
This model has been refined into the Emergency Evacuation Simulator
(EES) tool designed to assist emergency services in Australia respond to
natural disasters (Singh & Padgham, 2017). The tool features a web inter-
face which allows domain experts to build and test community evacuation
scenarios first-hand, with a view towards it being both a useful planning
tool and a decision support system during an emergency. As a currently
active project with a number of different stakeholders, the EES provides a
flexible and open platform in which to further test BDI bushfire behaviours
as they are developed in this thesis.
Development of the Jill BDI engine has recently focussed on expanding
the underlying capabilities of the BDI-ABM integration, and much of Jill's
architecture is constructed to interact directly with an ABM (Singh et al.,
2016). Taken as its own modular component in the integration, Jill allows
BDI agents to make decisions based on a goal-plan hierarchy. An agent holds
a number of belief states, which may be altered by information received from
the ABM environment. High-level goals are formed based on changes to
these belief states, and these are reasoned through via a set of conditional
plan options. A selected plan may involve a further set of more specific
goals, and this process continues until eventually a plan lands upon some
decision on how the counterpart agent should act and respond within the
ABM. This process is easily described via a goal-plan tree, which maps
the decision-making process into a flowchart that visualises and explicitly
orders the various goal and plan options that a BDI agent may have. Apart
from being useful in the process of designing a behavioural model, these
goal-plan trees are beneficial as explanatory tools that domain experts and
stakeholders can understand. An example goal-tree is provided in Figure 1.
18
Figure 1: An example of a goal-plan tree in a bushfire context (Singh et al.,
2016).
Goal
Plan
Action
M:msg
RespondBushfire
EvacuateHouse
Remain
PrepareMode
AssembleFamily MoveSafeLoc
GetCar
ArrangeLift
DriveToLoc
WaitPickUp
Walk(Car)Drive(Door)
M:ReqLift
DetermineLoc Drive(L)
For its ABM component the EES utilises MATSim, a multi-agent traf-
fic simulation framework that is implemented in Java (Horni et al., 2016).
The traffic flow model is designed for large-scale simulations and uses a
queue-based approach to move agents through a network graph made up of
weighted nodes and links. Storage and flow capacities in the network con-
strain and shape the movement of an agent, but the main impetus in the
model is an agent's desire to complete an individual plan containing a list of
activities with corresponding locations with end times, along with a travel
mode between each activity. This plan is initially provided as a per-agent
input to the simulation, and is iteratively updated by a co-evolutionary algo-
rithm (Popovici et al., 2012) that optimises the plan's 'score' an individual
level rather than aiming for system equilibrium. These plans are collectively
stored in the population.xml input file.
MATSim is a long-running and extendable open source project that now
features a large number of contributions augmenting its core function as a
traffic simulator. To facilitate these augmentations, MATSim plans can have
any number of contribution-specific attributes appended to them, which
allow additional parameters and inputs to be associated with each agent as
needed. See Figure 2 for an example of a MATSim population.xml file
with attributes added to each agent's plan.
BDI-MATSim contributions using BDI engines such as JACK and JADE
have included applications to taxi administration and bushfire evacuation
(Padgham & Singh, 2016).
In an emergency context, one of the distin-
guishing features of a BDI-MATSim simulation is that plans should not be
19
Figure 2: A MATSim population.xml file with attributes added to agent
plans.
< p o p u l a t i o n >
< person id = " 1 " >
< a t t r i b u t e s >
< a t t r i b u t e name = " B D I A g e n t T y p e " class = " java . lang
. String " > io . github . a g e n t s o z . ees . agents .
b u s h f i r e . R e s i d e n t </ a t t r i b u t e >
</ a t t r i b u t e s >
< plan s e l e c t e d = " yes " score = " 1 4 3 . 5 0 3 4 5 9 7 1 " >
< a c t i v i t y type = " home " x = " 7 6 6 7 2 8 . 6 1 7 3 8 0 2 3 9 " y = "
5 7 5 4 4 4 9 . 1 1 6 8 1 8 6 7 " e n d _ t i m e = " 0 9 : 4 0 : 0 0 " / >
< leg mode = " car " / >
< a c t i v i t y type = " work " x = " 7 8 9 8 7 2 . 4 7 2 6 1 4 3 8 " y = "
5 7 5 2 8 1 1 . 2 6 0 2 3 4 1 7 " e n d _ t i m e = " 1 6 : 3 4 : 0 0 " / >
< leg mode = " car " / >
< a c t i v i t y type = " home " x = " 7 6 6 7 2 8 . 6 1 7 3 8 0 2 3 9 " y = "
5 7 5 4 4 4 9 . 1 1 6 8 1 8 6 7 " / >
</ plan >
</ person >
< person id = " 2 " >
< a t t r i b u t e s >
< a t t r i b u t e name = " B D I A g e n t T y p e " class = " java . lang
. String " > io . github . a g e n t s o z . ees . agents .
b u s h f i r e . R e s i d e n t </ a t t r i b u t e >
</ a t t r i b u t e s >
< plan s e l e c t e d = " yes " score = " 9 3 . 2 9 8 7 7 2 1 " >
< a c t i v i t y type = " home " x = " 7 8 6 8 3 1 . 5 1 1 6 9 1 1 7 5 " y = "
5 7 7 0 0 8 3 . 8 1 3 1 6 7 1 7 " e n d _ t i m e = " 0 8 : 1 6 : 0 0 " / >
< a c t i v i t y type = " shops " x = " 7 5 9 2 9 9 . 7 8 1 6 9 5 2 0 1 " y = "
5 7 2 9 9 0 1 . 2 7 9 8 0 3 7 3 " e n d _ t i m e = " 1 0 : 3 4 : 0 0 " / >
< leg mode = " car " / >
< a c t i v i t y type = " beach " x = " 7 9 1 1 2 0 . 8 0 3 6 4 1 5 5 9 " y = "
5 7 5 2 8 2 0 . 3 5 2 9 6 1 9 8 " e n d _ t i m e = " 1 3 : 0 7 : 0 0 " / >
< leg mode = " car " / >
< a c t i v i t y type = " home " x = " 7 8 6 8 3 1 . 5 1 1 6 9 1 1 7 5 " y = "
5 7 7 0 0 8 3 . 8 1 3 1 6 7 1 7 " / >
</ plan >
</ person >
< p o p u l a t i o n >
20
iteratively improved; all decision-making must be performed 'on the fly' by
the BDI agent. This represents the idea that people in an emergency cannot
rely on their experience or prior learned behaviours, and must react to new
situations as they are presented. With this arrangement in place, the initial
plan inputs become more significant, because they will directly dictate where
agents are prior to the emergency situation. A large part of the contribution
in this thesis is providing a method to generate these plans, including the
BDI-specific parameters that are required for each agent. Together, these
input procedures will shape the resultant simulation, so establishing this
process is an essential design step for the EES project.
3 Population Generation Algorithm
A person's response to a bushfire will inevitably be influenced by their plan
for the day, including what they are doing when they become aware of the
threat and what they plan to do next. In this section, we describe a plan
generation algorithm that allows us to easily generate the planned activities
for every individual in a population on a given day and apply attributes that
reflect how they will respond to a bushfire. If we want to represent these
individuals as agents in an EES simulation then their plans need to follow the
same semantics that MATSim requires. The algorithm output for an agent
plan should consist of a set of sequenced activities, each with a location and
start-time, along with travel legs linking them. Any additional parameters
should be included as attributes per agent plan, with the collection of these
plans forming a population.xml file as in Figure 2.
Although there are a number of established methods for synthesising
populations from census data (Harland et al., 2012), we are interested in
generating populations in a range of diverse settings that may differ fun-
damentally in their underlying demography. The make-up of a population
and its vulnerability to a bushfire threat will vary depending on the weather
conditions, the time of year, which day of the week and even the time of day
that the fire strikes (Reid & Beilin, 2014). These variations are exacerbated
when the considered population has a significant transient element, or in a
scenario where a special event is occurring in the region that skews both
the location and movement patterns for a large portion of the population.
We therefore require a flexible set of inputs that can be manipulated to re-
flect the population in a specific scenario. Additionally, inputs should be
simple enough for emergency personnel to understand and validate against
observed data.
21
This section will proceed by first outlining in a generalised form the
inputs that the algorithm requires to generate agent plans. We then de-
scribe the algorithmic process for creating an agent plan, which includes
assigning activities to an agent, sequencing these activities into a day plan,
and giving each activity in the sequence an appropriate location. Once plan
generation is established, a bushfire behaviour Belief-Desire-Intention model
is formulated. The necessary BDI attribute parameters alongside the plan
generation inputs constitute a set of behaviour profiles that broadly define
how a given agent will react in a bushfire emergency scenario.
3.1
Inputs
We first consider the inputs that the algorithm requires for a given bushfire
scenario with region R and population of agents P over a time period T :
• A set of subgroups S and a size Ps for each s ∈ S that partitions P
up into categories based on an agent's relationship to R.
• A set of activities A (of size K) that an agent can be assigned to at
any time t ∈ T .
• A choice of time-step size T
N , where N denotes the number of time-
steps.
• An activity distribution ∆s for each subgroup s that defines the ex-
pected proportion of s-agents engaged in activity αk ∈ A at each
time-step tn for n = 1, 2, ..., N .
• A duration weighting ds,αk for each (s, αk) pair that dictates how long
an s-agent will take to complete activity αk.
• A set of locations L in R defined by a suitable coordinate reference
system W .
• A location mapping Ms,αk for each (s, αk) pair that defines the set of
possible locations in L where an s-agent can complete the activity αk.
• An allocation number a(cid:96) for each location (cid:96) ∈ L that defines the
expected number of agents attending (cid:96) during the given scenario.
• a family {Li} (indexed by I) of localities that partitions L into neigh-
bourhoods.
• A travel factor gs for each s that governs how likely an s-agent is to
choose a new location outside of its current locality.
22
3.1.1 Subgroups
We partition the population of agents P into subgroups that broadly reflect
the diversity of people in the area and where they have come from. These
will differ depending on R and the given bushfire scenario, but broadly there
should be a focus on knowledge of the area, intended duration of stay in the
area, and connection to the community through friendships, relationships
and assets. The most basic input is thus a description of how the population
should be broken up into distinct groups. Let S be the set of all subgroups
s. We also denote
{Pss ∈ S}
as the partition of the population of agents P into subgroups. One simple
example of a subgroup partition would be breaking the population up into
residents and visitors.
3.1.2 Activity Distributions
Subgroups are mainly formed around expected behaviours during a bushfire,
but they are also useful in defining differing intentions for people in their
daily activities. For a given bushfire scenario, a set of activities that a signif-
icant number of the population is expected to participate in is determined.
These could include going to the beach, shopping, eating at a restaurant or
attending a large event. Any desired granularity is allowable at this step,
but it is better to tend towards more general activity labels here, with more
specific sub-activities distinguished by location (see Section 3.1.3). Addi-
tionally, activities that are only undertaken by a minority can be grouped
together into an all encompassing other activity. The only enforced activ-
ity condition is that α1 should specify a home location for each agent. This
plays a fundamental role in how an agent's plan develops, as we shall see.
We denote
A = {α1, α2, ..., αk, ..., αK}
as the set of K possible activities αk that an agent can be assigned to
at time t ∈ T .
The time period T is then discretised into N steps, and for each time-step
tn, for n = 1, 2, ..., N , and each subgroup s, a map
δs,tn : A −→ [0, 1]
23
K(cid:88)
k=0
s.t.
δs,tn(αk) = 1
defines the expected proportion of subgroup s to be engaged in activity αk
during time-step tn. The set
∆s = {δs,tnn = 1, 2, ..., N}
of N mappings for each subgroup then defines the distribution of that sub-
group over the day. The activity distribution input is most easily understood
in terms of the chart in Figure 3, with each activity being assigned to a pro-
portion of Ps at every tn.
1
)
k
α
(
n
,
s
δ
k
0.5
(cid:80)
0
t1
t2
t3
t4
t5
t6
t7
t8
t9
t10
t11
t12
time-step
α1 α2 α3 α4 α5
Figure 3: Activity distribution table for subgroup s with K = 5 and N = 12.
In Section 3.2.1 we will derive start times from these input distributions
using each activity's duration. We do not impose the start times as an input
because the concept of describing what a population is doing across various
time-steps is simpler to input than specifying exactly when a proportion of
people begin and end activities. Certainly the former is more intuitive for
the user, and seems a lot more feasible from a validation and measurement
perspective.
3.1.3 Location Mapping
We have now defined the types of activities available to an agent, and the ac-
tivity distributions which dictate when they can occur. It remains to specify
24
for all activities their possible locations in the region. Again, this mapping
is dependent on the subgroup an agent belongs to, which allows variations
within activities to be expressed according to who is performing them. For
instance, we may have visitor agents who have their other activity mapped
to popular tourist destinations in the region, whilst resident agents would
have schools, golf clubs etc. as other locations. Location mapping is par-
ticularly important in distinguishing the choice of home location for each
s-agent.
With a set of coordinates W (in a suitable coordinate reference system5)
that cover R, we define a location as (cid:96) = (x, y) ∈ W , and let
L = {(cid:96) (cid:96) ∈ R ∪ W}
denote the set of all locations in the region. Then for each subgroup s and
each activity αk,
Ms,αk ⊂ L
defines the allowable locations for s and αk. Note that it we may have
the case
Ms,αk = ∅ ⇐⇒ δs,n(αk) = 0 ∀n
implying that that an agent of type s will never perform activity αk at any
point during the time period T .
In summary for each subgroup s there are three main inputs that must
be specified by the user: the distribution of activities, the locations those
activities can occur and the number of s-agents. There are other more
specific input parameters that will be introduced at the the relevant step of
the algorithm explained below.
3.2 Activity Assignment
We now explain the algorithm's process per agent plan. The first step in
constructing an s-agent plan is to decide which activities will be performed
at each time-step. These activities are drawn from the corresponding distri-
bution table and then sequenced according to their assigned duration.
5We assume a Cartesian coordinate system e.g. Universal Transverse Mercator (UTM).
25
3.2.1 Selecting a Feasible Activity Set
For each subgroup s, the input distributions must be converted into a cu-
mulative probability matrix. The distributions only tell us where the agents
should be at a given time, and not when they begin that activity. Therefore
the start times must be derived, and to do this we must know the expected
duration of each activity. These durations will be some multiple of the length
of each time-step T
N . We denote a duration weight
ds,αk ∈ {1, 2, ..., N}
such that T
N ds,αk gives the corresponding duration for subgroup s performing
activity αk. Then at each time-step, we can determine the number of s-
agents who are expected to start activity αk by subtracting the proportion
who started αk during any of the ds,αk − 1 previous time-steps. We form
the recursive function
ξs,tn(αk) = δs,tn(αk) −(cid:88)
j
ξs,tn−j (αk)
j ∈ (0, ds,αk − 1] ∩ N
(1)
Since for n ≤ 0, ξs,tn = 0 (we assume no agent has started an activity
outside of the time period), ξs,t1 = δs,t1 for all αk and thus we can recursively
define the starting proportion for each activity at each time period using
(1).
Now, for each s we can view the collection of ξs,tn(αk) as a K×N matrix
Ξs that describes all start time proportions for that subgroup:
ξs,t2(α1)
ξs,t2(α2)
...
ξs,t2(αK)
···
···
. . .
···
ξs,tN (α1)
ξs,tN (α2)
...
ξs,tN (αK)
Ξs =
...
ξs,t1(α2)
ξs,t1(α1)
k(cid:88)
ξs,t1(αK)
Note that the sum of the nth column in Ξs,
ξs,tn(αk) ≤ 1
j=1
(the remaining proportion represents s-agents who are currently engaged in
an activity at tn), and so we normalise Ξs
26
Ξ∗
s =
=
...
...
ξs,t1 (α1)
j=1 ξs,t1 (αj )
ξs,t1 (α2)
j=1 ξs,t1 (αj )
(cid:80)k
(cid:80)k
(cid:80)k
ξs,t1 (αk)
j=1 ξs,t1 (αj )
ξ∗
ξ∗
ξ∗
ξ∗
ξs,t2 (α1)
j=1 ξs,t2 (αj )
ξs,t2 (α2)
j=1 ξs,t2 (αj )
(cid:80)k
(cid:80)k
(cid:80)k
ξs,t2 (αk)
j=1 ξs,t2 (αj )
···
···
...
. . .
s,t2,K ···
ξ∗
ξ∗
...
ξ∗
s,t1,K ξ∗
ξ∗
s,t2,2
s,t2,1
...
s,t1,1
s,t1,2
s,tN ,1
s,tN ,2
s,tN ,K
···
···
. . .
···
(cid:80)k
(cid:80)k
(cid:80)k
ξs,tN (α1)
j=1 ξs,tN (αj )
ξs,tN (α2)
j=1 ξs,tN (αj )
...
ξs,tN (αk)
j=1 ξs,tN (αj )
where each (k, n) entry now represents the probability of an s-agent who has
ended any activity at time-step tN beginning new activity αk. Finally, we
take the cumulative sum of each column to attain a cumulative probability
matrix:
s,t1,1
(cid:80)2
ξ∗
k=1 ξ∗
(cid:80)K−1
...
k=1 ξ∗
1
Cs =
s,t2,1
(cid:80)2
ξ∗
k=1 ξ∗
(cid:80)K−1
...
k=1 ξ∗
1
s,tN ,1
···
··· (cid:80)2
··· (cid:80)K−1
ξ∗
k=1 ξ∗
...
k=1 ξ∗
1
···
. . .
s,tN ,k
s,tN ,k
s,t1,k
s,t2,k
s,t1,k
s,t2,k
Now that we have a well-defined probabilistic description of when an
agent will perform an activity, we can begin to construct the plan by selecting
activities from Cs.
For a given s-agent, we generate a 1 × N vector U of values drawn from
the uniform distribution on (0, 1). Then, conditionally checking the nth
value of U , un against the nth column of Cs, we select one activity for each
tn. Form B, a K × N matrix with each bk,n entry defined by
b1,n =
bk,n =
1 un < c1,n
0 un ≥ c1,n
1
0 un ≥ ck,n
ck−1,n ≤ un < ck,n
k = 1, 1 ≤ n ≤ N
2 ≤ k ≤ K, 1 ≤ n ≤ N
For each column, we only have one non-zero value, so B is a Boolean
matrix representing an activity at each time-step tn. These pre-selections
27
(cid:40)
(cid:40)
can be thought of as potential activities at that time, dependent on the
duration of previous activities.
defining B(cid:48):
To actualise these activities, we inject the duration weights into B by
(cid:40)
b(cid:48)
k,n =
ds,αk when bk,n = 1
0
when bk,n = 0
1 ≤ k ≤ K, 1 ≤ n ≤ N
We now cycle through each time-step to determine its affect on the next.
Algorithm 1 illustrates this process whereby the duration weight ds,αk blocks
out subsequent time-steps until the full duration of αk has been performed.
Note that we define
to be the nth column of B.
b·,n
Algorithm 1: Activity selection for an s-agent
1 Form B, B(cid:48) for s-agent
2 Set block_flag=0
3 for n ∈ {1, 2, ..., N} do
4
5
6
7
8
if block flag > 0 then
b·,n = b·,n−1
block flag=block flag −1
else
block flag=(cid:80)K
k=1 b(cid:48)
k,n − 1
9 return B
At the end of this process, the resultant Boolean matrix B will contain
a feasible plan which describes what the s-agent will be doing at each time-
step, in line with the format of the initial distributions (see Figure 4).
3.2.2 Sequencing the Daily Plan
It remains to translate the tn at which the agent changes activity into mean-
ingful time and provide an ordered list of start times for each activity in the
plan. We define a recursive function on the columns of B
f (b·,n) = b·,n − b·,n−1
2 ≤ n ≤ N
28
Figure 4: An example output from Algorithm 1 with K = 5 and N = 12.
B =
1 1 1 0 0 0 0 0 0 0 1 1
0 0 0 1 1 1 1 0 0 0 0 0
0 0 0 0 0 0 0 1 1 0 0 0
0 0 0 0 0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 0 0 0 0 0
(cid:40)
f(cid:48)
k,n =
with f (b·,1) = b·,1.
We then set
F = f (B) = f ([b·,1, b·,2, . . . , b·,N ]) = [f (b·,1), f (b·,2), . . . , f (b·,N )]
which will have entries from {−1, 0, 1}. The only elements of F that will be
greater than 0 i.e. unaffected are those where the agent has changed plan
from tn−1 to tn. Therefore we can form F (cid:48) such that each entry is defined
1
0
fk,n > 0
fk,n ≤ 0
1 ≤ k ≤ K, 1 ≤ n ≤ N
Then, the sum of the nth column in F (cid:48) defines a Boolean true or false
as to whether an activity begins at time-step tn, and the row k containing
the non-zero entry in column n determines the activity αk that begins then.
This process is illustrated in Algorithm 2.
Line 11 gives the activity a start time at a random point within the length
of the time-step tn. This is important because it ensures an even distribution
of start times for activities over the population; if the times were fixed to
each tn, then we would have unnatural 'waves' of trips occurring periodically.
Note that runif(-1,1) on this line denotes a uniformly distributed random
variable in the interval (−1, 1). Another purpose served in Algorithm 2 is
the allocation of an additional activity α1 at the start and end of the plan.
This is a requirement of MATSim's re-planning features, and it ensures that
agents always return to their home the end of the time period.
3.3 Assigning Locations to Activities
Once an s-agent's time-ordered list of activities is determined, locations
must be assigned to each activity. Activity locations are chosen based on a
number of factors, including the distance required to travel, and the relative
29
Algorithm 2: Algorithm for selecting times that activities begin
1 Form F (cid:48) for s-agent
2 create empty list plan
3 activity = α1
4 time = 0
5 add (time, activity) to plan
6 for n ∈ {1, 2, ..., N} do
if (cid:80)K
7
8
9
10
11
12
k=1 f(cid:48)
for k ∈ {1, 2, ..., K} do
k,n = 1 then
if k = 1 then
activity = αk
time = tn + runif(-1,1) T
2n
add (time, activity) to plan
13 activity = α1
14 time = T
15 add (time, activity) to plan
16 return plan
popularity of the location. In this section we will often introduce probabili-
ties that are (inversely) proportionate to distances. In the algorithm, these
probabilities are properly formed by normalising the inverse distances over
the sum of all inverse distances (in much the same way that we normalised
Ξs in Section 3.2.1). To notate this at each step would be cumbersome and
distracting (see the end of Section 3.3.1 for evidence of this), so we maintain
the proportionate notation throughout.
3.3.1 Localities
If we first assume that an s-agent is at a certain location (cid:96)0, then the question
is how to choose the location (cid:96)∗ of the next activity αk in the list. This could
be done based on inverse distance:
Pr(αk occurs at (cid:96)) ∝
1
dist((cid:96)0, (cid:96))
where dist((cid:96)0, (cid:96)) denotes the Euclidean distance from (cid:96)0 to (cid:96) for some
(cid:96) ∈ Ms,αk .
30
However, when there are a large number of location options for an activity,
it becomes costly to calculate the distance for every agent to every potential
destination, and scalability to larger areas or increasing the detail of location
data becomes problematic. Further, a purely distance-based metric tends to
result in trips that cluster around the agent's home location; for each new
trip, the agent will generally take the closest option to their current location.
This is amplified in built-up areas, where an agent is likely to have more
options close to them. Whilst intuitively this makes sense and agrees with
many of the assessments made by Song et al. (2010), over a population it does
result in movement patterns that are almost too optimal, with agents either
commuting between communities or not at all. It is desirable to encourage
some intra-community trips which disperse traffic away from major arterial
routes, and reflect the possibility (also discussed by Song et al.) that people
within a neighbourhood may have preferences that do not depend entirely
on distance.
Both of the above issues can be addressed if we group locations by some
locality area that they lie within. These localities can be based on the
existing suburban/postal districts in the area, by ABS statistical measures,
or by any other grouping appropriate to the community in question. The
localities {Li} of a region are defined as a family of I sets such that
Li ⊂ L
∀i ∈ I, and (cid:91)
i∈I
Li = L,
(cid:92)
i∈I
Li = ∅
i.e. {Li} is a partition of L.
In practical terms, we draw a polygon cover of R and denote all locations
under polygon Li as belonging to locality Li (illustrated in Figure 5).
For each Li, we define a centroid location (cid:96)i as
(cid:88)
(cid:96)∈Li
(cid:96)i = (
1
Li
x(cid:96),
1
Li
(cid:88)
(cid:96)∈Li
y(cid:96))
where Li is the number of locations in Li and x(cid:96) and y(cid:96) denote the x and
y co-ordinates of (cid:96) ∈ W ∪ Li. Now if we take the current locality to be
Li0 (cid:51) (cid:96)0, we can choose a new locality Li with probability proportional to
the inverse distance from centroid (cid:96)i0 to centroid (cid:96)i:
31
Figure 5: Example of location points (pink) being grouped by locality poly-
gons (green).
Pr(αk occurs in Li) ∝
1
dist((cid:96)i0, (cid:96)i)
It is then a simple matter to randomly choose from the selected locality
Li∗ a location
(cid:96)∗ ∈ L∗
i ∩ Ms,αk
for αk, the next activity in the plan list. A major benefit of this process
is that the distances between each centroid are fixed and only need to be
calculated once during the entire algorithm runtime. Thus we can drastically
cut down on distance calculation requirements whilst also gaining the ability
probabilistically encourage mildly suboptimal choices.
One problem this presents is that the most likely choices for (cid:96)∗ should be
located within Li0, but the distance to them in terms of locality centroids is
dist((cid:96)i0, (cid:96)i0) = 0
so the probability of selecting from the current locality is undefined. How-
ever, this provides an opportunity to introduce a new parameter to the al-
gorithm. For each subgroup s we define a travel factor gs, which determines
32
how likely an s-agent is to leave its current locality. This allows for agents
in some subgroups to be more likely to remain within their neighbourhood,
and alternatively can encourage other subgroups to move around the region
throughout the day. Under pure locality-based selection, gs directly repre-
sents the probability of an s-agent moving to another locality for αk. This
definition for gs implies that
and logically,
gs =
Pr(αk occurs in Li0) = 1 − gs
(cid:88)
Pr(αk occurs in Li) ∝(cid:88)
i∈I\{i0}
1
dist((cid:96)i0, (cid:96)i)
i∈X
(2)
where X = {i ∈ I \ {i0}Li ∩ Ms,αk (cid:54)= ∅} i.e. X only consists of i such that
locality Li contains valid locations at which activity αk can be performed
by an s-agent. So we can write (2) as
1 − gs
gs
Pr(αk occurs in Li0) =
Pr(αk occurs in Li)
(cid:88)
i∈I\{i0}
and this allows us to define a pseudo 'distance'
dist0((cid:96)i0, (cid:96)i0) =
(1 − gs)(cid:80)
gs
i∈X
1
dist((cid:96)i0 ,(cid:96)i)
(3)
relative to all other locality distances that contain locations in Ms,αk . Ob-
serve that this definition recovers (2) after normalisation:
Pr(αk occurs in Li0) =
dist0((cid:96)i0 ,(cid:96)i0 )
=
1−gs
gs
dist((cid:96)i0 ,(cid:96)i)
1
1
1
1−gs
gs
dist0((cid:96)i0 ,(cid:96)i0 )
i∈X
i∈X
+(cid:80)
(cid:80)
dist((cid:96)i0 ,(cid:96)i) +(cid:80)
(cid:80)
dist((cid:96)i0 ,(cid:96)i) +(cid:80)
i∈X
1−gs
gs
1
1
dist((cid:96)i0 ,(cid:96)i)
i∈X
1
dist((cid:96)i0 ,(cid:96)i)
i∈X
i∈X
i∈X
1
(cid:80)
(cid:80)
1
dist((cid:96)i0 ,(cid:96)i)
1
dist((cid:96)i0 ,(cid:96)i)
=
=
=
1−gs
gs
1−gs
gs
1−gs
gs
+ 1
1 − gs
1 − gs + gs
33
= 1 − gs
Unlike the other locality distances, the pseudo-distance is dependent on
αk and s, so it must be recalculated at every new location assignment step.
This definition for dist0((cid:96)i0, (cid:96)i0) will be important in Section 3.3.2.
3.3.2 Allocation Considerations
Another complication that is still present with locality-based selection is
that each probability is still roughly proportional to the inverse distance
from the agent's current location. There is no consideration for the size
of each locality Li, so locations in smaller localities are actually favoured
because they are generally competing in a smaller pool (once their locality
has been chosen). This seems counter-intuitive, and we should expect a
heavier traffic flow towards more bigger localities rather than away from
them. Whilst this is mitigated by a larger proportion of agents having their
home in built up areas, it can still cause a significant and unusual build up
of traffic in areas that should not be so busy. This can be exacerbated if
there are a lot of s-agents with a gs that encourages them to travel; imagine
a one pub town that suddenly and inexplicably has an influx of hundreds of
tourists! Whilst this is a plausible scenario, it should not exist by default,
and we would prefer to be able to control this.
We can add this control by introducing an additional weighting to each
locality that reflects its overall size. Not only should we account for the
number of locations in the locality, but also have some measure of their
popularity and expected usage. Each location (cid:96) is assigned an allocation
number a(cid:96). Conceptually, a(cid:96) represents the expected number of agents that
will visit (cid:96) during the given scenario, but its main purpose here is to weight
the probabilities for each locality. We now give the probability that the next
activity αk occurs in locality Li as:
Pr(αk occurs in Li) ∝
(cid:88)
(4)
1
dist((cid:96)i0, (cid:96)i)
(cid:96)∈Li∩Ms,αk
a(cid:96)
Localities with higher overall allocations for the relevant activity are
given a higher weighting. Importantly, the current locality is equally affected
by this metric, with a 'distance' being prescribed as in (3) (note that gs is
now not a proper probability, but a factor that determines the pseudo-
distance) so an agent's decision to stay within their current locality will also
consider the options on offer there. This method of selecting destinations
is a version of the gravity model of transportation (J. E. Anderson, 2011).
It allows the dispersion of agents throughout the day to be more tightly
34
controlled, and for hotspots to be set according to the prevailing conditions.
For instance, if on a given day there is an event at one particular beach, then
the allocation for that beach should be raised to encourage those agents with
beach activities to go there.
3.3.3 Home Selection
We have discussed how agents will select a new location based on their
current location, but have not yet established where they are initialised.
For most agents, the location they begin at will heavily influence all further
trip locations. Each plan has been explicitly constructed to ensure that an
agent begins and ends their day with the same α1 activity, so unlike with
other activities, we guarantee that an s-agent's home location remains the
same every time it is assigned (note that as defined in Section 3.1.2, home and
α1 are synonymous). As it will always be the first activity in an s-agent's
ordered daily plan, α1 is also the only activity that needs to be assigned
without the context of a prior (cid:96)0 location. Instead, we randomly choose Li
from the set of eligible localities
{LiLi ∩ Ms,α1 (cid:54)= ∅}.
and then consequently we can choose (cid:96) ∈ Li ∩ Ms,α1 as the location for α1.
The allocation numbers a(cid:96) for (cid:96) ∈ Ms,α1 play a special role in the home
selection process. In this case, the conceptual understanding of a(cid:96) as the
expected number of people attending the location during the scenario does
impose a hard cap on the number of agents that can reside at (cid:96). Whenever
an agent (from any subgroup) is assigned (cid:96) as their home location we deduct
1 from a(cid:96), globally affecting the likelihood of that location being assigned to
another agent. If a(cid:96) = 0, then (cid:96) is removed from L. This ensures that the
population is initially distributed in a manner that matches the region. It
is therefore especially important that all specified home locations
(cid:91)
s∈S
Ms,α1
and the associated allocation numbers reflect the true distribution and den-
sity of people in each area of interest, because this will dictate where traffic
is heaviest for all other activities. Algorithm 3 describes the location section
of the algorithm. The function choose randomly samples from the specified
set.
35
Algorithm 3: Choosing activity locations for all agents
1 Form plan for each agent∈ {1, 2, ..., P} via Algorithm 2
2 for s ∈ S do
for agent ∈ {1, 2, ..., Ps} do
3
4
5
6
7
8
9
10
11
12
home locality =choose {LiLi ∩ Ms,α1 (cid:54)= ∅}
home location =choose {(cid:96) ∈ home locality ∩ Ms,α1a(cid:96) > 0}
a(cid:96) = a(cid:96) − 1
if a(cid:96) == 0 then
L = L \ {(cid:96)}
(cid:96)0 = home location
for activity ∈ planagent \ {α1} do
locality =choose Li ∈ R with weighted prob. from (4)
activity location =choose (cid:96) ∈ Li ∩ Ms,activity
3.4 Adding Bushfire Behaviours
The above algorithm creates a population.xml that is ready for MATSim
to use on a regular day, but it will not contain any of the information re-
quired to describe how the agents will act in a bushfire. We now describe a
behaviour model that allows an agent to disrupt these usual daily plans by
taking the exceptional circumstances into account. This model is designed
to be implemented in the BDI-ABM integration framework described in
(Singh et al., 2016), where agents running in the BDI engine Jill serve as
cognitive counterparts to the active ABM agents in MATSim (see Figure 6).
The behaviour model thus follows Jill's architecture, where goal-plan hi-
erarchies filter abstract belief states down to actions for the ABM agent
execute. The design of the hierarchies is based around the same broad
input subgroups that govern the population generation algorithm, and ac-
cordingly we initialise behaviours as appended attributes in the generated
MATSim population.xml. These BDI attributes -- together with the asso-
ciated characteristics that emerge from the population generation inputs --
form a behaviour profile that describes how an agent is going to act in a
given bushfire scenario.6
6As we move into discussing the bushfire behaviour model, our language will shift
towards a more descriptive tone. From here on, the focus is less on formally defining
concepts, and more on explaining the methods and rationale behind our attempt to capture
human behaviour in the model.
36
actions
percepts
status
A1
BDI System
A2
A3
A1
ABM System
A2
A3
Figure 6: A diagram representing the BDI-ABM integration (Singh et al.,
2016).
3.4.1 Threshold Model
Since the 2009 Black Saturday bushfires in Victoria, most Australian re-
search regarding bushfire behaviour has concerned if and when people will
evacuate in the face of the threat. This decision process is central to the be-
havioural model. We need to accommodate the temporal aspect of incoming
information in an unfolding situation, where multiple warning sources can
have a cumulative effect on a person's actions. The scheme should also en-
sure that there is variation in behaviour even within subgroups; some people
will intrinsically be more inclined to leave early, whereas others might never
evacuate even in the face of severe danger.
For this reason, we employ a threshold-based model, where each agent
has an intrinsic tolerance value which determines how risk averse they are,
and therefore how likely they are to respond to the new information about
the bushfire threat. Incoming information, or percepts, are broken up into
two categories. Environmental alerts pertain to information gleaned di-
rectly from the surroundings. They will include embers, smoke, flames, the
movement of neighbouring agents, and the actions of emergency services
in the vicinity. Transmitted alerts include those that are issued by the
emergency authorities before and during a bushfire emergency and other
messages received via the media or other communications networks. Let E
and T denote the sets of all alerts e and τ for each category. Both sets are
37
totally ordered
e1 ≤ e2 ≤ ... ≤ eq ≤ ... ≤ eQ,
τ1 ≤ τ2 ≤ ... ≤ τm ≤ ... ≤ τM
i.e. we denote the severity or value of each percept relative to the oth-
ers in its category. Whenever an agent receives a percept, it will update
its 'barometer' for the corresponding alert type. The two categories dif-
fer here in that the Environmental barometer only updates monotonically
when the new alert is of greater severity than the previous alert, whereas the
Transmitted barometer can also be decreased based on a reduced level of
threat. This reflects the nature of the different alert types; if you see smoke,
it doesn't reduce the threat of the flames you had seen earlier, but if you
receive a less severe warning from an authority it is likely to decrease the
concern caused by an earlier message. We now have a way of capturing the
different types of threat percepts that might trigger an evacuation, defined
a method of updating the severity level of the threat, and separated the
percepts into groups that broadly maintain a distinction between primary
and secondary information.
We now must combine these two barometers into one value that reflects
a person's overall threat level. If we consider
E × T = {{0, 0},{e1, 0},{0, τ1},{e1, τ1}, ...,{eQ, τM}}
representing the set of all combinations of alerts that an agent may have
registered as their barometer at some time (with 0 representing no threat
of that type being received). There is naturally a partial ordering on the
ordered pairs for each fixed value of eq or tm (stemming from the ordering in
the other category set). However, there is no defined order of, say, {e1, t3}
and {e2, t2}. We therefore need to define a mapping
r : E × T −→ (0, 1)
which maintains all partial orders and also defines an order between each
undefined element pair. Note that the value set that r maps to is arbitrary,
and serves only to quantify the ordering into a ranking score. We may then
have the case that, say, r({e1, t3}) = r({e2, t2}) i.e. an agent can arrive at a
ranking score through various different combinations of alerts. The ranking r
of the current barometer levels is then compared against an agent's intrinsic
tolerance score tol, and if r > tol then the agent will change their daily
plan.
38
3.4.2 Possible Bushfire Behaviours
We now show how changes of plan are decided upon by an agentusing the
Belief-Desire-Intention model of cognition. We design and implement be-
haviours based on what is understood to occur in a bushfire situation. Fun-
damentally, we have constrained variation of behaviour to three end goals:
• Go home now: The agent will take a route towards their home location.
It is important that every agent has a home location, as specified in
3.1.2.
• Go to dependant now: The agent will take a route towards the loca-
tion of a dependant. Here a dependant may represent a child, elderly
relative, farm animal or some other responsibility that an agent may
have that will require them to detour before going home or before
leaving.
• Leave now: The agent will decide on an evacuation destination and
begin to move there.
These goals are triggered as the result of the reasoning process that
accounts for current location, whether the agent is also planning on going
home, and if there are dependants to attend to. We do this using two
threshold levels for each agent. The first, tol=INIT represents the level at
which an agent decides to stop their daily plan, but has not yet committed
to seeking shelter, and triggers the high-level Initial Response goal. The
second, tol=ACT, defines the level at which an agent will decide to leave, and
will trigger the Act Now goal. Both high-level goal structures are described
as goal-plan tree diagrams in Figure 7 and Figure 8.
Both INIT, ACT ∈ [0, 1], with 0 and 1 indicating that the threshold is
always or never broken respectively. It is clear too that the two decisions
happen in order, so INIT ≤ ACT. The distinction between the two thresholds
is important as it allows a diverse range of responses. An agent may have its
INIT threshold broken, returning to home and then never exceeding its ACT
threshold; this would represent someone who has chosen to stay and defend
their property. Likewise, we can define agents who will never be allowed to
do this by setting INIT = ACT i.e. both thresholds will be triggered at the
same time. Note that if the agent has a dependant, then they are guaranteed
to execute the Go to dependant goal even in the case that INIT = ACT; this
is illustrated in the goal-plan tree.
39
Initial
Response
dep=T
dep=F
1. If dependant is
nearby then visit them
else
Go home with some
probability then
visit dependant
2. Go home with
some probability
If ACT threshold met
continue daily plan
else
Go home
Go Home
Go To
Dependant
Figure 7: Initial Response goal-plan tree (goals are in green, plans are in
yellow).
40
Act Now
dep=F
dep=T
1. Go home with
some probability
2. Leave now
Leave now
Go Home
Leave Now
Figure 8: Final Response goal-plan tree (goals are in green, plans are in
yellow).
3.4.3 Assigning Attributes
Despite the simplicity of the BDI structure above, it can result in a range
of outcomes for an agent depending on their threshold levels, whether they
have a dependant, and on the probability that dictates whether the agent
should go home. This is further diversified when you consider the number
of destination options implicit in the Leave now goal. To tie this model to
the population generation algorithm, we choose an allocation of these pos-
41
sibilities that matches the expected behaviour of the predefined subgroups
in a bushfire.
For a subgroup s, we define the following parameters:
• prob of dependants: a probability determining how likely an s-agent
is to have a dependant.
• stays: a Boolean variable determining whether an s-agent will be per-
mitted to stay and defend i.e. whether the strict inequality INIT<ACT
is possible.
• prob of go homes: a probability determining how likely an s-agent is
to return home before leaving.
• [threshold mins,threshold maxs]: an interval within the chosen rank-
ing range which determines where INIT and ACT should lie within.
We can then attribute to each agent their INIT and ACT thresholds,
determine if they have a dependant and indicate whether they will decide
to Go home now.
Lastly, we must assign dependant, invacuation and evacuation locations.
The dependant location can either be chosen within some radius of home, or
if there are particular areas likely to have dependants (e.g. schools, retire-
ment homes), these can be specified as dependant locations and selected in
the same manner as in Section 3.3. The invacuation and evacuation points
will determine preferred destinations to evacuate to once ACT has been trig-
gered. An agent will first attempt to reach its preferred evacuation location,
but if that is not possible, it will seek refuge within its current locality (with
the invacuation point as preference if possible). Clearly there are many po-
tential situations that might prevent the agent accessing either location, but
having these defined preferences allows behaviours to be further shaped on
the subgroup level. To determine the known invacuation and evacuation
points, we also use a similar method as in Section 3.3. We want invacuation
points to be close to the agent's home, so we enforce a strict preference for
invacuation locations within the locality of the agent's home. Conversely,
evacuation points should be far from the home, so we use probabilities pro-
portional to distance rather than inverse distance. In this case, a person
is more interested in getting out of the area than evacuating to a specific
location, so entire localities are weighted by their total capacity, and the
destination point is the central node of the locality.
42
Figure 9 provides an example of these attributes within a MATSim
population.xml file. We conclude this section by highlighting how the cho-
sen subgroups S have fundamentally shaped both the population generation
and the assignment of BDI attributes. Indeed, a particular agent belonging
to a subgroup s will have a listable set of potential behaviours that com-
pletely capture how it might act during a given bushfire scenario. Within
these behaviours there is enough variation to ensure that over a population
of interacting agents the simulation will not be deterministic, and instead
will approximate the complex and unpredictable nature of a real bushfire.
< person id = " 0 " >
< a t t r i b u t e s >
< a t t r i b u t e name = " B D I A g e n t T y p e " class = " java . lang
. String " > io . github . a g e n t s o z . ees . agents .
b u s h f i r e . R e s i d e n t </ a t t r i b u t e >
< a t t r i b u t e name = " H a s x s A t L o c a t i o n " class = " java .
lang . String " >
7 6 6 7 2 2 . 4 8 0 9 7 7 6 0 7 , 5 7 5 4 4 5 5 . 4 8 8 8 5 1 7 5 </ a t t r i b u t e
>
< a t t r i b u t e name = " I n i t i a l R e s p o n s e T h r e s h o l d "
class = " java . lang . Double " > 0.1 </ a t t r i b u t e >
< a t t r i b u t e name = " F i n a l R e s p o n s e T h r e s h o l d "
class = " java . lang . Double " > 0.1 </ a t t r i b u t e >
< a t t r i b u t e name = "
W i l l G o H o m e A f t e r V i s i t i n g D e p e n d a n t s " class = "
java . lang . Boolean " > true </ a t t r i b u t e >
< a t t r i b u t e name = " W i l l G o H o m e B e f o r e L e a v i n g " class
= " java . lang . Boolean " > false </ a t t r i b u t e >
< a t t r i b u t e name = " E v a c L o c a t i o n P r e f e r e n c e " class =
" java . lang . String " >
, 7 9 1 1 1 8 . 1 0 2 5 6 9 2 9 3 , 5 7 5 3 4 6 1 . 0 9 8 7 6 6 3 5 </
a t t r i b u t e >
< a t t r i b u t e name = " I n v a c L o c a t i o n P r e f e r e n c e " class
= " java . lang . String " >
, 7 6 6 7 2 8 . 6 1 8 1 9 5 6 1 7 , 5 7 5 4 4 4 9 . 1 1 6 6 6 8 2 3 </
a t t r i b u t e >
</ a t t r i b u t e s >
...
</ person >
Figure 9: An agent plan with attributes appended.
43
4 Applications
We will now apply the population generation model to the Surf Coast Shire
region, an area that has been affected by a number of severe fire incidents
(Black Saturday, Ash Wednesday) and contains a community that is regu-
larly exposed to the bushfire threat. We construct a population for a typ-
ical summer weekday in the region, analysing the algorithm's performance
and examining the resultant simulation runs to verify that the outputted
agent plans adequately reflect the input data. The specific inputs for the
scenario have been co-developed with domain experts from the Surf Coast
Shire Council (SCSC) and the Department of Environment, Land, Water
and Planning (DELWP).
4.1 Algorithm Inputs in Surf Coast Shire
In this section we describe the inputs that the population generation al-
gorithm takes and justify them within the context of the Surf Coast shire
region. The chosen inputs are designed to easily translate across a large
number of potential scenarios, and for this reason the number of parameters
is kept to a minimum while at the same time giving sufficient flexibility in
the kinds of populations that can be generated. Another motivation for this
is to provide a user-friendly input process using data and expert opinion
most likely to be available to emergency personnel.
4.1.1 Subgroups in Surf Coast Shire
The fire season in Victoria falls between November-April and coincides with
the busiest tourism period on the Surf Coast, and this also brings a large
transient and seasonal population of workers, semi-permanent residents and
holiday home owners. Small townships along the Great Ocean Road can
have up to 6-7 times the resident population, with permanent residents far
outnumbered by tourists and other short-term visitors. This gives a basis
upon which to split the population into subgroups, with visitors not only
having very different reasons and priorities in forming their daily plans, but
also a reduced knowledge of the area and routes between locations.
The five identified subgroups for the region are:
• Resident (R): those who live in the region, have local knowledge of
roads and places of congregation. They are connected to the commu-
nity, are likely to have a concern for others, pets, and property, and
are most likely to defend property.
44
• ResidentPartTime (RP ): those who own property in the region, but
may only live there for several months of the year. Are somewhat
familiar with the area, but do not have a large community network.
They are less likely to be prepared for a bushfire threat and less likely
to defend property.
• VisitorRegular (V R): those who have visited the region on several
occasions and may have a holiday home there. Will know the area
somewhat, but are unlikely to have pets or relatives. Very unlikely to
stay and defend property.
• VisitorOvernight(V O): those who are unfamiliar with the area and
living in short term accommodation. They will not defend property
but are likely to gather belongings and then follow instructions or leave
the region.
• VisitorDaytime (V D): those who are mostly unfamiliar with local
roads and places of congregation. They will either ignore a fire alert
or leave region.
Thus we set
S = {R, RP, V R, V O, V D}
and require as a first input the number of required agents for each s ∈ S:
PR = 10000
PRP = 5000
PV R = 5000
PV O = 15000
PV D = 15000
giving a total population of P = 50000 in the region.
4.1.2 Activities in Surf Coast Shire
We now consider the most likely or significant types of activities that occur
during summer in the region, taking into account both where people might
congregate and which trips are important in shaping a person's daily plan.
45
Whilst the actual nature and location of trips made by individuals of dif-
ferent subgroups may differ significantly, it is useful to maintain a broad
description of activities here and add granularity via the location maps for
each subgroup as this makes the activity distributions easier to construct
and understand.
The five activities that broadly capture the types of trips that an agent
(of any subgroup) will make are:
• α1 = home: as required by the algorithm; home locations will range
from: residential addresses for Resident, ResidentPartTime and a
percentage of VisitorRegular; to hotels, caravan parks and other
short-term accommodation options for VisitorOvernight and the re-
maining VisitorRegular; to specific 'source' locations outside of the
region for VisitorDaytime. Duration is set to 2 hours for all sub-
groups. Note that duration here
• α2 = work: the work activity will only apply to Resident and
ResidentPartTime, and will occur at specified place-of-business loca-
tions. Duration is set to 4 hours for all subgroups, which allows for
people having breaks and/or working multiple jobs. This activity re-
quires the most consideration when it is being distributed throughout
the day, with peak work hours occurring during a typical 9am-5pm
window.
• α3 = shop: similarly, these will be designated in places specified as
shopping districts. The focus will be on larger shop areas, where people
may tend to gather at times in the day. These locations are also more
relevant from an emergency perspective as many of these shopping
centres will be places of congregation for invacuation. Duration is set
to 2 hours for all subgroups.
• α4 = beach: this is a major activity during summer and will be par-
ticularly popular with all of the visitor subgroups. Here the peak
times in the distribution should be quite clear based on the forecast
and could potentially consider tidal information on a per-day basis.
Resident agents may have patterns for the beach activity that avoid
peak tourist times. Duration is set to 2 hours for all subgroups.
• α5 = other: will serve as a catch-all for other less significant activities
that occur during the day.
It will allow for more diversity in plans
between subgroups, and also ensure that there are not too many agents
46
evacuating from the same spots in an emergency. Should usually be
maintained as a minor but constant presence at all time steps to allow
for this diversity. Duration is set to 2 hours for all subgroups.
We therefore define the set of activities
A = {α1, α2, α3, α4, α5}.
A is mapped per subgroup to the set of locations L, which are provided as
a ESRI shapefile (.shp) by SCSC. Also included in this file are the locality
and allocation number for each location. Apart from the distinction in
home locations described above, some other interesting differences between
subgroup activity mappings include only mapping more popular beaches to
the visitor groups, assigning schools to other for the Resident subgroup
and assigning landmarks and national parks to other for VisitorDaytime
and VisitorOvernight.
4.1.3 Distributions in Surf Coast Shire
We now take the set of activities A and formulate distributions for each
subgroup. In keeping with our choices for durations, the time-step size is
set to 2 hour blocks, meaning that there are 12 'bins' that need to be es-
tablished for each subgroup. Note that the durations listed for each activity
do not necessarily reflect their typical duration, but indicate the approxi-
mate time interval at which an agent will re-evaluate its activity. It is quite
possible that a Resident agent could spend the entire day performing the
home activity if during activity selection home is chosen for every time-step.
Figures 10 and 11 show some examples of activity distributions on a typical
summer day for Resident and VisitorDaytime:
Figure 10 shows that for a Resident, the major working time is be-
tween 8 a.m. and 6 p.m. in keeping with a regular weekday. In both cases,
shopping activity increases in the afternoon and continues until dinnertime,
when people (particularly visitors) are likely to be at restaurants and bars.
Note that these distributions have not been formally validated by SCSC, and
rather represent a first attempt to capture the dynamics of typical weekday
activities in the region. These serve for demonstration purposes, but activ-
ity distributions will ultimately be derived from actual daily trip data.
47
Resident
e
g
a
t
n
e
c
r
e
P
100
50
0
1
3
5
7
9
11 13 15 17 19 21 23
Time of day
home work beach shops other
Figure 10: Activity distribution for Resident subgroup on a typical summer
day.
4.1.4 Locations in Surf Coast Shire
Locations are supplied by SCSC and exist in a number of different forms.
At the most basic level, a small number of source nodes throughout the
region serve to distribute the population, with a limited set of polygons
describing areas of work, beaches etc. This approach is crude but useful
for quickly describing patterns of movement. However, to create accurate
day plans that vary across the population (and thus give realistic traffic
movement) a greater level of detail is required. Per-address data allows a
population to be precisely mapped at an appropriate density to residential
areas and tourist spots. It also allows the possibility of specific, niche trips
(e.g. to a golf club, national park or fishing pier) that further diversify the
population's movements or capture known patterns that may be of concern
in an emergency. Figure 12 illustrates why realistic address data can have
a significant impact on resultant traffic flow.
The trade-off is that the increased detail increases algorithm runtime.
The locality-based approach (see Section 3.3.1 mitigates this somewhat, and
also allows us to frame our subgroups in terms of how likely they are to
move through the region. The final input we need to form agent plans is a
travel factor measure for each subgroup. We would expect that visitors are
48
VisitorDaytime
e
g
a
t
n
e
c
r
e
P
100
50
0
1
3
5
7
9
11 13 15 17 19 21 23
Time of day
home work beach shops other
Figure 11: Distribution table for VisitorDaytime subgroup on a typical
summer day.
more likely to move between localities than residents, and so we assign the
following provisional probabilities:
gR = 0.2
gRP = 0.3
gV R = 0.4
gV O = 0.6
gV D = 0.8
4.2 Results of the Algorithm
We now run the algorithm with the above inputs using the programming
language R to generate a viable MATSim population.xml file. For testing
we used a 2013 Macbook Air with 2 cores and 8GB of RAM. With 50, 000
agents, the process takes 25.14 minutes, but this duration is largely due to
the fine granularity of the full locations .shp file (∼ 75, 000 address points).
If we reduce that down to ∼ 40, 000 addresses, which is still very fine-grain,
we get a runtime of 9.30 minutes. It is therefore important to balance a re-
49
Figure 12: Per-address location data on the Surf Coast Shire. The location
points (pink) cluster around larger coastal communities, with hardly any
addresses in the national park area (green).
alistic representation of locations in the region with restricting the location
options down to match the number required by the given population (note
that many visitor agents will be mapped to the same home nodes; in fact we
may have 10,000-15,000 all originating from the same out-of-region source
node, so we generally require less nodes than agents). We proceed with the
full locations data as the most general case.
4.2.1 Plan Times
To verify that the generated plans reflect the required population dynam-
ics, we compare the input distributions, which detail the expected number
of agents performing each activity during each time bin, to the generated
number of agents completing those activities in the population.xml file.
In Table 2 and Table 3, we see for the two resident subgroups that the
cumulative effect of work having a longer duration than other activities
results in a ∼ 5% discrepancy towards the end of the peak working period.
In the other subgroups (Table 4-6) where work is not a valid activity, we see
50
Table 2: Percentage variation from input distribution at each time step for
Resident subgroups.
Resident
Activity
home
work
beach
shops
other
Activity
home
work
beach
shops
other
8am
12am 2am 4am 6am
10am
-2am -4am -6am -8am -10am -12pm
0.03
0.09
0.04
-0.41
-0.18
-0.15
0.2
0.19
-0.12
-0.08
-0.13
0.14
0.36
-0.25
0.16
-0.41
0.13
0.05
-0.3
0.19
-0.18
0.25
0.38
0
0
0
0
0
0
0
8pm
12pm 2pm 4pm 6pm
10pm
-2pm -4pm -6pm -8pm -10pm -12am
-0.2
-0.09
-0.39
0.19
0.08
0.52
-0.01
-0.03
-0.49
0.79
-0.26
-0.01
0.35
-0.2
-0.2
0.12
-0.07
-3.83
5.04
-1.07
-0.14
-0.57
0.19
-0.01
0.39
-0.1
0
0
0
0
Table 3: Percentage variation from input distribution at each time step for
the ResidentPartTime subgroup.
ResidentPartTime
Activity
home
work
beach
shops
other
Activity
home
work
beach
shops
other
12am 2am 4am 6am
10am
-2am -4am -6am -8am -10am -12pm
-0.12
8am
0.1
-0.08
0.18
0.04
-0.24
0.48
-0.02
0.24
-0.3
-0.4
-0.5
-0.2
-0.24
-0.02
0.96
0
0
0
0.12
0
0
0
0
0
0.04
-0.16
0
0
0.12
8pm
12pm 2pm 4pm 6pm
10pm
-2pm -4pm -6pm -8pm -10pm -12am
-0.2
-0.92
0.24
-1.2
1.18
-0.02
-2.06
4.96
0.28
-1.42
-1.76
-0.34
0.3
0.06
0.58
-0.6
0
0
0.9
0.02
-0.64
0.02
-0.24
-0.2
0.42
0.02
0.56
0.06
0
0
51
that at all time-steps the generated plans match the input distributions to
a maximum error of ∼ 1%. The difference in outcome for the two activity
distributions for Resident and VisitorDaytime, first introduced in Figures
10 and 11, is emphasised visually in Figure 13.
Table 4: Percentage variation from input distribution at each time step for
VisitorRegular subgroup.
VisitorRegular
Activity
home
beach
shops
other
Activity
home
beach
shops
other
12am 2am 4am 6am
10am
-2am -4am -6am -8am -10am -12pm
0.22
8am
0.16
0.26
0
0
0
0
0
0
-0.22
-0.16
-0.26
0.12
-0.5
0.38
0
-1.2
0.46
0.1
0.64
-0.76
0.26
-0.16
0.66
-0.5
-0.32
0.6
0.22
8pm
12pm 2pm 4pm 6pm
10pm
-2pm -4pm -6pm -8pm -10pm -12am
-0.34
0.74
0.94
-0.6
0
0.2
0.08
0.12
-0.4
0.3
0.02
-0.2
-0.12
-0.38
-0.36
0.24
0.48
-0.52
-0.2
0
52
Table 5: Percentage variation from input distribution at each time step for
the VisitorOvernight subgroup.
VisitorOvernight
Activity
home
beach
shops
other
Activity
home
beach
shops
other
12am 2am 4am 6am
10am
-2am -4am -6am -8am -10am -12pm
0.11
-0.04
8am
0
0
0
-0.11
0.04
0
0
-0.41
0.37
0
0.03
0.09
0.11
-0.34
0.14
0.23
0.13
-0.48
0.12
-0.21
-0.01
0.21
12pm 2pm 4pm 6pm
10pm
-2pm -4pm -6pm -8pm -10pm -12am
8pm
0
-0.21
0.15
0.07
0.03
0.23
0
-0.26
-0.03
-0.25
0.09
0.19
0.03
0.31
-0.09
-0.25
-0.13
0.05
0.19
-0.11
0
0
0
0
Table 6: Percentage variation from input distribution at each time step for
the VisitorDaytime subgroup.
VisitorDaytime
Activity
home
beach
shops
other
Activity
home
beach
shops
other
8am
12am 2am 4am 6am
10am
-2am -4am -6am -8am -10am -12pm
-0.04
0.53
-0.27
-0.21
-0.19
0.22
0.32
-0.35
0.38
-0.11
-0.36
0.09
-0.05
-0.03
0
0
0
0
0
0
0
0
0.08
0
8pm
12pm 2pm 4pm 6pm
10pm
-2pm -4pm -6pm -8pm -10pm -12am
0.07
-0.32
0.19
0.07
0.25
-0.26
0.43
-0.41
-0.23
-0.08
0.23
0.07
-0.63
-0.35
0.54
0.45
0.03
0.09
0.12
-0.25
0
0
0
0
53
Resident activity error (output-input)
11
33
55
77
99
1111 1313 1515 1717 1919 2121 2323
Time of day
home work beach shops other
VisitorDaytime activity error (output-input)
e
g
a
t
n
e
c
r
e
P
e
g
a
t
n
e
c
r
e
P
5
0
−5
1
0
−1
11
33
55
77
99
1111 1313 1515 1717 1919 2121 2323
Time of day
home work beach shops other
Figure 13: Percentage error in both the Resident and VisitorDaytime
subgroup. The former has a very obvious error outlier. The latter has a
maximum error < 1%, with the error spread evenly throughout the 'busy'
part of the day.
54
4.2.2 Plan Locations
We must also analyse a) how successfully the algorithm distributes activ-
ities, and b) if the trips between activities are realistic enough to provide
an appropriate traffic background. To do this, we run the plan in MAT-
Sim and observe the output events data. Figures 14-18 provide a snapshot
run-through of how the morning plays out in this scenario for a particu-
lar community in the region (Torquay). In these maps, location nodes are
coloured if they are being used by an agent for an activity. The colour
mapping is
home −→ blue
work −→ green
beach −→ yellow
shops −→ orange
other −→ white
To verify that the background traffic is working as planned, we view
snapshots of traffic flow in Figures 19 and 20. The colour mapping here
is the same as previously, but now the triangles represent moving vehicles
on the road that are travelling to a location to complete the activity corre-
sponding to their colour.
55
Figure 14: Agent activity locations at midnight. The vast majority of the
population are at home (blue), which is not surprising just after initialisation
(recall that every agent begins at home). We can see one or two agents
have already started performing other activities (white) and one worker
has already begun their shift (green). We can note that the algorithm has
successfully allocated a concentration of people to individual address points
in residential areas, with CBD areas less populated during off-peak times.
56
Figure 15: Activity locations at 3am. More other activities have begun,
and a few more agents are now at work. We also see the introduction of
a few early morning shops activities (orange). In this visualisation, glow-
ing or bright points indicate that activity is starting at that location, with
continuing activities maintaining their colour until they are terminated.
57
Figure 16: Activity locations at 6am. There are a few more people at
work, but the most significant change is that dawn has brought out a few
beachgoers (yellow). Notice how both work and beach activities are con-
centrated within one or two areas. The input location data used here tends
to focus on a few busy or popular areas as centres for particular locations.
58
Figure 17: At 9am the activity locations have reached a point where all 5
activities are well represented in the snapshot. We note that there is now a
significant number of agents performing work, coinciding with the beginning
of the peak work times for a typical summer weekday. Beach activitiy is also
peaking here, with many activities occurring along the coastal strip.
59
Figure 18: By midday we have less people beginning work, but there is
still a steady set of working agents. Trips to shops have increased signifi-
cantly, and are almost the most popular activity around lunch time. Many
of those people who began work during the first peak work starting time
would now be finished, and may have decided to go to the shops after work.
The percentage of people at home has now very noticeably dropped from
earlier in the morning, indicating that many agents are out and travelling
to/undertaking other activities.
60
Figure 19: Here we have an overview of the entire region at 9am. We can see
that a significant flow of traffic has already begun to come in from the East
(from Melbourne), with smaller contributions also coming from the west. As
we would expect, the most congested road in the region is along the coast
(the Great Ocean Rd). Most of the work traffic seems to be focussed in the
area we observed already in Torquay.
61
Figure 20: Zooming in on commuters between Anglesea and Torquay, we can
see in greater detail how congested the coastal routes can get. Congestion
within Anglesea seems comparatively light, and it seems the majority of
travel is through Anglesea rather than to or from it.
62
4.3 Adding BDI Attributes
We now append the attributes that will determine how each agent will act
once their plans change due to a bushfire threat in the SCS region. This
involves first establishing sets E and T of Environmental and Transmitted
alerts, defining a ranking r on E × T , and determining the parameters for
each subgroup that will assign attributes.
4.3.1 Alerts in the Surf Coast Shire
There are any number of events that might trigger a person to evacuate
during a bushfire, but here we focus only on two environmental cues, smoke
and fire, and three core message types that may be issued by emergency
services: 'Advice', 'Watch and Act', and 'Evacuate Now'. These warnings
can be ordered in terms of severity, and we assign both types of alert values
in [0, 1] as follows:
Table 7: Barometer scores for each alert.
Environmental Transmitted Value
0.1
0.2
0.3
0.4
Watch and Act
Evacuate Now
Advice
-
-
-
Smoke
Fire
We then define the total barometer score to be
r(e, τ ) = e + τ
e.g. an agent who has both seen smoke and received a 'Watch and Act'
message will have a total barometer score of 0.5.
There are a number of assumptions that we make with these messages.
Firstly, it is assumed that all messages are able to be sent to specific areas
of the region. We then also assume that any agent who is in that area when
the message is sent will see the message. The same holds for environmental
alerts. We assume that people only see and respond to the visual cues when
they are within a 5km of the fire front (for smoke) or 1km (for fire). Smoke
is particularly hard to quantify in this sense, because in a bushfire it is very
likely that people will be able to see smoke from further than 5km away.
But seeing smoke from afar is also not as likely to invoke a response. We
may conceptualise the smoke alert here as the point when smoke begins
63
to interfere with people's line of sight. Indeed, it may be useful in future
work to have an additional, lesser, smoke alert which represents this smoke
awareness.
4.3.2 Attribute Parameters
The only other inputs that the BDI model requires are the per-subgroup
parameters which determine their attributes. These are:
• prob of dependants: the probability that an s − agent has a depen-
dant.
• stays: true if the strict inequality INIT<ACT is possible i.e. deter-
mines whether an s-agent can stay and defend.
• prob of go homes: the probability that an s-agent will return home
before leaving.
• [threshold mins,threshold maxs]: the interval from which the two
threshold score INIT and ACT are drawn from.
The values that we use for these variables here are given in Table 8.
Table 8: Attribute parameters for each subgroup.
Subgroup
prob of -
dependant
0.3
Resident
ResidentPartTime 0.05
0
VisitorRegular
VisitorOvernight 0
0
VisitorDaytime
Subgroup
threshold -
min
0.1
Resident
ResidentPartTime 0.1
0.1
VisitorRegular
VisitorOvernight 0.2
0.3
VisitorDaytime
stay
1
1
1
0
0
prob of -
go home
0.5
0.4
0.4
0.8
0
threshold -
max
0.8
0.6
0.4
0.4
0.7
With these inputs, we can now add the BDI attributes. Using a separate
R script, we assign Boolean values to each agents' plan that reflect whether
64
Figure 21: An agent plan with attributes appended.
< person id = " 11154 " >
< a t t r i b u t e s >
< a t t r i b u t e name = " B D I A g e n t T y p e " class = " java . lang .
String " > io . github . a g e n t s o z . ees . agents .
b u s h f i r e . R e s i d e n t P a r t T i m e </ a t t r i b u t e >
< a t t r i b u t e name = " H a s D e p e n d a n t s A t L o c a t i o n " class = "
java . lang . String " > </ a t t r i b u t e >
< a t t r i b u t e name = " I n i t i a l R e s p o n s e T h r e s h o l d " class =
" java . lang . Double " > 0.5 </ a t t r i b u t e >
< a t t r i b u t e name = " F i n a l R e s p o n s e T h r e s h o l d "
class =
" java . lang . Double " > 0.6 </ a t t r i b u t e >
< a t t r i b u t e name = "
W i l l G o H o m e A f t e r V i s i t i n g D e p e n d a n t s " class = " java
. lang . Boolean " > false </ a t t r i b u t e >
< a t t r i b u t e name = " W i l l G o H o m e B e f o r e L e a v i n g " class = "
java . lang . Boolean " > false </ a t t r i b u t e >
< a t t r i b u t e name = " E v a c L o c a t i o n P r e f e r e n c e " class = "
java . lang . String " > Lorne
3 4 , 7 5 9 2 7 1 . 0 9 3 7 1 8 8 3 , 5 7 2 9 8 5 1 . 2 9 6 7 0 2 9 </ a t t r i b u t e >
< a t t r i b u t e name = " I n v a c L o c a t i o n P r e f e r e n c e " class = "
java . lang . String " >
, 7 9 4 0 2 0 . 8 9 3 9 8 4 8 0 2 , 5 7 7 1 8 4 5 . 0 8 2 2 4 7 8 4 </ a t t r i b u t e >
</ a t t r i b u t e s >
< plan s e l e c t e d = " yes " >
< a c t i v i t y type = " home " x = " 7 9 4 3 6 3 . 9 7 6 3 4 9 2 1 5 " y = "
5 7 7 1 3 6 7 . 1 8 2 7 4 8 8 8 " e n d _ t i m e = " 0 8 : 0 4 : 0 0 " / >
< leg mode = " car " / >
< a c t i v i t y type = " shops " x = " 7 9 0 5 2 4 . 0 4 7 7 8 6 2 9 8 " y = "
5 7 5 2 2 1 9 . 8 3 8 2 8 9 7 2 " e n d _ t i m e = " 1 0 : 4 4 : 0 0 " / >
< leg mode = " car " / >
< a c t i v i t y type = " other " x = " 7 8 9 8 1 6 . 4 2 1 6 7 4 4 4 7 " y = "
5 7 5 3 2 1 3 . 3 3 5 9 2 4 0 3 " e n d _ t i m e = " 1 2 : 4 9 : 0 0 " / >
< leg mode = " car " / >
< a c t i v i t y type = " work " x = " 7 8 9 7 7 8 . 1 4 4 7 5 4 0 2 2 " y = "
5 7 5 2 6 0 8 . 5 9 3 2 4 6 9 " e n d _ t i m e = " 1 6 : 0 1 : 0 0 " / >
< leg mode = " car " / >
< a c t i v i t y type = " shops " x = " 7 7 1 1 1 7 . 5 1 1 4 2 9 0 5 6 " y = "
5 7 3 8 6 8 0 . 5 7 7 3 0 3 4 6 " e n d _ t i m e = " 1 9 : 0 6 : 0 0 " / >
< leg mode = " car " / >
< a c t i v i t y type = " home " x = " 7 9 4 3 6 3 . 9 7 6 3 4 9 2 1 5 " y = "
5 7 7 1 3 6 7 . 1 8 2 7 4 8 8 8 " / >
</ plan >
</ person >
65
it has a dependant and whether it will go home (note that this decision
depends on whether the agent has a dependant). We also determine a INIT
and ACT tolerance and choose a location for the dependant (if applicable), as
well as evacuation and invacuation preference locations. These last two are
determined by a separate Refuges.shp file provided by SCSC. The resulting
attributes appear at the start of the agent's MATSim plan, as shown in
Figure 21.
Note that the blank HasdependantsAtLocation field occurs whenever
the agent does not have a dependant. The InitialResponseThreshold
and FinalResponseThreshold fields represent INIT and ACT respectively.
WillGoHomeAfterVisitingDependants and WillGoHomeBeforeLeaving en-
compass two variations of the same behaviour, and at most one of them will
be true -- this is determined via the prob of go home parameter.
4.4 SCS Application
To fully demonstrate the usefulness of the algorithm, we apply the plans
with attributes in a bushfire evacuation simulation. Using the EES simu-
lation model, this will show both the importance of having the population
realistically situated and engaged in daily plans, and the ability for BDI
attributes to directly impact an agent's response to a bushfire threat. In the
following scenario, the response of a community in Anglesea is tested under
two different scenarios. Each scenario has the same generated plan and the
same bushfire threat that ignites at midday and eventually encroaches into
parts of the town. In one test, the agents are left to respond to the fire on
their own (with an eventual 'Evacuate Now' warning at 3pm). In the other,
a sequence of messages of increasing severity are delivered to the population,
directing them to act earlier. We will proceed again with a series of snap-
shots. This time, vehicles are coloured on a spectrum based on their speed
in the network, where green represents free-speed and red represents heavy
congestion. The previous activity location node colours are also present in
the background. We proceed via Figures 22-25.
66
Figure 22: At 12pm when the fire ignites, the state is exactly the same in
each scenario. We observe the familiar flow of traffic through Anglesea, and
that there are still many people at home. As was the case in Torquay, midday
sees a high level of shops and beach activity around the coastal strip.
Figure 23: By 2pm, we see the two scenarios have diverged significantly.
On the left, traffic has increased in the middle of Anglesea proper, but the
traffic flow in and out has not changed. The purple activity location nodes
indicate that some have decided to take action, but a lot of people are still
at home oblivious to or unconcerned by the threat. On the right side, we
see that congestion increases as people attempt to evacuate from the danger
zone. Three messages have already been issued, with the severity raised at
1.30pm. There are hardly any people left in their homes, except those that
have returned there to prepare and/or defend.
67
Figure 24: By 4pm, we begin to see the fire encroach upon the outskirts
of Anglesea. The smoke front will now have alerted most people that they
need to act even if the messages did not. The community on the left has by
now received an 'Evacuate Now' message, and we see that they are roughly
at the level of congestion stage that the other community was two hours
ago. In the scenario on the right, the congestion has eased somewhat by
now towards the south of Anglesea, and pretty much everyone, bar a few
defenders, has left or is trying to leave.
Figure 25: The fire is now engulfing the edge of the town. By 6pm, everyone
in the community on the right who wanted to escape has done so, and there
is no congestion on the roads leaving Anglesea. On the left, however, the
situation is far more urgent, with lots of people still stuck on the road as
the fire approaches.
68
Whether or not the stark contrast between outcomes in the previous
EES example totally reflects the effectiveness that emergency warnings can
have, we did see the capability of agents to take in information from various
sources and react to a threat accordingly. Further, we saw variation in
behaviour, even within each scenario, where a fair number of local agents
decided to stay and defend, at least until it became clear that the danger was
too great. The movement of people prior to the evacuation situation was
also important, as it meant that the main arterial roads out of Anglesea were
already busy. This affected the ability of people to evacuate, particularly
on the left scenario. Thus this application provides some useful insights
into what benefits our population generation model with appended BDI
attributes can provide.
69
4.5 Discussion
The error tables (Tables 2 and 6) reveal that variation in activity durations
can affect the ability of the algorithm to reproduce the input distributions.
Whilst the algorithm takes into account the durations of each activity and
adjusts the probability matrix accordingly, it still relies on the input distri-
butions giving a 'feasible path' for agents to complete activities according to
their durations. If the distribution of people working drops off too dramati-
cally from one time step to the next, the algorithm may not be able to shift
enough agents away from the work activity. This is because the distributions
are essentially decoupled from activity durations. One solution is to smooth
the changes between time-steps in the distribution. However, if the distri-
bution data does accurately reflect the whereabouts of the population, then
modifying them essentially results in a discrepancy anyway, so this process
should be minimised. The other alternative is to make small adjustments
to the durations of each activity, in conjunction with smaller time-steps (i.e.
more detailed activity distributions). This requires a more complicated set
of distribution inputs, and these may be difficult or impossible to accurately
obtain. We therefore must also weigh the need for the output to match the
distributions exactly against the accuracy of the distributions themselves
and the appropriate duration of each activity. If there was no level of du-
ration control, then total correspondence with the distributions would be
attained (agents are always where they should be) but with the drawback
that there will be more movement between activities (because each agent
can change activity at every time-step). This may result in background traf-
fic that does not realistically represent the region. To strike an appropriate
balance between these needs, a maximum error of ∼ 5% was decided upon,
which we observe in the tables.
One potential improvement to the population generation process would
be to apply MATsim's iterative re-planning capabilities to the generated
population.xml file. The agent plans will then converge to a traffic equi-
librium, ensuring that the daily plans we apply prior to the bushfire threat
are collectively feasible and optimised at the individual level. However, this
process could serve to compound the error that we observe above, as the
iterative process would likely diverge away from the idealised expected ver-
sion of events presented by the input distributions. Similarly, attempting to
calibrate the model outputs to actual traffic counts would add to the tension
that exists between matching the static inputs to a dynamic output process
that is designed to replicate an even more complex reality. This calibra-
tion approach is planned for future work, and will likely see the algorithm
70
modified to factor in these traffic counts as part of the generation process.
The alert ranking system provides a way of ordering alerts that is easy
to understand, but is somewhat arbitrary in the way it orders the possible
combinations of environmental and transmitted alerts. It results in some
equivalences that may not exist, e.g.
(Smoke, Advice) ≡ (Fire, 0)
A more thorough ranking system would explicitly order every possible
combination of environmental and transmitted alert. With the number of
alerts we consider during this demonstration this would be a simple task,
but in practice there will be other alerts of both types that may or may
not be used in a given scenario. This current system may be viewed as
a useful temporary measure to allow quick interchanging or introduction
of new alerts without necessarily needing to place every new combination
into some order. Once the full set of alerts is established, a more concrete
ranking system can be applied. Some other potential environmental alerts
are observing embers, observing neighbours leaving and aural cues. There
are also a number possible of transmitted messages that are possible, often
as variations of the three included here.
5 Conclusion
Bushfires threaten communities every year, and due to their sometimes rapid
onset and unpredictable nature it is imperative that at-risk regions prepare
themselves to respond. Planning that accounts for nuances in a specific
population and region can help improve resilience and preparedness of both
the authorities and the community in the face of a threat. In this research
we have presented a method for capturing these dynamics and measuring
their significance through bushfire evacuation simulation.
Our contribution is an algorithm that takes a set of activity profiles and
generates agent daily plans in the form of a MATSim population.xml file.
The algorithm was designed to generate synthetic populations for regions
that are not easily approximated by incremental sampling techniques due
to an actual population that has a significant transient element. Further,
we have implemented this algorithm as a tool to generate populations for
emergency evacuation simulations. This requires a set of flexible inputs
that can be modified to suit a particular scenario and that are be easily
understandable to domain experts.
71
We have also utilised the BDI cognitive framework to construct a bush-
fire behaviour model that allows our generated populations to reason and
respond to a bushfire threat. Using threshold parameters, we have intro-
duced a method to vary these responses across our populations in a manner
that mimics the way that humans process new information. Input attributes
of this bushfire behaviour model have been integrated into the population
generation algorithm to provide a single tool that can create synthetic pop-
ulations for a wide range of different bushfire scenarios.
Lastly, we have applied this tool to Surf Coast Shire in Victoria and
tested it in a number of scenarios. The algorithm has contributed to the
wider project by bringing a dynamic element to the notion of evacuation
messaging, and the behaviour model will allow emergency personnel to test
different approaches to bushfire response when the population has a varying
level of engagement. Ultimately it is hoped that the algorithm and associ-
ated model presented here will improve understanding of the risks associated
with bushfire evacuation.
5.1 Future Work
This research is part of a larger project looking at improving the evacuation
planning process for at-risk communities. This thesis presents work to date
in the wider context of this project, with the ultimate aim being to develop
a model that gives a realistic representation of any population, both in the
form of their daily plans and their eventual bushfire responses. Most of the
input data used in Section 4 has only been informally validated and does not
necessarily reflect the actual make-up of the population in the Surf Coast
Shire. In this sense, much future work will be devoted to ensuring that the
model is validated and calibrated. For the algorithm, this will mean using
traffic counts and travel times to match observed data with the movements
of the synthetic population. For the behavioural model, we plan to continue
to work with behavioural experts, especially in the area of behaviour profiles
and archetypes (Strahan et al., 2018). On particular area of interest is the
introduction of social networks as another means for agents to become aware
of the threat level. This could be implemented as a third alert type in the
threshold model, and allow a bigger set of alert combinations to diversify
the kinds of possible responses further.
In terms of application, the project aims to encompass the whole of Vic-
toria. The required inputs therefore must be easily obtainable and general
enough to apply to each region, whilst still maintaining the ability to distin-
guish communities by their demographic make-up. This broader approach
72
will also necessitate that the model allows for interaction between regions,
and how evacuation movements in one place can have a flow-on effect in
another.
References
Adam, C., Danet, G., Thangarajah, J., & Dugdale, J. (2016). BDI mod-
elling and simulation of human behaviours in bushfires. In Information
Systems for Crisis Response and Management in Mediterranean Coun-
tries. Madrid: Springer Germany.
7
Adam, C., & Gaudou, B. (2017). Modelling Human Behaviours in Disasters
from Interviews: Application to Melbourne Bushfires. Journal of Artificial
Societies and Social Simulation, 20 (3). doi: 10.18564/jasss.3395
17
Anderson, J. E. (2011). The Gravity Model. Annual Review of Economics,
3 (1), 133-160. doi: 10.1146/annurev-economics-111809-125114
34
Anderson, J. R. (1996). ACT: A simple theory of complex cognition. Amer-
ican Psychologist, 51 (4), 355-365. doi: 10.1037/0003-066X.51.4.355
12
Antonovics, J., Iwasa, Y., & Hassell, M. P. (1995, May). A Generalized
Model of Parasitoid, Venereal, and Vector-Based Transmission Processes.
The American Naturalist, 145 (5), 661-675. doi: 10.1086/285761
10
Auchincloss, A. H., & Garcia, L. M. T.
(2015, November). Brief intro-
ductory guide to agent-based modeling and an illustration from urban
health research. Cadernos de saude publica, 31 (Suppl 1), 65-78. doi:
10.1590/0102-311X00051615
14
Axelrod, R.
(1997). Advancing the Art of Simulation in the Social Sci-
ences. In R. Conte, R. Hegselmann, & P. Terna (Eds.), Simulating Social
Phenomena (p. 21-40). Springer Berlin Heidelberg.
13
73
Ballas, D., Clarke, G., Dorling, D., Eyre, H., Thomas, B., & Rossiter,
D.
(2005, January). SimBritain: A spatial microsimulation approach
to population dynamics. Population, Space and Place, 11 (1), 13-34. doi:
10.1002/psp.351
15
Barrett, B., Ran, B., & Pillai, R.
(2000, January). Developing a Dy-
namic Traffic Management Modeling Framework for Hurricane Evacu-
ation. Transportation Research Record: Journal of the Transportation
Research Board , 1733 , 115-121. doi: 10.3141/1733-15
17
Bazzan, A. L. C., Wahle, J., & Klugl, F. (1999). Agents in Traffic Modelling
- from Reactive to Social Behaviour.
15
Bellifemine, F., Poggi, A., & Rimassa, G. (2001). Developing multi-agent
systems with a FIPA-compliant agent framework. Software: Practice and
Experience, 31 (2), 103-128. doi: 10.1002/1097-024X(200102)31:2(cid:104)103::
AID-SPE358(cid:105)3.0.CO;2-O
12
Beloglazov, A., Almashor, M., Abebe, E., Richter, J., & Steer, K. C. B.
(2016, January). Simulation of wildfire evacuation with dynamic factors
and model composition. Simulation Modelling Practice and Theory, 60 ,
144-159. doi: 10.1016/j.simpat.2015.10.002
11
Birkin, M., & Clarke, M.
(1988, December). Synthesis -- A Synthetic
Spatial Information System for Urban and Regional Analysis: Methods
and Examples. Environment and Planning A, 20 (12), 1645-1671. doi:
10.1068/a201645
15
Bonabeau, E. (2002, May). Agent-based modeling: Methods and techniques
for simulating human systems. Proceedings of the National Academy of
Sciences of the United States of America, 99 (Suppl 3), 7280-7287. doi:
10.1073/pnas.082080899
4, 14
Bordini, R. H., Hubner, J. F., & Vieira, R. (2005). Jason and the Golden
Fleece of Agent-Oriented Programming. In R. H. Bordini, M. Dastani,
74
J. Dix, & A. El Fallah Seghrouchni (Eds.), Multi-Agent Programming:
Languages, Platforms and Applications (p. 3-37). Boston, MA: Springer
US. doi: 10.1007/0-387-26350-0-1
12
Bratman, M. E., Israel, D. J., & Pollack, M. E. (1988, September). Plans
and resource-bounded practical reasoning. Computational Intelligence,
4 (3), 349-355. doi: 10.1111/j.1467-8640.1988.tb00284.x
12
Busetta, P., Ronnquist, R., Hodgson, A., & Lucas, A. (1999). JACK Intel-
ligent Agents - Components for Intelligent Agents in Java.
12
Chan, W. K. V., Son, Y. J., & Macal, C. M. (2010, December). Agent-
based simulation tutorial - simulation of emergent behavior and differ-
ences between agent-based simulation and discrete-event simulation. In
Proceedings of the 2010 Winter Simulation Conference (p. 135-150). doi:
10.1109/WSC.2010.5679168
14
Chen, X., & Zhan, F. B.
(2008, January). Agent-based modelling and
simulation of urban evacuation: Relative effectiveness of simultaneous
and staged evacuation strategies. Journal of the Operational Research
Society, 59 (1), 25-33. doi: 10.1057/palgrave.jors.2602321
4, 14
Christakis, N. A. (2007). The Spread of Obesity in a Large Social Network
over 32 Years. The New England Journal of Medicine, 370-379. doi:
10.1056/NEJMsa066082
11
{Country Fire Authority}, & {Sweeney Research}. (2009, September). A
Qualitative Report on CFA Community Engagement (Tech. Rep.). Coun-
try Fire Authority.
4
Crooks, A. T., & Heppenstall, A. J. (2012). Introduction to Agent-Based
Modelling. In Agent-Based Models of Geographical Systems (p. 85-105).
Springer, Dordrecht.
14
75
Dash, N., & Gladwin, H.
(2007, August). Evacuation Decision Making
and Behavioral Responses: Individual and Household. Natural Hazards
Review , 8 (3), 69-77. doi: 10.1061/(ASCE)1527-6988(2007)8:3(69)
7
Davis, S., Trapman, P., Leirs, H., Begon, M., & Heesterbeek, J. a. P. (2008,
July). The abundance threshold for plague as a critical percolation phe-
nomenon. Nature, 454 (7204), 634-637. doi: 10.1038/nature07053
11
Dawson, R. J., Peppe, R., & Wang, M. (2011, October). An agent-based
model for risk-based flood incident management. Natural Hazards, 59 (1),
167-189. doi: 10.1007/s11069-011-9745-4
14
de Boer, F. S., Hindriks, K. V., van der Hoek, W., & Meyer, J.-J. C. (2002,
July). Agent Programming with Declarative Goals. arXiv:cs/0207008 .
12
D'Inverno, M., Luck, M., Georgeff, M., Kinny, D., & Wooldridge, M. (2004,
July). The dMARS Architecture: A Specification of the Distributed
Multi-Agent Reasoning System. Autonomous Agents and Multi-Agent
Systems, 9 (1-2), 5-53. doi: 10.1023/B:AGNT.0000019688.11109.19
12
Epstein, J. M. (2009, August). Modelling to contain pandemics. Nature,
460 (7256), 687-687. doi: 10.1038/460687a
14
Farmer, J. D., & Foley, D. (2009, August). The economy needs agent-based
modelling. Nature, 460 (7256), 685-686. doi: 10.1038/460685a
11
Funk, S., Salath´e, M., & Jansen, V. A. A.
(2010, May). Modelling
the influence of human behaviour on the spread of infectious diseases:
A review. Journal of The Royal Society Interface, rsif20100142. doi:
10.1098/rsif.2010.0142
10
Galea, S., Riddle, M., & Kaplan, G. A. (2010, February). Causal thinking
and complex system approaches in epidemiology. International Journal
of Epidemiology, 39 (1), 97-106. doi: 10.1093/ije/dyp296
11
76
Gilbert, N. (2006). When does social simulation need cognitive models?
In R. Sun (Ed.), Cognition and Multi-Agent Interaction: From Cogni-
tive Modeling to Social Simulation (p. 428-432). Cambridge: Cambridge
University Press.
11, 12, 15
Gilbert, N., & Troitzsch, K. (2011). Simulation for the Social Scientist (2nd
ed.). UK: Open University Press.
13
Gode, D. K., & Sunder, S. (1993, February). Allocative Efficiency of Mar-
kets with Zero-Intelligence Traders: Market as a Partial Substitute for
Individual Rationality. Journal of Political Economy, 101 (1), 119-137.
doi: 10.1086/261868
13
Grimm, V., & Railsback, S. (2005). Individual-based Modelling and Ecology.
USA: Princeton University Press.
15
Harland, K., Heppenstall, A., Smith, D., & Birkin, M. H. (2012). Creating
realistic synthetic populations at varying spatial scales: A comparative
critique of population synthesis techniques. Journal of Artificial Societies
and Social Simulation, 15 .
15, 21
Haynes, K., Handmer, J., McAneney, J., Tibbits, A., & Coates, L. (2010,
May). Australian bushfire fatalities 1900 -- 2008: Exploring trends in rela-
tion to the 'Prepare, stay and defend or leave early' policy. Environmental
Science & Policy, 13 (3), 185-194. doi: 10.1016/j.envsci.2010.03.002
7
Horni, A., Nagel, K., & Axhausen, K. (2016, August). Introducing MATSim.
In A. Horni, K. Nagel, & K. Axhausen (Eds.), The Multi-Agent Transport
Simulation MATSim (p. 3-8). London: Ubiquity Press.
19
Ingrand, F. F., Georgeff, M. P., & Rao, A. S. (1992, December). An archi-
tecture for real-time reasoning and system control. IEEE Expert, 7 (6),
34-44. doi: 10.1109/64.180407
12
77
Jain, S., Ronald, N., & Winter, S. (2015, December). Creating a Synthetic
In Conference of Transportation
Population: A Comparison of Tools.
Research Group of India (p. 13). India.
15
Jarvis, J., Jarvis, D., & Ronnquist, R. (2008). Holonic Execution: A BDI
Approach. Springer Science & Business Media.
12
Johnson, P. F., Johnson, C. E., & Sutherland, C.
Human Behavior in Major Evacuations.
Dynamics (p. 675-684). Springer, Boston, MA.
7
(2011). Stay or Go?
In Pedestrian and Evacuation
Jung, C. G. (1969). Collected Works of C.G. Jung, Volume 9 (Part 1):
Archetypes and the Collective Unconscious (G. Adler & R. F. C. Hull,
Eds.). Princeton University Press.
8
Kennedy, W. G.
(2012). Modelling Human Behaviour in Agent-Based
In Agent-Based Models of Geographical Systems (p. 167-179).
Models.
Springer, Dordrecht.
11
Kermack, W. O., & McKendrick, A. G. (1927, August). A Contribution to
the Mathematical Theory of Epidemics. Proc. R. Soc. Lond. A, 115 (772),
700-721. doi: 10.1098/rspa.1927.0118
10
Knoop, V. L. (2009). Road Incidents and Network Dynamics: Effects on
driving behaviour and traffic congestion. 7
Kulash, W. (1990, October). Traditional Neighbourhood Development: Will
the Traffic Work? Bellevue WA.
10
Laird, J. E. (2007, July). Extending the Soar Cognitive Architecture: (Tech.
Rep.). Fort Belvoir, VA: Defense Technical Information Center. doi:
10.21236/ADA473738
12
78
Lee, S., Son, Y.-J., & Jin, J.
(2010, November). An Integrated Human
Decision Making Model for Evacuation Scenarios Under a BDI Frame-
work. ACM Trans. Model. Comput. Simul., 20 (4), 23:1 -- 23:24. doi:
10.1145/1842722.1842728
5, 16
Lindell, M. K., & Prater, C. S. (2007, March). Critical Behavioral Assump-
tions in Evacuation Time Estimate Analysis for Private Vehicles: Exam-
ples from Hurricane Research and Planning. Journal of Urban Planning
and Development, 133 (1), 18-29. doi: 10.1061/(ASCE)0733-9488(2007)
133:1(18)
7
Lloyd-Smith, J. O., Getz, W. M., & Westerhoff, H. V. (2004, March). Fre-
quency -- dependent incidence in models of sexually transmitted diseases:
Portrayal of pair -- based transmission and effects of illness on contact be-
haviour. Proceedings of the Royal Society of London B: Biological Sci-
ences, 271 (1539), 625-634. doi: 10.1098/rspb.2003.2632
10
Macal, C. M., & North, M. J. (2005, December). Tutorial on agent-based
modeling and simulation. In Proceedings of the Winter Simulation Con-
ference, 2005. (p. 14 pp.-). doi: 10.1109/WSC.2005.1574234
14
McCaffrey, S. M., Rhodes, A., & Stidham, M.
(2015, April). Wildfire
evacuation and its alternatives: Perspectives from four United States'
communities.
International Journal of Wildland Fire, 24 (2), 170-178.
doi: 10.1071/WF13050
7
McCaffrey, S. M., Stidham, M., Toman, E., & Shindler, B. (2011, Septem-
ber). Outreach Programs, Peer Pressure, and Common Sense: What
Motivates Homeowners to Mitigate Wildfire Risk? Environmental Man-
agement, 48 (3), 475-488. doi: 10.1007/s00267-011-9704-6
10
McClelland, J. L. (1988). Connectionist models and psychological evidence.
Journal of Memory and Language, 27 (2), 107-123. doi: 10.1016/0749
-596X(88)90069-1
12
79
McLennan, J., Elliott, G., & Omodei, M. (2012). Householder decision-
making under imminent wildfire threat: Stay and defend or leave? Inter-
national Journal of Wildland Fire, 21 (7), 915. doi: 10.1071/WF11061
8
Medsker, L. (1994). Design and development of hybrid neural network and
expert systems. In Proceedings of 1994 IEEE International Conference on
Neural Networks (ICNN'94) (Vol. 3, p. 1470-1474). Orlando, FL, USA:
IEEE. doi: 10.1109/ICNN.1994.374503
12
Miller, J., & Page, S. (2007). Complex Adaptive Systems: An Introduction to
Computational Models of Social Life. USA: Princeton University Press.
13
Moeckel, R., Spiekermann, K., & Wegener, M.
(2003, May). Creating
In Computers in Urban Planning and Urban
a Synthetic Population.
Management (p. 18). Sendai, Japan.
15
Moore, C., & Newman, M. E. J. (2000, May). Epidemics and percolation in
small-world networks. Physical Review E , 61 (5), 5678-5682. doi: 10.1103/
PhysRevE.61.5678
11
Moritz, M. A., Batllori, E., Bradstock, R. A., Gill, A. M., Handmer, J.,
Hessburg, P. F., . . . Syphard, A. D. (2014, November). Learning to coexist
with wildfire. Nature, 515 (7525), 58-66. doi: 10.1038/nature13946
7
Norling, E., Sonenberg, L., & Ronnquist, R. (2000, July). Enhancing Multi-
Agent Based Simulation with Human-Like Decision Making Strategies. In
Multi-Agent-Based Simulation (p. 214-228). Springer, Berlin, Heidelberg.
doi: 10.1007/3-540-44561-7-16
4, 12
Okaya, M., & Takahashi, T. (2011). BDI Agent Model Based Evacuation
Simulation. In The 10th International Conference on Autonomous Agents
and Multiagent Systems - Volume 3 (pp. 1297 -- 1298). Richland, SC: In-
ternational Foundation for Autonomous Agents and Multiagent Systems.
16
80
Padgham, L., Nagel, K., Singh, D., & Chen, Q. (2014). Integrating BDI
Agents into a MATSim Simulation.
In Proceedings of the Twenty-first
European Conference on Artificial Intelligence (pp. 681 -- 686). Amster-
dam, The Netherlands, The Netherlands:
IOS Press. doi: 10.3233/
978-1-61499-419-0-681
17
Padgham, L., & Singh, D. (2016, August). Making MATSim Agents Smarter
with the Belief-Desire-Intention Framework. In K. Nagel, K. Axhausen, &
A. Horni (Eds.), The Multi-Agent Transport Simulation MATSim (p. 3-8).
London: Ubiquity Press.
19
Paveglio, T., Carroll, M. S., & Jakes, P. J. (2008, March). Alternatives to
Evacuation -- Protecting Public Safety during Wildland Fire. Journal of
Forestry, 106 (2), 65-70. doi: 10.1093/jof/106.2.65
7
Pel, A., Bliemer, M., & Hoogendoorn, S. (2011). Modelling Traveller Be-
haviour under Emergency Evacuation Conditions. European Journal of
Transport and Infrastructure Research, 11 (2), 166-193.
5, 6, 7
Pillac, V., Van Hentenryck, P., & Even, C. (2016, January). A conflict-
based path-generation heuristic for evacuation planning. Transportation
Research Part B: Methodological , 83 , 136-150. doi: 10.1016/j.trb.2015.09
.008
16
Popovici, E., Bucci, A., Wiegand, R. P., & De Jong, E. D. (2012). Coevolu-
tionary Principles. In G. Rozenberg, T. Back, & J. N. Kok (Eds.), Hand-
book of Natural Computing (p. 987-1033). Berlin, Heidelberg: Springer
Berlin Heidelberg. doi: 10.1007/978-3-540-92910-9-31
19
Rao, A. S. (1996). AgentSpeak(L): BDI agents speak out in a logical com-
putable language. In J. G. Carbonell et al. (Eds.), Agents Breaking Away
(Vol. 1038, p. 42-55). Berlin, Heidelberg: Springer Berlin Heidelberg. doi:
10.1007/BFb0031845
12
Rao, A. S., & Georgeff, M. P. (1991). Modeling Rational Agents within
a BDI-Architecture. In J. Allen, E. Sandewall, & R. Fikes (Eds.), 2nd
81
International Conference on Principles of Knowledge Representation and
Reasoning (p. 473-484). San Mateo: Morgan Kaufmann Publishers, Inc.
4, 12
Rao, A. S., & Georgeff, M. P. (1995). BDI Agents: From Theory to Practice.
In In Proceedings of the First International Conference on Multi-Agent
Systems (icmas-95 (pp. 312 -- 319).
12
Reid, K., & Beilin, R. (2014, February). Where's the Fire? Co-Constructing
Bushfire in the Everyday Landscape. Society & Natural Resources, 27 (2),
140-154. doi: 10.1080/08941920.2013.840815
4, 8, 21
Reynolds, B. T. (2017). A History of the Prepare, Stay and Defend or Leave
Early Policy in Victoria. , 312.
7
Rust, J., Miller, J. H., & Palmer, R.
(1994, January). Characterizing
effective trading strategies: Insights from a computerized double auction
tournament. Journal of Economic Dynamics and Control , 18 (1), 61-96.
doi: 10.1016/0165-1889(94)90069-8
13
Sadri Arif Mohaimin, Ukkusuri Satish V., Murray-Tuite Pamela, & Glad-
win Hugh. (2014, January). How to Evacuate: Model for Understanding
the Routing Strategies during Hurricane Evacuation. Journal of Trans-
portation Engineering, 140 (1), 61-69. doi: 10.1061/(ASCE)TE.1943-5436
.0000613
16
Scerri, D., Gouw, F., Hickmott, S., Yehuda, I., Zambetta, F., & Padgham,
L. (2010). Bushfire BLOCKS: A Modular Agent-based Simulation. In
Proceedings of the 9th International Conference on Autonomous Agents
and Multiagent Systems: Volume 1 - Volume 1 (pp. 1643 -- 1644). Rich-
land, SC: International Foundation for Autonomous Agents and Multia-
gent Systems.
16
Shahparvari, S., Chhetri, P., Abareshi, A., & Abbasi, B. (2015, September).
Multi-Objective Decision Analytics for Short-Notice Bushfire Evacuation:
An Australian Case Study. Australasian Journal of Information Systems,
19 (0). doi: 10.3127/ajis.v19i0.1181
82
16
Singh, D., & Padgham, L. (2017, August). Emergency Evacuation Sim-
ulator (EES) - a Tool for Planning Community Evacuations in Aus-
tralia. In Proceedings of the Twenty-Sixth International Joint Conference
on Artificial Intelligence (p. 5249-5251). Melbourne, Australia: Inter-
national Joint Conferences on Artificial Intelligence Organization. doi:
10.24963/ijcai.2017/780
18
Singh, D., Padgham, L., & Logan, B. (2016, November). Integrating BDI
Agents with Agent-Based Simulation Platforms. Autonomous Agents and
Multi-Agent Systems, 30 (6), 1050-1071. doi: 10.1007/s10458-016-9332-x
5, 18, 19, 36, 37
Song, C., Koren, T., Wang, P., & Barab´asi, A.-L. (2010, October). Modelling
the scaling properties of human mobility. Nature Physics, 6 (10), 818-823.
doi: 10.1038/nphys1760
10, 31
Strahan, K., Whittaker, J., & Handmer, J. (2018, March). Self-evacuation
archetypes in Australian bushfire. International Journal of Disaster Risk
Reduction, 27 , 307-316. doi: 10.1016/j.ijdrr.2017.10.016
5, 8, 9, 72
Thompson, K. R., Haigh, L., & Smith, B. P. (2018, March). Planned and
ultimate actions of horse owners facing a bushfire threat: Implications for
natural disaster preparedness and survivability. International Journal of
Disaster Risk Reduction, 27 , 490-498. doi: 10.1016/j.ijdrr.2017.11.013
10
Tibbits, A., & Whittaker, J. (2007, January). Stay and defend or leave early:
Policy problems and experiences during the 2003 Victorian bushfires. En-
vironmental Hazards, 7 (4), 283-290. doi: 10.1016/j.envhaz.2007.08.001
8
Tu, H., Tamminga, G., Drolenga, H., de Wit, J., & van der Berg, W. (2010).
Evacuation plan of the city of almere: Simulating the impact of driving
behavior on evacuation clearance time. Procedia Engineering, 3 , 67-75.
doi: 10.1016/j.proeng.2010.07.008
7
83
Waddell, P. (2002). UrbanSim: Modeling urban development for land use,
transportation, and environmental planning. American Planning Asso-
ciation. Journal of the American Planning Association; Chicago, 68 (3),
297-314.
13
Whittaker, J., & Handmer, J. (2010, October). Community Bushfire Safety:
A Review of Post-black Saturday Research. Australian Journal of Emer-
gency Management, The, 25 (4), 7.
11
Whittaker, J., Haynes, K., Handmer, J., & McLennan, J. (2013, September).
Community safety during the 2009 Australian 'Black Saturday' bushfires:
An analysis of household preparedness and response. International Jour-
nal of Wildland Fire, 22 (6), 841-849. doi: 10.1071/WF12010
5, 8
Wickramasinghe, B. N., Singh, D., & Padgham, L. (2017). Heuristic Data
Merging for Constructing Initial Agent Populations.
In G. Sukthankar
& J. A. Rodriguez-Aguilar (Eds.), Autonomous Agents and Multiagent
Systems (p. 174-193). Springer International Publishing.
15
Yuan, S., Chun, S. A., Spinelli, B., Liu, Y., Zhang, H., & Adam, N. R.
(2017, May). Traffic evacuation simulation based on multi-level driving
decision model. Transportation Research Part C: Emerging Technologies,
78 .
16
84
|
1709.08235 | 1 | 1709 | 2017-09-24T18:40:21 | Dynamic Path Planning and Movement Control in Pedestrian Simulation | [
"cs.MA"
] | Modeling and simulation of pedestrian behavior is used in applications such as planning large buildings, disaster management, or urban planning. Realistically simulating pedestrian behavior is challenging, due to the complexity of individual behavior as well as the complexity of interactions of pedestrians with each other and with the environment. This work-in-progress paper addresses the tactical (path planning) and the operational level (movement control) of pedestrian simulation from the perspective of multiagent-based modeling. We propose (1) an novel extension of the JPS routing algorithm for tactical planning, and (2) an architecture how path planning can be integrated with a social-force based movement control. The architecture is inspired by layered architectures for robot planning and control. We validate correctness and efficiency of our approach through simulation runs. | cs.MA | cs |
Dynamic Path Planning and Movement Control
in Pedestrian Simulation
Fatema T. Johora∗, Philipp Kraus, and Jorg P. Muller
Department of Informatics, Clausthal University of Technology,
Julius-Albert-Str. 4 D-38678 Clausthal-Zellerfeld, Germany
September 26, 2017
Abstract
Modeling and simulation of pedestrian behavior is used in applica-
tions such as planning large buildings, disaster management, or urban
planning. Realistically simulating pedestrian behavior is challenging, due
to the complexity of individual behavior as well as the complexity of in-
teractions of pedestrians with each other and with the environment. This
work-in-progress paper addresses the tactical (path planning) and the op-
erational level (movement control) of pedestrian simulation from the per-
spective of multiagent-based modeling. We propose (1) an novel extension
of the JPS routing algorithm for tactical planning, and (2) an architecture
how path planning can be integrated with a social-force based movement
control. The architecture is inspired by layered architectures for robot
planning and control. We validate correctness and efficiency of our ap-
proach through simulation runs.
1
Introduction
Over the past two decades, pedestrian simulation has received considerable at-
tention. Modeling and analysing the behavior of individual pedestrians provides
insights for solving challenging problems such as optimizing building layouts to
assure crowd safety. Also, fueled by the paradigm of multiagent-based mod-
eling and by ever increasing computing power, recently we see a trend from
macroscopic to microscopic approaches of modeling pedestrians. Crociani et
al. classify the overall behavior of pedestrians in three levels [3]:
• At the strategic level, pedestrians maintain abstract plans that motivate
their overall behaviour.
• At the tactical level, pedestrians organize their activities, choose activity
areas and plan routes to reach their destinations.
• Pedestrians' physical movements are handled at the operational level.
∗Email: [email protected]
1
A comprehensive model of complex pedestrian behavior needs to address each
level as well as the interactions among the levels. The literature on pedestrian
simulation contains numerous approaches to handle these levels individually,
but there is not much research in integrating these levels. Well-known hierar-
chical approaches from robot planning (e.g. [7]) and hybrid agent architectures
[13] have hardly been applied to pedestrian modeling and simulation, possibly
due to their computational complexity and limited scalability to large numbers
of pedestrians. Our research aims to address this gap to a certain extent. This
work-in-progress paper provides a first small step towards applying cognitive
In this paper1 we fo-
layered control architectures to pedestrian simulation.
cus on tactical (i.e. path planning) and operational (i.e., movement control)
pedestrian behavior and their interplay. We propose and develop a model to
dynamically plan routes for pedestrians and also control their movements dur-
ing plan execution. We propose a conceptual integration architecture. For path
planning, we propose an efficient extension of the JPS algorithm; for movement
control, a social force model is used. For a proof of concept, we provide a sim-
ple initial implementation of the architecture to integrate path planning and
movement control.
The structure of this paper is as follows: Several works related to this paper
are described in Section 2. Section 3 gives an overview of the basic assumptions
and requirements underlying our approach. The proposed model is described
in Section 4. In Section 5, we report some experimental findings. We conclude
and discuss future work in Section 6.
2 State of the Art
Tactical path planning for pedestrian and controlling their individual motions
are two important problems in pedestrian modelling. There is a large body of
research in this area; some related approaches are reviewed in this section. The
concept of steering behaviors, to model realistic movements of an individual
character within an environment was introduced by Reynolds [14]. This model
proposes different behavioral characteristics such as goal seeking or collision
avoidance. The social force model by Helbing [10] is a physical model describes
the movement of pedestrians by combining different types of attractive and
repulsive forces.
In [12], Li and Chou propose an approach to plan the motion of robots in
crowded environment. Rymill and Dodgson proposed an approach to model the
motion behavior of crowds in real-time [15]. However, this work mainly focuses
on the visual perception and collision avoidance techniques of humans. A model
that can plan path for crowds using A* and also supports collision avoidance
for controlling their movements is proposed in [5]. Cherif and Chighoub [2]
present a behavioral model of pedestrian movement based on the pedestrian's
socio-psychological state.
It can be noted that most of the state-of-the-art research discussed focusses
on motion control. Although, notably, the approach in [5] implements both
path planning and a part of motion control, however, the authors do not ex-
1The level of detail in this paper is restricted by space limitations. The master's thesis
of the first author [11] provides a detailed account of the model described in this paper.
Accessible at https://goo.gl/rmIBik.
2
plain how these two levels of behaviour are integrated. However, in research on
mobile robot control and agent architectures, there are many concepts for inte-
grating planning with execution, such as the ATLANTIS architecture proposed
by Gat [7], or the layered InteRRaP architecture described in [13]. These
approaches motivate us to consider the topic of integrating path planning and
movement control behaviors of pedestrians.
3 Problems and Requirements
Normally, pedestrians plan their path based on the current state of their work-
ing environment or sometimes even before entering their working environment.
They do not consider the dynamically changing environment including other
pedestrians, during path planning. However, due to the changes in their local
environment such as extreme crowd, they may need to re-plan their path or
somehow try to adapt the changes [6]. Our assumption is that pedestrian do
not always plan path using straight minimal-length path metric. But we use this
metric to design our overall architecture and the algorithms with a motivation
that we can start with simple objective functions and, in future work, replace
them by more realistic objective functions, without the need of replacing the
path planning algorithm. To simulate the walking behavior of pedestrians, two
levels of behaviors namely planning (tactical level) and execution are required
to be integrated (operational level).
To plan the shortest path for pedestrians, a path planning algorithm is
needed. A mobility model is also needed to control the movement of pedestrians
by satisfying the dynamic aspects of their environment. Finally, a handler is
needed to handle the interaction between path planning and movement control.
4
Integrated Control Architecture
We propose a conceptual architecture that implements pedestrian path planning
and movement control and handles the collaboration between the two levels.
This architecture consists of three modules which are described in the follow-
ing subsections. The path planner (Section 4.1) provides (deliberative) route
planning on a graph or grid model (in this paper, we assume a grid-based rep-
resentation). The motion controller (cf. Section 4.2) handles the operational
(reactive) moving decisions of pedestrians. The control flow between these two
modules is orchestrated the third component: The transition handler (Sec-
tion 4.3). In the approach described in this paper, we assume that the path
planner operates at a high level of abstraction, not taking dynamic aspects of
the environment into consideration, while the motion controller accommodates
these dynamic aspects.
4.1 Path Planner
For pedestrian path planning, we propose and develop an optimization of the
Jump Point Search (JPS) algorithm [8]. We start by explaining the basic model
underlying JPS. JPS combines an A* algorithm with two sets of rules, i. e. prun-
ing rules and jumping rules:
3
Pruning Rules: JPS chooses a single path among many equivalent paths
using the following pruning strategy:
• If current node x can be reached through a straight move from its par-
ent p(x), then JPS prunes any neighbor of x that satisfies the constraint
length(p(x), ..., n) − x) ≤ length(p(x), x, n)
• If x is a diagonal move away from p(x), JPS prunes any neighbor of x that
satisfies the constraint length(p(x), ..., n) − x) < length(p(x), x, n)
If x does not have any neighbor which is an obstacle, then the remain-
ing nodes after applying pruning rules are called its natural neighbors. If any
neighbor of x isan obstacle then any remaining node n after pruning is a forced
neighbor of x, if (1) n is not a natural neighbor of the current node x, and (2)
length(p(x), x, n) < length(p(x), ..., n) − x).
Jumping Rules: JPS does not expand all non-pruned neighbors of x as its
successors. It continues moving in the direction of each non-pruned neighbor
until it finds jump points (JPs). Any non-pruned neighbor y of x, located in a
direction (cid:126)d, can be a JP if y minimizes the number of steps k such that y = x+k (cid:126)d
and also satisfies one of the following constraints:
1. y is the goal node
2. y has at least one forced neighbor
3. If (cid:126)d is a diagonal move and z = y + ki
satisfying constraint 1 or 2. Here, ki is a real number and (cid:126)di ∈ ( (cid:126)d1, (cid:126)d2)
(cid:126)di is a jump point from node y by
Figure 1: JPS-S path planning returns a sequence of jump points. Orange nodes
indicate expanded nodes. Nodes with SJ denote pre-calculated JPs, gray nodes
denote JP successors of expanded nodes.
JPS first expands the start node (S) to get all its JP successors, then recur-
sively expands the successors, until it finds the target (G). At each iteration, it
4
chooses one node to expand among all successors, based on the estimated cost
to go to G from S via each successor. In a more recent paper, the authors of
JPS introduced Jump Point Search Plus (JPS+) [9], which performs a dynamic
(and rather resource-intensive) calculation of jump points.
To increase the efficiency of JPS / JPS+ for larger simulations, we propose
an optimization named JPS with Static Jump Points (JPS-S).In JPS-S, we add
an offline pre-processing step to JPS. Here, offline preprocessing stands for all
the calculations which are performed before starting the actual path finding.
Unlike JPS+, JPS-S pre-calculates the JPs surrounding static obstacles on the
grid map; these JPs are called static jump points (SJPs) as they remain un-
changed unless the map changes. JPS+ pre-calculates the first JPs in eight
possible directions for every traversable node on the grid map, which enhances
the performance of JPS but also increases the memory usages significantly. Our
approach JPS-S speeds up JPS, but with minimized memory usage (see Sec-
tion 5).
Before starting path planning, JPS-S filters the SJP list using the given start
and goal positions to get all SJPs within the search range. The path searching
procedures of JPS and JPS-S algorithms are similar, except that while searching
the next JP in diagonal direction, JPS-S is searching recursively in the same
direction that JPS does, but, in contrast to JPS, it checks the SJP list before. If
it finds any SJP in that direction, it considers it as the successor of the currently
expanding node.
4.2 Motion Controller
We select the social force model [10] for controlling the movement of pedes-
trian. In the social force model, the motion of a pedestrian is conducted by the
following set of simple forces that a pedestrian faces at a specific time:
Acceleration: The model assumes that pedestrians move towards their des-
tination taking the shortest possible way. The pedestrian α reaches his/her des-
tination with a desired velocity (cid:126)wα(t) unless the movement of α is distracted.
If his/her velocity deviates to (cid:126)vα(t) then he/she accelerates to achieve (cid:126)wα(t)
again in a certain relaxation time τ . This acceleration can be computed by the
formula:
(cid:126)f o
α( (cid:126)wα(t), (cid:126)vα(t)) =
(cid:126)wα(t) − (cid:126)vα(t)
τ
Repulsive Forces from Other Pedestrians and Boundaries: Pedestrians
maintain some distance from unknown pedestrians (β), boundaries or static
obstacles (B). These behaviors can be represented by the two force vectors
(1)
(2)
(3)
(cid:21)
(cid:20)−(cid:126)dαβ(t)
(cid:20)−(cid:126)dαB(t)
σ
(cid:21)
(cid:126)nαβFαβ
αβ ( (cid:126)dαβ(t)) = V o
f soc
αβ exp
and
(cid:126)f soc
αB ((cid:126)dαB(t)) = U o
αB exp
(cid:126)nαB.
R
αβ and U o
Here, V o
αB denote the interaction strengths and σ and R indicates the
range of these repulsive interactions. (cid:126)dαβ(t) and (cid:126)dαB(t) are the distances from
5
α to β, and α to B at a specific time. (cid:126)nαβ and (cid:126)nαB denote the normalized
vectors. Fαβ represents the anisotropic behavior of pedestrian i.e. pedestrians
are mostly influenced by the objects which capture within their angle of view.
Attractive Forces towards Point of Interests: Pedestrians can be at-
tracted to their point of interests such as street artists. This attractive force
(cid:126)f soc
is time-dependent and linearly decreased to zero. It is calculated by the
αi
same equation as (cid:126)f soc
αβ . As compared to repulsive interactions, the interaction
range of (cid:126)f soc
is larger and the strength of this interaction is smaller, negative
αi
and time-dependent.
Joining Behavior: Pedestrians aim to stay together as a group with their
family or friends. The joining behavior represents this behavior [4]. It is calcu-
lated by the formula of the interaction strength Cαβ and the normalized vector
(cid:126)nβα. Combination of these forces describes the temporal change of pedestrian's
velocity.
4.3 Transition Handler
The control flow between path planner and motion controller can be realized
in largely differing ways: In a naive model, planner and controller may work
sequentially. In a more dynamic model, they may be represented as states of
a finite state machine, and take control alternatively. In a very sophisticated
model, they may run concurrently with a vertical or horizontal flow regulating
access to perception and action (see e.g. [13]). In this paper, we propose a con-
ceptual control architecture where transition handler (TH) component decides
when control is shifted between path planner and motion controller. This can
be modeled by a simple finite state machine with two states, where only one
state has control of the pedestrian at a specific time (see Figure 2). State tran-
sition is determined based on the currently active state, perceiving input and
transition conditions. If the perceptual input trigger any transition condition
of the current state then that transition will be conducted. The TH receives a
set of jump points specifying the desired path from the path planner. Then, it
switches to the controller state; the movement of a pedestrian from its current
position to the next jump point is then handled by the motion controller.
The transition handler will also be responsible for plan monitoring: If during
plan execution, pedestrians deviate from their pre-calculated path, the TH will
recognize this and shift the locus of control to the path planner state, where
the path will be recalculated. For identifying a deviation, we use an algorithm
very similar to the path following behavior in Reynolds' steering model [14]. It
represents a path as a spine and a radius; a spine thus represents a sequence of
line segments. If the distance between the next position of the pedestrian and
the nearest point on the path is greater than the path radius, then a deviation
is detected.
5 Evaluation
To evaluate our proposed model, we tried to answer the following questions:
6
Figure 2: Control architecture
• Does JPS-S plan optimal-length paths? Does it optimize the JPS algo-
rithm?
• Does the social force model accommodate the movement behaviors of
pedestrians?
• How to integrate the path planning and movement control behaviors of
pedestrian in a desirable way?
To answer these questions, we present an argumentation on the optimality of
JPS-S in terms of minimal length and also perform computer simulations. Sim-
ulations are implemented using the Java (version 8) programming language on
an Intel(cid:114) CoreTMi5 processor with 3 GB RAM.
Qualitative Evaluation of Minimality of Paths Found by JPS-S: Both
JPS and JPS-S algorithms give a sequence of jump points (JPs) as the optimal-
length path. The proof that the optimal path between two points (if exists) in
a grid map can be found by only expanding JPs during search, is given in [8].
In this paragraph, we argue that the extension of the pre-processing technique
(SJP) with JPS maintains optimality and completeness of JPS, as follows:
Let π = { S, JP1, JP2, JP3....JPn(G) } be the optimal-length path be-
tween S and G, which is calculated by JPS. Let further π(cid:48) = { S, JP1
(cid:48), JP2
(cid:48),
(cid:48)(G) } be the path between S and G, which is calculated by JPS-S.
(cid:48)....JPn
JP3
Let length(JPx, JPy) denote the euclidean distance between JPx and JPy.
(cid:48) is straight. If the
Any JPx = JPx
(cid:48) is diagonal then for d = length(JPx−1,JPx) and
direction from JPx−1
(cid:48)), d ≤ d(cid:48) holds. For each JPx and JPx
(cid:48), if d(cid:48) = d
d(cid:48) = length(JPx−1
then the number of JPs in π and π(cid:48) are equal. Otherwise, the number of JPs
in π(cid:48) is smaller than the number of JPs in π because length(JPx−1
(cid:48),JPx
(cid:48)) =
length(JPx−1,JPx) + length(JPx,JPx+1). The length of π and π(cid:48) is always the
same.
(cid:48), if the direction from JPx−1
(cid:48) to JPx
(cid:48), JPx
(cid:48) to JPx
JPS-S always maintains the basic properties of JPS: the pruning rules, the
jumping rules, all moves to go from one JP to another adjacent JP always
7
involve traveling in the same direction, and the line between two adjacent JPs
does not collide with any obstacle.
Performance Test: A grid-based map with 250 × 250 cells is used as the
environment for performing path planning using JPS-S. We measure perfor-
mance in terms of the relative improvement to the needed time to solve a given
problem. We perform tests on both the JPS and JPS-S algorithms, to find the
optimal path for 10000 pedestrians in parallel at a time. This test is performed
30 times. Figure 3 shows the results of this test. On average, the test takes
approximately 9% less computation time for JPS-S. In the best case, the test
takes over 16% extra time for JPS than for JPS-S.
We use the following simulation benchmarks to test the efficiency of the
social force model and all of these simulations terminates successfully as each
pedestrian reaches their goal point.
Narrow Walkway: Two oppositely directed groups of 50 pedestrians walking
through a narrow walkway to reach their goal by avoiding collisions.
Narrow Passage: A group of 100 pedestrians are trying to pass a narrow door
to reach their goal by avoiding collisions with others.
Path Following: Two groups of 100 pedestrians are trying to reach similar
goal by moving along their planned path and avoiding collisions with oth-
ers.
Figure 3: JPS-S outperforms JPS with respect to time efficiency (CPU time).
8
Interpretation of Results: The results of the above small experiments show
that the performance of JPS-S is better than that of JPS in terms of needed
time; also in the benchmark scenarios, the social force model suitably accommo-
dates the movement behavior of pedestrians. The qualitative evaluation of the
optimality of JPS-S is also given in page 7. Therefore, the first two questions
which are described in the beginning of this section are answered.
JPS-S is time efficient and the social force model efficiently enables pedes-
trians to move along their planned path which is validated from the result of
path following benchmark. Hence, we conclude that our architecture integrates
the path planning and movement control behaviors of pedestrian in a correct
and efficient way.
6 Conclusion and Outlook
This paper reports work-in-progress towards a multiagent-based model for pedes-
trian simulation that integrates tactical and operational behaviour. The con-
tribution of this paper is threefold: First, we propose a conceptual control
architecture to integrate pedestrian path planning and movement control. Sec-
ond, we propose an optimization JPS-S of the well-known JPS algorithm for
shortest path planning. We show that JPS-S maintains the optimality proper-
ties of JPS, and we compare it experimentally against JPS and JPS+, showing
its advantages in performance and scalability. Third, we describe a movement
control algorithm based on the social force model [10], and implemented a proof-
of-concept of the overall system combining JPS-S with the social force model.
Simulation results from this proof-of-concept indicate the feasibility and validity
of the overall model, but also some limitations of the approach.
So far we only implemented the naive integration architecture without re-
planning. Also, the fact that the JPS-S algorithm places the static jump points
close to static obstacles may not work well with the social force model which
tends to steers pedestrians away from obstacles. Dedicated experiments are
needed to explore this relationship. Also so far, the simulation is implemented
in plain Java. We are working to port it to the LightJason BDI agent platform [1]
for better scalability. Finally, so far we have not yet compared the output of our
model with benchmarks of real pedestrian behaviours. This is another necessary
task for future work.
References
[1] Malte Aschermann, Philipp Kraus, and Jorg P. Muller. LightJason: A
BDI Framework inspired by Jason. In Multi-Agent Systems and Agreement
Technologies: 14th Europ. Conf., EUMAS 2016, and 4rd Int. Conf., AT
2016, Valencia, Spain, 2016, volume 10207 of Lecture Notes in COmputer
Science, pages 58–66. Springer International Publishing, 2017.
[2] F Cherif and R Chighoub. Crowd simulation influenced by agent's socio-
psychological state. arXiv preprint arXiv:1004.4454, 2010.
9
[3] Luca Crociani, Andrea Piazzoni, Giuseppe Vizzari, and Stefania Bandini.
When reactive agents are not enough: Tactical level decisions in pedestrian
simulation. Intelligenza Artificiale, 9(2):163–177, 2015.
[4] AK Dewdney. Diverse personalities search for social equilibrium at a com-
puter party. Scientific American, 257(3):112–115, 1987.
[5] Cherif Foudil, Djedi Noureddine, Cedric Sanza, and Yves Duthen. Path
finding and collision avoidance in crowd simulation. CIT. Journal of Com-
puting and Information Technology, 17(3):217–228, 2009.
[6] Michel Fr´emond.
Collisions Engineering: Theory and Applications.
Springer, 2016.
[7] Erann Gat.
Integrating planning and reacting in a heterogeneous asyn-
chronous architecture for controlling real-world mobile robots. In Proceed-
ings of the 10th National Conference on Artificial Intelligence (AAAI),
pages 809–815, San Jose, CA, USA, 1992. AAAI Press.
[8] Daniel Damir Harabor, Alban Grastien, et al. Online graph pruning for
pathfinding on grid maps. In Proceedings of the 25th National Conference
on Artificial Intelligence (AAAI), pages 1114–1119, San Francisco, CA,
USA, 2011. AAAI Press.
[9] Daniel Damir Harabor, Alban Grastien, et al. Improving jump point search.
In Proceedings of the 24th International Conference on Automated Planning
and Scheduling (ICAPS), pages 128–135, Portsmouth, NH, USA, 2014.
AAAI Press.
[10] Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics.
Physical review E, 51(5):4282, 1995.
[11] Fatema T. Johora. Dynamic path planning and movement control in pedes-
trian simulation. Master's thesis, Technische Universitat Clausthal, Depart-
ment of Informatics, 2017.
[12] Tsai-Yen Li and Hsu-Chi Chou. Motion planning for a crowd of robots. In
Robotics and Automation, 2003. Proceedings. ICRA'03. IEEE International
Conference on, volume 3, pages 4215–4221. IEEE, 2003.
[13] Jorg P. Muller. The Design of Intelligent Agents, volume 1177 of Lecture
Notes in Artificial Intelligence. Springer-Verlag, 1996.
[14] Craig W Reynolds. Steering behaviors for autonomous characters. In Game
developers conference, volume 1999, pages 763–782, 1999.
[15] Stephen J Rymill and Neil A Dodgson. Psychologically-based vision and
attention for the simulation of human behaviour. In Proceedings of the 3rd
international conference on Computer graphics and interactive techniques
in Australasia and South East Asia, pages 229–236. ACM, 2005.
10
|
0710.2659 | 1 | 0710 | 2007-10-14T13:51:09 | Rigidity and persistence for ensuring shape maintenance of multiagent meta formations (ext'd version) | [
"cs.MA",
"cs.DM"
] | This paper treats the problem of the merging of formations, where the underlying model of a formation is graphical. We first analyze the rigidity and persistence of meta-formations, which are formations obtained by connecting several rigid or persistent formations. Persistence is a generalization to directed graphs of the undirected notion of rigidity. In the context of moving autonomous agent formations, persistence characterizes the efficacy of a directed structure of unilateral distance constraints seeking to preserve a formation shape. We derive then, for agents evolving in a two- or three-dimensional space, the conditions under which a set of persistent formations can be merged into a persistent meta-formation, and give the minimal number of interconnections needed for such a merging. We also give conditions for a meta-formation obtained by merging several persistent formations to be persistent. | cs.MA | cs |
RIGIDITY AND PERSISTENCE FOR ENSURING
SHAPE MAINTENANCE OF MULTIAGENT
META FORMATIONS (EXT’D VERSION)
JULIEN M. HENDRICKX, CHANGBIN YU, BARIS¸ FIDAN AND BRIAN D.O. ANDERSON
Abstract. This paper treats the problem of the merging of formations, where the underly-
ing model of a formation is graphical. We first analyze the rigidity and persistence of meta-
formations, which are formations obtained by connecting several rigid or persistent formations.
Persistence is a generalization to directed graphs of the undirected notion of rigidity. In the
context of moving autonomous agent formations, persistence characterizes the efficacy of a di-
rected structure of unilateral distance constraints seeking to preserve a formation shape. We
derive then, for agents evolving in a two- or three-dimensional space, the conditions under
which a set of persistent formations can be merged into a persistent meta-formation, and give
the minimal number of interconnections needed for such a merging. We also give conditions for
a meta-formation obtained by merging several persistent formations to be persistent.
Keywords: Formations, Meta-formations, Rigidity, Persistence, Autonomous Agents
1. Introduction
Recently, significant interest has been shown on the behavior of autonomous agent formations
(groups of autonomous agents interacting which each other) [2, 4, 7, 9, 19], and more recently on
meta-formations, which is the name ascribed to an interconnection of formations, generally
with the individual formations being separate [1, 25]. By autonomous agent, we mean here
any human-controlled or unmanned vehicle moving by itself and having a local intelligence or
computing capacity, such as ground robots, air vehicles or underwater vehicles. Many reasons
such as obstacle avoidance and dealing with a predator can indeed lead a (meta-)formation to
be split into smaller formations which are later re-merged. Those smaller formations need to
be organized in such a way that they can behave autonomously when the formation is split.
Conversely, some formations may need to be temporarily merged into a meta-formation to ac-
complish a certain task, this meta-formation being split afterwards.
The particular property of formations and meta-formations which we analyze here is persis-
tence . This graph-theoretical notion which generalizes the notion of rigidity to directed graphs
was introduced in [9] to analyze the behavior of autonomous agent formations governed by uni-
lateral distance constraints: Many applications require the shape of a multi-agent formation
J. M. Hendrickx is with Department of Mathematical Engineering, Universit´e catholique de Louvain, Avenue
Georges Lemaitre 4, B-1348 Louvain-la-Neuve, Belgium; [email protected]. His work is sup-
ported by the Belgian Programme on Interuniversity Attraction Poles initiated by the Belgian Federal Science
Policy Office, and the Concerted Research Action (ARC) “Large Graphs and Networks” of the French Commu-
nity of Belgium. The scientific responsibility rests with its authors. Julien Hendrickx holds a FNRS fellowship
(Belgian Fund for Scientific Research).
C. Yu, B. Fidan and B. Anderson are with Australian National University and National ICT Australia, 216
Northbourne Ave, Canberra ACT 2601 Australia ; brad.yu,baris.fidan,[email protected].
Their work is supported by an Australian Research Council Discovery Pro ject Grant and by National ICT Aus-
tralia, which is funded by the Australian Government’s Department of Communications, Information Technology
and the Arts and the Australian Research Council through the Backing Australia’s Ability Initiative.
to be preserved during a continuous move. For example, target localization by a group of un-
manned airborne vehicles (UAVs) using either angle of arrival data or time difference of arrival
information appears to be best achieved (in the sense of minimizing localization error) when the
UAVs are located at the vertices of a regular polygon [5]. Other examples of optimal placements
for groups of moving sensors can be found in [17]. This ob jective can be achieved by explic-
itly keeping some inter-agent distances constant.
In other words, some inter-agent distances
are explicitly maintained constant so that all the inter-agent distances remain constant. The
information structure arising from such a system can be efficiently modelled by a graph, where
agents are abstracted by vertices and actively constrained inter-agent distances by edges.
We assume here that those constraints are unilateral, i.e., that the responsibility for maintain-
ing a distance is not shared by the two concerned agents but relies on only one of them. This
unilateral character can be a consequence of the technological limitations of the autonomous
agents. Some UAV’s can for example not efficiently sense ob jects that are behind them or have
an angular sensing range smaller than 360◦ [3, 8, 20]. Also, some of the authors of this paper
are working with agents in which optical sensors have blind three dimensional cones. It can also
be desired to ease the tra jectory control of the formation, as it allows so-called leader-follower
formations [2, 6, 21]. In such a formation, one agent (leader) is free of inter-agent distance con-
straints and is only constrained by the desired tra jectory of the formation, and a second agent
(first follower) is responsible for only one distance constraint and can set the relative orientation
of the formation. The other agents have no decision power and are forced by their distance
constraints to follow the two first agents.
This asymmetry is modelled using directed edges in the graph. Intuitively, an information
structure is persistent if, provided that each agent is trying to satisfy all the distance constraints
for which it is responsible, it can do so, with all the inter-agent distances then remaining con-
stant, and as a result the formation shape is preserved. A necessary but not sufficient condition
for persistence is rigidity [9], which intuitively means that, provided that all the prescribed
distance constraints are satisfied during a continuous displacement, all the inter-agent distances
remain constant (These concepts of persistence and rigidity are more formally reviewed in the
next section). The above notion of rigidity can also be applied to structural frameworks where
the vertices correspond to joints and the edges to bars. The main difference between rigidity and
persistence is that rigidity assumes all the constraints to be satisfied, as if they were enforced
by an external agency or through some mechanical properties, while persistence considers each
constraint to be the responsibility of a single agent. As explained in [9], persistence implies
rigidity, but it also implies that the responsibilities imposed on each agent are not inconsistent,
for there can indeed be situations where this is so, and they must be avoided. Rigidity is thus
an undirected notion (not depending on the edge directions), while persistence is a directed one.
Both rigidity and persistence can be analyzed from a graph-theoretical point of view, and it can
be proved [9, 22, 28] that if a formation is rigid (resp. persistent), then almost all formations
represented by the same graph are rigid (resp. persistent).
As stated in [1], the problem of merging rigid formations into a rigid meta-formation has been
considered in a number of places. In [18, 23], the rigidity of a multi-graph (a graph in which
some vertices are abstractions of smaller graphs) is analyzed. In two dimensions, the vertices of
a multi-graph can be thought as two dimensional solid bodies at the boundary of which some
bars can be attached; two vertices are then connected by an edge if the corresponding bodies are
attached to the same bar. Of course, the idea extends obviously to three dimensions. Opera-
tional ways to merge two rigid formations into a larger rigid formation can also be found in [7, 26].
1
7
8
3
4
6
5
2
(c)
(a)
(b)
Figure 1. In ℜ2 , the graph represented in (a) is not rigid because it can be
deformed (dashed line), while the one in (b) is rigid. The graph (c) satisfies the
first two conditions of Theorem 2 but not the third one, and is therefore not rigid
in ℜ3 : the two parts of the graph can rotate around the axis defined by 1 and 2.
In this paper, we treat the problem of determining whether a given meta-formation obtained
by merging several persistent formations is persistent. For this purpose, we first consider the
above mentioned problem of determining whether a meta-formation obtained by merging rigid
formations is rigid. We also analyze the conditions under which a collection of persistent forma-
tions can be merged into a persistent meta-formation. Conditions are then given on the minimal
number of additional links that are needed to achieve such a merging. Note that throughout all
the paper, we always assume that the internal structure of the formations cannot be modified.
Moreover, we use a convenient graph theoretical formalism, abstracting agents by vertices and
(unilateral) distance constraints by (directed) edges.
After reviewing some properties of rigidity and persistence of graphs in Section 2, we examine
in Section 3 the issues mentioned above for agents evolving in a two-dimensional space. We
show in Section 4 how our results can be generalized in a three-dimensional space, and explain
why this generalization can only partially be achieved. Note that some proofs are omitted for
three-dimensional space when they are direct generalization of results on two-dimensional space.
The paper ends with the concluding remarks in Section 5.
This paper is an extended version of [12] in which some proofs are omitted for space reasons.
Some preliminary results have also been published in [11] without proofs, and are included here
at a greater level of details. Moreover, Propositions 7 and 8 correct the unproven Proposition 5
in [11], which did not take the case described in Proposition 7 into account.
2. Review of Rigidity and Persistence
2.1. Rigidity. As explained in Section 1, the rigidity of a graph has the following intuitive
meaning: Suppose that each vertex represents an agent in a formation, and each edge represents
an inter-agent distance constraint enforced by an external observer. The graph is rigid if for
almost every such structure, the only possible continuous moves are those which preserve every
inter-agent distance, as shown in Fig. 1(a) and (b). For a more formal definition, the reader is
referred to [9, 22]. In ℜ2 , that is, if the agents represented by the vertices of the graph evolve in
two dimensions, there exists a combinatorial criterion to check if a given graph is rigid:
Theorem 1 (Laman [15, 24]). A graph G = (V , E ), with V > 1, is rigid in ℜ2 if and only if
there is a sub-set E ′ ⊆ E such that
(i) E ′ = 2 V − 3.
(ii) For al l non-empty E ′′ ⊆ E ′ there holds
E ′′ ≤ 2 V (E ′′ ) − 3,
where V (E ′′ ) is the set of vertices incident to edges of E ′′ .
Unfortunately, the analogous criterion in ℜ3 is only necessary.
Theorem 2. If a graph G = (V , E ), with V > 2, is rigid in ℜ3 , there exists E ′ ⊆ E such that
(i) E ′ = 3 V − 6.
(ii) For al l non-empty E ′′ ⊆ E ′ , there holds
E ′′ ≤ 3 V (E ′′ ) − 6, where V (E ′′ ) is the set of vertices incident to edges of E ′′ .
(iii) The graph G′ (V , E ′ ) is 3-connected (i.e. remains connected after removal of any pair of
vertices).
Condition (iii), which also implies the 3-connectivity of G, is not usually stated but is inde-
pendently necessary even if the two first conditions are satisfied. Fig. 1(c) shows for example a
non-rigid graph for which (i) and (ii) are satisfied, but not (iii). Intuitively, the graph G′ in the
theorem needs to be sufficient to ensure “alone” the rigidity of G. 3-connectivity is then needed
as otherwise two or more parts of the graph could rotate around the axis defined by any pair of
vertices whose removal would disconnect the graph. Note that such connectivity condition is not
necessary in 2-dimensional spaces, as the counting conditions (i) and (ii) of Theorem 1 imply the
2-connectivity. For more information on necessary conditions for rigidity in three-dimensional
spaces, we refer the reader to [16].
We say that a graph is minimal ly rigid if it is rigid and if no single edge can be removed
without losing rigidity. It follows from the results above that a graph is minimally rigid in ℜ2
(resp. in ℜ3 ) if and only if it is rigid and contains 2 V − 3 (resp. 3 V − 6) edges [22]. Therefore
we have the following characterization of minimal rigidity in ℜ2 .
Theorem 3 (Laman [15, 24]). A graph G = (V , E ), with V > 1, is minimal ly rigid in ℜ2 if
and only if it is rigid and contains 2 V − 3 edges, or equivalently if and only if
(i) E = 2 V − 3.
(ii) For al l non-empty E ′′ ⊆ E there holds
E ′′ ≤ 2 V (E ′′ ) − 3, where V (E ′′ ) is the set of vertices incident to edges of E ′′ .
The notion of rigidity can also be described from a linear algebraic point of view, using the
so-called rigidity matrix. Suppose that a position pi ∈ ℜd (with d = 2, 3) is given to each vertex
i of a graph G = (V , E ), and let p ∈ ℜdV be the juxtapositions of all positions. For each vertex,
consider now an infinitesimal displacement δpi , and let δp be a vector obtained by juxtaposing
these displacements. Since with infinitesimal displacements one can neglect higher order terms,
the distance between the positions of two vertices i and j is preserved by the set of infinitesimal
displacements if
(1)
(pi − pj )T (δpi − δpj ) = 0.
Hence, if each edge represents a distance constraint, a set of infinitesimal displacements is al-
lowed if and only if (1) is satisfied for any edge (i, j ) ∈ E . This set of linear constraints can
be conveniently re-expressed in a condensed form as RG δp = 0 where RG ∈ ℜE ×dV is the
rigidity matrix, which contains one row for each edge and d columns for each vertex. In the row
corresponding to the edge (i, j ), the d(i − 1) + 1st to dith columns are (pi − pj )T , the d(j − 1) + 1st
to dj th columns are (pj − pi )T , and all other columns are 0. A graph G is rigid if for almost all
position assignment its rigidity matrix has a rank d V − f (d, V ), where f (d, V ) is the number
2
2
4
1
4
1
3
(a)
3
(b)
Figure 2. In ℜ2 , the graph represented in (a) is rigid but not persistent. For
almost all uncoordinated displacements of 2, 3 and 4 (even if they satisfy their
constraints), 4 is indeed unable to satisfy its three constraints. This problem
cannot happen for the graph represented in (b), which is persistent.
of degrees of freedom in a d−dimensional space of a min(V − 1, d)-dimensional rigid body (Ob-
serve that min(V − 1, d) is the largest possible dimension of a graph on V vertices embedded
in a d-dimension space). In a 2-dimensional space, a single point has two DOFs f (2, 1) = 2, and
any one or two-dimensional body has three DOFs. In a three-dimensional space, a single point
has three DOFs, a one-dimensional ob ject has five DOFs, and any other ob ject has six DOFs.
A subgraph G′ (V ′ , E ′ ) ⊆ G(V , E ) is rigid if the restriction RG′ of RG to the rows and columns
corresponding to E ′ and V ′ has a rank d V ′ − f (d, V ′ ). Note that the rank d V ′ − f (d, V ′ ) is
the maximal that can be attained by a rigidity (sub-)matrix. In a minimally rigid (sub-)graph,
this rank is attained with a minimal number of edges and all rows of the rigidity matrix are thus
linearly independent. For more information on the rigidity matrix, we refer the reader to [22].
2.2. Persistence. Consider now that the constraints are not enforced by an external entity,
but that each constraint is the responsibility of one agent to enforce. To each agent, one assigns
a (possibly empty) set of unilateral distance constraints represented by directed edges: the
−−→
notation
(i, j ) for a directed edge connotes that the agent i has to maintain its distance to
j constant during any continuous move. As explained in the Introduction, the persistence of
the directed graph means that provided that each agent is trying to satisfy its constraints, the
distance between any pair of connected or non-connected agents is maintained constant during
any continuous move, and as a consequence the shape of the formation is preserved. Note though
that the assignments given to an agent may be impossible to fulfill, in which case persistence is
not achieved. An example of a persistent and a non-persistent graph having the same underlying
undirected graph is shown in Fig. 2. For a more formal definition of persistence, the reader is
referred to [9, 28], where are also proved the rigidity of all persistent graphs and the following
criterion to check persistence:
Theorem 4. A graph G is persistent in ℜ2 (resp. ℜ3 ) if and only if every subgraph obtained
from G by removing edges leaving vertices whose out-degree is greater than 2 (resp. 3) until no
such vertex is present anymore in the graph is rigid.
A key result in the proof of Theorem 4 [9, 28] is the following:
Proposition 1. A persistent graph ℜ2 (resp. ℜ3 ) remains persistent after removal of an edge
leaving a vertex whose out-degree is larger than 2 (resp. 3).
We use the term number of degrees of freedom of a vertex i to denote the (generic) dimension
of the set in which the corresponding agent can choose its position (all the other agents being
fixed). Thus it represents in some sense the decision power of this agent. In a three-dimensional
space, an agent being responsible for one distance constraint can for example freely move on the
surface of a sphere centered on the agent from which the distance needs to be maintained, and
has thus two degrees of freedom. The number of degrees of freedom of a vertex i in ℜ2 (resp. ℜ3 )
is given by max (0, 2 − d+ (i)) (resp. max (0, 3 − d+ (i))), where d+ (i) represent the out-degree
of the vertex i. A vertex having a maximal number of degrees of freedom (i.e. an out-degree
0) is called a leader since the corresponding agent does not have any distance constraint to
satisfy. We call the number of degrees of freedom of a graph the sum of the numbers of degrees
of freedom over all its vertices. It is proved in [9, 28] that this quantity cannot exceed 3 in ℜ2
and 6 in ℜ3 . Note that those numbers correspond to the number of independent translations
and rotations in ℜ2 and ℜ3 . In the sequel we abbreviate degree of freedom by DOF.
As explained in [28], although the concept of persistence is applicable in three and larger
dimensions, it is not sufficient to imply the desired stability of the formation shape. For the
shape stability, the graph corresponding to a three-dimensional formation needs in addition to
be structural ly persistent. In ℜ3 , a graph is structurally persistent if and only if it is persistent
and contains at most one leader, i.e. at most one vertex with no outgoing edge. In ℜ2 , persis-
tence and structural persistence are equivalent.
Similarly to minimal rigidity, we say that a graph is minimal ly persistent if it is persistent
and if no single edge can be removed without losing persistence. It is proved in [9, 28] that a
graph is minimally persistent if and only if it is persistent and minimally rigid. The number of
edges of such a graph is thus uniquely determined by the number of its vertices as it is the case
for minimally rigid graphs.
3. Rigidity and Persistence of 2D Meta-Formations
3.1. Rigidity. Consider a set N of disjoint rigid (in ℜ2 ) graphs G1 , . . . , GN having at least
two vertices each, and a set S of single-vertex graphs GN +1 , . . . , GN +S . In the sequel, those
graphs are called meta-vertices, and it is assumed that no modification can be made on their
internal structure: no internal edge or vertex can be added to or removed from a meta-vertex.
We define the merged graph G by taking the union of all the meta-vertices, and of some addi-
tional edges EM each of which has end-points belonging to different meta-vertices.
The conditions under which the merging of two meta-vertices leads to a rigid graph are
detailed in [26]: If both meta-vertices contain more than one vertex, the merged graph is rigid
if and only if EM contains at least three edges, the aggregate of which are incident to at least
two vertices of each meta-vertex. This is actually a particular case of the following result for an
arbitrary number of graphs (analogous to a result in [18] which is obtained under the assumption
that no vertex of any meta-vertex is incident to more than one edge of EM ):
Theorem 5. If it contains at least two vertices, G = (cid:16)SN ,S Gi(cid:17) ∪ EM (with N and S as defined
at the beginning of this section) is rigid if and only if there exists E ′
M ⊆ EM such that
(i) E ′
M = 3 N + 2 S − 3.
(ii) For al l non-empty E ′′
M ⊆ E ′
M , there holds
E ′′
M ≤ 3 I (E ′′
M ) + 2 J (E ′′
M ) − 3,
where I (E ′′
M ) is the set of meta-vertices such that there are at least two vertices within the meta-
vertex al l incident to edges of E ′′
M , and J (E ′′
M ) is the set of meta-vertices such that there is
precisely one vertex within the meta-vertex that is incident to one or several edges of E ′′
M . Note
that in each case, there can be an arbitrary number of vertices in the meta-vertex which are not
incident on any edge of E ′′
M .
To prove this theorem, we first need the following lemma, which we shall prove the both for
ℜ2 and ℜ3 , intending to use the ℜ3 result in the next section.
1 , . . . G′
Lemma 1. Let G(V , E ) be a rigid graph (in ℜ2 or ℜ3 ), and G′
N be minimal ly rigid
subgraphs of G having distinct vertices. Then there exists a minimal ly rigid subgraph G′ (V , E ′ )
of G containing al l vertices of G and al l subgraphs Gi .
Proof. For simplicity, let us first consider the 2-dimensional case. Consider the rigidity matrix
RG of G. Since G is rigid, it has (for almost all positions) a rank 2 V − 3. Since each G′
i is
of RG to the rows and columns corresponding to the edges
minimally rigid, the restriction RG′
i
and vertices of G′
i has 2 Vi − 3 linearly independent rows (or is an empty matrix if Vi = 1).
Also, since the vertices of the different G′
i are distinct, there can be no dependence between rows
corresponding to edges of different subgraphs G′
, corresponding to
i . Therefore, all rows of RS G′
i
all edges of S G′
i , are linearly independent. Since the rank of RG is 2 V −3, it is a standard result
in linear algebra that RS G′
can be completed by the addition of further rows of RG to obtain a
i
subset of 2 V − 3 linearly independent rows of RG . Letting E ′ be the set of edges corresponding
to this set of rows, the graph G′ (V , E ′ ) is a minimally rigid subgraph of G containing all G′
i .
This completes the proof for the 2-dimensional case. The proof for the 3-dimensional case is
established following the same steps above, but replacing 2 V − 3 by 3 V − 6 and adding a
special case for Vi = 2 in addition to the case where Vi = 1.
(cid:3)
We can now prove Theorem 5.
Proof. For every Gi , let G′
i be a minimally rigid subgraph of Gi on the same vertices (The ex-
istence of such subgraphs follows directly from the definition of minimal rigidity, and they can
be obtained by successively removing edges from the initial graph). Since they are minimally
rigid, they contain 2 Vi − 3 edges if Gi ⊆ N and no edge if Gi ⊆ S .
We first suppose that there exists a set E ′
M as described in the theorem and prove the rigidity
of G, by proving the minimal rigidity of one of its subgraph viz., G′ = (V , E ′ ) = (cid:16)SN ,S G′
i(cid:17) ∪ E ′
M
which contains all its vertices. The number of edges in G′ is
M + PGi ∈N E ′
E ′ = E ′
i
= 3 N + 2 S − 3 + PGi∈N (2 Vi − 3)
= 2 V − 3,
since V = S + PGi∈N Vi . To show that G′ satisfies the second condition of Theorem 1,
suppose that there exists a subset of edges E ′′ ⊂ E ′ such that E ′′ > 2 V (E ′′ ) − 3, let I be
the set of meta-vertices containing at least two vertices of V (E ′′ ) and J the set of meta-vertices
containing only one vertex of V (E ′′ ). Let now E ′′
M = EM ∩ E ′′ and for each i, V ′′
i = V (E ′′ ) ∩ Vi
M + PGi∈I E ′′
i . There holds V (E ′′ ) = PGi ∈I V ′′
and E ′′
i + J , and E ′′ = E ′′
i = E ′′ ∩ E ′′
i .
Moreover, since each G′
i is minimally rigid, it follows from Theorem 3 that E ′′
i ≤ 2 V ′′
i − 3.
We have then
M = E ′′ − PGi ∈I E ′′
E ′′
i
> 2 V ′′ − 3 − PGi ∈I (2 V ′′
i − 3)
= 3 I + 2 J − 3,
M ⊆ E ′
so that this E ′′
M does not satisfy condition (ii) in the theorem.
We now suppose that G is rigid. It follows from Lemma 1 that there is a minimally rigid
subgraph G′ (V , E ′ ) ⊆ G containing all G′
i . Let E ′
M = E ′ ∩ EM ; we prove that E ′
M satisfies the
condition of this theorem. Since G′ is minimally rigid, there holds E ′ = 2 V − 3. Moreover,
i , and V = PGi∈N V ′
M + Pi∈N E ′
we have E ′ = E ′
i + S , so that
M (cid:12)(cid:12) = 2 V − 3 − X
(cid:12)(cid:12)E ′
(2 Vi − 3) = 3 N + 2 S − 3.
Gi ∈N
E ′
M contains thus the predicted number of edges. We suppose now that there is a set E ′′
M such
that E ′′
M > 3 I (E ′′
M ) + 2 J (EM ) − 3 and show that this contradicts the minimal rigidity of
G′ . Let us build E ′′ by taking the union of E ′′
M and all E ′
i for which Gi ∈ I (E ′′
M ). There holds
V (E ′′ ) = J (E ′′
M ) + PGi ∈I (E ′′
M ) Vi . Therefore, we have
M ) E ′
E ′′ = E ′′
M + Pi∈I (E ′′
i
M ) + 2 J (E ′′
> 3 I (E ′′
M ) − 3
+ PGi∈I (E ′′
M ) (2 Vi − 3)
= 2V (E ′′ ) − 3.
By Theorem 3, this contradicts the minimal rigidity of G′ (V , E ′ ) as E ′′ ⊆ E ′ .
(cid:3)
This criterion can be checked by a quadratic time algorithm (with respect to the number of
meta-vertices) which would be a simple adaptation of the pebble game algorithm that is used
for rigid graphs (see [14]), or even faster [18].
For a given collection of meta-vertices, we say that G is an edge-optimal rigid merging if no
single edge of EM can be removed without losing rigidity. Notice that a single graph can be an
edge-optimal rigid merging with respect to a certain collection of meta-vertices, and not with
respect to another one, as shown in Fig. 3. If all meta-vertices are minimally rigid, then an
edge-optimal rigid merging is also a minimally rigid graph. From Theorem 5, one can deduce
the following characterization of edge-optimal rigid merging.
Theorem 6. G = (cid:16)SN ,S Gi(cid:17) ∪ EM (with N and S as defined at the beginning of this section)
containing at least two vertices is an edge-optimal rigid merging if and only if it is rigid and
satisfies EM = 3 N + 2 S − 3. Moreover, each rigid merging contains an edge-optimal rigid
merging on the same set of meta-vertices.
Proof. Observe first that Theorem 5 requires a rigid merged graph G to satisfy EM ≥ 3 N +
2 S − 3. Therefore a rigid merged graph for which EM = 3 N + 2 S − 3 is an edge-optimal
merging. Let now G be a rigid merged graph. By Theorem 5 there exists E ′
M ⊂ EM with
E ′
M = 3 N + 2 S − 3 satisfying condition (ii) of this same theorem. One can see, again
using Theorem 5, that G′ = (cid:16)SN ,S Gi(cid:17) ∪ E ′
M is rigid, as the set E ′
M trivially contains itself
It follows then from the size of E ′
and satisfies both conditions (i) and (ii).
M and from the
discussion above that G′ is an edge-optimal rigid merging. We have thus proved that any rigid
merged graph G contains an edge-optimal rigid merged graph G′ on the same meta-vertices
satisfying E ′
M = 3 N + 2 S − 3. Therefore it cannot contain less than 3 N + 2 S − 3 edges,
and if it contains more of them, it is not edge-optimal. It is thus edge-optimal if and only if
EM = 3 N + 2 S − 3.
(cid:3)
3.2. Persistence. Next we analyze the case where the meta-vertices Gi are directed persistent
graphs, and adapt the definitions of N and S in consequence. If it is possible to merge them
into a persistent graph, then it is possible to do so in such a way that all the edges of EM leave
vertices which have an out-degree not greater than 2 in G: a set of edges EM that would make
G
1
G
1
G
2
G’
2
G’
3
(a)
(b)
Figure 3. The graph represented in (a) and (b) is an edge-optimal rigid merge
if it is obtained by merging G1 and G2 (a) but not if it is obtained by merging
2 and G′
G1 , G′
3 (b). The dashed edges represent the edges of EM
G persistent but that would not satisfy this property could indeed be reduced by Proposition 1
until it satisfies it. Moreover, we have the following proposition.
Proposition 2. Let G = (cid:16)SN ,S Gi(cid:17) ∪ EM with N and S as defined at the beginning of this
section, and with al l Gi persistent. If no vertex left by an edge of EM has an out-degree larger
than 2, then G is persistent if and only if it is rigid.
Proof. Rigidity is a necessary condition for persistence, so we just have to prove that it is here
sufficient. Let G′ be a (directed) graph obtained from G by removing edges leaving vertices with
out-degree larger than 2 until no such vertex exists in the graph. It follows from Theorem 4 that
we just need to prove the rigidity of any such G′ . For every i, let G′
i be the restriction of G′ to
the meta-vertex Gi . Since in G, every edge of EM leaves a vertex with an out-degree at most 2,
there holds G′ = (cid:16)SN ,S G′
i(cid:17) ∪ EM as no edge of EM is removed when building G′ . Moreover, for
every i, G′
i can be obtained from Gi by removing edges leaving vertices with an out-degree larger
than 2 until no such vertex exists in the graph anymore. The only vertices that are not left by
exactly the same edges in G as in Gi are indeed those left by edges of EM , which by hypothesis
have an out-degree at most 2 and are therefore unaffected by the edge-removal procedure. It
follows then from the persistence of all Gi and from Theorem 4 that all G′
i are rigid. And since
G is rigid, EM satisfies the necessary and sufficient conditions of Theorem 5. Therefore, the
graph G′ = (cid:16)SN ,S G′
i(cid:17) ∪ EM is also rigid, as the conditions of Theorem 5 do not depend on the
edges inside the different meta-vertices. As explained above, this implies the persistence of G′ .
(cid:3)
The condition on the out-degrees of the vertices with an outgoing edge of EM can be conve-
niently re-expressed in terms of degrees of freedom: To each DOF (within a single meta-vertex)
of any vertex there corresponds at most one outgoing edge of EM . By an abuse of language,
we say that such edges leave a vertex with one or more local DOFs, i.e. a vertex which inside
its meta-vertex has one or more DOFs and which is then left by no more edges of EM than the
number of DOFs is has. This allows reformulating Proposition 2, the proof of which can directly
be extended to any dimension, in a dimension-free way:
Theorem 7. A col lection of persistent meta-vertices can be merged into a persistent graph if
and only if it can be merged into a persistent graph by adding edges leaving vertices with one or
more local DOFs, the number of added edges not exceeding the number of local DOFs. In that
G
1
G
2
**
*
G
3
G
4
(a)
(b)
Figure 4. Merging of the persistent meta-vertices G1 and G2 into a persistent
graph in ℜ2 (a). The symbol “*” represents one DOF (with respect to the meta-
vertex). (b) represents two persistent meta-vertex that cannot be merged into a
persistent graph in ℜ2 by addition of interconnecting edges because none of their
vertices has a DOF.
case, the merged graph is persistent if and only if it is rigid.
If one or more edges of EM do leave a vertex with an out-degree larger than 2, no criterion has
been found yet to determine whether the merged graph is persistent or not, while also taking
advantage of the fact that the graph is obtained by merging several persistent meta-vertices.
Tying Theorem 7 together with what is known and reviewed above regarding the merging of
two rigid meta-vertices, we conclude: two persistent meta-vertices Ga and Gb each having two
or more vertices can be merged into a persistent graph if and only if three edges leaving vertices
with local DOFs can be added in such a way that they are incident to at least two vertices in
each meta-vertex. There must thus be at least three local DOFs available among the vertices
in Ga and Gb . Conversely, if there are available three local DOFs among the vertices of Ga and
Gb , since no vertex can have more than two DOFs, it is possible to add a total of at least three
edges leaving at least two vertices of Ga ∪ Gb . The vertices to which those edges arrive can then
be chosen in such a way that at least two vertices of both Ga and Gb are incident to edges of
EM , as in the example shown in Fig. 4. It follows then from Theorem 5 that this graph is rigid,
which by Theorem 7 implies that the merged graph is persistent:
Proposition 3. Two persistent meta-vertices each having two or more vertices can be merged
into a persistent graph if and only if the sum of their DOF numbers is at least 3. At least three
edges are needed to perform this merging, and merging can always be done with exactly three
edges.
If one or two of the meta-vertices are single vertex graphs, the result still holds, but the min-
imal number of added edges (and therefore the number of needed DOFs) are then respectively
2 and 1. We define the number of missing DOFs (mDOF ) to be the maximal number of DOFs
that any graph with the same number of vertices can have, less the number of DOFs the graph
In ℜ2 , this maximal number is 2 for the single vertex graphs, and 3 for other
actually has.
persistent graphs. There is an interesting consequence: when the minimal number of edges is
used to merge two meta-vertices Ga and Gb , the number of missing DOFs is preserved through
the process, i.e. mDOF (Ga ∪ Gb ∪ EM ) = mDOF (Ga ) + mDOF (Gb ).
Consider now an arbitrary number of persistent meta-vertices, possibly containing single-
vertex graphs, but such that the total number of vertices is at least 2.
If the sum of their
number of missing DOFs is no greater than 3, it follows from Proposition 3 that any two of
them can be merged in such a way that the obtained graph is persistent and that the total
number of missing DOFs remains unchanged. Any pair of those meta-vertices would indeed
contain at least the required number of DOFs. Doing this recursively, it is possible to merge all
these meta-vertices into a single persistent graph. In case there are more than 3 missing DOFs,
the total DOF number is by definition smaller than 3 N + 2 S − 3, which is the minimal number
of edges required to make the merged graph rigid. It follows then from Theorem 7 that such
meta-vertices cannot be merged in a persistent graph by addition of interconnecting edges. We
have thus proved the following result:
Proposition 4. A col lection of persistent meta-vertices N ∪ S (with N and S as defined in the
beginning of this section) can be merged into a persistent graph if and only if the total number
of missing DOFs is no greater than 3, or equivalently if the total number of local DOF in N ∪ S
is at least 3 N + 2 S − 3. At least 3 N + 2 S − 3 edges are needed to perform this merging,
and merging can always be done with exactly this number of edges.
As when merging rigid meta-vertices, we say that G is an edge-optimal persistent merging if
no single edge of EM can be removed without losing persistence. Again, if all meta-vertices are
minimally persistent, then G is an edge-optimal persistent merging if and only if it is minimally
persistent.
Theorem 8. G = (cid:16)SN ,S Gi(cid:17) ∪ EM (with N and S as defined at the beginning of this section and
with al l Gi persistent) is an edge-optimal persistent merging if and only if it is an edge-optimal
rigid merging and al l edges of EM leave vertices with local DOFs.
Proof. Let G be a persistent merging. If there is an edge that lies in EM leaving a vertex with
no local DOF, then it follows from Proposition 1 that the graph obtained by removing this edge
would also be persistent, and thus that G is not an edge-optimal persistent merging.
Now if G is a persistent merging for which all edges of EM leave local DOFs but which is not
an edge-optimal rigid merging, then by removing one edge of EM it is possible to obtain a rigid
graph which by Proposition 2 is also persistent, so that G is not an edge-optimal persistent
merging.
There remains to prove that an edge-optimal rigid merging G where all edges of EM leave local
DOFs is an edge-optimal persistent merging. Since such G is rigid, it follows from Proposition
2 that it is also persistent. Moreover, since it is an edge-optimal rigid merging, removing any
edge of EM destroys rigidity and therefore persistence.
(cid:3)
Tying Theorem 8 with Theorem 6 leads to the following more explicit characterization of
edge-optimal persistent merging.
Theorem 9. G = (cid:16)SN ,S Gi(cid:17) ∪ EM (with N and S as defined at the beginning of this section
and with al l Gi persistent) containing at least two vertices is an edge-optimal persistent merging
in ℜ2 if and only if the fol lowing conditions al l hold:
(i) EM = 3 N + 2 S − 3.
(ii) For al l non-empty E ′′
M ⊆ E ′
M , there holds
E ′′
M ≤ 3 I (E ′′
M ) + 2 J (E ′′
M ) − 3
with I (E ′′
M ) and J (E ′
M ) as defined in Theorem 5
(iii) Al l edges of EM leave vertices with local DOFs.
Notice that an efficient way to obtain such a merging is provided in the discussion immediately
preceding Proposition 4.
2
2 ≥ 3
1
1
1
Va
2 ≥ 3 ≥ 3
2 ≥ 3
1
Vb
4
5
6
2
3
1
min EM
Table 1. Minimal number of edges required to merge two rigid graphs Ga and
Gb into a single rigid graph in ℜ3 .
4. Rigidity and Persistence of 3D Meta-Formations
4.1. Rigidity. We now consider a set N of disjoint rigid (in ℜ3 ) graphs G1 , . . . , GN having at
least three vertices each, a set D of graphs containing two (connected) vertices GN +1 , . . . , GN +D ,
and a set S of single-vertex graphs GN +D +1 , . . . , GN +D +S . As in Section 3, these graphs
are called meta-vertices, and we define the merged graph G by taking the union of all the meta-
vertices, and of some additional edges EM each of which has end-points belonging to different
meta-vertices.
The merging of two rigid meta-vertices, each containing more than two vertices, is treated
in [26]: At least six edges are needed, and they must be incident to at least three vertices of each
meta-vertex (which is necessary for 3-connectivity). But these conditions are only necessary, as
they do not imply 3-connectivity. For example, the so-called “double-banana” graph in Fig. 1(c)
can be obtained by merging two distinct rigid tetrahedral meta-vertices (1,3,4,5) and (2,5,7,8)
using a total of six edges incident to four vertices of each meta-vertex. However, it is always
possible to achieve a rigid merging using exactly six edges incident to exactly three vertices
of each meta-vertex, with no single vertex having more than three incident edges out of the
six. With a minor modification, the merging result above holds in the cases where at least one
meta-vertex has less than 3 vertices: The required number of edges is different, as summarized
in Table 1 where min EM represents the minimal number of edges required to merge the meta-
vertices Ga (Va , Ea ) and Gb (Vb , Eb ) into a rigid graph. Also, if a meta-vertex has less than 3
vertices, all of them should be incident to edges of EM , otherwise at least 3 of them should
be. When merging several meta-vertices, there is no available necessary and sufficient condition
for the rigidity of G. Determining whether a merged graph is rigid in ℜ3 is indeed a more
general problem than determining whether a given graph is rigid (for which it suffices to take
N = D = ∅) and there is no known set of combinatorial necessary and sufficient conditions for
this. We can however prove that the rigidity of the merged graph G only depends on EM , on
the vertices to which nodes of EM are incident and on the belonging of the Gi to N , D or S .
Proposition 5. Let G = (cid:16)SN ,D ,S Gi(cid:17) ∪ EM with N , D , S as defined at the beginning of this
section. Suppose that a meta-vertex Gi is replaced by a meta vertex G′
i with the same set of
vertices incident to EM , with the same set membership, N , S or D, as Gi , but otherwise with
different internal structure. Let G’ be the graph so obtained. Then G′ is rigid if and only if G
is rigid.
Proof. This could be proved using algebraic arguments based on the rigidity matrix, but we
prefer the following more intuitive argument.
The result is trivial for meta-vertices of D and S as they are entirely determined by their be-
longing to these classes; we assume therefore that Gi ∈ N . We also assume that the set Vi (EM )
of vertices of Gi (and G′
i ) that are incident on edges of EM contains at least three vertices. In
case this assumption is not verified, both G and G′ fail to be 3-connected and therefore rigid (by
Theorem 2), so that the result is also trivial. We then prove that the non-rigidity of G implies
the non-rigidity of G′ . Since the roles of G and G′ can be exchanged, this is sufficient to prove
the theorem.
Suppose that G is not rigid, and give positions in ℜ3 to its vertices. Then there is a smooth
motion M (satisfying the distance constraints corresponding to edges in G) of the vertices of G
apart from pure translation or rotation. Because Gi is rigid, the restriction of M to the vertices
of Gi is a rigid motion, that is a translation and/or rotation, which we call T . Therefore, the
restriction of M to (G \ Gi ) ∪ Vi (EM ) is not a rigid motion. Otherwise all distances would
be preserved by M apart from some distances between vertices of Gi \ Vi (EM ) and vertices of
G \ Gi . We would then have two vertices whose relative distance is not preserved while their
relative distance with respect to all the three or more vertices of Vi (EM ) are preserved, which is
impossible. We call M ∗ this restriction to (G \ Gi ) ∪ Vi (EM ). Let now M ′ be a smooth motion
of the vertices of G′ , which for the vertices of G′
i is the translation and/or rotation T , and for
the vertices of (G′ \ G′
i ) ∪ Vi (EM ) is the motion M ∗ (observe that that the two motions are
identical on V (EM ) which is the intersection of the two sets on which M ′ is defined). Since
M ∗ is a non-rigid motion (not preserving all distances), so is M ′ . Therefore, we just need to
prove that M ′ satisfies all distance constraints on vertices connected by edges in G′ to prove the
non-rigidity of G′ . Consider a pair of vertices. If they both belong to (G′ \ G′
i ) ∪ Vi (EM ), their
constraint in G′ is the same as in G, and their motion is defined by M ∗ which satisfies all distance
constraints. If they do not both belong to (G′ \ G′
i ) ∪ Vi (EM ), then due to the structure of the
graph they necessarily both belong to G′
i , and their motion is the rotation and/or translation
which by essence preserve all distances.
(cid:3)
Moreover, we have the following necessary condition:
Theorem 10. Let Gi for i = 1, 2, ...., N + D + S be rigid meta-vertices, and suppose
G = (cid:16)SN ,D ,S Gi(cid:17) ∪ EM (with N , D , S as defined at the beginning of this section) is rigid in
ℜ3 and contains at least three vertices. Then there exists E ′
M ⊆ EM such that
(i) E ′
M = 6 N + 5 D + 3 S − 6
(ii) For al l non-empty E ′′
M ⊆ E ′
M , there holds
E ′′
M ≤ 6 I (E ′′
M ) + 5 J (E ′′
M ) + 3 K (E ′′
M ) − 6,
where I (E ′′
M ) is the set of meta-vertices such that either there are at least three vertices within
the meta-vertex al l incident to edges of E ′′
M , or precisely two vertices within the meta-vertex
which are unconnected and both incident to edges of E ′′
M . J (E ′′
M ) is the set of meta-vertices such
that there are precisely two vertices within the meta-vertex which are connected and both incident
to edges of E ′′
M ; K (E ′′
M ) is the set of meta vertices such that there is precisely one vertex within
the meta-vertex that is incident to one or several edges of E ′′
M . Note that in each case, there
can be an arbitrary number of vertices in the meta-vertex which are not incident on any edge of
E ′′
M .
Moreover, the graph (cid:16)SN ,D ,S Gi(cid:17) ∪ E ′
M is rigid.
Proof. The proof is similar to the one of Theorem 5 (necessary part). For every Gi , let G′
i be a
minimally rigid subgraph of Gi on the same vertices, which therefore contains 3 Vi − 6 edges if
Gi ∈ N , one edge if Gi ∈ D and no edge if Gi ⊆ S . As mentioned in its proof, Lemma 1 can also
be applied in a three-dimensional space. So if G is rigid, there is a minimally rigid subgraph
G′ (V , E ′ ) ⊆ G containing all G′
i . Let E ′
M = E ′ ∩ EM ; we shall prove that E ′
M satisfies the
condition of this theorem. Since G′ is minimally rigid, there holds E ′ = 3 V − 6. Moreover,
we have E ′ = E ′
M + PGi∈N E ′
i + D , and V = PGi ∈N V ′
i + 2 D + S , so that
E ′
M = 3 V − 6 − PGi∈N (3 Vi − 6) − D
= 6 N + 5 D + 3 S − 6.
M contains thus the predicted number of edges. We suppose now that there is a set E ′′
E ′
M such
that E ′′
M > 6 I (E ′′
M ) + 5 J (E ′′
M ) + 3 K (E ′′
M ) − 6 and show that this contradicts the minimal
rigidity of G′ . Let us then build E ′′ by taking the union of E ′′
M and all E ′
i for which i ∈ I (E ′′
M ),
M in all meta-vertices in J (E ′′
and the edge connecting the two vertices incident to E ′′
M ). There
holds V (E ′′ ) = K (E ′′
M ) + 2 J (E ′′
M ) + PGi ∈I (E ′′
M ) Vi . Therefore, we have
M ) E ′
i + J (E ′′
E ′′ = E ′′
M + PGi∈I (E ′′
M )
> 6 I (E ′′
M ) + 5 J (E ′′
M ) + 3 K (E ′′
M ) − 6
M ) (3 Vi − 6) + J (E ′′
+ PGi∈I (E ′′
M )
= 3 V (E ′′ ) − 6.
This however contradicts the minimal rigidity of G′ as E ′′ ⊆ E ′ . Finally, since G′ = (cid:16)SN ,D ,S G′
i(cid:17)∪
M is rigid, it follows from several applications of Proposition 5 that (cid:16)SN ,D ,S Gi(cid:17) ∪ E ′
E ′
M is also
rigid.
(cid:3)
Note that the rigidity of (S Gi ) ∪ E ′
M is explicitly mentioned here and not in Theorem 5,
because in a two-dimensional space it follows directly from sufficiency of the counting conditions.
But, the counting conditions of Theorem 10 are not sufficient for rigidity, as the non-rigid graph
of Fig. 1(c) which can be obtained by merging two rigid tetrahedral meta-vertices (1,3,4,5) and
(2,6,7,8) would indeed satisfy them. Nevertheless, one can deduce from Theorem 10 that G is
an edge-optimal rigid merging in ℜ3 if and only if it is rigid and EM = 6 N + 5 D + 3 S − 6,
using E ′
M exactly in the same way as in Theorem 6.
4.2. Persistence. We consider now that all meta-vertices Gi are persistent graphs, and adapt
the definitions of N , D and S in consequence. Theorem 7 can be generalized to three dimensions,
as it follows from Proposition 2, the proof of which can be immediately extended to three
dimensions.
Theorem 11. A col lection of (structural ly) persistent meta-vertices can be merged into a (struc-
tural ly) persistent graph if and only if it can be merged into a (structural ly) persistent graph by
adding edges leaving vertices with one or more local DOFs. In that case, the merged graph is
persistent if and only if it is rigid.
Proof. Suppose first that a collection of persistent meta-vertices can be merged into a persistent
graph in such a way that some edges do not leave local DOFs. Then, it follows from Proposition
1 that these edges can be removed without destroying the persistence of the merged graph, so
that the same collections of meta-vertices can be merged without having connecting edges that
do not leave local DOFs. In case the meta-vertices are structurally persistent and are merged
into a structurally persistent graph, the result still holds as removing edges that do not leave
local DOFs never destroys structural persistence. The reverse implication is trivial.
The proof of the rest of the result is done exactly as in Theorem 2, using Proposition 5 instead
of Theorem 5.
(cid:3)
Merging two meta-vertices into a persistent graph is however a more complicated problem
in ℜ3 than in ℜ2 . Consider indeed a meta-vertex Ga without any DOF, and a meta-vertex Gb
which is not structurally persistent, i.e. which is persistent and contains two vertices (leaders)
having three DOFs. The number of available DOFs is equal to the minimal number of edges that
should be added to obtain a rigid merged graph. However, the only way to add six edges leaving
local DOFs is to add three edges leaving each leader of Gb and arriving in Ga , as represented
by the example in Fig. 5(a). Only two vertices of Gb would thus be incident to the added
G
a
***
G
b
***
G
c
***
G
d
***
***
***
(a)
(b)
Figure 5. Example of a persistent but not structurally persistent meta-vertex
Gb which cannot be merged into a persistent or rigid graph with the meta-vertex
Ga , the latter being persistent but having no DOF. (b) shows how two non-
structurally persistent meta-vertices can be merged into a structurally persistent
graph. The symbol “*” represents one DOF, and the dashed edges are the edges
of EM .
G
a
***
G
b
***
Figure 6. Ga and Gb both have all their DOFs concentrated on one leader. As
a result they cannot be merged into a persistent graph. The only way to add 6
edges leaving local DOFs is depicted and does not lead to a rigid graph, because
the overall graph is not 3-connected. The symbol “*” represents one DOF, and
the dashed edges are the edges of EM .
edges, which prevents the merged graph from being rigid and therefore persistent as it is thus
not 3-connected. We have thus proved the following condition:
Proposition 6. If two persistent meta-vertices are such that one is not structural ly persistent
and the other does not have any DOF, they cannot be merged into a persistent graph by addition
of interconnecting edges.
Another problem appears when Ga and Gb each have one leader (having three DOFs) and no
other vertex has DOFs. Again, the number of available DOFs is equal to the minimal number
of edges that should be added to obtain a rigid merged graph, but the only way to add six edges
leaving local DOFs does not lead to a rigid graph. One can indeed only add three edges leaving
each leader as shown in Fig. 6. This results in a graph that is not 3-connected and therefore not
rigid by Theorem 2, as the removal of the two ex-leaders would render the graph unconnected.
We have thus proved the following condition:
Proposition 7. If two persistent meta-vertices have each one leader (with 3 DOFs) and no
other DOF, they cannot be merged into a persistent graph by addition of interconnecting edges.
However, these are the only cases for which the argument used in establishing Proposition 3
cannot be generalized to establish an analogous property in ℜ3 :
Proposition 8. Two persistent meta-vertices (each with three or more vertices) can be merged
into a persistent graph by addition of directed connecting edges if and only if the sum of their
DOFs is at least 6 and the DOFs are located on more than two vertices. At least six edges are
needed to perform this merging, and merging can always be done with exactly six edges and in
such a way that the graph obtained is structural ly persistent and does not have al l its DOFs
located on leaders.
Proof. Consider two meta-vertices each having more than 2 vertices. It follows from Theorem
11 that they can be merged into a persistent graph if and only if it is possible to add directed
edges leaving local DOFs in such a way that the obtained graph is rigid.
Suppose first that the total number of available DOFs is 6. If all these DOFs are located on
two leaders, the two graphs satisfy the conditions of either Proposition 6 or Proposition 7, so
that they cannot be merged into a persistent graph. If the 6 DOFs are located on more than
2 vertices, an exhaustive verification (see Appendix) show that the two graphs can always be
merged into a rigid graph by adding 6 edges, each leaving a vertex with a local DOF, with at
least one DOF for each edge. Note that this exhaustive verification is needed as no sufficient
condition for rigidity of a graph obtained by connecting two rigid graphs is known which is
sufficiently weak to be helpful for this proof.
If the total number of DOFs is larger than 6, they are located on at least 3 vertices, as a
vertex has at most 3 DOFs. It is therefore possible to select a subset of 6 DOFs located on at
least 3 vertices, and to apply the result obtained above for 6 DOFs.
There remains to prove that the merging can always be done in such a way that the obtained
graph does not have all its DOFs located on leaders, or in other words the obtained graph
has only vertices with 0 or 3 DOFs (This also implies that the graph obtained is structurally
persistent, as the only persistent graphs that are not structurally persistent are those with two
leaders and therefore no other DOF). Such a situation, i.e. the obtained graph has only vertices
with 0 or 3 DOFs, could only happen if this graph has exactly 3 or 6 DOFs, and thus if 9 or
12 DOFs are initially available, as the merge is done by addition of 6 edges. A simple way of
avoiding having all remaining DOFs on leaders is then to select the 6 DOFs that are going to
be removed in the merging process in such a way that a number of DOFs different from 3 and
6 is left in each of the initial graphs. At least one vertex has then indeed one or two DOFs.
(cid:3)
In case at least one of the two meta-vertices has less than 3 vertices, an exhaustive consider-
ation of all possible cases (see Appendix) shows that the result still holds, but with a different
required number of edges in EM and therefore of available DOFs: these minimal numbers
are both equal to min EM in Table 1 (for the merging of a graph Ga (Va , Ea ) with a graph
Gb (Vb , Eb )). Observe that as in the 2-dimensional case, the merge can be done in such a way
that the number of missing DOFs is preserved, the number of missing DOFs being defined in the
same way as in Section 3.2, with maximal number of DOFs being 6, 5 and 3 for meta-vertices
of respectively N , D and S . It is worth noting that even if one or both of the meta-vertices
are not structurally persistent, it is possible to obtain a structurally persistent merged graph,
as represented in Fig. 5(b). This has already been observed in [28] for the case where one
meta-vertex is a single vertex graph.
Consider now a collection of meta-vertices such that the total number of vertices is at least 3.
Unless the collection consists in two meta-vertices satisfying the hypotheses of Proposition 6 or 7,
all the graphs that compose it can be merged into one large persistent graph by addition of edges.
Proposition 9. A col lection of persistent meta-vertices N ∪ D ∪ S (with N , D , S as defined in
the beginning of this section) containing in total at least three vertices and that does not consist
of only two meta-vertices satisfying the condition of Proposition 6 or 7 can be merged into a
persistent graph if and only if the total number of missing DOFs is no greater than 6, or equiv-
alently if the total number of local DOFs in N ∪ D ∪ S is at least 6 N + 5 D + 3 S − 6. At
least 6 N + 5 D + 3 S − 6 edges are needed to perform this merging. Merging can always be
done with exactly this number of edges, and in such a way that the merged graph is structural ly
persistent.
Proof. The proof is similar to the one of Proposition 4. If a pair of meta-vertices can be merged
into a persistent graph, this merging can be done in such a way that the number of missing
DOFs is preserved, and by adding only edges leaving vertices with local DOFs (with at most
one edge for each DOF). Doing this recursively, we eventually obtain a single persistent graph
that has the same number of missing DOFs as the initial collection of graphs. The number of
added edges is then equal to the number of DOFs that have disappeared during the merging
process, that is 6 N + 5 D + 3 S − 6.
There remains to prove that these mergings can actually be done, and that the obtained graph
is structurally persistent. By Proposition 8 (and its extension to graphs with 1 or 2 vertices),
when their number of missing DOFs is smaller than 6, two persistent graphs can always be
merged into a structurally persistent graph, unless either one of them is not structurally persis-
tent while the other has no DOF (case of Proposition 6), or both of them have one leader and
no other DOF (case of Proposition 7). In these two cases, the two “problematic” meta-vertices
have at least three vertices each.
Suppose first that one meta-vertex has no DOF (and that the rest of the meta-vertices collec-
tion does not consist in one single non structurally persistent meta-vertex). Then since the total
number of missing DOF is 6, no other meta-vertex has a missing DOF, and by hypothesis there
are at least two other meta-vertices (or possibly exactly one structurally persistent meta-vertex).
It follows then from successive applications of Proposition 8 that they all can be merged into a
structurally persistent graph that still does not have any missing DOF. This latter graph can
then be merged with the graph that has no DOF, and the graph obtained is also structurally
persistent.
Suppose now that two meta-vertices have exactly one leader and no other DOF. It follows then
from the hypotheses that there is at least one other meta-vertex in the collection. And again,
no other meta-vertex has any missing DOF. Temporarily isolating one of the meta-vertices with
one leader and no other DOF, it follows again from successive applications of Proposition 8 that
all other graphs can be merged into a persistent graph that does not have all its DOFs located
on one single leader, and this graph can then be merged with the temporarily isolated graph
into a structurally persistent graph.
(cid:3)
As in the two-dimensional case, a merged graph is an edge-optimal persistent merging if and
only if it is an edge-optimal rigid merging and all edges in EM (such as defined in the beginning of
this subsection) leave local DOFs. The proof of this is an immediate generalization of Theorem
8. However, due to the absence of necessary and sufficient conditions allowing a combinatorial
checking of the rigidity of a graph or of a merged graph in ℜ3 , the result cannot be expressed in
a purely combinatorial way. Since the number of edges in EM in an edge-optimal rigid merging
is fixed, the above criterion can be re-expressed as
Theorem 12. G = (cid:16)SN ,D ,S Gi(cid:17) ∪ EM (with N , D , S as defined at the beginning of this section
and with al l Gi persistent) containing in total at least three vertices is an edge-optimal persistent
merging in ℜ3 if and only if the fol lowing conditions al l hold:
(i) G is rigid.
(ii) Al l edges of EM leave local DOFs.
(iii) EM = 6 N + 5 D + 3 S − 6.
Again, an efficient way to obtain an edge-optimal persistent merging from a collection of
meta-vertices satisfying the hypotheses of Proposition 9 is to first merge two of them and then
to iterate, as in the discussion of Propositions 4 and 9.
5. Conclusions
We have analyzed the conditions under which a formation resulting from the merging of
several persistent formations is itself persistent. Necessary and sufficient conditions were found
to determine which collections of persistent formations could be merged into a larger persistent
formation. We first treated these issues in ℜ2 . Our analysis was then generalized to ℜ3 and to
structural persistence, leading to somewhat less powerful results. This is especially the case for
those which rely on the sufficient character of Laman’s conditions for rigidity in ℜ2 (Theorem 1),
no equivalent condition being known in ℜ3 . Following this work, we plan to develop systematic
ways to build all possible optimally merged persistent formations, similarly to what has been
done for minimally persistent formations [10] and for minimally rigid merged formations [27].
These references canvas generalizations of the Henneberg sequence concept [13, 22] for building
all minimally rigid graphs in two dimensions.
References
[1] B.D.O. Anderson, C. Yu, B. Fidan, and J.M. Hendrickx. Use of meta-formations for cooperative control.
In Proceedings of the 17th International Symposium on Mathematical Theory of Networks and Systems
(MTNS2006), pages 2381–2387, Kyoto (Japan), July 2006.
[2] J. Baillieul and A. Suri. Information patterns and hedging brockett’s theorem in controlling vehicle forma-
tions. In Proc. of the 42nd IEEE Conf. on Decision and Control, volume 1, pages 556–563, Hawaii, December
2003.
[3] J.M. Borky. Payload technologies and applications for uninhabited air vehicles (uavs). In Proc. of the IEEE
Aerospace Conf., volume 3, pages 267–283, Aspen (CO), USA, February 1997.
[4] A. Das, R. Fierro, V. Kumar, and Ostrowski J.P. A vison based formation control framework. IEEE trans.
on Robotics and Automation, 18(5):813–825.
[5] K. Dogancay. Optimal receiver tra jectories for scan-based radar localization. In Proc. Information, Decision
and Control Conference 2007, Adelaide (SA), Australia, February 2007.
[6] T. Eren, B.D.O. Anderson, A.S. Morse, Whiteley W., and P.N. Belhumeur. Information structures to secure
control of rigid formations with leader-follower structure. In Proc. of the American Control Conference, pages
2966–2971, Portland, Oregon, June 2005.
[7] T. Eren, B.D.O. Anderson, W. Whiteley, A.S. Morse, and P.N. Belhumeur. Operations on rigid formations
of autonomous agents. Communications in Onformations and Systems, 3(4):223–258, 2004.
[8] H.R Everett. Sensors for Mobile Robots: Theory and Application. A.K. Peters, 1995.
[9] J.M. Hendrickx, B.D.O. Anderson, J.-C. Delvenne, and V.D. Blondel. Directed graphs for the analysis of
rigidity and persistence in autonomous agents systems. International Journal of Robust and Nonlinear Con-
trol, 17:960–981, 2007.
[10] J.M. Hendrickx, B. Fidan, C. Yu, B.D.O. Anderson, and V.D. Blondel. Primitive operations for the construc-
tion and reorganization of minimally persistent formations. To appear in IEEE Transactions on Automatic
Control.
[11] J.M. Hendrickx, C. Yu, B. Fidan, and B.D.O. Anderson. Rigidity and persistence of meta-formations. In
Proceedings of the 45th IEEE Conference on Decision and Control, pages 4567–4663, San Diego (CA), USA,
December 2006.
[12] J.M. Hendrickx, C. Yu, B. Fidan, and B.D.O. Anderson. Rigidity and persistence for ensuring shape main-
tenance in multiagent meta-formations. To appear in Asian Journal of control’s special issue on Col lective
Behavior and Control of Multi-Agent Systems, 2008.
[13] L. Henneberg. Die graphische Statik der starren Systeme. Leipzig, 1911.
[14] Donald J. Jacobs and Bruce Hendrickson. An algorithm for two-dimensional rigidity percolation: the pebble
game. J. Comput. Phys., 137(2):346–365, 1997.
[15] G. Laman. On graphs and rigidity of plane skeletal structures. J. Engrg. Math., 4:331–340, 1970.
[16] A. Mantler and J. Snoeyinkn. Banana spiders: A study of connectivity in 3d combinatorial rigidity. In Pro-
ceedings of the 16th Canadian Conference on Computational Geometry (CCCG 2004), pages 44–47, Montr´eal,
Canada, August 2004.
[17] S. Martinez and F. Bullo. Optimal sensor placement and motion coordination for target tracking. Automatica,
42(3):661–668, 2006.
[18] C. Moukarzel. An efficient algorithm for testing the generic rigidity of graphs in the plane. J. Phys. A,
29(24):8079–8098, 1996.
[19] R. Olfati-Saber and R.M Murray. Graph rigidity and distributed formation stabilization of multi-vehicle
systems. In Proceedings of the 41st Conference on Decision and Control, volume 3, pages 2965–2971, Las
Vegas, NV, December 2002.
[20] M. Pachter and J. Hebert. Cooperative aircraft control for minimum radar exposure, pages 199–211. Kluwer
Academic, 2002.
[21] H.G. Tanner, G.J. Pappas, and V. Kumar. Leader-to-formation stability. IEEE Transactions on Robotics and
Automation, 4(3):433–455, 2004.
[22] T. Tay and W. Whiteley. Generating isostatic frameworks. Structural Topology, (11):21–69, 1985.
[23] T.S. Tay. Rigidity of multigraphs. I. Linking rigid bodies in n-space. J. Combin. Theory Ser. B, 36(1):95–112,
1984.
[24] W. Whiteley. Some matroids from discrete applied geometry. In Matroid theory (Seattle, WA, 1995), volume
197 of Contemp. Math., pages 171–311. Amer. Math. Soc., Providence, RI, 1996.
[25] A. Williams, S. Glavaski, and T. Samad. Formations of formations: Hierarchy and stability. In Proc. of the
2004 American Control Conference, pages 2992–2997, Boston (MA), USA, July 2004.
[26] C. Yu, B. Fidan, and B.D.O. Anderson. Principles to control autonoumous formation merging. Proc. of the
2006 American Control Conference, pages 762–768, June 2006.
[27] C. Yu, B. Fidan, J.M. Hendrickx, and B.D.O. Anderson. Multiple formations merging: a meta-formation
prospective. In Proceedings of the 45th IEEE Conference on Decision and Control, pages 4657–4663, San
Diego (CA, USA), December 2006.
[28] C. Yu, J.M. Hendrickx, B. Fidan, B.D.O. Anderson, and V.D. Blondel. Three and higher dimensional au-
tonomous formations: Rigidity, persistence and structural persistence. Automatica, 43:387–402, 2007.
In this appendix, we complete the proof of Proposition 8 on the merging of persistent meta-
vertices in a 3-dimensional space, and extend this proposition to cases where one of the meta-
vertices has less than 3 vertices. We have to prove that two persistent graphs (in a three-
dimensional space) Ga and Gb having in total 6 DOFs located on at least three vertices can
always be merged into a rigid graph by addition of six edges leaving vertices with local DOFs,
with at least one local DOF for each added edge.
For this purpose, we use the following lemma, which summarizes results obtained in [26].
Lemma 2. Let Ga and Gb be two (initial ly distinct) rigid graphs each with three or more ver-
tices. Performing a sequence of three or more operations selected among the two fol lowing types
of operations results in merging Ga and Gb into a rigid graph by addition of 6 edges.
Operation (v): Taking a vertex i of Ga not connected yet to any vertex of Gb , and connecting it
to 3 − t vertices of Gb , where t is the number of operations already performed.
Operation (e): Taking a vertex i of Ga not connected yet to any vertex of Gb , and an edge (k , j )
with k ∈ VA and j ∈ VB . Replacing the edge (k , j ) by (i, j ) and connecting i to 2 − t other
Ga
Gb
d+
EM
3
2
1
Ga
Gb
Operations: (v)(v)(v)
(a)
(b)
Figure 7. (a) represents a rigid merged graph obtained by performing three
operations (v). It is shown in (b) how directions can be given to the edges in
such a way that three vertices of Ga are left by respectively 3, 2 and 1 edges.
vertices in Gb , where t is the number of operations already performed.
Without loss of generality, we suppose that Ga has at least as many DOFs as Gb . The par-
tition of DOFs can thus be 6-0, 5-1, 4-2 or 3-3. In the sequel, we prove the result for each of
these particular cases, starting with Ga having 6 DOFs.
It follows from Lemma 2 that the merged graph represented in Fig. 7(a) is rigid.
It can
indeed be obtained by three applications of the operation (v). Moreover, one can see in Fig.
7(b) that directions can be given to the connecting edges in such a way that the out-degree
distribution (with respect to the connecting edges) is (3, 2, 1), that is one vertex of Ga is left by
three connecting edges, one by two, and one by one. Suppose now that Ga is a persistent graph
with 6 DOFs with a DOF allocation (3, 2, 1), that is a persistent graph having one vertex having
3 DOFs, one 2 DOFs, and one 1 DOF. Then it can be merged with Gb into a rigid graph by
adding 6 edges leaving vertices with local DOF (with one DOF for each edge). It suffices indeed
to take the edges represented in Fig. 7(b), identifying each vertex with δ DOFs with a vertex
left by δ connecting edges.
We now treat a DOF allocation (2, 2, 2).
It follows again from Lemma 2 that the merged
graph represented in Fig. 8(a) is rigid, as it can be obtained by two applications of the opera-
tion (v) followed by one application of operation (e). Moreover, Fig. 8(b) shows that directions
can be assigned to the edges in such a way that the out-degree distribution (with respect to the
connecting edges) is (2, 2, 2). For the same reason as above, Ga can thus be merged with Gb
into a rigid graph by adding 6 edges leaving vertices with local DOF (with one DOF for each
edge) if its DOF distribution is (2, 2, 2).
Next we show that such construction can be obtained in all other cases, except those where
the 6 DOFs are all located on two vertices. When Ga has 6 DOFs, the remaining possible DOF
distributions are (3, 1, 1, 1), (2, 2, 1, 1), (2, 1, 1, 1, 1) and (1, 1, 1, 1, 1, 1), the case (3, 3) does not
satisfy the hypotheses. The construction in these four cases are obtained by performing the
operation (e) of Lemma 2 (up to three times) on the constructions detailed above for (3, 2, 1)
and (2, 2, 2). They are represented in Fig. 9
If Ga has 5 DOFs and Gb one DOF, the required construction can always be obtained from
one of the construction for the case where Ga has 6 DOFs. It suffices indeed to use one of the
Ga
Gb
d+
EM
2
2
2
Ga
Gb
Operations: (v)(v)(e)
(a)
(b)
Figure 8. (a) represents a rigid merged graph obtained by performing three
operations (v). It is shown in (b) how directions can be given to the edges in
such a way that the three vertices of Ga are each left by 2 edges.
d+
EM 3
1
1
1
d+
EM 2
1
1
1
1
Ga
Gb
Ga
Gb
Operations: (e) from (3,2,1)
Operations: (e) from (3,1,1,1)
d+
EM
2
2
1
1
d+
EM
1
1
1
1
1
1
Ga
Gb
Ga
Gb
Operations: (e) from (2,2,2)
Operations: (e) from (2,1,1,1,1)
Figure 9. Representations of how a rigid graph can be obtained by merging two
persistent graphs Ga and Gb where Gb has no DOF and where the DOF allocation
of Ga is (3, 1, 1, 1), (2, 2, 1, 1), (2, 1, 1, 1, 1) or (1, 1, 1, 1, 1, 1). The operations (v)
and (e) used to obtain the structure are also mentioned.
constructions already provided by temporarily adding one vertex with one DOF to the distribu-
tion of 5 DOFs in Ga , executing the appropriate construction from the group above, and then
reversing the direction of the edge leaving a vertex with one DOF, as shown in Fig. 10 for a
DOF distribution (3, 2).
Suppose now that Ga has 4 DOFs, and Gb 2 DOFs. Then the possible DOF distribution for
Ga are (3, 1), (2, 2), (2, 1, 1) and (1, 1, 1, 1). For Gb , they are (2) and (1, 1). The construction
proving the result for these eight cases are shown in Fig. 11.
Finally, if both graphs have 3 DOFs, the possible distribution for each are (3), (2, 1) and
(1, 1, 1). The case where they both have a distribution (3) does not satisfy the hypotheses of
this Proposition, and three other cases do not need to be treated for symmetry reasons. The
d+
EM
3
2
1
d+
EM
3
2
Ga
Gb
Ga
Gb
d+
EM
1
Figure 10. Representation of how the construction for a DOF partition 5-1
between Ga and Gb can be obtained from a construction for a partition 6-0.
d+
EM
3
1
Ga
d+
EM
3
1
Ga
d+
EM
2
2
Ga
Gb
d+
EM
Ga
Gb
d+
EM
Gb
d+
EM
2
2
2
d+
EM
2
Ga
1
1
Gb
d+
EM
1
1
1
1
2
Gb
d+
EM
d+
EM
2
1
Ga
2
1
Gb
d+
EM
1
1
d+
EM
1
1
1
1
d+
EM 1
1
1
1
Ga
Gb
d+
EM
2
Ga
Gb
d+
EM
1
1
Figure 11. Constructions for the eight possible DOF allocations when Ga has 4
DOFs and Gb 2 DOFs. The graphs are all rigid are they have the same undirected
underlying graphs as construction in Fig. 7, 8 or 9.
construction for the remaining 5 cases is shown in Fig. 12.
We now suppose that at least one of the graphs has less than 3 vertices, and show that a
rigid graph can be obtained by adding directed edges leaving vertices with local DOFs, the
number of these edges being provided in Table 1. Observe that a graph consisting of one single
vertex always has 3 DOFs, and thus that it is never needed to use any DOF of the other graph.
Similarly, each vertex of a graph containing two vertices has at least 2 DOFs, so that at most
one DOF of the other graph needs to be used, and only when the other graph has three or
more vertices. Fig. 13 shows how these mergings can be performed. Note that the rigidity of
d+
EM
1
2
Ga
2
1
d+
EM
3
Ga
Ga
Gb
Gb
d+
EM
1
2
d+
EM
1
1
1
Gb
d+
EM
1
2
d+
EM
1
1
1
d+
EM
1
1
1
Ga
Gb
Gb
d+
EM
1
1
1
Ga
d+
EM
3
Figure 12. Constructions for the five different DOF allocations satisfying the
hypothesis of Proposition 8 when each of Ga and Gb has 3 DOFs. The graphs
are all rigid are they have the same undirected underlying graphs as construction
in Fig. 7, 8 or 9 or rotated versions of them.
d+
EM
= 1
d+
EM
= 2
Ga
Ga
d+
EM 2
Ga
2
Gb
Gb
Gb
d+
EM
= 3
Ga
2
d+
EM
Ga
2
Gb
Gb
d+
EM
= 1
Figure 13. Illustration of the merging between two graphs, one of which at least
has less than 3 vertices. The dashed line represent the internal edge(s) of graphs
with two vertices, the orientation of which is not relevant for our purpose. The
vertex count in Gb is precisely 1,2 and 2 for the first three and a minimum of 3
for the last two.
the three first graphs is immediate as they are complete graphs. The rigidity of the other two
follows from the fact that they can be obtained from Gb by performing one of two operations
(v), which guarantees the rigidity of the graph obtained [22].
|
1210.7422 | 1 | 1210 | 2012-10-28T08:08:15 | Sensor networks security based on sensitive robots agents. A conceptual model | [
"cs.MA"
] | Multi-agent systems are currently applied to solve complex problems. The security of networks is an eloquent example of a complex and difficult problem. A new model-concept Hybrid Sensitive Robot Metaheuristic for Intrusion Detection is introduced in the current paper. The proposed technique could be used with machine learning based intrusion detection techniques. The new model uses the reaction of virtual sensitive robots to different stigmergic variables in order to keep the tracks of the intruders when securing a sensor network. | cs.MA | cs | Sensor networks security based on sensitive robots agents. A conceptual model
Camelia-M. Pinteaa, Petrica C. Popa
aTechnical University Cluj Napoca, North University Center Baia Mare, Romania
2
1
0
2
t
c
O
8
2
]
A
M
.
s
c
[
1
v
2
2
4
7
.
0
1
2
1
:
v
i
X
r
a
Abstract
Multi-agent systems are currently applied to solve complex problems. The security of networks is an eloquent example of
a complex and difficult problem. A new model-concept Hybrid Sensitive Robot Metaheuristic for Intrusion Detection is
introduced in the current paper. The proposed technique could be used with machine learning based intrusion detection
techniques. The new model uses the reaction of virtual sensitive robots to different stigmergic variables in order to keep
the tracks of the intruders when securing a sensor network.
Keywords:
intrusion detection, sensor network, intelligent agents
1. Introduction
Prevention and detection of intruders in a secure net-
work is nowadays a challenging issue. The intrusion detec-
tion system based on computational intelligence (CI) has
proved in time to have huge advantages over traditional
detection systems due to characteristics of CI methods:
adaptation, fault tolerance, high computational speed etc.
It is essential to design efficient Intrusion Detection Sys-
tems (IDS) especially for open medium networks as wire-
less sensor devices.
The intrusions could be missue intrusions and anomaly
intrusions. Missue intrusions are the attacks knowing the
weak points of a system. Anomaly intrusions are based
on observations of normal system usage patterns and de-
tecting deviations from the given norm. The mentioned
intrusions are hard to quantify because there are no fixed
patterns that can be monitored and as a result a more
fuzzy approach is often required.
The Intrusion Preventing Systems (IPS) are network
security appliances that monitor network and/or system
activities for malicious activities.
IPS is a device used
to block all the unwanted access to the targeted host, to
remove malicious part of packets and as well it may re-
configure the network device where an attack is detected
[3].
Social autonomic cooperative colonies as ants, bees and
others have the capability to coordinate and construct
complex systems [4]. Using their behavior, engineers have
built real collective robotic systems. The metaheuristics
based on classes of specialized robots provide feasible solu-
tions for nowadays complex problems. One of these tech-
niques is Sensitive Robot Metaheuristic developed by Pin-
tea et al. [19, 21]. The sensitive model was introduced and
Email addresses: [email protected] (Camelia-M. Pintea),
[email protected] (Petrica C. Pop)
explained in [5, 6, 21] and used to solve complex problems
in [7, 20, 21]. The SRM model was implemented first to
solve a large drilling problem but it has the potential to
solve other NP-hard problems including intrusion detec-
tion. The model ensure a balance between diversification
and intensification in searching.
The aim of the current paper is to provide an effective
stigmergic-based technique for IDS in a sensor network
graph, that consist of multiple detection stations called
sensor nodes. The new Hybrid Sensitive Robot Metaheuris-
tic for Intrusion Detection (HSRM-ID) model uses a col-
lection of robots endowed with a stigmergic sensitivity
level. The sensitivity of robots allow them to detect and
react to different stigmergic variables involving the attacks
into a secure network. The hybrid model combines ele-
ments from Sensitive Robot Metaheuristic (SRM) [19] as
Ant Colony System (ACS) [10], autonomous mobile robots
and the intrusion detection based on emotional ants for
sensors (IDEAS) [2].
2. Sensitive Stigmergic Robots
The metaheuristic Sensitive Robot Metaheuristic (SRM)
[19] combining the concepts of stigmergic communication
and autonomous robot search is used to solve NP-hard
optimization problems. The basic concepts are defined
and described further in this section, see for more details
[4, 5, 6, 7, 19, 20, 21].
Definition 1. Stigmergy occurs when an action of an in-
sect is determined or influenced by the consequences of the
previous action of another insect.
Definition 2. Sensitive robots refers to artificial entities
with a Stigmergic Sensitivity Level (SSL) expressed by a
real number in the unit interval [0, 1].
Definition 3. Environment explorers'robots are sensitive
robots with small Stigmergic Sensitivity Level (sSSL) with
the potential to autonomously discover new promising re-
gions of the search space.
Definition 8. Intrusion Prevention System (IPS) is ac-
tive, in-line device in the network that can drop packets
or stop malicious connection before reaching the targeted
system.
Definition 4. Environment exploiters robots are sensitive
robots with high Stigmergic Sensitivity Level (hSSL) em-
phasizing search intensification.
An important characteristic of stigmery is that indi-
vidual behavior modifies the environment, which in turn
modifies the behavior of other individuals [11]. The SRM
technique attempts to address the coupling between per-
ception and action as direct as possible in an intelligent
stigmergic manner.
As it is known, robot communication relies on local en-
vironmental modifications that can trigger specific actions.
The set of the rules defining actions (stimuli pairs) used by
a homogeneous group of stigmergic robots defines their be-
havior and determines the type of structure the robots will
create [4, 26]. Robot stigmergic communication does not
rely on chemical deposition as it is for artificial ant-based
colonies [10]. A stigmergic robot action is determined by
the environmental modifications caused by prior actions of
other robots. The value of quantitative stigmergy modify
the future actions of robots. Discrete stimulus are involved
in qualitative stigmergy and the action is switched to a dif-
ferent action [4, 26].
Some real-life applications of the behavior-based ap-
proach, including autonomous robots, are in data min-
ing, military applications, industry and agriculture, waste
management, health care.
3. Intrusion detection techniques using Artificial
Intelligence
At first are introduced the main concepts of IDS fol-
lowed by a survey of Artificial Intelligence-based existing
models for computer security.
3.1. Intrusion Detection System
Due to increasing incidents of computer attacks, it is
essential to build efficient intrusion detection mechanisms.
The definitions of the main concepts related to this domain
are given in what it follows, see for example [8, 13].
Definition 5. Intrusion detection technology is a technol-
ogy designed to monitor computer activities for the purpose
of finding security violations.
Definition 6. Intrusion detection system (IDS) is a sys-
tem that implements intrusion detection technology.
Definition 7. A security violation of a system is any de-
liberate activity that is not wanted including denial of ser-
vice attacks, port scans, gaining of system administrator
access and exploiting system security holes.
IPS is able to detect and prevent attacks but it has
not deeper detection capabilities of IDS. Neither of Intru-
sion Detecting System and Intrusion Prevention System
is capable to provide in depth security. Intrusion Detect-
ing and Prevention System (IDPS), a combinations of IDS
and IPS, is a more effective system capable of detection
and prevention [22]. Based on the placement, the IDPS is
divided into four classes as follows:
1. a network-based system, which is able to monitor
traffic of network or its particular segment and iden-
tify different network attacks.
An example of network-based system is Snort [14].
Snort is an open source network intrusion prevention
and detection system - nowadays a standard for IPS -
that combines the benefits of signature, protocol and
anomaly-based inspection. A number of problems
associated with Network-based system according to
[17] are:
-- they cannot fully detect novel attacks;
-- variations of known attacks are not fully de-
tected;
-- they generate a large amount of alerts, as well
as a large number of false alerts;
-- the existing IDS is focus on low-level attacks or
anomalies and do not identify logical steps or
strategies behind these attacks.
2. host-based systems describe the class of software able
to monitor a single system, analyse characteristics
and log to at one host. These systems are deployed
on critical hosts.
3. wireless-based systems analyse wireless traffic to mon-
itor intrusion or any suspicious activity. They scan
traffic but are not able to identify attack in the ap-
plication layer or higher layer network protocols as
UDP and TCP. It may be deployed at the point
where unauthorized wireless network could be ac-
cessed.
4. behavior-based systems are used for examining net-
work traffic in order to identify attacks (e.g. Denial
of Service attacks). These systems are deployed to
monitor flow of network or flow between internal and
external network.
3.2. Artificial Intelligence in Intrusion Detection System
The current paper deals with an artificial intelligent
approach for intrusion detections. A short review of the
main AI techniques already used and their benefits for
detecting intrusion in network systems follows.
2
According to Beg et al.
[3], the intrusion detection
classical algorithms have the following disadvantages: false
alarm rate and constant updates of database with new sig-
natures. The network administrator responds to alarms
and updates the signatures that increases in time. For
example, in the already mentioned Snort signatures in-
creased from 1500 to 2800 over two years [14]. In order
to improve the administrator work, reducing the number
of false alarms and better intrusion detection are intro-
duced artificial intelligence mechanisms [23]. Some of AI
techniques used in intrusion detection are data mining,
genetic algorithm, neural network, multi-agents, ant-net
miner, etc.
Lee et al. [15] introduced a data mining classification
mechanism with association rules from the audit data -
knowledge present in a knowledge base - providing gaudi-
ness for data gathering and feature selection. In order to
detect abnormal behavior one can use genetic algorithms,
see for example [1]. In [18], neural networks use back prop-
agation MLP for a small network in order to detect anoma-
lies and identify user profiles after end of each log session.
It shall also be remarked that several of the leading
methods for detecting intrusions and detecting intrusions
are hybrid artificial approaches, which combine different
AI solution techniques [9, 16, 25]. Some hybrid meth-
ods used in the literature are data mining and fuzzy logic
techniques [16], data mining and genetic algorithm select-
ing the best rules for the system [9]. In the future could
be implemented hybrid models involving intelligent evolu-
tionary agents [12] and dynamic decision boundary using
Support Vector Machine [24] for handle a large number of
features.
Banerjee et al.
[2] introduced an intrusion detection
based on emotional ants for sensors (IDEAS), which could
keep track of the intruder trials. This technique is able to
work in conjunction with the conventional machine learn-
ing based intrusion detection techniques to secure the sen-
sor networks.
4. Hybrid Sensitive Robot Metaheuristic for In-
trusion Detection
In this section we introduce a new hybrid metaheuristic
in order to detect the intruders in a sensor network. The
new model is called Hybrid Sensitive Robot Metaheuristic
for Intrusion Detection (HSRM-ID), is based on Sensitive
Robot Metaheuristic (SRM) introduced in [19] and uses a
specific rule in order to generate a state of thinking or the
choice of an intruder [2].
The proposed (HSRM) can be modelled using two dis-
tinct groups of sensitive stigmergic robots. The first group
of robots-agents is endowed with small sensitive values SSL
and they are sensitive-explorers (sSSL: small SSL-robots).
They can sustain diversification in intruders searching. In
the second group are the robots-agents with high sensi-
tive stigmergic values (hSSL: high SSL-robots). They are
sensitive-exploiters and could exploit intensively the re-
gions already identified with attacks from intruders.
In
time, based on the experience of robots-agents, the sensi-
tive stigmergic level SSL can increase or decrease.
The pseudo-code description of the Hybrid Sensitive
Robot Metaheuristic for Intrusion Detection is described
in what it follows.
Algorithm 1 Hybrid Sensitive Robot Algorithm for In-
trusion Detection
Set parameters; initialize stigmergic values of the trails;
for k=1 to m do
Place robot k on a randomly chosen node of a sensor
network;
for i=1 to Niter do
Each robot incrementally builds a solution based on
the autonomous search sensitivity;
The sSSL robots choose the next node based on the
attack probability (1);
A hSSL-robot uses the information supplied by the
sSSL robots to chose the new node (2);
Apply a local stigmergic updating rule (3);
Apply the rule generating a state of thinking or the
choice of an intruder (4):
A global updating rule is applied (5);
Validate the path and detect intruder;
end for
end for
The stigmergic value of an edge is τ and the visibility
value is η. A tabu list with the already visited nodes is
maintained, see [10] for more details. In order to divide
the colony of m robots in two groups it is used a random
variable uniformly distributed over [0, 1].
Let q be a realization of this random variable and q0 a
constant 0 ≤ q0 ≤ 1. If the inequality q > q0 stands the
robots are endowed with small sensitive stigmergic value
sSSL robots and otherwise they are highly sensitive stig-
mergic robots (hSSL). A hSSL-robot uses the information
supplied by the sSSL robots.
In order to define the rule to generate a state of think-
ing or the choice of an intruder we use the same notations
as in Banerjee et al. [2]:
• A(I, s, t) denotes the tendency of an intruder I to be
assigned to the sensor node s at moment t.
• I1(intruder1) C(I, s, t) is the potential to generate
the state of choice to a particular path in the network
sensor graph.
• I C(I, s, t) is the intensity of the attack,
• f C(.) is a function specific of the thinking of in-
truder
• T c(I, t) is the threshold value.
The new hybrid model (HSRM-ID) for identifying the
affected path of a sensor network graph is described fur-
ther.
3
• Initially the SSL robots are placed randomly in the
network space. The parameters of the algorithm are
initialized.
• A SSL robot chooses the next move with a proba-
bility based on the distance to the candidate node
and the stigmergic intensity on the connecting edge.
In order to stop the stigmergic intensity increasing
unbounded each time unit evaporation takes place.
• Let i be the current node. The next node is chosen
probabilistically. Let J k
i be the unvisited successors
of node i by robot k and u ∈ J k
i. As in Ant Colony
System technique [10] the probability of choosing the
next node u, possible to be attacked, is shown in (1).
pk
iu(t) =
[τiu(t)][ηiu(t)]β
Σo∈J k
i[τio(t)][ηio(t)]β
,
(1)
where β is a positive parameter, τiu(t) is the stigmer-
gic intensity and ηiu(t) is the inverse of the distance
on edge (i, u) at moment t.
• The new node j is choose by hSSL robots using (2):
j = argmaxu∈J k
i
{τiu(t)[ηiu(t)]β},
(2)
where β determines the relative importance of stig-
mergy versus heuristic information.
• Update trail stigmergic intensity by local stigmergic
rule (3):
τij (t + 1) = q2
0τij (t) + (1 − q0)2 · τ0.
(3)
where (i, j) are the edges belonging to the most suc-
cessful traversing across sensor nodes.
• Equation (4) illustrates the rule to generate a state
of thinking or the choice of an intruder [2].
If I C(I, s, t) = I C(I, s, t) − T C(I, t)
then l C(l, s, t) > I C(l, t)
else I C(I, s, t) = 0.
(4)
• A global updating rule is applied [2] as in (5) and is
used a tabu list where to store the track and edge
details.
τij(t + 1) = q2
0τij (t) + (1 − q0)2 ·
k
Xj=1
∆sj τij(t), (5)
where
∆sjtij = (cid:26) f (sj)
0
if sj contributes to τij
otherwise
(6)
and where q0 is the evaporation rate, ∆sjτij is the
combination of a solution sj with the update for
pheromone value τ ij; f (sj) is the function specific
to the thinking of the intruder and k is the number
of solution used for updating the pheromones.
4
Table 1: Analyze the action of agents-robots based on the pheromone
level on the edges of the sensor network graph.
Agents
Intruders Pheromone Detecting
intrusion
searching
Level
Action
Type
type
sSSL
robots
explorers
hSSL
robots
exploiters
low
high
low
high
no
continue
to explore
possibly
intruders
the attack
is not
certified
notify
hSSL
robots
update
pheromone
trails
attack
is highly
present
identify
affected
path
• Update the intensity of attack value I C(I, s, t) through
validating the path and detect intruder.
The output of the algorithm is the most affected path of
a sensor network with n nodes. Termination criteria is
given by the number of iterations, denoted by Niter. The
complexity of the proposed algorithm is O(n2 · m · Niter).
5. The analyze of the new concept
In the following is performed an analyze of the Hybrid
Sensitive Robot Algorithm for Intrusion Detection. The
artificial pheromone from the edges of the sensor network
graph reveals as the attacked zone within the network.
Each bio-inspired robot uses his one specific properties as
his level of sensitivity in order to detect the intruders and
the artificial stigmergy in order to find the attacked edges.
Table 1 illustrates the behavior of different groups of sensi-
tive bio-inspired virtual robots when investigate the sensor
network in search of intrusion. As a concept, the intro-
duced model Hybrid Sensitive Robot Algorithm for Intru-
sion Detection has more chances to improve the intrusion
detection systems comparing with the existing approaches
from the literature, due to the sensitivity property of the
bio-inspired robots. As well the diversity of robots groups
implies also different values of virtual pheromone trail val-
ues. The robots with small stigmergic value are constantly
sustaining diversification in intruders searching and as a
complementary action, the robots with high sensitive stig-
mergic values are testing the already identified networks
attacked regions. In the future we will perform numeri-
cal experiments to assess the performance of the proposed
algorithm.
[15] W. Lee, S. Stolfo and K. Mok. Mining audit data to build intru-
sion detection models. Knowledge Discovery and Data Mining,
New York, AAAI Press, 66 -- 72 (1998)
[16] J. Luo. Integrating Fuzzy Logic and Data Mining Methods for
Intrusion detection. Master thesis, Mississippi State University,
1999.
[17] S. Northcutt. Network Intrusion Detection. New Riders Pub-
lishers, 2002.
[18] J. Ryan, M-J. Lin and R. Miikkulainen.
Intrusion Detection
with Neural Networks. Advances in Neural Information Pro-
cessing Systems 10, MIT Press, 1998.
[19] C-M. Pintea, C. Chira, D. Dumitrescu and P.C. Pop. A sen-
sitive metaheuristic for solving a large optimization problem.
SOFSEM 2008, LNCS 4910, 551 -- 559, 2008.
[20] C-M. Pintea, C. Chira and D. Dumitrescu. Sensitive ants: In-
ducing diversity in colony. NICSO 2008, Studies in Computa-
tional Intelligence, 236:15 -- 24, 2009.
[21] C-M. Pintea. Combinatorial optimization with bio-inspired
computing. PhD Thesis, Babes-Bolyai University, EduSoft Pub-
lisher, 2010.
[22] K. Scarfone and P. Mell.
tection
http://csrc.nist.gov/publications/nistpubs/800-94/SP800-94.pdf
Prevention
and
Guide
Systems
to Intrusion De-
2007).
(IDPS),
[23] S. Selvakani and R. S. Rajes. Genetic algorithm for forming
rules for intrusion detection. Int. J. of Computer Science and
Network Security 7(11):285 -- 290, 2007.
[24] R. Stoean, M. Preuss, C. Stoean, E. El-Darzi and D. Du-
mitrescu. Support Vector Machine Learning with an Evolution-
ary Engine. J. Operational Research Society 60(8):1116 -- 1122,
2009.
[25] J. T. Yao, S. L. Zhao and L. V. Saxton. A study on fuzzy
intrusion detection, SPIE, Data Mining, Intrusion Detection.
Information Assurance and Data Networks Security, 5812:23 --
30, 2005.
[26] T. White.
Expert Assessment
Report
http://www.scs.carleton.ca/~arpwhite/stigmergy-report.pdf
the Department
for
of Stigmergy.
A
of National Defence.
6. Conclusions
Nowadays the networks are threatened by security at-
tacks and resource limitations. In order to deal with this
security network problem efficient intruders detection and
prevention systems are used. Within this paper we in-
troduce a new concept Hybrid Sensitive Robot Algorithm
for Intrusion Detection based on bio-inspired robots.
It
is used a qualitative stigmergic mechanism, each robot is
endowed with a stigmergic sensitivity level facilitating the
exploration and exploitation of the search space. In the
future some computational tests will be proposed and fur-
ther hybrid AI techniques will be involved for securing the
networks.
Acknowledgement.
This research is supported by Grant PN II TE 113/2011,
New hybrid metaheuristics for solving complex network
design problems, funded by CNCS Romania.
References
[1] L. Alhazzaa. Intrusion Detection Systems using Genetic Algo-
rithms. 2007.
[2] S. Banerjee, C. Grosan and A. Abraham. IDEAS: Intrusion De-
tection based on Emotional Ants for Sensors. Intelligent Sys-
tems Design and Applications. IEEE C.S. 344 -- 349, 2005.
[3] S. Beg, U. Naru, M. Ashraf and S. Mohsin. Feasibility of Intru-
sion Detection System with High Performance Computing: A
Survey. Int. J. for Advances in Computer Science. 1(1):26 -- 35,
2010.
[4] E. Bonabeau, M. Dorigo and G. Tehraulaz. Swarm intelligence
from natural to artificial systems. Oxford, UK: Oxford Univ.
Press, 1999.
[5] C. Chira, C-M. Pintea and D. Dumitrescu. Sensitive stigmergic
agent systems: a hybrid approach to combinatorial optimiza-
tion.
Innovations in Hybrid Intelligent Systems, Advances in
Soft Computing, 44:33 -- 39, 2008.
[6] C. Chira, C-M. Pintea and D. Dumitrescu. Cooperative learning
sensitive agent system for combinatorial optimization. NICSO
2007, Studies in Computational Intelligence, 129:347 -- 355, 2008.
[7] C. Chira, D. Dumitrescu and C-M. Pintea. Learning sensitive
stigmergic agents for solving complex problems. Computing and
Informatics, 29(3):337 -- 356, 2010.
[8] T. Crothers. Implementing Intrusion Detection Systems, Wiley,
2003.
[9] Y. Dhanalakshmi and I. R. Babu. Intrusion detection using data
mining along fuzzy logic and genetic algorithms. Int. J. of Com-
puter Science and Network Security, 8(2):27 -- 32, 2008.
[10] M. Dorigo and L. M. Gambardella. Ant Colony System:A coop-
erative learning approach to the Traveling Salesman Problem.
IEEE Trans.Evol.Comp. 1:53 -- 66, 1997.
[11] Grass´e, P.-P. La Reconstruction du Nid et Les Coordinations In-
terindividuelles Chez Bellicositermes Natalensis et Cubitermes.
Insect Soc. 6:41 -- 80, 1959.
[12] B. Iantovics and C. Enachescu. Intelligent Complex Evolution-
ary Agent-based Systems. Development of Intelligent and Com-
plex Systems. AIP, 116-124, 2009.
[13] N. Ierace, C. Urrutia and R. Bassett. Intrusion Prevention Sys-
tems. Ubiquity, ACM, 2 -- 2, 2005.
[14] Kim, B., Yoon, S., Oh, J. Multi-hash based Pattern Matching
Mechanism for High-Performance Intrusion Detection. Int. J.
of Computers 1(3):115 -- 124, 2009.
5
|
1511.07349 | 1 | 1511 | 2015-11-23T18:17:05 | MOS-2: A Two-Dimension Space for Positioning MAS Organizational Models | [
"cs.MA"
] | The increased complexity and dynamism of present and future Multi-Agent Systems (MAS) enforce the need for considering both of their static (design-time) and the dynamic (run-time) aspects. A type of balance between the two aspects can definitely give better results related to system stability and adaptivity. MAS organization is the research area that is concerned with these issues and it is currently a very active and interesting research area. Designing a MAS with an initial organization and giving it the ability to dynamically reorganize to adapt the dynamic changes of its unpredictable and uncertain environment, is the feasible way to survive and to run effectively. Normally, MAS organization is tackled by what is called, MAS organizational models, which are concerned with the description (formally or informally) of the structural and dynamical aspects of agent organizations. This paper proposes a two-dimension space, called MOS-2, for positioning and assessing MAS organizational models based on two dimensions: their adopted engineering viewpoint (agent-centered or organization-centered) as the vertical dimension and the agents awareness/unawareness of the existence of the organizational level as the horizontal dimension. The MOS-2 space is applied for positioning a number of familiar organizational models. Its future trends and possible improvements are highlighted. They include the following, (1) adding Time as a dimension, (2) increasing the considered dimensions, (3) providing a quantitative approach for positioning MAS organizational models. | cs.MA | cs | MOS-2: A Two-Dimension Space for Positioning MAS
Organizational Models
1
Department of Electrical Engineering, Assiut University, Assiut, Egypt,
2
Department of Computer Engineering, Cairo University, Giza, Egypt,
1*
Hosny Abbas
2
, Samir Shaheen
Email: [email protected]
Email: [email protected]
Abstract- The increased complexity and dynamism of present and future Multi-Agent Systems (MAS) enforce the need for
considering both of their static (design-time) and the dynamic (run-time) aspects. A type of balance between the two aspects can
definitely give better results related to system stability and adaptivity. MAS organization is the research area that is concerned with
these issues and it is currently a very active and interesting research area. Designing a MAS with an initial organization and giving it
the ability to dynamically reorganize to adapt the dynamic changes of its unpredictable and uncertain environment, is the feasible way
to survive and to run effectively. Normally, MAS organization is tackled by what is called, MAS organizational models, which are
concerned with the description (formally or informally) of the structural and dynamical aspects of agent organizations. This paper
proposes a two-dimension space, called MOS-2, for positioning and assessing MAS organizational models based on two dimensions:
their adopted engineering viewpoint (agent-centered or organization-centered) as the vertical dimension and the agents
awareness/unawareness of the existence of the organizational level as the horizontal dimension. The MOS-2 space is applied for
positioning a number of familiar organizational models. Its future trends and possible improvements are highlighted. They include the
following, (1) adding Time as a dimension, (2) increasing the considered dimensions, (3) providing a quantitative approach for
positioning MAS organizational models.
Keywords - multi-agent systems (MAS), MAS organization, dynamic reorganization, organizational models, stability, adaptivity
1. INTRODUCTION
In contrast to initial MAS research, which concerned individual agents’ aspects such as agents’ architectures, agents’ mental
capabilities, behaviors, etc, the current research trend of MAS is actively interested in the adaptivity, environment, openness and the
dynamics of these systems. In open environments, agents must be able to adapt towards the most appropriate organizations according
to the state of the environment, which changes in an unpredictable manner. MAS organization [1] is currently considered as an
emergent area of MAS research that relies on the notion of openness and heterogeneity of MAS and imposes new demands on
traditional MAS models [2]. Considering MAS with no real structure isn’t suitable for handling current software systems complexity,
and higher order abstractions should be used and some way of structuring the society is typically needed to reduce system complexity,
to increase system efficiency, and to more accurately model the problem being tackled [3].
Horling et al. [4] stated that our real world getting more complex and highly distributed and that should be reflected in new
software engineering paradigms such as MAS. Therefore, the adoption of higher order abstract concepts like organizations, societies,
communities, and groups of agents can reduce complexity, increase efficiency, and improve system scalability. Shehory [5] defined
MAS organization as the way in which multiple agents are organized to form a multi-agent system. The relationships and interactions
among the agents and specific roles of agents within the organizations are the focus of MAS organization. Dignum [26] pointed out
that MAS organization can be understood from two perspectives, (1) organization as a process, and (2) organization as an entity. The
first perspective considers agents organization as the process of organizing a set of individual agents, thus in this sense it is used to
refer to constraints (structures, norms and patterns) found in a social context that shape the actions and interactions of agents. On the
other hand, the second perspective considers agents organization as an entity in itself, with its own requirements and objectives and is
represented by (but not identical to) a group of agents. In fact, MAS organization demands the integration of both perspectives and
relies for a great extent on the notion of openness and heterogeneity of MAS.
There are two familiar viewpoints of MAS engineering, the first one is the agent-centered MAS (ACMAS) in which the focus is
given to individual agents. With this viewpoint, the designer concerns the local behaviors of agents and also their interactions without
concerning the global structure of the system. The global required function of the system is supposed to emerge as a result of the lower
level individual agents interactions. The agent-centered approach takes the agents as the “engine” for the system organization, and
agent organizations implicitly exist as observable emergent phenomena, which states a unified bottom-up and objective global view of
the pattern of cooperation between agents [6]. Ant colony [7] is a natural example of the ACMAS viewpoint, where there is no
organizational behavior and constraints are explicitly and directly defined inside the ants. The main idea is that the organization is the
result of the collective emergent behavior due to how agents act their individual behaviors and interact in a common shared and
dynamic environment. In ACMAS, the MAS organization is actually a process not an entity; there is a consensus to call this process as
self-organization [28][29][30].
The second viewpoint of MAS engineering is what is called organization-centered MAS (OCMAS) in which the structure of the
system is given a bigger attention through the explicit abstraction of agent organizations. With that approach, the designer designs the
entire organization and coordination patterns on the one hand, and the agents’ local behaviors on the other hand. It is considered as a
top-down approach because the organization abstraction imposes some rules or norms used by agents to coordinate their local
behaviors and interactions with other agents. In OCMAS, the MAS organization is actually an explicit entity not a process and to
distinguish it from the ACMAS approach, the change in system organization is often called dynamic reorganization [6][24], which is
a more general name than self-organization.
When a researcher proposes an approach to dynamically reorganize a multi-agent system to adapt environments’ changes, he
actually proposes what the MAS community agreed to call it as an organizational model [19]. MAS organizational models will play a
critical role in the development of future larger and more complex MAS. The main concern of organizational models is to describe the
structural and dynamical aspects of organizations [8]. They have proven to be a useful tool for the analysis and design of multi-agent
systems. Furthermore, they provide a framework to manage and engineer agent organizations, dynamic reorganization, self-
organization, emergence, and autonomy within MAS.
Picard et al. [6] added the agents’ awareness /unawareness of the existence of the organization structure as a dimension of the
organization modification process and he identified four cases, (1) the agents don’t represent the organization, although the observer
can see an emergent organization. In some sense, they are unaware that they are part of an organization, (2) each agent has an internal
and local representation of cooperation patterns which it follows when deciding what to do, this local representation is obtained either
by perception, communication or explicit reasoning, (3) the organization exists as a specified and formalized schema, made by a
designer but agents don’t know anything about it and even do not reason about it. They simply comply with it as if the organizational
constraints were hard-coded inside them, (4) agents have an explicit representation of the organization which has been defined, the
agents are able to reason about it and to use it in order to initiate cooperation with other agents in the system.
In this paper, we propose a two-dimension space inspired from a previous work of Picard et al. [6], for positioning and comparing
MAS organizational models. The proposed space is similar to a two-dimension Cartesian coordinate system, where the
ACMAS/OCMAS are represented by the vertical axis and the agents’ awareness/unawareness of the existence of the organizational
level are represented by the horizontal axis as shown in Figure 1. A MAS organizational model is represented as a point (small circle)
or an area (oval) as demonstrated in the figure and as will be explained later. A number of familiar organizational models are
positioned and compared using the proposed MAS organization space; called MOS-2, where MOS stands for MAS Organization
Space and the number 2 indicates that the space is a two dimension space.
Figure 1: The MOS-2 MAS Organization Space
The central circle shown in the MOS-2 space can be seen as the unity circle (its radius is one). By this way it will be possible to
precisely position an organizational model by determining the extent (i.e., percentage) to which it is an ACMAS or OCMAS approach
and also the extent to which the individual agents are aware or unaware about the organizational aspects. In other words, the MOS-2
space can be used quantitatively not just qualitatively; this point is left as a future work but will be highlighted later in this paper.
Furthermore, in the proposed space, the number of dimensions is two (Aware/Unaware, ACMAS/OCMAS), but it is also possible to
increase the considered dimensions by adding new comparison aspects to address a fine-grained classes of MAS organizational
models. In case the number of the considered comparison aspects (dimensions) is increased to be N, then the resultant space will be an
N dimension space and can be called, MOS-N. Actually, this paper proposes only the two dimension case (MOS-2); other versions of
higher dimensions are left as a future work.
The remaining of this paper is organized as follows: Section 2 explores the related work. Section 3 provides a background related
to the two adopted comparison aspects. Section 4 presents the proposed MOS-2 space for positioning MAS organizational models.
Section 5 demonstrates the applicability of MOS-2 on a number of familiar MAS organizational models and approaches. Section 6
MAS organization was and still a very active and interesting research point in MAS. Concepts like organization, dynamic
reorganization, self-organization, and emergence have attracted great attention in the last few years. The reason is related to the
increasing complexity and highly distribution of modern real-life applications. This paper is not aimed to explain in details these
concepts but interested readers can inspect their related references, for example [22][23][24][25][26][7][27]. The goal of this paper is
to propose a visual semi-formal method for positioning and comparing a broad band of MAS organizational models. Organizational
models are the tools to design methods, techniques, and approaches for managing the static and dynamic organization of MAS. Many
researchers provided valuable narrative surveys and reviews contain informal analysis and evaluation of MAS organizational models.
For instance, Picard et al. [6] aimed to study and propose a comprehensive view of how one could make multi-agent organizations
adapted to dynamics, openness and large-scale environment. The authors proposed an analysis grid of different MAS organization
approaches. The proposed grid has two dimensions: the vertical dimension identifies if the considered organization approach adopts
the ACMAS, OCMAS, viewpoints. The second dimension is concerned with the awareness/unawareness of the individual agents of
the organizational aspects. The authors claimed that the two dimensions of their grid are continuous, and it is completely possible to
identify approaches that are at the boundary of two categories. Our propose work is inspired from Picard et al. work but in a more
formal, visual and interesting way.
Alberola [18] provided, as a part of her PhD thesis, an analysis of how current reorganization approaches in MAS provide
support to agent designers in order to develop adaptive agent societies. She described in detail some of the most relevant existing
approaches, in order to show the advantages and limitations of each one. Alberola suggested that reorganization in agent societies can
be represented as a loop process composed by different phases: Monitor, Design, Selection, and Evaluation phases. She also studied a
number of familiar MAS organizational models by identifying the techniques adopted in each phase for each organizational model.
Alberola study is valuable and it benefits us a lot but it is informal and contains intensive information and that makes it difficult to be
captured by students and beginners.
V. Dignum [19] edited a handbook of research in MAS aimed to provide an overview of current work in agent organizations,
from several perspectives, and focus on different aspects of the organizational spectrum. The handbook explored a number of familiar
MAS organizational models and what makes it interesting is that the authors of the selected models wrote themselves the chapter that
tackled their model. From the other hand, the handbook did not provide a general comparison or any type of positioning of the
considered models.
Jensen et al. [20] investigated the agent-centered and organization-centered approaches to designing and implementing multi-
agent systems. The authors have developed and evaluated two teams of agents for a variant of the well-known Bomberman computer
game. One team is based on the basic Jason system, which is an implementation in Java of an extension of the logic-based agent-
oriented programming language AgentSpeak. The other team is based on the organizational model Moise+, which is combined with
Jason in the middleware called J-Moise+. They concluded that the agent-criented approach has a number of advantages when it comes
to game-like scenarios with just a few different character types.
Horling and Lesser [21] also stated that organizational design employed by an agent system can have a significant, quantitative
effect on its performance characteristics, and they surveyed the major organizational paradigms used in multi-agent systems. These
include hierarchies, holarchies, coalitions, teams, congregations, societies, federations, markets, and matrix organizations. Also, they
provided a description of each paradigm, and discuss its advantages and disadvantages, further, they provided examples of how each
organization paradigm may be instantiated and maintained. But their work was not targeted to organizational models, which concerns
both of static and dynamic aspects; they just concerned how to structure MAS with different paradigms.
In nutshell, the related work was valuable for us in designing the MOS-2 space for positioning MAS organizational models and
approaches in a visual, semi-formal, easy to understand way. Providing a 2-dimension space for identifying an organizational model
features and limitations in the scope of two or more dimensions is a good idea. It provides an effective tool to compare visually an
organizational model with other models. This way enables designers and beginners to quickly capture a certain model in their minds
and allows them to remember easily the considered model and its features or limitations relative to other models.
3. BACKGROUND
discuss the results and provides a complete view based on these results. Section 7 highlights the possible future trends. Finally,
Section 8 concludes this paper.
2. RELATED WORK
This paper is not concerned with the promotion of one MAS engineering viewpoint (ACMAS or OCMAS) relative to the other one,
but the main concern is to position MAS organizational models and show the extent to which a certain model benefits from the
adoption of each of these viewpoints. We have to emphasize here that both of the MAS engineering viewpoints (ACMAS/OCMAS)
are generally not mutually exclusive and have led to different approaches in the domain [6]. In other words, it is possible to mix both
viewpoints in one organizational model to take benefit of their pros and avoid their cons. Table 2 provides a comparison between the
two viewpoints by presenting the characteristics and the shortcomings of each one. Also, Figure 2 provides the advantages and
disadvantages of the adoption of each viewpoint relative to the individual agents’ awareness or unawareness of the higher level
organizational aspects.
Figure 2 represents the basis of the proposed two-dimension space as it orthogonally aligns the two comparison aspects into a
vertical (ACMAS/OCMAS) and a horizontal (Awareness/Unawareness) axes. As demonstrated in the figure, four quadrants are
resulted: ACMAS-Awareness, ACMAS-Unawareness, OCMAS-Unawareness, and OCMAS-Awareness. Therefore, a MAS
organizational model that is positioned in one of these quadrants will simultaneously benefit and suffer from the advantages and
disadvantages of this quadrant respectively. Note also that, an organizational model can be positioned into more than one quadrant.
Table 1: Comparison between the ACMAS and OCMAS viewpoints
Shortcomings
• Unpredictability and Uncertainty:
lead to unreliability
• Lack of Modularity: all agents are
accessible from everywhere
• Undesirable Emergent Behavior:
can impact system performance
• Dual Responsibility: agents have to
manage simultaneously both the
functional an the organizational
aspects
• Computational / Communication
• Reduce overall flexibility or
overhead
reactivity
• Add additional layer of complexity
ACMAS
OCMAS
Characteristics
• Organization is a process (self-
organization)
• Informal/Bottom-
up/Emergent/Endogenous
• The focus is given to individual agents
• Agents are the “engine” for the system
• An agent may communicate with any other
organization
agent
• An agent is responsible to define its
relations with other agents
• An agent is responsible to constrain its
accessibility from other agents
• Agents are autonomous and no constraint
is placed on the way they interact
• An agent provides a set of services
available to other agents
• Organization is often an explicit entity
• Support dynamic reorganization
• Formal/ To-down/ Pre-exist organization
• Reduce system complexity
• Increase system efficiency
• Improve system scalability
• Provide Effective coordination
• Limit the scope of interactions
• Tuning of the agents autonomy
• Structuring of agents interactions
• Separation of concerns
• Modularity/Reliability
4. THE PROPOSED MOS-2 SPACE
Figure 3 presents the proposed MOS-2 space. As shown in the figure, the axes of a two-dimensional Cartesian system divide the space
into four infinite regions, called quadrants, each bounded by two half-axes. In mathematics, these are often numbered from 1st to 4th
and denoted by Roman numerals, lets take the same naming conversion, and thus the four quadrants can be identified as follows:
1. The I symbol identifies the ACMAS-Aware space quadrant
2. The II symbol identifies the ACMAS-Unaware space quadrant
3. The III symbol identifies the OCMAS-Unaware space quadrant
4. The IV symbol identifies the OCMAS-Aware space quadrant
Figure 2: Advantages and Disadvantages of MOS-2 Quadrants
Therefore, to position a MAS organizational model it should be studied and explored to see to which quadrant in the MOS-2 space it is
best fit according to its characteristics and properties. If the considered model fits one of MOS-2 quadrants, then its position will be
represented by a small circle as shown in Figure 4-a. But, if the model fits two quadrants, then its position is represented by a small
oval shape expanded along the two space quadrants as demonstrated in Figure 4-b. Note that the oval part appeared in each quadrant
should be relative to the extent to which the model realizes the characterizes of the MAS organization class represented by that
quadrant. If the considered model realizes (partially or fully) the characteristics of three space quadrants then the position of this
model can be represented as a half-circle expanded along the three space quadrants as shown in Figure 4-c. Finally, in case the
considered model fits the whole space (it is rare but possible) then a circle expanded along the four space quadrants can be used to
represent that perfect organizational model!, this case is demonstrated in Figure 4-d. Note also that the MOS-2 space can be enlarged
to position simultaneously many organizational models. The next section applies the proposed space for positioning a number of
familiar MAS organizational models.
Figure 3: The MOS-2 space with the characteristics of each organization class
Figure 4: positioning of different types of organizational models
5. THE APPLICABILITY OF MOS-2 SPACE: CASE STUDIES
This section applies the MOS-2 space to position and compare a number of familiar MAS organizational models. The selected models
are: AGR, MACODO, MOISE, Swarm-based approaches, Contract Net coordination model, and Gaia development methodology.
Actually, there are a large number of MAS organizational models and approaches but we found that the selected ones are enough to
demonstrate the usage of the proposed space and we leave to the reader the mission of trying to position any other model found in the
literature or proposed by him.
AGR
The AGR [8][9] is a MAS organizational model that adopts the OCMAS viewpoint. This model is influenced by both AOSE and
social reasoning, in the sense that organization is used by designer to specify the system-to-be and by the agents that can dynamically
perform organizational acts and possibly modify the organization. The designer uses abstract concepts such as Group Structure and
Organizational Structure to specify application in design-time. The group structure is an abstract representation of the roles required in
this group and their interaction relationships and protocols. The organization structure is the set of group structures expressing the
design of a multi-agent organizational scheme. In run-time, the agents can reason on the organizational aspects and can modify the
application structure by the dynamic creation of agents groups (agents partitioning) and dynamic forming of hierarchies of groups.
Therefore, the AGR model fits well with the III and IV quadrants of MOS-2 space and can be positioned as shown in Figure 5-a.
Figure 5: Positioning of AGR and MACODO organizational models
MACODO
The MACODO [10][11] is a MAS organizational model that is really an interesting model because it tries to realize (although
partially) both of the MAS engineering viewpoints: ACMAS and OCMAS. It, to a large extent, belongs to the OCMAS philosophy,
which provides an explicit representation of agent organizations. Although it provides a formal predefined specification for system
dynamic reorganization, it allows the agents to (according to the environment context) form a type of short-lived coalitions by
presenting a cooperative behavior with each other. The agents’ cooperative behavior is supervised and controlled by the organization
controllers. The organizational model is part of an integrated approach, called MACODO (Middleware Architecture for COntext-
driven Dynamic agent Organizations); in this model, the life-cycle management of dynamic organizations is separated from the
agents, organizations are first-class citizens, and their dynamics are governed by laws. We see that, the MACODO organizational
model is best fit with two space quadrants, IV and I of the MOS-2 space, as presented in Figure 5-b.
MOISE
Hannoun et al. [12] proposed MOISE (Model of Organization for multI-agent SystEms); for modeling organizational aspects of MAS.
The MOISE model is aimed at providing support in order to adapt an agent organization to its environment and to help it to efficiently
achieve its goals. This model defines an organization which is composed by agents, roles, missions, and the deontic dimension. Each
role represents a set of constraints that an agent follows when it plays this role. These constraints represent the structure dimension
(relations between roles) and the functional dimension (missions, deontic dimension). A mission is a set of coherent goals that an
agent can commit to. The deontic dimension specifies the permissions and obligations of a role in a mission. MOISE adopts a
proactive reorganization carried out in a distributed way by monitor agents. The logic for reorganization is implemented at design time
and cannot be changed during runtime. But the agents can modify the MAS organization according to the predefined logic. For
example, the agents can change their roles or give a new obligation; a new role can be added to the system, etc. We see the MOISE
model similar to the AGR model and can be positioned in the same way through quadrants III and IV of the MOS-2 space, as shown
in Figure 6-a.
Figure 6: Positioning of MOISE and Swarm-based organizational models
Swarm-Based Approaches
The swarm-based approaches adopt the ACMAS viewpoint where the agents are unaware of any higher level structure. The system
organization is dynamic and informal, it is an emergent phenomena appears to the observer in the higher level as a result of the
individual agents lower level interactions directly (in a peer to peer fashion) or indirectly (through environment). In this type of agent
systems, the designer concerns only the individuals and the environments, and he doesn’t give any attention to the global
organizational level. When the designer develops an individual agent he put in his mind the application domain and its environment
only. In these systems, the individuals are purely reactive that simply react to environment changes. The ant colony [13] represents a
realistic natural example of these systems. In the ant colony there is no organizational behavior and constraints are explicitly and
directly defined inside the ants. The main idea is that the organization is the result of the collective emergent behavior due to how
agents act their individual behaviors and interact in a common shared and dynamic environment. This class of systems represents the
pure ACMAS viewpoint, which is located in the II quadrant in the MOS-2 space, as shown in Figure 6-b.
Contract Net
Contract Net Protocol (CNP) [14] is a task-sharing protocol in multi-agent systems, consisting of a collection of nodes or software
agents that form a purposeful coalition. The agents are pre-augmented by the designer with some social rules, interaction models, and
dependency models to be able to participate in purposeful coalitions. The organization is implicit and depends on the situation faced
by the agents. The agents may be not able to reason about the global system organization, but they just follow the predefined social
rules. The CNP model is best fit with the I-quadrant in the MOS-2 space, and can be represented by a small circle as shown in Figure
7-a.
Figure 7: Positioning of Contract Net and Gaia organizational models
GAIA
The Gaia [15] is an agent-oriented software engineering methodology (AOSE) [16]; it considers the system organization at the design-
time. Organizations are specified before encoding the agents. Agents can reason on the organization at run-time but cannot be able of
modifying it. In other words, In MAS that are modeled by the Gaia methodology, the system structure is defined in design-time and
doesn’t change in run-time (fixed structure). The Gaia-based MAS can be positioned in the III quadrant in the MOS-2 space, which
represents the AOSE engineering approaches, as shown in Figure 7-b.
NOSHAPE
The NOSHAPE MAS organizational model [17] is a recent model, although it is not matured yet, but it provides a novel approach for
engineering complex and highly distributed MAS. Like the MACODO model, the NOSHAPE model tries to adopt the two MAS
engineering viewpoints: the ACMAS and the OCMAS. The NOSHAPE model allows individual agents to loosely reshape the higher
level system organization by emitting triggers (triggers can be seen as the pheromones released in the environment by ants in the ant
colony as a type of indirect interaction). According to these triggers, the organizational level changes the system organization by
establishing overlap relationships among higher order entities (agents’ organizations, organizations’ worlds, and worlds’ universes).
There are no any constraints imposed on the individual agents (except for mobility). The relationship between the individual agents’
level and the organizational level is loose and depends only on the agents triggers. The organizational level can be seen as a helper or
a guide to the agents. So, in the NOSHAPE model, the system structure emerges as a result of the agents triggers which are managed
in a service-oriented manner, it is not possible to predict the next shape (structure) of the system, there are a pre-defined specification,
and agents can modify indirectly the whole system structure by just emitting triggers. The NOSHAPE model aims to provide
generality (relative to systems scale) and also aim to make the relation between the level of individual agents and the organizational
level loosely coupled, so the agents can behave according to the ACMAS viewpoint (where agents are unaware of the organizational
aspects) and the organizational level behaves according to the OCMAS viewpoint independently. Therefore, the NOSHAPE model
can be positioned along the quadrants II, III, IV of MOS-2 space, as shown in Figure 8.
Figure 8: Positioning of the NOSHAPE MAS organizational model
COMPLETE VIEW
All of the selected MAS organization models can be positioned and represented on one space diagram as shown in Figure 9. Based on
the complete view shown in Figure 9, we can claim that the more space quadrants an organizational model visits, the more it has of
features. In other words, it will possess the advantages of each space quadrants. Not only this, but also it will have the chance to match
the disadvantages of one space quadrants to the advantages of another one. According to these results, we claim that the NOSHAPE
model is a promising one because it visits three space quadrants.
6. FUTURE TRENDS
Figure 9: Complete view of positioning MAS organizational models
Increasing the considered dimensions
The proposed MOS-2 space can be evolved to precisely position and compare MAS organizational models and approaches by
considering the following issues:
1. Adding Time as a Dimension
2.
3. The quantitative Approach
In the following subsections, these issues will be highlighted and some suggestions and ideas will be proposed to be addressed in
future.
Adding Time as a Dimension
Based on the proposed MOS-2 space for MAS organization classification, is it possible to ask this question “can an organizational
model evolves dynamically and changes to a novel emergent model”? May be it seems like a pure fantastical idea but as we all know
“science is not about why? It’s about why not?” What this question means is to add a third dimension to the proposed space, it is
Time, so it may seem logical to call it MOS-2T, let’s illustrate this amazing idea! Consider the organization space shown in Figure 10-
a, as shown there is a model M0 positioned in the III quadrant (the AOSE quadrant), with the addition of time as a third dimension, lets
denote the absolute position of a model as a tuple <P, T> where P is the observed position of the considered model on the space and T
is the time where we observed the position, then:
Figure 10: Dynamic evolution of organizational models
At T0 the position of model M0 is <P0, T0>, we may write it like:
P(M0)= <P0, T0> where P: M (cid:214) P x T where P = {P0, P1, P2, ….Pk} or the set of possible positions. And M={M0, M1,M2,….Mn} is
the set of possible models (assume it is a finite set).
If at time T1 the model M0 changed to a new emergent model M1 (we can say that it is evolved, ignore questions like why? When?
How? At least for now) then:
P(M1)= <P1, T1>
And at time T2 also M1 evolved to M2 as follows
P(M2)=<P2,T2> As demonstrated in Figure 10-a.
Also let’s denote the organization of a MAS when the organization model M0 is active at T0 by
O(M0, T0)= M0.O0 where O: M x T (cid:214) M.O where M.O={M.O0, M.O1, …, M.Om} or the set of possible organizations under the
umbrella of a certain organizational model M.
So with M0 is active, a dynamic organization change can happen as follows:
M0.O0 (cid:198) M0.O1
That is normal and expected according to the specification of the organization model M0. But what is not normal is when the model
M0 is no longer active because it evolved to a novel emergent model:
M0 (cid:198) M1
As demonstrated in Figure 10-b, then it is expected to see a new organization of the system that is not planned by the designer and
may violate the designer specifications, that can only happen if the agents are intelligent and have learning capabilities, so they may
cause this amazing change in the organizational model as an emergent phenomena. As we see in Figure 10-b, after the dynamic
emergent change of the organizational model takes place, the system organization M0.O1 relative to M0 becomes M1.O0 relative the
emergent model M1, in other words M0.O1 becomes the initial organization of M1, and the new model (M1) will cause the system
organization to change to a novel organization that is not planned by the designer. Dynamic evolution of pre-defined organizational
models because of agents’ intelligence and their ability to learn is expected to be the next mainstream research area in MAS discipline.
Increasing the Considered Dimensions
The dimension of the proposed MOS-2 MAS organization space can be increased by adding other characteristics or properties of
MAS organizational models. For example, a new dimension to represent the extent to which the model tackles the inter-reorganization
and intra-reorganization can be added. The inter-reorganization is concerned with the organizational level interactions (i.e.,
interactions between groups of agents), and the intra-reorganization is concerned with individual agents interactions inside one group
of agents. In this case the organization space will be a 3-dimension space as shown in Figure 11. Also in this case the name of space
will be MOS-3 and if the time dimension is added, its name will be MOS-3T. The space shown in Figure 11 demonstrates how a
model can be positioned in MOS-3 by considering three defined dimensions (x, y, z) correspond to classification criteria of MAS
organizational models.
More dimensions can be added and the more dimensions added, the more fine-grained classification of MAS organizational
models can be achieved. Therefore, the general name of the proposed space can be written as MOS-NT, where N is the number of
used dimensions and T is the time dimension (the T should be removed if the time dimension is not considered).
Figure 11: A 3-dimension space for MAS organization
The Quantitative Approach
In the previous sections, we tackled MAS organizational models in a qualitative way. In fact, the qualitative positioning of MAS
organizational models is largely limited by the imagination of the researcher. In the qualitative approach, the researcher depends on
descriptions and observations, but sometimes it is better to provide a quantitative data based on rigorous measurements and
calculations. So, it is possible to quantify the position of a MAS organizational model by finding numerically the extent to which the
model realizes a certain dimension (i.e., ACMAS). Consider the MOS-2 space shown in Figure 12, what if we considered the central
circle shown in the space as the unity circle, so the maximum value of a dimension is 1 and the minimum value is -1 (i.e., - ACMAS=
OCMAS). Thus, when positioning a model, we should find a way to accurately calculate the extent to which the model can be
considered to realize a certain dimension.
Assume that there is a function V that is designed to calculate numerically the extent to which a model M is ACMAS/OCMAS. The
function V can have the following signature:
Figure 12: MOS-2 with a quantitative approach
V: M (cid:214) [-1,1]
Similarly assume that there is a function W that is designed to calculate numerically the extent to which the agents in a model M are
Aware/Unaware. The function W can have the following signature:
W:M (cid:214) [-1,1]
Therefore the quantitative position of the model M in MOS-2 space can be determined as follows:
P(M)=(V(M), W(M))
The example in Figure 12 shows that:
V(M)= -0.7 and W(M)= 0.6
So the position of the model M can be written as follows:
P(M) = (V(M), W(M)) = (-0.7, 0.6)
The problem now is how to implement the functions V and W. The complexity of these functions will increase as the number of
considered dimensions increases. Not only this but also the number of this functions will increase because each considered dimension
will need a function to quantify the extent to which this dimension is realized by the MAS organizational model. Moreover, their
implementation will be more complex if the time dimension is considered. This research trend is highlighted as future work.
7. CONCLUSIONS
This paper proposes a new method for positioning and comparing of MAS organizational models based on a two dimension space,
called MOS-2. MOS-2 uses two comparison aspects: the MAS engineering viewpoint (ACMAS/OCMAS) and the agents’
awareness/unawareness of the existence of a higher organizational level. The proposed space provides a feasible, effective, visual, and
semi-formal positioning method for comparing MAS organizational methods using an orthogonal coordinate system similar to the
familiar Cartesian coordinate system. The applicability of the proposed MAS organization space have been demonstrated by using it
to position and compare a number of familiar and recent MAS organizational models, other models can be positioned similarly. Future
trends are highlighted and discussed in Section 7.
REFERENCES
[1] Hosny Ahmed Abbas, Samir Ibrahim Shaheen, Mohammed Hussein Amin. Organization of Multi-Agent Systems: An Overview.
International Journal of Intelligent Information Systems. Vol. 4, No. 3, 2015, pp. 46-57. doi: 10.11648/j.ijiis.20150403.11
[2] Sichman, J. S., Dignum, V., & Castelfranchi, C. (2005). Agents' organizations: a concise overview. Journal of the Brazilian
[3] Jennings, N. R., & Wooldridge, M. Agent-Oriented Software Engineering. in Bradshaw, J. ed. Handbook of Agent Technology,
Computer Society, 11(1), 3-8.
AAAI/MIT Press, 2000.
[4] Horling, B., & Lesser, V. (2004). A survey of multi-agent organizational paradigms. The Knowledge Engineering Review, 19(4),
281–316.doi:10.1017/ S0269888905000317.
[5] Shehory, O. Architectural properties of multi-agent systems. Technical Report CMU-RI-TR-98-28, The Robotics Institute,
Carnegie Mellon University, Pittsburgh, Pennsylvania 15213, 1998.
[6] Picard, G., Hübner, J. F., Boissier, O., & Gleizes, M. P. (2009, June). Reorganisation and self-organisation in multi-agent
systems. In 1st International Workshop on Organizational Modeling, ORGMOD (pp. 66-80).
[7] A. Drogoul, B. Corbara, and S. Lalande. MANTA: New experimental results on the emergence of (artificial) ant societies. In
Nigel Gilbert and Rosaria Conte, editors, Artificial Societies: the Computer Simulation of Social Life, pages 119–221. UCL
Press, London, 1995.
[8] Ferber, J., Michel, F., & Báez, J. (2005). AGRE: Integrating environments with organizations. In Environments for multi-agent
systems (pp. 48-56). Springer Berlin Heidelberg.
[9] Ferber, J., & Gutknecht, O. (1998, July). A meta-model for the analysis and design of organizations in multi-agent systems.
In Multi Agent Systems, 1998. Proceedings. International Conference on (pp. 128-135). IEEE.
[10] Weyns, D., Haesevoets, R., & Helleboogh, A. (2010). The MACODO organization model for context-driven dynamic agent
organizations. ACM Transactions on Autonomous and Adaptive Systems (TAAS), 5(4), 16.
[11] Weyns, D., Haesevoets, R., Helleboogh, A., Holvoet, T., & Joosen, W. (2010). The MACODO middleware for context-driven
dynamic agent organizations.ACM Transactions on Autonomous and Adaptive Systems (TAAS), 5(1), 3.
[12] Hannoun, M., Boissier, O., Sichman, J. S., & Sayettat, C. (2000). MOISE: An organizational model for multi-agent systems. In
Advances in Artificial Intelligence (pp. 156-165). Springer Berlin Heidelberg.
[13] A. Drogoul, B. Corbara, and S. Lalande. MANTA: New experimental results on the emergence of (artificial) ant societies. In
Nigel Gilbert and Rosaria Conte, editors, Artificial Societies: the Computer Simulation of Social Life, pages 119–221. UCL
Press, London, 1995.
[14] Smith, R. G. (1980). The contract net protocol: High-level communication and control in a distributed problem solver. IEEE
Transactions on computers, (12), 1104-1113.
[15] Wooldridge, M., Jennings, N. R., & Kinny, D. (2000). The Gaia methodology for agent-oriented analysis and design.
Autonomous Agents and Multi-Agent Systems, 3(3), 285-312.
[16] Giorgini, P., & Henderson-Sellers, B. (2005). Agent-oriented methodologies: an introduction. Agent-oriented Methodologies, 1-
[17] Abbas, H. A. (2014). Exploiting the Overlapping of Higher Order Entities within Multi-Agent Systems. International Journal of
Agent Technologies and Systems (IJATS), 6(3), 32-57. doi:10.4018/ijats.2014070102.
[18] Alberola Oltra, J. M. (2013). Reorganization in Dynamic Agent Societies (Doctoral dissertation), CH 2.
[19] Dignum, V. (Ed.). (2009). Handbook of Research on Multi-Agent Systems: Semantics and Dynamics of Organizational Models:
Semantics and Dynamics of Organizational Models. IGI Global.
[20] Jensen, A. S., & Villadsen, J. (2013). A comparison of organization-centered and agent-centered multi-agent systems. Artificial
[21] Horling, B., & Lesser, V. (2004). A survey of multi-agent organizational paradigms. The Knowledge Engineering Review, 19(4),
Intelligence Research, 2(3), p59.
281–316.doi:10.1017/ S0269888905000317.
19.
30(1).
[22] Giovanna Di Marzo Serugendo et al (2011), “Self-organizing Software, From Natural to Artificial Adaptation”, Springer.
[23] Giovanna di Marzoserugendo et al, “Self-organization in multi-agent systems”, The Knowledge Engineering Review, Vol. 20:2,
165–189., 2005, Cambridge University Press.
[24] Dignum, V., Dignum, F., & Sonenberg, L. (2004, September). Towards dynamic reorganization of agent societies. In Proceedings
of Workshop on Coordination in Emergent Agent Societies at ECAI (pp. 22-27).
[25] Sichman, J. S., Dignum, V., & Castelfranchi, C. (2005). Agents' organizations: a concise overview. Journal of the Brazilian
[26] Dignum, V. (2009). The role of organization in agent systems. Handbook of Research on Multi-Agent Systems: Semantics and
Computer Society, 11(1), 3-8.
Dynamics of Organizational Models, 1-16.
[27] Hosny Ahmed Abbas, Samir Ibrahim Shaheen, Mohammed Hussein Amin. Organization of Multi-Agent Systems: An Overview.
International Journal of Intelligent Information Systems. Vol. 4, No. 3, 2015, pp. 46-57. doi: 10.11648/j.ijiis.20150403.11
[28] De Wolf, T., & Holvoet, T. (2005). Emergence versus self-organisation: Different concepts but promising when combined. In
Engineering self-organising systems (pp. 1-15). Springer Berlin Heidelberg.
[29] Serugendo, G. D. M., Irit, M. P., & Karageorgos, A. (2006). Self-organisation and emergence in MAS: An overview. Informatica,
[30] Schillo, M., Fley, B., Florian, M., Hillebrandt, F., & Hinck, D. (2002, August). Self-organization in multiagent systems: from
agent interaction to agent organization. In Third International Workshop on Modelling Artificial Societies and Hybrid
Organisations (MASHO) (pp. 37-46).
|
1303.3828 | 1 | 1303 | 2013-03-15T16:58:00 | Using Serious Games to Train Evacuation Behaviour | [
"cs.MA"
] | Emergency evacuation plans and evacuation drills are mandatory in public buildings in many countries. Their importance is considerable when it comes to guarantee safety and protection during a crisis. However, sometimes discrepancies arise between the goals of the plan and its outcomes, because people find it hard to take them very seriously, or due to the financial and time resources required. Serious games are a possible solution to tackle this problem. They have been successfully applied in different areas such as health care and education, since they can simulate an environment/task quite accurately, making them a practical alternative to real-life simulations. This paper presents a serious game developed using Unity3D to recreate a virtual fire evacuation training tool. The prototype application was deployed which allowed the validation by user testing. A sample of 30 individuals tested the evacuating scenario, having to leave the building during a fire in the shortest time possible. Results have shown that users effectively end up learning some evacuation procedures from the activity, even if only to look for emergency signs indicating the best evacuation paths. It was also evidenced that users with higher video game experience had a significantly better performance. | cs.MA | cs | Using Serious Games to Train Evacuation Behaviour
João Ribeiro1, João Emílio Almeida1†, Rosaldo J. F. Rossetti1†, António Coelho1‡, António Leça Coelho2
1Department of Informatics Engineering
†LIACC – Laboratory of Artificial Intelligence and Computer Science
‡INESC TEC – INESC Technology and Science
Faculty of Engineering, University of Porto
Rua Roberto Frias, S/N, 4200-465, Porto, Portugal
{joao.pedro.ribeiro, joao.emilio.almeida, rossetti, acoelho}@fe.up.pt
2LNEC – National Laboratory of Civil Engineering
Av. Brasil, 101, 1700-066, Lisboa, Portugal
[email protected]
immersion feeling, to make users acquire important, real-world
knowledge. This software approach can be useful in order to
study possible outcomes
to unusual situations and
the
behaviour of whole populations, but do not supply much
interaction; therefore, they can rapidly become uninteresting
for users, instead of helping them as intended. Through the use
of video games, these difficulties can be addressed and
minimised. Video games by definition involve and require user
interaction. They are often quite immersive due to their
multimedia nature and by creating feedback through video,
audio and sometimes even haptic feedback. Video games
usually present some characteristics that make them an
interesting resource: users stay interested in its execution, they
are challenged to master some skills and their performance is
evaluated. These applications have recently been used with
positive results in several and distinct fields of study, what
makes them an option to consider.
The attitude towards them has also changed dramatically:
although previously seen as a "waste of time" and "highly
addictive", their use in education and training has increased in
the past years. It should be noted that video games have been a
major driving
the development of computer
in
force
technology. They now provide realistic scenarios, forces and
interactions, and also user friendly interfaces. Besides, they run
on regular home computers, greatly reducing the necessary
investment in order to make use of them.
In this paper we explore the concept of serious games to
build a virtual fire drill simulator as an attempt to address some
of the issues we have identified within real-world fire drills. It
is our intention to improve the way people participate in such
experiments enhancing their experience in many different
ways. We have adapted and custom the environment of a game
engine, in this case Unity3D, to support simulation features that
enable users to be tracked and assessed while playing.
The remaining of this paper is organised as follows. In
Section II we briefly present some related concepts of subjects
that concern this project, such as serious games, pedestrian
simulation, and crowd behavioural modelling. We then discuss
on applying serious games to evacuation training, in Section
III, where we present and formalise our problem and propose
the approach implemented in this paper. A preliminary
Abstract: Emergency evacuation plans and evacuation drills are
mandatory
in public buildings
in many countries. Their
importance is considerable when it comes to guarantee safety and
protection during a crisis. However, sometimes discrepancies
arise between the goals of the plan and its outcomes, because
people find it hard to take them very seriously, or due to the
financial and time resources required. Serious games are a
possible solution to tackle this problem. They have been
successfully applied in different areas such as health care and
education, since they can simulate an environment/task quite
accurately, making them a practical alternative to real-life
simulations. This paper presents a serious game developed using
Unity3D to recreate a virtual fire evacuation training tool. The
prototype application was deployed which allowed the validation
by user testing. A sample of 30 individuals tested the evacuating
scenario, having to leave the building during a fire in the shortest
time possible. Results have shown that users effectively end up
learning some evacuation procedures from the activity, even if
only to look for emergency signs indicating the best evacuation
paths. It was also evidenced that users with higher video game
experience had a significantly better performance.
Keywords: Evacuation simulator, fire drill, modelling and
simulation, serious games
I.
INTRODUCTION
Everyone who has gone through an emergency drill
exercise has surely noticed some of its downsides. It is hard to
take them seriously, to fully participate, to keep focused on the
exercise and to learn all the necessary information in order to
evacuate correctly. The disruption of the regular functioning of
an enterprise or institution easily upsets their occupants, and so
they tend to distrust these exercises. For the same reasons, and
additionally due to the employees need to move away from
their workstations, it is not usual for managers to accept and
face fire drills very easily, who generally take this kind of
exercises as a loss of time.
A virtual simulation system is a possible alternative to
emergency drills, and can indeed produce some useful
information. To accomplish a good virtual simulation
environment, it is required to provide a realistic scenario and
experiment using our prototype is presented in Section IV,
where some results are also discussed in Section V. We finally
draw some conclusions in Section VI and suggest some further
steps in this research.
II. BACKGROUND
Before we can go further in the explanation of our approach
to create a virtual environment to aid and enhance traditional
fire drills, it is important to introduce some valuable concepts
and some basics of subjects that underlie the implementation of
this project.
A. Serious Games
is an “artificially
According
to Hays [1], a game
constructed, competitive activity with a specific goal, a set of
rules and constraints that is located in a specific context.”
Despite their origins being based purely with entertainment in
mind, video games have been recently used to achieve other
goals and objectives. Serious Games refer to video games
whose application is focused on supporting activities such as
education, training, health, advertising, or social change.
Video games present some characteristics that make them
helpful as a resource. As sentiments like enjoyment and
fulfilment are felt, users are more likely to stay motivated while
doing their tasks. Freitas states in [2] a few benefits from
combining them with other training activities:
•
the learners’ motivation is higher;
•
completion rates are higher;
•
possibility of accepting new learners;
•
possibility of creating collaborative activities;
•
learn through doing and acquiring experience.
Other aspects that draw video game players’ attention are
fantasy elements, challenging situations and the ability to keep
them curious about the outcomes of their possible actions [3].
According to experts from the European Centre for
Children’s Products [4], Serious Games can be classified in
five categories: Edutainment, Advergaming, Edumarket
Games, Political Games and Training and Simulation Games.
Each of these categories can find their practical application in a
diverse range or areas, proving to be an invaluable asset and
instrument for training and influencing behavioural patterns.
B. Pedestrian Simulation
Pedestrian Simulation is a field of computing that has
gained a considerable dimension more recently. Its applications
lean mostly towards planning urban areas, studying and testing
the safety of buildings or open spaces and evaluating
emergency systems and evacuation routes. Other uses include
crowd animation for films or video games, as well.
This field of study involves different technology areas, as
Artificial Intelligence and Crowd Simulation, but also Decision
Making and Behaviour Analysis. One major issue the scientific
community faces though is how to accurately model and
implement human behaviour and feature simulated entities
with rather real attitudes and reactions. This has been addressed
through many different approaches, mainly based on
psychological experiments. Even so, it is also hard to keep
subjects motivated and committed to this kind of experiments.
In this area too, serious games seem to have a great potential as
a research
tool, fostering
the understanding of human
behaviour and the decision-making process carried out by
individuals in a variety of circumstances.
C. Crowd Representation Models
Although many approaches exist to virtually simulate the
behaviour of crowds with varying levels of realism and
computational efficiency, from the analysis of different sources
[5, 6, 7] three models seem to be the most used:
• Cellular Automata Models - treats individuals as
separate objects within a space that is discretely
represented
through a matrix of cells. Besides
allowing only discrete movement and positioning of
the elements in the space, with high-density crowds
the base individual areas are exposed. Even so, it can
still be a valid option as
it demands
lower
computational resources and still provides reasonably
realistic crowd behaviours macroscopically.
• Forces-based Models - use mathematical formulas
to calculate the position variations of the individual
elements through the application of forces. Its most
explored subtypes consider Magnetic Forces and
Social Forces. The former model assigns polarities to
individuals and objects and calculates attractions,
directions and movement based on these values. As
for the Social Forces approach, it is based on the first
one and draws on its concept of polarity to manage
individuals' objectives and wishes.
• Artificial Intelligence-based Models - the decisions
are taken by the individual elements that compose the
crowd. One of the most common subtypes of this
model is the Agent-based Model, whose unit is the
Agent. Agents are elements capable of individually
deciding
their actions according
to
their own
perceptions and interactions. If used in a high enough
number, these models can be a good approximation
to a crowd representation. Most importantly, agent-
based models are a good metaphor for representing
interactions and behaviour, which
social
is of
paramount importance
in crowd modelling and
simulation.
D. Crowd Behaviour
Groups of individuals present different behaviour patterns
under normal and emergency events [8, 9]. One important
factor that has great influence over people's movement and
actions is the layout of a building or surrounding areas.
Helbing [8] considers that the following attributes should
be accounted for and definitely included in the representation
of each virtual individual to simulate their behaviour:
• State - including its position, health and mobility;
•
• Speed - possible moving speeds;
• Vision - field of vision, influenced by smoke;
• Reaction Time - elapsed time before changing
from normal to evacuation behaviour;
• Collaboration - influences the probability of
helping other people;
Insistence - probability of maintaining the same
objective over time;
• Knowledge - familiarity with the building and the
necessary actions to take to evacuate.
Concerning the simulation of large groups under stress
because of an emergency situation, Cordeiro [10] provides
some insight into additional possible effects such situations
might have on the group. It has been observed, for instance,
that "herding" (or "flocking") happens due to people following
other individuals instinctively (a reaction thoroughly explored
by Reynolds in [11]), believing that they will be able to
evacuate faster and more effectively. However, in conditions of
low visibility or little knowledge of the surroundings this can
provoke flocks of wandering people, contributing to the panic
and confusion of the whole group, which is also a social
reaction rather to be avoided if possible.
III. USING SERIOUS GAMES IN EVACUATION TRAINING
Whereas pedestrian and crowd simulations are an important
instrument to help us better understanding social and group
behaviour in a vast number of different circumstances, they are
not good enough to persuade people to behave correctly. On
the other hand, traditional fire drills as they are implemented in
reality albeit constituting a very useful training resource have
also some drawbacks to be sorted out. In our approach, we
combine pedestrian and crowd simulation with the concept of
serious games to build upon an aid tool that can be used to
influence subjects’ behaviour
towards a more effective
instinctive reaction when responding to stressing situations
such as the need to evacuate their current environment.
A. Problem definition
Performing real simulation exercises is an important
activity nowadays, as it helps buildings’ users train evacuation
routes, learn useful behaviours and prepare themselves in case
there is an emergency situation. Having the occupants of a
building well trained in the necessary procedures can help
prevent casualties in case of an uncommon event (which is
even more relevant for large, crowded buildings and ones that
receive many visitors). In the end of a fire drill, it is expected to
have identified some problems with the building’s safety routes
and attempt to correct them. These simulations create the
conditions
for assessing
implement recent methods
to
buildings’ safety levels (such as Performance Based Design -
PBD), replacing more traditional methods that present the
adequate solution for a given situation. In some countries they
are mandatory and even required at certain intervals.
Despite this, due to being uncommon these exercises can be
extremely disruptive to the usual activities and operation of an
institution, more so when they are unannounced. A briefing
must be provided so everyone knows what is going to happen,
how they should proceed and what is expected from them all.
They are required to abandon their workstations for some
moments, halt their work and leave the building before
returning to their places and resuming their activities. Alarm
mechanisms are frequently activated for the duration of the
exercise, emergency authorities are often required to be present
and evacuation
times are measured. For management
executives, who also have to perform these activities, all the
trouble and derangement might be seen as a complete loss of
time and a factor disrupting production.
In these simulations, it is almost impossible for someone to
grasp all the necessary information in order to evacuate from a
building. One might remember later which path was taken
previously, but there are also other numerous details, ranging
from building plans to all possible evacuation routes and gather
points. Thus, it is difficult to retain all this knowledge and have
it present in times of real necessity.
It should also be considered that in order to maintain high
levels of realism in these activities a high investment must be
made to recreate threats, hazards and uncommon events.
Objects that represent fire and smoke machines can be used,
for instance. Nonetheless, unpredictable hazards cannot be
fully recreated, due to their dangerous nature. This is a typical
example to which dealing directly with the real environment
can ultimately compromise its normal functioning.
Some simulation software, known as egress or evacuation
simulators, has been used in recent years to solve part of the
difficulties aforementioned. Indeed they can be useful to
substitute some of the real activities, by providing a less
expensive alternative, which also benefits from being more
easily started, planned and executed. They might be shown to
several occupants at once, minimising the time needed for the
exercise, and can be of help in identifying best routes for
leaving the building, finding bottlenecks and obstacles, and
passing general information. Nevertheless, their aim is usually
more to provide theoretical feedback on the current situation of
a building regarding evacuation. They can be of use for a
building in project, in order to correctly plan exits, paths, and
illumination, or for existing buildings so as to evaluate how its
population could behave in case of an emergency and
unexpected hazardous events. Even the highly realistic ones do
not provide an immersive experience to the users, as they
cannot be put them in charge of controlling a virtual person of
that simulation.
B. The proposed approach
Video games can be a choice to consider in order to address
the aforementioned issues. Their potential to make tedious
tasks become enjoyable is considerable and they have been
recently applied to different ends other than solely providing
entertainment to their users.
As McGonigal presents it [12], it takes a clear goal that we
can achieve and direct feedback for work to feel positive and
even pleasant, and games excel at providing both. In fact, the
author even states that by only bestowing our best efforts into
overcoming a game challenge, we feel more rewarded.
According to Davidson [13], a key quality of video games
is that they focus not only on learning and knowing some
information but actually on using it appropriately. One other
aspect that is relevant is that video games are useful by
providing an experience not far from real life with lower cost
and without compromising safety [14]. Hazards can also be
controlled more easily
to provide randomness
to
the
simulations. This project’s main objective is to study the
viability of using a game engine to create an application that is
simultaneously a pedestrian simulator, a game and a fire
evacuation training tool.
C. Implementation
A Game Engine is a set of tools that provide necessary
features to the development of video games. The required
features in our specific situation as listed below:
•
3D renderer;
• Audio playback;
• Physics engine;
• Graphical User Interfaces (GUI);
• Collision detection;
• Support for artificial intelligence.
After comparing various game engines, Unity3D was
selected. In fact, it presents among other features a powerful
interface that allows visual object placement and property
changing during runtime (especially useful to rapidly create
new scenarios from existing models and assets and quick
tweaking of script variables). The framework
is also
customisable, giving the developer the ability to create code in
JavaScript, C# or Boo. Finally, it also provides a simple project
deployment environment for multiple platforms, with no need
for additional configuration, including the web (which makes it
possible to run any game on a web browser).
Figure 1. DEI plan and 3D model
Currently, our framework contains a single simulation
scenario. It takes place in FEUP’s Informatics Engineering
Department (DEI). The model was created using Blender based
on the actual blueprints of the building so as to recreate it as
realistically as possible in terms of topology, dimensions, scale
and proportions Figure 1. The 3D environment can be seen in
Figure 2, which gives a good idea of how realistic such an
environment can represent the actual place under study.
D. Execution of a evacuation scenario
The player starts in a predefined room and, upon starting
the evacuation event, a fire appears in a random room, forcing
the alarm to ring. At this moment, the timer starts and the
player must then traverse the building in order to go outside as
quickly as possible, choosing from one of the two possible
exits. Several emergency signs are in place so as to help the
player to identify the nearest exit to his/her current spot.
Figure 2. 3D rendered view of the building
This attempts to recreate the experience of the user being
physically there and exploring their surroundings. To steer
his/her digital avatar in a 3D virtual world the user has the
controls that are commonly standard for the FPS game genre
(First Person Shooter games characterised by placing players in
a 3D virtual world which is seen through the eyes of an avatar).
IV. EVALUATION OF PRELIMINARY RESULTS
In order to test our approach and evaluate its potentials, a
simple experiment was executed and some results analysed.
Our findings have been quite encouraging and promising,
demonstrating the viability of our methodology.
A. Population Sample
A total of 30 individuals participated in this experiment and
were invited to test the developed prototype. Testers can be
individually classified according to different dimensions,
namely whether they are regular video game players and
whether they had previous knowledge of the building.
According
to
these characteristics,
four groups were
considered. The different groups were labelled A, B, C and D
as follows:
• A - previous building knowledge and frequent game
players;
• B - previous building knowledge and little experience
with games;
• C - no previous building knowledge and frequent game
players;
• D - no previous building knowledge and little
experience with games;
The distribution of our sample into each of the groups
previously described is presented in Table 1.
TABLE I.
Frequent game
player
Yes
USER GROUPS
Previous building
knowledge
Yes
No
Yes
No
Yes
No
No
Population
8
6
5
11
Group
A
B
C
D
B. Test Setup
Before the evacuation test started, all players had a few
moments to try the controls and get used to them (this was
done without the possibility of exploring the building or even
leaving the starting room). Players with little to no familiarity
with the building additionally had a small demonstration of the
way from the main entrance to the room where the test would
start, to have in mind at least one way out.
All users had to perform the same task in order to generate
comparable results. Starting in a room inside DEI, the player’s
objective was then to reach the outside of the building in the
shortest time possible. The alarm would sound and computer-
controlled agents would also start to move towards the nearest
exit, crowding the corridors and causing the natural disruptions
we can usually find when facing this kind of situations. With
time, fire would spread all over the place and become an even
more serious obstacle. While navigating inside the building,
players could refer to the emergency signs pointing to the
nearest emergency exits.
The only measured value was the time taken to evacuate the
building, as mistakes made while finding the way out would be
automatically reflected in a longer elapsed time.
V. RESULTS AND DISCUSSION
While carrying out our experiments, some performance
measures were logged for each subject. Figure 3 depicts the
times for each player.
Very curiously, the average time to leave the building was
higher than expected. After all, in real life it only takes around
22 seconds to do the same way. This might have been
originated by the difficulties verified by some users to control
the avatar, mostly ones who demonstrated little to no
acquaintance with video games on a regular basis. Some of
them required several seconds to get used to controlling mouse
and keyboard at the same time; in some cases they even gave
up trying to do it simultaneously and opted instead by
alternating between pointing at the desired direction and then
moving on that way.
500
450
400
350
300
250
200
150
100
50
0
)
s
d
n
o
c
e
s
(
e
m
i
T
Figure 3. User times for each player
Average evacuation times computed by category are shown
in Table II, below.
TABLE II.
Group
A
B
C
D
AVERAGE TIMES BY CATEGORY
Average time
(seconds)
23.9
43.9
58.0
145.1
It is not surprising that the group with better performance
was that of regular video game players who were also familiar
with the building (group A). It was also expected to see the
opposite group – group D, people that are neither familiar with
DEI’s building nor regular gamers – have the lowest score. It is
the two middle scores that deserve more attention.
Not knowing the building or not being used to the game
controls: both groups present similar results. Whilst the
frequent game players where faster in controlling the avatar,
they missed the shortest path to outside and choose the usual
way due to their poor building knowledge (group C); the group
with little game experience but good building familiarity
(group B) although using the emergency exit took more time
due to little experience with games. Frey et al. [15] have even
shown that by making users get used to a game, the difficulties
with controls are quickly overcome and the subjects differences
on this point mitigated. To the benefit of the experience,
players were allowed to play only once. When repeating the
game, all would select the emergency exit and evacuation times
decreased substantially, so results would be biased. It was
precisely the novelty of facing such a new situation that was
the aim of this experience.
[3]
[9]
[4]
REFERENCES
[1] RT Hays. The Effectiveness of Instructional Games: a Literature Review
and Discussion. Technical report , Naval Air Wwarfare Center Training
Systems Division Orlando, FL , 2005.
[2] Sara Freita s. Using Games and Simulations for Supporting Learning.
Learning, Media and Technology, 31(4):343–358, 2006.
John Kirriemuir and A. McFarlane. Literature Review in Games and
Learning. Technical report , futurelab, 2004.
Julian Alvarez, Olivier Rampnoux, Jean-Pierre Jessel, and G illes
Methel . Seriou s Game: Ju st a Qu estion of Posture? In Artificial &
Ambient Inteligence,AISB’07, pages 420–423, 2007.
[5] Laure Heïgeas, Annie Luciani , Joelle Thollot, and Nicola s Ca stagné. A
Physically-Ba sed Particle Model of Emergent Crowd Behaviors. In
Graphicon, 2003.
[6] N. Pelechano and A. Malkawi. Comparison of Crowd Simulation for
Building Evacuation and an Alternative Approach. In Building
Simulation, pages1514–1521, 2007.
[7] Gabriel Santos and Benigno E Aguirre. A critical review of emergency
evacuation simulation models. Critical Review, (1032):25–50, 2004.
[8] Dirk Helbing. Simulation of Pedestrian Crowds in Normal and
Evacuation Situations. In Pedestrian and Evacuation Dynamics, pages
21–58, 2002.
J. E . Almeida, Rosaldo Rosseti, and A. L . Coelho. Crowd Simulation
Modeling Applied to Emergency and Evacuation Simulations using
Multi-Agent Systems. In 6th Doctoral Symposium on Informatics
Engineer ing, 6th Doctoral Symposium on Informatics Engineering,
DSIE’11, Por to, Por to, 2011.
[10] Elisabete Cordeiro, António Leça Coelho, Rosaldo J F Rossetti, and
João Emílio Almeida. 2011. Human Behavior under Fire Situations – A
ca se– study in the Portugu ese Society. In ADVANCED RESEARCH
WORKSHOP: Evacuation and Human Behavior
in Emergency
Situations, 63-80. Santander, Spain: GIDAI . Universidad de Cantabr ia.
[11] Reynolds, C . W . (1987). Flock s, herds and schools: A distributed
behavioral model . ACM SIGGRAPH Computer Graphics, 25-34.
[12] Jane McGonigal. Reality Is Broken: Why Games Make Us Better and
How They Can Change the Wor ld, volume 22. Penguin Press HC, The,
2011.
[13] Drew Davidson. Beyond Fun, volume 11. ETC Press, 2008.
[14] K. Corti. Games-Based Learning; a Serious Bu siness Application.
Technical report , 2006.
[15] A Frey, J Har tig, A Ketzel , A Zink ernagel, and H Moosbrugger. The
Use of Virtual Environments Based on a Modification of the Computer
Game Quake III Arena R in Psychological Experimenting. Computers in
Human Behavior, 23(4):2026–2039, July 2007.
VI. CONCLUSIONS AND FUTURE WORK
In this work we explore the concept of serious games as an
important asset to aid and improve traditional fire drills. It does
neither completely replace nor avoid the need for in-site drills
to train people in emergency situations, such as with the
prospect of fire in an office building, for instance. Nonetheless,
game environments can be very attractive in many different
ways, and have proven to be an invaluable tool for training.
Additionally, we build our approach on the potential of such a
concept to ease and improve the understanding of human
behaviour in such situations, as subjects are monitored during
their playing the game and some performance measures are
logged to be further analysed later on.
We have implemented our prototype with Unity3D, a
popular game engine, which provided us with a customisable
framework and allowed us to insert features of a serious game
platform into the virtual environment. We invited some
subjects to use the game and collected some preliminary results
that demonstrate the viability of the approach. We have then
conceived a methodology which is both instrumental as an aid
invaluable
to
train people and an
instrument
to help
practitioners and scientists
to better understand group
behaviour and the social phenomenon in a vast range of
circumstances.
the
include
this research
in
The very next steps
improvement of the prototype featuring it with tools for rapidly
setting up simulation environments from CAD blueprints of
buildings. We also intend to include other performance
measures
in
individual and social behaviour
to study
circumstances other the hazardous scenarios. Ultimately, this
tool is also expected to be used as an imperative decision
support tool, providing necessary and additional insights into
evacuation plans, building layouts, and other design criteria to
enhance places where people usually gather and interact rather
socially, such as shopping malls, stadiums, airports, and so on.
ACKNOWLEDGMENT
This project has been partially supported by FCT
(Fundação para a Ciência e a Tecnologia), the Portuguese
Agency for R&D, under grant SFRH/BD/72946/2010.
|
1910.09442 | 1 | 1910 | 2019-10-08T16:15:57 | Data assimilation in Agent-based models using creation and annihilation operators | [
"cs.MA",
"cs.LG"
] | Agent-based models are a powerful tool for studying the behaviour of complex systems that can be described in terms of multiple, interacting ``agents''. However, because of their inherently discrete and often highly non-linear nature, it is very difficult to reason about the relationship between the state of the model, on the one hand, and our observations of the real world on the other. In this paper we consider agents that have a discrete set of states that, at any instant, act with a probability that may depend on the environment or the state of other agents. Given this, we show how the mathematical apparatus of quantum field theory can be used to reason probabilistically about the state and dynamics the model, and describe an algorithm to update our belief in the state of the model in the light of new, real-world observations. Using a simple predator-prey model on a 2-dimensional spatial grid as an example, we demonstrate the assimilation of incomplete, noisy observations and show that this leads to an increase in the mutual information between the actual state of the observed system and the posterior distribution given the observations, when compared to a null model. | cs.MA | cs | DATA ASSIMILATION IN AGENT-BASED MODELS USING
CREATION AND ANNIHILATION OPERATORS
Daniel Tang
Leeds Institute for Data Analytics
University of Leeds
Leeds, UK
[email protected]
9
1
0
2
t
c
O
8
]
A
M
.
s
c
[
1
v
2
4
4
9
0
.
0
1
9
1
:
v
i
X
r
a
October 22, 2019
ABSTRACT
Agent-based models are a powerful tool for studying the behaviour of complex systems that can be
described in terms of multiple, interacting "agents". However, because of their inherently discrete
and often highly non-linear nature, it is very difficult to reason about the relationship between the
state of the model, on the one hand, and our observations of the real world on the other. In this paper
we consider agents that have a discrete set of states that, at any instant, act with a probability that may
depend on the environment or the state of other agents. Given this, we show how the mathematical
apparatus of quantum field theory can be used to reason probabilistically about the state and dynamics
the model, and describe an algorithm to update our belief in the state of the model in the light of
new, real-world observations. Using a simple predator-prey model on a 2-dimensional spatial grid as
an example, we demonstrate the assimilation of incomplete, noisy observations and show that this
leads to an increase in the mutual information between the actual state of the observed system and the
posterior distribution given the observations, when compared to a null model.
Keywords Data assimilation, Agent based model, Quantum field theory, Probabilistic programming, Fock space
1 Introduction
P (sΩ) ∝ P (Ωs)P (s)
When modelling real-world phenomena we often use the model to make forecasts, only to be later faced with new
observations that weren't available at the time of the forecast. This leads to the problem of how to "assimilate" the new
observations into an updated forecast and, perhaps, another forecast for a later time. There exists an extensive literature
on this problem as applied to weather forecasting models[1][2] which describes how the problem can be reduced using
Bayes' rule:
(1)
Here, P (s) is our original forecast (this is a probability distribution since there is uncertainty in the forecast) P (Ωs) is
the likelihood of the observation, Ω, given a forecast, s, and P (sΩ) is our updated forecast. However, the algorithms
developed to do this in weather forecasting models depend on the model's governing equations having some degree of
smoothness, a condition not generally satisfied by agent-based models.
If there are only a very small number of observations, particle filtering or Markov Chain Monte Carlo methods may
have some success[3], but as the number of observations increases, these methods tend to quickly fail. This is because
as the number of observations increases, the proportion of the model's phase-space that has non-zero likelihood often
shrinks exponentially, so if the dimensionality of the phase-space is not trivially small, finding a non-zero start point for
a Markov chain becomes exponentially harder. Similarly for particle filters; the number of particles needed to prevent
all particles ending up with zero probability, increases exponentially with the number of observations.
For these reasons, the problem of data assimilation in agent-based models remains an unsolved problem, and this
imposes a severe restriction on the utility of these models when applied to real world problems.
2 Quantum field theory applied to agent-based models
In order to apply Bayes' rule to agent-based models (ABMs) we must first devise a way to represent a probability
distribution over the possible states of an agent-based model. The task is complicated somewhat by the fact that, in
Data assimilation in Agent-based models using creation and annihilation operators
general, the exact number of agents in the model is unknown and variable, so the set of model states must span models
with any number of agents. Luckily, we can use the mathematical formalism of quantum field theory to do this. In a
quantum field there are an unknown number of interacting particles and the quantum field is represented as a quantum
amplitude over the possible configurations of particles; to apply this to ABMs, we just need to interpret the particles as
agents and replace the quantum amplitudes with probabilities.
2.1 The creation operator
†
Let ∅ represent the empty model state (i.e. the state with no agents present), and let a
i be an operator on model states
that has the effect of adding an agent in state i to the model state, this is called the creation operator. So, for example,
†
i∅ is the state with a singe agent in state i. Since the result of a creation operator is itself a state, we can apply another
a
†2
†
†
i ∅
i∅ to represent a state with two agents, one in state i and one in state j, or even a
creation operator, for example, a
ja
to represent a state with two agents in state i. We can go on applying operators in this way to construct any desired state.
If we now allow states to be multiplied by real numbers and added together, a vector space is induced which can
†
†
i∅ + 0.5a
j∅ would represent a probability
represent probability distributions over model states. For example 0.5a
distribution with 0.5 probability that there's a single agent in state i and 0.5 probability that there's a single agent in
state j (with no probability that both agents are present at the same time).
Since each term in a probability distribution always ends in the empty state, ∅, we can just as easily consider the
additions and multiplications to be part of the operator, so the example distribution above could be equivalently written
†
†
j)∅. Given this, composure of operators works in the expected way, for example, a†
i + a
0.5(a
†
†
j)∅.
0.5(a†
i + a†
za
za
With this representation, we need a way to express the fact that it doesn't matter in what order the creation operators are
†
†
†
†
i∅ = a
j∅. Since this is true for all states, not just ∅, we can drop the ∅ and say1
applied, so a
i a
ja
(cid:17)∅ =
(cid:16)
†
†
i + a
0.5(a
j)
z
This form of equation is known as a commutation relation or commutator and has a shorthand notation:
†
†
i , a
[a
j] = 0
†
†
†
†
j − a
a
i a
ja
i = 0
2.2 The annihilation operator
Now let ai be the annihilation operator which has the following properties:
ai∅ = 0
and
where δij is the Kronecker delta function.
Given just these properties, we can show that
[ai, aj] = 0
†
[ai, a
j] = δij
This is a recurrence relationship which, using equations 3 and 5, can be solved to give
†n−1
†
†n
i ∅ = aia
i a
i
†n−1
†
†
∅ = (a
i ai + [ai, a
i ])a
i
∅
aia
†n
i ∅ = na
aia
†n−1
i
∅
(2)
(3)
(4)
(5)
So, the annihilation operator has the effect of multiplying by the number of agents in a state, then removing one of
those agents.
Notice that, using the commutation relations in this way we can transform any sequence of operations on ∅ into an
equivalent form that contains no annihilation operators. We'll call this the reduced form of the operator.
3 Equations of motion of a probabilistic ABM
Armed with just the creation and annihilation operators we now show that if the behaviour of agents in an ABM can be
described as a set of propensities to act and interact at any instant then we can transform the ABM into a "Hamiltonian"
operator, containing only addition, multiplication, creation and annihilation operators. The Hamiltonian operator turns a
1to be correct, the 0 is interpreted as the "multiply by zero" operator or alternatively 0I where I is the identity operator
2
Data assimilation in Agent-based models using creation and annihilation operators
probability distribution over ABM states into a rate of change of that distribution. This is significant in that it allows us
to express the equations of motion of a probability distribution as an operator, and so describe probabilistically how the
ABM evolves through time. That is, we've defined a "probabilistic ABM" that captures all the behaviour of the original
ABM in a smooth, differentiable way which, as we'll show, makes it much easier to perform inference on the ABM.
To do this, start with the definition of an agent's behaviour: this must be representable as a set of actions and interactions.
An action consists of a pair of functions, (ρ(i), A(i)) where ρ is a rate per unit time and A is a set of agents. An agent
in state i is said to have behaviour (ρ(i), A(i)) if in any infinitesimal time-slice, δt, it replaces itself by the agents
in A(i) with probability ρ(i)δt. An interaction consists of a pair of functions, (ρ(i, j), A(i, j)) such that an agent in
state i, in the presence of another agent in state j replaces both itself and the other agent with A(i, j) with probability
ρ(i, j)δt. Although this is not a common way to think about agent-based models, a very large class of models can easily
be described in this way.
Without loss of generality, we assume all agents in a model have the same set of behaviours (different behaviours among
agents can be achieved by putting an "agent type" marker in the agent's state and making the behaviour rates zero for
certain agent types).
Consider now an action, (ρ(i), A(i)) and let [ki1 . . . kini] = A(i) be the results of the action. The Hamiltonian for this
action would be
(cid:89)
k∈A(i)
ai
†
†
k − a
a
i
H(ρ,A) = ρ(i)
†
The product term expresses the rate of increase in probability in the state resulting from the action, while the a
i term
expresses the (equal in magnitude) rate of decrease in probability in the initial state due to the agent being removed
when it acts. Because the annihilation operator, ai, multiplies by the number of agents in a given state, the rate of
change is also multiplied if there are multiple agents to which the behaviour applies (or multiplied by zero if there are
no applicable agents).
Now consider an interaction (ρ(i, j), I(i, j)). The Hamiltonian for this interaction would be
(cid:89)
k∈I(i,j)
− a
†
k
a
†
i a
†
j
ajai
H(ρ,I) = ρ(i, j)
The double annihilation operator, ajai, has the effect of multiplying by the number of i, j agent pairs so that each agent in
†n
∅,
i ∅ = n(n− 1)a
state i gets a chance to interact with each agent in state j. This also works when j = i, since a2
i a
and the number of i, i pairs is n(n − 1).
This pattern can easily be extended to interactions between three or more agents, but in this paper we'll consider only
binary interactions.
If an agent has more than one behaviour, the Hamiltonian is simply the sum of the individual behaviours, so in
this way we can build a Hamiltonian for a whole ABM. So, an ABM whose agents have a set of action behaviours
{(ρ1(i), A1(i)) . . . (ρα(i), Aα(i))} and a set of interaction behaviours {(ρ1(i, j), I1(i, j)) . . . (ρβ(i, j), Iβ(i, j))} has
the Hamiltonian
†n−2
i
(cid:89)
− a
†
i
ai +
†
a
k
k∈An(i)
β(cid:88)
(cid:88)
(cid:88)
m=1
i
j
ρm(i, j)
(cid:89)
k∈Im(i,j)
− a
†
a
k
†
†
i a
j
ajai
α(cid:88)
(cid:88)
n=1
i
H =
ρn(i)
Although the number of terms in the Hamiltonian can become very large, we'll see later that we only ever need to deal
computationally with the much smaller commutation relations [X, H].
Given the Hamiltonian, the time evolution of a probability distribution over ABM states, ψ, is by definition
This has the analytical solution
dψ
dt
= Hψ
where ψ0 is the initial state, ψt is the state at time t and the exponential of an operator, tH, can be defined as
ψt = etH ψ0
∞(cid:88)
n=0
tnH n
n!
etH =
(6)
Equation 6 gives us a prior forecast which can be inserted into Bayes' rule (equation 1) in order to do data assimilation:
P (stΩt) =
(cid:80)
P (Ωtst)P (etH ψ0 = st)
s P (Ωts)P (etH ψ0 = s)
3
(7)
Data assimilation in Agent-based models using creation and annihilation operators
where st is a state at time t, Ωt are the observations at time t and ψ0 is the probability distribution of the ABM at time
t = 0 given all observations up to that time.
So, we have a succinct, mathematical solution to the problem of data assimilation in an ABM. We now need to translate
equation 7 into a numerical algorithm that will execute in a reasonable time.
4 The Deselby distribution
As a first step towards a practical numerical algorithm, we introduce the Deselby distribution which we define as
∞(cid:88)
(k)∆λk−∆e−λ
†k
i ∅
a
Diλ∆ =
k=0
k!
where ∆ is an integer, λ is a real number and (k)∆ denotes the ∆th-order falling factorial of k.
The Deselby distributions have the useful property that
and so
It can also be seen that
†
a
i Diλ∆ = Diλ(∆+1)
†∆
i Diλ0 = Diλ∆
a
aiDiλ0 = λDiλ0
Now imagine a product of Deselby distributions over all agent states
(cid:89)
DΛ =
Diλi0
(8)
(9)
so that, for any i
i
aiDΛ = λiDΛ
So, in the same way as with ∅, any sequence of operations on DΛ can be reduced to a sequence of creation operations
on DΛ, i.e. a reduced form, and we can consider DΛ as a "ground" state in just the way we have been using ∅ so far.
So, if we express a probability distribution in terms of operators on DΛ we can apply the Hamiltonian to this to get a
rate of change in in terms of DΛ. This will be much more computationally efficient than dealing directly with operators
on ∅.
Note also that when ∆ = 0, the Deselby distribution is just the Poisson distribution, so it is particularly convenient if
we can express our prior beliefs about the number of agents in each state of an ABM in terms of Poisson distributions.
4.1 Data assimilation with Deselby distributions
As a concrete example, suppose there are a number of identical agents moving around a 2-dimensional grid, and
we observe the number of agents in the ith grid-square. To make this slightly more general, suppose our method of
observation may be imperfect and there is a probability, r, that an agent is detected given that it is there, such that the
likelihood function is just the Binomial distribution
P (mk) =
rm(1 − r)k−m
(cid:18) k
(cid:19)
m
where m is the observed number of agents and k is the actual number present. Appendix A shows that, if our prior is a
Deselby distribution Diλ∆, then the posterior is given by
†m
i am
i LgirDiλ∆
P ∝ a
(10)
Where we introduce the operator Lgir which has the properties
†
†
i ] = −ra
i Lgir
[Lgir, a
[Lgir, ai] =
rai
1 − r
Lgir
and
LgirDΛ = DΛ(cid:48)
where Λ(cid:48) is Λ with the ith component, λi, replaced by (1 − r)λi.
4
(11)
(12)
(13)
Data assimilation in Agent-based models using creation and annihilation operators
Since euqation 10 is true for any Deselby distribution, it is also true for any state described as an operator on DΛ. So,
given ψ0 (a distribution at time t = 0 expressed as an operator on DΛ) and an observation of m agents in state i at time
t, the postrior distribution at time t is given by
Multiple observations Ω = {(i1, m1)...(in, mn)} can be treated by applying operators for each observation
ψt ∝ a
†m
i am
i LgiretH ψ0
(cid:89)
(i,m)∈Ω
etH ψ0
†m
i am
a
i Lgir
(14)
(15)
ψt ∝
4.2 Deselby state approximation
Even when working with Deselby distributions, as time progresses the size of the reduced form of the distribution will
become large and sooner or later. In order to maintain computational tractability, we'll need to somehow approximate
the distribution. There are many ways of doing this, but a common way is to minimise the Kulback-Leibler divergence
between the original distribution and the approximation[4] over a family of possible approximations. This has the effect
of minimising the loss of information due to the approximation.
If P∅ is the original distribution and A∅ the approximation, and we choose to minimise DKL(P∅A∅) the resulting
approximation has a particularly simple form.
By definition the KL-divergence is given by
where we define
and
DKL(P∅A∅) = ∅ · P ∗ (log(P∅) − log(A∅))
log(cid:0)(c1a†l1 + ··· + cna†ln )∅(cid:1) = (log(c1)a
†l1
1 + ··· + log(cn)a†ln
n )∅
∅ · a†n∅ = δn0
where δn0 is the Kronecker delta function.
P ∗ is defined to be the conjugate of P , where the conjugation operator is defined as having the properties
†n
i )∗ =
(a
an
i
n!
(A + B)∗ = A∗ + B∗
and
(AB)∗ = A∗B∗
and c∗ = c if c is a constant.
If we choose A to be a Deselby distribution, and minimise the KL divergence with respect to Λ we have, for all i
Solving this for λi gives
Where the operator
∀i : 0 = ∅ · P ∗ ∂ log(Diλi∆i∅)
∂λi
∀i : λi = ∅ · e
(cid:80)
j aj aiP∅ − ∆i
j aj = ∅ ·(cid:89)
∞(cid:88)
an
j
n!
j
n=0
(cid:80)
∅ · e
(16)
has the effect of summing over a distribution (i.e. summing the coefficients of an operator expressed in reduced form on
∅).
This can be interpreted as saying that the Deselby distribution, D, that minimises DKL(P∅D) is the one whose mean
number of agents in state i matches the mean of P∅ for all i.
This gives us values for λi given ∆i. To calculate ∆i, notice that for the KL-divergence to be finite, the support of P∅
must be a subset of the support of D. So if ∆i has value n, the probability that there are any less than n agents in state
i in P∅ must be zero. So, we can choose ∆i to be the highest value for which there are definitely at least that many
agents in state i in P∅.
5
− ∆i
(17)
(18)
(19)
(20)
n, H] is considerably smaller
To calculate this, we need an algorithm to calculate expressions of the form
To increase numerical stability, we transform this to
where I is the identity operator. Expanding the exponential in H gives
(cid:80)
y = ∅ · e
(cid:80)
j aj XetH ψ0
y = e−t∅ · e
j aj Xet(I+H)ψ0
∞(cid:88)
(cid:80)
∅ · e
tn
n!
y = e−t
j aj Znψ0
where Zn = X(I + H)n. However,
n=0
Zn+1 = Zn(I + H) = Zn + HZn + [Zn, H]
but since H is the Hamiltonian, and the Hamiltonian preserves probability mass,
j aj HZnψ0 = 0
for all Zn and ψ0 so we can replace Z by Z(cid:48) in equation 19 where Z(cid:48)
n, H]
Z(cid:48)
n+1 = Z(cid:48)
n + [Z(cid:48)
0 = X and
(cid:80)
∅ · e
without affecting the value of y.
This is very convenient computationally, because although H is in general very large, [Z(cid:48)
so we have a much more efficient way to calculate the terms of the expansion.
(cid:80)
The computation can be made even more efficient by noting that
†
i X = ∅ · e
j aj a
∅ · e
j aj X
(cid:80)
Data assimilation in Agent-based models using creation and annihilation operators
5 A practical algorithm for data assimilation
Using equations 15 and 16 we can perform a data assimilation cycle that consists of the following steps: take an initial
state ψ0, integrate forward in time, apply the set of observations Ω = {(i1, m1) . . . (in, mn)}, then finally approximate
to the Deselby distribution that minimises the KL-divergence.
Since we've just observed m agents in state i for each (i, m) ∈ Ω we can define the ∆ of the approximation to be
given this, λi is given by
∀i : λi =
0
otherwise
(cid:26)m if (i, m) ∈ Ω
(cid:16)(cid:81)
(cid:16)(cid:81)
†m
j am
†m
j am
(j,m)∈Ω a
(j,m)∈Ω a
∆i =
(cid:80)
∅ · e
(cid:80)
∅ · e
j aj ai
j aj
(cid:17)
(cid:17)
j Lgjr
etH ψ0
j Lgjr
etH ψ0
for all i. So when calculating each Z(cid:48)
them to the left, leaving only annihilation operators.
In practice, we only need to calculate the first few terms in the expansion in 19 so we end up with
n we can use the commutation relations to remove all creation operators by moving
(21)
The Z(cid:48)
n terms, can be calculated consecutively using equation 20 with
for the numerator in equation 18 and
Z(cid:48)
0 =
for the denominator.
†m
j am
a
j Lgjr
y = e−t
(cid:80)
j aj Z(cid:48)
∅ · e
nψ0
Z(cid:48)
0 = ai
†m
j am
a
j Lgjr
n=0
tn
n!
N(cid:88)
(cid:89)
(cid:89)
(j,m)∈Ω
6
(j,m)∈Ω
Data assimilation in Agent-based models using creation and annihilation operators
Aget type Description
Prey
death
reproduction
movement
Rate per unit time
0.1
0.15
1.0
Predator
death
predation/reproduction (per prey)
movement
0.1
0.5
1.0
Table 1: The rates of each behaviour in the predator-prey model
We then just need to multiply Z(cid:48)
reduced form and sum the coefficients (in order to perform the ∅ · e
Numerical experiments showed that, when calculating the terms in equation 21, a good estimate of the error obtained by
truncating the series at the N th term can be calculated as
n by ψ0 (i.e. the Deselby distribution from the previous assimilation cycle), convert to
j aj operation)2.
where
¯x =
t
N
(22)
(cid:12)(cid:12)(cid:12) 1
n
nψ0
(cid:80)
(cid:33)
≈ e−t
(cid:32)
e¯x − N(cid:88)
(cid:12)(cid:12)(cid:12)∅ · e
N(cid:88)
(cid:80)
j aj Z(cid:48)
n=0
¯xn
n!
n=1
This can be calculated along with the terms of the expansion until the error estimate falls below some desired threshold.
A final computational simplification can be made by noticing that when calculating the numerator of equation 18,
observations far away from the ith grid-square will have very little influence on its posterior (depending on t and the
speed of information flow through the ABM).
This intuition is formalised in Appendix B where we show that if we let Φ(X) be the set of all states operated on by
operator X, and let
[nX, H] = [...[[X, H], H]..., H]
denote the n-fold application of a commutation, then if the observations in Ω can be partitioned into two sets ΩA and
ΩB in such a way that
N(cid:88)
n=0
Φ
[nai
(cid:89)
(m,j)∈ΩA
∩ Φ
N(cid:88)
[n (cid:89)
n=0
(m,j)∈ΩB
†m
j am
a
j , H]
= ∅
†m
j am
a
j , H]
(where ∅ above is the empty set, not the ground state) we can factorise out the effect of all observations in ΩB (to
N th-order accuracy) in both the numerator and denominator so that they cancel and we can effectively remove the
observations in ΩB (note that in the above equation we've removed the Lgjr operator as it has no effect under the Φ
operation).
6 Spatial predator-prey example model
In order to provide a numerical demonstration of the above techniques, we simulated a 32x32 grid of agents. Each
agent could be either a "predator" or a "prey". At any time a prey had a propensity to move to an adjacent grid-square
(i.e. up, down, left or right), to die or to give birth to another prey on the an adjacent grid-square. Predators also had
propensities to move to adjacent grid-squares or die, but in the presence of a prey on an adjacent grid-square, they also
had a propensity to eat the prey and reproduce into an adjacent grid-square. Perhaps surprisingly, even a model with
such a simple set of behaviours exhibits quite complex emergent properties that are difficult to predict. The rate per unit
time of each behaviour is shown in table 1.
A non-probabilistic, forward simulation of the ABM was performed in order to create simulated observation data. The
initial state of the simulation was constructed by looping over all grid-squares and choosing the number of predators in
that square from a Poisson distribution with rate parameter λ = 0.03 and the number of prey from a Poisson distribution
with λ = 0.06. The model was then simulated forward in time and at time intervals of 0.5 units an observation of
the state of the simulation was recorded. An observation was performed as follows: for each grid-square, with 0.02
probability, the number of prey was observed. If an observation was made, each prey in that grid-square was counted
with a 0.9 probability. The same procedure was then repeated for predators.
2We can also sum coefficients when working with a Deselby ground state since the sum of the coefficients of a Deselby
distribution, expressed as an operator on ∅, is 1
7
Data assimilation in Agent-based models using creation and annihilation operators
Figure 1: Plot showing the increase of mutual information between the computed posterior distribution and the actual
state for the first 64 assimilation windows of 10 simulations. The zero-information point is taken to be that of the
posterior given the current window's observations and the prior at the start of the simulation.
Once the observations were generated, a sequence of data assimilation cycles were performed as described in equations
17 and 18, with a starting state of the Deselby ground state DΛ with λi = 0.03 for all i representing a predator and
λi = 0.06 for all i representing a prey. The terms in the expansion were calculated until the expected truncation error,
calculated using equation 22 fell below 0.2% of the sum of terms up that point.
6.1 Results
The mutual information between the assimilated state and the real state (where the real state is represented by a Dirac
delta distribution) was calculated at the end of each assimilation cycle. In order to distinguish between the information
accumulated by the data assimilation cycles and the information contained in the current time's observation and the
prior at the start of the simulation, we also calculated a reference value equal to the mutual information between
(cid:16)(cid:81)
(cid:80)
i ai
∅ · e
(cid:17)
DΛ
(cid:17)
(j,m)∈Ωt
(cid:16)(cid:81)
†m
j am
a
j Lgir
†m
j am
a
(j,m)∈Ωt
j Lgir
DΛ
and the real state (i.e. the posterior given only the initial prior and the current window's observations).
Figure 1 shows the difference between the mutual information of the assimilated state and the reference value for
each assimilation cycle of 10 separate simulations. The upward trend in this plot shows that the assimilation cycle is
successfully accumulating information about the real state.
Figure 2 shows snapshots of the real state of the non-probabilistic simulation superimposed on the assimilated state, for
three of the simulations after 64 assimilation cycles. Inspection of these snapshots seems to suggest that the mutual
information tends to be higher when agents are arranged in clusters rather than being more scattered, which is not
surprising. Also, when a grid-square containing an agent is observed but (because of the 0.9 probability of observation)
the number of observed agents happens to be zero, this will reduce the mutual information.
7 Previous work
Applying the methods of quantum field theory to other domains is not new; for example Dodd and Ferguson[5] suggest
applying it to epidemiology, O'Dwyer and Green[6] apply it to simulations of biodiversity while Santos et.al.[7] applies
it to enzymatic reations. Abarbanel[8] comes closest to our work in that he talks about quantum field theory in the
context of data assimilation. However, to our knowledge, nobody has before developed quantum field theory to apply to
such a general class of ABMs, nor have they developed it into an algorithm for data assimilation. Also, the various
8
-20 0 20 40 60 80 100 120 0 10 20 30 40 50 60 70Posterior / real state mutual information (bits)Assimilation window Data assimilation in Agent-based models using creation and annihilation operators
Figure 2: Posterior, assimilated probability (background colour) plotted against actual agent positions (dots) for the
highest (top), average (middle) and lowest (bottom) mutual information (out of 10 simulations) after 64 assimilation
cycles. Red dots are real prey positions, blue dots are real predator positions. The background colour shows the posterior
probability: higher intensity red corresponds to higher probability of prey, higher intensity of blue corresponds to higher
probability of predator.
9
Data assimilation in Agent-based models using creation and annihilation operators
numerical strategies presented here that make the technique computationally tractable, including our presentation of the
Deselby distribution is, we believe, new.
8 Discussion and Conclusion
This paper presents some early results of a promising new technique in data assimilation for agent based models.
However, there's plenty of work left to be done and many opportunities for further development. For example, more
research into flexible ways of approximating the state or computationally efficient ways of representing a state as a data
structure, would be worthwhile. The example model presented here is very small and simple, the next step would be
to apply this technique to larger and more complex models. In addition, in this paper we only consider agents with a
discrete set of states, the extension to states including continuous variables is fairly straightforward and work on this is
currently under way.
At present there are very few techniques that can be used to successfully perform data assimilation in agent based
models. We believe the results presented here provide a significant first step towards filling that gap and hope that it will
form the foundation of a fruitful approach to the problem.
The code used to create the results in this paper can be found at https://github.com/deselby-research/ProbabilisticABM,
where you'll also find a library of code we're developing in order to experiment with new techniques in this area.
References
[1] Eugenia Kalnay. Atmospheric modeling, data assimilation and predictability. Cambridge university press, 2003.
[2] Alberto Carrassi, Marc Bocquet, Laurent Bertino, and Geir Evensen. Data assimilation in the geosciences: An
overview of methods, issues, and perspectives. Wiley Interdisciplinary Reviews: Climate Change, 9(5):e535, 2018.
[3] Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall Press, 2009.
[4] Kevin P Murphy. Machine learning: a probabilistic perspective. MIT press, 2012.
[5] Peter J Dodd and Neil M Ferguson. A many-body field theory approach to stochastic models in population biology.
PloS one, 4(9):e6855, 2009.
[6] James P O'Dwyer and Jessica L Green. Field theory for biogeography: a spatially explicit model for predicting
patterns of biodiversity. Ecology letters, 13(1):87 -- 95, 2010.
[7] Fernando AN Santos, Hermes Gadêlha, and Eamonn A Gaffney. Fock space, symbolic algebra, and analytical
solutions for small stochastic systems. Physical Review E, 92(6):062714, 2015.
[8] Henry Abarbanel. Predicting the future: completing models of observed complex systems. Springer, 2013.
10
Data assimilation in Agent-based models using creation and annihilation operators
A Appendix: Binomial observation
The product of a Deselby distribution and a binomial distribution is given by
(cid:18) k
(cid:19)
rm(1 − r)k−m (k)∆λk−∆e−λ
k!
P (km) ∝
rearranging and dropping constants
m
P (km) ∝ (1 − r)∆ (k)m
(k)∆((1 − r)λ)k−∆e−(1−r)λ
k!
(23)
expanding the product of falling factorials
min(m,∆)(cid:88)
q=0
P (km) = (1 − r)∆
m!∆! ((1 − r)λ)(m−q)
q!(m − q)!(∆ − q)!
(k)m+∆−q((1 − r)λ)k−(m+∆−q)e−(1−r)λ
k!
Expressing this as an operator on ∅
∞(cid:88)
(1 − r)∆
min(m,∆)(cid:88)
k=0
q=0
P ∝
so
m!∆! ((1 − r)λ)(m−q)
q!(m − q)!(∆ − q)!
(k)m+∆−q((1 − r)λ)k−(m+∆−q)e−(1−r)λ
k!
a†k∅
P ∝ (1 − r)∆
m!∆! ((1 − r)λ)(m−q)
q!(m − q)!(∆ − q)!
a†(m+∆−q)Di((1−r)λ) 0
But, from the commutation relations, we have the identity
ama†∆ =
m!∆!
q!(m − q)!(∆ − q)!
a†∆−qam−q
min(m,∆)(cid:88)
q=0
min(m,∆)(cid:88)
q=0
so
P ∝ (1 − r)∆a†mama†∆D((1−r)λ) 0
Let Lgr be an operator that has the properties
and
then
So
[Lgr, a†] = −ra†Lgr
[Lgr, a] =
ra
1 − r
Lgr
LgrDλ0 = D(1−r)λ0
Lgra†∆ = (1 − r)∆a†∆Lgr
(24)
(25)
P ∝ a
†m
i am
i LgirDiλ∆
11
Data assimilation in Agent-based models using creation and annihilation operators
B Appendix: Factorisation of compound observations
Consider an expression of the form
∞(cid:88)
n=0
(cid:80)
∅ · e
tn
n!
j aj ABH n
j aj ABetH =
where A, B and H are operators, and H is a Hamiltonian.
Due to the form of a Hamiltonian, ∅ · e
j aj HX = 0 for all X, so
(cid:80)
j aj ABH n = ∅ · e
(cid:80)
j aj [nAB, H]
Where we introduce the notation
[nX, H] = [...[[X, H], H]..., H]
to denote the n-fold application of a commutation.
So
(cid:80)
∅ · e
(cid:80)
∅ · e
(cid:80)
∅ · e
(cid:80)
∅ · e
tn
n!
∞(cid:88)
(cid:19)
(cid:18) n
n=0
j aj ABetH =
j aj [nAB, H]
n(cid:88)
(cid:19)(cid:0)[mA, H][n−m+1B, H] + [m+1A, H][n−mB, H](cid:1)
[mA, H][n−mB, H]
m=0
m
Cn(A, B) =
(cid:18) n
m
n(cid:88)
m=0
n+1(cid:88)
(cid:18) n
n+1(cid:88)
(cid:19)
(cid:18)n + 1
m(cid:48) − 1
(cid:19) m(cid:48)
m(cid:48)
n + 1
m(cid:48)=0
[mA, H][n+1−mB, H]
m(cid:48)=1
[mA, H][n+1−mB, H] +
n+1(cid:88)
(cid:18)n + 1
(cid:19)
m
m=0
=
[mA, H][n+1−mB, H] +
[m(cid:48)
A, H][n+1−m(cid:48)
B, H]
[Cn(A, B), H] = Cn+1(A, B)
[m(cid:48)
A, H][n+1−m(cid:48)
B, H]
[Cn(A, B), H] =
(cid:18) n
(cid:19)
n(cid:88)
(cid:18)n + 1
(cid:19) n + 1 − m
m=0
m
=
m
n + 1
(cid:19)
m
(cid:18) 1
(cid:19)
(cid:18) n
m=0
1(cid:88)
(cid:18) n
n(cid:88)
m
n(cid:88)
∞(cid:88)
m=0
[AB, H] = A[B, H] + [A, H]B =
[mA, H][1−mB, H] = C1(A, B)
[nAB, H] =
n=0
∞(cid:88)
n(cid:88)
∞(cid:88)
(cid:80)
m=0
n=0
=
[nAB, H]tn
=
n!
m
[mA, H]tm[n−mB, H]tn−m
m=0
n=0
m!(n − m!)
(cid:80)
∅ · e
j aj ABetH ψDΛ ≈ ∅ · e
j aj
[mA, H][n−mB, H]
(cid:19) [mA, H][n−mB, H]tn
∞(cid:88)
∞(cid:88)
[jA, H]tj
n!
=
[kB, H]tk
N(cid:88)
j=0
j=0
[jA, H]tj
j!
j!
N(cid:88)
k=0
k!
k=0
[kB, H]tk
k!
ψDΛ
Now let
so
n+1(cid:88)
m=0
=
So
But
so
So
So
where ψ is a sequence of creation operators and DΛ is the Deselby ground state.
If we let Φ(X) be the set of all states operated on by operator X, then if
(cid:33)
(cid:32) N(cid:88)
(cid:33)
(cid:32) N(cid:88)
n=0
n=0
12
Φ
[nA, H]
∩ Φ
[nB, H]
= ∅
(26)
Data assimilation in Agent-based models using creation and annihilation operators
where ∅ above is the empty set (not the ground state) then ψ can easily be factorised into ψAψB such that
N(cid:88)
(cid:80)
∅ · e
j aj
N(cid:88)
[jA, H]tj
[kB, H]tk
j!
j=0
k=0
k!
ψDΛ = ∅ · e
j aj
[jA, H]tj
[kB, H]tk
j!
j=0
ψA
k!
k=0
ψBDΛ
(cid:80)
N(cid:88)
N(cid:88)
But if X consists only of annihilation operators (which we can arrange by using commutation relations to move all
creation operators to the left and then removing them), then
(cid:80)
∅ · e
j aj XY D0 =
j aj XD0
j aj Y D0
+ ∅ · e
j aj [X, Y ]D0
but if equation 26 is satisfied then
(cid:80)
(cid:16)∅ · e
(cid:34) N(cid:88)
∅ · e
n=0
(cid:80)
(cid:17)
(cid:80)
(cid:80)
(cid:17)(cid:16)∅ · e
N(cid:88)
n=0
(cid:35)
(cid:32)
irrespective of whether we strip the first term of creation operators, so, if 26 is satisfied
(cid:80)
j aj ABetH ψDΛ ≈
∅ · e
[nA, H]ψA,
[nB, H]ψB
= 0
N(cid:88)
j aj
[jA, H]tj
ψDΛ
j!
j=0
N(cid:88)
(cid:80)
∅ · e
j aj
[kB, H]tk
k!
k=0
(cid:33)
ψDΛ
where we've removed the factorisation of ψ as it doesn't affect the sum.
13
|
1609.07015 | 2 | 1609 | 2016-10-25T21:01:50 | Distributed Consistent Data Association | [
"cs.MA",
"cs.RO",
"eess.SY"
] | Data association is one of the fundamental problems in multi-sensor systems. Most current techniques rely on pairwise data associations which can be spurious even after the employment of outlier rejection schemes. Considering multiple pairwise associations at once significantly increases accuracy and leads to consistency. In this work, we propose two fully decentralized methods for consistent global data association from pairwise data associations. The first method is a consensus algorithm on the set of doubly stochastic matrices. The second method is a decentralization of the spectral method proposed by Pachauri et al.. We demonstrate the effectiveness of both methods using theoretical analysis and experimental evaluation. | cs.MA | cs | Distributed Consistent Data Association
Spyridon Leonardos1, Xiaowei Zhou2 and Kostas Daniilidis3
6
1
0
2
t
c
O
5
2
]
A
M
.
s
c
[
2
v
5
1
0
7
0
.
9
0
6
1
:
v
i
X
r
a
Abstract -- Data association is one of the fundamental prob-
lems in multi-sensor systems. Most current techniques rely on
pairwise data associations which can be spurious even after the
employment of outlier rejection schemes. Considering multiple
pairwise associations at once significantly increases accuracy
and leads to consistency. In this work, we propose two fully
decentralized methods for consistent global data association
from pairwise data associations. The first method is a consensus
algorithm on the set of doubly stochastic matrices. The second
method is a decentralization of the spectral method proposed by
Pachauri et al. [28]. We demonstrate the effectiveness of both
methods using theoretical analysis and experimental evaluation.
I. INTRODUCTION
Multi-sensor data association has been a long standing
problem in robotics and computer vision. It refers to the
problem of establishing correspondences between feature
points, regions, or objects observed by different sensors
and serves as the basis for many high-level tasks such as
localization, mapping and planning. Most of the efforts in
previous works have been dedicated to improving the data as-
sociation by designing new feature detectors, descriptors, and
outlier removal algorithms in a pairwise setting. However,
the problem setting in practice is often multi-way if the data
are collected by a sensor network or a multi-robot system,
and how to establish consistent data association for multiple
sensors and leverage the global reasoning to improve the
association has drawn increasing attention in recent years.
A necessary condition for good data association among
multiple sensors is the cycle consistency meaning that the
composition of associations along a cycle of sensors should
be identity. The cycle consistency is often violated in practice
due to outliers and how to use such a constraint to remove
the false associations has been studied in robotics, computer
vision and graphics. The related work will be discussed
in Section II. While promising empirical and theoretical
results have been reported in previous works, most of them
addressed the problem in a centralized manner and assumed
that
the observations and states of all sensors could be
accessed at the same time. This assumption is impractical
in a distributed system where the data is processed on
local sensors with limited computational and communication
resources. In this paper, we aim to develop distributed
algorithms that can efficiently operate on each sensor, re-
quiring only local information and communication with its
1,2,3The authors are with the Department of Computer and Informa-
tion Science, University of Pennsylvania Philadelphia, PA 19104, USA
{spyridon,xiaowz,kostas}@seas.upenn.edu
Support by the grants ARL MAST-CTA W911NF-08-2-0004 and ARL
RCTA W911NF-10-2-0016 is gratefully acknowledged.
Fig. 1.
Illustration of consistent data association.
neighbors, and finally producing globally consistent data
association.
Our contributions are the following. 1) We propose a novel
consensus-like algorithm for distributed data association.
2) We propose a decentralized version of the centralized
state-of-art spectral method [28]. 3) We show that both
methods provably converge, do not depend on initialization,
are parameter-free and guarantee global consistency. 4) We
demonstrate the validity of proposed methods through both
theoretical analysis and experimentation with synthetic and
real data.
The rest of this paper is organized as follows. Section II
includes a brief review of prior works. Section III introduces
the reader to notation and preliminaries essential to under-
standing the main results of this paper. A rigorous problem
formulation is the topic of Section IV. The two proposed
methods are presented in Sections V and VI respectively.
Finally, experimental evaluation is presented in Section VII.
II. RELATED WORK
The cycle consistency has been explored in a variety of
computer vision applications. For example, Zach et al. [37]
studied the problem of multi-view reconstruction and pro-
posed to identify unreliable geometric relations (e.g. relative
poses) between views by measuring the consistencies along a
number of cycles. Nguyen et al. [26] addressed the problem
of finding point-wise maps among a collection of shapes
and proposed to use the cycle consistency to identify the
correctness of maps and replace the incorrect maps with the
compositions of correct ones. Yan et al. [35], [36] consid-
ered the multiple graph matching problem and imposed the
cycle consistency by penalizing the inconsistency during the
optimization. Zhou et al. [38] developed a discrete searching
algorithm that maximizes the cycle consistency to improve
the dense optical flow in an image collection. Instead of
iteratively optimizing the pairwise matches, [15], [21], [28]
showed that finding globally consistent associations from
noisy pairwise associations can be formulated as a quadratic
integer program and relaxed into a generalized Rayleigh
problem and solved by the eigenvalue decomposition. More
recently, Huang and Guibas [14] and Chen et al. [6] the-
oretically proved that the consistent associations could be
exactly recovered from noisy measurements under certain
conditions by convex relaxation and solved the problem by
semidefinite programming. Zhou et al. [39] proposed to solve
the problem by rank minimization and developed a more
efficient algorithm. All of these works addressed problem
in a centralized setting, i.e., all pairwise measurements are
available and optimized jointly.
In the robotics community, numerous approaches for as-
sociating elements of two sets have been proposed and
extensively used in single robot Simultaneous Localization
And Mapping (SLAM) settings. These include but are not
limited to Nearest Neighbor (NN) and Maximum Likelihood
(ML) approaches [18], [40], Iterative Closest Point (ICP)
[5], RANSAC [8], Joint Compatibility Branch and Bound
(JCBB) [25], Maximum Common Subgraph (MCS) [3],
Random Finite Sets (RFS) [33]. More recently, multi-robot
localization and mapping algorithms have been developed. In
almost all of them, the data association is addressed either in
a pairwise fashion by directly generalizing data association
techniques from the single robot case [7], [9], [34] or by
considering all associations jointly in a centralized manner
[16]. Closer to our approach is the work by Aragues et al. [1]
which is a decentralized method for finding inconsistencies
based on cycle detection. However, the proposed inconsis-
tency resolution algorithm comes without any optimality or
guarantees.
III. PRELIMINARIES, NOTATION AND
DEFINITIONS
A. Graph Theory
In this subsection, we review some elementary facts from
graph theory. For an in depth analysis, we refer the reader
to standard texts [10], [24]. An undirected graph or simply
a graph is denoted by the pair G = (V,E), where V =
{1, 2, . . . , n} is the set of vertices and E ⊆ [V]2 is the
set of edges, where [V]2 denotes the set of unordered pairs
of elements of V. The neighborhood Ni of the vertex i is
the subset of V defined by Ni = {j ∈ V {i, j} ∈ E}. A
path is a sequence i0, i1, . . . , im of distinct vertices such that
{ik−1, ik} ∈ E for all k = 1, . . . , m. A graph is connected
if there is a path between any two vertices. Given a graph
G = (V,E), its adjacency matrix is defined by
if {i, j} ∈ E
(cid:26) 1,
[A(G)]ij =
(1)
0, otherwise
The degree matrix ∆(G) is the diagonal matrix such that
[∆(G)]ii = Ni, where · denotes the cardinality of a set.
A directed graph or digraph is denoted by the pair G =
(V,E), where E ⊆ V×V. A weighted digraph G = (V,E, w)
is a graph along with a function w : E → R+. The adjacency
matrix of a weighted digraph is defined by
if (j, i) ∈ E
otherwise
[A(G)]ij =
0,
(2)
Intuitively, if [A(G)]ij > 0 there is information flow from
vertex j to vertex i. The neighborhood Ni of the vertex i is
the subset of V defined by Ni = {j ∈ V (j, i) ∈ E}. The
degree matrix ∆(G)
(cid:26) w(j, i),
(cid:88)
j∈Ni
[∆(G)]ii = din(i)
.
=
[A(G)]ij
(3)
A directed path is a sequence i0, i1, . . . , im of distinct ver-
tices such that (ik−1, ik) ∈ E for all k = 1, . . . , m. A digraph
is strongly connected if there is a directed path between any
two vertices. A digraph is a rooted out-branching tree if it has
a vertex to which all other vertices are path connected and
does not contain any cycles. A digraph is balanced if the in-
(i,j)∈E [A(G)]ji
degree din(i) and the out-degree dout(i)
are equal for all i ∈ V. From this point and on, all graphs
under consideration are assumed to be directed unless stated
otherwise. The graph Laplacian L(G) is defined as
=(cid:80)
.
L(G) = ∆(G) − A(G)
(4)
By construction, L(G)1 = 0. A digraph on n vertices
contains a rooted-out branching if and only the rank of its
Laplacian is n − 1. A matrix closely related to the graph
Laplacian is defined by
F (G)
= (I + ∆(G))−1(I + A(G))
.
(5)
B. Stochastic Matrices and Permutations
Stochastic matrices and their properties have been well
studied in the area of distributed dynamical systems [4], [17],
[27] and in probability theory in the context of computing
invariant distributions of Markov chains [2], [4]. A nonneg-
ative square matrix is stochastic if all its row sums are equal
to 1. The spectral radius ρ(A) of a stochastic matrix A is
equal to 1 and it is an eigenvalue of A since A1 = 1. The
existence and the form of the limit limk→∞ Ak depends on
the multiplicity of 1 as an eigenvalue. A nonegative square
matrix A is irreducible if its associated graph [13] is strongly
connected. An irreducible stochastic matrix is primitive if it
has only one eigenvalue of maximum modulus.
A nonnegative square matrix is doubly stochastic matrix
if both its row sums and column sums are equal
to 1.
A doubly stochastic matrix is a permutation matrix if its
elements are either 0 or 1. The sets of stochastic matrices
and doubly stochastic matrices are compact convex sets. This
is a rather useful property since it implies closure under
convex combinations. For more details, regarding the theory
of stochastic matrices, we refer the reader to [13].
= {1, 2, . . . , n} for some positive integer n. A
.
if it
is a permutation of [n]
[n] → [n]
mapping π :
Let [n]
(cid:26) 1,
is bijective. The set of all permutations of [n] forms a
group under composition, termed the symmetric group Sn. A
permutation π ∈ Sn is represented 1 by a n× n permutation
matrix Π defined by
[Π]ij =
if π(j) = i
0, otherwise
(6)
or equivalently Πej = eπ(j), where ej is the jth canonical
basis vector. The simplest choice for a distance on Sn is
given by
d(π1, π2) = d(e, π−1
1 π2)
= n − (cid:104)Π1, Π2(cid:105)
.
(7)
where (cid:104)A, B(cid:105) .
= tr(AT B), e is the identity map and Π1, Π2
are the matrix representations of the permutations π1 and π2
respectively. The distance function defined above is simply
the number of labels assigned differently by permutations π1
and π2. With a slight abuse of notation, we write Π ∈ Sn
meaning that Π is an n × n permutation matrix.
C. Consensus Algorithms
Consensus algorithms have been extensively studied in the
control community [4], [17], [27]. In its simplest form, a
consensus algorithm is a decentralized protocol in which the
agents, modeled as vertices of a graph, try to reach agreement
by communicating only with their neighbors. More formally,
let xi(t) ∈ R denote the state of agent i at time time t. The,
the simplest consensus protocol is given by
(cid:88)
j∈Ni∪i
aijxj(t)
(8)
xi(t + 1) =
where aij ≥ 0 and (cid:80)
j aij = 1. A popular consensus
protocol [27], which converges to the average of the initial
values provided that G is balanced, is described by
x(t + 1) = (I − L(G))x(t)
(9)
where x = [x1, . . . , xn]T and 0 < < 1/ maxi[∆(G)]ii.
IV. PROBLEM FORMALIZATION
In this section, we formalize the problem of consistent
data association. We assume there are n sensors observing
m targets. For instance, assume we have a set of n cameras
observing a scene in the world described by a set of m
feature points. Sensors communicate only with a subset of
all sensors. Communication constraints between sensors are
encoded by the sensor graph. The sensor graph is the digraph
G = (V,E) where V = {1, 2, . . . , n} and (i, j) ∈ E if there
is information flow from sensor i to j. The pairwise associa-
tion πij ∈ Sm is defined as follows: we have that πij(l) = k
if the lth target in jth sensor corresponds to the kth target in
ith sensor. Observe that the pairwise associations πij can be
written as πij = πi ◦ π−1
, where πi ∈ Sm is the mapping
from the labels of sensor i to some global labels, termed
the "universe of features" in some works [6], [39]. This is
analogous to the absolute position in Euclidean spaces in
1A representation of a group G on Rn is a map ρ : G → GL(n)
j
satisfying ρ(g1g2) = ρ(g1)ρ(g2).
S1
S2
S3
1
2
3
1
2
3
1
2
3
1
2
3
1
2
3
1
2
3
.
= ρ(πi).
Fig. 2. Example with n = 3 sensors S1, S2, S3 observing m = 3 targets.
Top: consistent data association. Bottom: inconsistent data association since
π12 ◦ π23(2) = 3 but π31(2) = 2.
localization problems. We denote by(cid:101)πij ∈ Sm the, possibly
and j and by (cid:101)Πij the corresponding matrix representation.
erroneous, estimated pairwise association between sensor i
First, we need a precise definition of consistency.
Definition 4.1 (Consistency): A set of pairwise associa-
Moreover, let Πi
Related to the sensor graph is the data association graph
D = (VD,ED, wD), where VD = V × {1, 2, . . . , m}. There
is an edge from (i, k) to (j, l) if and only if (i, j) ∈ E and
[(cid:101)Πij]kl > 0. The corresponding edge weight is simply equal
to [(cid:101)Πij]kl.
tions {(cid:101)πij}(i,j)∈E is consistent if
{(cid:101)πij}(i,j)∈E, if(cid:101)πij = πi ◦ π−1
(10)
for all valid indices i, j, k. A set of labels {πi}n
i=1 is con-
sistent with respect to some consistent pairwise associations
(11)
Based on the definition of consistency, the problem of
consistent data association is naturally defined as follows.
(cid:101)πij ◦(cid:101)πjk =(cid:101)πik
∀(i, j) ∈ E
,
j
Definition 4.2 (Consistent data association):
associations {(cid:101)πij}(i,j)∈E, find labels
Given pairwise
π1, . . . , πn ∈ Sm, such that
(cid:101)πij = πi ◦ π−1
j
∀ (i, j) ∈ E
,
(12)
Remark 1: Under the presence of noise, it might not be
i=1 satisfying (12) exactly. There-
i=1 satisfying
possible to find labels {πi}n
fore, in practice we are looking for labels {πi}n
(12) as much as possible according to some criterion.
V. DISTRIBUTED AVERAGING
In the classic consensus algorithms, each agent updates his
estimate of a collective quantity by taking convex combina-
tions of the estimates of his neighbors. The problem at hand
is different: we do not want all πi's to converge to the same
value, rather to converge to a value satisfying consistency.
The first obstacle one encounters is the combinatorial nature
of the problem at hand. Therefore, we relax the domain of
problem and work with doubly stochastic matrices instead
of permutation matrices. This is a natural choice since the
convex hull of the set of m × m permutation matrices is
exactly the set of m × m doubly stochastic matrices.
A. Algorithm
Based on the previous discussion, we propose the follow-
ing update rule:
Πi(t + 1) =
1
Ni + 1
(cid:0)Πi(t) +
(cid:88)
j∈Ni
(cid:101)ΠijΠj(t)(cid:1)
(13)
Let Π = [ΠT
1 , . . . , ΠT
n ]T . The update rule (13) can be
collectively written as
Π(t + 1) = F (D)Π(t)
(14)
where D is the data association graph and F (D) is defined
in (5). The convergence properties of the update rule (13) de-
pend solely on limk→∞ F (D)k and Π(0). These properties
will be presented in Section V-B.
After the convergence of Protocol (13), we discretize the
solution by solving the assignment problem, formally defined
by
(cid:104)(cid:101)Πi, Πi(cid:105)
maximize
subject to Πi ∈ Sm
Πi
(15)
(16)
(17)
where (cid:101)Πi
is the result of Protocol (13). The assignment
problem can be solved using the Hungarian algorithm [22]
in O(m3) time.
Remark 2: The update rule (13) can be seen as a discrete-
time system with state Πi and control input Ui, i.e.
Πi(t + 1) = Πi(t) + Ui(t)
Ui(t) =
1
Ni + 1
(cid:0)(cid:101)ΠijΠj(t) − Πi(t)(cid:1)
(cid:88)
j∈Ni
Remark 3: (Intrinsic ambiguity) Let {Πi}n
i=1 ∈ (Sm)n
be a set of consistent labels. Then, for any Π0 ∈ Sm,
{ΠiΠ0}n
i=1 are consistent as well. This intrinsic ambiguity is
analogous to the ambiguity of the global reference frame in
rotation localization [30]. To remove this ambiguity, one can
fix the label of one sensor, say the first sensor, by modifying
the sensor graph G so that the first sensor has only outgoing
edges resulting in Π1(t) = Π1(0) ∈ Sm for all t ≥ 0. This
approach is necessary in the presence of noise and we will
refer to it as Protocol (13) with a distinguished vertex.
Remark 4: (Equivalence to consensus) In the noiseless
j0 for some consistent set
i=1 ∈ (Sm)n . By defining new variables
case, we can write (cid:101)Πij = Πi0Π−1
(cid:0)Π(cid:48)
= Π−1
.
i0 Πi, we obtain
Π(cid:48)
i(t + 1) =
of labels {Πi0}n
Π(cid:48)
j(t)(cid:1)
(cid:88)
i(t) +
(18)
Π(cid:48)
1
i
Ni + 1
j∈Ni
which is a component-wise Vicsek model [17], [32].
B. Properties
First of all, as the sets of stochastic and doubly stochastic
matrices are convex and thus, closed under convex combi-
nations, we naturally have the following lemma.
Lemma 5.1: If all(cid:101)Πij and Πi(0) are stochastic then, Πi(t)
is stochastic for all t ≥ 0. Similarly, if all (cid:101)Πij and Πi(0)
are doubly stochastic then, Πi(t) is doubly stochastic for all
t ≥ 0.
Next, we show that in the noiseless case Protocol (13)
with a distinguished vertex converges to a consistent solution
under mild coniditions on the sensor graph. To prove this,
we need the following lemma:
Lemma 5.2: Given a digraph G that contains a rooted-out
branching tree, we have that the matrix F (G) as defined
in (5) satisfies: (a) ρ(F (G)) = 1, (b) 1 is an algebraically
simple eigenvalue of F (G) with corresponding eigenvector
√
1/
n and (c) the limit F (G) exists and is given by
k→∞ F (G)k = 1cT ,
lim
(19)
where c ≥ 0 and cT 1 = 1.
A proof of lemma 5.2 can be found in Appendix IX-A. Next,
we present the theorem of convergence of Protocol 13 to a
consistent labeling from arbitrary initialization. A proof of
theorem 5.3 is presented in Appendix IX-B and heavily relies
on lemma 5.2.
Theorem 5.3: If the sensor graph G contains a rooted-
out branching and the pairwise associations are noiseless,
then the consensus protocol with the distinguished vertex
converges to a set of consistent
labels up to a global
permutation, that is
(20)
lim
t→∞ Πi(t) = Πi0Π0
where {Πi0}n
to (cid:101)Πij) and Π0 is an arbitrary permutation. Furthermore, if
i=1 is a consistent set of labels (with respect
vertex 1 is the distinguished vertex, then Π0 = ΠT
10Π1(0).
However, in the general case, pairwise associations will
contain noise. Therefore, we should prove convergence with-
out assuming perfect associations. We generalize lemma 5.2
for the case of interest. Consider a digraph D with n vertices
and m < n distinguished vertices that have only outgoing
edges. Assume that for every non-distinguished vertex there
is a (directed) path from at least one distinguished vertex.
Then, the succeeding lemma provides us with the limit of
F (D).
Lemma 5.4: The matrix F (D) as defined in (5) satisfies:
(a) ρ(F (D)) = 1, (b) F (D) has 1 as an eigenvalue with
algebraic and geometric multiplicities equal to the number
of distinguished vertices, (c) it asymptotically converges to
a limit of the form
k→∞ F (D)k =
lim
(21)
A proof of lemma 5.4 is presented in Appendix IX-C.
An immediate consequence of lemma 5.4 is the following
theorem.
Theorem 5.5: If the sensor graph G contains a rooted-out
branching, then the consensus protocol with a distinguished
(cid:20) Im 0
(cid:21)
(cid:101)F21
0
vertex asymptotically converges to a solution that depends
only on {(cid:101)Πij}(i,j)∈E and on the initial value of the labels of
the distinguished vertex.
Remark 5: Although Protocol (13) is not guaranteed to
converge to the true solution in the presence of noise, it is
verified experimentally (see Section VII) that for moderate
levels of noise the true solution is still recovered.
VI. DISTRIBUTED ORTHOGONAL ITERATION
In this section, we present a decentralized implementation
of the spectral method proposed by Pachauri et al. [28]. First,
we briefly review the main concept of [28]. Based on the
discussion in Section IV, the multi-sensor data association
can be cast as the following optimization problem:
(22)
(23)
minimize
π1,π2,...,πn∈Sm
which is equivalent to
maximize
Π1,Π2,...,Πn∈Sm
(cid:88)
(cid:88)
(i,j)∈E
(i,j)∈E
)
j
d((cid:101)πij, πi ◦ π−1
i (cid:101)ΠijΠj)
tr(ΠT
In the above formulation, each Πi is a permutation matrix
which makes the problem computationally intractable. How-
ever, if the individual constraints are relaxed into a single col-
lective constraint of the form ΠT Π = mIm as proposed in
[28], the relaxed problem becomes computationally tractable
since it becomes a Rayleigh quotient problem:
maximize
tr(ΠT PΠ)
Π
(24)
where [P]ij = (cid:101)Πij. It is a well-known fact that the solution
subject to ΠT Π = mIm
of problem (24) is given by the matrix having as columns
m.
the m leading eigenvectors of P scaled by a factor of
Numerically, the optimal solution in a centralized setting can
be computed using Orthogonal Iteration [11] (see Algorithm
1). Intuitively, each iteration consists of two main steps:
a power step and an orthogonalization step based on QR-
decomposition. For details on the convergence of Orthogonal
Iteration we refer the reader to [11] and references therein.
√
Algorithm 1 Orthogonal Iteration
Require: Input matrix P, initial Π(0)
1: for t = 0, . . . , N do
2:
3:
4:
5: end for
Y(t) = PΠ(t)
Q(t)R(t) = Y(t)
Π(t + 1) = Q(t)
The Orthogonal Iteration method is not readily decentral-
izable since the QR-decomposition step requires information
from the entire sensor graph. Nevertheless, as observed
by Kempe and McSherry [20],
n ]T
where each sub-matrix Yi is available at sensor i, we obtain
i Yi can
be computed in a decentralized manner, then we can recover
R(t) using Cholesky factorization. Let Zi
i Yi and
RT R = YT Y =(cid:80)n
i Yi. If the sum (cid:80)n
i=1 Y T
.
= nY T
if Y = [Y T
1 , . . . , Y T
i=1 Y T
(cid:80)n
.
= 1
n
i=1 Zi = RT R. Based on this observation, if all
Z
sensors know n and solve a consensus problem for com-
puting the average Z = RT R, then the orthogonalization
step can be performed locally using Cholesky decomposition.
This is summarized in Algorithm 2 where Z i denotes the
i Yi
can be computed using the Push-Sum algorithm [19] as in
[20].
estimate of Z by sensor i. Alternatively, the sum(cid:80)n
i=1 Y T
Remark 6: Kempe and McSherry [20] have shown that if
the error in the computation of Ri can be made arbitrarily
small in finite iterations, then Distributed Orthogonal Iter-
ation converges asymptotically an invariant subspace of the
input matrix. A rigorous analysis on the number of iterations
of the inner consensus is the subject of future work.
Algorithm 2 Distributed Orthogonal Iteration
Require: Input associations {(cid:101)Πij}(i,j)∈E, initial {Πi(0)}i∈V
Yi(t) =(cid:80)
j∈Ni∪{i}(cid:101)ΠijΠj(t)
1: for t = 0, . . . , N do
2:
3:
4:
5:
6: end for
Estimate Z i(t) by consensus.
Ri(t) = chol(Z i(t))
Πi(t + 1) = Yi(t)Ri(t)−1
Remark 7: Due to the nonlinear dependence of Zi on
Yi, dynamic consensus algorithms [29] cannot be readily
incorporated into the current approach.
The solution obtained by the Orthogonal Iteration is opti-
mal up to an arbitrary m×m orthogonal transformation. That
is, if Π(cid:63) is an optimal solution of (24), then any matrix of the
form Π(cid:63)Q with Q ∈ O(m) is optimal as well. Finally, the
corrective orthogonal transformation is computed by solving
the following Orthogonal Procrustes [11] problem:
minimize
Q∈O(m)
(cid:107)I − Π1Q(cid:107)2
F
(25)
It is well known [11] that if Π1 = U SV T is an SVD of Π1,
then Q(cid:63) = V U T . As a final step, we apply the corrective
transformation Q to all remaining Πi, i = 2, . . . , n and then
find a discrete solution by the Hungarian algorithm.
VII. EXPERIMENTS
A. Synthetic data
First, we experiment with synthetic data to validate the
accuracy of proposed methods across different problem
settings. We generate instances with m = 50 targets and
n ∈ {20, 50, 100} sensors. As a performance criterion we use
the accuracy of the obtained labels, which is simply defined
as the percentage of correct labels. We vary the percentage
of erroneous associations in each pairwise association from
10% to 90% to test the robustness of each method against
the presence of outliers. We first experiment with a fully
connected graph, see Fig. 3 (a)-(c), and then we remove half
of the edges, see Fig. 3 (d)-(f). In the case of a fully con-
nected graph, the consensus protocol achieves exact recovery
of the labels for up to approximately 40% of outliers. The
spectral method clearly outperforms the consensus protocol
and is able to achieve exact recovery for up to 80%-90%
of outliers. By reducing either the number of edges in the
sensor graph or the number of sensors, the accuracy of both
methods slightly drops. As expected, the benefit of using
joint matching in accuracy increases with the size of the
network and the number of edges.
B. Real data
First, we experiment with the CMU hotel sequence 2
which consists of 111 frames and the CMU house 3 sequence
which consists of 101 frames. We detect corner features
[12] in the first image that remain visible across the entire
sequence. The ground-truth associations are obtained by
tracking the features across the entire sequence. Specifically,
we used 89 points in the hotel sequence and 55 in the house
sequence. Pairwise matches are computed between an image
and at most 10 previous images and 10 subsequent images
in the sequence, thus the sensor graph is relatively sparse.
Pairwise matches are obtained by extracting SIFT descriptors
[23] and then solving the assignment problem using the
Hungarian algorithm. We used the SIFT implementation
of [31]. The baseline method consists of obtaining labels
by directly matching with the first image in the sequence.
The ground-truth associations are extracted by tracking the
features over the sequence. The results are presented in
Table I. Both methods significantly outperform the baseline
which shows that joint matching can significantly improve
the overall matching performance.
TABLE I
CMU DATASETS RESULTS
Dataset
House
Hotel
Baseline
0.716
0.666
Spectral
0.861
0.872
Consensus
0.887
0.878
Next, we experiment with the Affine Covariant Regions
Datasets 4 which consist of sequences of 6 images with
significant overlap but with viewpoint variability. We detect
approximately 300 SIFT features [23] in each sequence
that are visible across the entire sequence. The ground-
truth associations are extracted using the available ground-
truth homographies. The results are presented in Table II.
Both methods significantly outperform the baseline in the
two challenging sequences Graffiti and Wall. However, in
the remaining sequences the input pairwise associations are
already very accurate resulting in only minor improvements
if any. It
the spectral method and
the consensus method have similar performance in all real
sequences despite the fact that the spectral method is com-
putationally more expensive. Therefore, in situations where
the computational resources are limited, for instance swarms
of UAVs with on-board cameras, the consensus algorithm
would be preferable.
is worth noting that
2 http://vasc.ri.cmu.edu//idb/html/motion/hotel/
3http://vasc.ri.cmu.edu//idb/html/motion/house
4http://www.robots.ox.ac.uk/vgg/data/data-aff.
html
TABLE II
GRAFFITI DATASETS RESULTS
Dataset
Graffiti
Wall
Trees
Leuven
Bikes
UBC
Baseline
0.544
0.570
0.912
0.962
0.986
0.996
Spectral
0.639
0.646
0.902
0.972
0.987
0.996
Consensus
0.647
0.658
0.909
0.979
0.987
0.996
VIII. CONCLUSIONS AND FUTURE WORK
In this paper, we proposed two fully decentralized methods
for the problem of data association in sensor networks.
The first was a computationally inexpensive consensus-like
algorithm and the second a decentralized spectral method.
We presented experimental results in the context of camera
sensor networks. We believe that the results of this work
will find applications in other settings as well. In the future,
we plan on generalizing these algorithms in order to handle
different number of targets in each sensor and time-varying
pairwise associations.
IX. APPENDIX
A. Proof of lemma 5.2
First, we need the following lemma:
Lemma 9.1: Let Li(G) be the matrix obtained from the
Laplacian L(G) by removing the ith row and ith column.
Define Fi(G) analogously. Then,
det(I − Fi(G)) (cid:54)= 0
det(Li(G)) (cid:54)= 0 iff
(26)
Proof: By construction L(G) = (I + ∆(G))(I − F (G))
and thus, Li(G) = (I + ∆i(G))(I − Fi(G)) since ∆(G) is
diagonal. Therefore, since det(I + ∆i(G)) > 0, we have that
det(Li(G)) (cid:54)= 0 if and only if det(I − Fi(G)) (cid:54)= 0.
Now, we are ready to proceed to the proof of lemma 5.2
By Gersgorin's discs theorem, we know that the eigenvalues
of F (G) lie in the unit circle and since it is stochastic, that is
F 1 = 1, it can be immediately concluded that ρ(F (G)) = 1.
From Proposition 3.8 page 51 of [24], we have that if
the (directed) graph G contains a rooted-out branching as a
subgraph, then rank L(G) = n − 1. Combining this result
with 9.1, we conclude that the existence of a rooted-out
branching implies that rank(I−F (G)) = n−1. We conclude
that 1 is a simple eigenvalue of F (G).
Let λ1, λ2, . . . , λn be an enumeration of the eigenvalues of
F (G) such that λ1 = 1. Moreover, let F (G) = P J(Λ)P −1
be the Jordan decomposition of F (G). From, F (G)P =
P J(Λ), it is easy to see that the the first column of P is in
the span of 1 and thus, limk→∞ F (G)k = 1cT . Moreover,
since F (G)k is stochastic for all positive integers k, it follows
that c ≥ 0 and cT 1 = 1.
B. Proof of theorem 5.3
In the noiseless case, observe that
F (D) = Π0 (F (G) ⊗ Im) ΠT
0 ,
(27)
1
0.8
0.6
0.4
0.2
1
0.8
0.6
0.4
0.2
y
c
a
r
u
c
c
a
y
c
a
r
u
c
c
a
CS
SP
CS
SP
y
c
a
r
u
c
c
a
y
c
a
r
u
c
c
a
1
0.8
0.6
0.4
0.2
0
0
0.2
0.6
0.4
0.8
Fraction of outliers
1
(b) n = 50
1
0.8
0.6
0.4
0.2
0
0
0.2
0.6
0.4
0.8
Fraction of outliers
1
(e) n = 50
CS
SP
CS
SP
y
c
a
r
u
c
c
a
y
c
a
r
u
c
c
a
CS
SP
CS
SP
1
0.8
0.6
0.4
0.2
0
0
0.2
0.6
0.4
0.8
Fraction of outliers
1
(c) n = 100
1
0.8
0.6
0.4
0.2
0
0
0.2
0.6
0.4
0.8
Fraction of outliers
1
(f) n = 100
0
0.2
0.6
0.4
0.8
Fraction of outliers
1
(a) n = 20
0
0.2
0.6
0.4
0.8
Fraction of outliers
1
(d) n = 20
Fig. 3. Accuracy of consensus protocol (CS) and spectral method (SP) versus percentage of outliers in pairwise data associations.
Fig. 4.
consensus algorithm.
Top: initial matches. Blue color corresponds to inlier matches whereas red color corresponds to outlier matches. Bottom: matches after the
where Π0 = diag(Π10, . . . , Πn0), each Πi0 ∈ Sm satisfies
Πi0ΠT
j0 = (cid:101)Πij.
k→∞ F (D)k = Π0 (F (G)∞ ⊗ Im) ΠT
lim
0
(28)
since limk→∞ F (G)k exists as we proved earlier. It follows
that F (D)k converges to a limit as well. Any asymp-
totic solution of the consensus protocol satisfies Π(∞) =
F (D)Π(∞). Then, for all i ∈ {1, . . . , n} we have Πi(∞) =
j0, we get
(1/Ni)(cid:80)
j∈Ni(cid:101)ΠijΠj(∞). Using (cid:101)Πij = Πi0ΠT
= 0
ΠT
(L(G) ⊗ Im)
10Π1(∞)
...
n0Πn(∞)
ΠT
Under the assumption that G is contains a rooted-out branch-
ing and thus, null(L(G)) = span(1), we obtain Πi(∞) =
Πi0Q where Q is an arbitrary matrix Q. Moreover, if one
enforces Π1(t) = Π1(0), Π1(0) ∈ Sm, for all t ≥ 0, it
10Π1(0),
follows that Q = ΠT
that is equal to any valid labels up to a global permutation.
(cid:4)
10Π1(0 and Πi(∞) = Πi0ΠT
C. Proof of lemma 5.4
Part (a) follows from the fact that F (G) is stochastic.
Without loss of generality, assume that vertices {1, 2, . . . , m}
are the distinguished ones. Then, F (G) takes the form
(cid:20) Im
(cid:21)
F (G) =
0
F21 F22
(29)
For all k = 1, 2, . . . we have by a simple induction:
(cid:20) Im
(cid:21)
F (G)k =
0
F21,k F k
22
(30)
where F21,1 = F21, F21,k+1 = F21,k + F k
22F21
By the assumption that every non-distinguished vertex is
path-connected to some distinguished vertex, it follows that
for k > 0 large enough, each row of F21,k contains at least
one positive entry. Thus, all eigenvalues of F k
22 lie strictly
(cid:4)
inside the unit circle and thus, limk→∞ F k
22 = 0.
REFERENCES
[1] Rosario Aragues, Eduardo Montijano, and Carlos Sagues. Consistent
data association in multi-robot systems with limited communications.
In Robotics: Science and Systems, pages 97 -- 104, 2011.
[2] Søren Asmussen. Applied probability and queues, volume 51. Springer
Science & Business Media, 2008.
[3] Tim Bailey and Eduardo Nebot. Localisation in large-scale environ-
ments. Robotics and autonomous systems, 37(4):261 -- 281, 2001.
[4] Dimitri P Bertsekas and John N Tsitsiklis. Parallel and distributed
computation: numerical methods, volume 23. Prentice hall Englewood
Cliffs, NJ, 1989.
[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes.
In Robotics-DL tentative, pages 586 -- 606. International Society for
Optics and Photonics, 1992.
[6] Yuxin Chen, Leonidas Guibas, and Qixing Huang. Near-optimal joint
In International Conference
object matching via convex relaxation.
on Machine Learning, 2014.
[7] Alexander Cunningham, Kai M Wurm, Wolfram Burgard, and Frank
Dellaert. Fully distributed scalable smoothing and mapping with robust
multi-robot data association. In Robotics and Automation (ICRA), 2012
IEEE International Conference on, pages 1093 -- 1100. IEEE, 2012.
[8] Martin A Fischler and Robert C Bolles. Random sample consensus:
a paradigm for model fitting with applications to image analysis and
automated cartography. Communications of the ACM, 24(6):381 -- 395,
1981.
[9] Dieter Fox, Jonathan Ko, Kurt Konolige, Benson Limketkai, Dirk
Schulz, and Benjamin Stewart. Distributed multirobot exploration and
mapping. Proceedings of the IEEE, 94(7):1325 -- 1339, 2006.
[10] Chris Godsil and Gordon F Royle. Algebraic graph theory, volume
207. Springer Science & Business Media, 2013.
[11] Gene H Golub and Charles F Van Loan. Matrix computations,
volume 3. JHU Press, 2012.
[12] Chris Harris and Mike Stephens. A combined corner and edge detector.
In Alvey vision conference, volume 15, page 50. Citeseer, 1988.
[13] Roger A Horn and Charles R Johnson. Matrix analysis. Cambridge
university press, 2012.
[14] Qi-Xing Huang and Leonidas Guibas. Consistent shape maps via
semidefinite programming. Computer Graphics Forum, 32(5):177 --
186, 2013.
[15] Qi-Xing Huang, Guo-Xin Zhang, Lin Gao, Shi-Min Hu, Adrian
Butscher, and Leonidas Guibas. An optimization approach for ex-
tracting and encoding consistent maps in a shape collection. ACM
Transactions on Graphics, 31(6):167, 2012.
[16] Vadim Indelman, Erik Nelson, Nathan Michael, and Frank Dellaert.
Multi-robot pose graph localization and data association from un-
In 2014
known initial relative poses via expectation maximization.
IEEE International Conference on Robotics and Automation (ICRA),
pages 593 -- 600. IEEE, 2014.
[17] Ali Jadbabaie, Jie Lin, and A Stephen Morse. Coordination of groups
IEEE
of mobile autonomous agents using nearest neighbor rules.
Transactions on automatic control, 48(6):988 -- 1001, 2003.
[18] Michael Kaess and Frank Dellaert. Covariance recovery from a square
root information matrix for data association. Robotics and autonomous
systems, 57(12):1198 -- 1210, 2009.
[19] David Kempe, Alin Dobra, and Johannes Gehrke. Gossip-based
In Foundations of Computer
computation of aggregate information.
Science, 2003. Proceedings. 44th Annual IEEE Symposium on, pages
482 -- 491. IEEE, 2003.
[20] David Kempe and Frank McSherry. A decentralized algorithm for
In Proceedings of the thirty-sixth annual ACM
spectral analysis.
symposium on Theory of computing, pages 561 -- 568. ACM, 2004.
[21] Vladimir G Kim, Wilmot Li, Niloy J Mitra, Stephen DiVerdi, and
Thomas A Funkhouser. Exploring collections of 3d models using
fuzzy correspondences. ACM Transactions on Graphics, 31(4):54,
2012.
[22] Harold W Kuhn. The hungarian method for the assignment problem.
Naval research logistics quarterly, 2(1-2):83 -- 97, 1955.
[23] David G Lowe. Distinctive image features from scale-invariant
International journal of computer vision, 60(2):91 -- 110,
keypoints.
2004.
[24] Mehran Mesbahi and Magnus Egerstedt. Graph theoretic methods in
multiagent networks. Princeton University Press, 2010.
[25] Jos´e Neira and Juan D Tard´os. Data association in stochastic mapping
using the joint compatibility test. IEEE Transactions on robotics and
automation, 17(6):890 -- 897, 2001.
[26] Andy Nguyen, Mirela Ben-Chen, Katarzyna Welnicka, Yinyu Ye, and
Leonidas Guibas. An optimization approach to improving collections
of shape maps. Computer Graphics Forum, 30(5):1481 -- 1491, 2011.
[27] Reza Olfati-Saber, J Alex Fax, and Richard M Murray. Consensus
and cooperation in networked multi-agent systems. Proceedings of
the IEEE, 95(1):215 -- 233, 2007.
[28] Deepti Pachauri, Risi Kondor, and Vikas Singh. Solving the multi-
way matching problem by permutation synchronization. In Advances
in neural information processing systems, pages 1860 -- 1868, 2013.
[29] Demetri P Spanos, Reza Olfati-Saber, and Richard M Murray. Dy-
namic consensus on mobile networks. In IFAC world congress, pages
1 -- 6. Prague Czech Republic, 2005.
[30] Roberto Tron and Rene Vidal. Distributed 3-d localization of camera
sensor networks from 2-d image measurements. Automatic Control,
IEEE Transactions on, 59(12):3325 -- 3340, 2014.
[31] Andrea Vedaldi and Brian Fulkerson. Vlfeat: An open and portable
library of computer vision algorithms. In Proceedings of the 18th ACM
international conference on Multimedia, pages 1469 -- 1472. ACM,
2010.
[32] Tam´as Vicsek, Andr´as Czir´ok, Eshel Ben-Jacob, Inon Cohen, and Ofer
Shochet. Novel type of phase transition in a system of self-driven
particles. Physical review letters, 75(6):1226, 1995.
[33] B-N Vo, Sumeetpal Singh, and Arnaud Doucet. Sequential monte
carlo methods for multitarget filtering with random finite sets. IEEE
Transactions on Aerospace and electronic systems, 41(4):1224 -- 1245,
2005.
[34] Stefan B Williams, Gamini Dissanayake, and Hugh Durrant-Whyte.
Towards multi-vehicle simultaneous localisation and mapping.
In
Robotics and Automation, 2002. Proceedings. ICRA'02. IEEE Inter-
national Conference on, volume 3, pages 2743 -- 2748. IEEE, 2002.
[35] Junchi Yan, Minsu Cho, Hongyuan Zha, Xiaokang Yang, and
Stephen M Chu. Multi-graph matching via affinity optimization with
IEEE transactions on pattern
graduated consistency regularization.
analysis and machine intelligence, 38(6):1228 -- 1242, 2016.
[36] Junchi Yan, Jun Wang, Hongyuan Zha, Xiaokang Yang, and Stephen
Chu.
Consistency-driven alternating optimization for multigraph
matching: A unified approach. IEEE Transactions on Image Process-
ing, 24(3):994 -- 1009, 2015.
[37] Christopher Zach, Manfred Klopschitz, and Manfred Pollefeys. Dis-
ambiguating visual relations using loop constraints. In IEEE Confer-
ence on Computer Vision and Pattern Recognition, 2010.
[38] Tinghui Zhou, Yong Jae Lee, X Yu Stella, and Alexei A Efros.
Flowweb: Joint image set alignment by weaving consistent, pixel-wise
correspondences. In IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 2015.
[39] Xiaowei Zhou, Menglong Zhu, and Kostas Daniilidis. Multi-image
In Proceedings of the
matching via fast alternating minimization.
IEEE International Conference on Computer Vision, pages 4032 -- 4040,
2015.
[40] Xun S Zhou and Stergios I Roumeliotis. Multi-robot slam with
unknown initial correspondence: The robot rendezvous case. In 2006
IEEE/RSJ International Conference on Intelligent Robots and Systems,
pages 1785 -- 1792. IEEE, 2006.
|
1909.13204 | 1 | 1909 | 2019-09-29T05:41:50 | Clustering Strategies of Cooperative Adaptive Cruise Control: Impacts on Human-driven Vehicles | [
"cs.MA"
] | As a promising application of connected and automated vehicles (CAVs), Cooperative Adaptive Cruise Control (CACC) is expected to be deployed on the public road in the near term. Thus far the majority of the CACC studies have been focusing on the overall network performance with limited insight on the potential impact of CAVs on human-driven vehicles (HVs). This paper aims to quantify the influence of CAVs on HVs by studying the high-resolution vehicle trajectory data that is obtained from microscopic simulation. Two clustering strategies for CACC are implemented: an ad hoc coordination one and a local coordination one. Results show that the local coordination outperforms the ad hoc coordination across all tested market penetration rates (MPRs) in terms of network throughput and productivity. The greatest performance difference between the two strategies is observed at 30% and 40% MPR for throughput and productivity, respectively. However, the distributions of the hard braking observations (as a potential safety impact) for HVs change significantly under local coordination strategy. Regardless of the clustering strategy, CAVs increase the average lane change frequency for HVs. 30% MPR is the break-even point for local coordination, after which the average lane change frequency decreases from the peak 5.42 to 5.38. Such inverse relationship to MPR is not found in the ah hoc case and the average lane change frequency reaches the highest 5.48 at 40% MPR. | cs.MA | cs |
Clustering Strategies of Cooperative Adaptive
Cruise Control: Impacts on Human-driven Vehicles
Zijia Zhong
Mark Nejad
Earl E. Lee II
Department of Civil and Environmental Engineering
University of Delaware
Newark, DE, USA
Email: {zzhong, elee, nejad}@udel.edu
Joyoung Lee
John A. Reif, Jr. Department of Civil and
Environmental Engineering
New Jersey Institute of Technology
Newark, NJ, USA
Email: [email protected]
Abstract -- As a promising application of connected and au-
tomated vehicles (CAVs), Cooperative Adaptive Cruise Control
(CACC) is expected to be deployed on the public road in the
near term. Thus far the majority of the CACC studies have
been focusing on the overall network performance with limited
insight on the potential impact of CAVs on human-driven vehicles
(HVs). This paper aims to quantify the influence of CAVs on HVs
by studying the high-resolution vehicle trajectory data that is
obtained from microscopic simulation. Two clustering strategies
for CACC are implemented: an ad hoc coordination one and a
local coordination one. Results show that the local coordination
outperforms the ad hoc coordination across all tested market
penetration rates (MPRs) in terms of network throughput and
productivity. The greatest performance difference between the
two strategies is observed at 30% and 40% MPR for throughput
and productivity, respectively. However, the distributions of the
hard braking observations (as a potential safety impact) for HVs
change significantly under local coordination strategy. Regardless
of the clustering strategy, CAVs increase the average lane change
frequency for HVs. 30% MPR is the break-even point for local
coordination, after which the average lane change frequency
decreases from the peak 5.42 to 5.38. Such inverse relationship
to MPR is not found in the ah hoc case and the average lane
change frequency reaches the highest 5.48 at 40% MPR.
Index Terms -- CAV Clustering, Cooperative Adaptive Cruise
Control, Vehicle Trajectory Analysis, Mixed Traffic Condition,
Traffic Flow Characteristics
I. INTRODUCTION
Cooperative adaptive cruise control
(CACC) enables
closely-coupled vehicular platoons by the extra layers of
communication and automation. Being one of the most-studied
application of CAVs, CACC is expected to drastically increase
mobility, decrease emission, while providing a safer and more
convenient way for occupants. The CACC evaluation thus far
has been focusing on the benefits that CAV could potentially
bring to our transportation network. The potential impact on
non-equipped vehicles (i.e. Human-driven vehicles(HVs)) has
been overlooked.
The motivation of the study is twofold. First, studying
CACC under mixed traffic conditions in anticipation of its
near-term deployment has gained an increasing amount of
attention. It is the consensus that CACC can improve the
performance of our transportation system. However, the pos-
sible change of the flow characteristic of HVs has often
been overlooked. CACC may alter the behavior of the HVs
especially when active clustering strategy is employed. The
expected impact on HVs from CACC could include: i) the ad-
ditional weaving during CACC clustering, ii) the lane change
by HVs induced by CACC clustering, iii) increase collision
risk, and iv) lane blockage by CACC platoon. Furthermore, the
majority of CACC studies focused on the longitudinal move-
ment and put less emphasis on the lateral movement of CAVs
which is vital in cooperative driving. The lateral control when
it comes to local coordination of forming CACC platoons has
to be taken into account, as the development of CACC vehicle
progresses. By implementing local coordination, the CACC
platoon ratio and consequently the network performance can
be increased.
In this paper, the comparison among ad hoc coordination
and local coordination for CACC are made according to
network-wide performance measures. The traffic flow charac-
teristic of HVs was investigated with the emphasis on the in-
teraction between HVs and CAVs based on the high-resolution
vehicle trajectory data extracted from microscopic simulation.
The proposed methodology developed is also suitable for
extracting the driving behavioral data for field deployment
and modeling heterogeneous traffic flow that is consisted of
HVs and CAVs. Three scenarios are evaluated: i) base without
CAVs, ii) CAVs with ad hoc coordination, and iii) CAVs with
local coordination algorithm (e.g., active platoon formation).
The organization of the remainder of the paper is as follows.
Relevant research regarding research of CACC in mixed traffic
will be reviewed in Section II, followed by the microscopic
simulation framework in Section III. The simulation results
are presented and discussed in Section IV. Lastly, findings
and recommendations are discussed in Section V.
II. CACC RESEARCH IN MIXED TRAFFIC
A. Impact of CACC
CACC can positively increase the traffic performance with
sufficient presence in the traffic flow. The reduced time
headway and following distance have been recognized as the
primary benefits of CACC. Arnaout and Arnaout [1] proposed
the Flexible Agent-based Simulator of Traffic framework for
evaluating CACC with ad hoc coordination. Moderate, satu-
rated, and over-saturated demand scenarios were tested on a
hypothetical four-lane high highway under various of CACC
market penetration rates (MPRs). The advantage of CACC
showed when MPR was above 40% and the network was able
to serve 9,400 out of 10,000 vehicles per hour (vph). Lee et
al. [2] evaluated the potential benefits for both mobility and
safety under a wide range of traffic scenarios. It was found
that the mobility benefits of CACC were shown at as less as
30% MPR.
Songchitruksa et al. [3] evaluated the improvement of
CACC on a 26-mile segment of the Dallas I-30 freeway. For
simplification, zero demand from the on-ramps was assumed.
The highest throughput was observed as 4,400 vph with local
coordination, where only rear-join to a platoon was allowed.
Van Arem et al. [4] assessed the impact of CACC on freeway
traffic flow on a 6-km, one-lane freeway with ramps distributed
with 1.6-km interval. They found the capacity reached 4,250
vph per lane (vphpl) with full CACC penetration. With the
same CACC model, Shladover et al. [5] studied the impact
of CACC on freeway traffic flow on a one-lane freeway with
no demand from ramps. The lane capacity was found to reach
3,600 vphpl at 90% MPR of CACC.
All of the above studies assessed the potential benefits
brought by CACC on the overall network. However, none
of them has investigated the impact that CACC could bring
to HVs, especially under local coordination in which a free-
agent CAV is actively seeking and subsequently performing
lane change to join a platoon.
Nowakowski et al. [6] studied the acceptance of the short
following distance (ranging from 0.6 s to 1.1 s) enabled by
CACC. As discovered, while all the drivers showed the will-
ingness to accept the shorter following gaps, male participants
were more likely to choose a shorter following distance. The
carry-over effect of the short headway in manual driving was
exhibited even after the disengagement from the platoons in
the KONVOI project [7]. Gouy et al. [8] investigated the
behavioral adaptation effect that potentially caused by the
short headway of CACC platoon using driving simulator.
Participants were instructed to driver alongside with two
CACC platoon configurations: i) 10-truck platoon with 0.3
s intra-platoon headway, and ii) 3-truck platoon with 1.4 s
intra-platoon headway. They found that smaller average time
headway was observed when in the short headway scenario. In
the first platoon scenarios, participants spent more time under
a 1-second headway, which is deemed unsafe [9].
Lin et al. [10] studied the time-gap of bus ACC system
in three aspects: i) preferred time-gaps for expressway driv-
ing, ii) time-gaps that maximized safety, and iii) the influence
of the secondary tasks to time-gaps. Calvert and Lint [11]
studied the negative effect of ACC on the system capacity
using the propose ACC control. The simulation scenarios
contained various traffic flow conditions in terms of demand
and flow composition (e.g., ACC vehicles, trucks, etc.). They
concluded that the small negative effect on road capacity did
exist and it was caused by the higher gap times of ACC.
B. CACC Coordination
There are three types of clustering strategies: ad hoc coordi-
nation, local coordination, and global coordination [5]. In this
study, only the former two clustering strategies are evaluated
due to the scope. Since global coordination requires advance
planning for the travel demand at an origin-designation level.
CAVs are coordinated to enter the highway in platoons. Global
coordination is likely to be sub-optimal without a robust
planning system to couple with the logistical challenge under
dynamic traffic conditions.
1) Ad hoc coordination: Ad hoc coordination assumes
random arrival of CAVs and no coordination among them.
Therefore, the probability of driving behind another CAV
is highly correlated to the market penetration rate (MPR).
Ad hoc coordination has been observed in the majority of
the research due to its simplicity in implementation and the
lack of the conceptual framework for CAV platoon formation.
However, the ad hoc coordination is not likely to harness
the full potential of CACC, as it does not fully utilize the
short intra-platoon headway enabled by the CAV technology.
Exclusive lane, or other forms of managed lane, has been
employed to aid the ad hoc clustering [12]. In a sense, a CAV
lane is a special form of coordination where the target lane is
the CAV lane.
2) Local coordination: Local coordination facilities the
platoon formation, where free-agent CAVs are actively seeking
clustering opportunity in their surroundings. The subject CAV,
as well as the surrounding CAVs can be coordinated to change
trajectory to facilitate clustering. There are four basic types of
lane change: i) free-agent-to-free-agent lane change, ii) free-a-
gent-to-platoon lane change,
iii) platoon-to-free-agent lane
change, and iv) platoon-to-platoon lane change [13].
Lee et al. [2] developed a local coordination scheme which
allows three ways to form a platoon: i) front-join, ii) mid-join,
and iii) rear-join. The longitudinal control was a rule-based
acceleration selection. Developed from Lee et al.'s coordi-
nation model, Zhong et al. [14] implemented the MIXIC
[4] as car following model to study the CAV benefits for
arterials. Lee et al.'s [2] CACC algorithm was updated to using
the EIDM for simulating CACC string behavior. The string
formation and dispersion mechanism were enhanced, including
preferential lane logic, platoon size restriction [15]. Zhong
[16] developed a CACC control model by combining E-IDM
and the MOBIL model [17]. The MOBIL model is adopted
as the mechanism to prevent lane changing of a free-agent
CAV that may be potentially disruptive to the surrounding
traffic. When a potential platooning opportunity is identified
via V2V communication, the CACC system estimates the
impacts on the immediate vehicles based on MOBIL should
the lane change be initiated. The Lane-change Model with
Relaxation and Synchronization (LMRS) model, proposed in
[18], gives a normalized strategic lane-change score by taking
into consideration of route, gain speed, and lane preference.
For a higher desire score, the driver is willing to accept smaller
headway and to decelerate more in LMRS. Calvert and Lint
[11] adopted the LMRS model in conjunction with the IDM+
[19] for evaluating the ACC.
III. EVALUATION FRAMEWORK
A. Simulation of Human Driving Behavior
Human drivers can take into account more input stimulus
(e.g., brake lights, next-nearest neighbors, etc.) with anticipa-
tion of the situation for the next few seconds [20]. All of these
aspects can be formulated in terms of psycho-physiological
models, such as Gipps' model [21] and Weidemann model
[22]. Weidemann model was re-calibrated in [23] using an
instrumented vehicle to measure the thresholds among differ-
ence driving states. The Weidemann model is used by Vissim
as the as the default car-following model.
The Vissim car-following model also includes the tactical
driving behavior, which carries certain planning in advance
with a temporal horizon (multiple time steps) or spatial
horizon that is beyond neighboring vehicles [24]. There are
four different driving states in the Widemann model. i) free
driving, ii) approaching, iii) following,and iv) braking. The
acceleration is primarily determined by the current speed,
speed difference, and gap to the preceding vehicle for each of
the four driving state. The Wiedemann-99 model, suitable for
freeway application, has ten calibratable parameters to repre-
sent a wide range of driver behavior. Therefore, the Wiedmann
model has to be calibrated to specific traffic stream data [25],
[26], as it was initially developed on limited available data.
The objective of the calibration process is to minimize the
difference between the measured driving behavior in the field
and the driving behavior simulated.
There are two types of lane change in the Wiedemann
model: necessary lane change and free lane change. The
former focus on the hard constraint of the lane change, such
as lane drop. The latter type is the focus of this study. Such
lane change is performed when more space and higher speed
is desired for a vehicle. As such, the safety distance plays
an important role in the lane change behavior. A suitable gap
is found based on i) the speed of the vehicle changing lane,
and ii) the approaching speed of the vehicle from behind on
the lane [27]
B. Quantify Impact to Human Driver
The trajectory data for HVs are also collected. A before-
and-after study is the most straightforward way to assess the
changes that are brought by CAVs. The ultimate goal is to
determine whether the changes are of statistical significance.
Figure 1 illustrates the study methodology. First, the human
driving behavior is calibrated by multiple sources of data that
were collected from the roadway segment of interest. The
calibration effort was conducted in [28]. With a calibrated
behavior model, we then treat the car-following model as a
black box. The input and the output to the human behavior
model is stimuli (local
traffic condition) and reactions of
HVs, respectively. Then on a collective level, the traffic flow
characteristics and vehicle trajectories are analyzed.
C. Simulation of CAV Driving Behavior
Longitudinal control (or car-following) and lateral control
are the two main components for simulating the behavior of
CACC. Numerous car-following models have been proposed.
Among them, the Gipps [21], Wiedemann [22], the IDM [29]
and its variants have been widely used.
1) Longitudinal Control: The Enhanced Intelligent Driver
Model (E-IDM) [30] is adapt and the longitudinal control
model, which is expressed in (1).
x
xdes
)δ − ( s∗( x, xlead)
a[1 − (
if x = xIDM ≥ xCAH
(1 − c)xIDM + c[xCAH + b · tanh( xIDM −xCAH
)]
s0
b
)]
otherwise
x =
s∗( x, xlead) = s0 + xT +
x( x − xlead)
2√ab
xCAH =
x2
·min(xlead,x)
−2x·min(xlead,x)
x2
lead
if xlead( x − xlead) ≤ −2x min(xlead, x)
min(xlead, x) − ( x− xlead)2Θ( x− xlead)
2x
otherwise
(1)
where a is the maximum acceleration; b is the desired
deceleration; c is the coolness factor; δ is the free acceleration
exponent; x is the current speed of the subject vehicle; xdes
is the desired speed, xlead is the speed of the lead vehicle; s0
is the minimal distance; x is the acceleration of the subject
vehicle; xlead is the acceleration of the lead vehicle; xIDM is
the acceleration calculated by the original IDM model [29];
T is the desired time gap; and xCAH is the acceleration
calculated by the CAH component as shown in Eq.1 where
Θ is the Heaviside step function.
2) Lateral Control and Coordination: We assume all the
CAVs are equipped with automated longitudinal control. Each
CACC vehicle is able to detect the surrounding traffic and
discern CAVs from HVs. Three cases are tested:
i) Base case: case without CAV traffic. This is the baseline
for the case study network. The I-66 network has been
calibrated.
ii) Ad hoc coordination: the CACC system controls only
the longitudinal control based on the Enhanced-IDM
model. Lateral movement was controlled by the calibrated
Wiedenmann model.
iii) Local coordination: the CACC control cover both logi-
tudinal and lateral aspects. It was developed in [2]. The
longitudinal control of the is replaced with E-IDM and
the lateral control is assumed to be done by the automated
driving system as well.
A break-down of the cases is shown in Table I. Since the focus
of the paper is the near-term deployment of CACC in mixed
traffic condition, the MPR is set at a medium-to-low range
from 0% to 40%.
Fig. 1: Potential Impact to Non-equipped Vehicles
TABLE I: Simulation Case
Cases
Base
Ad hoc coordination
Local coordination
Longitudinal Control
calibrated Weidemann
E-IDM
E-IDM
Lateral Control
calibrated Weidemann
calibrated Weidemann
gap acceptance-based [2]
• a calibrated driving behavior model in Vissim with real-
world data constitutes a good representation of the human
driver.
• vehicle-to-vehicle (V2V) communication is perfect (no
interference or packet loss).
• human drivers do not differentiate CAVs and other HVs
D. Case Study Network
as followers
In this study, an 8-km (5-mile) segment (Fig. 2) of Interstate
Highway I-66 outside of the beltway (I-495) of Washington
D.C. is used. This freeway segment has recurring congestion
during weekdays, specifically in the eastbound direction in the
morning and westbound direction in the afternoon. The road-
way is with four lanes in each direction. The leftmost lane is an
HOV 2+ lane with 1500 vphpl peak volume [31]. Currently, no
physical barrier is between the HOV lane and its adjacent GP
lane. The Vissim simulation network is available via the U.S.
DOT Open Source Application Development Portal [32]. The
calibration was conducted with two independent data sources
(i.e., INRIX TMC travel time and RTMS flow data).
Fig. 2: I-66 Simulation Testbed
In anticipation of the increase of the traffic demand over
time, we assume a 30% growth in traffic demand from the
baseline of the calibrated network. Each deployment scenarios
was run five times with different random seeds to factor in
the variability of the traffic condition. The duration of the
simulation is 3900 s, the first 300 s of which is used to
load the network with traffic. No data is collected during this
period. The evaluation of the impact is based on the following
assumptions:
• a low-level vehicle controller for longitudinal (e.g., throt-
tle, brake) and lateral (e.g., steering) control is available.
This study only focuses on the tactical driving aspect of
vehicle operation.
IV. RESULTS & DISCUSSION
The vehicle trajectory data are collected every 0.5 s, which
offers a snapshot of the prevailing traffic condition. Not only
vehicle dynamic data, but also the interaction state including
driving state, interacting vehicle, etc. is available.
A. Network Performance
Fig. 3 shows the benefits that CACC bring to the trans-
portation network. Fig. 3(a) shows the ratio of vehicle miles
traveled (VMT) and vehicle-hours traveled (VHT). VMT is the
output of a transportation system, whereas VHT is considered
the input to a transportation system. The ratio of VMT and
VHT ratio is referred as Q [33], which represents the output
of a transportation system with the unit value of the input.
In short, the higher the value of the Q, the more productive
a transportation system is. Both of the ad hoc and local
coordination strategies exhibit an increasing trend as the MPR
increases. It is notable that the benefits gained by ad hoc
coordination show a diminishing increase after 30% MPR.
In comparison, local coordination displays a linear increasing
pattern for the performance gain.
When it comes to network throughput, Fig. 3(b) shows
that the ad hoc coordination does not increase the network
throughput at 10% MPR: the throughput remains as 9398
vph. After 10% MPR, the throughput for ad hoc coordination
increases with a liner pattern, which matches the underlying
operational implication of ad hoc coordination. The throughput
reaches the highest 10167 vph at 40%. With local coordination,
additional throughput is observed even at 10% MPR. Then
the slope of the throughput curve is greater at the MPR range
between 10% and 30% than other tested MPR values. It also
shows that the rate of increase of the throughput decrease
after 30% MPR. Moreover, the vertical distance indicates the
magnitude that the local coordination outperforms the ad hoc
coordination at each level of MPR. The greatest difference is
observed at 30% MPR.
(a) productivity
(a) hard braking observation CDF
(b) throughput
Fig. 3: Network Performance
B. Hard Braking Observations
Hard braking observation has been used as a safety measure.
Abrupt braking is likely an indication of a hazardous traffic sit-
uation that drivers respond to [34]. Hard braking observations
are recorded when the acceleration of a vehicle is less than
-3 m/s2. Recall our primary focus is the HVs. There are two
types of hard braking: the first one occurs when an HV inter-
acts with another HV; whereas the second type occurs when
an HV interacts with a CAV. The hard braking observation
for HVs when they interact with other HVs is shown in Fig.
4(a). Similar patterns of the cumulative distribution function
(CDF) curves for hard braking are observed across the testing
scenarios. Fig. 4(b) shows the sample size for each scenario.
The primary factor for the decreasing trend is the reduction
of HVs the traffic stream. The linear trend also infers that the
likelihood of hard braking remains at the same level.
Fig. 5(a) shows the CDF curves of the hard breaking
observations recorded for HVs when they interact with CAVs
under each scenario. The CDFs show two distinctive patterns
between two coordination strategies. In the ad hoc coordina-
tion cases, the CDFs are with similar distributions. On the
other hand, the CDF curves of the local coordination are
more sensitive to MPR. The probability of hard braking in
(b) hard braking observation
Fig. 4: Hard braking observation when interacting with HVs
the range from -6.5 to -3.5 m/s2 drastically increases even
at 10% MPR. The occurrence of hard braking event keeps at
the same level in ad hoc coordination; whereas the occurrence
of coordination strategy shows an increasing trend until 30%
MPR where the value peaks. The sample size is shown in Fig.
5(b). Both strategies exhibit an increasing trend until 30 %
MPR, then a declining occurrence after 30% MPR. With the
same amount of CAVs, the hard breaking is more sensitive to
MPR in Local coordination then that in ad hoc strategy.
Two-sample Kolmogorov-Smirnov (K-S) test is adapted to
further analyze the CDF curves. The K-S test is a powerful tool
for testing the hypothesis that whether two random samples
have been drawn from the same population [35]. It is a non-
parametric test where no assumption is made regarding the
distribution of the variables [36]. The null hypothesis (H0) of
the K-S test is that the comparing two sample sets are from
the same continuous distribution. The the two-sample K-S test
is conducted for hard braking for each pair of the scenarios
at 5% significance level. The hypothesis tests show that any
pair of the scenarios rejects the null hypothesis and accept the
and 5.48 at 30% and 40%, respectively. 30% is the break-
even point when it comes to average lane change frequency.
(a) hard braking event CDF
Fig. 6: Lane Change Activity of HVs
In summary, the local coordination strategy outperforms
levels of MPR in
ah hoc coordination strategy across all
terms of network throughput and productivity. However, the
induced hard braking for HV should not be overlooked. The
distribution of hard braking for HVs changes substantially
when they interacting with CAVs under local coordination.
Compared to ad hoc coordination at
the
probability of the hard braking event in the range of -7.3 to -
6.5 m/s2 is higher. Local coordination causes a higher average
lane change frequency for HV at low MPRs (i.e., 10% and
20%). It starts to decrease after reaching 30% MPR, whereas
in the ad hoc strategy the average lane change frequency
maintains the increasing trend.
the same MPR,
V. CONCLUSION
In this paper, we investigate the two types of coordination
strategies for CAV platoon formation. Platoon clustering strat-
egy is a crucial aspect when it comes to deploying CAV in
mixed traffic condition in the near term. Agreeing with previ-
ous studies, CAV is able to bring benefits to the transportation
network even with ad hoc coordination. Adapting local coor-
dination can further increase the benefits. The impact on HVs
is quantified. The distribution of the hard braking observation
for HVs, when interacting with CAVs, change substantially
with local coordination strategy for platoon formation. In
comparison, the distributions for HVs when interacting with
other HV show no substantial changes. The average lane
change for HVs increases with the presence of CAVs until
30% MPR is reached.
Future research would be focused on the following areas.
First, the lateral control is an underexplored area compared to
longitudinal control of CAVs. Further investigation of platoon
formation in mixed traffic is desired. Currently, there are only
a few platoon coordination algorithms, most of which are rule-
based. Second, the aggressiveness of the lane change for CAVs
when forming a platoon is also an important aspect. As shown,
the characteristic of the HVs traffic could be altered. Some of
the changes could pose safety concerns for HVs. Third, the
(b) hard braking event sample size
Fig. 5: Hard Braking Event when Interacting with CAVs
alternative hypothesis that the two samples are not from the
same distribution. The H0 is not rejected only in the ad hoc
coordination case at 20% and 30% MPR.
C. Lane Change Activity
Fig. 6 shows the accumulative lane change instance recorded
at ever 0.5 s for all HVs. The lane change activity of CACC is
not recorded as the scope of the paper is confined to HVs. The
lane change activity decrease as the MPR of CACC increase
in either of the coordination strategies. Local coordination
is marginally higher than the ad hoc one when the MPR is
low. At 40% MPR, they reach the same level of lane change
activity. However, recall that the number of HV within the
network decreases as the MPR of CACC increase. As such,
the average lane change frequency for each HV is calculated
and plotted in Fig. 6 as well. The local coordination strategy
shows a higher average lane change frequency at 10% and 20%
MPR. The average lane change frequency peaks at 30%, then
reduced to 5.38 from 5.42. On the contrary, the increasing
trend for ah hoc clustering keeps increasing and reach 5.46
Board,
http://trrjournalonline.trb.org/doi/10.3141/1999-10
86 -- 94,
2007.
1999,
vol.
pp.
[Online]. Available:
[18] W. J. Schakel, V. L. Knoop, and B. van Arem, "Integrated Lane Change
Model with Relaxation and Synchronization," Transportation Research
Record, vol. 2316, no. 1, pp. 47 -- 57, jan 2012. [Online]. Available:
https://doi.org/10.3141/2316-06
[19] W. J. Schakel, B. van Arem, and B. D. Netten, "Effects of cooperative
adaptive cruise control on traffic flow stability," in 13th International
IEEE Conference on Intelligent Transportation Systems, Sep. 2010, pp.
759 -- 764.
[20] M. Treiber and A. Kesting, "Traffic flow dynamics," Traffic Flow
Dynamics: Data, Models and Simulation, Springer-Verlag Berlin Hei-
delberg, 2013.
"A behavioural
[21] P. Gipps,
simulation,"
vol. 15, no. 2, pp. 105 -- 111, 1981.
http://www.sciencedirect.com/science/article/pii/0191261581900370
computer
Transportation Research Part B: Methodological,
[Online]. Available:
car-following model
for
[22] R. Wiedemann, "Simulation des Straenverkehrsflusses," Ph.D. disserta-
tion, Karlsruhe, 1974.
[23] U. Reiter, "Empirical studies as basis for traffic flow models," in
PROCEEDINGS OF THE SECOND INTERNATIONAL SYMPOSIUM
ON HIGHWAY CAPACITY, VOLUME 2, 1994.
[24] J. Barcel´o et al., Fundamentals of traffic simulation.
Springer, 2010,
vol. 145.
[25] B. Higgs, M. M. Abbas, and A. Medina, "Analysis of the wiedemann
car following model over different speeds using naturalistic data," in
Procedia of RSS Conference, 2011, pp. 1 -- 22.
[26] U. Durrani, C. Lee, and H. Maoh, "Calibrating the wiedemanns vehicle-
following model using mixed vehicle-pair interactions," Transportation
research part C: emerging technologies, vol. 67, pp. 227 -- 242, 2016.
[27] PTV Group, "Ptv vissim 11 user manual," Germany: PTV GROUP,
2018.
[28] Leidos, "Simulation of evolutionary introduction of cooperative adaptive
cruise control equipped vehicles into traffic," Saxton Transportation
Operations Laboratory, Tech. Rep., 2016.
[29] M. Treiber, A. Hennecke, and D. Helbing, "Congested Traffic States in
Empirical Observations and Microscopic Simulations," pp. 1805 -- 1824,
2000.
[30] A. Kesting, M. Treiber, and D. Helbing, "Enhanced intelligent driver
model to access the impact of driving strategies on traffic capacity,"
Philosophical Transactions of the Royal Society A: Mathematical, Phys-
ical and Engineering Sciences, vol. 368, no. 1928, pp. 4585 -- 4605, 2010.
[31] X.-Y. Lu, J. Lee, D. Chen, J. Bared, D. Dailey, and S. E. Shladover,
"Freeway micro-simulation calibration: Case study using aimsun and
vissim with detailed field data," in 93rd Annual Meeting of the Trans-
portation Research Board, Washington, DC, 2014.
OSDAP,
for
[32] FHWA
add-on
now available
https://www.itsforge.net/index.php/vedm-cac-announced
"Vissim external
download,"
automated
connected
vehicle
2015.
for
driver model
(vedm)
applications
[Online]. Available:
(cav)
[33] Caltrans Division of Traffic Operations, "A guide to using the caltrans
performance measurement system (pems) in transportation concept
reports (tcr's)," Tech. Rep., 2013.
[34] O. Bagdadi and A. V´arhelyi, "Jerky drivingan indicator of accident
proneness?" Accident Analysis & Prevention, vol. 43, no. 4, pp. 1359 --
1363, 2011.
[35] L. A. Goodman, "Kolmogorov-smirnov tests for psychological research."
Psychological bulletin, vol. 51, no. 2, p. 160, 1954.
[36] I. T. Young, "Proof without prejudice: use of the kolmogorov-smirnov
test for the analysis of histograms from flow systems and other sources."
Journal of Histochemistry & Cytochemistry, vol. 25, no. 7, pp. 935 -- 941,
1977.
comparison between clustering strategies should be expanded
to include additional local coordination strategies.
REFERENCES
[1] G. M. Arnaout and J.-P. Arnaout, "Exploring the effects of cooperative
adaptive cruise control on highway traffic flow using microscopic traffic
simulation," Transportation Planning and Technology, vol. 37, no. 2, pp.
186 -- 199, 2014.
[2] J. Lee, J. Bared, and B. Park, "Mobility impacts of cooperative adaptive
cruise control (cacc) under mixed traffic conditions," in 93rd Annual
Meeting of the Transportation Research Board, Washington, DC, 2014.
[3] P. Songchitruksa, A. Bibeka, L. I. Lin, Y. Zhang et al., "Incorporating
driver behaviors into connected and automated vehicle simulation,"
Center for Advancing Transportation Leadership and Safety (ATLAS
Center), Tech. Rep., 2016.
[4] B. Van Arem, C. J. Van Driel, and R. Visser, "The impact of cooperative
adaptive cruise control on traffic-flow characteristics," IEEE Transac-
tions on intelligent transportation systems, vol. 7, no. 4, pp. 429 -- 436,
2006.
[5] S. E. Shladover, D. Su, and X.-Y. Lu, "Impacts of cooperative adaptive
cruise control on freeway traffic flow," Transportation Research Record,
vol. 2324, no. 1, pp. 63 -- 70, 2012.
[6] C. Nowakowski, J. O'Connell, S. E. Shladover, and D. Cody, "Co-
operative adaptive cruise control: Driver acceptance of following gap
settings less than one second," in Proceedings of the Human Factors
and Ergonomics Society Annual Meeting, vol. 54, no. 24.
SAGE
Publications Sage CA: Los Angeles, CA, 2010, pp. 2033 -- 2037.
[7] S. M. Casey and A. K. Lund, "Changes in speed and speed adaptation
following increase in national maximum speed limit," Journal of safety
research, vol. 23, no. 3, pp. 135 -- 146, 1992.
[8] M. Gouy, K. Wiedemann, A.
"Driving
time
Stevens, G. Brunett,
and
platoons:
drivers
control?" Transportation Research Part F: Traffic
and Behaviour, vol. 27, pp. 264 -- 273, 2014,
[Online]. Available:
N. Reed,
How do
longitudinal
Psychology
vehicle Automation and Driver Behaviour.
http://www.sciencedirect.com/science/article/pii/S1369847814000345
automated
influence
next
to
headways
non-platoon
vehicle
short
[9] S. H. Fairclough, A. J. May, and C. Carter, "The effect of time
headway feedback on following behaviour," Accident Analysis and
Prevention, vol. 29, no. 3, pp. 387 -- 397, 1997. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0001457597000055
[10] T.-W. Lin, S.-L. Hwang, and P. A. Green, "Effects of
time-gap
(acc) on driving performance
settings of adaptive cruise control
and subjective
in a bus driving simulator," Safety
Science, vol. 47, no. 5, pp. 620 -- 625, 2009. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0925753508001422
acceptance
[11] W. J. S. S. C. Calvert and J. W. C. van Lint, "Will automated vehicles
negatively impact traffic flow," Journal of Advanced Transportation, vol.
2017, 2017.
[12] M. Segata, F. Dressler, R. Lo Cigno, and M. Gerla, "A simulation tool
for automated platooning in mixed highway scenarios," in Proceedings
of the 18th annual international conference on Mobile computing and
networking. ACM, 2012, pp. 389 -- 392.
[13] Z. Wang, G. Wu, P. Hao, K. Boriboonsomsin, and M. Barth, "Developing
a platoon-wide eco-cooperative adaptive cruise control (cacc) system,"
in 2017 IEEE Intelligent Vehicles Symposium (IV).
IEEE, 2017, pp.
1256 -- 1261.
[14] Z. Zhong, L. Joyoung, and L. Zhao, "Evaluations of Managed Lane
Strategies for Arterial Deployment of Cooperative Adaptive Cruise
Control," in 96th Transportation Research Board Annual Meeting,
Washington DC, USA, 2017.
[15] Transportation Research Board, Dedicating Lanes for Priority or
Exclusive Use by Connected and Automated Vehicles, B. A.
I. of Technology, Eds. Washington,
Hamilton, WSP, and N. J.
DC: The National Academies Press, 2018.
[Online]. Available:
https://www.nap.edu/catalog/25366
[16] Z. Zhong, "Assessing the effectiveness of managed lane strategies for
the rapid deployment of cooperative adaptive cruise control technology,"
Ph.D. dissertation, New Jersey Institute of Technology, 2018.
[17] A. Kesting, M. Treiber,
"General Lane-
Changing Model MOBIL for Car-Following Models," Transportation
Research
Research
and D. Helbing,
Transportation
Record:
Journal
the
of
|
1801.07583 | 1 | 1801 | 2018-01-20T15:06:03 | An Agent-Based Simulation Model for Optimization of the Signalized Intersection Connected to Freeway On-Ramp | [
"cs.MA"
] | Unlike most existing studies on off-ramp traffic signal control, this paper focuses on the optimization problem of the signalized intersection connected to freeway on-ramps. Conflicts are often observed between the traffic heading to an on-ramp and the traffic continuing straight which leads to issues such as intersection overflow, increased delay, and concerns about safety. For studying this problem, a real-world signalized intersection in Buffalo, New York was chosen, which has two through lanes and one short shared (through and right-turn) lane. At the downstream of the intersection are two following on-ramps, one to the highway I-290 West and the other to I-290 East. During peak hours, the shared lane often observes a long queue, which furthermore blocks the through traffic on the parallel lane. To solve this problem, a VISSIM agent-based simulation model was built and calibrated based on field observations. Three potential optimization solutions were proposed and tested with the help of VISSIM: (1) increasing the length of the short shared through and right-turn lane; (2) making the short shared through and right-turn lane right-turn only, and (3) adding a new diverge lane for the right-turn vehicles. According to the simulation results, solution (3) performs the best, resulting in the least vehicle delay time. | cs.MA | cs | An Agent-Based Simulation Model for Optimization of the Signalized
Intersection Connected to Freeway On-Ramp
Xuejin Wen1
1 Ohio State University; E-Mail: [email protected]
ABSTRACT
Unlike most existing studies on off-ramp traffic signal control, this paper
focuses on the optimization problem of the signalized intersection connected to
freeway on-ramps. Conflicts are often observed between the traffic heading to an on-
ramp and the traffic continuing straight which leads to issues such as intersection
overflow, increased delay, and concerns about safety. For studying this problem, a real-
world signalized intersection in Buffalo, New York was chosen, which has two through
lanes and one short shared (through and right-turn) lane. At the downstream of the
intersection are two following on-ramps, one to the highway I-290 West and the other
to I-290 East. During peak hours, the shared lane often observes a long queue, which
furthermore blocks the through traffic on the parallel lane. To solve this problem, a
VISSIM agent-based simulation model was built and calibrated based on field
observations. Three potential optimization solutions were proposed and tested with the
help of VISSIM: (1) increasing the length of the short shared through and right-turn
lane; (2) making the short shared through and right-turn lane right-turn only, and (3)
adding a new diverge lane for the right-turn vehicles. According to the simulation
results, solution (3) performs the best, resulting in the least vehicle delay time.
1
INTRODUCTION
Intersections controlled by traffic signal lights play a major role in road
transportation system. Since the first manually operated traffic signal in the 19th
century (Mueller, 1970), the topic has been extensively studied. One important
principle of signal light optimization is to maximize the throughput and decrease the
traffic accident risks simultaneously.
Intersections are one of the core elements of a transportation network. How to
optimize the signal timing of one isolated intersection has been extensively studied.
For example, Trabia et al. (1999) presented a two-stage fuzzy logic controller for an
isolated intersection. In the first stage, it estimates the relative traffic intensities and the
second stage determines the extension or termination of the current signal phase (Trabia
et al., 1999). Niu et al. (2009) also applied fuzzy control model for the traffic lights at
a single intersection.
However, only optimizing one isolated intersection may not solve the
optimization problem of the whole transport network. Therefore, the coordinated
control system with relevant intersections has attracted more public attention. Split,
cycle, and offset optimization technique (SCOOT) is a coordination system in England
which can analyze the data from vehicle detectors, predict, and adjust the signal timings
to minimize congestion (Hunt et al., 1981). The integrated control of highway off-ramp
and its downstream intersection is also necessary to mitigate the congestion. For
example, Günther et al. (2012) proposed a methodology which can determine the flow
of competing vehicles in the freeway to be detoured to underutilized local roads to
improve the systems' capacity. This is extremely useful for the traffic management
during the inclement weather (Lin et al., 2015) and traffic accidents (Lin et al., 2016).
Li et al. (2016) built mathematical models to study the effect of left-turn vehicles at a
signalized intersection with the presence of pedestrian crossings. Accordingly, the
macroscopic parameters of traffic flow, such as average flow rate, speed, and density
can be determined.
Besides the signal control optimization problem, the specific intersection
optimization is also vital to increase the capacity. Although some reference handbooks
like the America Highway Capacity Manual (HCM) have provided default capacity
calculation methods for intersections, researchers have shown that the realistic cases
are different. Tian and Wu (2006) built a probabilistic model for a signalized
intersection with a short right-turn lane and validated the model by traffic simulation
because HCM would overestimate the capacity and thus underestimate delay. Wu
(2007) built a general capacity model for intersections with short shared lanes based
on database generated from simulation package VISSIM. Ring and Sadek (2012)
studied signalized intersections with auxiliary through lanes and predicted the lane
utilization based on the data collected in Buffalo, New York.
2
As can be seen, traffic simulation is one of the important ways in intersection
studies, through which the alternative traffic signal plans and intersection optimizations
can be tested. VISSIM is one popular microscopic simulation tool which has been
applied for intersection simulation (Ring and Sadek, 2012; Wu, 2007) and border
crossing studies (Lin et al., 2013a; Lin et al., 2013b; Lin et al., 2014a; Lin et al., 2014b)
and so on. It can customize intersections according to the user's needs by using a link-
connector-based network procedure. It also has a strong ability to simulate the traffic
signal controllers (Ring and Sadek 2012; Wu, 2007).
This paper will study an isolated intersection with a short shared through and
right-turn lane. More interesting is that the intersection locates at the upstream of two
consecutive on-ramps, one to the interstate highway I-290 West and the other to the I-
290 East. During peak hours, a long queue observed at the short shared lane blocks the
through traffic. After the field data collection, a VISSIM model is built and calibrated.
Then a few optimization solutions are tested.
The rest arrangement of the paper is as following. The data collection part will
introduce the data collection and give a preliminary analysis. The next model
development section talks about how to build the intersection model and sets the traffic
signal controller in VISSIM. In the results section, different optimization solutions are
tested, and the results are shown. The conclusion and future study are discussed in the
last section.
DATA COLLECTION
As shown in Figure 1, the intersection of Twin Cities Memorial Highway and
W Colvin Blvd locates at the upstream of two on-ramps to I-290 West and I-290 East.
There are two through lanes and one short shared through and right-turn lane from the
northwest to the southeast direction (referred to hereafter as the left lane, the middle
lane, and the short lane separately). The left lane and middle lane connects different
streets downstream. One is to Eggert Road, and the other is to the Colvin Road. This
short lane is where the long queue forms and spills back to block the through traffic.
3
Figure 1. The intersection in Google map.
Before building the simulation model, some field observations must be
conducted. Firstly, this intersection takes actuated signal control and the minimum
green time, maximum green time, yellow time and so on have to be obtained. Secondly,
the traffic flow rate for each direction and each lane during peak hour should also be
collected. After field recording from 08:16 AM to 09:15 AM of three days 10/16/2013-
10/18/2013, Table 1 and Table 2 show the results.
According to the filed observations in Table 1, the traffic signal at this
intersection consists of five phases. For Phase 1 the left-turn and through traffic enter
the intersection in the same phase from northeast entrance. Phase 2 is similar that the
left-turn and through traffic are from southwest entrance. For southeast entrance, the
left-turn and through traffic are controlled by two separate signal phases Phase 3 and
Phase 4. Finally, the through traffic from northwest entrance is guided by Phase 5. No
left-turn traffic is allowed from the northwest entrance.
Table 1. Traffic Signal Parameters (unit: second)
Phase Index
1
2
3
4
5
Phase sequence
Minimum green
Maximum green
Yellow time
Red clearance
5
10
3
1
10
15
3
1
20
30
3
1
20
30
3
1
5
10
3
1
4
Table 2 shows the lane distribution of traffic flow rate for northwest entrance.
As can be seen, more than half of the total traffic flow was going through the short lane.
Because of the limitation of experiment equipment, here we only recorded the traffic
flow rates for the northwest entrance. For the other entrances, different flow rates will
be set as the vehicle inputs of the simulation model, through which it will also be
interesting to test the optimization solutions for various scenarios. The details are
introduced in next section.
Table 2. Traffic Flow Rate for Each Lane at Northwest Entrance during Peak
Hours 08:16 AM -09:15 AM (unit: pcu/h)
Lanes
Left Lane
Mid Lane
Short Lane
Total
Northwest Entrance
401
343
836
1,580
MODEL DEVELOPMENT
Signal Parameter Setting
With the help of Google Map, the layout of this intersection was built in
VISSIM as Figure 2.
Twin City Highway
West Colvin Boulevard
Figure 2. The intersection in VISSIM.
For the signal setting, VISSIM takes the Ring Barrier Controller. This controller
can simulate fully actuated signal control by defining different signal groups or phases.
5
Through the field observations, we already know there are five signal phases, in the
simulation model they are shown as the red bars in Figure 2. The green rectangles are
the detectors for the corresponding signal groups.
The signal parameters including the maximum/minimum green time, yellow
time, red clearance time are set according to Table 1. Besides that, the "vehicle
extension" in Ring Barrier Controller assigns 2 seconds for West Colvin Boulevard
direction and 3 seconds for Twin City Highway direction. The "vehicle extension"
extends the green interval time based on the detector status once the phase is green. All
five signal groups are flagged with "minimum recall" which means they will receive
an automatic vehicle call when they are not green (Mcelroy 2008). At last, the blue bars
are stop signs which are used to simulate the right-turn behavior during the red lights.
Vehicle Input Setting
For the northwest entrance, the vehicle inputs are the same as the field
observations shown in Table 2. Based on the real observations, the right-turn proportion
for the vehicles at the short lane is assumed as 10%. For each of the other three
entrances, although accurate observations are missing, three traffic flow rates are
assumed: low, medium, and high. Meanwhile, the left turn, through, and right turn
traffic probabilities are assumed to be equal. The inputs for the simulation model are
shown in Table 3.
Table 3. Assumed Traffic Flow Rates for the Other Three Entrances (unit: pcu/h)
Traffic flow rates
High
Medium
low
Southeast Northeast Southwest
1,500
1,000
500
1,000
600
200
1,000
600
200
There are 27 possible situations based on the combination of the entrance vehicle
inputs. For each situation, the models with different optimizing settings are simulated
three times of 3,600 seconds with various random seeds. The results section will show
the performances of the optimizing settings.
RESULTS
In this paper, three possible optimization solutions are proposed as following:
(1) increasing the length of the short lane; (2) making the short shared through and
right-turn lane right-turn only; (3) adding a new diverging lane for the right-turn
vehicles. For the 27 possible situations, to make it simple, a three-digit number is
applied to represent each situation; the three digits represent the traffic flow rates of
southeast, northeast, and southwest entrances in order. Each digit has three values,
("1"-high, "2"-"medium", and "3"-"low"). So for example, "112" means the traffic
6
flow rates of southeast and northeast entrances are high, and the flow rate of the
southwest entrance is medium. The performances of the three solutions were discussed
in details below.
(1) Increasing the length of the short lane
The current length of the short shared lane is 212.24 feet, which can
accommodate nine vehicles. Now suppose the short shared lane increases into 313.71
feet, which accommodating 14 vehicles. The new design of the intersection is
demonstrated in Figure 3.
Point A
Figure 3. The intersection with increased length of the short lane.
It is reasonable to suppose that the vehicles in the left lane will not be affected.
For the vehicles in the middle lane and the short lane, the delay times were compared
with that of the original intersection. The delay time in VISSIM was calculated as
compared to the ideal travel time (free flow condition, no signal control), the mean
time delay calculated from all vehicles observed on a single or several link sections
(PTV 2011). In this study, the link section of delay time is from the vehicle input point
of that link (point A in Figure 3) to the traffic lights location (red bar 8 in Figure 3) at
its corresponding lane. The comparison result is shown in Figure 4.
7
Figure 4. Delay time comparison of middle and short lanes after increasing the
length of the short lane.
As can be seen, the delay time at middle lane decreases regardless of the 27
situations. However, for the short lane, it shows that the delay time will become longer
than the original intersection. The blocking of vehicles at the middle lane may lead to
the increased delay time at the short lane. To find out the reason, the maximum queue
length at the middle lane is also recorded through VISSIM, and the results are shown
as in Figure 5.
Figure 5. Maximum queue length changes of middle lane after increasing the
short lane length
Figure 5 shows that as the length of the short shared lane increases, most of the
8
time the maximum queue length is around 250 feet, which is longer than that of the
original intersection. However, in VISSIM model, the diverging point of the length
increased short lane is 313 feet, which is far away from the queue length detector. It
means that the queue at middle lane with a maximum length of around 250 feet will
not block the vehicles going to the short shared lane.
Another reason for the increased delay time at short lane is that when the length
of the short lane increases, more vehicles will wait in that lane. While in the original
intersection model, more vehicles are waiting for the start point of the delay time
detector, and in VISSIM this period is not counted as the delay time if the vehicles have
not passed the start point (point A in Figure 3). For testing this, the input of vehicles is
controlled through VISSIM COM interface (PTV 2011) to make sure the same number
of vehicles arrive at the intersection at the same time. Finally, the adjusted delay time
at the short lane is shown in Figure 6.
Figure 6. Delay time comparison of short lane after increasing the short lane
length with vehicles input control of VISSIM COM
(2) Making the short shared through and right-turn lane as the right-turn-
only lane
Previously it was assumed that 10% of the vehicles at the short lane would turn
to the right. Now this solution sets the short shared lane to a right-turn-only lane, so the
rest 90% of the vehicles at the short lane need to go through the middle lane. The second
column in Table 5 shows the new traffic flow rate for each lane in this scenario I (the
traffic flow rate at the middle lane will be the sum of the original traffic flow rate at the
middle lane and the transferred flow rate from the short lane). The delay times at the
short lane and the middle lane are compared with those of the original intersection in
Figure 7.
9
Table 5. The New Traffic Flow Rate for Each Lane at Northwest Entrance after
Making the Short lane right-turn only (unit: pcu/h)
Lanes
scenario I
scenario II
Left Lane
Mid Lane
Short Lane
Total
401
1095
84
1,580
744
752
84
1,580
Figure 7. Delay time comparison of middle and short lanes after making the short
lane right-turn only.
Using this solution, the delay time at the short right lane decreased a lot.
However, the delay time at the middle lane increases. It is reasonable because all the
vehicles planning to head to the highway are going through the middle lane, and the
original vehicles at the middle lane will not change to the left lane, the vehicles heading
to the highway are occupying the middle lane. Now assume an extreme scenario II that
all the original vehicles at the middle lane will go to the left lane (they may change
back to the middle lane after passing the intersection because of the different connected
streets downstream), and the vehicles going to the highway go through the middle lane.
The new traffic flow rates are shown in the third column in Table 4; the VISSIM model
was simulated again. The result is shown in Figure 8.
10
Figure 8. Delay time comparison of left, middle and short lanes after letting the
original vehicles at middle lane go to the left lane.
If the entire original through traffic changed to the left lane, the delay time at
the middle lane is still longer than the original intersection, and the vehicles in the left
lane will also experience a much higher delay. Therefore, this solution only improves
the quality of service for the right-turn traffic.
(3) Adding a new diverging lane for the right-turn vehicles
Solution (3) is to add a diverging link for the right turn vehicles. The new design
layout of the intersection in VISSIM is shown in Figure 9.
11
Point A
Diverging lane
Stop sign B
Figure 9. The new intersection with a diverging lane.
Then the delay time for the diverging lane (from the vehicle input point A to the
stop sign B), short lane, and middle lane are recorded. The comparison with the original
intersection is presented in Figure 10.
Figure 10. Delay time comparison of the diverging, short, and middle lanes after
adding a new diverging lane for the right-turn vehicles
The delay time at diverging lane compared with that at the short lane of the
original intersection is introduced. It is observed that this solution can prevent the right-
turn vehicles being blocked by the vehicles going to the highway. Instead, they can go
to West Colvin Boulevard much faster. At the same time, this can save some space at
12
the short lane, which also relieves the pressure to the middle lane for the further step.
Therefore, the delay time at the short lane and middle lane are both deceased.
With the help of VISSIM, three solutions improving the current congestion
status are observed. More accurate, increasing the length of the short shared through
and right-turn lane decreases the delay time at the middle lane, but it has no impacts
for the vehicles at the short shared lane. Also, turning the short shared lane as a right-
turn-only lane makes the delay time at the short lane shorter, but may cause the delay
time at middle and left lanes longer. The third one adding a new diverging lane for the
right-turn vehicles performs best which can be beneficial for the right-turn vehicles, as
well as the through traffic.
CONCLUSION AND FUTURE WORK
This paper proposed three optimal solutions for the long queue problem at one
specific intersection caused by the vehicles going to the I-290 highway through the
downstream on-ramps. Each solution is tested by building corresponding VISSIM
models and compared with the original intersection. A few observations include:
1. An intersection with actuated signal control VISSIM model can be
established easily. Besides, VISSIM also provides a few evaluation criteria, like the
queue length, delay time;
2. For assessment and comparison of the VISSIM intersection models, the
vehicle input should be controlled by COM interface to make sure the number of
vehicles arriving at the intersection is equal at the same time;
3. Adding a new diverging link for the right-turn only vehicles can effectively
improve the quality of service of this intersection. Meanwhile, the other two solutions,
increasing the length of the short shared lane and making the short shared lane right-
turn, can only reduce the delay time for the vehicles in one lane.
For the future study, the nearby intersections close to this one will be included
in the VISSIM model, and the coordinated signal control will be tested as well. Besides
that, this paper focuses on how to improve the layout of the intersection, the
configuration of traffic signal could also be considered.
REFERENCES
Günther, G., Coeymans, J. E., Muñoz, J. C., and Herrera, J. C. (2012). "Mitigating
freeway off-ramp congestion: A surface streets coordinated approach."
Transportation Research Part C: Emerging Technologies, 20(1), 112–125.
Hunt, P. B., Robertson, D. I., Bretherton, R. D., and Winton, R. I. (1981). "SCOOT - A
Traffic Responsive Method of Coordinating Signals." Publication of: Transport
and Road Research Laboratory.
Li, H., Li, S., Li, H., Qin, L., Li, S., and Zhang, Z. (2016). "Modeling Left-Turn Driving
13
Behavior at Signalized Intersections with Mixed Traffic Conditions."
Mathematical Problems in Engineering, Hindawi Publishing Corporation, 2016,
1–11.
Lin, L., Li, Y., & Sadek, A. (2013a). "A k nearest neighbor based local linear wavelet
volume
neural
prediction." Procedia-Social and Behavioral Sciences, 96, 2066-2077.
network model
traffic
for
on-line
short-term
Lin, L., Wang, Q., & Sadek, A. (2013b). "Short-term forecasting of traffic volume:
evaluating models based on multiple data sets and data diagnosis
measures." Transportation Research Record: Journal of the Transportation
Research Board, (2392), 40-47.
Lin, L., Wang, Q., Huang, S., & Sadek, A. W. (2014a). "On-line prediction of border
crossing traffic using an enhanced Spinning Network method." Transportation
Research Part C: Emerging Technologies, 43, 158-173.
Lin, L., Wang, Q., & Sadek, A. W. (2014b). "Border crossing delay prediction using
transient multi-server queueing models." Transportation Research Part A:
Policy and Practice, 64, 65-91.
Lin, L., Ni, M., He, Q., Gao, J., & Sadek, A. W. (2015). "Modeling the impacts of
inclement weather on freeway traffic speed: exploratory study with social
media data." Transportation Research Record: Journal of the Transportation
Research Board, (2482), 82-89.
Lin, L., Wang, Q., & Sadek, A. W. (2016). "A combined M5P tree and hazard-based
duration model for predicting urban freeway traffic accident durations."
Accident Analysis & Prevention, 91, 114-126.
Mcelroy, R. (2008). "Traffic Signal Timing Manual."
Mueller, E. A. (1970). "Aspects of the History of Traffic Signals." IEEE Transactions
on Vehicular Technology.
Niu, H., Li, G., and Lin, L. (2009). "Fuzzy control modeling and simulation for urban
traffic lights at single intersection." Transport Standardization, 17.
PTV. (2011). VISSIM 5.40 User Manual. epubli GmbH, Karlsruhe.
Ring, J. B., and Sadek, A. W. (2012). "Predicting Lane Utilization and Merge Behavior
at Signalized Intersections with Auxiliary Lanes in Buffalo, New York."
Journal of Transportation Engineering, American Society of Civil Engineers,
138(9), 1143–1150.
Tian, Z. Z., and Wu, N. (2006). "Probabilistic Model for Signalized Intersection
Capacity with a Short Right-Turn Lane." Journal of Transportation
Engineering, 132(3), 205–212.
Trabia, M. B., Kaseko, M. S., and Ande, M. (1999). "A two-stage fuzzy logic controller
for traffic signals." Transportation Research Part C: Emerging Technologies,
Elsevier Science Ltd, 7(6), 353–367.
14
Wu, N. (2007). "Total Approach Capacity at Signalized Intersections with Shared and
Short Lanes: Generalized Model Based on a Simulation Study." Transportation
Research Record: Journal of the Transportation Research Board, 2027(2027),
19–26.
15
|
1805.00787 | 1 | 1805 | 2018-04-09T04:12:08 | Cognition in Dynamical Systems, Second Edition | [
"cs.MA",
"cs.AI",
"cs.GT",
"cs.LG"
] | Cognition is the process of knowing. As carried out by a dynamical system, it is the process by which the system absorbs information into its state. A complex network of agents cognizes knowledge about its environment, internal dynamics and initial state by forming emergent, macro-level patterns. Such patterns require each agent to find its place while partially aware of the whole pattern. Such partial awareness can be achieved by separating the system dynamics into two parts by timescale: the propagation dynamics and the pattern dynamics. The fast propagation dynamics describe the spread of signals across the network. If they converge to a fixed point for any quasi-static state of the slow pattern dynamics, that fixed point represents an aggregate of macro-level information. On longer timescales, agents coordinate via positive feedback to form patterns, which are defined using closed walks in the graph of agents. Patterns can be coherent, in that every part of the pattern depends on every other part for context. Coherent patterns are acausal, in that (a) they cannot be predicted and (b) no part of the stored knowledge can be mapped to any part of the pattern, or vice versa. A cognitive network's knowledge is encoded or embodied by the selection of patterns which emerge. The theory of cognition summarized here can model autocatalytic reaction-diffusion systems, artificial neural networks, market economies and ant colony optimization, among many other real and virtual systems. This theory suggests a new understanding of complexity as a lattice of contexts rather than a single measure. | cs.MA | cs |
Cognition in Dynamical Systems
Second Edition
Jack Hall
May 3, 2018
Abstract
Cognition is the process of knowing. As carried out by a dynamical
system, it is the process by which the system absorbs information into its
state. A complex network of agents cognizes knowledge about its environ-
ment, internal dynamics and initial state by forming emergent, macro-level
patterns. Such patterns require each agent to find its place while partially
aware of the whole pattern. Such partial awareness can be achieved by sep-
arating the system dynamics into two parts by timescale: the propagation
dynamics and the pattern dynamics. The fast propagation dynamics describe
the spread of signals across the network. If they converge to a fixed point for
any quasi-static state of the slow pattern dynamics, that fixed point repre-
sents an aggregate of macro-level information. On longer timescales, agents
coordinate via positive feedback to form patterns, which are defined using
closed walks in the graph of agents. Patterns can be coherent, in that every
part of the pattern depends on every other part for context. Coherent pat-
terns are acausal, in that a) they cannot be predicted and b) no part of the
stored knowledge can be mapped to any part of the pattern, or vice versa.
A cognitive network's knowledge is encoded or embodied by the selection of
patterns which emerge. The theory of cognition summarized here can model
autocatalytic reaction-diffusion systems, artificial neural networks, market
economies and ant colony optimization, among many other real and virtual
systems. This theory suggests a new understanding of complexity as a lattice
of contexts rather than a single measure.
Acknowledgments
Many professors have been very patient with me.
Dawn Tilbury and Benito Fernandez set me on my path.
Raul Longoria kept me in graduate school.
Matthew Campbell and Maxwell Stinchcombe made me explain myself.
Luis Sentis balanced my realism with optimism.
My friends were a boon.
Courtney Shell and Nick Paine kept me sane.
Felipe Lopez, Prashant Rao and Taylor Niehues kept me laughing and
made sure my glass was full.
Gray Thomas always listened to my latest strange idea.
My mother and sisters held me up and reminded me that they love me
anyway.
Thank you all.
Contents
Chapter 1 Introduction
1
Chapter 2 Signal Propagation
5
5
2.1 A Naive Approach . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Propagation Dynamics . . . . . . . . . . . . . . . . . . . . . .
7
2.3 Existence of an Aggregation . . . . . . . . . . . . . . . . . . . 10
2.4 Separating Propagation from Pattern . . . . . . . . . . . . . . 13
2.5 Retrospective on Propagation . . . . . . . . . . . . . . . . . . 16
Chapter 3 Pattern and Feedback
19
3.1 Existing Approaches to Pattern . . . . . . . . . . . . . . . . . 19
3.2 Feedback Loops . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Coherence and Causality . . . . . . . . . . . . . . . . . . . . . 26
Chapter 4 Applications
29
4.1 Dissipative Structure . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Artificial Neural Networks . . . . . . . . . . . . . . . . . . . . 32
4.3 Market Economies
. . . . . . . . . . . . . . . . . . . . . . . . 37
4.4 Ant Colony Optimization . . . . . . . . . . . . . . . . . . . . . 39
Chapter 5 Conclusions
43
5.1 Context in Cognition . . . . . . . . . . . . . . . . . . . . . . . 43
5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
References
49
Chapter 1
Introduction
Ants are both familiar and mysterious. You can find a colony almost
anywhere, but how do they dig such intricate tunnels, forage as a group and
wage war? How do they decide which of these they should all be doing,
when each ant - including the queen - is quite stupid? Ants signal each
other by touch, sound and pheremone, and these signals coordinate them in
ways no individual ant could fathom. The signals propagate from ant to ant,
inhibiting, modifying, amplifying and directing each other so that the ants,
in obeying them, act for the colony's good.
This complexity would not be necessary in a simple or isolated en-
vironment. A colony must simultaneously find enough food, secure its nest
against predators, dispose of waste and care for its young. Despite their caste
system most ants perform many tasks. A worker may retrieve some food,
store it, repair a collapsed tunnel, then go back to harvesting food. Each ac-
tion is a response to stimulus from the environment and from other ants. An
ant that signaled for more food may become sidetracked to help the queen
brood, but the hunger signal remains and propagates to other ants, some of
whom respond. The colony as a whole constantly responds to myriad stimuli,
allocating resources and collectively making decisions.
Asked to explain this collective cognition, you could point to evolu-
tion by natural selection. Every generation, each ant's response to stimulus
changed slightly, which tweaked the collective behavior. Successful tweaks
survived, and ant societies grew more complex. But this story tells us little
about how each ant colony reacts to its environment. How do those entan-
1
gled patterns of signaling bridge environmental stimulus to colony survival?
How might they be engineered?
Cognition is the process of knowing.
While cognition commonly refers only to human thought, this definition
comes from cybernetics. For a dynamical system to know, information must
be encoded in its state. That information may come from the environment
or quirks of the system's own nature.
A system cognizes when it absorbes information, internal or
external, into its state.
To store any significant amount of information, the system must have
a very "large" state space. Real numbers can technically hold an arbitrary
amount of information, and the theory of dynamical systems whose state
are vectors of reals is well-developed. Could we use that theory to describe
cognition? Unfortunately, no. To store a message, the system must find
a corresponding region of its state space and remain there. The system's
dynamics function as an encoder of the message. To increase the size of the
message, a state space would need to be split into increasingly small stable
regions, and the dynamics would become increasingly difficult to analyze.
Furthermore, each such region would border on a limited number of others,
making it hard for the system to encode an arbitrary message starting from
arbitrary initial conditions.1
Alternatively, we can increase the dimensionality of the state space.
To keep the dynamics tractable, we need to restrict the ways in which the
state variables affect each other. Complex networks fit the bill. The state
space of a complex network has an arbitrary number of dimensions, and each
agent affects only itself and its neighbors. Certain complex networks are
already known to absorb information from their environment.2 All living
systems are complex networks, as are some nonliving systems.
A complex network is a system of agents that selectively inter-
act with each other in ways that are hard to understand solely
from knowledge of the individual agents.
1If the system starts with no prior information about the message, the initial conditions
will certainly be arbitrary.
2This is usually described as acquiring order from the environment.
2
An agent is an indivisible entity which acts of its own accord.
If it was not already clear, I will not investigate ant colonies in par-
ticular. Instead, I will look at an entire class of complex networks - to which
ant colonies belong - that produce unpredictable but stable and coherent
patterns. Each of these words - unpredictable, stable, coherent - is both vi-
tal and frustratingly vague. If the patterns were predictable, we would have
no use for the complex system. If the system did not find a stable region of
its state space, then it cannot store information. And to store a single large
message rather than many unrelated and trivial messages, agents must spe-
cialize and coordinate - even though each agent cannot directly affect most
others.
A pattern is coherent if each part of it depends on the rest.
Unfortunately, no mathematics exist that describe pattern or stability
in a complex system.[13] Without such mathematics, any reasoning about
complex systems will be as fuzzy as the words I defined in the last paragraph.
I will articulate, in this thesis, mathematics that underpin cognition. These
mathematics could be used to:
• control large distributed systems like smart grids and reconfigurable
factories
• solve hard optimization problems and automate engineering design
• coordinate and plan complicated robotic motions
The mathematics of cognition will be necessary - though likely not suffi-
cient - for growing strong artificial intelligence. We may also come to better
understand natural systems like market economies and ecosystems. That un-
derstanding could help us to develop better public policy and safely manage
the environment. And there are many other ways to apply cognition; I am
sure that I will never see them all.
The next two chapters will construct this new theory. I will begin with
the interactions between agents, studying the way information propagates
across a network. Then, in Chapter 3, I will set forth how agents coordinate
and form patterns through the use of positive feedback.
In Chapter 4, I
will apply the mathematics of the first two chapters to a set of more concrete
3
models. My goal is a set of abstractions that explain a wide range of concrete
cognitive systems.
In my last chapter, I will examine why I chose those abstractions and
not others. The concept of context will prove to be a strong guiding principle.
It may seem, by the end, that I present some of this material out of order. I
do apologize for any difficulties in understanding my text, but we are all at
the mercy of cognition itself. It emerges whole, of its own accord.
4
Chapter 2
Signal Propagation
This chapter will describe how information propagates across a com-
plex network. By the end, I will have explained how agents gain access to
macro-level information through the aggregations each agent performs for its
neighbors.
2.1 A Naive Approach
The standard model of a complex network is a directed graph G =
(V, E) where the vertices V are the agents. For convenience, I will define two
functions, T (v) = {(v′, v) : (v′, v) ∈ E} and F (v) = {(v′, v) : (v′, v) ∈ E},
to give the arcs incident to and incident from agent v, respectively. If there
is no chance of ambiguity, these can be written T v and F v. The state of
each agent v is governed by some function ψt
v. The graph specifies which
agents directly affect the state of which others; the state of agent v depends
on the state of agent v′ if and only if (v′, v) ∈ E. This means that if time is
discrete (t ∈ Z), then ψ1
v maps the states of agents {v} ∩ T v to the state of
agent v, and if time is continuous (t ∈ R), then dψv/dt is a function of those
same agent states. By changing their states over time and influencing their
neighbors, agents are said to act per their definition.
For agents to coordinate with each other, they must exchange informa-
tion. Unfortunately, we cannot study solely how information spreads across
the network, because the standard model ties up that spread with changes
in agent state. To separate the two, I will complicate the model slightly.
5
b
f
a
e
c
h
d
g
i
Figure 2.1: A directed graph. As an example of my notation, T b =
{(a, b), (d, b)} and F b = {(b, c), (b, d)}.
A signal is the action of one agent on another.
Let every signal be a point in X. The signal xe ∈ X moves across arc e ∈ E.
The vector of signals in a set of arcs A ⊆ E is a point xA ∈ X A, the ·
notation being the cardinality of A.
I will use different subscripts on the
same variable to select subvectors, so xA is a subvector of xE ∈ X E, and
zA is a subvector of zE ∈ X E. Agent v's state now depends the signals xT v
rather than the states of agents in v′ : (v′, v) ∈ T v. I will take the unusual
step of identifying each agent's state with a function that maps input signals
to output signals.
fv : X T v 7→ X F v
fv ∈ Fv
(2.1)
Fv is the state space through which each agent v moves. Having now adjusted
the standard model, how do agents exchange information?
I will briefly take the role of a single agent. If I am to join a coherent
pattern, I must coordinate with other agents to which I am not necessarily
adjacent. But I am only a simple agent, so my state cannot specifically reflect
these distant agents.1 Any information I exchange with distant agents must
therefore flow through my neighbors. I can tell my neighbors apart, but there
is no way to know the ultimate source of their information - or for them to
1I use "simple" here in a very narrow sense. An agent can itself be complex, but by
this assumption it cannot cognize the whole network of which it is a part.
6
know the ultimate destination of mine. Nonetheless the information channel
between me and any distant agent consists of chains of adjacent agents. All
agents constantly move, which changes the way they direct my signals. By
the time my signal reaches a distant agent, and that agent responds, the
information channel between us has changed in response to other signals as
well as its own collective state. My state has changed too, for that matter.
You might argue that agent movements in and of themselves do not
doom my communication with the distant agent. If agents move cyclically,
then I may be able to sychronize to some extent with that distant agent.
But if agents only move in cycles, the whole network will simply vibrate
indefinitely, making no progress towards a coherent pattern.
Insofar as a
cycle is steady, it amounts to a fixed state. And if the cycle changes over
time (again at a rate similar to that of signal travel), then our information
channel again loses integrity.
So coordinating with distant agents seems unlikely. But how then can
they form the coherent patterns necessary to store large messages?
2.2 Propagation Dynamics
To foster coherent patterns, let us try slowing down agent movement
to the point where they are quasi-static on signaling timescales. Each agent,
if its function is fixed, simply maps input signals to output signals. Our single
complex system is now two interrelated systems: one for the propagation of
signals across the network and one for changes in agent state. For brevity,
let FV = Qv∈V Fv. A point fV ∈ FV is a vector of functions fv for every
agent v ∈ V . Similar to the convention of signal notation, fv is a component
of fV , and gv is a component of gV .
Definition 2.1. The propagation dynamics on G = (V, E) are a flow (or
an iterated function) Γt : FV × X E
7→ X E parameterized by the agent
functions fV . A point zE is a fixed point of the propagation dynamics only
if zF v = fv(zT v) for all agents v ∈ V .
The propagation dynamics define the way information spreads across
the network. Usually they are stochastic and discrete, although as you will see
in Chapter 4, the propagation dynamics can be stochastic or deterministic,
discrete or continuous.
7
Our second system - the slow counterpart to the propagation dynamics
- describes the emergence of patterns. To limit the speed of agent motion,
there must be a metric on Fv. I will call it dv.
Definition 2.2. The pattern dynamics on G = (V, E) are a continuous
flow ψt
V : FV 7→ FV that describes the motion of every agent function over
time, starting from any point in FV . The motion of a single agent v is
denoted ψt
v and is Lipschitz continuous (with a Lipschitz constant L ≥ 0):
t ∈ R,
∀fv ∈ Fv,
∀v ∈ V
(2.2)
dv(cid:0)fv, ψt
v(fv)(cid:1) ≤ Lt
An agent's motion at any instant is determined only by its current state and
the signals it receives.
I would like to make the last statement in that definition more formal,
but I cannot do so quite yet. If ψv were differentiable, then its time derivative
would be a function of the current fv and current xT v. Unfortunately, re-
quiring differentiability would cripple a theory of cognition; there are many
cognitive networks with non-differentiable states.
I will discuss several of
them in Chapter 4. For now, think of each continuous agent flow ψv as ap-
proximated by an iterated function of the current state fv and the current
input signals xT v, where each iteration moves the agent infintesimally.
The propagation and pattern dynamics must be kept separate. But
what does this mean? Say that the propagation dynamics always converge
to a fixed point. According to Definition 2.1, the fixed point depends on
the agent functions. Let p : FV 7→ X E map the agent functions to a fixed
point of the corresponding propagation dynamics. As the pattern dynamics
progress, the fixed point of the propagation dynamics would move. To be
separate from the pattern dynamics, the propagation dynamics should stay
arbitrarily close to their fixed point. That way, each agent v will change its
state as a function of p(fV )T v.
"Closeness" implies metric on X E. Let (X, ρ) be a complete metric
space and A ⊆ E. Define a metric ρA(xA, yA) = Pe∈A ρ(xe, ye) on X A.
(X E, ρE) is also a complete metric space.2 Resist the temptation to think
of signals as real numbers. This interpretation is quite limiting, as you will
see in Chapter 4. A better interpretation of signals is as measureable sets of
stimuli. If µ : X 7→ [0, ∞) is the measure and △ is the symmetric difference,
then ρ(x, y) = µ(x△y).
2I have chosen the L1 norm to combine vectors of signals, but there are other options.
8
Definition 2.3. At time t = 0, let xE be the state of the propagation
dynamics and fV be the state of the pattern dynamics. The pattern dynamics
are separated from the propagation dynamics at time t if there exists an
arbitrarily small ǫ > 0 such that:
(2.3)
ρE(cid:16)Γt(cid:0)ψt
V (fV ), xE(cid:1), p(cid:0)ψt
V (fV )(cid:1)(cid:17) ≤ ǫ
Agent states map to signals via p, which in turn determine how those
agent states change over time. I will provide an example proof of separation
in Section 2.4.
The propagation dynamics, separated from the pattern dynamics,
solve a major standing problem in complex systems. Emergence is gener-
ally agreed to come from feedback loops between the micro and macro levels
of a complex system[3], but most models have trouble conveying macro-
level information to individual agents. They mostly choose between two
approaches.3 The first - taken by cellular automata - is to let information
percolate on its own. As we saw in Section 2.1, this is no different from
ignoring the problem and hoping for a miracle.
The second approach is to compute a single, crude aggregate of all the
agent states and make that single aggregate available to every agent. These
crude aggregations force a tradeoff between agent complexity and pattern
coherence. On one end of the scale, the crude aggregate perfectly summarizes
the state of the whole system. Each supposedly simple agent must then be
complex enough to parse this information and understand its part in the
pattern - a clear contradiction. Multi-agent systems usually operate at or
near this extreme. On the other end of the scale, the aggregate contains
no useful information about macro-level pattern. Without it, agents cannot
specialize. Without specialization, a pattern cannot be coherent. Agent-
based models often use these uninformative aggregations.
Propagation dynamics sidestep the complexity/coherence tradeoff by
distributing the work of aggregation. Each agent aggregates information for
its neighbors. This concept is not new; Freidrich Hayek had noticed by 1945
that actors in a market aggregate information about the scarcity of each
product by setting prices.[7] To know how much of a good to make, the
producers of that good need not trace all the ways in which it is used or
3The few exceptions - artifical neural networks come to mind - invariably mimic specific
natural systems. No one has explained precisely why they work.
9
transformed by the rest of the economy; they need only observe changes in
price. Likewise to know how much of each good to use, consumers need not
keep track of improvements or disruptions in every supply chain. No human
could ever be fully aware of all the decisions made in the economy, yet those
decisions cohere.
Definition 2.4. The aggregate of the graph G and agent states fV is given
by a function p : FV 7→ X E such that p(fV ) is always a fixed point of the
propagation dynamics on G. The function p is the aggregation.
In the course of modeling any particular cognitive network, it's natural
to ask which information propagates and which doesn't. To keep my theory
as general as possible, I have avoided this question and tried not to assume
anything about the answer.
2.3 Existence of an Aggregation
The aggregation p only exists if the propagation dynamics always have
a stable fixed point, preferably a unique one. Any proof of convergence to
a fixed point will depend on the particular application of this theory. This
section contains two simple examples.
Consider first the following propagation dynamics: every agent v, si-
multaneously and at intervals in time, applies fv to its inputs xT v and updates
its outputs xF v with the result.
Γ1(fV , xE) =(cid:0)fv(xT v)(cid:1)v∈V
Definition 2.5. Let α ∈ (0, 1) and xT v, yT v ∈ X T v. An agent function fv
contracts if:
αρT v(xT v, yT v) ≥ ρF v(cid:0)fv(xT v), fv(yT v)(cid:1)
Theorem 2.1. If every agent function contracts, then the propagation dy-
namics defined by Equation 2.4 converge exponentially to unique fixed point.
(2.4)
(2.5)
10
Proof.
ρE(Γ1(fV , xE), Γ1(fV , yE)) =Xe∈E
=Xv∈V Xj∈F v
=Xv∈V
≤Xv∈V
ρ(cid:0)Γ1(fV , xE)e, Γ1(fV , yE)e(cid:1)
ρ(cid:0)fv(xT v)j, fv(yT v)j(cid:1)
ρF v(cid:0)fv(xT v), fv(yT v)(cid:1)
αρT v(xT v, yT v)
(2.6)
Γ1 is a contraction mapping. All contraction mappings converge exponen-
tially to a unique fixed point.
ρE(cid:0)Γ1(fV , xE), Γ1(fV , yE)(cid:1) ≤ αρE(xE, yE)
Updating all agents simultaneously requires the agents to synchronize,
which - according to my assumption that agents only coordinate via signals
- they are unable to do unless the propagation dynamics converge. The
synchronous propagation dynamics in Equation 2.4 are physically unlikely.
Consider now the propagation dynamics in which each agent updates
randomly according to a Poisson process in time. The Poisson processes are
independent and identically distributed with a rate λ. These asynchronous
propagation dynamics Λ are a jump process on X E whose jumps are them-
selves Poisson-distributed in time, albeit with a much faster rate than λ.
Letting xAkxB = xA∪B concatenate two vectors indexed by disjoint sets A
and B, each jump moves xE to xE−F vkfv(xT v) for a randomly drawn agent
v.
Theorem 2.2. If every agent function contracts, Λ contracts to a unique
fixed point.
Proof. Since Γ has only one fixed point zE, then by Definition 2.1 so does Λ
have the same fixed point. Let the interval [a, b) be an interval in time over
which every agent has updated at least once. The expected length of this
interval can be derived from the Poisson distribution.
1/2 = (1 − P (N(b − a) = 0))V
1/2 = (1 − e−λ(b−a))V
11
E(b − a) = −
1
λ
ln(1 − 2−1/V )
(2.7)
Because this interval grows only logarithmically with the number of agents,
the number of agents matters little. The actual size of each interval has a
probability distribution, but the law of large numbers asserts that the time-
average of interval lengths gets arbitrarily close to the expectation of that
length as time goes on.
The propagation dynamics contract at least as much over the interval
[a, b) as they would in one synchronous update. The distance between the
trajectories from xE and yE over time is therefore bounded from above by
an exponential decay.
αt/(b−a)ρE(xE, yE) ≥ ρE(cid:0)Λt(fV , xE), Λt(fV , yE)(cid:1)
(2.8)
Having shown that aggregations exist for propagation dynamics Γ and
Λ, I now ask: How complete is the aggregate? For two distant agents to co-
ordinate, the aggregate available to each must be sensitive to the state of
the other. In other words, the signals available to each agent should include
at least some information about every other agent. Given the restriction on
fixed points in Definition 2.1, this property depends only on the agent func-
tions (and the graph) rather than propagation dynamics they parameterize.
I can see two scenarios in which information fails to spread properly.
If the graph grows too large and sparse, the shortest path between
a pair of agents can grow arbitrarily long. Agent contraction does not just
bound the time in which signals converge, it also limits how far they can
spread through the graph. Agents can (and usually will) prioritize certain
signals over others, which will cause the former to propagate further than
the latter. Nevertheless, contractive agents cannot pass on more than a
fraction of the signals they receive. Over a long enough path, even the
strongest and most important of signals must fade. One solution is to give
large networks the small-world property4, thereby putting an upper bound
on the shortest path between any two agents. Another solution is to relax
the rate of convergence from exponential to asymptotic.
4in which the number of arcs to and from any particular agent has a geometric prob-
ability distribution. In simple terms: while most agents have relatively few arcs, a few
agents have many, and a very few have a huge number.
12
Signals can also fail to propagate if agents direct them based not on
their content but on their origin. Obviously a closed community of agents -
one with no arcs to the rest of the graph - will never send any signals to the
rest of the graph. The same thing can happen without an explicitly closed
community if agents make some of their output signals insensitive to changes
in their input signals. While this behavior can sometimes be desireable, a
"virtual" community of agents whose signals to the rest of the network are all
insensitive to changes within the community is no different from an explictly
closed community.
2.4 Separating Propagation from Pattern
I defined separation in Definition 2.3 as the propagation dynamics
staying close to an aggregate as that aggregate moves in response to the
pattern dynamics. In this section, I will discuss how stability of the propa-
gation dynamics together with Lipschitz continuity of the pattern dynamics
can maintain separation.
The aggregate should not move too fast, compared to the rate at
which the propagation dynamics converge.
I will use a differentiable and
decreasing function of time φ : [0, ∞) 7→ [0, 1] to characterize the rate at
which the propagation dynamics converge.
φ(t)ρE(xE, p(fV )) ≥ ρE(Γt(fV , xE), p(fV ))
t ≥ 0
(2.9)
Because Γ0(fV , xE) = xE, φ(0) = 1. If the propagation dynamics are asymp-
totically stable, then limt→∞ φ(t) = 0.
Since Definition 2.2 limits the speed at which the agent functions fV
move, the aggregation should also be Lipschitz continuous. Having supposed
a metric dv on Fv in Section 2.2, I will use the average of those distances as
a metric on FV .
dV (fV , gV ) =
dv(fv, gv)
(2.10)
1
V Xv∈V
This particular dV will make Theorem 2.3 slightly easier to prove, but there
are other viable choices. As far as dv is concerned, a natural choice is:
dv(fv, gv) = sup
xT v
ρF v(cid:0)fv(xT v), gv(xT v)(cid:1)
13
(2.11)
Theorem 2.3. Let the aggregate p(fV ) = zE. The aggregation p is Lipschitz
continuous if there exists an M > 0 such that:
M(1 − φ(t)) dV (fV , f ′
V ) ≥ ρE(cid:0)zE, Γt (f ′
Proof. Let z′
E = p(f ′
V ). Starting from the triangle inequality:
t ≥ 0
V , zE)(cid:1)
E(cid:1) ≥ ρE(cid:0)zE, z′
E(cid:1)
E)
t ≥ 0
(2.12)
(2.13)
ρE(cid:0)zE, Γt(f ′
V , zE)(cid:1) + ρE(cid:0)Γt(f ′
V ) + φ(t)ρE(zE, z′
V , zE), z′
E) ≥ ρE(zE, z′
M(1 − φ(t)) dV (fV , f ′
This inequality is satisfied trivially at φ(0) = 1. At all other points in time,
M dV (fV , f ′
V ) ≥ ρE(zE, z′
E)
M dV (fV , f ′
V ) ≥ ρE(cid:0)p(fV ), p(f ′
V )(cid:1)
(2.14)
Equation 2.12 limits the rate at which the propagation dynamics, after
a change in the agent functions, can diverge from the previous fixed point.
Equation 2.13 pits this rate against the rate of convergence. It should not
actually be necessary for these two rates to cancel out, so long as the rate of
convergence exceeds the rate of divergence.
Theorem 2.4. If the aggregate p is Lipschitz continuous and Equation 2.9
holds, there exists some t′ > 0 such that Γ and ψV are separate for all t > t′.
Proof. Let xE, fV be the state of the cognitive network at t = 0, so zE =
p(fV ) is the initial aggregate. Given some small δ > 0, let x′
E = Γδ(fV , xE)
and z′
V (fV )). Starting again from a triangle inequality:
E = p(ψδ
ρE(x′
E, zE) + ρE(zE, z′
E) ≥ ρE(x′
φ(δ)ρE(xE, zE) + MLδ ≥ ρE(x′
E, z′
E, z′
E, z′
ρE(x′
E)
ρE(xE, zE)
ρE(xE, zE)
φ(δ) +
MLδ
≥
E)
E)
(2.15)
(2.16)
If the right-hand side of Equation 2.16 is greater than one, the signals
get further away from the aggregate over the interval δ. Less than one, and
they get closer. If ρE(xE, zE) = ǫ/2 and the right-hand side is one, then:
1
2
ǫ = lim
δ→0
MLδ
1 − φ(δ)
14
(2.17)
We can ensure this limit exists by requiring that:
d
dt
MLt ≤
d
dt
(1 − φ(t))
t ∈ [0, δ]
ML ≤ −
dφ
dt
t ∈ [0, δ]
(2.18)
By decreasing L further, we can make ǫ arbitrarily small.
Equation 2.15 describes a single step through time. If ρ(xE, zE) > ǫ/2,
then the signals Γδ(fV , xE) step closer to p(ψδ
V (fV )). Adding more steps and
taking the limit as δ → 0, the signals will approach the (ǫ/2)-ball around the
aggregate. At some finite time t′ they will be within ǫ of the aggregate, and
they will never leave that ǫ-ball.
ρE(cid:16)Γt(cid:0)ψt
V (fV ), xE(cid:1), p(cid:0)ψt
V (fV )(cid:1)(cid:17) ≤ ǫ
t > t′
As long as the propagation and pattern dynamics are separated, agent
v reacts only to p(ψt
V (fV ))T v, or at least to arbitrarily similar signals. If we
let ǫ grow, ψV will become stochastic; the propagation dynamics are often
stochastic themselves, and there is no way for an agent to know where in the
ǫ-ball the aggregate actually is. It will react to whatever signals it sees.
In fact, the fixed point of propagation dynamics need not be unique.
As long as the propagation dynamics stay near one of the fixed points and
the fixed points don't pass within 2ǫ of each other, then the system's choice of
a fixed point is unambiguous, and the aggregation is well-defined. The price
we would pay for the extra flexibility is hysteresis in the pattern dynamics.
If multiple fixed points exist, then there is no way to know which one the
system will choose without knowing which fixed point it has chosen in the
past.
The separation proof can also help us to simulate cognitive networks
with a computer. The constants M and L along with the convergence func-
tion φ imply an upper limit on the time step δ, beyond which Equation 2.18
no longer holds.
15
2.5 Retrospective on Propagation
We started with a goal: to design dynamical systems that cognize -
that perform the process of knowing. A cognitive system, simply by obeying
its own dynamics, absorbs information and stores that information in its
state. We focused on complex networks because their state space has many
dimensions, and high-dimensional spaces can store more information. But to
make full use of that state space, the agents in the network must coordinate
with distant counterparts and specialize.
This is quite difficult, because an agent cannot be explicitly aware
of any other agents with which it does not share an arc. Agents may only
communicate through other agents. Any change in the agents thus alters
their communication channels.
To keep those communication channels stable, we slowed down
changes in agent state, giving the propagation dynamics time to converge
to a fixed point. Each component of that fixed point is a signal traveling
over an arc in the directed graph. That single signal can reflect, to varying
degrees, the state of every agent. Because the propagation dynamics are
parameterized by the agent states, the fixed point moves as the agents do,
and the propagation dynamics never quite reach it. But the closer they are
to their fixed point, the more certain each agent is about the information it
receives. Therefore the slower the agents move relative to the propagation
dynamics, the more deterministic the agents' motion.
Although each signal goes from one agent to another, the communi-
cation between agents is many-to-many rather than one-to-one. Every agent
sees a few components of the aggregate, as produced by its neighbors. These
were in turn produced from other components of the aggregate, as produced
by their neighbors. The propagation dynamics are the process of aggregation
- the process by which the micro level becomes aware of the macro level.
I identify cognition with the emergence of patterns, and the emergence
of patterns is widely identified with feedback between the micro and macro
levels of a complex network[3]. I have managed to explicitly link the micro
level to the macro level - one half of the required feedback process. To my
knowledge, this has never been done before with simple agents and without
any omniscient view of the network.
As the agents move in response to that aggregate (at least the com-
16
ponents they can see), the aggregate responds to their motion in turn. The
conditions for distributed aggregation are few and broad. They come in two
classes, which I call primary and secondary. Primary conditions must apply
to any network, regardless of the fixed point proof used for the propagation
dynamics.
• Agents must be quasi-static on the signaling timescale. (L is small.)
• Signals must have a metric (ρ).
• The propagation dynamics must converge. (At least as fast as φ.)
• The propagation dynamics Γt(gV , p(fV )) must not diverge too quickly
from p(fV ). (Equation 2.12)
The first two conditions apply to the micro-level, which makes them easy to
enforce. While the particular reasons vary from field to field, the heart of
the matter is that macro-level behavior is difficult to understand or control.
To try is usually to fail or to diminish the network's cognitive ability. For
instance, what if I were to stipulate that the propagation dynamics contract
without requiring that each agent contract? Some agents would be allowed
an α ≥ 1, but which ones? It would only take two (mutually adjacent) ex-
pansive agents to prevent convergence. Such a scheme is not impossible - the
agents could coordinate their contraction or expansion as part of the pattern
dynamics. But it would be much more difficult to analyze and control.
The second two primary conditions avoid this difficulty through the
use of abstraction. They are metaconditions - conditions about conditions.
Any concrete model of cognition will add its own set of conditions - the
secondary conditions. To give an example, the proofs in Section 2.3 both
require that each agent contract individually. That secondary condition need
not hold for every cognitive network; the proof of a metacondition can differ
for every application.
There is another metacondition hidden in Definition 2.1: the require-
ment that, at a fixed point zE of Γt(fV , ·), zF v = fv(zT v) for every agent v. It
means that the set of potential fixed points is determined on the micro level;
the propagation dynamics - a macro phenomenon - only select a fixed point
from that set. The two versions of the propagation dynamics in Section 2.3
(Γ1 and Λt) actually have the same fixed points.
The three metaconditions I've created so far are relatively easy to
tackle because they only restrict the propagation dynamics. Because the
17
propagation dynamics are often specific to a problem domain anyway, the
secondary conditions necessary in the special case sacrifice no descriptive
power or ease of application in the general case.
I have striven to keep
the my primary conditions as few and as broad as possible. The agents,
for example, are still black boxes. Their state is not parameterized, which
frees us to use nearly any parameterization we like. They need not even
be uniform. Another example are the signals, which only need a notion of
distance.
Most study of complex networks centers on the agents, trying to an-
swer questions like: What do they do? How do they interact? My theory
shifts focus away from the agents and toward their signals: How do signals
propagate? What information do they carry? Aggregation (as in Definition
2.4) creates a many-to-many communications channel, and Section 2.4 veri-
fies that channel's integrity. That still leaves us asking what information it
conveys. There are as many answers to this question as there are problem
domains, but many domains have commonalities. For example, an agent
that can be connected to an arbitrary number of others must combine an
arbitrary number of signals. A natural way to do this would be to recur-
sively fold pairs of signals together - to give them algebraic structure. Signal
algebras could help determine which information spreads across the network
and which does not.
In addition to shifting focus from agents to signals, I have also changed
my view of - for lack of a word - that which cognizes. Artificial intelligence
researchers usually try to create cognitive algorithms. Creating an algorithm
means breaking down a task or process into discrete steps which mostly occur
in linear order. Aside from the fact that writing a cognitive algorithm has
proven to be quite a nasty problem, this approach fosters bad habits of mind.
Because the default point of view is omniscient, we rarely ask key questions
about information - how it spreads, how it's stored, where it gets used. We
get used to thinking about information in discrete, hierarchically-organized
packets rather than in diffuse gestalts.
Instead of writing cognitive algorithms, I am designing cognitive dy-
namical systems. That means accounting for any transmission of information
via signaling and limiting each agent's use of information. It also creates more
questions: How do states affect each others' motion? Can the system's dy-
namics lead to gestalt configurations of state? When are those configurations
stable? I will start to answer these questions in Chapter 3.
18
Chapter 3
Pattern and Feedback
So now I have separated signaling (the propagation dynamics) from
the motion of agents (the pattern dynamics), and the agents have time to
collectively aggregate macro-level information. As they move, the aggrega-
tion allows them to coordinate their state, together processing information
beyond any individual agent's capability. That information is stored in pat-
terns of agents. In this chapter I will formally define pattern in a cognitive
network, and I will explain how patterns arise from feedback between agent
state on the micro level and the aggregate on the macro level. After that,
I will discuss in broad terms how patterns relate to the information they
store.
3.1 Existing Approaches to Pattern
Patterns encode a complex network's knowledge, and their emergence
is precisely the process of cognition. Despite their importance, most funda-
mental questions about patterns remain unanswered. When do they arise,
and when do they not? If patterns do arise, which ones? We may yet answer
these questions, but not until we know what a pattern is. I danced around
their definition in Chapter 1, when I said that patterns are large configura-
tions of agents and their relationships. These configurations, if the system is
to cognize large chunks of information, must be commensurately large, and
they must also be unpredictable, stable and coherent.
We have tried to identify the building blocks of patterns. Network
19
motifs[15] are small subgraphs with fixed numbers of vertices that repeat
across the network. For a fixed n ∈ N, there are a limited number of possible
n-vertex motifs. The network motif approach starts by identifying, through
the use of statistical tests, which three-vertex motifs occur unusually often.
It then moves on to larger motifs, controlling for any smaller motifs already
found. The approach usually stops at five-vertex motifs. Networks in the
same problem domain exhibit similar motifs, which can teach us about the
style of patterns in that problem domain. But if all networks in a domain have
the same motifs, then those motifs do not encode an individual network's
knowledge. Rather, networks store information by combining motifs into
patterns.
It remains unclear how motifs come to be combined. It's also not clear
that I should be asking how motifs combine instead of asking how agents
combine. To ask how patterns form is to ask how an agent coordinates with
distant neighbors of whom it cannot be aware. I could allow agents to know
the motifs of which they are a part - relaxing my simplicity axiom - but I
could not explain how motifs coordinate without being aware of each other
any more than I could explain how agents do the same. Furthermore, because
I am not working within any particular problem domain, I know less about
individual motifs than I know about individual agents. I conclude that while
motifs are a useful tool on the micro-level, shifting focus from agents to motifs
would add complication without answering my fundamental questions about
cognition and pattern emergence.
Another approach is to form patterns by explicit consensus1 of the
constituent agents. There is a large body of work on consensus networks,
ably reviewed in [17]. The consensus definition of consensus is "to reach an
agreement regarding a certain quantity of interest that depends on the state
of all agents." This usually means that "the state of all agents asymptotically
be[comes] the same." To the degree that agent states do not become the
same, they do not consense. A pattern does not follow automatically from
consensus, but the agents can agree on a pattern to form together. In this
case, the state they agree on must be sufficient to select and describe that
pattern. Because each agent must store this state, we arrive back at a tradeoff
similar to the one I described in Section 2.2.
1Any kind of pattern will constitute at least an implicit consensus.
20
That tradeoff pitted the simplicity2 of the agents against their need for
macro-level information. I resolved it by using the propagation dynamics to
compute an aggregate, thereby distributing the work of aggregation among all
the agents. Consensus networks fail to distribute the storage of information
among all the agents, which pits the simplicity of the agents against the
network's ability to store information. In other words, consensus networks
must trade off agent simplicity against the network's cognitive ability. An
agent that can - by itself - describe large, coherent patterns is no longer simple
relative to the network. Thus consensus and coherence are at odds, and we
find ourselves trading off agent simplicity against the scope and coherence of
pattern.
I seek to sidestep this tradeoff as I did the last... or to sidestep the
same tradeoff again. Are these tradeoffs distinct, or are they aspects of the
same fundamental paradox?
3.2 Feedback Loops
It is generally agreed that patterns in a complex network emerge as
feedback loops between micro and macro behavior[3]. But what is a "feed-
back loop between micro and macro behavior"? In a cognitive network,
the agents (micro level components) collectively compute an aggregate (at
the macro level), which in turn determines agent motion (on the micro level).
That interplay creates the possibility for micro-macro feedback, but it doesn't
guarantee that feedback occurs. What would make that guarantee?
I will start with a pair of agents u, v ∈ V . Each agent has an arc to
the other - (u, v), (v, u) ∈ E. If they coordinate, then z(u,v) should reinforce
z(v,u) and vice versa.3 Reinforcement suggests at least a partial ordering of
signals, which I will denote with "≺".4 This ordering may vary according to
the micro-level context.
Mutual reinforcement also suggests causality. An increase in z(u,v)
should cause agent v to change its function fv in a way that increases z(v,u).
2Narrowly: relative to the network as a whole.
3I'm using z rather than x to remind you that agents react the aggregate.
4If you're still thinking about signals as measureable sets like I suggested in Section
2.2, then perhaps for x, y ∈ X, x ≺ y if and only if x ⊂ y. A stronger ordering might be
x ≺ y ⇐⇒ µ(x) < µ(y).
21
The converse should also be true.
optimizer of some real function qv : X T v+F v 7→ R. For a positive δ → 0,
I will treat each agent v as a greedy
ψδ
v(fv) = arg max
f ′
v∈Fv (cid:8)qv(cid:0)zT v, f ′
v(zT v)(cid:1) : dv(fv, f ′
v) ≤ Lδ(cid:9)
(3.1)
You might protest that making every agent a greedy optimizer is too severe
a restriction, that it would at the very least exclude networks made up of
people. But notice that I have not defined qv.
I am not choosing which
kinds of agents to analyze; I am abstracting the way any agent reacts to
its surroundings. Studying the motion of agents would normally require
calculus, which would in turn require that the state spaces Fv and X be
differentiable. As you will see in Chapter 4, differentiability really is too
restrictive. Equation 3.1 only requires that a utility function qv exist that
agent v greedily optimizes. Game theory will replace calculus.
Greedy optimization can describe a very wide variety of agents and
behaviors. Agents in a market optimize profit, for instance, and molecules
seek their lowest-energy state. Even in more complicated situations, the for-
malism can still be useful. Take the market example. A person might choose
less-profitable work because it lets him or her spend more time with fam-
ily. In that case, utility should reflect both profit and family time. Another
might sacrifice short-term earnings in exchange for long-term gains. We can
express such priorities by reformulating the signals as functions of time.5
For two agents to coordinate, their utility functions must be comple-
mentary in the game-theoretic sense. More simply: each agent must increase
the marginal gain its partner reaps by coordinating. Let xE\x′
e = xE−{e}kx′
e
and more generally xA\x′
A∩B. With that, and given a current
aggregate zE, let xE, x′
E, zE) < ǫ,
where ǫ is the separation tolerance from Definition 2.3. If x(u,v) ≺ x′
(u,v) and
x(v,u) ≺ x′
E ∈ X E such that ρE(xE, zE) < ǫ and ρE(x′
B = xA−Bkx′
(v,u),
(u,v), xF v\x′
qv(cid:0)xT v\x′
(v,u)(cid:1) − qv(cid:0)xT v\x′
≥ qv(cid:0)xT v, xF v\x′
(u,v), xF v(cid:1)
(v,u)(cid:1) − qv(cid:0)xT v, xF v(cid:1)
(3.2)
5Computer simulation of existing cognitive networks may prove to be impractical.
There may not exist a good finite approximation of the signals, or the agents' utility func-
tions may be impossible to observe in full. The optimization carried out by each agent
may defy our power to solve it on a large scale. This potential limitation is troubling, but
I must ignore it in order to continue.
22
b
f
a
e
c
h
d
g
i
Figure 3.1: Feedback between agents 'h' and 'f'.
(v,u), xF u\x′
qu(cid:0)xT u\x′
(3.3)
(u,v)(cid:1) − qu(cid:0)xT u\x′
≥ qu(cid:0)xT u, xF u\x′
(v,u), xF u(cid:1)
(u,v)(cid:1) − qu(xT u, xF u)
This condition - called supermodularity - will not always lead the signals z(u,v)
and z(v,u) to both increase or both decrease. It may be that no fv ∈ Fv exists
that increases z(v,u), in which case agent v would like to continue the feedback
process but cannot. The same could also be true for agent u. Nevertheless,
supermodularity is necessary for positive feedback between our two agents.
Many pairs of agents are not mutually connected. Even if they are,
they may be involved in a larger feedback process that excludes the arcs
that connect them directly. Distant agents communicate through chains of
adjacent agents. A walk should therefore exist from each member of a pair
to its counterpart. Concatenating these walks makes a single closed walk.
Because a closed walk can start from any of its vertices and end up back where
it started, it can be expressed as a subgraph - rather than as a sequence of
alternating vertices and arcs - without losing any mathematical structure.
Thus we can say that for two agents u and v to coordinate, there must exist
some closed walk C = (VC, EC) such that u, v ∈ VC. In the case of Equations
3.2 and 3.3, VC = {u, v} and EC = {(u, v), (v, u)}.
To say that the signals between u and v reinforce each other is to say
that the signals along an entire closed walk reinforce each other. It follows
that if this pair of distant agents coordinates, so too must every agent in
some closed walk that includes the pair. That closed walk could include
23
b
f
a
e
c
h
d
g
i
Figure 3.2: Feedback between (say) agents 'h' and 'd'. Agents 'b', 'c', 'f' and
'h' are equal participants in the feedback loop.
any number of walks between u and v, so long as the entire walk keeps
reinforcing itself. In fact, C could be any strongly connected subgraph of G.
Remember that each agent v is only explicitly aware of itself and the signals
zT v. All distant agents are anonymous and also of indeterminate number
and connectivity. Therefore any feedback process between distant agents is a
micro-macro feedback loop. Patterns arise from positive feedback over closed
walks in the graph.
Definition 3.1. Let C = (VC, EC) be a closed walk in G and zE = p(fV ).
Let the pattern and propagation dynamics be separated by ǫ. Take any
xE, x′
e for all
e ∈ EC. C is a feedback loop if, for all v ∈ VC,
E ∈ X E such that ρE(xE, zE) < ǫ and ρE(x′
E, zE) < ǫ. xe ≺ x′
(3.4)
EC , xF v\x′
qv(cid:0)xT v\x′
EC(cid:1) − qv(cid:0)xT v\x′
EC , xF v(cid:1)
≥ qv(cid:0)xT v, xF v\x′
EC(cid:1) − qv(xT v, xF v)
Definition 3.2. A feedback loop is robust if Equation 3.4 holds for any
fV −VC ∈ FV −VC .
I define robustness here mostly to point out that the agents V − VC
can affect the feedback condition in Equation 3.4. There are likely many
degrees and kinds of robustness, but that analysis is out of my scope.
24
b
f
a
e
c
h
d
g
i
Figure 3.3: Depending on the feedback condition for agent 'f', agent 'h' may
not be necessary to maintain feedback in the rest of the loop. The coherent
feedback loop (with 'h' removed) is pictured.
The union of contiguous closed walks is also a closed walk, and the
If a feedback loop
same is true for unions of contiguous feedback loops.
breaks, parts of it may either remain or not.
Definition 3.3. A feedback loop is coherent if it has no subgraphs that are
feedback loops in their own right.
A feedback loop that is not coherent may include virtual communities
of the kind I mentioned briefly in Section 2.3.
Definition 3.4. A pattern is a feedback loop in which the aggregate sub-
vector zEC is Lyapunov stable.
Definition 3.4 is similar to a Nash equilibrium[6], but it does not
require the entire system to be stable. Even the agents VC in the pattern can
continue to move, so long as the aggregate zEC remains at rest. But how does
the aggregate become stable if it undergoes positive feedback? A runaway
feedback loop will not reach a steady state. The answer is to bound the image
of the aggregation p(FV ) ⊂ X E. Any positive feedback loop in a compact
state space will either reach those bounds or break its own supermodularity
condition first. Since a pattern is a stable feedback loop, it must always lie
on a bound of the aggregate space. In fact, the pattern could not be stable
25
p
FV
p(FV )
X E
Figure 3.4: Bounding the image of the aggregation p. Return to Section 2.2
for details about the aggregation.
without the feedback loop holding it against the boundary. The image of the
aggregation can be manipulated via the agent function spaces Fv.
3.3 Coherence and Causality
Let us return to the specifications we set for pattern in the intro-
duction: stability, coherence and unpredictability. Patterns are stable by
definition. Some are coherent. I will argue that the definition of coherence
in Section 3.2 is a good one, and that coherence implies unpredictability.
If cognition is the process by which a system knows or is aware of
something, then patterns embody and encode that knowledge.
I said in
Chapter 1 that a pattern is coherent if every part of it depends on the rest. By
Definitions 3.1, 3.3 and 3.4, every ze, e ∈ EC can depend in some way on every
state fv, v ∈ VC. Each agent's role in the pattern is defined by its neighbors,
whose roles are in turn defined by their neighbors. A pattern, being a stable
configuration of agent relationships, constitutes an implicit consensus of the
agents VC. Rather than rule out explicit consensus, Definition 3.4 renders it
a trivial special case.6
6See Section 4.4 for an example.
26
In general, a pattern encodes knowledge that no individual agent or
motif can. Furthermore, because no agent could produce the stable zF v∩EC
without the proper zT v∩EC , no one agent or relationship encodes any inde-
pendently useful knowledge. Instead, the smallest useful unit of knowledge
is the coherent pattern. Not all patterns are coherent; a pattern can be com-
posed of smaller, contiguous patterns. But if a pattern is coherent, then any
closed loops within it depend on the rest of the pattern in the same way that
individual agents do.
Where does the knowledge in a coherent pattern come from? I will
examine the various sources of information which a network may cognize:
initial state, the system's own dynamics and the environment. Each of them
will influence most patterns in most networks to some degree.
The environment is the most obvious of the three influences. By the
second law of thermodynamics, no living network can subsist without draw-
ing order from its environment. Despite the close ties between a cognitive
network and its environment, I have not yet specified how the two are re-
lated. Let the graph G have at least one special vertex w, which represents
the environment rather than an agent. Any signals coming from this vertex
are said to enter the network, and any signals sent to it are said to leave the
network. In order for feedback loops including w to stabilize into patterns,
the function fw must be quasi-static on pattern timescales. Since w has no
utility function, the definition of a feedback loop must exclude it. Making the
environment quasi-static is not as restrictive as it sounds, for two reasons.
First: signals can represent probability distributions or cycles of interactions
over time. Second: feedback relations with the environment can be modeled
by simply giving w a utility function and allowing it to move, thereby folding
it into the model.
Some of a network's knowledge comes from its own dynamics.
It
will tend to form certain patterns and not others, based on the objective
functions qV and the search spaces FV . Some aspects of this style can be
captured by network motifs.
Insofar as motifs are common across many
instances of a network model, they isolate style from environmental influence
or prior information. Of course the details of a network need not repeat to
be cognized - feedback loops can amplify the information available to even
one agent. These non-repeating details will not affect the network's motifs.
Lastly, patterns can reflect the network's initial state. They usually
27
will, unless there is only one viable pattern in the whole state space. Initial
states may exist which do not lead to a pattern, even if viable patterns
exist somewhere in the state space. An engineer can prevent this case by
ensuring that the pattern dynamics are topologically mixing in the case that
no pattern arises. In the search for a pattern, the system should eventually
visit any particular region of its state space.
If a coherent pattern reflects knowledge of all three influences above
- initial state, the environment and the system's own dynamics - then these
influences are interwoven and cannot be isolated from one another. Our
inability to correspond subsets of knowledge with subgraphs of a coherent
pattern prevents us from predicting coherent patterns. Specifically, it dis-
rupts the causal relationships between patterns (effects) and the conditions
from which they arise (causes). We cannot know which pattern will arise
from any given set of conditions, and any given feedback loop could have
begun in any number of ways. A system in which cause and effect cannot be
linked is acausal. This is curious because - assuming the pattern and prop-
agation dynamics are perfectly separated - the pattern dynamics are both
acausal and deterministic.
In Chapter 3, I have continued to shift focus away from agents and
towards signals. I defined feedback loops and patterns not in terms of the
agent states but in terms of the signals between the agents. The signals in
a pattern must be stable, yet the stability of the agents in the pattern is
unimportant in and of itself. And despite the focus on signals, my theory
derives much of its power from keeping those signals unrestricted. Formally,
they must have a metric and a partial ordering - nothing more. The partial
ordering can even be unique to each pair of agents (with only minor changes
to Section 3.2). Every concrete example in Chapter 4 will take advantage of
the flexibile definition of a signal.
28
Chapter 4
Applications
This chapter will explain how various domains of study fit into our new
theory of cognition. Its purpose is to better explain cognition by providing
exemplars, rather than to learn about those exemplars per se. Most formal
proofs are also out of scope - a matter for future work. That said, I expect
that developing these applications further will bear fruit in their respective
domains.
Each section will begin with a short overview of the exemplar's model
before drawing an analogy between that model and my theory of cognition.
I will start off with Prigogine's dissipative structures.[18] They will help me
introduce a concept common to many applications: the two-way networks.
Unfortunately, a formal theory of two-way networks must wait for another
time.
Where dissipative structures are physical systems, cognition can also
apply to information processing. Artificial neural networks will serve as an
exemplar. They will be followed by market economies and then by ant colony
optimization. Cognition should apply naturally to ecology as well, but I know
too little of that field to make the attempt.
4.1 Dissipative Structure
As I said in Section 3.3, many cognitive networks absorb their knowl-
edge from their environment. Often, that knowledge reflects two aspects of
29
the environment that are in disequilibrium, by which I mean that the en-
vironment has a high degree of order derived from the separation of those
two aspects. Knowing the disequilibrium allows the cognitive network to
alleviate it by bridging the two aspects, transporting or transforming some-
thing from one aspect to the other. That something, which I will call the
"throughput", may be matter or or information or anything else that can be
expressed as a measurable set. Autocatalytic reaction-diffusion systems are
a simple example of a two-way network.
A reaction-diffusion system is a set of reactions occurring in a set of
places. For our purposes, each reaction in a particular place is an agent,
and sets of reactants make up the signals. The products of each reaction
either react again in the same place, diffuse to another place, or leave the
system. Alan Turing and Ilya Prigogine analyzed reaction-diffusion systems
that produce their own catalysts[19][18]. They found that when concentra-
tions of certain reactants are kept high and the concentrations of certain
products are kept low, the reaction-diffusion system produces its own cata-
lysts so as to increase the rate at which the abundant reactants are trans-
formed into the scarce products. This is autocatalysis. The newly-catalyzed
reaction produces entropy in the throughput faster than before. At the same
time, the entropy in the system's state decreases. The second law of ther-
modynamics remains unbroken because the system produces more entropy
in the throughput than it loses by cognizing the disequilibrium. Eventually
the system stablilizes with high concentrations of the catalysts, far from the
thermodynamic equilibrium.
An autocatalytic loop is a feedback loop - a closed walk in the graph
of reactions over which an increase in each reaction's rate causes the rate of
subsequent reactions to increase. Through these feedback loops, the reaction-
diffision system cognizes the chemical potential between certain of its original
reactants and ultimate products. When the feedback stabilizes, the system
has stored information about that chemical potential as a pattern of reactions
and the reactants they exchange.
I have just described the pattern dynamics of a reaction-diffusion sys-
tem. They looked straighforward; what then makes the reaction-diffusion
system a two-way network? The answer has to do with the flow of infor-
mation. By the law of mass action, rates of reaction depend on the concen-
trations of both reactants and products. As the reactants travel "forward",
information about the products must travel "backward". On fast timescales
30
- while the various chemical concentrations are quasi-static - the flow of in-
formation forward and backward works itself out into deterministic rates of
reaction. The propagation dynamics are the process by which this two-way
flow stabilizes.
Let us call the space of forward signals X and the space of backward
signals Y . Each type of signal in a two-way network travels over its own
graph. The two graphs are mirror images - wherever one has an arc (u, v),
the other has an arc (v, u). The vertex set is the same for both graphs.
Because each graph implies the other, we can still talk about a single graph
G by choosing a convention. Agent functions map both kinds of inputs to
both kinds of outputs.
fv : X T v × Y F v 7→ X F v × Y T v
(4.1)
Proving that the propagation dynamics contract becomes more complicated
than in Section 2.3, and not just because both kinds of signal must converge.
Specifically, there can be no positive feedback between the two types of signal
on those timescales. Positive feedback would create acausality1 where the
fixed point of the propagation dynamics should be causal. If the relationship
between agent states fV and the signals X E were acausal, the aggregation
p : FV 7→ X E × Y E could not exist.
As I have labeled the flow of reactants as forward and the flow of
products as backward, I will label the source of reactants as the "entry" and
the sink of products as the "exit". There is nothing special mathematically
about forward signals or backward signals; these labels are a matter of con-
vention. The entry and exit are special vertices in the graph representing the
environment. They may even be the same vertex - w - that I introduced in
Section 3.3. Like I discussed in that section, fw should not change over time.
Turing and Prigogine, in their work on what Prigogine termed dissi-
pative structures, identified some physical mechanics of cognition. But these
physical mechanics do not describe all cognitive systems. Most living things,
for instance, are cognitive networks in which each agent is itself a cognitive
network. Obviously these cannot be modeled using reaction-diffusion equa-
tions. Cognition is more an abstract pattern than a concrete phenomenon.
Turing and Prigogine's work was also limited by the need to write
down all the equations of motion. Doing this for large networks or interlock-
1I argued in Section 3.3 that pattern emergence is acausal.
31
ing feedback loops is impractical. Adrian Bejan sidestepped this limitation in
his related constructal law[2], but he substituted another: constructal theory
is limited to trees (as opposed to graphs). Bejan's theory was also limited
by its lack of formality.
4.2 Artificial Neural Networks
Artificial neural networks are a computational and mathematical ab-
straction of biological neurons[1]. The "neurons" are arranged in a directed
graph. At any particular time, every neuron is either firing or inactive. It ac-
tivates if the weighted sum of other neurons' activity passes a threshold[14].
Let x ∈ X = {0, 1}M be a neuron's firing activity for M data points. For an
agent v with a bias bv and a weight w for every input link,
xj = H bv + Xi∈T v
wixi!
j ∈ F v
(4.2)
where H is the Heaviside step function. To use an artificial neural network,
one sets the firing activity of a subset of neurons (called the input neurons),
waits for changes in firing activity to propagate through the network and,
finally, reads the firing activity of another subset of neurons (the output
neurons). The mapping between input and output for the network is thus
determined by:
• the graph
• the weights of connections between neurons
• the neuron biases
• the sequence in which neurons update their firing activity
Researchers have studied many, many ways to connect the neurons, sequence
their updates and train the weights and biases. The most popular schemes
find ways to serialize the process (making the computation easier) and keep
it predictable.
One such scheme is the multilayer perceptron with "vanilla" backprop-
agation. This scheme is a two-way cognitive network in which the neurons
are agents. In a multilayer perceptron, the neurons are arranged in K layers,
where Lk ⊂ V for each k ∈ {1, . . . , K}. The input neurons L1 are the entry
32
vertices. To fit my model, an exit vertex v′ must be added to the perceptron;
that exit vertex has one arc from each neuron in the output layer LK.
V = [k≤K
E = [k<K
Lk! ∪ {v′}
{(u, v) : u ∈ Lk, v ∈ Lk+1}! ∪ {(u, v′) : u ∈ LK}
(4.3)
(4.4)
Each neuron has an arc from every neuron in the previous layer and an arc
to every neuron in the next layer.
Since the neurons have no other connections, the firing activity for
each data pattern can be computed layer-by-layer with increasing k, and
each neuron need only update once for the forward aspect of the propagation
dynamics to converge. This layered graph also allows an error gradient to
be computed for each weight and bias and data point, using the chain rule
to propagate partial derivatives backward from the outputs[21]. Of course,
the activation function must be softened to a sigmoid rather than a hard
threshold, and each signal x ∈ X becomes a vector of M reals. This softening
does not affect the network's computational ability[10]. In the equations for
fv (v 6= v′ and v /∈ L1) below, S(x) is the logistic function. Equation 4.5
computes firing activity and Equation 4.6 computes gradients.
xj = S bv +Xi∈T v
yi = wi ·
wixi!
dx bv +Xi∈T v
dS
wixi! · Xj∈F v
j ∈ F v
(4.5)
yj
i ∈ T v
(4.6)
Just like with forward propagation, each neuron need only compute partial
derivatives once. The state space of each hidden neuron Fv (again, v 6= v′
and v /∈ L1) corresponds one-to-one with RT v+1 - the space of its input
weights and its bias. The firing activity of each input neuron v ∈ L1 is set by
the environment and does not change. The gradients from v′ are computed
by comparing each xi, i ∈ T v′ to a desired firing activity xd
i , i ∈ T v′ from the
training set. They are then backpropagated using Equation 4.6.
yi = xd
i − xi
i ∈ T v′
(4.7)
33
This is the gradient of a squared error function:
Err =
y2
i
1
2 Xi∈T v′
(4.8)
The total propagation dynamics Γ1 : FV × X E
7→ X E consist of
evaluating Equation 4.5 for each neuron in L1 and then L2 and so on until
LK, at which point v′ computes error gradients using Equation 4.8, each
neuron in LK computes error gradients from Equation 4.6, LK−1 does the
same, and so on until L2. The input layer L1 has no need of error gradients.
V : FV 7→ FV consist of each neuron greedily
minimizing error by descending the error gradient in (ostensibly) small steps.
The gradient for a bias bv is:
The pattern dynamics ψ1
∂Err
∂bv
=
dS
dx bv + Xi∈T v
wixi! · Xj∈F v
yj
The gradient for a weight we, e ∈ T v is:
∂Err
∂we
=
∂Err
∂bv
· ye
(4.9)
(4.10)
Here is the objective function for each neuron. Again, there are M
data points, and m is an index that selects a data point.
v 6= v′ and v /∈ L1
(4.11)
qv = − Xm≤M (cid:12)(cid:12)(cid:12)(cid:12)
∂Err
∂bv (cid:12)(cid:12)(cid:12)(cid:12)m
+ Xi∈T v(cid:12)(cid:12)(cid:12)(cid:12)
∂Err
∂we (cid:12)(cid:12)(cid:12)(cid:12)m!
The change in each weight or bias is the product of some step size η > 0 and
the negation of the gradient for that parameter. Unfortunately the gradients
generally become unbounded as they propagate backwards - due to repeated
multiplication by numbers potentially greater than one - so there exists no
Lipschitz constant L that limits the rate of change of each fv. Because the
propagation dynamics are discrete, this violation of Equation 3.1 threatens
the separation between propagation and pattern dynamics. This violation is
usually refered to as the "exploding gradients" problem.
If a weight increases as the pattern dynamics progress, the receiving
neuron becomes more sensitive to the firing of the sending neuron. Since
34
the weight increased in order to decrease the receiver's error, weights of the
receiver's outgoing arcs may increase in turn, as the receiver's firing activity
has become more valuable. At the same time, the sender becomes more
sensitive to gradients sent back from the receiver. The sender may henceforth
adjust its activity to be of more use the receiver. These relationships are often
crudely summarized as "Neurons that fire together, wire together." If these
relationships hold along a walk from the input layer to the output layer, the
corresponding round trip - forward over one graph and backward over the
other - constitutes a feedback loop. A feedback loop becomes a pattern when
the gradients along the loop become zero, which happens at a local minimum
in qv.
Feedback relations in a neural network are difficult to formulate. For
e iff x′
backward signals, y ≺ y′ ⇐⇒ Pm≤M y < Pm≤M y′. For the forward
signals in a link e ∈ E, xe ≺ x′
e decreases the gradient ye relative to
xe. Unfortunately, there is no way express the forward ordering using only
micro-level information. The effect of a change in firing activity must be
propagated all the way to v′ and back, so the forward ordering would be
recursive and span entire layers of the network. In simple terms: no neuron
actually knows what it wants;
it only sees a gradient that points it in a
(hopefully) good direction through Fv. Thus there is no closed-form way to
verify Equation 3.4 for a multilayer perceptron. If there was, I conjecture
that the proof would start from Equation 4.12 - a version of Equation 3.4
modified for two-way cognitive networks like this:
EC , yF v\y′
EC , xF v\x′
EC , yT v\y′
≥ qv(cid:0)xT v, yF v, xF v\x′
EC(cid:1) − qv(cid:0)xT v\x′
EC , yT v\y′
EC , yF v\y′
EC , xF v, yT v(cid:1)
EC(cid:1) − qv (xT v, yF v, xF v, yT v)
qv(cid:0)xT v\x′
(4.12)
In English, this equation states that the neuron v's utility becomes more
sensitive to the output signals xF v∩EC and yT v∩EC as the input signals xT v∩EC
and yF v∩EC strengthen.
I am not yet sure that Equation 4.12 is properly
formulated. Such a formulation would be core to a formal theory of two-way
cognitive networks.
Not all walks from input to output are feedback loops. A marginal
improvement may not be enough to make the receiver (the one described
above) more valuable than the other neurons in its layer. And the sender's
increased sensitivity to the receiver's gradients may not surpass the sender's
sensitivity to the gradients of other neurons in the receiver's layer. To identify
the selection of closed walks that undergo positive feedback, Equations 4.5,
35
4.6 and 4.7 should be plugged into a feedback condition like Equation 4.12. Of
course, it will be necessary to modify Equations 4.6 and 4.7 so that sensitivity
of the backward signals to the forward signals in a link (on pattern timescales)
could be expressed in closed form. This method of analyzing neural learning
rules could prove quite useful in future work.
Neurons train without any knowledge of the error itself. They only
see the error gradient. By descending that gradient only in small steps, the
weights and biases try to be quasi-static on signaling timescales. Unfortu-
nately, the network runs into trouble here. The gradients in a given layer are
roughly proportional to the weights in layers closer to the exit. Positive feed-
back loops can cause the gradients to increase exponentially - to "explode"
- as the backpropagate. Because the steps taken by the weights and biases
are proportional to the gradient, the pattern dynamics can violate their own
Definition 2.2 and thus the separation between propagation and pattern.
Just like positive feedback leads to exploding gradients, negative feed-
back can lead to "fading" gradients. As I warned at the end of Section
2.3, overly-aggressive contraction can prevent signals from spreading prop-
erly across the network. Also, small gradients lead to only small changes in
weights and biases, eventually halting the pattern dynamics.
Multilayer perceptrons largely try to avoid problems with exploding
and fading gradients by keeping the graph feedforward with few layers. If
each component of the gradient only makes two or three jumps backward,
it can't grow or shrink much. Gradient descent causes more problems in
networks with many layers, like deep belief networks. The problem is worst
in recurrent networks, which have directed cycles. A signal in a recurrent
network may make an infinite number of jumps. If the gradients explode,
then the propagation dynamics of such a network will never converge, and
without convergence, the aggregation p does not exist. A gentle fading will
make the propagation dynamics contract, but excessive fading still prevents
signal propagation.
In contrast, the forward signals do contract. Although I will not
prove this statement, Hopfield provided good inductive evidence[9] that the
propagation of firing activity mostly converges. Intuitively, a neuron rejects
any changes in its input that do not push it over its threshold - from inactive
to firing or vice versa.
36
4.3 Market Economies
Most naturally-occuring cognitive networks are composed of, inter-
twined with and assemble into other cognitive networks. Modeling such
systems is tricky and intellectually risky. The modeler must pick and choose
what to represent, and any omissions are bound to be criticized by those
with different priorities. In this section, I run a greater risk than most: I am
not an economist. Remember that my purpose is to explain cognition, not
the finer points of economics. The broad pattern of an economy will suffice.
Among the many contexts in which it can be understood, I will model the
specialization of market niches and the relationships between them. I will
try to describe how niches develop and coordinate to best acquire, process
and allocate resources.
Every model defines a context within the heterarchy of cognitive net-
works - the heterarchy that extends from individual cells to the biosphere.
A theory of cognition can guide the creation of models and render them in a
common language, enabling us to draw clearer lines between contexts and to
avoid using a model outside its context. The content of signals in the model
circumscribes what it can cognize, and the choice of agents determines the
model's granularity. A consequence of acausality is that even a good model
may not predict which patterns emerge in its referent. That does not, how-
ever, mean the model is useless. The model might still predict what kinds of
pattern can emerge, as well as the stability or coherence of patterns which
have already emerged. In other words, the model might still predict what
can be known or is likely to be known.
Each niche is an agent characterized by:
• a product
• the producers of that product
• the method of production
The market economy is a two-way cognitive network. Goods and services are
the forward signals while money and preferences are the backward signals.
By transforming goods and services into other goods and services, niches
(as agents) collectively bridge the available resources to consumer demand.
37
Those available resources are the entry vertices. Consumers are the exit
vertices.
The propagation dynamics are the process by which the market works
out the price of each good and the volume bought by each receiving niche.
The niches are quasi-static, which is to say the products, sets of producers and
the methods of production are all fixed on short timescales. This limits each
niche's flexibility. Say, for instance, that the demand for a good increases.
At some point the increase will test the capacity of the producing niche. The
price will rise, decreasing demand. Conversely, a drop in demand will cause a
drop in price, which raises demand. This negative feedback - driven by profit
and limits on flexibility - ensures that the propagation dynamics converge.
The pattern dynamics are the adaptation of goods to fit buyer pref-
erences and the adaptation of the means of production to fit the available
resources. Producers can leave or join niches, or they may invest in or liq-
uidate means of production. There are two kinds of positive feedback. The
first - analagous to autocatalytic loops - occurs when a supply chain produces
some of its own means of production. The second comes from economies of
scale and is analagous to the feedback loops in a multi-layer perceptron. For
many goods, the more that gets produced, the easier it is to produce. The
easier a good is to produce, the more profit that can be made, which draws
more people and companies to produce that good. This kind of feedback loop
is a supply chain from raw material to consumer. Like the feedback loops in
an ANN, the backward walk mirrors the forward walk.
In general, goods in a supply chain have uses both inside and outside
the supply chain. As positive feedback progresses, those goods will become
more suited to their use inside the supply chain as opposed to outside it. If
the graph can change over time, niches will split and differentiate to take
advantage of their own economies of scale.2
In this way, the supply chain
may become an industry.
Like any model, this one leaves a lot of interesting behavior out of
context. Here is a short, non-exhaustive list.
• Modeling niches as agents ignores the differences between large com-
panies, small businesses and individuals.
2I have not yet addressed changes to the graph, but such a study would make for
fascinating future work.
38
• Modeling the physical environment as the entry ignores environmental
damage.
• Treating the consumer as the exit node omits anything that alters con-
sumer demand, like culture or advertising.
I have also left out the precise definition of a good. The simplest definition is
the set of uses for the good. To model uncertainty - e.g. risk or asymmetries
in information - I could make every signal a probability distribution over
the power set of possible uses for a good.3 That still leaves out temporal
behavior, like borrowing money or training the workforce. To model those,
each good would become a function of its uses over a cycle in time.
That should suffice for an example of cognition. The rest, I will leave
to economists.
4.4 Ant Colony Optimization
Ant colony optimization (ACO) is a class of algorithms for finding a
"short" path from one vertex in a graph to another[4]. It may seem very
specialized, but many combinatoric optimization problems can be expressed
as such a path search. ACO works by imitating the method by which ants
settle on a path between their colony and a food source. Some number of ants
start at the colony and wander through a graph of waypoints, guided toward
the food by smell (a search heuristic). After they reach the food, they retrace
their steps back to the colony, laying pheremone uniformly along their path.
Each ant has a fixed amount of pheremone to lay, so it will lay pheremone
more densely on shorter paths. Ants that walk the graph thereafter will favor
arcs4 with more pheremone. When these ants lay their pheremones in turn,
they will tend to lay it on these arcs, which will draw more ants. As time
goes on, the ants settle on a single short path through the graph, and this
path maps to a solution of the combinatoric optimization problem.
An ACO is a two-way cognitive network in which the waypoints are the
agents. An agent also includes the pheremone concentrations of its outgoing
arcs. The location of the colony is the entry vertex, and the location of the
food source is the exit vertex. The forward signal in an arc is a probability
3Similar to Eigen and Schuster's treatment of genetic risk[5].
4or edges, in the undirected case
39
distribution over all the forward walks an ant could take from the entry vertex
through that arc (before reaching the exit). The backward signal is the same
but from the exit vertex and over the backward graph. These walks are not
uniformly random; ants are guided first by the search heuristic and then by
pheremone concentrations.5 The heuristic - the smell of food - has a fixed
value for each arc. When a single ant walks randomly across the graph, it
samples a probability distribution for each arc in its path.
The propagation dynamics consist of ants walking across the network
while the pheremone concentrations are quasi-static. That condition can
be achieved by keeping each ant's pheremone payload small compared to
the ants' ability to smell. Most implementations pass a population of ants
through the network before laying all their pheremones at once, but this
is more a matter of convenience than functionality. More important is the
impracticality of covering the entire support of every path distribution. We
cannot say the propagation dynamics have converged until individual ants
cease to shift the colony's estimate of these distributions. But every path
would need to be traversed by at least one ant, and the point of ACO is
to be faster than simply trying all the paths. That contradiction sounds
catastrophic, yet ACO works. Why?
It works because ACO need not maintain perfect separation between
propagation and pattern dynamics. Relaxing separation makes the pattern
dynamics stochastic with an increasing variance. That stochasticity doesn't
cripple the algorithm because there are usually many good paths. ACO does
try to improve separation by keeping the number of likely paths small. First,
it prohibits ants from revisiting a vertex before reaching the exit. Second,
the search heuristic attracts the ants to a smaller set of paths, concentrating
their probability mass on a subset of its support. Finally, positive feedback
helps minimize the number of likely paths. As time proceeds, a subset of
paths comes to dominate the rest, and then a subset of that subset, and so
on. This serves to shrink the effective support of the path distributions. The
result of these steps is an algorithm that reaches a good solution quickly. For
very hard problems, this is often better than taking a long time to find the
best solution.
The pattern dynamics are the process by which pheremone is laid.
5Some instances of ACO don't use a heuristic. Before pheremones are laid, the ants in
those colonies do make uniformly random pathing choices.
40
Having already described the process from the ants' perspective, here is the
waypoints' perspective as agents. An agent maximizes the number of ants
on the outgoing arc with the most ants.6 If an agent increases the number
of ants it sends along the (expected) shortest path, the expected path length
of its ants will decrease. That gives the waypoints from which it receives
ants reason to favor it, because they have the same incentive. By this rea-
soning, the shortest walk through the graph should coincide with the sole
feedback loop. The catch is that pheremone concentration is similar but not
identical to the inverse of expected path length. At a certain point in the
feedback, when pheremones have become concentrated enough, the ants will
choose well-worn paths simply because they are well-worn. Like with many
stochastic optimization algorithms, one must balance speed of convergence
with the quality of the solution.
Pheremones evaporate at a rate proportional to their concentration.
Even if all the ants choose the same path, eventually the rate at which those
ants lay pheremone equals the evaporation rate. The evaporation rate is not
necessary for a feedback loop to become a pattern - that loop stabilized as
soon as its pheremone levels got high enough that nearly all the ants chose
it. Evaporation is necessary if one wishes ants to explore after a pattern
emerges. Such exploration is only useful if the algorithm is elitist7; otherwise
the other ants will ignore any new paths an explorer finds. Elitism involves
computation external to the network, so it has no analog in a cognitive
network.
Unlike the other exemplars I've introduced, ant colonies primarily
cognize themselves. Ants enter the network without external information.
Instead, the information encoded in a pattern reflects only the graph, the
lengths of the arcs, the search heuristic and the random choices of the earliest
ants.
Waypoints are not usually seen as agents; the ants are. These ant-
agents come to an explicit consensus on the best path. ACO occupies a
curious middle ground between cognitive networks and consensus algorithms.
This dual interpretation possible only because feedback in ACO is winner-
take-all. Seeing it as a consensus algorithm, each agent must know the whole
pattern. Seeing it as a cognitive network, the set of possible patterns is very
6This does not mean sending all ants over the same arc as soon as that arc becomes
popular; such a shift would be a discontinuous jump through the agent's state space.
7An elitist algorithm always selects and remembers the best solution found so far.
41
restricted. These two apparent problems seem to offset each other.
42
Chapter 5
Conclusions
I have now outlined a new formal theory of cognition. It is does not
focus on humans, on artificial intelligence or on any other specific domain.
Instead, it lays out the abstract process by which a complex network acquires
knowledge. My theory is still limited to fixed graphs1 with two levels - micro
and macro. A complete theory should describe the dynamics of the graph
and explain how an arbitrary number of levels interact when arranged in a
heterarchy.
5.1 Context in Cognition
A cognitive system is a network of agents. These agents signal
each other on relatively short timescales and change their state over longer
timescales. For every configuration of agent states, the signal propagation
dynamics find a stable fixed point, called the aggregate. Since signals prop-
agate faster than the agents move, the propagation dynamics closely track
this fixed point, which means the agents react only to the aggregate.
In
the pattern dynamics, each agent maximizes a utility function of its input
and output signals. This behavior may lead subsets of agents coordinate
across closed walks through the graph by reinforcing their signals to each
other. A stronger signal over any arc in the walk causes the receiving agent
to strengthen the signal it sends over the next arc. Eventually this wave
1where the signals or agents states move but not the vertex or arc sets
43
of influence returns to strengthen the first signal further, and the closed
walk becomes a feedback loop. If the signals in a feedback loop stablize, the
feedback loop becomes a pattern. The selection of closed walks that undergo
feedback and then stabilize into patterns, together with the nature of the
signals in the pattern, encode the network's knowledge. Despite being ad
hoc and of arbitrary scale, these encodings are coherent.
So far, I have mostly avoided the use of "complex" and "simple"
as adjectives, because they are only useful in a relative sense. Recall the
(biological) ant colony I sketched on page one.
If I modeled the colony,
I would probably say that each ant is "simple". Yet an ant is incredibly
complex in its own right. Its brain is a cognitive network of cells, and those
cells are cognitive networks of molecules. Retreating to the perspective of an
ecosystem, the colony itself looks like a simple agent. Every cognitive system
operates in some context or set of contexts.
A context is defined by a σ-algebra on a set of things that can
be known in that context.
A context can be very broad but never universal. Billy Vaughn Koen
would say that all knowledge is heuristic, in that it "provides a plausible
aid or direction in the solution of a problem but is in the final analysis
unjustified, incapable of justification, and potentially fallible."[12] Since any
unit of knowledge is part of some contexts but not others, it follows that
objective knowledge is a contradiction. It also follows that knowledge can
only be measured relative to a given context, with that context being the
domain of the measure. Since no context exists which includes all possible
knowledge, no objective measure of knowledge exists.
Similar reasoning applies to measures of "complexity". As a field,
we have been asking how complex behavior arises from simple parts. But
without an objective notion of complexity, this question loses its meaning.
I will reframe it. By saying that an agent is simple, I am saying that its
context includes the signals it receives (and sends) but not the whole network.
Likewise, the network sees an agent's output signals but not the potentially-
complex arg max problem an agent solves internally. Thus an agent is both
simple and complex relative to the whole network, and the whole network
is both simple and complex with respect to each agent. Every cognitive
process implies one or more contexts. Each context is a σ-algebra of things
the cognitive system can know.
44
Figure 5.1: On the left: A single monolithic context divided into micro con-
texts. In the center: Overlapping micro (thin) and macro (bold) contexts
in a cognitive network. Parts of each context are shared, and others are
hidden. On the right: The micro contexts have subsumed the macro con-
text. This could involve a centralized aggregation procedure (as opposed to
a propagation dynamics) or an algorithm for the explicit consensus of agents.
The paradox of "complex behavior from simple parts" comes from
thinking about a cognitive network in only one context at a time. By consid-
ering agents in their own micro-level contexts, I have reframed our question:
How do the micro and macro contexts of a cognitive network relate to each
other? The macro context consists of feedback loops and patterns, and each
micro context consists of an individual agent's utility maximization problem.
My theory links these contexts by means of the aggregation (recall Definition
2.4) and of complementarity between agent utility functions. This framework
begins to explain - formally - how a consciousness relates to the substrate
from which it emerges.
The field of cybernetics tried to explain cognition qualitatively. Mat-
urana and Varela, for example, introduced the concept of autopoiesis, or
self-creation, in living systems[20]. Douglas Hofstadter introduced strange
loops, which are configurations of symbols that take on circular meaning,
thereby becoming self-aware[8]. Unfortunately, cybernetics researchers were
only able to express their ideas with words rather than with mathematics.
While their work fascinated many, none could agree on what it meant or
how it should be applied. We need formality if we are to reliably convey our
understanding of cognition from mind to mind.
Most other work on cognition falls into one of two categories. The
first is a top-down approach that focuses more on what a cognitive system
does than how it cognizes. Psychology falls into this category, and so does
45
the study of symbolic artificial intelligence. This approach struggles with
context. Psychological studies are difficult to control because human thought
takes place in many contexts, and those contexts are very hard to ascertain
or control. A well-controlled study may yield clear quantitative results, but
those results only apply in the narrow context created for the study.
In
artificial intelligence, Alan Newell[16] framed cognition as a rational, goal-
oriented search through a problem space. But a well-posed search calls for
a well-defined goal, which cannot be defined a priori for a whole cognitive
network. And rationality - being an application of knowledge - is just as
context-dependent as the knowledge itself. Unfortunately, the context for all
known artificial intelligences must either be specified by a human designer or
generated by an algorithm written by a human designer. The context(s) of
such a system will always be circumscribed, and the system will always face
an intractable and ill-defined meta-problem - figuring out when its knowledge
is or is not useful. A top-down approach requires a clear context that doesn't
often exist.
The second category is characterized by a bottom-up approach which
seeks the micro-level mechanics of cognition. This kind of cognitive architec-
ture adapts to the context it finds itself in. Artifical neural networks are a
good example, as are Pentti Kanerva's sparse distributed memories[11]. Neu-
roscience searches for the mechanics of human (or animal) cognition. Like
cybernetics, work in this category asks how something learns. But where
cybernetics tends to be informal and abstract, bottom-up approaches are
formal and concrete. Every cognitive architecture modeled with a bottom-
up approach either arose by itself or explicitly mimicks a system that did.
Until now, no one has been able extract or abstract the essence of cognition
for novel use in other systems.
My goal has been a formal, abstract theory of cognition that deals
properly with context. Like the bottom-up approach, I have sought formal
mechanics of cognition. Like cybernetics, I have tried to keep those mechan-
ics abstract. At the end of Section 3.1, I pointed out that the usual tradeoffs
between agents and the macro-level - a) agent simplicity versus the process-
ing of macro-level information and b) agent simplicity versus the storage
of macro-level information - could be two aspects of the same fundamental
paradox. That paradox is how complex behavior arises from simple parts.
I resolved it by finding a better question: Where do the micro and macro
contexts overlap, and where do they not? By treating signal propagation as
46
a dynamical system unto itself, the micro-level contexts aggregate the macro
context without subsuming it. By defining pattern in terms of closed walks
of complementarity between agents, the micro-level contexts coordinate to
store information in the macro context while remaining distinct from the
macro context.
5.2 Future Work
Almost every aspect of my theory could be developed further.
Regarding aggregation, I have not yet explained what information
propagates and what doesn't. Abstract algebra may provide the proper tools
to model and control the spread of signals, given how they join and split at
each agent. If signals can be measured, ergodic theory may also be useful.
Two-way networks deserve a formal definition and analysis. How do
forwards and backward signals interact? What, precisely, constitutes a feed-
back loop through a pair of mirrored graphs? Such work will allow us to
apply the mathematics of cognition to a much wider range of systems.
The topology of the pattern dynamics is still fairly mysterious. How
robust are patterns, and what kinds of robustness are there? Can patterns
merge or split, and how do these events affect their coherence? Can parts of
a coherent pattern survive if the pattern loses stability? These questions are
fundamental to the design of cognitive networks.
Even though most natural cognitive networks are related heterarchi-
cally, my theory analyzes only two levels of behavior at a time - a micro-level
and a macro-level. Can a formal definition of context allows us to model the
relationships between many contexts? An answer to this question is vital for
modeling natural systems.
How does the graph of a cognitive network change over time without
interfering with separation or breaking feedback loops? In particular, how
can an agent create new arcs when its context is limited to the arcs it already
has? This question is very pertinent to artificial neural networks as well as
to design automation.
Finally, there's a lot of application work to do. Cognition can help
design new metaheuristics for optimization, and it might provide a better
understanding of the ones we already have - genetic algorithms, ACO and
47
particle swarms in particular. New ways to train and structure artificial
neural networks should present themselves. And finally, it should be possible
to improve the distributed control of large industrial systems like smart grids
and reconfigurable manufacturing systems.
Many of these applications will inevitably inform the more abstract
inquiries, just like the systems in Chapter 4 guided me to the theory in
Chapters 2 and 3. The social and biological sciences could be a fertile ground
for new quantitative models, although we will probably need a formal theory
of context in order to comprehensively model market economies or ecologies.
A better understanding of context could likewise provide a formal language
for comparing and contrasting cognitive architectures.
Curiosity aside, we should be cautious about modeling social systems.
As we study them, we must remember that social theories can affect their
referents if or when they become public knowledge. These observer effects
can be unpredictable or even dangerous. Given the training necessary to fully
understand theories of cognition, these theories are unlikely to find their way
(intact) into the public consciousness. Rogue, poorly applied social theories
have in the past created much upheaval.
48
References
[1]
I. A. Basheer and M. Hajmeer. "Artificial neural networks: fundamen-
tals, computing, design and application". In: Journal of Microbiological
Methods 43 (2000), pp. 3–31.
[2] Adrian Bejan and Sylvie Lorente. "Constructal theory of generation
of configuration in nature and engineering". In: Journal of Applied
Physics 100.041301 (Aug. 2006).
[3] Tom De Wolf and Tom Holvoet. "Emergence Versus Self-organisation:
Different Concepts but Promising When Combined". In: Engineering
Self-Organising Systems. Springer Berlin Heidelberg, 2005, pp. 1–15.
[4] Marco Dorigo and Christian Blum. "Ant colony optimization theory:
A survey". In: Theoretical Computer Science 344 (Nov. 2005), pp. 243–
278.
[5] Manfred Eigen and Peter Schuster. "The Hypercycle: A Principle of
Natural Self-Organization, Part A: Emergence of the Hypercycle". In:
Die Naturwissenschaften 64 (1977), pp. 541–565.
[6] Drew Fudenberg and Jean Tirole. Game Theory. The MIT Press, 1991.
[7] Friedrich A. Hayek. "The Use of Knowledge in Society". In: American
Economic Review 35.4 (Sept. 1945), pp. 519–30.
[8] Douglas Hofstadter. Godel, Escher, Bach: An Eternal Golden Braid.
Basic Books, 1979.
[9] John J. Hopfield. "Neural network and physical systems with emer-
gent collective computational abilities". In: Proceedings of the National
Academy of Sciences of the United States of America 79.8 (Apr. 1982),
pp. 2554–2558.
49
[10] John J. Hopfield. "Neurons with graded response have collective com-
putational properties like those of two-state neurons". In: Proceedings
of the National Academy of Sciences of the United States of America
81.10 (May 1984), pp. 3088–3092.
[11] Pentti Kanerva. Sparse Distributed Memory. The MIT Press, Cam-
bridge, MA, 1988. isbn: 0-262-11132-2.
[12] Billy Vaughn Koen. Discussion of the Method: Conducting the Engi-
neer's Approach to Problem Solving. Oxford University Press, 2003.
[13] Steven M. Manson. "Simplifying complexity: a review of complexity
theory". In: Geoforum 32 (2001), pp. 405–414.
[14] Warren S. McCulloch and Walter H. Pitts. "A Logical Calculus of the
Ideas Immanent in Nervous Activity". In: Bulletin of Mathematical
Biophysics 5.4 (Dec. 1943), pp. 115–133.
[15] R. Milo et al. "Network Motifs: Simple Building Blocks of Complex
Networks". In: Science 298 (Oct. 2002), pp. 824–827.
[16] Alan Newell. Unified Theories of Cognition. Harvard University Press,
1990.
[17] Reza Olfati-Saber, Alex J. Fax, and Richard M. Murray. "Consensus
and Cooperation in Networked Multi-Agent Systems". In: Proceedings
of the IEEE 95.1 (Jan. 2007), pp. 215–233.
[18]
Ilya Prigogine and Gregoire Nicolis. "On Symmetry-Breaking Instabil-
ities in Dissipative Systems". In: The Journal of Chemical Physics 46.9
(May 1967), pp. 3542–3550.
[19] Alan Turing. "The Chemical Basis of Morphogenesis". In: Philosoph-
ical Transactions of the Royal Society of London, Series B, Biological
Sciences 237.641 (Aug. 1952), pp. 37–72.
[20] F.G. Varela, H.R. Maturana, and R. Uribe. "Autopoiesis: The Organi-
zation of Living Systems, Its Characterization and a Model". In: Cur-
rents in Modern Biology 5.4 (1974), pp. 187–196.
[21] Paul J. Werbos. "Beyond regression: New tools for prediction and anal-
ysis in the behavioral sciences". PhD thesis. Harvard University, 1974.
50
|
1212.6298 | 1 | 1212 | 2012-12-27T04:07:01 | Design of Intelligent Agents Based System for Commodity Market Simulation with JADE | [
"cs.MA",
"cs.AI"
] | A market of potato commodity for industry scale usage is engaging several types of actors. They are farmers, middlemen, and industries. A multi-agent system has been built to simulate these actors into agent entities, based on manually given parameters within a simulation scenario file. Each type of agents has its own fuzzy logic representing actual actors' knowledge, to be used to interpreting values and take appropriated decision of it while on simulation. The system will simulate market activities with programmed behaviors then produce the results as spreadsheet and chart graph files. These results consist of each agent's yearly finance and commodity data. The system will also predict each of next value from these outputs. | cs.MA | cs | Design of Intelligent Agents Based System
for Commodity Market Simulation with JADE
R.Refianti 1, A.B. Mutiara2, H. Gunawan3
Faculty of Computer Science and Information Technology,
Gunadarma University, Indonesia
E-mail: 1,2{rina,amutiara}@staff.gunadarma.ac.id, 3 [email protected]
Abstract
A market of potato commodity for industry scale usage is engaging several types of
actors. They are farmers, middlemen, and industries. A multi-agent system has been built to
simulate these actors into agent entities, based on manually given parameters within a
simulation scenario file. Each type of agents has its own fuzzy logic representing actual
actors' knowledge, to be used to interpreting values and take appropriated decision of it
while on simulation. The system will simulate market activities with programmed
behaviors then produce the results as spreadsheet and chart graph files. These results
consist of each agent's yearly finance and commodity data. The system will also predict
each of next value from these outputs.
Keywords: Agent, JADE, Java, Fuzzy logic, Back propagation, Potato
1. Introduction
In potato commodity market, industries and farmers, as end buyer and raw producer respectively, hold
big role in market activities that can give impact to each other. Industries need the farmers to fulfill
their raw commodity requirement for continuous production. For industries, their production is vital
activities to gain profit. At other side, farmers need to keep producing and sell their harvest revenue to
have income. However, both farmers and industries are not the only one “seller” and “buyer” in
market. Middlemen hold those both two roles, positioning themselves a competitor to farmers and
industries and gain profit from it.
In market competition, each actor's assets, knowledge, and behaviors can be different to others
thus giving different measures on the same situation. A simulation system will help by simulating the
market activities and the produce the data to see if their current conditions can give positive impact in
achieving their goal.
Agent Oriented Programming (AOP) is one of the best approaches to declare actual actors as
system entities, called agent, and simulate their actions. AOP offers several advantages like message
based communication, multi-behaviors support, life cycle management, and more [2]. The capability of
agents can be improved by posing artificial intelligent backup to help them sense conditions of
environment where they live and take appropriated actions regarding to it. Thus this type of agents are
called intelligent agent [1]. Users have to supply actors' knowledge and scenario to the system. The
system will be limited to give output data as it is, without any conclusion.
2. Methodology
The methodology that is used consists four phases in order [2]. Each phase has multiple steps inside
them to be done in order too. The flow, phases, and steps of the methodology are illustrated in Fig. 1.
Figure 1. Methodology diagram [2]
This methodology is used corresponding to JADE (Java Agent Development) library which is
used to built system's architecture. JADE is Java based library with ACL (Agent Communication
Language) that is defined by FIPA (Foundation for Intelligent Physical Agent). In general, this ACL
declares the messages performative and service type “yellow paging” as basic way for agents to
communicate.
The methodology serves as a guide for the system designer when developing a system. In
general, a software development methodology may comprise of [2]:
A process, i.e. a sequence of phases and steps that guide the developer in building the system.
A set of heuristic rules that support the developer in making relevant choices.
A number of artifacts, i.e. diagrams, schemas or documents representing in graphical or textual
form one or more models of the system.
A suitable notation to be used in the artifacts.
A set of patterns that can be applied to solve common situations.
One or more tools that: automate, as much as possible, the phases and step specified in the
process; force consistency between the models produced; highlight problems arising from
incorrect design choices, when possible; generate code and documentation, etc.
The focus of the methodology is on the process and the artifacts that are produced. The
described process covers the analysis phase and the design phase and is shown in Fig.1. The analysis
phase is general in nature and independent of the adopted platform. Conversely, the design phase
specifically assumes JADE as the implementation platform and focuses directly on the classes and
concepts provided by JADE. Observing Fig.1, it can be seen that there is no strict boundary between
the analysis and design phases. Moreover, the methodology is of an iterative nature, thus allowing the
designer to move back and forth between the analysis and design phases and the steps therein.
At the end of the design phase, the developer should be able to progress straight to the
implementation, which is where the actual coding occurs. In addition, most of this phase can probably
be carried out by means of a proper tool which automates the implementation process. The planning
stage, like implementation and testing, is not formally addressed in the methodology. However, for the
sake of the methodology, a question is included (see Fig.1), which initially asks if the designer has
made a rational decision on whether to use an agent-based solution. If the answer is yes, the designer
moves on the analysis, while if the answer is no, the designer should seek an alternative solution.
3. Results and Discussion
3.1. Planning Phase
The system uses intelligent agent and Fuzzy logic is chosen as knowledge implementation since the
actual actors may interpret some values into natural language. The fuzzy rules are written as FCL
(Fuzzy Control Language) script and applied with help of jFuzzyLogic library and its built-in fuzzy
inference system. The required input is simulation scenario describing some parameters including FCL
scripts' location in form of spreadsheet file. At start, system will ask user to select this file in order to
run. Finally, each agent will produce their finance and commodity data at the end of simulation, both in
form of spreadsheet files (tabular) and picture files (charts) by using jExcel and jFreeChart libraries.
3.2. Analysis Phase
The following Use case diagrams are showing relation between farmer-middlemen, farmer-industries,
and middlemen-industries respectively.
Figure 2. Use Case Diagram between Farmers and Middlemen
Figure 3. Use Case Diagram between Farmers and Industries
Figure 4. Use Case Diagram between Middlemen and Industries
Actual actors are represented by three kind of agents based on service they provide. Farmers,
middlemen, and industries are “ProdusenAgent”, “DistributorAgent”, and “KonsumenAgent” agents
respectively. Since the system needs to access the external resource scenario and FCL files, a
transducer agent that not representing actual actors “MainAgent” is created. Fig. 2 shows relation
between these agents kind.
Table 1 below shows those agents' tasks.
Figure 5. Agents relation
Table 1: Tasks of agents
Agent Type
ProdusenAgent
DistributorAgent
KonsumenAgent
MainAgent
Tasks
Plant seed to field
Harvest
Respond to commodity request
Respond to pledge deal
Offer own commodity
Search highest price for commodity
Give harvest revenue to pledge dealer
Sort commodity according to its standard
Respond to commodity offers
Respond to price request
Respond to commodity giving
Respond to commodity request
Offer own commodity
Perform monthly production
Sort commodity according to its standard
Respond to commodity offers
Respond to price request
Respond to commodity giving
Make pledge deal with ProdusenAgent
Request for commodity
Load scenario file
Parse scenario's parameters
Do the dating count
Synchronize date to all active agents
3.3. Design Phase
This phase offers chance to improve agents relation and interaction protocol in case FIPA doesn't cover
the need. After specify agents relation, a interaction specification table including messages template is
made as shown in Table 2. All capitalized word refers to messages performative.
Table 2:
Agent
ProdusenAgent
Role With When
R
D, K Get REQUEST message
IP
Offer
Interaction specification table
Interaction
Respond to commodity
request
Respond to pledge deal
Seed
R
K
Get PROPOSE message
Offer own commodity
Offer
I
D, K Every month change and has
stock
Search highest price for
Pledge
I
D, K Harvest time
Template
QUERY_IF
ACCEPT_PRO
POSAL,
REJECT_PROP
OSAL
INFORM,
INFORM_IF,
AGREE,
REFUSE
CFP
DistributorAgent
commodity
Give harvest revenue to
pledge dealer
Respond to commodity offers Offer
R
Pledge R
Respond to price request
Pledge
I
Respond to commodity
giving
Respond to commodity
request
Offer own commodity
Pledge R
Offer
Offer
R
I
P
P
P
K
K
D, K Deal on given price
CONFIRM
Get INFORM message
Get CFP message
Get CONFIRM message
INFORM_IF
REQUEST_W
HEN
DISCONFIRM
Get REQUEST message
QUERY_IF
Every month change and has
stock
INFORM,
INFORM_IF,
AGREE,
REFUSE
INFORM_IF
REQUEST_W
HEN
DISCONFIRM
PROPOSE,
AGREE,
REFUSE
REQUEST,
AGREE,
REFUSE
SUBSCRIBE
KonsumenAgent
Respond to commodity offer Offer
R
Pledge R
Respond to price request
P, D Get INFORM message
P
Get CFP message
Respond to commodity
giving
Make pledge deal with
ProdusenAgent
Pledge R
Seed
Request for commodity
Offer
MainAgent
Synchronize date to all active
agents
Detailed
Year
P
P
Get CONFIRM message
Every month
P, D Every month when stock
isn't enough to perform
production
Every month
P, D,
K
I
I
I
To interact with user, the system has its own graphical user interface (GUI) showing finance
and commodity charts. The GUI also shows some agent's actions log for debugging case and agent's
current status as described in stock, money, and field values. User gives the scenario input through
“MainAgent”. The scenario file has two sections of parameters. The first section as described on Table
3 is used to determine how the simulation will work globally. A scenario file only has one of these
sections.
Table 3: Global parameters
Description
Title
Specification block counts Number of how many specification blocks that need to load
Start year
Year of start
Simulation duration (year) How long the simulation
Harvest ratio (Revenue/seed) Number of harvest revenue (Kg) produced by 1 Kg seed
Number of seed (Kg) planted in 1 Ha field
Kg seed per Ha
Autonom (boolean)
Specify whether simulation will run automatically (1) or manually (0)
The second section is used to give parameters to agent. This section can appear multiple times
in one scenario file. The number of this section must be specified at the start of scenario, in other
words first global parameter. The section is described in Table 4.
Table 4: Agents parameters
Name
Type of agent
Description
Specify whether agent(s) has GUI (1) or not (0)
GUI
Name for agent(s)
Name
Service type of agent(s): “produsen” (0), “distributor” (2), “konsumen” (3)
Role
Number of stock at start
Stock at start (kg)
Number of money at start
Money at start (Rp)
Number of seed at start
Seed at start (Kg)
Lower value of diameter to pass selection
Minimum diameter (cm)
Upper value of diameter to pass selection
Maximum diameter (cm)
Number of commodity needed to perform one time production
Production usage (kg)
Price for 1 Kg of processed commodity
Production income (Rp/kg)
Price for 1 Kg of commodity when sold in traditional market
Market income (Rp/kg)
Starting price when buying
Normal buy price (Rp/Kg)
Normal pledge price (Rp/Kg) Starting price when pledging
Probability for harvest revenue number to randomly goes down
Harvest failure chance (%)
Field (Ha)
Owned field
Cost to plant in 1 Ha field
Plant cost (Rp/Ha)
File path relatively to system's location for FCL script
FCL path
Number of agents
Number of agent(s) using this specification
Each agent has some instance of JADE's Behavior class to describe an action of actual actor
into algorithm. These behaviors can be executed in parallel or sequence depends on the need. The lists
of behaviors representing actual actors’ actions are mentioned by Table 5.
Table 5: Agent’s actor representing behaviors
Behavior's
Represented Action
Name
InformingBev Offer own commodity
InformedBev
Respond to commodity offer
Request for commodity
QueryingBev
Respond to commodity request
QueriedBev
SellBev
Offer own commodity to respond request
Make pledge deal
PledgingBev
Respond to pledge deal
PledgedBev
GivingBev
Give commodity
Respond to commodity giving
GivedBev
Search highest price
DivergingBev
DivergedBev
Respond to price request
Beside those behaviors, there are some which aren't representing actual actors' actions. These
behaviors are related to agents' activities like calculate date and pool received message. They are listed
in Table 6.
ProdusenAgent, DistributorAgent
DistributorAgent, KonsumenAgent
KonsumenAgent
ProdusenAgent, DistributorAgent
ProdusenAgent, DistributorAgent
KonsumenAgent
ProdusenAgent
ProdusenAgent
DistributorAgent, KonsumenAgent
ProdusenAgent
DistributorAgent, KonsumenAgent
Table 6: Agent’s system related behaviors
Behavior's Name
CyclicBehavior
DelayBehavior
Purpose
Used
its
received message and check
to pool
performative then call appropriate behavior to handle it.
Used to update date by increasing month count after
some time. It'll be done as long as simulation running.
Type of agent
ProdusenAgent, DistributorAgent,
KonsumenAgent
MainAgent
For ontology, the system will use object serialization as its protocol. To do this, some classes
which will be content language like Price, Offer, and Seed have to implement Serializable interface.
Instance of these classes will be set to ACLMessage before sent. ACLMessage itself is JADE's class
that support performative and object use as content rather plain string.
3.4. Implementation and Testing Phase
A test is performed by using a dummy scenario and three kinds of knowledge bases, one for each type
of agents. The scenario specifies six block specification for six agents (one block per agent) and put
them in group of two. Therefore each type of agents has two agents instance and shares same
knowledge base. The complete list of scenario can be seen in Table 6. This table uses scenario's
dummy values for test. Note that this table doesn't represent how to write them on the actual scenario
file because there's difference in format. In order to optimize, several operation parameters for Java
Virtual Machine (JVM) and JADE class loader are given as listed in Table 7.
Table 6: Complete list of scenario’s dummy
Parameters
Specification block counts
Start year
Simulation duration (year)
Harvest ratio (Revenue/seed)
Kg seed per Ha
Autonom (boolean)
6
2002
8
2
1000
1
Values
Group
GUI
Name
Role
Stock at start (kg)
Money at start (Rp)
Seed at start (Kg)
Minimum diameter (cm)
Maximum diameter (cm)
Production usage (kg)
Production income (Rp/kg)
Market income (Rp/kg)
Normal buy price (Rp/Kg)
Normal pledge price (Rp/Kg)
Harvest failure chance (%)
Produsen
1
1
P1
P2
0
10
35000
20
0
0
0
0
1000
5300
0
0.45
0
5
100000
34
0
0
0
0
1600
5100
0
0.57
Field (Ha)
Plant cost (Rp/Ha)
FCL path
Number of agents
Group
GUI
Name
Role
Stock at start (kg)
Money at start (Rp)
Seed at start (Kg)
Minimum diameter (cm)
Maximum diameter (cm)
Production usage (kg)
Production income (Rp/kg)
Market income (Rp/kg)
Normal buy price (Rp/Kg)
Normal pledge price (Rp/Kg)
Harvest failure chance (%)
Field (Ha)
Plant cost (Rp/Ha)
FCL path
Number of agents
Group
GUI
Name
Role
Stock at start (kg)
Money at start (Rp)
Seed at start (Kg)
Minimum diameter (cm)
Maximum diameter (cm)
Production usage (kg)
Production income (Rp/kg)
Market income (Rp/kg)
Normal buy price (Rp/Kg)
Normal pledge price (Rp/Kg)
Harvest failure chance (%)
Field (Ha)
Plant cost (Rp/Ha)
FCL path
2.3
50000
script/produsen.fcl
1
Distributor
1
D3
D4
1
12
250000
0
4.9
6.3
0
0
1200
5700
0
0
0
0
script/distributor.fcl
1
Konsumen
1
5.3
56400
1
1
1
3
37600
0
3.9
5.7
0
0
1200
5500
0
0
0
0
1
1
K5
K6
2
500
400000
80
5.1
6
150
3000
1500
6000
6200
0
0
0
script/konsumen.fcl
2
830
100000
220
6.5
8.79
125
4100
900
6400
7500
0
0
0
Number of agents
Table 7: Operation parameters
Parameter
-Xms512m
-Xmx1024m
jade_core_messaging_MessageManager
_maxqueuesize 50000000
1
1
Description
Allocate 512MB of RAM to JVM
Allocate maximal 1024MB of RAM to JVM for additional usage
Tell JADE to allocate 50MB of RAM as message queue
The results of simulation for two agents of “ProdusenAgent” are shown in below figures. Fig.6
shows complete GUI of the agents and commodity charts at same time. The finance charts hidden
inside the scroll area can be seen in Fig.7.
Figure 6. Commodity charts of “ProdusenAgent” agents
Figure 7. Finance charts of “ProdusenAgent” agents
The results of simulation for two agents of “DistributorAgent” are shown both in commodity (Fig. 8)
and in finance (Fig. 9) charts too.
Figure 8. Commodity charts of “DistributorAgent” agents
Figure 9. Finance charts of “DistributorAgent” agents
Finally, the output of “KonsumenAgent” agents are shown by Fig.10 and Fig.11.
Figure 10. Commodity charts of “KonsumenAgent” agents
Figure 11. Finance charts of “KonsumenAgent” agents
For additional test, the system is also strained by simulating some numbers of agents instances (Table
8), including previous dummy test as test no. I. This is conducted to see how well the system's performance is.
Table 9 shows specification of the testing machine where all test performed. Memory and processor (CPU)
usages that are used during the test are listed by Table 10.
Table 8: Number of agent’s instances for strain test
Number of
Test No.
Number of
“ProdusenAgent”
“DistributorAgent”
instance
instance
2
2
7
7
12
12
17
17
22
22
Number of
“KonsumenAgent”
instance
2
7
12
17
22
I
II
III
IV
V
Total
6
21
36
51
66
Table 9: Testing machine
Parameter
Operating System
CPU
RAM
Storage
JDK/JRE version
Description
Windows 7 Ultimate 32 bit
AMD E-450 2 CPU @1.6GHz
2048MB, 384MB Shared
ST950032 SATA
Java 6 update 29
Table 10: Strain test results
Memory usage
Test No.
CPU usage
Duration
24,90%
98,30%
100,00%
100,00%
100,00%
5 m 46 s
5 m 59 s
5 m59 s
6 m 44 s
19 m 54 s
Min (Byte) Max (Byte)
21.635.680
156.129.920
159.294.232
19.316.896
164.188.432
17.112.774
24.967.448
591.403.360
49.914.848 1.360.807.024
I
II
III
IV
V
4. Conclusions
The built system can run simulation based on given scenario then produce output each agent's yearly finance and
commodity data in form of spreadsheet and chart graph files. The system is also able to add additional value for
spreadsheet files as prediction. The output is limited to raw data without any further analysis or conclusions,
which are expected to come from appropriate market experts. Inheriting advantage of agent based application,
system extension efforts like adding new agents or new behaviors is possible. For further research, some steps
can be taken like implementing self adapting logic, performance optimization, and improvement for better
quality of results.
References
[1]
Fabio Bellifemine., Giovanni Caire., Dominic Greenwood., 2007, Developing multi-agent
systems with JADE, John Wiley and Sons, Ltd, West Sussex, England.
[2] Magid Nikraz., Giovanni Caire., Parisa A. Bahri., 2xxx, A methodology for the Analysis and
Design of Multi-Agent Systems using JADE, Telecom Italia Lab, Via Reiss Romoli, Turin, Italy
10148.
|
1004.3527 | 1 | 1004 | 2010-04-20T18:43:48 | On Asymptotic Consensus Value in Directed Random Networks | [
"cs.MA",
"math.OC"
] | We study the asymptotic properties of distributed consensus algorithms over switching directed random networks. More specifically, we focus on consensus algorithms over independent and identically distributed, directed random graphs, where each agent can communicate with any other agent with some exogenously specified probability. While different aspects of consensus algorithms over random switching networks have been widely studied, a complete characterization of the distribution of the asymptotic value for general \textit{asymmetric} random consensus algorithms remains an open problem. In this paper, we derive closed-form expressions for the mean and an upper bound for the variance of the asymptotic consensus value, when the underlying network evolves according to an i.i.d. \textit{directed} random graph process. We also provide numerical simulations that illustrate our results. | cs.MA | cs | On Asymptotic Consensus Value in Directed Random Networks
Victor M. Preciado, Alireza Tahbaz-Salehi, and Ali Jadbabaie
0
1
0
2
r
p
A
0
2
]
A
M
.
s
c
[
1
v
7
2
5
3
.
4
0
0
1
:
v
i
X
r
a
Abstract -- We study the asymptotic properties of distributed
consensus algorithms over switching directed random networks.
More specifically, we focus on consensus algorithms over inde-
pendent and identically distributed, directed random graphs,
where each agent can communicate with any other agent with
some exogenously specified probability. While different aspects
of consensus algorithms over random switching networks have
been widely studied, a complete characterization of the distri-
bution of the asymptotic value for general asymmetric random
consensus algorithms remains an open problem. In this paper,
we derive closed-form expressions for the mean and an upper
bound for the variance of the asymptotic consensus value, when
the underlying network evolves according to an i.i.d. directed
random graph process. We also provide numerical simulations
that illustrate our results.
I. INTRODUCTION
Distributed consensus algorithms have attracted a signif-
icant amount of attention in the past few years. Besides
their wide range of applications in distributed and parallel
computation [1], distributed control [2], [3] and robotics [4],
they have also been used as models of opinion dynamics and
belief formation in social networks [5], [6]. The central focus
in this vast body of literature is to study whether a group of
agents in a network, with local communication capabilities
can reach a global agreement, using simple, deterministic
information exchange protocols.1
In recent years, there has also been some interest in un-
derstanding the behavior of consensus algorithms in random
settings [7] -- [12]. The randomness can be either due to the
choice of a randomized network communication protocol
or simply caused by the potential unpredictability of the
environment in which the distributed consensus algorithm
is implemented [13]. It is recently shown that consensus
algorithms over i.i.d. random networks lead to a global agree-
ment on a possibly random value, as long as the network
is connected in expectation [10]. While different aspects of
consensus algorithms over random switching networks, such
as conditions for convergence [7] -- [10] and the speed of con-
vergence [13], have been widely studied, a characterization
This research is supported in parts by the following grants: DARPA/DSO
SToMP, NSF ECS-0347285, ONR MURI N000140810747, and AFOSR:
Complex Networks Program.
Victor Preciado is with Department of Electrical and Systems En-
gineering, University of Pennsylvania, Philadelphia, PA 19104 (e-mail:
[email protected])
Alireza Tahbaz-Salehi is with Department of Electrical Engineering and
Computer Science, Massachusetts Institute of Technology, Cambridge, MA
02139 (e-mail: [email protected]).
Ali Jadbabaie is with the General Robotics, Automation, Sensing and
Perception (GRASP) Laboratory, Department of Electrical and Systems
Engineering, University of Pennsylvania, Philadelphia, PA 19104 (e-mail:
[email protected]).
1For a survey on the most recent works in this area see [3].
of the distribution of the asymptotic consensus value has
attracted little attention. Two notable exceptions are Boyd et
al. [14], who study the asymptotic behavior of the random
consensus value in the special case of symmetric networks,
and Tahbaz-Salehi and Jadbabaie [15], who compute the
mean and variance of the consensus value for general i.i.d.
graph processes. Nevertheless, a complete characterization of
the distribution of the asymptotic value for general asymmet-
ric random consensus algorithms remains an open problem.
In this paper, we study asymptotic properties of consensus
algorithms over a general class of switching, directed random
graphs. More specifically, building on the results of [15],
we derive closed-form expressions for the mean and an
upper-bound for the variance of the asymptotic consensus
value, when the underlying network evolves according to
an i.i.d. directed random graph process. In our model, at
each time period, a directed communication link is estab-
lished between two agents with some exogenously specified
probability. Due to the potential asymmetry in pairwise
communications between different agents,
the asymptotic
value of consensus is not guaranteed to be the average of the
initial conditions. Instead, agents will asymptotically agree
on some random value in the convex hull of the initial
conditions. Furthermore, our closed-form characterization
of the variance provides a quantitative measure of how
dispersed the random agreement point is around the average
of the initial conditions in terms of the fundamentals of the
model, namely, the structure of the network, the exogenous
probabilities of communication, and the initial conditions.
The rest of the paper is organized as follows. In the
next section, we describe our model of random consensus
algorithms. In Sections III and IV, we derive an explicit
expression for the mean and an upper bound for the variance
of the limiting consensus value over switching directed
random graphs, respectively. Section V contains simulations
of our results and Section VI concludes the paper.
II. CONSENSUS OVER SWITCHING DIRECTED RANDOM
GRAPHS
Consider the discrete-time linear dynamical system
x (k) = Wkx (k − 1),
(1)
where k ∈ {1,2, . . . } is the discrete time index, x(k) ∈ Rn
is the state vector at time k, and {Wk}
k=1 is a sequence of
stochastic matrices. We interpret (1) as a distributed scheme
where a collection of agents, labeled 1 through n, update
their state values as a convex combination of the state values
of their neighbors at
the previous time step. Given this
interpretation, xi(k) corresponds to the state value of agent i
¥
at time k, and Wk captures the neighborhood relation between
different agents at time k: the i j element of Wk is positive
only if agent i has access to the state of agent
j. For the
remainder of the paper, we assume that the weight matrices
Wk are randomly generated by an independent and identically
distributed matrix process.
We say that the dynamical system (1) reaches consensus
k=1, if along that path,
asymptotically on some path {Wk}
there exists x∗ ∈ R such that xi(k) → x∗ for all i as k → ¥
.
We refer to x∗ as the asymptotic consensus value. It is well-
known that for i.i.d. random networks, the dynamical system
(1) reaches consensus on almost all paths if and only if the
graph corresponding to the communications between agents
is connected in expectation. More precisely, it was shown
in [10] that Wk . . .W2W1 −→ 1dT almost surely − where d
is some random vector − if and only if the second largest
eigenvalue modulus of EWk is subunit. Clearly, under such
conditions, the dynamical system in (1) reaches consensus
with probability one where the consensus value is a random
variable equal to x∗ = dT x(0), where x(0) is the vector of
initial conditions.
A complete characterization of the random consensus
value x∗ is an open problem. However, it is possible to
compute its mean and variance in terms of the first two
moments of the i.i.d. weight matrix process. In [15], the
authors prove that
the conditional mean of the random
consensus value is given by the random consensus value are
given by
Ex∗ = x(0)T v1(EWk),
and its conditional variance is equal to
var(x∗) = [x(0) ⊗ x(0)]T vec(cov(d))
= [x(0) ⊗ x(0)]T v1(E [Wk ⊗Wk]) − [x(0)T v1(EWk)]2
where v1 (·) denotes the normalized left eigenvector corre-
sponding to the unit eigenvalue, and ⊗ denotes the Kronecker
product. In [16], the authors derive an explicit expression
for the mean and variance of x∗ in the particular case
of a switching Erdos-R´enyi random graph process. In the
following, we shall use (3) to extend these results and derive
an explicit expression for the mean and an upper-bound for
the variance of the asymptotic consensus value over a wider
class of switching, directed random graph processes.
III. MEAN ANALYSIS FOR DIRECTED RANDOM GRAPHS
PROCESSES
A. Directed Random Graph Process
Consider a connected undirected graph Gc = (V,Ec) with
a fixed set of vertices V = [n] , and unweighted edges (no
self-loops allowed). Each undirected edge in Ec represents
a potential communication channel between nodes i and j,
where this channel can be used to send information in both
directions. In this paper, we focus on directed communica-
tions, i.e., the event of node i sending information towards
node j is independent from the even of node j sending
information towards i. In this context, it is convenient to
interpret an undirected edge {i, j} ∈ Ec as the union of two
(2)
(3)
independent directed edges, {(i, j) , ( j,i)}, where the ordered
pair (i, j) represents a directed link from node i to node j.
In this paper, we study randomized time-switching con-
sensus processes. In particular, in each discrete time slot
k ≥ 1, we construct a random directed graph Gk = (V, Ek),
with Ek ⊆ Ec, such that the existence of a directed edge
(u,v) ∈ Ek is determined randomly and independently of all
other directed edges (including the reciprocal edge (v,u))
with a probability puv ∈ (0,1) for (u,v) ∈ Ec, and puv = 0 for
(u,v) 6∈ Ep. In other words, in each time slot, we randomly
select a subset Ek of directed links chosen from a set of
candidate (directed) links in Ec. We are specially interested
in the case in which the probability puv of existence of a
directed link (u,v) depends exclusively on the node that
receives information via that link, i.e., Pr ((u,v) ∈ Ek) = pv,
where pv ∈ (0,1). In this setting, we can model the ability of
a node to 'listen' to their neighboring nodes. For example,
in the context of opinion dynamics in social networks [5],
[6], [12], the probability pv can represent the tendency of the
individual at node v to take into account the opinion of her
neighbors (which could depend, for example, on how many
acquaintances the individual has).
Let us denote by Ac the symmetric adjacency matrix
of the graph Gc, where entries ai j = 1 if {i, j} ∈ Ec, and
ai j = 0 otherwise. We define the degree of node i as di =
(cid:229) n
j=1 a ji, and the associated degree matrix as Dc = diag(di).
We also denote the random (nonsymmetric) adjacency matrix
associated with Gk as Ak =h a(k)
uvi, which can be described
as
a(k)
uv =(cid:26) auv, w.p. pv,
0,
w.p. 1 − pv.
We denote the in-degrees of Gk as d(k)
uv , and the in-
degree matrix as Dk = diag( d(k)
v ). From the definition of Gk,
the in-degrees are independent Bernoulli random variables
d(k)
v ∼ Ber (dv, pv), i.e.,
v = (cid:229) u a(k)
Pr(cid:16) d(k)
d(cid:19)pd
v = d(cid:17) =(cid:18)dv
v (1 − pv)dv−d .
We describe the consensus dynamics in (1) via a sequence
k=1 associated to the sequence of
of stochastic matrices {Wk}
random directed graphs {Gk = (V, Ek)}
k=1 as follows:
(5)
Wk =(cid:0) Dk + I(cid:1)−1(cid:0) Ak + I(cid:1)T , for k ≥ 1.
Notice that adding the identity matrix to the adjacency in
(5) is equivalent to adding a self-loop to every vertex in V .
We include these self-loops to avoid singularities associated
with the existence of isolated nodes in Gk (for which di = 0,
and Dk would not invertible). Since Gc is connected and
pv > 0 for all v ∈ V , the expected communications graph is
connected and the stochastic dynamical system in (1) reaches
consensus on almost all paths, although the asymptotic
consensus value x∗ is a random variable (not the initial
average).
In the following subsections, we first derive closed-form
expressions for the mean of x∗, and an upper-bound for the
variance var (x∗) in terms of the following elements:
(4)
¥
¥
¥
1) The set of initial conditions, {xu (0)}u∈V ,
2) the set of nodes properties, {(pu,du)}u∈V , and
3) the network topology, via the eigenvalues of the ex-
pected matrix EWk.
As we shall show in Section IV, our expression for the
variance has a nice interpretation, since it separates the
influence of each one of the above elements into three
multiplicative terms. In the next subsection we provide the
details regarding our analysis of the expectation of x∗.
B. Mean of Consensus Value
We use (2) to study the mean of the consensus value.
We first derive an expression for EWk, and then study its
dominant left eigenvector v1 (EWk). For notational conve-
di ∼ Ber (di, pi), and denote its first and second moments as
M(1)
nience, we define the random variable zi , 1/(cid:0) di + 1(cid:1) where
i(cid:1). The diagonal entries of EWk
are then given by Ewii = E(cid:2)1/(cid:0) di + 1(cid:1)(cid:3), which present the
following explicit expression (see Appendix I for details):
, E (zi) and M(2)
, E(cid:0)z2
i
i
Ewii = M(1)
i =
1 − qdi+1
pi (di + 1)
i
.
(6)
where qi = 1 − pi. Furthermore, the off-diagonal entries of
EWk are equal to (see Appendix I for details):
qdi+1
i + pi (di + 1) − 1
Ewi j = a ji
pi (di + 1)di
= a ji
1 − M(1)
i
di
.
(7)
Taking (6) and (7) into account, we can write EWk as follows:
EWk = S + (I − S )D−1
c AT
c ,
(8)
where S , diaghM(1)
i i. As expected, it is easy to check
i.e., (EWk) 1n = S 1n +
is a stochastic matrix,
that EWk
(I − S )D−1
c AT
c 1n = 1n.
Based on (8) we can write E (x∗) explicitly in terms of di
and pi, as follows:
Theorem 1: Consider the random adjacency matrix Ak in
(4) and the associated (random) stochastic matrix Wk in (5).
The expectation of the asymptotic consensus value of (1) is
given by
where
wi (pi,di) ,
E (x∗) =
n(cid:229)
r wi xi (0) ,
i=1
pi (di + 1)di
pi (di + 1) − 1 − qdi+1
i
, and
r (pi,di) , (cid:229)
i
wi (pi,di)!−1
.
Proof: Our proof is based on computing v1 (EWk)
and applying (2). Let us define v , v1 (EWk) and w ,
(I − S )v. From (8), we have that the eigenvalue equation
corresponding to the dominant left eigenvector of EWk is
vT(cid:0)S + (I − S )D−1
c AT
c(cid:1) = vT , which can be rewritten as
c = vT (I − S ). This last equation can be
vT (I − S )D−1
c AT
written as wT D−1
c AT
c = wT . The solution to this equation is
the stationary distribution of the Markov chain with transition
matrix D−1
i di, where d =
(d1, ...,dn)T . Hence, the solution to the eigenvalue equation
is
c , which is equal to p = d/(cid:229)
c AT
v1 (EWk) = s (I − S )−1 d,
where we include the normalizing parameter
s = (cid:229)
i
di
i !−1
1 − M(1)
such that kv1 (EWk)k1 = 1. Hence, from (2) we have
E (x∗) =
n(cid:229)
s
di
i=1
1 − M(1)
xi (0) .
i
Substituting the expression for M(1)
simple algebraic simplifications.
i
in (6), we reach (9) via
In general, the asymptotic mean E (x∗) does not coincide
with the initial average ¯x0 = 1
i xi (0). There is a simple
n
technique, based on Theorem 1, that allows us to make the
expected consensus value to be equal to the initial average.
This technique consists of using yi (0) = (r wi)−1 xi (0) as
initial conditions in (1). Hence, one can easily check that the
asymptotic consensus value E (y∗) equals the initial average
¯x0.
IV. VARIANCE OF THE ASYMPTOTIC CONSENSUS VALUE
In this section, we derive an expression that explicitly
relates the variance var (x∗) with the three elements that
influences it, namely, the set of initial conditions {xu (0)}u∈V ,
the nodes properties {(pu,du)}u∈V , and the network structure
(via the eigenvalues of the expected matrix EWk). For sim-
plicity in notation, we denote E (Wk ⊗Wk) and (EWk ⊗ EWk)
by R and Q, respectively. Our analysis starts in expression
(3), which can be rewritten as
var (x∗) = [x(0) ⊗ x(0)]T [v1 (R) − v1 (Q)] .
Hence, we can upper-bound the variance of the asymptotic
consensus value as follows:
(9)
var (x∗) ≤ kx(0) ⊗ x(0)k1 kv1 (R) − v1 (Q)k¥
.
(10)
From the rules of Kronecker multiplication, we can write the
first factor in terms of the initial conditions as
kx(0) ⊗ x(0)k1 = (cid:229)
1≤i, j≤n(cid:12)(cid:12)xi (0)x j (0)(cid:12)(cid:12) .
(11)
In the following, we derive an upper bound for the sec-
ond factor kv1 (R) − v1 (Q)k¥
in terms of the nodes prop-
erties and the network structure. Our approach to bound
is based on the observation that both R
kv1 (R) − v1 (Q)k¥
and Q are n2 × n2 stochastic matrices, and the dominant left
eigenvectors v1 (R) and v1 (Q) are stationary distributions of
the Markov chains with transition matrices R and Q. We
(cid:229)
matrix D = R − Q also follows the same pattern as R and
Q. Hence, comparing the entries of Q and R we can easily
deduce the following seven cases for the entries of D
:
D 2 = D 5 = D 7 = 0,
D 1 = M2 (zi) − M2
D 3 =
a ji
1 (zi) ,
di (cid:0)−M2 (zi) + M2
1 (zi)(cid:1) ,
D 4 = a ji M1 (zi) − M2 (zi)
D 6 =( (cid:16) (1−Mi)2
di
− 1+2Vi−3Mi
−a jiasi (1 − M1 (zi))2 ,
di(di−1) (cid:17) ,
d2
i
−
(1 − M1 (zi))2
d2
i
! ,
for di > 1,
for di = 1.
i
, E(cid:2)1/( di + 1)2(cid:3) can be written as a hyperge-
where M(2)
ometric function that depends on pi and di (see Appendix
III). From the above entries, we can compute via simple, but
tedious, algebraic manipulations the following expression for
the infinity norm of the perturbation:
kR − Qk¥ = kD k¥ = max
1≤i≤n
{Si} ,
(13)
where
Si = 2(1 − M(1)
i
)(cid:20)M(1)
i +
1
di(cid:16)M(1)
i − 1(cid:17)(cid:21) ,
Note that Si is a function that depends exclusively on the
sequence of nodes properties {(pu,du)}u∈V ; hence, kR − Qk¥
depends exclusively on the set of nodes degrees and probabil-
ities, but it is independent on how these nodes interconnect.
In the following subsection, we show how the pattern of
interconnection among nodes influences the upper-bound of
the variance in (10) via the condition number k s in (12).
B. Perturbation-Based Bound for the Variance
In this subsection we derive an explicit relationship be-
tween the condition number k s and the network structure
via the spectral properties of EWk = S + (I − S )D−1
c . This
result will then be used to bound the variance of x∗ in (10).
We base our analysis on a the following bound, derived by
Meyer in [18], relating k s with the eigenvalues of Q, denoted
as {m i}n2
c AT
i=1:
2(cid:0)n2 − 1(cid:1)
k=2 (1 − m k)
n2
.
(14)
g†
max
i j(cid:12)(cid:12)(cid:12) <
i, j (cid:12)(cid:12)(cid:12)
j(cid:9)n2
i=1 and(cid:8)m
(cid:12)(cid:12)1 − l
Before we present our main result, we need some notation.
Denote by {l i}n
j=1 the set of eigenvalues of EWk
and EWk ⊗ EWk, respectively. The ordering of the eigenvalue
sequences is determined by their distance to 1, i.e., 1 − l i ≤
value of (1) can be upper-bounded by
Theorem 3: The variance of the asymptotic consensus
j(cid:12)(cid:12) for i ≤ j. Hence, our result can be stated as follows:
j)!
2(cid:0)n2 − 1(cid:1)
{Si}(cid:19)
(cid:18) max
}
{z
}
{z
var (x∗) ≤ kx(0) ⊗ x(0)k1
(1 − l il
{z
(15)
1≤i≤n
}
(C)
(A)
(B)
,
Fig. 1. The pattern of E[Wk ⊗Wk] for n = 3. The numbers in parentheses
represent the value of each entry defined in the Appendices.
denote these distributions by v1 (R) , p and v1 (Q) , p , re-
spectively. In this setting, we can apply the following lemma
from [17] which studies the sensitivity of the stationary
distribution of Markov chains:
Lemma 2: Consider two Markov chains with transition
matrices Q and R, and stationary distributions p and p , re-
spectively. We define G = I −Q, and denote its pseudoinverse
by G† = [g†
i j]. Hence,
,
(12)
k p − p k¥ ≤ k s kR − Qk¥
g†
is called the condition number of the
where k s = maxi, j(cid:12)(cid:12)(cid:12)
i j(cid:12)(cid:12)(cid:12)
chain described by the transition matrix Q.
In the next subsections, apply the above lemma to bound
the factor kv1 (R) − v1 (Q)k¥
. In the first subsection we com-
pute an explicit expression for the norm of the perturbation
kR − Qk¥ as a function of the properties of the nodes. In the
second subsection, we study the coefficient k s in terms of
the eigenvalues of EWk.
A. Infinity Norm of the Perturbation kR − Qk¥
Our approach is based on studying the entries of the n2 ×
n2 matrix R = E [Wk ⊗Wk], and compare them with the entries
of Q = EWk ⊗ EWk. The entries of Q and R are of the form
E(wi j)E(wrs) and E(wi jwrs), respectively, with i, j,r, and s
ranging from 1 to n. These entries can be classified into
seven different cases depending on the relations between the
indices. In the Appendices II and III, we present explicit
expressions for each one of these cases. A key observation
in our approach is to notice the pattern that the above entries
induces in the matrices R and Q. For sake of clarity, we
illustrate this pattern for n = 3 in Fig. 1, where the numbers
in parenthesis correspond to each one of the seven cases
identified in the Appendices.
Since the entries of R follow the same pattern as the entries
of Q (although these entries are different), the perturbation
(cid:213)
(cid:213)
where {l i}n
(1 − l il
i=1 are the eigenvalues of EWk, and the product
j) = (cid:213)
(i, j) s.t. (i, j)6=(1,1) (1 − l il
j) .
Proof: We start our proof from (10)
var (x∗) ≤ kv1 (R) − v1 (Q)k¥ kx(0) ⊗ x(0)k1
(a)
(b)
≤ k s kR − Qk¥ kx(0) ⊗ x(0)k1
= k s(cid:18) max
{Si}(cid:19) kx(0) ⊗ x(0)k1
2(cid:0)n2 − 1(cid:1)
(1 − m k)
max
1≤i≤n
1≤i≤n
(c)
<
{Si} kx(0) ⊗ x(0)k1 ,
where we have used Lemma 2 in inequality (a), the expres-
sion for kR − Qk¥
in (13) in equality (b), and the upper
bound in (14) in step (c). We obtain the statement of the
theorem by applying the following standard property of the
Kronecker product: {m k}n2
k=1 =(cid:8)l il
j(cid:9)1≤i, j<n.
Remark 4: The bound in (15) separates the variance into
three multiplicative terms representing each one of the fol-
lowing elements (for convenience, we have underlined each
one of these terms in (15)):
(A)
(B)
(C)
This first term exclusively depends on the initial
condition as indicated by (11).
The second term depends solely on the nodes
properties di and pi.
The last term represents the influence from the
overall graph structure via the spectral properties
of EWk.
c AT
Remark 5: It is specially interesting to study the implica-
tions of Term (C) in the asymptotic variance. For example,
given the sequences of degrees and probabilities, {di}n
i=1
and {pi}n
i=1, the influence of the network structure on the
variance is given via Term (C). Since the eigenvalues of
EWk are key in this term, it is interesting to briefly describe
the homogeneous Markov chain with transition matrix P ,
EWk = S + (I − S )D−1
c . This Markov chain presents a self-
loop in each one of its n states with transition probability
Pii = M(1)
, as well as a link from i to j with transition
for (i, j) ∈ Ec. From an
analysis proposed by Meyer in [18], we conclude that Term
(C) is primarily governed by how close the subdominant
eigenvalues of EWk are from 1. In particular, the further
the subdominant eigenvalues of EWk are from 1, the lower
the upper bound in (15). In the next section, we illustrate
the relationship between spectral properties of EWk and the
variance of the asymptotic consensus value with several
numerical simulations.
probability Pi j = (cid:16)1 − M(1)
i (cid:17) d−1
i
i
V. NUMERICAL SIMULATIONS
In this section we present several numerical simulations
illustrating our results. In the first subsection, we numerically
verify the result in Theorem 1. In the second subsection
we present some examples to illustrate the implications of
Theorem 3 in the variance of the asymptotic consensus value.
Fig. 2. The above figure represents a chain of leader nodes, denoted as 1,2,
and 3, with a set of followers. Connections between leaders are represented
as pairs of directed links with existence probabilities puv and pvu, with
u,v ∈ {1,2,3}. We have also included in this figure the connection of each
leader and one of its followers (drawn as a smaller circle on top of each
leader). In this case, the directed link incoming the follower (leader) exists
with probability pl f (p f l). The rest of followers are represented under each
leader, where we represent each pair of directed links as a single undirected
edge for clarity.
A. Expectation of the Asymptotic Consensus
In our first simulation we take a graph Gc composed by 3
stars connected in a chain (see Fig. 2). This graph is intended
to represent, in a social context, three leaders with a set of
followers. In Fig. 2, we represent the leaders using large
circles marked as 1,2, and 3, and the followers using smaller
circles. We assume that each follower only listen to one
leader (the center of a star) and nobody else. In this particular
example, the first, second and third leaders have 4, 8, and 16
followers, respectively. In each time step, a directed random
graph Gk is built by choosing a set of directed communication
links from Gc. We have fixed the probability of existence of a
directed link coming into followers to be equal to pl f = 1/2,
for all such links (see Fig. 2). Also,
the probability of
existence of a directed link coming into a leader is inversely
proportional to the degree of the leader. More specifically,
we have chosen puv = 1/dv, for v = 1,2, and 3, where puv
represents the existence of a directed communication link
from the u-th node to the v-th leader.
In this example, we compute the asymptotic consensus
value for 100 realizations of a random consensus algorithm
with initial conditions xi (0) = i/N. We represent the his-
togram of these realizations in Fig. 3, where the empirical
average of the 100 realizations equals 0.5077. The theoretical
expectation for the asymptotic consensus value applying
i r wi (i/N) =0.4595, which is
Theorem 1 equals E (x∗) = (cid:229)
in great accordance with the empirical value. For future
reference, we have also computed the empirical standard
deviation of the 100 realizations to be 0.0298, and the three
eigenvalues of EWk closest to 1 are {0.9823,0.9449,0.75}.
B. Expectation of the Asymptotic Variance
In this subsection, we numerically illustrate some of the
implications of Theorem 3 in the variance of the asymptotic
consensus value. Although the upper bound stated in this
Theorem is not tight, there are some qualitative implications
(cid:213)
(cid:213)
Fig. 3. Empirical histogram for 100 realizations of the random directed
consensus over the graph in Fig. 2. The empirical average and the empirical
standard deviation of the realizations are 0.4595 and 0.0298, respectively.
that are consistent throughout our numerical experiments.
For example, as we mentioned in Remark 5, given a set of
initial conditions and nodes properties, the upper bound in
(15) is primarily governed by how close the subdominant
eigenvalues of EWk are from 1. In particular, the further
the subdominant eigenvalues of EWk are from 1, the lower
the upper bound. We illustrate in the following numerical
experiments that the lower the upper bound, the lower one
should expect the variance to be.
In the first experiment, we slightly modify the network
described in the previous subsection and study the influence
of this modification on the eigenvalues of EWk and the
variance of x∗. Our first modification is a change in the
probabilities of existence of a directed link from the u-
th node to the v-th leader without changing the network
topology. In particular, we choose the new probabilities to be
puv = 3/dv, for v = 1,2, and 3. In the modified network, the
eigenvalues of EWk closest to 1 are {0.9789,0.9372,0.75}.
Hence, since the effect of our modification on the eigenvalues
is to move them away from 1, we should expect the variance
of x∗ to be reduced according to Remark 5. In fact, running
100 random consensus algorithms with the same initial
conditions, xi (0) = i/N, using our new probabilities, we
obtain a standard deviation 0.0286 (which is less than the
0.0298 obtained before).
In the second experiment, we illustrate how the larger
the gap between the eigenvalues of EWk and 1, the smaller
the variance of x∗. In this case, apart from keeping the
new set of probabilities puv = 3/dv, we convert the 3-chain
of leaders into a 3-ring of leaders, as depicted in Fig.
4. In this case,
the three eigenvalues of EWk closest to
1 are {0.9577,0.9212,0.75}, which are even further away
from 1 than in the second example. Hence, as expected,
the standard deviation of 100 random consensus algorithms
is even smaller than in the second example, in particular
0.0274. In conclusion, our simulations are consistent with
Theorem 3 and with the qualitative behavior described in
Remark 5.
Fig. 4. Ring of 3 leaders and corresponding followers.
VI. CONCLUSIONS AND FUTURE WORK
We have studied the asymptotic properties of the consen-
sus value in distributed consensus algorithms over switching,
directed random graphs. While different aspects of consensus
algorithms over random switching networks, such as condi-
tions for convergence and the speed of convergence, have
been widely studied, a characterization of the distribution
of the asymptotic consensus for general asymmetric random
consensus algorithms remains an open problem.
In this paper, we have derived closed-form expressions
for the expectation of the asymptotic consensus value as a
function of the set of initial conditions, {xu (0)}u∈V , and the
set of nodes properties, {(pu,du)}u∈V , as stated in Theorem
1. We have also studied the variance of the asymptotic
consensus value in terms of several elements that influence
it, namely, (i) the initial conditions, (ii) nodes properties, and
(iii) the network topology. In Theorem 3, we have derived
an upper bound for the variance of the asymptotic consensus
value that explicitly describes the influence of each one of
these elements. We also provide an interpretation of the
influce of the network topology on the variance in terms
of the eigenvalues of the expected matrix EWk. From our
analysis we conclude that, in most cases, the variance of
x∗ is primarily governed by how close the subdominant
eigenvalues of EWk are from 1. We have checked the validity
of our predictions with several numerical simulations.
APPENDIX
COMPUTING THE ENTRIES OF EWk
We start by computing the entries of EWk. The diagonal
entries of EWk are given by:
Ewii = E(cid:20) 1
di
1 + di(cid:21) =
k(cid:19)pk
k + 1(cid:18)di
1
k=0
k=0
1 − qdi+1
pi (di + 1)
i
, M(1)
i
di
1
k + 1
P( di = k)
i (1 − pi)di−k
=
=
(cid:229)
(cid:229)
Similarly, the non-diagonal entries of EWk result in:
Similarly, we also have:
Ewi j = E(cid:20) a ji
1 + di(cid:21)
qdi+1
i + pi (di + 1) − 1
= a ji
= a ji
pi (di + 1)di
1 − M(1)
i
di
.
APPENDIX
ENTRIES OF EWk ⊗ EWk
We now compute the possible entries in Q = EWk ⊗ EWk.
The entries of the Kronecker matrix EWk ⊗ EWk, with EWk =
S + (I − S )D−1AT , present entries that can be classified into
seven different cases depending on the relations between the
indices. These are the cases, where we assume that all four
indices i, j,r, and s are distinct:
,
Q4 , E (wi j) E (wi j) =
Q2 , E (wii) E (w j j) = M(1)
Q3 , E (wii) E (wi j) =
i (cid:17)2
Q1 , E (wii) E (wii) =(cid:16)M(1)
i (cid:16)1 − M(1)
i (cid:17) ,
i (cid:17)2
i (cid:16)1 − M(1)
r (cid:17) ,
i (cid:16) 1 − M(1)
i (cid:17)2
i (cid:16)1 − M(1)
r (cid:17) .
i (cid:17)(cid:16) 1 − M(1)
didr (cid:16)1 − M(1)
,
i M(1)
j
a ji
M(1)
di
a ji
d2
asr
dr
a jiasi
d2
a jiasr
Q7 , E (wi j) E (wrs) =
Q6 , E (wi j) E (wis) =
Q5 , E (wii) E (wrs) =
M(1)
,
,
APPENDIX
ENTRIES OF EWk ⊗Wk
We now turn to the computation of the elements of E[Wk ⊗
Wk], which are of the form E(wi jwrs). Again, we classify the
entries into seven different cases depending on the relations
between the subindices:
R1 , Ew2
1
ii = E(cid:20)
(k + 1)2(cid:18)di
( di + 1)2(cid:21)
k(cid:19)pk
1
i qdi−k
i
=
di
k=0
= qdiH(pi,di) , M(2)
i
R2 , E (wiiw j j) = M(1)
R3 , E (wiiwi j) =
R4 , E (wi jwi j) =
R5 , E (wiiwrs) =
R6 , E (wi jwis)
,
j
i M(1)
a ji
di (cid:16)M(1)
di (cid:16)M(1)
a ji
M(1)
asr
dr
i (cid:17) ,
i − M(2)
i (cid:17) ,
i − M(2)
r (cid:17) ,
i (cid:16) 1 − M(1)
i (cid:17) ,
i − 3M(1)
= ( a jiasi
di(di−1)(cid:16)1 + 2M(2)
0,
for di > 1,
for di = 1,
R7 , E(wi jw ji) = E(wi jw js) = E(wi jwri)
= E(wi jwr j) = E(wi jwrs)
=
a jiasr
didr (cid:16)1 − M(1)
i (cid:17)(cid:16) 1 − M(1)
r (cid:17) .
REFERENCES
[1] J. N. Tsitsiklis, Problems in decentralized decision making and com-
putation, Ph.D. dissertation, Massachusetts Institute of Technology,
Cambridge, MA, 1984.
[2] A. Jadbabaie, J. Lin, and A. S. Morse, "Coordination of groups
of mobile autonomous agents using nearest neighbor rules," IEEE
Transactions on Automatic Control, vol. 48, no. 6, pp. 988-1001, 2003.
[3] R. Olfati-Saber, J. A. Fax, and R. M. Murray, "Consensus and
cooperation in networked multi-agent systems," Proceedings of the
IEEE, vol. 95, no. 1, 2007.
[4] J. Cortes, S. Martinez, and F. Bullo, "Analysis and design tools
for distributed motion coordination," in Proceedings of the American
Control Conference, Portland, OR, pp. 1680-1685, 2005.
[5] M. H. DeGroot, "Reaching a Consensus," Journal of American Sta-
tistical Association, vol. 69, no. 345, pp. 118-121, 1974.
[6] B. Golub and M. O. Jackson, "Naive learning in social networks: Con-
vergence, influence, and the wisdom of crowds," 2007, unpublished
Manuscript.
[7] Y. Hatano and M. Mesbahi, "Agreement over random networks," IEEE
Transactions on Automatic Control, vol. 50, no. 11, pp. 1867-1872,
2005.
[8] C. W. Wu, "Synchronization and convergence of linear dynamics in
random directed networks," IEEE Transactions on Automatic Control,
vol. 51, no. 7, pp. 1207-1210, 2006.
[9] M. Porfiri and D. J. Stilwell, "Consensus seeking over random
weighted directed graphs," IEEE Transactions on Automatic Control,
vol. 52, no. 9, pp. 1767-1773, 2007.
[10] A. Tahbaz-Salehi and A. Jadbabaie, "A necessary and sufficient
condition for consensus over random networks," IEEE Transactions
on Automatic Control, vol. 53, no. 3, pp. 791-795, 2008.
[11] G. Picci and T. Taylor, "Almost sure convergence of random gossip
algorithms," in Proceedings of the 46th IEEE Conference on Decision
and Control, New Orleans, LA, pp. 282-287, 2007.
[12] D. Acemoglu, A. Ozdaglar, and A. ParandehGheibi, "Spread of
(mis)information in social networks," Massachusetts Institute of Tech-
nology, LIDS Report 2812, May 2009, submitted for Publication.
[13] F. Fagnani and S. Zampieri, "Randomized consensus algorithms over
large scale networks," IEEE Journal on Selected Areas in Communi-
cations, vol. 26, no. 4, pp. 634-649, 2008.
[14] S. Boyd, A. Gosh, B. Prabhakar, and D. Shah, "Randomized gossip
algorithms," Special issue of IEEE Transactions on Information Theory
and IEEE/ACM Transactions on Networking, vol. 52, no. 6, pp. 2508-
2530, 2006.
[15] A. Tahbaz-Salehi and A. Jadbabaie, "Consensus over ergodic station-
ary graph processes," IEEE Transactions on Automatic Control, vol.
55, no. 1, pp. 225-230, 2010.
[16] V.M. Preciado, A. Tahbaz-Salehi, and A. Jadbabaie, "Variance analysis
of randomized consensus in switching directed networks,"in Proceed-
ings of the IEEE American Control Conference, 2010, accepted for
Publication.
(cid:229)
[17] R.E. Funderlic and C.D. Meyer, "Sensitivity of the stationary dis-
tribution vector for an ergodic Markov chain," Linear Algebra and
Applications, vol. 76, pp. 1-17, 1986.
[18] C.D. Meyer, "Sensitivity of the stationary distribution of a Markov
chain," SIAM J. Matrix Appl., vol. 15, pp. 715-728, 1994.
|
1901.09837 | 1 | 1901 | 2019-01-28T17:33:45 | Designing a Multi-Objective Reward Function for Creating Teams of Robotic Bodyguards Using Deep Reinforcement Learning | [
"cs.MA",
"cs.LG"
] | We are considering a scenario where a team of bodyguard robots provides physical protection to a VIP in a crowded public space. We use deep reinforcement learning to learn the policy to be followed by the robots. As the robot bodyguards need to follow several difficult-to-reconcile goals, we study several primitive and composite reward functions and their impact on the overall behavior of the robotic bodyguards. | cs.MA | cs | Designing a Multi-Objective Reward Function for Creating Teams of Robotic
Bodyguards Using Deep Reinforcement Learning
9
1
0
2
n
a
J
8
2
]
A
M
.
s
c
[
1
v
7
3
8
9
0
.
1
0
9
1
:
v
i
X
r
a
Hassam Ullah Sheikh 1 Ladislau Boloni 1
Abstract
We are considering a scenario where a team of
bodyguard robots provides physical protection to
a VIP in a crowded public space. We use deep
reinforcement learning to learn the policy to be
followed by the robots. As the robot bodyguards
need to follow several difficult-to-reconcile goals,
we study several primitive and composite reward
functions and their impact on the overall behavior
of the robotic bodyguards.
1. Introduction
Recent progress in the field of autonomous robots makes it
feasible for robots to interact with multiple humans in public
spaces.
In this paper, we are considering a human VIP
moving in a crowded market environment who is protected
from physical assault by a team of bodyguard robots. The
robots must take into account the position and movement of
the VIP, the bystanders and other robots. Previous work in
similar problems relied on explicitly programmed behaviors.
Recent research in deep reinforcement learning (Levine
et al., 2016; Silver et al., 2016) and imitation learn-
ing (Rahmatizadeh et al., 2018) applied to the robotics has
raised the possibility that learning algorithms might lead to
better algorithms than explicit programming. In this paper
we explore deep reinforcement learning approaches for our
scenario. We aim to simultaneously learn communication
and coordination techniques between the agent and the task
oriented behavior.
We aim to develop a general task framework, which can
generalize to other types of desired behaviors, beyond body-
guard protection.
In order to achieve these goals, we need to specify: the envi-
ronment in which the agents will perform, the environment
1Department of Computer Science, University of Central
Florida, Orlando, Florida, United States. Correspondence to: Has-
sam Ullah Sheikh <[email protected]>.
Accepted at the 1st Workshop on Goal Specifications for Reinforce-
ment Learning, FAIM 2018, Stockholm, Sweden, 2018. Copyright
2018 by the author(s).
representation in the robot that forms the basis of learning,
the reward function that describes the desired behavior, and
the reinforcement algorithms deployed. For the bodyguard
task, the design of the reward function is especially chal-
lenging, because it is task specific, and it needs to reconcile
multiple conflicting objectives the maximum protection,
while minimizing interference with the crowd and being un-
obtrusive. In Section 4.2 we discuss several different reward
functions that reflect the different aspects of the desired
behavior.
We describe several experiments using the Multi-agent Deep
Deterministic Policy Gradient (MADDPG) algorithms over
several choices of reward functions. We found that commu-
nication penalization reward functions captures better the
collaborative nature of the scenario, and thus it performs
better in experiments.
2. Related Work
The use of robots as a bodyguards is related to several dif-
ferent area of research and has received significant attention
lately. Several different studies such as (Richard Klima,
2016; Yasuyuki et al., 2015) considered using robots and
multi-agent reinforcement learning for security related tasks
such as patrolling and team coordination by placing check-
points to provide protection against imminent adversaries.
A multi-robot patrolling framework was proposed by (Khan
et al., 2012) that analyzes the behavior pattern of the sol-
diers and the robot and generates a patrolling schedule. The
control of robots for providing maximal protection to a VIP
was well studied in (Bhatia et al., 2016) where they intro-
duced the concept of threat vector resolution and quadrant
load balancing.
3. Background
We consider the problem of providing maximal physical pro-
tection as a standard reinforcement learning setup with N
agent interacting with the environment in E discrete steps us-
ing real valued continuous actions at such that at ∈ Rd. At
each timestep t, the agents receive an observation xt, takes
the action at, and receive a scalar reward rt. Generally, the
environment can be partially observable i.e we may need the
Multi-Objective Reward for Teams of Robotic Bodyguards
entire past history of observations, action pairs to represent
the current state such that st= (x1, a1, x2, a2, . . . , xt). For
our problem, we have assumed that the environment is fully
observable so we will represent st = xt
The behavior of each agent is represented by its own policy
π that takes the state st as an input and outputs a probability
distribution over all the actions i.e π (s) → P (A). Since
the environment is stochastic, we will model it as Markov
Decision Process with a state space S, action space A, re-
ward function R (st, at, st+1) and the transition dynamics
p(st+1st, at).
The return Gt from state s at timestep t is defined as the
discounted cummulative reward that the agent accumulates
starting from state s at timestep t and represented as Gt =
γi−tR (si, ai, si+1) where γ is the discounting factor
γ ∈ [0, 1]. The goal of the reinforcement learning is find
an optimum policy π∗ that maximizes the expected return
starting from state s. We denote the trajectory for state
visitation for the policy π as ρπ.
T(cid:88)
i=t
3.1. Policy Gradients
Policy gradient methods have been shown to learn the opti-
mal policy in variety of reinforcement learning tasks. The
main idea behind policy gradient methods is instead of pa-
rameterizing the Q-function to extract the policy, we param-
eterize the policy using the parameters θ to maximize the
objective which is represented as J (θ) = E [Gt] by taking a
step in the direction of ∇J (θ) where ∇J (θ) is defined as:
∇J (θ) = E [∇θ log πθ (as) Qπ (s, a)]
The policy gradient methods are prone to high variance
problem. Several different methods such as (Wu et al.,
2018; Schulman et al., 2017) have been shown to reduce
the variability in policy gradient methods by introducing a
critic which is a Q-function that tells about the goodness of
a reward by working as a baseline.
3.2. Deep Deterministic Policy Gradients
In (Silver et al., 2014) has shown that it is possible to extend
policy gradient framework to deterministic policies i.e. πθ :
S → A. In particular we can write ∇J (θ) as
∇J (θ) = E(cid:2)∇θπ (as)∇aQπ (s, a)a=π(s)
(cid:3)
Deep Deterministic Policy Gradients (Lillicrap et al., 2015)
is an off-policy algorithm and is a modification of the DPG
method introduced in (Silver et al., 2014) in which the pol-
icy π and the critic Qπ is approximated using deep neural
networks. DDPG also uses an experience replay buffer
alongside with a target network to stabilize the training.
3.3. Multi-Agent Deep Deterministic Policy Gradients
Multi-agent deep deterministic policy gradients (Lowe et al.,
2017) is the extension of the DDPG for the multi-agent
setting where each agent has it's own policy. The gradient
∇J (θi) of each policy is written as
∇J (θi) = E [∇θi log πi (aisi) Qπ
i (si, a1, . . . , aN )]
where Qπ
i (si, a1, . . . , aN ) is a centralized action-value
function that takes the actions of all the agents in addition
to the state of the agent to estimate the Q-value for agent
i. Since every agent has it's own Q-function, its allows the
agents to have different action space and reward functions.
The primary motivation behind MADDPG is that know-
ing all the actions of other agents makes the environment
stationary, even though their policy changes.
3.4. Policy Gradients
Policy gradient methods have been shown to learn the opti-
mal policy in variety of reinforcement learning tasks. The
main idea behind policy gradient methods is instead of pa-
rameterizing the Q-function to extract the policy, we param-
eterize the policy using the parameters θ to maximize the
objective which is represented as J (θ) = E [Gt] by taking a
step in the direction of ∇J (θ) where ∇J (θ) is defined as:
∇J (θ) = E [∇θ log πθ (as) Qπ (s, a)]
The policy gradient methods are prone to high variance
problem. Several different methods such as (Wu et al.,
2018; Schulman et al., 2017) have been shown to reduce
the variability in policy gradient methods by introducing a
critic which is a Q-function that tells about the goodness of
a reward by working as a baseline.
3.5. Deep Deterministic Policy Gradients
In (Silver et al., 2014) has shown that it is possible to extend
policy gradient framework to deterministic policies i.e. πθ :
S → A. In particular we can write ∇J (θ) as
∇J (θ) = E(cid:2)∇θπ (as)∇aQπ (s, a)a=π(s)
(cid:3)
Deep Deterministic Policy Gradients (Lillicrap et al., 2015)
is an off-policy algorithm and is a modification of the DPG
method introduced in (Silver et al., 2014) in which the pol-
icy π and the critic Qπ is approximated using deep neural
networks. DDPG also uses an experience replay buffer
alongside with a target network to stabilize the training.
3.6. Multi-Agent Deep Deterministic Policy Gradients
Multi-agent deep deterministic policy gradients (Lowe et al.,
2017) is the extension of the DDPG for the multi-agent
setting where each agent has it's own policy. The gradient
Multi-Objective Reward for Teams of Robotic Bodyguards
∇J (θi) of each policy is written as
∇J (θi) = E [∇θi log πi (aisi) Qπ
i (si, a1, . . . , aN )]
where Qπ
i (si, a1, . . . , aN ) is a centralized action-value
function that takes the actions of all the agents in addition
to the state of the agent to estimate the Q-value for agent
i. Since every agent has it's own Q-function, its allows the
agents to have different action space and reward functions.
The primary motivation behind MADDPG is that know-
ing all the actions of other agents makes the environment
stationary, even though their policy changes.
4. Problem Formulation
The setting we are considering for providing maximal phys-
ical protection to the VIP in crowded environment is a co-
operative Markov game which becomes a natural extension
of the single agent MDP for multi-agent systems. A multi-
agent MDP is defined as state space S that decribes all the
configurations of all the agents, an action space A that de-
scribes the action space of every agent A1, . . . ,AN . The
transitions are defined as T = S1 × A1 × . . . × SN × AN .
For each agent i, the reward function is defined as ri =
R (Si,Ai).
In this problem, we are assuming that all bodyguards have
same state space and they are following an identical policy.
For this problem we are considering a finite horizon problem
where each episode is terminated after T steps. Since this
problem is a cooperative setting, the goal of all the agents is
to find individual policies that increase the collected payoff.
4.1. The Environment Model
For the emergence of interesting behaviors in a multi-agent
setting, grounded communication in a physical environ-
ment is considered to be a crucial component. For per-
forming the experiments, we used Multi-Agent Particle
Environment (Mordatch & Abbeel, 2017) which is a two-
dimensional physically simulated environment in a discrete
time and continuous space. The environment consists of N
agent and M landmarks, both possessing physical attributes
such as location, velocity and size etc. Agents can act and
move independently with their own policies.
In addition to the ability of performing physical actions in
the environment, the agents also have the ability to utter
verbal symbols over the communication channel at every
timestamp. The utterances are symbolic in nature and does
not carry any meaning. At each timestamp, every agent
utters a categorical variable that is observed by every other
agent and it is up to the agents to infer a meaning of these
symbols during the training time. Every utterance carry a
small penalty and the agent can decide not to utter at every
timestamp. We denote the utterance which is a one-hot
vector by c.
The complete observation of the environment is given by
o = [x1,...N +M , c1,...N ] ∈ O. The state of each agent is
the physical state of all the entities in the environment and
verbal utterances of all the agents. Formally, the state of
each agent is defined as si = [xj,...N +M , ck,...N ] where xj
is the observation of the entity j from the perspective of
agent i and ck is the verbal utterance of the agent k.
4.2. Reward Functions
In (Bhatia et al., 2016) has defined a metric that quantifies
the threat to the VIP from each crowd member bi at each
timestep t. This metric can be extended to a reward function.
Since the threat level metric gives a probability. We can
conclude that when the distance between the VIP and the
crowd member is 0, the threat to the VIP is maximum, i.e
1, conversely when the distance between the VIP and the
crowd member bi is more than the safe distance, the threat
to the VIP is 0. We can model this phenomenon as an
exponential decay. Thus the fundamental reward function
can be defined as
Rt (B, VIP ) = −1 +
(1 − TL (VIP , bi))
(1)
where
i=1
TL (VIP , bi) = exp−A(Dist(VIP,bi))/B
(2)
In the following we will derive reward functions and ex-
plain the motivation behind them that were derived from
equation 1.
The baseline Threat-Only Reward Function penalizes each
agent with the threat to the VIP at each time step as men-
tioned in (Bhatia et al., 2016).
Rt (B, VIP ) = −1 +
(1 − TL (VIP , bi))
(3)
k(cid:89)
k(cid:89)
i=1
The Binary Threat Reward Function penalizes each agent
for the threat with a negative binary reward, in addition,
each agent is also penalized for not maintaining a suitable
distance from the VIP:
L (VIP ,B) =
if − 1 +
(1 − TL (VIP , bi)) (cid:54)= 0
otherwise
m ≤ (cid:107)xi − VIP(cid:107)2 ≤ d
otherwise
(4)
(5)
D (VIP , xi) =
−1
k(cid:89)
i=1
−1
0
0
Multi-Objective Reward for Teams of Robotic Bodyguards
where m is the minimum distance the bodyguard has to
maintain from VIP and d is the safe distance. The final
reward function is represented as
Rt (B, VIP , xi) = L (VIP ,B) + D (VIP , xi)
(6)
The Composite Reward Function is the composition of the
threat only reward function and the penalty for not maintain-
ing a suitable distance from the VIP
k(cid:89)
i=1
+ D (VIP , xi)
Rt (B, VIP , xi) = − 1 +
(1 − TL (VIP , bi))
(7)
The Communication Penalization Reward Function aug-
ments the composite reward by adding a small penalty p
every time the bodyguard performs an utterance, as recom-
mended in (Mordatch & Abbeel, 2017).
5. Experiments
We performed our experiments using the Multi-Agent Parti-
cle Environment (MPE) (Mordatch & Abbeel, 2017). The
performance was measured using the threat metric defined
in (Bhatia et al., 2016) over one episode. The experiments
were performed with 2-4 bodyguards ranging from 2-4, and
a constant number of 10 bystanders. For all of the experi-
ments, we have trained the agents for 10,000 episodes and
limiting the length of the episode to 25 steps.
Figure 1 shows examples of the resulting bodyguard behav-
ior for the composite reward function (left) and the threat
only reward function (right). Notice that for the threat only
behavior, the bodyguards are not in the close proximity of
the VIP - they have found ways to keep the threat low by
"attacking" the crowd.
Figure 1. The emerging collaborative bodyguard behavior, using
the composite reward function from Equation 7 (left) and the
threat-only function from Equation 3 (right). The VIP is brown,
bodyguards blue, bystanders red and landmarks grey.
Figure 2 shows the threat levels obtained by different reward
functions. The communication penalty function appears the
Figure 2. The overall threat level achieved by agents trained using
4 different reward functions (threat only, composite, binary threat
and communication penalty). The scenario involved 4 agents and
10 bystanders.
clear winner, with the lowest threat level obtained over the
course of the scenario.
Acknowledgement: This research was sponsored by the
Army Research Laboratory and was accomplished under
Cooperative Agreement Number W911NF-10-2-0016. The
views and conclusions contained in this document are those
of the authors and should not be interpreted as represent-
ing the official policies, either expressed or implied, of the
Army Research Laboratory or the U.S. Government. The
U.S. Government is authorized to reproduce and distribute
reprints for Government purposes notwithstanding any copy-
right notation herein.
References
Bhatia, T.S., Solmaz, G., Turgut, D., and Boloni, L. Con-
trolling the movement of robotic bodyguards for maximal
physical protection. In Proc of the 29th International
FLAIRS Conference, pp. 380 -- 385, May 2016.
Khan, S. A., Bhatia, T.S., Parker, S., and Boloni, L. Mod-
eling human-robot interaction for a market patrol task.
In Proc. of 25th International FLAIRS Conference, pp.
50 -- 55, May 2012.
Levine, Sergey, Finn, Chelsea, Darrell, Trevor, and Abbeel,
Pieter. End-to-end training of deep visuomotor policies.
Journal of Machine Learning Research, 17(1):1334 -- 1373,
January 2016.
Lillicrap, Timothy P., Hunt, Jonathan J., Pritzel, Alexander,
Heess, Nicolas, Erez, Tom, Tassa, Yuval, Silver, David,
and Wierstra, Daan. Continuous control with deep rein-
forcement learning. In Proceedings of the International
Conference on Learning Representations (ICLR), 2015.
Multi-Objective Reward for Teams of Robotic Bodyguards
Lowe, Ryan, Wu, Yi, Tamar, Aviv, Harb, Jean, Abbeel,
Pieter, and Mordatch, Igor. Multi-agent actor-critic for
mixed cooperative-competitive environments. Neural
Information Processing Systems (NIPS), 2017.
Mordatch, Igor and Abbeel, Pieter. Emergence of grounded
compositional language in multi-agent populations. arXiv
preprint arXiv:1703.04908, 2017.
Rahmatizadeh, Rouhollah, Abolghasemi, Pooya, Boloni,
Ladislau, and Levine, Sergey. Vision-based multi-task
manipulation for inexpensive robots using end-to-end
learning from demonstration. In International Conference
on Robotics and Automation (ICRA), 2018.
Richard Klima, Karl Tuyls, Frans Oliehoek. Markov se-
curity games: Learning in spatial security problems. In
The NIPS 2016 The Learning, Inference and Control of
Multi-Agent System, 2016.
Schulman, John, Wolski, Filip, Dhariwal, Prafulla, Radford,
Alec, and Klimov, Oleg. Proximal policy optimization
algorithms. CoRR, abs/1707.06347, 2017.
Silver, David, Lever, Guy, Heess, Nicolas, Degris, Thomas,
Wierstra, Daan, and Riedmiller, Martin. Deterministic
policy gradient algorithms. In Proceedings of the 31st
International Conference on Machine Learning, 2014.
Silver, David, Huang, Aja, Maddison, Christopher J., Guez,
Arthur, Sifre, Laurent, van den Driessche, George, Schrit-
twieser, Julian, Antonoglou, Ioannis, Panneershelvam,
Veda, Lanctot, Marc, Dieleman, Sander, Grewe, Dominik,
Nham, John, Kalchbrenner, Nal, Sutskever, Ilya, Lillicrap,
Timothy, Leach, Madeleine, Kavukcuoglu, Koray, Grae-
pel, Thore, and Hassabis, Demis. Mastering the game of
Go with deep neural networks and tree search. Nature,
529:484 -- 503, 2016.
Wu, Cathy, Rajeswaran, Aravind, Duan, Yan, Kumar,
Vikash, Bayen, Alexandre M, Kakade, Sham, Mordatch,
Igor, and Abbeel, Pieter. Variance reduction for policy
gradient with action-dependent factorized baselines. In
International Conference on Learning Representations,
2018.
Yasuyuki, S., Hirofumi, O., Tadashi, M., and Maya, H.
Cooperative capture by multi-agent using reinforcement
learning application for security patrol systems. In 2015
10th Asian Control Conference (ASCC), pp. 1 -- 6, May
2015.
|
1303.4692 | 1 | 1303 | 2013-03-15T15:58:26 | Crowd Simulation Modeling Applied to Emergency and Evacuation Simulations using Multi-Agent Systems | [
"cs.MA"
] | In recent years crowd modeling has become increasingly important both in the computer games industry and in emergency simulation. This paper discusses some aspects of what has been accomplished in this field, from social sciences to the computer implementation of modeling and simulation. Problem overview is described including some of the most common techniques used. Multi-Agent Systems is stated as the preferred approach for emergency evacuation simulations. A framework is proposed based on the work of Fangqin and Aizhu with extensions to include some BDI aspects. Future work includes expansion of the model's features and implementation of a prototype for validation of the propose methodology. | cs.MA | cs | |
1310.2290 | 1 | 1310 | 2013-10-08T21:56:33 | Modelling Complexity for Policy: Opportunities and Challenges | [
"cs.MA",
"cs.CY",
"nlin.AO",
"physics.soc-ph"
] | This chapter reviews the purpose and use of models from the field of complex systems and, in particular, the implications of trying to use models to understand or make decisions within complex situations, such as policy makers usually face. A discussion of the different dimensions one can formalise situations, the different purposes for models and the different kinds of relationship they can have with the policy making process, is followed by an examination of the compromises forced by the complexity of the target issues. Several modelling approaches from complexity science are briefly described, with notes as to their abilities and limitations. These approaches include system dynamics, network theory, information theory, cellular automata, and agent-based modelling. Some examples of policy models are presented and discussed in the context of the previous analysis. Finally we conclude by outlining some of the major pitfalls facing those wishing to use such models for policy evaluation. | cs.MA | cs | Modelling
for
Complexity
Opportunities
and
Challenges
Policy:
Bruce Edmonds, Manchester Metropolitan University
[email protected]
Carlos Gershenson, Universidad Nacional Autónoma de México
[email protected]
This chapter reviews the purpose and use of models from the field
of complex systems and, in particular, the implications of trying to
use models to understand or make decisions within complex
situations, such as policy makers usually face. A discussion of the
different dimensions one can formalise situations, the different
purposes for models and the different kinds of relationship they
can have with the policy making process, is followed by an
examination of the compromises forced by the complexity of the
target issues. Several modelling approaches from complexity
science are briefly described, with notes as to their abilities and
limitations. These approaches include system dynamics, network
theory, information theory, cellular automata, and agent-based
modelling. Some examples of policy models are presented and
discussed in the context of the previous analysis. Finally we
conclude by outlining some of the major pitfalls facing those
wishing to use such models for policy evaluation.
Introduction
For policy and decision-making, models can be an essential
component, as models allow the description of a situation, the
exploration of future scenarios, the valuation of different
outcomes and the establishment of possible explanations for what
is observed. The principle problem with this is the sheer
complexity of what is being modelled. A response to this is to use
more expressive modelling approaches, drawn from the “sciences
of complexity”—use more complex models to try and get a hold
on the complexity we face. However, this approach has potential
pitfalls as well as opportunities, and it is these that this chapter will
attempt to make clear. Thus, we hope to show that more complex
modelling approaches can be useful, but also to help people
“fooling themselves” in the process.
The chapter starts with an examination of the different kinds
of model that exist, so that these kinds might be clearly
distinguished and not confused. A section follows on the kinds of
uses to which such models can be put. Then we look at some of
the consequences of the fact that what we are modelling is
complex and the kinds of compromises this forces us into,
followed by some examples of models applied to policy issues. We
conclude by summarising some of
the key danger and
opportunities for using complex modelling for policy analysis.
Kinds
of
Model
A model is an abstraction of a phenomenon. A useful model has to be
simpler than the phenomenon but to capture the relevant aspects
of it. However, what needs to be represented in a model and what
can be safely left out is often a matter of great subtlety. Since
relevance changes with context, some models will be useful in
some circumstances and useless in others. Also, a model that is
useful for one purpose may well be useless for another. Many of
the problems associated with the use of models to aid the
formulation and steering of policy derive from an assumption that
a model will have value for a purpose or in a context different
form the one the model was established and validated for. In
other words, the value of a model is seen to be in its representation
(e.g. simulation code) and not in its social embedding.
Generally speaking, all of epistemology deals with models.
That is to say that all our descriptions can be thought of as models:
they are abstractions of the phenomena with which we deal. These
abstractions are required in order to understand and communicate
about the complex phenomena that we have to deal with.
However, models, in this most general sense, are not necessarily
either precise or formal. Indeed, most of the models we use in
everyday life are informal and couched in a language that is open
to a considerable degree of interpretation. Two dimensions of
formality can be distinguished:
a) the extent to which the referents of the representation are
constrained, e.g. by definition (“specificity of reference” or
S R ),
b) the extent to which the ways in which instantiations of the
representation can be manipulated are constrained, e.g. by
rules of logical deduction (“specificity of manipulation” or
S M ).
These two dimensions are illustrated in figure 1, below.
Figure 1. Two dimensions of formality
For example, an analogy expressed in natural language has a low
S R since, what its parts refer to are reconstructed by each hearer
in each situation. For example, the phrase “a tidal wave of crime”
implies that concerted and highly coordinated action is needed in
order to prevent people being engulfed, but the level of danger and
what (if anything) is necessary to do must be determined by each
listener. In contrast to this is a detailed description where what it
refers to is severely limited by its content, e.g. “Recorded burglaries in
London rose by 15% compared to the previous year”.
A system of abstract logic, mathematics or computer code has
high SM since the ways these can be manipulated is determined by
precise rules—what one person infers from them can be exactly
replicated by another. This in contrast to a piece of natural
language which can be used to draw inferences in many different
ways, only limited by the manipulators’ imagination and linguistic
ability. However, just because a representation has high SM does
not mean that the meaning of its parts in terms of what it
represents is well determined. Many simulations, for example, do
not represent anything we observe directly, but are rather
explorations of ideas. We, as intelligent interpreters, may mentally
fill in what it might refer to in any particular context, but these
“mappings” to reality are not well defined. Such models are more
in the nature of an analogy, albeit one in formal form – they are
not testable in a scientific manner since it is not clear as to
precisely what they represent. Thus simulations, especially agent-
based simulations, can give a false impression of their applicability
because they are readily interpretable (but informally). This does
not mean they are useless for all purposes. For example, Schelling’s
abstract simulation of racial segregation did not have any direct
referents in terms of anything measurable1, but it was an effective
1 Subsequent elaborations of this model have tried to make the relationship to what is observed
counter-example
that
that an assumption
that can show
segregation must be caused by strong racial prejudice was
unsound. Thus such ‘analogical models’ (those with low S R ) can
give useful insights, they can inform thought, but they can not give
reliable forecasts or explanations as to what is observed.
Formal models are a key aspect of science, since scientific
models aim at describing and understanding phenomena. Their
formality is important because that means that both their inference
and meaning is (a) checkable by others and (b) stable. This makes
it possible for a community of researchers and others to work
collectively with the same models, confident that they are not each
interpreting them in different ways. As the above discussion
should have made clear, they can be formal in (at least) two
different ways. For example data is a formal model of some aspect
of what we observe, in the sense that it abstracts but in a well
defined way – its meaning is precise. Data is not formal in terms
of S M however, and one could make very different inferences
from the same set of data. Usually “formal modelling” means that
the inference from a model is well specified, in other words it is a
representation with high S M . Thus scientific modelling is often
associated with mathematics or computer simulation. However, in
order to connect the formal inference to data it has to be formal in
the S R sense as well, there needs to be a precise mapping between
its parts and processes to what is observed, which is usually2 done
in terms of a map to some data.
The
Use
of
Models
There are many purposes for models, including: as a game, an
more direct, but the original model, however visually suggestive, was not related to any data.
2 It is possible to directly ‘wire’ something like a computational process to reality via sensors and
actuators, as happens in programmed trading, in this case it is not always clear the extent to which
the model is a representation of anything observed, but more an embedded participant in it.
aesthetic construction, or an illustration of some idea 3 . Most
scientific models claim to be predictive, i.e. they should allow us to
obtain information about the future of the phenomenon before it
occurs. For example, one can calculate and predict a ballistic
trajectory aiming at a target using a mechanical model. However,
on closer examination, many are more concerned with two other
goals: explanation or exploration, with that of prediction being left as
a theoretical possibility only4. There are many scientific models
that are not predictive, or which only predict abstract properties.
For example, the Gutenberg–Richter law describes the distribution
of earthquake intensities, but this does not tell us when might be
the next earthquake nor how intense it might be. Darwin’s theory
of evolution does not tell us what will evolve next, or even the
reasons why what has evolved did so, but it does predict the
relationship between genetic distance and the length of time since
species diverged. Unfortunately many reports about models are
not clear as to their purpose in this regard, indeed many seem to
deliberately conflate different purposes. Whilst models may have
more than one goal, one should be wary of a model that was
developed and tested for one purpose but is now being used for
another. A clear case of this is where a model is designed to
establish a theoretical counter-example (such as in the Schelling
case discussed below) but then is later claimed to be for prediction
(albeit in a modified form).
There is another, very basic, distinction in the way models are
used in practice. That is between models that (a) represent
something observed and ones which (b) are a component of an
adaptive strategy.
3 (Epstein, 2008) lists 16 different reasons, (Edmonds, et al. 2013) considers reasons that are more
connected with understanding human society.
4 It is common for papers describing them to list prediction as “future work” when the model is
more fully developed.
In the former case, there is a well-defined mapping between
the model and observational data/measurements, and the model is
judged as to the extent of its error in its predictions of its target
phenomena. Here the model is, to different degrees and ways,
either correct or not. In this case an examination of the model can
tell us something about the structure of what is modelled, for
example by exploring “what-if” questions using the model. For
sake of clarity we call this a “representational” model.
In the later case, the model is part of a decision making
process to select strategies for action. It takes (processed) inputs
from the world, for example indicators of success and the model is
changed depending on how well it is doing (for example by
depreciating the parts of the model that resulted in a poor
indicator). Outputs from the model are used in the determination
of interventions. Here the model is continually being adapted
according to events, it somehow encodes past successes or failures
for different courses of action given different observations of the
world. Here a useful model may not represent any aspect of the
world at all, but just be a useful intermediary in the process of
decision making. However, if the process of adaption is effective
it may come to encode knowledge as to what works and what does
not. We call this an “adaptive” model.
It may well be that an adaptive use of a model is more
effective in a particular setting, particular if a considerable period
of adaption has occurred, in effect training the model (given the
decision making structure it is embedded in) using a considerable
amount of feedback from its policy environment. If the model is
sufficiently flexible (i.e. has many adjustable internal parameters)
that it could indicate the correct action from the available inputs
(derived from observations of the environment) then, with enough
training, the model will eventually do so. However, this kind of
model adaption means that it is probably finely tuned to the
particular situation and will not be useful by others in similar
situations. Nor is it likely to be much use in exploring what would
happen in cases not yet observed; so if the situation changes in
some fundamental way, the model may well give totally the wrong
answers. Furthermore, it might not be apparent from an inspection
of the adapted model, why it works.
Representational models are usually hard to develop, taking
considerable time and effort, often by a team of experts somewhat
separate from those making policy decisions. Such models usually
rely on some theory of the system being modelled, whose
assumptions may be explicit. This kind of model, if it validates
well, might have some validity outside its original test situation,
and moreover, its assumptions and structure might give clues as to
when and how it might reliably be used. If the situation changes in
a way that is explicitly encoded into the model, one might be able
to change its settings to suit the new situation.
In practice, models are often used with a mixture of adaptive
and representational models, with adaptive models encapsulating
some
representational, and
somewhat
theory and being
representational models undergoing some process of model
adaption over time. In this case it is wise to know which aspects
of ones model are representational and which have been ‘tuned’ to
the particular situation or set of data.
Models are limited and using them carelessly can have
counterproductive consequences. One of the main limitations is
due to the complexity of their subject matter, which is what we
discuss next.
Complexity
and
its
Implications
for
Modelling
That society is complex may seem an obvious statement. Still, the
ways in which it is complex has implications for the use of models
for the planning and execution of policy. One problem is the lack
of agreement as to what “being complex” means. There are dozens
different definitions of complexity (Edmonds, 2000). Frequently
the word is used as a kind of negative. When available techniques
(or accepted techniques) fail we call what were trying to analyse,
“complex”—in this case it is a “dustbin concept”, a category to
use when others fail. Here, in order to obtain a common
understanding of the term, we can use its etymology. Complexity
comes from the Latin plexus, which means interwoven. Thus
something complex is difficult to separate out into separate
components or processes. This is because of relevant interactions
(Gershenson
and Heylighen, 2005; Gershenson, 2013a).
Interactions are relevant when the future of an element of a system
is partially determined by the interactions, in other words, if one
eliminates these interactions then the future would be significantly
different.
Traditionally, models have been reductionist, in the sense that
they study phenomena in isolation. By definition, interactions are
excluded. Either an element is modelled in isolation, or a whole
system is modelled, averaging the properties of its elements. In
other words, traditionally phenomena are modelled at a single
scale. This approach is suitable for simple systems, but it is not
sufficient for complex ones, where the properties of the system are
a consequence of the interactions of the elements. This requires
models to be multi-scale (Bar-Yam, 2004), and interactions must
be modelled to relate different scales.
In terms of policy models, interactions need to be included in
the model if the future projections are not to be distorted. Simple
models that do not include interactions will be unreliable. Such
interactions carry important implications for modelling complex
systems. These all make the ideal of the assessment of the impact
of policy interventions using a model difficult.
Firstly, it implies that elements cannot be studied in isolation.
Different social processes can interact to produce effects different
from those caused by each process singly. The outcome from a
population that is both disaffected and has access to an effective
medium for dissemination of views (e.g. twitter) might well be very
different to that of a disaffected population with only local gossip
or a satisfied population with something like twitter. The impact
of this is that separately analysing the impacts of different factors
upon the outcome might well be misleading; one has to consider
the outcomes from the whole system. This leads to the problem
that one might well not know how much one needs to include in
an model adequate for ones purposes, and that an approach that
starts with the simplest possible model and then experimentally
adds processes one at a time, might never get you to an adequate
model (Edmonds & Moss, 2005).
Secondly, it implies that interactions generate novel information,
which is not present in initial or boundary conditions. This new
information inherently limits the predictability of a complex
system. In other words, the results are at least partially ‘caused’ by
processes within the system, and not by external factors that can
be controlled for. At best, this may mean that one has to make do
with a broad distribution of outcomes as a forecast, or the
prediction of ‘weaker’, second-order properties of the outcome
(e.g. the volatility of the focus outcomes, or what will not happen).
At worst, it may mean that there is no well-defined distribution of
outcomes at all, with any measures upon the outcomes from a
model being due to artefacts (e.g. model size).
The result of such difficulties means that any policy model is
inevitably a compromise between different desirable modelling
goals. Figure 2 below illustrates some of these tensions in a simple
way.
Figure 2. Some of the tensions implicit in modelling complex systems.
These illustrated desiderata all refer to the model that is being
used. Simplicity is how simple the model is, the extent to which the
model itself can be completely understood. Analytically solvable
mathematical models, most statistical models and abstract
simulation models are at the relatively simple end of the spectrum.
Clearly a simple model has many advantages in terms of using the
model, checking it for bugs and mistakes (Galan, et al. 2009) and
communicating it. However, when modelling complex systems,
such as those policy makers face, such simplicity may not be worth
it if gaining it means a loss of other desirable properties. Generality
is the extent of the model scope: how many different kinds of
situation could the model be usefully applied. Clearly some level of
generality is desirable; otherwise one could only apply the model in
a single situation. Authors are often rather lax about making the
scope of their models clear—often implying a greater level of
generality that can be substantiated. Formality is what was called
specificity of reference (SR) above. Models where the meaning of
the model parts are well-defined have such formality, those which
do not, and are more in the way of a model of ideas about some
target system, rather than the system directly have less of this.
Finally validity means the extent to which the model outcomes
match what is observed to occur—it is what is established in the
process of model validation. This might be as close a match as a
point forecast, or as loose as projecting qualitative aspects of
possible outcomes.
What policy makers want, above all, is validity, with generality
(so they do not have to keep going back to the modellers) and
simplicity (so there is an accessible narrative to build support for
any associated policy) coming after this. Formality is for them is
not a virtue but more of a problem, they may be convinced it is
necessary (so as to provide the backing of ‘science’), but it means
that the model is inevitably somewhat opaque to them and not
entirely under their control. Modellers, usually, have very different
priorities. Formality is very important to them so that they can
replicate their results and so that the model can be unambiguously
passed to other researchers for examination, critique and further
development (Edmonds 2000). Simplicity and generality are nice if
you can get them, but one cannot assume that these are achievable
(Edmonds 2012). Validity should be an overwhelming priority for
modellers; otherwise they are not doing any sort of empirical
science. However, they often put this off into the future, preferring
the attractions of the apparent generality offered by analogical
models (Edmonds 2001, Edmonds 2012). Relatively simple
models that explore ideas rather than relate to any observed data
that give the illusion of generality are, unfortunately, common.
Another ramification of the complexity of what is being
modelled is in the goal of modelling—what sort of purpose the
model suitable for. One of the consequences of this is that
prediction of policy matters is hard, rare, and only obtained as a
result of the most specific and pragmatic kind of modelling
developed over relatively long periods of time5. It is more likely
that a model is appropriate for establishing and understanding
candidate explanations of what is happening, which will inform
policy making in a less exact manner than prediction, being part of
the mix of factors that a policy maker will take into account when
deciding action. It is common for policy people to want a
prediction of the impact of possible interventions “however
rough”, rather than settle for some level of understanding of what
is happening, however this can be illusory. If one really wanted a
prediction “however rough” one would settle for a random
prediction6 dressed up as a complicated “black box” model. If we
are wiser, we should accept the complexity of what we are dealing
and reject models that give us ill-founded predictions.
One feature of complex systems is that they can result in
completely unexpected outcomes, where due to the relevant
interactions in the system, a new kind of process has developed
resulting in qualitatively different results. It is for this reason that
complex models of these systems do not give probabilities (since
these may be meaningless, or worse be downright misleading) but
rather trace some (but not all) of the possible outcomes. This is
useful as one can then be as prepared as possible for such
outcomes, which otherwise would not have been thought of.
The effective use of models for policy formulation will thus
involve a clear focus as to its purpose and its manner of use
combined with some compromise between the factors discussed
above. However, the extent and impact of such compromises
should be openly and honestly made, as a proper balance is
necessary for reliable uses of the model. It is probable that a
each making different
related models,
combination of
5 For an account of actual forecasting and its reality, see Silver (2012).
6 Or other null model, such as “what happened last time” or “no change”.
compromises might be a productive way forward. However, this
requires extra work and care.
We now
look at a number of different approaches,
commenting upon the compromises and properties of each.
Tools
and
Approaches
System
Dynamics
System dynamics is an approach to modelling that represents a
system in terms of a set of interconnected feedback loops
(Forrester, 1971). It models these in terms of a series of flows
between stocks plus additional connections between variables and
flows. Crucially, it allows the representation of delays in such
feedback and
that
the outcomes of some variables can
control/effect the rate of other flows.
These flows and
relationships can then be simulated on a computer and (more
recently) visualized. Its advantages are that a complex set of
feedback relationships can be explored and hence better
understood. However, in practice, the variables it deals with are
themselves abstract entities, often representing abstract and
aggregate quantities. This approach is not well suited to the
modelling of systems where internal heterogeneity is significant in
terms of determining the outcomes.
Network
Theory
Networks naturally describe complex systems, representing
elements as nodes and their interactions explicitly as links. Only in
the last decade, there has been an explosion in the scientific
exploration of networks and their application to a broad range of
domains. Network theory has its roots in graph theory as
proposed by Euler in the eighteenth century. However, it is only
recently that its use has become widespread, in part because of the
large computing power and big data sets available.
Networks are useful for representing the structure of systems,
indicating how elements interact. However, they can also represent
the function of systems, with nodes representing states and directed
links representing transitions. Relating the structure and function
of systems
is one of
the most common questions for
understanding systems, i.e. how changes in the structure affect the
function of a system? Network theory can be used to study both
structure and function using the same formalism. Also, adaptive
and temporal networks (Gross & Sayama, 2009; Holme &
Saramäki, 2012) have been used to study the change in time of
network structure.
From the study of different natural and artificial networks, it
has been found that most of them do not have a trivial topology,
i.e. there is a relevant organization in their structure. Still, several
modelling approaches assume homogeneous topologies, as in
cellular automata (see below), or even a so called “well mixed”
population, i.e. there is no structure considered (only the macro
state). It has been shown that structure (micro scale) plays a crucial
role in the dynamics of such systems. For example, the same
system may change drastically its dynamics depending on whether
the local structure is considered or not (Shnerb et al., 2000).
Network models can be useful to study several aspects related
to policy and decision making. For example, random agent
networks (RANs) were proposed to model organizations such as
bureaucracies (Gershenson, 2008), showing how few modifications
to the structure of an organization can improve considerably its
performance. In general, “computing networks” (Gershenson,
2010) can be used to study and relate adaptability at different
scales. Since policy and decisions are usually made over changing
and uncertain scenarios, adaptability is a desired property of
models. However, the more networks change and the complexity
of the interactions represented over the links get, the less classic
network theory is applicable, and the closer to an individual-based
model one has.
Information
theory
Claude Shannon (1948) proposed information theory in the
context of telecommunications. He was interested on how a
message could be transmitted reliably over unreliable media. He
proposed a measure of information (equivalent to the Boltzmann-
Gibbs entropy in thermodynamics) where information is minimal
for regular strings, as new symbols do not carry new information.
Shannon information is maximal for random strings, as new
symbols carry all the new information, i.e. they are not predictable.
Several other measures have been derived from Shannon's
information, such as mutual information, predictive information,
excess entropy, and
information
transfer, among others
(Prokopenko et al., 2009).
Information theory has been used repeatedly to measure
complexity. However, there are two different views. One view
implies a similarity of information to complexity, where maximum
randomness
(Shannon
information) would have maximum
complexity. A more popular view poses that complexity is maximal
when a balance between regularity (order) and randomness (chaos)
is reached (Langton, 1990; Kauffman, 1993).
self-
emergence,
complexity,
Recently, measures of
organization, homeostasis, and autopoiesis were proposed based
on information theory (Fernández et al., 2014). These measures are
fast to compute and simple enough to be used by people without a
strong mathematical background, but can give insights into the
dynamics of systems. It has been argued (Edmonds 1999) that
there is not one such measure that can always be used, but rather
one has to choose a measure that gives meaningful results for the
kind of system that one is considering. Thus this approach
assumes that one has understood the target system sufficiently to
select the appropriate measure.
For decision making, it is vital to identify which type of
dynamics are followed by systems and their components, as
different decisions should be made depending on regular, complex,
or chaotic dynamics. Used correctly, these measures can provide
precisely this information and thus aid in knowing how to respond
to change in the systems.
Cellular
automata
Cellular automata (CA) can be seen as a particular type of network.
Each cell (node) has a state that depends on the states of its
neighbours (links) and its own previous state. Different CA models
can have different number of states and consider different number
of neighbours. Cells can also be arranged in one dimension (array),
two dimensions (lattice), three, or more dimensions.
Perhaps the most popular CA is Conway’s “Game of Life”
(Berkelamp, et al., 1982). Each cell can have one of two states: ‘0’
(dead) or ‘1’ (alive). Rules consider how many of the eight closest
neighbours are alive. For a live cell to continue living, it must have
two or three living neighbours (in any configuration). More than
three or less than two neighbours implies that the cell will die in
the next time step. New cells are born on empty cells when they
have exactly three neighbours. With these simple rules, several
complex structures emerge: stable structures of different shapes,
oscillators of different periods, moving structures
(gliders,
spaceships), eaters, glider guns, etc. The structures emerging with
the simple rules of the Game of Life can be used even to build a
Universal Turing Machine7. An example of the dynamics of the
Game of Life is shown in Figure 3.
7 To explore the Game of Life and other interesting CA, the reader is advised to download Golly at
http://golly.sourceforge.net
a
b
d
c
Figure 3. Evolution of the Game of Life from a random initial condition (a), where white cells are
“alive” and black cells are “dead”. After 410 steps (b), certain stable structures have been formed,
but there are still some active zones. After 861 steps (c), some structures have been destroyed and
some new ones have been created. Activity continues in the lower part of the lattice. After 1416
steps (d), the dynamics is periodic, with stable and oscillatory structures. Images created with
NetLogo (Wilensky, 1999). Figure initially published in Gershenson (2013b).
Cellular automata have been used in several urban and land-
use models (e.g. Portugali, 2000; Batty, 2005). However, “pure”
CA models tend to be too formal and abstract, so they have been
found to be more useful in hybrid models, in many cases
combining CA with agent-based modelling.
Individidual-‐
and
Agent-‐based
modelling
Individual-based modelling is given when social actors or entities
are represented by separate ‘objects’ within a computational
simulation. Each object can have different properties, so this
technique can represent heterogeneous collections of individuals.
The interactions between the actors are represented by messages
between the objects of the simulation. Thus, the mapping
between what is observed and the model can be very much more
straightforward with such simulations: each object modelling its
corresponding actor.
When the objects in the simulation have internal processes
representing their learning or decision making processes so that
these processes could be usefully interpreted as cognition, we call
the computational objects “agents” since they can act somewhat
independently—they have a simple form of agency. When the
agents are of this form, one has the technique of Agent-Based
Modelling (ABM). This technique is very flexible and puts few
constraints upon the modeller, so the simulations that result are
difficult to characterize in general but are of various kinds. An
accessible introduction to the approach is (Gilbert & Troitzsch,
2005) and a more comprehensive guide (Edmonds & Meyer,
2013).
In particular, simulations differ greatly as to their level of
detail, ranging from highly abstract and relatively simple
simulations to very specific and complicated ones. The key
difference here is whether the driver for model development is
simplicity or relevance (Edmonds & Moss, 2005). Agent-based
modelling has now been applied to a large number of policy-
relevant subjects, including (to take an arbitrary sample of recent
applications): energy infrastructure siting (Abdollahian, et al.,
2013), password behaviours within an organization (Renaud, et al.,
2013), mobile banking adoption (Wei, et al., 2013), China’s housing
market (Zhang, et al., 2013) and return migration (Biondo, et al.
2013).
The problem with ABM lies not in its expressiveness but in
the complexity of its models (which means that it may be hard to
understand
the models
themselves) and establishing
the
relationship between the models and what they represent (Moss &
Edmonds 2005).
Examples
Club
of
Rome’s
“Limits
to
Growth”
In the early 1970's, on behalf of an international group under the
name “The Club of Rome” a simulation study was published
(Meadows et al. 1972) with the attempt to convince humankind
that there were some serious issues facing it, in terms of a coming
population, resource, and pollution catastrophe. To do this they
developed a system dynamics model of the world. Thus, this is a
fairly simple kind of model that does not explicitly represent the
parts of a system or its interactions, but rather the feedback cycles
between key global factors. It was important to the authors to go
beyond simple statistical projections of the available data, since
that missed out the crucial delays in the usually self-correcting
feedback processes. The results of the simulations were a set of
computed curves showing such as pollution, population, etc. The
results indicated that there was a coming critical point in time and
that a lot of suffering would result, even if humankind managed to
survive it.
The book had a considerable impact, firmly establishing the
idea that it was possible that humankind could not continue to
grow
indefinitely. The book presented the results of the
simulations as predictions—a series of what-if scenarios. Whilst
they did add caveats and explore various possible versions of their
model, the overall intent of the book was unmistakable: that if we
did not change our lifestyles, disaster would result.
The authors clearly hoped that by using a simulation they
would be able to make the potential feedback loops real to people.
Thus this was a use of simulation to illustrate an understanding
that the authors had. It was thus a model of ideas rather than
directly of any such data. It did not, and could not, make
predictions about what will happen in the future, but rather
illustrate some possibilities. However, the model was not presented
as such, but as something more scientific in some sense. It was the
presentation as ‘scientific’ that made this book such a challenge but
also what laid it open to criticism (e.g. Cole, et al., 1973). An
examination of the model showed that some of its parameters
were very sensitive and thus had to be ‘tuned’ to get the published
results (Vermeulen & de Jongh, 1976). In other words, whilst the
models had an illustrative and exploratory purpose, they were
presented and criticised as if it was a predictive model.
The book made a considerable impact upon the general
consciousness of the problem, and did act to get people
questioning previously held assumptions (that we could keep on
growing economically and physically). However, it was also largely
discredited in the eyes of other modellers due to its perceived lack
of ‘rigor’. This was somewhat unfair as the alternatives were no
better in terms of validity or generality. However, a lack of
humility in terms of its results and the relative simplicity of their
model did lay it open to such attacks. The predictions of the book
have not yet come to pass, but it is not clear that a similar future
critical point and attendant suffering has been avoided.
Schelling’s
Model
of
Racial
Segregation
In addition to a host of simpler, analytically expressed models
(similar in kind to the Club of Rome’s systems dynamics models),
Schelling developed what we might recognize as a simple agent-
based model (Schelling, 1971; 1978). It did represent individuals
and their neighbourhoods explicitly, albeit abstractly in a 2D grid
with black and white ‘counters’ representing the people. The
simulation was very simple—counters were distributed randomly
to start with then each counter that had less than a given
percentage (c) of like neighbours moved to a new empty spot. The
simulation showed that segregation emerged even with relatively
low levels of racial bias (values of c down to 30%). This did not
relate to any particular data but was rather a counter-example to
the idea that the observed segregation must be due to strong racial
prejudice. In other words this was intended to be an exploration of
ideas to inform policy rather than a direct representation of what
was happening. It produced an understanding of possible
segregation processes, and so influenced local policies in Chicago,
away from focusing on prejudice as a cause of the extreme
segregation they had.
Employment
in
an
Arctic
Community
Berman et al. (2004) consider eight employment scenarios defined
by different policies for tourism and government spending, as well
as different climate futures, for an ABM case study of sustainability
in the Arctic community of Old Crow, in Canada. Scenarios were
developed with the input of local residents: tourism being a policy
option largely influenced by the autonomous community of Old
Crow (stemming from their land rights), and attracting great local
interest. Here the policy options were addressed as a certain type
of scenario, embedding the behaviour of actors within a few
possible future contexts. The simulation here ensured the
consistency of the scenarios, and helped to integrate the various
inputs into a coherent whole.
The merit of this model is that it can improve the reckoning of
human and social factors and information into the issues at stake;
allowing the exploration of some real possible outcomes. The
drawback is the multiplication of uncertainties, not least of which
is that we do not convincingly know how social actors might adapt
to new circumstances (even if the policy options are relatively
concrete).
A
detailed
model
of
HIV
spread
and
social
structure
Alam et al. (2007) investigate the outcomes indicated by a
complex, and detailed model of a particular village in the Limpopo
valley of South Africa. This model in particular looks at many
aspects of the situation, including: social network, family structure,
sexual network, HIV spread, death, birth, savings clubs,
government grants and local employment prospects. It concludes
with hypotheses about this particular case, showing that complex
destructive synergies between the spread of HIV and the
breakdown of social structure were possible, and could be
exacerbated by the influx of workers from outside due to the
granting of mining concessions. This does not mean that these
outcomes will actually occur, but this does provide a focus for
future field research and may provide thought for policy makers.
Unfortunately in this case, the conclusions of this study were not
what the local authority wanted to hear, and so the findings were
ignored. This was a model with a high degree of validation, but a
very specific and complex model taking 3 years to develop.
Evaluating
Pandemic
Preventive
Measures
Bajardi et al. (2011) combined in a model networks and agents to
model the epidemic spread of the H1N1 influenza in 2009. Nodes
represent regions, which are linked by the commercial flights
between their major airports. At each node, agents can be
susceptible, exposed, infected, or recovered (Anderson & May,
1992). Adjusting different parameters, the global spread of the
disease could be reproduced. Travel restrictions were imposed as a
preventive measure; reducing air travels by 40%. Simulations
showed that travel restrictions are ineffective to prevent the spread
of the disease. Comparing with scenarios with no travel restrictions
or with even more stringent travel restrictions, the authors found
that the disease reached a peak almost on the same day.
This is a theory-based model intended for predictive purposes.
Its validity depends upon the approximations and assumptions in
the model, including the characteristics of the social network used
in the model.
Prospects
and
Dangers
for
Complex
Policy
Modelling
All models have limits—not only limits in the accuracy of their
predictions but in the expression of the situations under which
such projections are based. The nature of complex situations
means that attempts to use models to aid policy formulation are
susceptible to some particular dangers and pitfalls, including:
• Confusing a model that has exploratory or explanatory
purposes for one that is predictive.
• Preferring a ‘black box’ model that seems to give definite
predictions despite neither understanding it nor knowing
that it is reliable for this purpose.
• Trying to use a model that has been adapted within a highly
specific situation out of its original context.
• Attempting very general or simple models of policy issues
probably means sacrificing direct validity for an indirect,
analogical relationship only.
Complex models of the kind described here (and elsewhere in
this book) have the potential to express a broader range of kinds
of situations than previous approaches. They are thus not so
limited by the kind of ‘brave’ assumptions that bedevil models
where analytic results are deemed necessary. They are also ideal
for the exploration and “laying bare” complex dynamics. For this
reason, they are prospective as an important tool in the exploration
and consideration of policy options. In particular, more descriptive
models can be directly related to what they are modelling, allowing
a greater range of data and input to be utilized in their specification
and validation (both high SM and SR in terms of the above
discussion).
Given the fact that models of complex systems will offer a
limited predictability, it is advisable to complement this lack of
predictability with adaptability (Gershenson, 2007). This will
enable decision makers to take the best choice for the specific
circumstances that are faced, as adaptability implies a distinction of
current circumstances that purely predictive models do not
consider.
Acknowledgements
Bruce Edmonds acknowledges the support of the EPSRC under
grant number EP/H02171X/1. Carlos Gershenson was partially
supported by SNI membership 47907 of CONACyT, Mexico.
References
Abdollahian, M., Yang, Z. & Nelson, H. (2013) 'Techno-Social Energy
Infrastructure Siting: Sustainable Energy Modeling Programming
(SEMPro)' Journal of Artificial Societies and Social Simulation 16(3):6
<http://jasss.soc.surrey.ac.uk/16/3/6.html>.
Alam, S.J., Meyer, R., Ziervogel, G. & Moss, S. (2007) The Impact of
HIV/AIDS in the Context of Socioeconomic Stressors: an Evidence-
Driven Approach. Journal of Artificial Societies and Social Simulation, 10(4):7
<http://jasss.soc.surrey.ac.uk/10/4/7.html>.
Anderson RM, May RM (1992) Infectious Diseases of Humans: Dynamics and
Control. Oxford: Oxford University Press.
Bajardi P, Poletto C, Ramasco JJ, Tizzoni M, Colizza V & Vespignani A
(2011). Human Mobility Networks, Travel Restrictions, and the Global
Spread of 2009 H1N1 Pandemic. PLoS ONE 6(1): e16591.
Bar-Yam, Y. (2004). Multiscale variety in complex systems. Complexity,
9(4):37–45.
Batty, M. (2005). Cities and complexity. MIT Press, Cambridge, MA, USA.
Berlekamp, E. R., Conway, J. H., and Guy, R. K. (1982). Winning Ways for
Your Mathematical Plays, volume 2: Games in Particular. Academic Press,
London.
Berman, M., Nicolson, C., Kofinas, G., Tetlichi, J. & Martin, S. (2004)
Adaptation and sustainability in a small arctic community: Results of an
agent-based simulation model. Arctic, 57(4), 401-414.
Cole, H.S.D., Freeman, C., Jahoda, M. & Pavitt, K.L. (Eds.) (1973) Models of
Doom: A Critique of the Limits to Growth. New York: Universe Books.
Edmonds, B. (1999). Syntactic Measures of Complexity. PhD Thesis, Centre for
Policy Modelling, MMU. http://bruce.edmonds.name/thesis/
Edmonds, B. (2000). Complexity and scientific modelling. Foundations of
Science, 5:379– 390.
Edmonds, B. (2001) The Use of Models - making MABS actually work. In
S. Moss & P. Davidsson (Eds.), Multi Agent Based Simulation (Lecture Notes
in Artificial Intelligence, 1979). Berlin, Springer: 15-32.
Edmonds, B. (2012) Complexity and Context-dependency. Foundations of
Science. DOI 10.1007/s10699-012-9303-x
Edmonds, B. & Meyer, R. (eds.) (2013) Simulating Social Complexity - A
Handbook. Springer.
Edmonds, B. & Moss, S. (2005) From KISS to KIDS – an ‘anti-simplistic’
modelling approach. In P. Davidsson et al. (Eds.), Multi Agent Based
Simulation 2004 (Lecture Notes in Artificial Intelligence, 3415). Berlin:
Springer: 130–144.
Edmonds, B., Lucas, P., Rouchier, J. & Taylor, R. (2013) Human Societies:
Understanding Observed Social Phenomena. In Edmonds, B. & Meyer, R.
(eds.) Simulating Social Complexity - A Handbook. Springer, 709-749.
Epstein, J. M. (2008) Why Model? Journal of Artificial Societies and Social
Simulation 11(4):12. <http://jasss.soc.surrey.ac.uk/11/4/12.html>
Fernández, N., Maldonado, C., & Gershenson, C. (2014). Information
measures of complexity, emergence, self-organization, homeostasis, and
autopoiesis. In Prokopenko, M., (ed.) Guided Self-Organization: Inception.
Springer. In Press.
Forrester, J. (1971). Counterintuitive behavior of social systems.
Technology Review 73(3):52–68.
Galán, J.M et al. (2009). Errors and Artefacts in Agent-Based Modelling.
Journal of Artificial Societies and Social Simulation, 12(1),
<http://jasss.soc.surrey.ac.uk/12/1/1.html>.
Gershenson, C. (2007). Design and Control of Self-organizing Systems. CopIt
Arxives, Mexico. http://tinyurl.com/DCSOS2007.
Gershenson, C. (2008). Towards self-organizing bureaucracies. International
Journal of Public Information Systems, 2008(1):1–24.
Gershenson, C. (2010). Computing networks: A general framework to
contrast neural and swarm cognitions. Paladyn, Journal of Behavioral Robotics,
1(2):147–153.
Gershenson, C. (2013a). The implications of interactions for science and
philosophy. Foundations of Science, Early View.
Gershenson, C. (2013b). Facing complexity: Prediction vs. adaptation. In
Massip, A. and Bastardas, A., editors, Complexity Perspectives on Language,
Communication and Society, pages 3–14. Springer, Berlin Heidelberg.
Gershenson, C. & Heylighen, F. (2005). How can we think the complex? In
Richardson, K., (ed.) Managing Organizational Complexity: Philosophy, Theory
and Application, chapter 3, pages 47–61. Information Age Publishing.
Gilbert, N. & Troitzsch, K. (2005) Simulation for the Social Scientist (2nd ed.).
Open University Press.
Gross, T. and Sayama, H., (eds.) (2009). Adaptive networks: Theory, Models and
Applications. Understanding Complex Systems. Springer, Berlin Heidelberg.
Holme, P. & Saramäki, J. (2012). Temporal networks. Physics Reports,
519(3):97 – 125.
Kauffman, S. A. (1993). The Origins of Order. Oxford University Press.
Langton, C. (1990). Computation at the edge of chaos: Phase transitions
and emergent computation. Physica D, 42:12–37.
Meadows, D.H, Meadows, D., Randers, J. & Behrens, W.W.III (1972) The
Limits to Growth: A Report for the Club of Rome’s Project on the Predicament of
Mankind. New York: Universe Books.
Moss, S. and Edmonds, B. (2005) Sociology and Simulation: - Statistical
and Qualitative Cross-Validation, American Journal of Sociology, 110(4) 1095-
1131.
Prokopenko, M., Boschetti, F., & Ryan, A. J. (2009). An information-
theoretic primer on complexity, self-organisation and emergence.
Complexity, 15(1):11–28.
Portugali, J. (2000). Self-organization and the City. Springer Verlag.
Renaud, K. & Mackenzie, L. (2013) 'SimPass: Quantifying the Impact of
Password Behaviours and Policy Directives on an Organisation's Systems'
Journal of Artificial Societies and Social Simulation 16 (3) 3
<http://jasss.soc.surrey.ac.uk/16/3/3.html>.
Schelling, T. C. (1971) Dynamic Models of Segregation, Journal of
Mathematical Sociology, 1(2):143-186.
Schelling, T. C. (1978). Micromotives and Macrobehavior, Norton.
Shannon, C. E. (1948). A mathematical theory of communication. Bell
System Technical Journal, 27:379–423 and 623–656.
Shnerb, N. M., Louzoun, Y., Bettelheim, E., & Solomon, S. (2000). The
importance of being discrete: Life always wins on the surface. Proceedings of
the National Academy of Sciences, 97(19):10322–10324.
Vermeulen, P.J. & de Jongh, D.C.J (1976) Parameter sensitivity of the
‘Limits to Growth’ world model. Applied Mathematical Modelling, 1(1):29-32.
Wei, X., Hu, B. & Carley, K. M. (2013) 'Combination of Empirical Study
with Qualitative Simulation for Optimization Problem in Mobile Banking
Adoption' Journal of Artificial Societies and Social Simulation 16(3):10
<http://jasss.soc.surrey.ac.uk/16/3/10.html>.
Zhang, H., Wang, Y., Lin, Y., Zhang, Y. & Seiler, M. J. (2013) 'Simulation
Analysis of the Blocking Effect of Transaction Costs in China's Housing
Market' Journal of Artificial Societies and Social Simulation 16(3):8
<http://jasss.soc.surrey.ac.uk/16/3/8.html>.
Biondo, A. Emanuele, P., Alessandro & Rapisarda, A. (2013) 'Return
Migration After Brain Drain: A Simulation Approach' Journal of Artificial
Societies and Social Simulation 16(2):11
<http://jasss.soc.surrey.ac.uk/16/2/11.html>.
Silver, N. (2013) The Signal and the Noise – Why Most Predictions Fail – but
Some Don't. Penguin.
|
1711.03948 | 1 | 1711 | 2017-11-10T18:21:05 | Manipulative Elicitation -- A New Attack on Elections with Incomplete Preferences | [
"cs.MA",
"cs.DS",
"cs.GT"
] | Lu and Boutilier proposed a novel approach based on "minimax regret" to use classical score based voting rules in the setting where preferences can be any partial (instead of complete) orders over the set of alternatives. We show here that such an approach is vulnerable to a new kind of manipulation which was not present in the classical (where preferences are complete orders) world of voting. We call this attack "manipulative elicitation." More specifically, it may be possible to (partially) elicit the preferences of the agents in a way that makes some distinguished alternative win the election who may not be a winner if we elicit every preference completely. More alarmingly, we show that the related computational task is polynomial time solvable for a large class of voting rules which includes all scoring rules, maximin, Copeland$^\alpha$ for every $\alpha\in[0,1]$, simplified Bucklin voting rules, etc. We then show that introducing a parameter per pair of alternatives which specifies the minimum number of partial preferences where this pair of alternatives must be comparable makes the related computational task of manipulative elicitation \NPC for all common voting rules including a class of scoring rules which includes the plurality, $k$-approval, $k$-veto, veto, and Borda voting rules, maximin, Copeland$^\alpha$ for every $\alpha\in[0,1]$, and simplified Bucklin voting rules. Hence, in this work, we discover a fundamental vulnerability in using minimax regret based approach in partial preferential setting and propose a novel way to tackle it. | cs.MA | cs |
Manipulative Elicitation – A New Attack on Elections
with Incomplete Preferences
Palash Dey
[email protected]
Tata Institute of Fundamental Research, Mumbai
Monday 13th November, 2017
Abstract
Lu and Boutilier [LB11] proposed a novel approach based on "minimax regret" to use
classical score based voting rules in the setting where preferences can be any partial (in-
stead of complete) orders over the set of alternatives. We show here that such an approach
is vulnerable to a new kind of manipulation which was not present in the classical (where
preferences are complete orders) world of voting. We call this attack "manipulative elicita-
tion." More specifically, it may be possible to (partially) elicit the preferences of the agents in
a way that makes some distinguished alternative win the election who may not be a winner
if we elicit every preference completely. More alarmingly, we show that the related com-
putational task is polynomial time solvable for a large class of voting rules which includes
all scoring rules, maximin, Copelandα for every α ∈ [0, 1], simplified Bucklin voting rules,
etc. We then show that introducing a parameter per pair of alternatives which specifies the
minimum number of partial preferences where this pair of alternatives must be comparable
makes the related computational task of manipulative elicitation NP-complete for all com-
mon voting rules including a class of scoring rules which includes the plurality, k-approval,
k-veto, veto, and Borda voting rules, maximin, Copelandα for every α ∈ [0, 1], and simplified
Bucklin voting rules. Hence, in this work, we discover a fundamental vulnerability in using
minimax regret based approach in partial preferential setting and propose a novel way to
tackle it.
1 Introduction
Aggregating preferences of a set of agents over a set of alternatives is a fundamental problem in
voting theory which has been used in many applications in AI for making various decisions.
Prominent examples of such applications include collaborative filtering [PHG00], similarity
search [FKS03], winner determination in sports competitions [BBN14], etc. [MBC+16].
In
a typical scenario of voting, we have a set of alternatives, a tuple of "preferences", called a pro-
file, over the set of alternatives, and a voting rule which chooses a set of alternatives as winners
based on the profile. Classically, preferences are often modeled as complete orders over the
set of alternatives. However, in typical applications of voting in AI, collaborative filtering for
example, the number of alternatives is huge and we have only partial orders over the set of
alternatives as preferences.
There have been many attempts to extend the use of voting theory in settings with incom-
plete preferences. The approach of Konczak and Lang [KL05] was to study the possible and
necessary winner problems.
In these problems, the input is a profile of partial preferences
and we want to compute the set of alternatives who wins (under some fixed voting rule) in at
1
least one completion of the profile for the possible winner problem; for the necessary winner
problem, we want to compute the set of alternatives who wins in every completion of the pro-
file. There have been substantial research effort in the last decade to better understand these
two problems [LPR+07, PRVW07, Wal07, XC11, BHN09, CLMM10, BBN10, BRR11, LPR+12,
FRRS14, DMN16b, DMN16a, DMN17, DMN15, DM17]. One of the main criticisms of this ap-
proach is that the definition of a necessary winner is so strong that none of the alternatives
may satisfy it whereas the definition of a possible winner is so relaxed that a large number of
alternatives may satisfy it. Moreover, the computational problem of finding the set of possible
winners is NP-hard for most of the common voting rules (finding the set of necessary winners
is also co-NP-hard for some voting rules, ranked pairs for example) [XC11].
Lu and Boutilier [LB11] took a completely different approach to handle incomplete pref-
erences and proposed a worst case regret based approach for score based voting rules. These
voting rules assign some score to every alternative based on the profile and select the alter-
natives with the maximum (or minimum) score as winners. Many popular voting rules, for
example, scoring rules, maximin, Copeland, etc. are score based voting rules. For score based
voting rules, intuitively speaking, the worst case regret, called maximum regret in [LB11], of
declaring an alternative w as a winner is the maximum possible difference between the score of
w and the score of a winning alternative in any completion of the input partial profile; the win-
ners of a partial profile are the set of alternatives with the minimum maximum (called minimax)
regret. A completion of a partial profile is another profile where every preference is complete
and it respects the orderings of the corresponding preference in the partial profile. The minimax
regret based approach is not only theoretically robust as argued in [LB11] but also practically
appealing since computing winners is polynomial time solvable for all commonly used voting
rules.
1.1 Motivation
Although the minimax regret based approach enjoys many exciting features, it introduces a
new (which was not present in the classical setting with complete preferences) kind of attack
on the election which we call "manipulative elicitation." That is, it may be possible to partially
elicit the preferences in such a way that makes some favorable alternative win the election. For
example, let us consider a plurality election E where an alternative, say w, is the top alternative
of one preference and another alternative, say x, is the top alternative of every other preference.
In a plurality election, the winners are the set of alternatives who appear as the top alternative
in the largest number of preferences. Hence, x is the unique winner in E. Let us now consider
a partial profile where, in every partial preference, only w and every other alternative who
is preferred less than w in the corresponding preference in E are comparable. Let us call the
resulting partial profile E′. If n is the number of preferences, then the minimax regret plurality
score of w in E′ is (n − 1) whereas the minimax regret plurality score of every other alternative
is n which makes w the unique winner of E′. We call this phenomenon manipulative elicitation.
The problem of manipulative elicitation is even more alarming in AI since, in many applications
(collaborative filtering for example), the parts of the preferences that will be elicited can often
be influenced and controlled in such settings.
1.2 Our Contribution
Our main contribution in this paper is the discovery of the manipulative elicitation attack in
regret based partial preferential setting. We also show that the corresponding computational
problem for manipulative elicitation is polynomial time solvable for every monotone voting rule
2
which includes all commonly used score based voting rules [Theorem 3.1 and Corollary 3.1].
Intuitively speaking, we call a score based voting rule monotone if improving the position of
some alternative in any (complete) preference can only improve its score; we defer its formal
definition till Section 2. To counter the negative result of Theorem 3.1, we introduce a param-
eter per pair of alternatives which specifies the minimum number of partial preferences where
these two alternatives should be comparable. We establish success of our approach by showing
that the new constraints make the corresponding computational task of manipulative elicitation
NP-complete for a large class of scoring rules [Theorem 4.1] which includes the plurality [The-
orem 4.2], veto [Theorem 4.3], k-approval for any k, and Borda voting rules [Corollary 4.1],
maximin [Theorem 4.4], Copelandα for every α ∈ [0, 1] [Theorem 4.5], and simplified Buck-
lin [Theorem 4.6] voting rules. We remark that there could be various ways to enforce lower
bounds on the number of partial preferences where a particular pair of alternatives is com-
parable. For example, this can be a feature in the applications which would allow users to
generate these bounds from some distribution which would in turn overrule the possibility of
such manipulation (due to our hardness results).
2 Preliminaries and Problem Formulation
For a positive integer k, we denote the set {1, 2, . . . , k} by [k]. Let A = {ai : i ∈ [m]} be a set of m
alternatives. We denote the set of all subsets of A of cardinality 2 by (cid:0)A
2 (cid:1). A complete order over
the set A of alternatives is called a (complete) preference. We say that an alternative a ∈ A is
placed at the ℓth position (from left or from top) in a preference ≻ if {b ∈ A : b ≻ a} = ℓ−1. We
denote the set of all possible preferences over A by L(A). A tuple ≻= (≻i)i∈[n] ∈ L(A)n of n
preferences is called a profile. An election E is a tuple (≻, A) where ≻ is a profile over a set A of
alternatives. If not mentioned otherwise, we denote the number of alternatives and the number
of preferences by m and n respectively. A map rc : ⊎n,A∈N+L(A)n −→ 2A \{∅} is called a voting
rule. Given an election E, we can construct from E a directed weighted graph GE which is called
the weighted majority graph of E. The set of vertices in GE is the set of alternatives in E. For any
two alternatives x and y, the weight of the edge (x, y) is DE(x, y) = NE(x, y) − NE(y, x), where
NE(a, b) is the number of preferences where the alternative a is preferred over the alternative
b for a, b ∈ A, a 6= b. Examples of some common voting rules are as follows.
⊲Positional scoring rules: An m-dimensional vector α = (α1, α2, . . . , αm) ∈ Nm with α1 >
α2 > . . . > αm and α1 > αm for every m ∈ N naturally defines a voting rule - an alternative
gets score αi from a preference if it is placed at the ith position, and the score of an alternative is
the sum of the scores it receives from all the preferences. The winners are the alternatives with
the maximum score. Scoring rules remain unchanged if we multiply every αi by any constant
λ > 0 and/or add any constant µ. Hence, we can assume without loss of generality that for any
score vector α, we have gcd((αi)i∈[m]) = 1 and there exists a j < m such that αℓ = 0 for all
ℓ > j. We call such an α a normalized score vector. If αi is 1 for i ∈ [k] and 0 otherwise, then, we
get the k-approval voting rule. The k-approval voting rule is also called the (m − k)-veto voting
rule. The 1-approval voting rule is called the plurality voting rule and the 1-veto voting rule is
called the veto voting rule. If αi = m − i for every i ∈ [m], then we get the Borda voting rule.
⊲Maximin: The maximin score of an alternative x is miny6=x NE(x, y). The winners are the
alternatives with the maximum maximin score.
⊲Copelandα: Given α ∈ [0, 1], the Copelandα score of an alternative x is {y 6= x : DE(x, y) >
0} + α{y 6= x : DE(x, y) = 0}. The winners are the alternatives with the maximum Copelandα
score.
⊲Simplified Bucklin: An alternative x's simplified Bucklin score is the minimum number ℓ such
3
that x is placed within the top ℓ positions in more than half of the preferences. The winners are
the alternatives with the lowest simplified Bucklin score.
We call a voting rule "score based" if the voting rule assigns some score to every alternative
based on the profile and chooses either the set of alternatives with the maximum score or
the set of alternatives with the minimum score as winners. All the above mentioned voting
rules are score based - positional scoring rules, maximin, and Copelandα for every α ∈ [0, 1]
select the set of alternatives with the maximum score as winners and the simplified Bucklin
voting rule selects the set of alternatives with the minimum score as winners. We now define
few properties of a score based voting rule which will be relevant to us. We say that a score
based voting rule s is monotone if, for every positive integer n, every two profiles (≻i)i∈[n]
and (≻′
i)i∈[n] over any finite set A of alternatives, and every alternative x ∈ A such that {y ∈
A : x ≻i y} ⊆ {y ∈ A : x ≻′
i)i∈[n]).
We call a voting rule r neutral if, for every positive integer n, every profile (≻i)i∈[n] over any
finite set A = {xi : i ∈ [m]} of m alternatives, and every permutation σ of [m], we have
σ(r((≻i)i∈[n])) = r((σ(≻i))i∈[n]) where σ(r((≻i)i∈[n])) is the image of r((≻i)i∈[n]) under σ
and σ(≻i) = xσ(1) ≻ xσ(2) ≻ · · · ≻ xσ(m) if ≻i= x1 ≻ x2 ≻ · · · ≻ xm. We call a voting rule
worst efficient if the worst possible score with n preferences over m alternatives can be computed
in a polynomial (in m and n) amount of time. We observe that all the voting rules mentioned
above are neutral, worst efficient, and monotone if, for the case of simplified Bucklin voting
rule, we replace the simplified Bucklin score with negative of that and choose the alternative
with the maximum score.
i y} for every i ∈ [n], we have s(x, (≻i)i∈[n]) 6 s(x, (≻′
2.1 Incomplete Election and Minimax Regret Extension of Score Based Voting
Rules
Although preferences are classically modeled as complete orders, in many scenarios, preferences
can be any partial order over A. We often denote a partial order R by the set {(a, b) : a, b ∈
A, aRb}. Given a profile P of partial preferences (which we call a partial profile), we denote the
set of all completions of P to complete orders by C(P). Lu and Boutilier [LB11] proposed a novel
approach to extend the use of score based voting rules for settings with partial profiles based
on a notion of regret. Let s be a score based voting rule so that the winner is an alternative
with the maximum score. Positional scoring rules, maximin, Copelandα for every α ∈ [0, 1],
etc. are prominent examples of such score based voting rules. Let us denote the score that a
score based voting rule s assigns to an alternative a ∈ A in a profile ≻∈ L(A)n by s(a, ≻).
We denote the minimax regret voting rule based on a voting rule s by s. For a profile ≻, let
s(≻) = argmaxa∈A{s(a, ≻)}. Given a partial profile P and a score based rule s, s(P) is defined
as follows.
s − Regret(a, ≻) = s(s(≻), ≻) − s(a, ≻)
s − MR(a, P) = max
≻∈C(P)
s − Regret(a, ≻)
s(P) = argmin
s − MR(a, P)
a∈A
We remark that Lu and Boutilier [LB11] defined s − Regret(a, ≻) without the absolute
operator on the right; we choose to do so to take care of the simplified Bucklin voting rule. For
a partial profile P and a minimax regret (MR for short) voting rule s, we say that an alternative
a ∈ A co-wins if a ∈ s(P) and wins uniquely if s(P) = {a}. For an alternative a ∈ A, if
s − MR(a, P) = s − Regret(a, ≻) for some ≻∈ C(P), then we call an alternative in s(≻) a
competing alternative of a in P.
4
We now formally define manipulative elicitation and the basic computational problem of
manipulative elicitation for a score based voting rule s.
Definition 2.1 (s-manipulative elicitation). For a profile ≻ over a set A of alternatives, we say
that a partial profile P is called a manipulative elicitation if ≻∈ C(P) and s(P) = {c}.
Definition 2.2 (s-MANIPULATIVE ELICITATION). Given a set A of alternatives, a profile ≻∈ L(A)n
of n preferences, and an alternative c ∈ A, compute if there exists a partial profile P such that
≻∈ C(P) and s(P) = {c}?
We will see in Theorem 3.1 that the s-MANIPULATIVE ELICITATION problem is polynomial
time solvable for every neutral, monotone, and worst efficient score based voting rule. This
shows that all the commonly used voting rule considered here are vulnerable under manipula-
tive elicitation. In the hope to counter this drawback, we extend the basic problem in Defini-
tion 2.2 to MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT in Definition 2.3. We will
indeed see in Section 4 that the MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT prob-
lem is NP-complete for all the voting rules that we consider in this paper. For a partial profile
≻= (≻i)i∈[n] and {a, b} ∈ (cid:0)A
2 (cid:1), we denote the number of partial preferences in ≻ where a and
b are comparable by p{a,b}(≻).
Definition 2.3 (s-MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT). Given a set A of
alternatives, a profile ≻∈ L(A)n of n voters, a function f : (cid:0)A
2 (cid:1) −→ N such that 0 6 f({a, b}) 6 n
for every {a, b} ∈ (cid:0)A
2 (cid:1), and an alternative x ∈ A, compute if there exists a partial profile P such
that ≻∈ C(P), f({a, b}) 6 p{a,b}(≻) for every {a, b} ∈ (cid:0)A
2 (cid:1) and s(P) = {x}?
We remark that both the computational problems in Definition 2.2 and 2.3 have been de-
fined for the unique winner case; we could as well define these problems for the co-winner
case also. It turns out that all our proofs (except Theorem 3.1) can be easily modified for the
co-winner counterpart and our choice for defining these problems in the unique winner setting
is only a matter of exposition.
3 Polynomial Time Algorithm for MANIPULATIVE ELICITATION
Our first result is Theorem 3.1 which shows that the MANIPULATIVE ELICITATION problem is
polynomial time solvable for a large class of voting rules.
Theorem 3.1. The MANIPULATIVE ELICITATION problem is polynomial time solvable for every
monotone, neutral, and worst efficient score based voting rule s.
Proof. Let (A, ≻= (≻i)i∈[n], c) be an arbitrary instance of s-MANIPULATIVE ELICITATION. We
first observe that every instance of s-MANIPULATIVE ELICITATION for the co-winner case is a
YES instance since, due to neutrality of s, c ∈ A = s((∅)n) and ≻∈ L(A)n = C((∅)n). So,
let us consider the unique winner case. Our algorithm is as follows. If c receives the worst
possible score in ≻, then we output NO; otherwise we output YES. Our algorithm runs in
polynomial time since s is worst efficient. To prove the correctness of our algorithm, we begin
with Claim 3.1 below.
Claim 3.1. If the score of c in ≻ is the worst possible score (say βn) that any alternative in A can
possibly receive in any profile with n preferences under the voting rule s, then the s-MANIPULATIVE
ELICITATION instance is a NO instance.
5
Proof. Suppose not, then let us assume that R = (Ri)i∈[n] be a partial profile such that ≻∈ C(R)
and s(R) = {c}. Let s − MR(c, R) = s(s(≻′), ≻′) − s(c, ≻′) for some ≻′= (≻′
i)i∈[n] ∈ C(R). We
now claim the following.
Claim 3.2. s(c, ≻′) = βn.
Proof. The idea of the proof is that if s(c, ≻′) > βn, then we can construct another profile which
can be used to calculate worse regret for c than ≻′ and this will contradict the choice of ≻′.
Formally, let us define another profile ¯≻ = ( ¯≻i)i∈[n] where ¯≻i is obtained from ≻′
i by "moving"
c immediately to the right of the alternatives that are on the left of c in either ≻′
i or ≻i for
i ∈ [n]; that is, for every i ∈ [n], ¯≻i is defined as follows.
¯≻i = {(a, b) : a, b ∈ A \ {c}, a ≻′
∪ {(c, a) : a ∈ A, c ≻′
∪ {(a, c) : a ∈ A, a ≻′
i b}
i a, c ≻i a}
i c or a ≻i c}
The profile ¯≻ ∈ C(R) since ≻∈ C(R) and ≻′∈ C(R). Due to monotonicity of s, the score of c
in ¯≻ is at most the score of c in ≻′ and the score of every other alternative in ¯≻ is at least their
score in ≻′. However, ≻′ has been used to calculate the MR score of c under s. Hence, we have
the following:
s(s(≻′), ≻′) = s(s( ¯≻), ≻′),
s(c, ≻′) = s(c, ¯≻)
We now have the following:
βn 6 s(c, ≻′) = s(c, ¯≻) 6 s(c, ≻) = βn
The first inequality follows from the definition of βn and the second inequality follows from
monotonicity of s.
Let y ∈ s(≻′) and s − MR(y, R) = s(s(≻′′), ≻′′) − s(y, ≻′′) for some ≻′′= (≻′′
i )i∈[n] ∈ C(R).
We now have the following claim.
Claim 3.3. s(s(≻′′), ≻′′) 6 s(s(≻′), ≻′)
Proof. The idea of the proof is along the same line as Claim 3.2. Let us define another profile
¯≻ = ( ¯≻i)i∈[n] where ¯≻i is obtained from ≻′′
i by "moving" c immediately to the right of the
alternatives that are on the left of c in either ≻′′
i for i ∈ [n]; that is, for every i ∈ [n], ¯≻i
is defined as follows.
i or ≻′
¯≻i = {(a, b) : a, b ∈ A \ {c}, a ≻′′
i b}
i a, c ≻′
i a}
i c or a ≻′
∪ {(c, a) : a ∈ A, c ≻′′
∪ {(a, c) : a ∈ A, a ≻′′
i c}
The profile ¯≻ ∈ C(R) since ≻′∈ C(R) and ≻′′∈ C(R). Due to monotonicity of s, the score of
c in ¯≻ is at most the score of c in ≻′′ and the score of every other alternative in ¯≻ is at least
their score in ≻′′. Again due to monotonicity of s, the score of c in ¯≻ is at most the score of
c in ≻′. However, ≻′ has been used to calculate the MR score of c under s. Hence, we have
s(s(≻′′), ≻′′) 6 s(s(≻′), ≻′); otherwise we will have s − Regret(c, ≻′) < s − Regret(c, ¯≻) which
is a contradiction.
6
We now combine Claim 3.2 and 3.3 as follows to prove the main claim.
s − MR(y, R) = s(s(≻′′), ≻′′) − s(y, ≻′′)
6 s(s(≻′), ≻′) − s(y, ≻′′)
6 s(s(≻′), ≻′) − βn
= s(s(≻′), ≻′) − s(c, ≻′)
= s − MR(c, R)
The second line follows from Claim 3.3, the third line follows from the definition of βn, and
the fourth line follows from Claim 3.2. Hence we have s − MR(y, R) 6 s − MR(c, R) which
contradicts our assumption that s(R) = {c}.
We now show that if c does not receive the worst possible score with n preferences over A
under s from the profile ≻, then the instance is a YES instance. To see this, let us consider the
partial profile P = (Pi)i∈[n] as Pi = {c ≻ y : c ≻i y} for every i ∈ [n]. Let R be any profile
in C(P). Since, the alternative c does not receive the worst possible score with n preferences
over A under s from the profile ≻, s(c, R) < βn. Hence, if α is the best possible score with
n preferences over A under s, we have s − MR(c, P) < α − βn. On the other hand, for any
alternative y ∈ A \ {c}, let us consider the profile Qy = (Qi)i∈[n] where Qi = c ≻ · · · ≻ y for
every i ∈ [n]. Now due to monotonicity of s, we have s(c, Qy) = α and s(y, Qy) = βn. Hence,
we have s − MR(y, P) = α − βn for every y ∈ A \ {c} and thus s(P) = {c}.
We remark that the proof of Theorem 3.1 for the co-winner case is trivial: every instance is
a YES instance since a partial profile where every preference is empty makes every alternative
win due to neutrality. Since scoring rules, maximin, Copelandα for every α ∈ [0, 1], and simpli-
fied Bucklin voting rules are monotone, neutral, and worst efficient, Theorem 3.1 immediately
implies the following corollary.
Corollary 3.1. The MANIPULATIVE ELICITATION problem is polynomial time solvable for scoring
rules, maximin, Copelandα for every α ∈ [0, 1], and simplified Bucklin voting rules.
Proof. The worst possible score with n preferences over m alternatives is 0 for scoring rules,
maximin, and Copeland α for every α ∈ [0, 1], and −m for the simplified Bucklin voting rule
(under modified but equivalent definition of simplified Bucklin voting rule).
4 Hardness Results for MANIPULATIVE ELICITATION WITH CANDIDATE
PAIR LIMIT
In this section we show that the MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT prob-
lem is NP-complete for maximin, Copelandα for every α ∈ [0, 1], simplified Bucklin, and a large
class of scoring rules which includes the k-approval voting rule for every k and the Borda voting
rule.
4.1 Scoring Rules
Let us define a restricted version of the classical set cover problem which we call SET COVER
FREQUENCY TWO. We will see in Lemma 4.1 that this problem is NP-complete by reducing
from the vertex cover problem which is well known to be NP-complete [GJ79]. Most of our
NP-hardness reductions are from this problem.
7
Definition 4.1 (SET COVER FREQUENCY TWO). Given a universe U of cardinality q, a family
S = {Si : i ∈ [t]} of t subsets of U such that for every a ∈ U, we have {i ∈ [t] : a ∈ Si} = 2,
and a positive integer ℓ, compute if there exists a subset G ⊆ S containing at most ℓ sets such that
∪A∈GA = U. We denote an arbitrary instance of SET COVER FREQUENCY TWO by (U, S, ℓ).
Lemma 4.1. SET COVER FREQUENCY TWO is NP-complete.
Proof. SET COVER FREQUENCY TWO clearly belongs to NP. To prove NP-hardness, we reduce
from VERTEX COVER to SET COVER FREQUENCY TWO. Let (G = (V, E), k) be an arbitrary instance
of VERTEX COVER. We construct and instance (U, S, ℓ) as follows:
U = {ae : e ∈ E}, S = {Sv : v ∈ V} where Sv = {ae : e is incident on v}, ℓ = k
Clearly every element ae ∈ U belongs to exactly two sets, namely Su and Sv if e = {u, v}.
Also the equivalence of two instances are straight forward.
We begin with showing that the MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT
problem is NP-complete for a large class of scoring rules which included the k-approval voting
rule for every 3 6 k 6 γm for any constant 0 < γ < 1 and the Borda voting rule. While
describing a (complete) preference, if we do not mention the order of any two alternatives,
they can be ordered arbitrarily. On the other hand, if we are describing a partial preference
and we do not mention the order of any two alternatives, then they should be assumed to be
incomparable.
Theorem 4.1. Let r be a normalized scoring rule such that there exists a function g : N −→ N such
that for every m ∈ N, we have 3m 6 g(m) 6 poly(m) and if α = (αi)i∈[g(m)], then there exists
a positive integer p such that 3 6 p 6 g(m) − m + 3, αp > αp+1 and αp−1 = poly(m). Then the
MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT problem is NP-complete for the scoring
rule r.
Proof. The MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT problem clearly belongs to
NP. To prove NP-hardness, we reduce from SET COVER FREQUENCY TWO to MANIPULATIVE
ELICITATION WITH CANDIDATE PAIR LIMIT for the scoring rule r. Let (U = {u1, . . . , uq}, S =
{Si : i ∈ [t]}, ℓ) be an arbitrary instance of SET COVER FREQUENCY TWO. Let us consider the
following instance (A, P, c, f) of MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT where
A is defined as follows.
A = {ai : i ∈ [q]} ∪ {c, d} ∪ W, where W = {w1, . . . , wg(q)−q−2}
The profile P consists of the following preferences. For an integer 0 6 k 6 g(q) − q − 2, we
denote the set {wi : i ∈ [k]} by Wk. Let κ = max{i ∈ [g(q)] : αi 6= 0}; we observe that ακ = 1
since r is normalized. For X ⊆ U, let us denote the set {aj : uj ∈ X} of alternatives by X to
simplify notation.
⊲ ∀i ∈ [t] : Wp−3 ≻ Si ≻ d ≻ c ≻ (U \ Si) ≻ (W \ Wp−3)
⊲ t − 2 copies of Wp−1 ≻ ai ≻ d ≻ (U \ {ai}) ≻ (W \ Wp−1) ≻ c
⊲ t(αp − αp+2) copies of W2 ≻ · · · ≻ c ≻ · · · d where the alternative c is placed at κ position
from left.
⊲ If αg(q)−1 = 1, then we add (q + 1)tαp−1 copies of · · · ≻ c ≻ d
⊲ Otherwise:
8
– If κ 6 g(q) − q − 1, then we add (q + 1)tαp−1 copies of W2 ≻ · · · ≻ c ≻ d ≻ U ≻ · · ·
where the alternative d is placed at κ + 1 position from left and we add, for every
i ∈ [q], (q + 1)tαp−1 copies of W2 ≻ · · · ≻ ai ≻ d ≻ c ≻ (U \ {ai}) ≻ · · · where the
alternative d is placed at κ + 1 position from the left.
– Otherwise we add (q + 1)tαp−1 copies of W2 ≻ · · · ≻ U ≻ c ≻ d ≻ · · · where the
alternative d is placed at κ + 1 position from the left.
For ease of reference, we call the above four groups as G1, G2, G3, and G4 respectively. Let n
be the number of preferences in P. We observe that n = poly(m) since αp−1 = poly(m). The
function f is defined as follows: f({d, x}) = n for every x ∈ A \ ({c, d}), f({d, c}) = n − ℓ; the value
of f be 0 for all other pairs of alternatives. This finishes the description of our reduced instance.
We now claim that the two instances are equivalent.
In one direction, let us assume that the SET COVER FREQUENCY TWO instance is a YES
instance; without loss of generality, let us assume (by renaming) that S1, . . . , Sℓ forms a set
cover of U. Let us consider the following partial profile Q with P ∈ L(Q).
⊲ Preferences in G1: ∀i ∈ [ℓ] : ((Wp−3 ∪ Si) ≻ d ≻ ((U \ Si) ∪ (W \ Wp−3)) S c ≻ ((U \ Si)) ∪
((W \ Wp−3)))
⊲ Preferences in G1: ∀i with ℓ + 1 6 i 6 t : (Wp−3 ∪ Si) ≻ d ≻ c ≻ ((U \ Si) ∪ (W \ Wp−3)
⊲ Preferences in G2: t − 2 copies of (Wp−1 ∪ ai) ≻ d ≻ ((U \ {ai}) ∪ (W \ Wp−1) ∪ {c})
⊲ Preferences in G3: t(αp − αp+2) copies of c ≻ X where the alternative X = {b ∈ A : c ≻
b in G3}.
⊲ Preferences in G4: for every preference in G4, we add c ≻ Y where the alternative Y =
{b ∈ A : c ≻ b in the corresponding preference in G4}.
Let ∆ be the score that the alternative w1 receives in P. We observe that the minimum scores
that the alternatives c and ai, i ∈ [q] receive in profile R with R ∈ L(Q) are all the same; let it
be λ. We summarize the MR score (based on r) of every alternative from Q in Table 1. Hence
the alternative c wins uniquely in Q.
Alternative MR-r score from Q Competing alternative
c
∆ − tαℓ − λ
ai, ∀i ∈ [q]
∆ − tαℓ + αℓ+1 − λ
w1(w2)
w ∈ W \ W2
∆
∆
d
> D − tαℓ
w1 (or w2)
w1 (or w2)
w2(w1)
w1
w1 (or w2)
Table 1: Summary of MR scores (based on r) of all the alternatives from the partial profile Q in
the proof of Theorem 4.1.
In the other direction, let us assume that the MANIPULATIVE ELICITATION WITH CANDIDATE
PAIR LIMIT instance (A, P, c, f) is a YES instance. Let Q be a partial profile such that P ∈ C(Q)
and the alternative c wins uniquely in Q under the MR scoring rule based on r. We observe that
if a preference profile Rc with Rc ∈ L(Q) is used to calculate the MR score of the alternative
based on r, then the MR score of c based on r is at least ∆ − tαℓ − λ using the alternative w1
as a competing alternative where λ and ∆ are as defined above. Let J ⊆ [t] be the set of i ∈ [t]
such that the corresponding partial preferences in the group G1 in Q leave the alternatives c and
9
d incomparable. Since f({d, c}) = n − ℓ, we have J 6 ℓ. We claim that {Sj : j ∈ J} forms a set
cover of U. Suppose not, then let uk ∈ U \ (∪j∈JSj). Then we observe that the MR score of ak
based on r is at least ∆ − tαℓ − λ using the alternative w1 as a competing alternative.However,
this contradicts our assumption that c is the unique MR-r winner of Q. Hence {Sj : j ∈ J} forms
a set cover of U and thus the SET COVER FREQUENCY TWO is a YES instance. This concludes the
proof of the theorem.
Theorem 4.1 immediately gives us the following corollary.
Corollary 4.1. The MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT problem is NP-
complete for the Borda and k-approval voting rules for every 3 6 k 6 γm for any constant
0 < γ < 1.
A drawback of Theorem 4.1 is that it does not cover the plurality, 2-approval, and the k-veto
voting rules for k = o(m). We will show that the MANIPULATIVE ELICITATION WITH CANDIDATE
PAIR LIMIT problem is NP-complete for the k-veto voting rule for any 1 6 k 6 γm for any
constant 0 < γ < 1 in Theorem 4.3. We now show in Theorem 4.2 that the MANIPULATIVE ELIC-
ITATION WITH CANDIDATE PAIR LIMIT problem is NP-complete for the plurality and 2-approval
voting rules by reducing it from the X3C problem which is defined as follows and known to be
NP-complete [GJ79].
Definition 4.2 (X3C). Given a universe U of cardinality q such that q is divisible by 3, a family
S = {Si : i ∈ [t]} of t subsets of U each of cardinality 3, compute if there exists a subset G ⊆ F of
q/3 sets such that ∪A∈GA = U. We denote an arbitrary instance of X3C by (U, S).
Theorem 4.2. The MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT problem is NP-
complete for the plurality and the 2-approval voting rules.
Proof. Let us first consider the plurality voting rule. The MANIPULATIVE ELICITATION WITH CAN-
DIDATE PAIR LIMIT problem for the plurality voting rule clearly belongs to NP. To prove NP-
hardness, we reduce from X3C to MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT for
the plurality voting rule. Let (U = {u1, . . . , uq}, S = {Si : i ∈ [t]}) be an arbitrary instance of
X3C. For every i ∈ [q] let us define fi = {j ∈ [t] : ui ∈ Sj}. Let us assume, without loss of
generality, that fi < t − q/2 (if not, then we add 3t new elements in U and t sets in S each of size
3 and collectively covering these new 3t elements). Let us assume, without loss of generality,
that q is divisible by 6; if not then we add 3 new elements in U and a set consisting of these
three new elements in S. Let us consider the following instance (A, P, c, f) of MANIPULATIVE
ELICITATION WITH CANDIDATE PAIR LIMIT where A is defined as follows.
A = {ai : i ∈ [q]} ∪ {c, d, w}
The profile P consists of the following preferences. For X ⊆ U, let us also denote, for the
sake of simplicity of notation, the set {aj : uj ∈ X} of alternatives by X.
⊲ ∀i ∈ [t] : d ≻ Si ≻ c ≻ (U \ Si) ≻ w
⊲ q/6 + 1 copies of c ≻ (A \ {c, w})
⊲ 1 copy of d ≻ w ≻ c ≻ (A \ {d, w})
For ease of reference, we call the above three groups as G1, G2, and G3 respectively. Let n be
the number of preferences in P. That is, n = t + q/6 + 2. The function f is defined as follows:
f({c, ai}) = n − fi + 1 for every i ∈ [q], f({w, d}) = n − 1, f({w, x}) = n for every x ∈ A \ {w, d};
10
the value of f be 0 for all other pairs of alternatives. This finishes the description of our reduced
instance. We now claim that the two instances are equivalent.
In one direction, let us assume that the X3C instance is a YES instance; without loss of
generality, let us assume (by renaming) that S1, . . . , Sq/3 forms a set cover of U. Let us consider
the following partial profile Q with P ∈ L(Q).
⊲ Preferences in G1: ∀i ∈ [q/3] : (Si ≻ c ≻ (U \ Si) ≻ w) S(d ≻ w)
⊲ Preferences in G1: ∀i with q/3 + 1 6 i 6 t : (c ≻ (U \ Si) ≻ w) S(d ≻ w)
⊲ Preferences in G2: q/6 + 1 copies of c ≻ (A \ {c, w})
⊲ Preferences in G3: 1 copy of w ≻ c ≻ (A \ {d, w})
We summarize the MR-plurality score of every alternative from Q in Table 2. Hence the
alternative c wins uniquely in Q.
Alternative MR-plurality score from Q Competing alternative
c
ai, ∀i ∈ [q]
d
t − q/6
t
t − q/6 + 1
d
d
c
Table 2: Summary of MR-plurality scores of all the alternatives from the partial profile Q in the
proof of Theorem 4.2.
In the other direction, let us assume that the MANIPULATIVE ELICITATION WITH CANDIDATE
PAIR LIMIT instance (A, P, c, f) is a YES instance. Let Q be a partial profile such that P ∈ C(Q)
and the alternative c wins uniquely in Q under the MR-plurality voting rule. Let J ⊆ [t] be the
set of i ∈ [t] such that the corresponding partial preferences in the group G1 in Q leave the
alternatives c and at least one alternative in Si incomparable. A key observation is that since
f({c, ai}) = n−fi+1 for i ∈ [q], we have ∪j∈JSj = U. Hence we have J > q/3. We now claim that
J 6 q/3. Suppose not, then the MR-plurality score of d is at most (t − q/3 − 1) + q/6 + 1 = t − q/6
using c as competing alternative (we observe that, since f({c, ai}) = n − fi + 1 for i ∈ [q],
using the alternative ai as a competing alternative for any i will lead to MR-plurality score of
d at most 2fi < t − q/6). Hence the MR-plurality score of d is at most t − q/6. However the
MR-plurality score of c is at least t + 1 − (q/6 + 1) = t − q/6. This contradicts our assumption
that c is the unique MR-plurality winner of Q. Hence {Sj : j ∈ J} forms a set cover of U and thus
the X3C is a YES instance. This concludes the proof of the theorem.
For the 2-approval voting rule, we can introduce n dummy alternatives each of which ap-
pears at the first position in exactly one preference and in the rest (n− 1) preferences, it appears
in the bottom (n − 1) positions. All other parameters of the reduction remain same. It is easy
to see that a similar argument will prove the result for the 2-approval voting rule.
We now show our hardness result for the k-veto voting rule by reducing from X3C.
Theorem 4.3. The MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT problem is NP-
complete for the k-veto voting rule for every 1 6 k 6 γm for any constant 0 < γ < 1.
Proof. Let us first consider the veto voting rule. The MANIPULATIVE ELICITATION WITH CANDI-
DATE PAIR LIMIT problem for the veto voting rule clearly belongs to NP. To prove NP-hardness,
we reduce from X3C to MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT for the veto
11
voting rule. Let (U = {u1, . . . , uq}, S = {Si : i ∈ [t]}) be an arbitrary instance of X3C. For every
i ∈ [q] let us define fi = {j ∈ [t] : ui ∈ Sj}. Let us assume, without loss of generality, that
fi < t − q/2 (if not, then we add 3t new elements in U and t sets in S each of size 3 and col-
lectively covering these new 3t elements). Let us assume, without loss of generality, that q is
divisible by 6; if not then we add 3 new elements in U and a set consisting of these three new
elements in S. We also assume that t is an odd integer; if not then we duplicate one set in S. Let
us consider the following instance (A, P, c, f) of MANIPULATIVE ELICITATION WITH CANDIDATE
PAIR LIMIT where A is defined as follows.
A = {ai : i ∈ [q]} ∪ {c, x, d, w1, w2}
The profile P consists of the following preferences. For X ⊆ U, let us also denote, for the
sake of simplicity of notation, the set {aj : uj ∈ X} of alternatives by X.
1. ∀i ∈ [t] : (U \ Si) ≻ w1 ≻ w2 ≻ d ≻ c ≻ Si ≻ x
2. (t−1)/2 + q/6 copies of (A \ {c, d}) ≻ d ≻ c
3. ∀j ∈ [q] : (t+1)/2 + q/6 + 1 − fj copies of (A \ {aj, d}) ≻ d ≻ aj
4. 10t copies of (A \ {d}) ≻ d
5. 10t copies of (A \ {d, w1, w2}) ≻ d ≻ w1 ≻ w2
6. ∀i ∈ [2] : q/3 copies of (A \ {wi, d}) ≻ d ≻ wi
For ease of reference, we call the above six groups as G1, G2, G3, G4, G5 and G6 respectively. Let n
be the number of preferences in P. The function f is defined as follows: f({d, z}) = n for every
z ∈ A \ {d}; the value of f be 0 for all other pairs of alternatives. This finishes the description of
our reduced instance. We now claim that the two instances are equivalent.
In one direction, let us assume that the SET COVER FREQUENCY TWO instance is a YES
instance; without loss of generality, let us assume (by renaming) that S1, . . . , Sq/3 forms a set
cover of U. Let us consider the following partial profile Q with P ∈ L(Q).
⊲ Preferences in G1: ∀i ∈ [q/3] : (U \ Si) ≻ w1 ≻ w2 ≻ d ≻ c ≻ Si ≻ x
⊲ Preferences in G1: ∀i with q/3 + 1 6 i 6 t : (U \ Si) ≻ w1 ≻ w2 ≻ d ≻ c ≻ (Si ∪ {x})
⊲ Preferences in G2: (t−1)/2 + q/6 copies of (A \ {c, d}) ≻ d ≻ c
⊲ Preferences in G3: ∀j ∈ [q] : (t+1)/2 + q/6 + 1 − fj copies of (A \ {aj, d}) ≻ d ≻ aj
⊲ Preferences in G4: 10t copies of (A \ {d, w1, w2}) ≻ d ≻ {w1, w2}
⊲ Preferences in G5: 10t copies of (A \ {d}) ≻ d
⊲ Preferences in G6: ∀i ∈ [2] : q/3 copies of (A \ {wi, d}) ≻ d ≻ wi
We summarize the MR-veto score of every alternative from Q in Table 3. Hence the alterna-
tive c wins uniquely in Q.
In the other direction, let us assume that the MANIPULATIVE ELICITATION WITH CANDIDATE
PAIR LIMIT instance (A, P, c, f) is a YES instance. Let Q be a partial profile such that P ∈ C(Q)
and the alternative c wins uniquely in Q under the MR-veto voting rule. We first observe that,
since f({d, z}) = n for every z ∈ A \ {d}, the alternative c is forced to receive a score of 0 from
every preference in Q corresponding to group G2 (and thus every other alternative is forced to
12
Alternative MR-veto score from Q Competing alternative
c
x
ai, ∀i ∈ [q]
d
w1(w2)
(t−1)/2 − q/6
(t+1)/2 − q/6
(t+1)/2 − q/6
> 5t
> 5t
x (or w1 or w2)
c
x (or w1 or w2)
x (or w1 or w2)
w2(w1)
Table 3: Summary of MR-veto scores of all the alternatives from the partial profile Q in the
proof of Theorem 4.3.
receive a score of 1), both the alternatives x and c are forced to receive a score of 1 from every
preference in Q corresponding to group G3, and the alternative d is forced to receive a score
of 0 from every preference in Q corresponding to group G4 (and thus every other alternative is
forced to receive a score of 1). We now claim that there must be at least q/3 preferences in Q
corresponding to the group G1 where the alternative x must be forced to receive a score of 0 for c
to win uniquely. Suppose not, then we observe that the MR-veto score of c is at least (t+1)/2−q/6
using x as a competing alternative whereas the MR-veto score of x is at most (t−1)/2− q/6 using c
as a competing alternative which contradicts our assumption that c wins the MR-veto election.
Let J ⊆ [t] be the set of i ∈ [t] such that the corresponding partial preferences in the group G1
in Q force the alternative x to receive a score of 0. Then we have J > q/3. We now claim that
J = q/3 and the sets {Sj : j ∈ J} forms an exact set cover of U. Suppose not, then there exists an
uk ∈ U which appears in at least two sets in {Sj : j ∈ J}. Then the alternative is forced to receive
a score of 1 from at least t − fk + 2 preferences in Q corresponding to the group G1 – t − fk
preferences where it appears on the left of d and at least 2 preferences where the alternative
x is forced to receive a score of 0. However, then the MR-veto score of the alternative ak is at
most (t−1)/2 − q/6 using w1 (or w2) as a competing alternative. This contradicts our assumption
that c wins the MR-veto election uniquely. Hence the sets {Sj : j ∈ J} forms an exact set cover
of U and thus the X3C is a YES instance. This concludes the proof of the theorem for the veto
voting rule.
For the k-veto voting rule, we add k − 1 "dummy" alternatives di, i ∈ [k − 1] at the bottom of
every preferences in P and this is easily verifiable that the proof for the k-veto voting rule goes
along the same line.
4.2 Maximin Voting Rule
We now show our hardness result for the maximin voting rule by reducing from SET COVER
FREQUENCY TWO.
Theorem 4.4. The MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT problem is NP-
complete for the maximin voting rule.
Proof. The MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT problem clearly belongs
to NP. To prove NP-hardness, we reduce from SET COVER FREQUENCY TWO to MANIPU-
LATIVE ELICITATION WITH CANDIDATE PAIR LIMIT for the maximin voting rule. Let (U =
{u1, . . . , uq}, S = {Si : i ∈ [t]}, ℓ) be an arbitrary instance of SET COVER FREQUENCY TWO. Let us
consider the following instance (A, P, c, f) of MANIPULATIVE ELICITATION WITH CANDIDATE PAIR
LIMIT where A is defined as follows.
A = {ai : i ∈ [q]} ∪ {c, w1, w2, d}
13
The profile P consists of the following preferences. For X ⊆ U, let us also denote, for the
sake of simplicity of notation, the set {aj : uj ∈ X} of alternatives by X.
⊲ ∀i ∈ [t] : w1 ≻ w2 ≻ Si ≻ d ≻ c ≻ (U \ Si)
⊲ 2 copies of c ≻ U ≻ d ≻ w1 ≻ w2
For ease of reference, we call the above two groups as G1 and G2 respectively. Let n be the
number of preferences in P. The function f is defined as follows: f({d, x}) = n for every x ∈
A \ ({c, d}), f({d, c}) = n − ℓ; the value of f be 0 for all other pairs of alternatives. This finishes
the description of our reduced instance. We now claim that the two instances are equivalent.
In one direction, let us assume that the SET COVER FREQUENCY TWO instance is a YES
instance; without loss of generality, let us assume (by renaming) that S1, . . . , Sℓ forms a set
cover of U. Let us consider the following partial profile Q with P ∈ L(Q).
⊲ Preferences in G1: ∀i ∈ [ℓ] : (({w1, w2} ∪ Si) ≻ d ≻ {U \ Si}) S(c ≻ {U \ Si})
⊲ Preferences in G1: ∀i with ℓ + 1 6 i 6 t : ({w1, w2} ∪ Si) ≻ d ≻ c ≻ {U \ Si}
⊲ Preferences in G2: 2 copies of c ≻ U ≻ d ≻ {w1, w2}
We summarize the MR-maximin score of every alternative from Q in Table 4. Hence the
alternative c wins uniquely in Q.
Alternative MR-maximin from Q Competing alternative
Comments
c
ai, ∀i ∈ [q]
w1(w2)
d
t − 2
t − 1
t
t
w1 (or w2)
w1 (or w2)
w2(w1)
w1 (or w2)
N(w1, c) − N(c, d)
N(w1, c) − N(ai, c)
N(w1, c) − N(w2, w1)
N(w1, c) − N(d, w1)
Table 4: Summary of MR-maximin scores of all the alternatives from the partial profile Q in the
proof of Theorem 4.4.
In the other direction, let us assume that the MANIPULATIVE ELICITATION WITH CANDIDATE
PAIR LIMIT instance (A, P, c, f) is a YES instance. Let Q be a partial profile such that P ∈ C(Q)
and the alternative c wins uniquely in Q under the MR-maximin voting rule. We observe that for
every R ∈ L(Q) which can be used for calculating the MR-maximin score of the alternative c,
we have NR(c, d) 6 2. Also, there are only two preferences (the preferences in G2) where there
exist some alternatives which are preferred over the alternative w1. Hence the MR-maximin
score of the alternative c in Q is at least t − 2. Let J ⊆ [t] be the set of i ∈ [t] such that the
corresponding partial preferences in the group G1 in Q leave the alternatives c and d incompa-
rable. Since f({d, c}) = n − ℓ, we have J 6 ℓ. We claim that {Sj : j ∈ J} forms a set cover of
U. Suppose not, then let uk ∈ U \ (∪j∈JSj). We observe that for every R′ ∈ L(Q), we have
NR′(ak, c) = 2. We also observe that NR′(ak, d) = NR′ (ak, wi) = 2 for every i ∈ [2] since
f({d, ui}) = n for every i ∈ [q] and f({d, w1}) = f({d, w2}) = n. Hence, the MR-maximin score
of the alternative ak is t − 2 where the alternative w1 plays the role of a competing alternative.
However, this contradicts our assumption that c is the unique MR-maximin winner of Q. Hence
{Sj : j ∈ J} forms a set cover of U and thus the SET COVER FREQUENCY TWO is a YES instance.
This concludes the proof of the theorem.
14
4.3 Copelandα Voting Rule
We show next our hardness result for the Copelandα voting rule for every α ∈ [0, 1].
Theorem 4.5. The MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT problem is NP-
complete for the Copelandα voting rule for every α ∈ [0, 1].
Proof. The MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT problem clearly belongs to
NP. To prove NP-hardness, we reduce from SET COVER FREQUENCY TWO to MANIPULATIVE
ELICITATION WITH CANDIDATE PAIR LIMIT for the Copelandα voting rule for every α ∈ [0, 1].
Let (U = {u1, . . . , uq}, S = {Si : i ∈ [t]}, ℓ) be an arbitrary instance of SET COVER FREQUENCY
TWO. Let us consider the following instance (A, P, c, f) of MANIPULATIVE ELICITATION WITH
CANDIDATE PAIR LIMIT where A is defined as follows.
A = {ai : i ∈ [q]} ∪ {c, x, y, d, w1, w2}}
The profile P consists of the following preferences. For X ⊆ U, let us also denote, for the
sake of simplicity of notation, the set {aj : uj ∈ X} of alternatives by X.
⊲ ∀i ∈ [t] : w1 ≻ w2 ≻ Si ≻ d ≻ c ≻ x ≻ y ≻ (U \ Si)
⊲ t − 3 copies of U ≻ c ≻ x ≻ y ≻ d ≻ w1 ≻ w2
For ease of reference, we call the above two groups as G1 and G2 respectively. Let n be the
number of preferences in P. We observe that n is an odd integer and thus the value of α is
irrelevant. Hence, from here on, we omit the parameter α. The function f is defined as follows:
f({d, x}) = n for every x ∈ A \ {c, d, x, w1, w2}, f({d, c}) = f({d, x}) = n − ℓ; the value of f be 0
for all other pairs of alternatives. This finishes the description of our reduced instance. We now
claim that the two instances are equivalent.
In one direction, let us assume that the SET COVER FREQUENCY TWO instance is a YES
instance; without loss of generality, let us assume (by renaming) that S1, . . . , Sℓ forms a set
cover of U. Let us consider the following partial profile Q.
⊲ Preferences in G1: ∀i ∈ [ℓ] : Si ≻ d ≻ y ≻ (U \ Si)) S(c ≻ {{x, y} ∪ (U \ Si)})
⊲ Preferences in G1: ∀i with ℓ + 1 6 i 6 t : Si ≻ d ≻ x ≻ y ≻ (U \ Si)
⊲ Preferences in G2: t − 3 copies of c ≻ {x, d, w1, w2}
We summarize the MR-Copeland score of every alternative from Q in Table 5. Hence the
alternative c wins uniquely in Q.
In the other direction, let us assume that the MANIPULATIVE ELICITATION WITH CANDIDATE
PAIR LIMIT instance (A, P, c, f) is a YES instance. Let Q be a partial profile such that P ∈ C(Q)
and the alternative c wins uniquely in Q under the MR-Copeland voting rule. We observe that
in P, the alternative c defeats only 2 alternatives namely x and y. Since, P ∈ L(Q), the MR-
Copeland score of the alternative c in Q is at least (A − 1) − 2 = A − 3. Let J ⊆ [t] be the
set of i ∈ [t] such that the corresponding partial preferences in Q leave the alternatives c and d
incomparable. Since f({d, c}) = n − ℓ for every i ∈ [t], we have J 6 ℓ. We claim that {Sj : j ∈ J}
forms a set cover of U. Suppose not, then let uk ∈ U \ (∪j∈JSj). We observe that in every profile
R ∈ L(Q), the alternative ak defeats the alternatives c and d and thus the MR-Copeland score of
the alternative ak in Q is at most (A− 3) where the alternative w1 plays the role of a competing
alternative. However, this contradicts our assumption that c is the unique MR-Copeland winner
of Q. Hence {Sj : j ∈ J} forms a set cover of U and thus the SET COVER FREQUENCY TWO is a YES
instance. This concludes the proof of the theorem.
15
Alternative MR-Copeland from Q Competing alternative
Comments
c
ai, ∀i ∈ [q]
w1(w2)
d
A − 3
A − 2
A − 1
A − 2
w1 (or w2)
w1 (or w2)
w2(w1)
w2(w1)
w1 defeats A \ {w1}
c defeats {x, y}
w1 defeats A \ {w1}
ai defeats d
w1 defeats A \ {w1}
w2 defeats none
w1 defeats A \ {w1}
d defeats y
Table 5: Summary of MR-Copeland scores of all the alternatives from the partial profile Q in the
proof of Theorem 4.5.
4.4 Simplified Bucklin Voting Rule
We now show our hardness result for the simplified Bucklin voting rule.
Theorem 4.6. The MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT problem is NP-
complete for the simplified Bucklin voting rule.
Proof. The MANIPULATIVE ELICITATION WITH CANDIDATE PAIR LIMIT problem clearly belongs
to NP. To prove NP-hardness, we reduce from SET COVER FREQUENCY TWO to MANIPU-
LATIVE ELICITATION WITH CANDIDATE PAIR LIMIT for the simplified Bucklin voting rule. Let
(U = {u1, . . . , uq}, S = {Si : i ∈ [t]}, ℓ) be an arbitrary instance of SET COVER FREQUENCY TWO.
Let us consider the following instance (A, P, c, f) of MANIPULATIVE ELICITATION WITH CANDI-
DATE PAIR LIMIT where A is defined as follows.
A = {ai : i ∈ [q]} ∪ {c, d} ∪ W, where W = {wi : i ∈ [q]}
The profile P consists of the following preferences. For X ⊆ U, let us also denote, for the
sake of simplicity of notation, the set {aj : uj ∈ X} of alternatives by X.
⊲ ∀i ∈ [t] : w1 ≻ w2 ≻ · · · ≻ wq−2 ≻ Si ≻ d ≻ c ≻ (U \ Si) ≻ wq−1 ≻ wq
⊲ t − 1 copies of U ≻ c ≻ d ≻ W
⊲ 2 copies of w1 ≻ w2 ≻ · · · ≻ wq ≻ c ≻ d ≻ U
For ease of reference, we call the above three groups as G1, G2, and G3 respectively. Let n
be the number of preferences in P. The function f is defined as follows: f({d, x}) = n for
every x ∈ A \ ({c, d, w1, . . . , wq−2}), f({d, c}) = n − ℓ; the value of f be 0 for all other pairs of
alternatives. This finishes the description of our reduced instance. We now claim that the two
instances are equivalent.
In one direction, let us assume that the SET COVER FREQUENCY TWO instance is a YES
instance; without loss of generality, let us assume (by renaming) that S1, . . . , Sℓ forms a set
cover of U. Let us consider the following partial profile Q with P ∈ L(Q).
⊲ Preferences
∀i ∈ [ℓ]
wq) S (c ≻ (U \ Si) ≻ wq−1 ≻ wq)
in G1:
:
(Si ≻ d ≻ (U \ Si) ≻ wq−1 ≻
⊲ Preferences in G1: ∀i with ℓ + 1 6 i 6 t : Si ≻ d ≻ c ≻ (U \ Si) ≻ wq−1 ≻ wq
16
⊲ Preferences in G2: t − 1 copies of U ≻ c ≻ d ≻ W
⊲ Preferences in G3: 2 copies of wq−1 ≻ wq ≻ c ≻ d ≻ U
We summarize the MR-simplified Bucklin score of every alternative from Q in Table 6. Hence
the alternative c wins uniquely in Q.
Alternative MR-simplified Bucklin score from Q Competing alternative
c
ai, ∀i ∈ [q]
w1
wi, 2 6 i 6 q
d
−q
−q − 1
−2q − 1
−2q − 1
−q − 1
w1
w1
w2
w1
w1
Table 6: Summary of MR-simplified Bucklin scores of all the alternatives from the partial profile
Q in the proof of Theorem 4.6.
In the other direction, let us assume that the MANIPULATIVE ELICITATION WITH CANDIDATE
PAIR LIMIT instance (A, P, c, f) is a YES instance. Let Q be a partial profile such that P ∈ C(Q)
and the alternative c wins uniquely in Q under the MR-simplified Bucklin voting rule. We can
assume without loss of generality that the alternative w1 is used as a competing alternative
to calculate MR-simplified Bucklin score of every alternative (other than w1 itself) since the
simplified Bucklin score of w1 is 1 (the minimum possible) in P and P ∈ C(Q). Let J ⊆ [t] be
the set of i ∈ [t] such that the corresponding partial preferences in the group G1 in Q leave
the alternatives c and d incomparable. Since f({d, c}) = n − ℓ, we have J 6 ℓ. We claim that
{Sj : j ∈ J} forms a set cover of U. Suppose not, then let uk ∈ U \ (∪j∈JSj). We observe that for
every R ∈ L(Q), the simplified Bucklin score of the alternative ak is at most q + 1. Hence the
MR-simplified Bucklin score of ak in Q is at least −q. On the other hand, since the simplified
Bucklin score of c in P is q + 1 and P ∈ C(Q), the MR-simplified Bucklin score of c in Q is at
most −q. However, this contradicts our assumption that c is the unique MR-simplified Bucklin
winner of Q. Hence {Sj : j ∈ J} forms a set cover of U and thus the SET COVER FREQUENCY TWO
is a YES instance. This concludes the proof of the theorem.
5 Conclusion and Future Work
In this work, we have discovered an important vulnerability, namely manipulative elicitation, in
the use of minimax regret based extension of classical voting rules in the incomplete preferential
setting. Moreover, we have shown that the related computational task is polynomial time solv-
able for many commonly used voting rules including all scoring rules, maximin, Copelandα for
every α ∈ [0, 1], simplified Bucklin voting rules, etc. Then we have shown that by introducing a
parameter per pair of alternatives which specifies the minimum number of partial preferences
where this pair of alternatives must be comparable makes the computational task of manipula-
tive elicitation NP-complete for all the above mentioned voting rules. We want to draw special
attention to the fact that our approach makes manipulative elicitation NP-complete even for the
plurality and veto voting rules which are vulnerable to most of the other manipulative attacks.
In summary, we have found an important vulnerability in the incomplete preferential setting
and proposed a novel approach to tackle it.
A drawback of our approach is that the parameters can be non-uniform – their values do not
need to be the same for every pair of alternatives. It would be interesting to study the computa-
17
tional complexity of the problem when the values of the parameters are all the same. In another
direction, it would be interesting to conduct extensive experimentation to study usefulness of
our approach in practice. This is specially important since computational intractability is known
to provide only a weak barrier in other forms of election manipulation [PR07].
References
[BBN10] Nadja Betzler, Robert Bredereck, and Rolf Niedermeier. Partial kernelization for
In Proc. International Symposium on
rank aggregation: theory and experiments.
Parameterized and Exact Computation (IPEC), pages 26–37. Springer, 2010.
[BBN14] Nadja Betzler, Robert Bredereck, and Rolf Niedermeier. Theoretical and empirical
evaluation of data reduction for exact kemeny rank aggregation. Auton. Agent Multi
Agent Syst., 28(5):721–748, 2014.
[BHN09] Nadja Betzler, Susanne Hemmann, and Rolf Niedermeier. A Multivariate Complex-
ity Analysis of Determining Possible Winners given Incomplete Votes. In Proc. Inter-
national Joint Conference on Artificial Intelligence (IJCAI), volume 9, pages 53–58,
2009.
[BRR11] Dorothea Baumeister, Magnus Roos, and Jorg Rothe. Computational complexity of
two variants of the possible winner problem. In Proc. 10th International Conference
on Autonomous Agents and Multiagent Systems (AAMAS), pages 853–860, 2011.
[CLMM10] Yann Chevaleyre, J´erome Lang, Nicolas Maudet, and J´erome Monnot. Possible
winners when new candidates are added: The case of scoring rules. In Proc. Inter-
national Conference on Artificial Intelligence (AAAI), pages 762–767, 2010.
[DM17]
Palash Dey and Neeldhara Misra. On the exact amount of missing information that
makes finding possible winners hard.
In Proc. 42nd International Symposium on
Mathematical Foundations of Computer Science, 2017.
[DMN15] Palash Dey, Neeldhara Misra, and Y. Narahari. Detecting possible manipulators in
elections.
In Proc. 2015 International Conference on Autonomous Agents and Mul-
tiagent Systems, AAMAS 2015, Istanbul, Turkey, May 4-8, 2015, pages 1441–1450,
2015.
[DMN16a] Palash Dey, Neeldhara Misra, and Y. Narahari. Complexity of manipulation with
partial information in voting. In Proc. Twenty-Fifth International Joint Conference on
Artificial Intelligence, IJCAI 2016, New York, NY, USA, 9-15 July 2016, pages 229–
235, 2016.
[DMN16b] Palash Dey, Neeldhara Misra, and Y. Narahari. Kernelization complexity of possi-
ble winner and coalitional manipulation problems in voting. Theor. Comput. Sci.,
616:111–125, 2016.
[DMN17] Palash Dey, Neeldhara Misra, and Y. Narahari. Frugal bribery in voting. Theor.
Comput. Sci., 676:15–32, 2017.
[FKS03]
Ronald Fagin, Ravi Kumar, and D. Sivakumar. Efficient similarity search and classi-
fication via rank aggregation. In Proc. 2003 ACM SIGMOD International Conference
on Management of Data, SIGMOD '03, pages 301–312, New York, NY, USA, 2003.
ACM.
18
[FRRS14] Piotr Faliszewski, Yannick Reisch, Jorg Rothe, and Lena Schend. Complexity of
manipulation, bribery, and campaign management in bucklin and fallback voting.
In Proc. 13th International Conference on Autonomous Agents and Multiagent Systems
(AAMAS), pages 1357–1358. International Foundation for Autonomous Agents and
Multiagent Systems, 2014.
[GJ79]
[KL05]
[LB11]
Michael R Garey and David S Johnson. Computers and Intractability, volume 174.
freeman New York, 1979.
Kathrin Konczak and J´erome Lang. Voting procedures with incomplete preferences.
In Proc. International Joint Conference on Artificial Intelligence-05 Multidisciplinary
Workshop on Advances in Preference Handling, volume 20, 2005.
Tyler Lu and Craig Boutilier. Robust approximation and incremental elicitation in
voting protocols. In Proc. 22nd International Joint Conference on Artificial Intelligence
(IJCAI), pages 287–293, 2011.
[LPR+07] J´erome Lang, Maria Silvia Pini, Francesca Rossi, Kristen Brent Venable, and Toby
Walsh. Winner determination in sequential majority voting. In Proc. 20th Interna-
tional Joint Conference on Artificial Intelligence (IJCAI), volume 7, pages 1372–1377,
2007.
[LPR+12] J´erome Lang, Maria Silvia Pini, Francesca Rossi, Domenico Salvagnin, Kristen Brent
Venable, and Toby Walsh. Winner determination in voting trees with incomplete
preferences and weighted votes. Auton. Agent Multi Agent Syst., 25(1):130–157,
2012.
[MBC+16] Herv´e Moulin, Felix Brandt, Vincent Conitzer, Ulle Endriss, J´erome Lang, and
Ariel D Procaccia. Handbook of Computational Social Choice. Cambridge Univer-
sity Press, 2016.
[PHG00] David M. Pennock, Eric Horvitz, and C. Lee Giles. Social choice theory and rec-
ommender systems: Analysis of the axiomatic foundations of collaborative filtering.
In Proc. Seventeenth National Conference on Artificial Intelligence and Twelfth Confer-
ence on on Innovative Applications of Artificial Intelligence, July 30 - August 3, 2000,
Austin, Texas, USA., pages 729–734, 2000.
[PR07]
Ariel D. Procaccia and Jeffrey S. Rosenschein. Junta distributions and the average-
case complexity of manipulating elections. J. Artif. Intell. Res., 28:157–181, 2007.
[PRVW07] Maria Silvia Pini, Francesca Rossi, Kristen Brent Venable, and Toby Walsh. Incom-
pleteness and incomparability in preference aggregation. In Proc. 20nd International
Joint Conference on Artificial Intelligence (IJCAI), volume 7, pages 1464–1469, 2007.
[Wal07]
Toby Walsh. Uncertainty in preference elicitation and aggregation. In Proc. Interna-
tional Conference on Artificial Intelligence (AAAI), volume 22, pages 3–8, 2007.
[XC11]
Lirong Xia and Vincent Conitzer. Determining possible and necessary winners under
common voting rules given partial orders. J. Artif. Intell. Res., 41(2):25–67, 2011.
19
|
1909.11650 | 1 | 1909 | 2019-09-25T17:54:03 | Explaining Agent-Based Financial Market Simulation | [
"cs.MA",
"q-fin.TR"
] | This paper is intended to explain, in simple terms, some of the mechanisms and agents common to multiagent financial market simulations. We first discuss the necessity to include an exogenous price time series ("the fundamental value") for each asset and three methods for generating that series. We then illustrate one process by which a Bayesian agent may receive limited observations of the fundamental series and estimate its current and future values. Finally, we present two such agents widely examined in the literature, the Zero Intelligence agent and the Heuristic Belief Learning agent, which implement different approaches to order placement. | cs.MA | cs |
EXPLAINING AGENT-BASED FINANCIAL MARKET SIMULATION
A PREPRINT
David Byrd
School of Interactive Computing
Georgia Institute of Technology
Atlanta, GA 30308
[email protected]
September 26, 2019
ABSTRACT
This paper is intended to explain, in simple terms, some of the mechanisms and agents common
to multiagent financial market simulations. We first discuss the necessity to include an exogenous
price time series ("the fundamental value") for each asset and three methods for generating that series.
We then illustrate one process by which a Bayesian agent may receive limited observations of the
fundamental series and estimate its current and future values. Finally, we present two such agents
widely examined in the literature, the Zero Intelligence agent and the Heuristic Belief Learning
agent, which implement different approaches to order placement.
The True Fundamental Value Times Series
A common feature of agent-based financial simulation is the presence of an exogenous (defined: "derived from ex-
ternal factors") price time series, representing some global consensus value for an asset at any time, arising from the
accumulation of all available news and information which could influence the valuation of the asset. A simple intuition
for the need of such an extrinsic value can be obtained by its omission. Imagine a simulated stock entering its opening
auction period, such that there is not yet a limit order book to examine. In the absence of an extrinsic value of some
kind, what would cause market agents to place bids or offers with any particular limit price? We could imagine using
the prior day's closing price as an anchor, but then what about the first day? There must be some concept of "what this
company is really worth" for at least some of the market participants, or there is no reason for the stock to trade in the
$5 range versus the $5,000 range. In real-world terms, we can imagine consultation of this extrinsic series to be the
distinguishing factor between value investors versus purely technical traders.
This "fundamental value series" requires an omniscient perspective that is not available to any market participant.
Thus at best, market participants may obtain a noisy observation of this value at some particular point in time. This
paper focuses on the set of mean-reverting fundamental value series with characteristics as described in Section 3 of
Chakraborty and Kearns [1].
Properties of the Fundamental
Intuitively, think of the fundamental as a consensus valuation of all people everywhere that somehow reflects the "real"
value of a company. This value changes over time in response to events, and of course no one can ever know precisely
what the global consensus value is.
• There is exactly one fundamental value time series for each equity in the simulation.
• It is a property of the equity and does not vary per agent.
• It is predetermined; agent actions do not affect it.
A PREPRINT - SEPTEMBER 26, 2019
Most any type of numeric time series data can serve as a fundamental value series for an equity. For example, one could
simply use historical market data. We consider three possibilities: a discrete mean reverting series, a continuous mean
reverting series based on the Ornstein-Uhlenbeck process, and a modification of that OU process with "megashock
events".
Discrete Mean Reverting Fundamental
We first consider a discrete mean reverting fundamental time series as presented in Section 4 of Wah et al. [2]
In this case, the fundamental value series for an equity is defined at all times t ∈ [0, T ] by:
rt = max{0, κ¯r + (1 − κ)rt−1 + ut}; r0 = ¯r
The mean fundamental value for an equity (¯r) is chosen as part of experimental configuration. The fundamental value
series r begins at this mean (r0 = ¯r) and is never allowed to become negative.
At each time step t ∈ [0, T ], the series makes a single-step reversion to the mean using parameter κ ∈ [0, 1], the mean
reversion rate. This functions similarly to the learning rate in many machine learning algorithms: α × new_value +
(1 − α) × old_value. Here, new_value is just ¯r (to produce mean reversion). Thus the series is always drifting back
toward the mean at a predictable rate, faster for high κ and slower for low κ. If κ = 0, of course, the time series does
not revert to the mean.
So far this looks very unlike a stock price. When the price deviates from the mean, it slowly and predictably returns to
the mean, then stays there forever. In fact, nothing thus far would make it leave the mean in the first place.
Enter ut ∼ N (0, σ2
s ), the "shock variance". This is a random number drawn at each time step t to perturb the
otherwise completely predictable fundamental value series. It erratically pushes the series around (including away
from the mean) to introduce noise into the price series and create a random walk.
Important: This ut "noise" is part of producing the single, true fundamental value series for an equity. This is not the
observation noise experienced by agents.
Agents are sometimes required to infer the fixed parameters of the mean reverting fundamental, but for simplicity it
is often assumed that agents have become "tuned" to their environment. In either case, agents do know the above
equations, but do not know the true fundamental value after time t = 0, because of the random draws for ut. In the
"tuned" case, agents additionally know that ut ∼ N (0, σ2
s ), and know the correct values for ¯r, κ, and σ2
s .
Ornstein-Uhlenbeck Process Fundamental
We next consider a continuous mean reverting time series as the fundamental, based on the Ornstein-Uhlenbeck (OU)
Process as described in Section 3.1 of Chakraborty and Kearns. [1]
A potential weakness in the discrete mean reverting series is that it must be computed at every discrete time step, in
order, for t ∈ [0, T ]. Time steps at which no agent activity occurs must still be evaluated to obtain future values. If there
are many time steps, and agents tend to arrive at the market infrequently, this can become a significant computational
burden for the simulation.
The OU Process, with values represented by Q, has the advantage that Qt can be computed for any future time t, given
value Q0. Because OU has the Markov property (i.e. Qt+1 is conditionally independent of Qt−1 given Qt) and Qt is
normally distributed, we can determine the expectation and variance of Qt and obtain the value by sampling, without
computing any of the values "in between". In the case of sparse agent arrivals, this can save considerable computation
effort for the simulation.
The OU process requires similar parameters and values to the discrete mean reverting process. The equivalent discrete
series parameter is given in parenthesis after each OU compoment. Specifically, the OU process requires a mean
fundamental µ (equiv: ¯r), a prior value Q0 (equiv: rt−1), a mean reversion rate γ (equiv: κ), and a volatility value σ
(equiv: σ2
s ).
2
A PREPRINT - SEPTEMBER 26, 2019
The value Qt of the OU process at time t can then be sampled from a normal distribution with mean: µ+(Q0 −µ)e−γt,
and variance: σ2
2γ (1 − e−2γt).
Like the discrete mean reverting process, the OU process can only move away from the mean through its volatility
value. If this value is kept small to represent high-frequency "noise" in the process, both processes will appear as
essentially single-scale accumulating white noise around the fundamental mean, and will not look very much like a
plot of, for example, a day of real transactions (or quotes) for some equity traded on a US exchange.
Megashock OU Fundamental
To obtain the simulation computation time improvements of the OU Process Fundamental, while producing a se-
ries that looks more like a real market equity price time series, we introduce the concept of "megashock events".
Megashock events are intended to represent extrinsic news of a substantial nature that occur relatively infrequently,
but have the potential to significantly alter the consensus valuation of a stock.
In technical terms, megashock events are layered on top of the preliminary OU fundamental, arriving via a Poisson
process, and are drawn from a bimodal distribution with mean zero. Each sub-distribution is a Gaussian with mean
significantly different from zero, and a high variance relative to the OU process base variance. The sub-distributions
are centered in positive and negative territory with mean equally far from zero, producing the desired overall mean
zero for the bimodal distribution.
Estimating the Final Fundamental Value
In the simulation described in Wang and Wellman [3], agents attempt to maximize the value of their portfolio at the
end of the simulation relative to the final fundamental value. This is what the closing price "should" be. It is not the
same as marking to market. Regardless, predicting this final fundamental value is an important part of making trading
decisions.
The agents arrive according to a Poisson distribution and a discrete mean reverting process is used for the fundamental.
Assume some agent wakes at time t and receives new (noisy) fundamental observation ot. This observation is unique
to each agent with ot = rt + nt, where rt is the true fundamental value and nt is random observation noise drawn
from N (0, σ2
n). The agent does know σ2
n.
The agent maintains an estimate rt of the current fundamental value rt, and an estimated variance σ2
Important: Again, σ2
(or loosely an "inverse confidence"), not an estimate of any market or simulation parameter.
t is the agent estimating the variance of its rt estimate. Thus, σ2
t of this estimate.
t is like an internal error metric
Let t′ be the last time the agent woke. At that time, it had some estimates (rt′ and σt′ ) of rt′ and its own estimation
error. First it must mentally advance time from t′ to t. With no observations while asleep, it can only apply mean
reversion at each time step:
rt′ ← (1 − (1 − κ)δ)¯r + (1 − κ)δ rt′
In the above update, δ = t − t′, the number of time steps since the agent last woke. Note that with δ = 1, it is exactly
the mean reversion equation used to compute the true fundamental value series rt, except: the agent does not know
rt−1 and so uses its previous estimate rt′ instead, and the agent does not know the random perturbations ut and so
uses the known mean value of zero instead.
With δ > 1, this is equivalent to applying mean reversion δ times. The first time the agent wakes, its previous wake
time is effectively t′ = 0. Thus it should assume rt′ = ¯r, as it knows the value of ¯r and that r0 = ¯r.
To complete the advancement of time from t′ to t, the agent must also update its internal error estimate σt′ , which is
initially zero. It does so as follows:
1 − (1 − κ)2δ
1 − (1 − κ)2 σ2
This update is a weighted mixture of the previous internal error estimate σ2
term for shock variance σ2
t′ ← (1 − κ)2δ σ2
σ2
s . The weight
s starts at 1 when δ = 1 and increases with δ. The weight term for previous internal error
t′ and the shock variance σ2
t′ +
s
3
A PREPRINT - SEPTEMBER 26, 2019
estimate σ2
the more time that has passed since the agent's last observation, the more weight is placed on shock variance σ2
the less weight is placed on the prior error estimate σ2
t′ .
t′ starts at (1 − κ)2 when δ = 1 and decreases as δ increases. Thus when computing the new error estimate,
s and
Having now advanced time from t′ to t, the agent must apply its new observation ot. This process is not unlike
a Dynamic Bayes Network, in which the agent alternates stepping forward in time (transition model) and applying
new evidence (sensor model). The current estimate of the fundamental value is a weighted combination of the prior
estimate with the new observation:
rt =
σ2
n
n + σ2
σ2
t′
rt′ +
σ2
t′
n + σ2
σ2
t′
ot
The weights are simply the observation noise σ2
t′ , normalized to sum to
one. The observation noise weights the agent's prior estimate of fundamental value rt′ and the internal error estimate
weights the new observation ot.
n and the agent's internal estimate of its error σ2
Important: The above weights are reversed from what might be initially expected. This is appropriate because a
higher variance term is resulting in a higher weight, which is the opposite of what we would want, so we switch the
weights. The higher our estimation of our internal error, the more weight we put on the new observation. The higher
the observation noise, the more weight we put on our previous estimates.
The agent must also finish updating its internal error estimate for use the next time it wakes:
σ2
t =
n σ2
σ2
t′
n + σ2
σ2
t′
Similarly to applying the new observation to rt, here we apply one "step" of the observation noise σ2
nal error estimate σ2
"certainty") but with higher observation noise, this effect is diminished.
n to our inter-
t . Making an observation of course tends to reduce our internal error estimate (or increase our
The agent now has updated estimates rt and σ2
t for the current fundamental value and its internal error estimation. At
last, it can compute rt, the final fundamental value rT as estimated at current time t. Having no future observations
(past time t) it can only advance time as before:
rt ← (1 − (1 − κ)T −t)¯r + (1 − κ)T −trt
This matches the previous "time advancement" process with δ = T − t, the number of time steps remaining in the
simulation. Again the agent must assume the random perturbations ut take on their mean value of zero.
Now the agent can use its estimate of the final fundamental value (or what the closing price should be) to make its
trading decisions.
Note that the above estimation method is not substantially different when the fundamental follows the OU process,
because the two processes map very closely.
Agent Preferences
It is important that each agent have individual preferences for holding certain quantities of stock, as we would want
each agent to behave differently even without forcing arbitrary randomness into the decision process. In the real world,
independent of any "consensus value" for INTC, some people will just really want to own Intel. Here we describe the
method as put forth in Wah and Wellman. [2]
These preferences are codified as a vector of incremental private values placed on the acquisition or release of one
additional unit of stock, given the agent's current holdings. If qmax is the holding limit, then the preferences for trading
agent i are the elements θq
i in:
Θi = (θ−qmax +1
i
, . . . , θ0
i , θ1
i , . . . , θqmax
i
)
where q is the quantity of stock currently held. Θi is drawn randomly from N (0, σ2
P V is a selected
experimental parameter. The values are sorted in descending order, ensuring that each additional unit of stock acquired
is valued less than the one before it.
P V ), where σ2
4
A PREPRINT - SEPTEMBER 26, 2019
For example, consider a private value vector for some particular agent:
Θq = (0.5, 0.3, 0.2, 0.1, −0.2, −0.4)
for q ∈ [−2, 3], and assume this agent's rt estimate of the final fundamental value is 100. Since each agent's total
valuation for its qth unit of stock is rt + θq, we can compute the following total valuations:
V q = (100.5, 100.3, 100.2, 100.1, 99.8, 99.6)
If the agent currently holds zero units of stock, it would pay (see caveat in next paragraph) $100.10 to go long one
unit, or demand $100.20 to go short one unit. If the agent is currently short two units, it would pay $100.30 to buy
back a unit, or demand $100.50 to sell another unit. If the agent is long two units, it would pay $99.60 to buy a third
unit, or require $99.80 to sell a unit.
It is this total valuation of a unit (incremental private preference plus estimated final fundamental value) that governs
each trader's limit prices on its orders. However an agent should not bid/offer exactly the limit prices listed above,
as that would result in zero surplus (gain) versus its valuation. Instead it will shade its bids lower or offers higher to
ensure a positive surplus if the order is filled.
Zero Intelligence Agent
The Zero Intelligence (ZI) agent has been a part of the financial literature from at least 1993, when Gode and Sunder
studied the allocative efficiency of constrained or unconstrained ZI agents with that of a population of human traders,
finding that efficiency arose as a consequence of market structure rather than human intelligence or motivation. [4]
Today, the term Zero Intelligence (or Zero Intelligence Plus) is extended to more complex agents, so long as they place
orders priced substantially at random, at times selected substantially at random, and do not possess significant memory,
intelligence, or visibility into the order stream.
Here we again describe the particular flavor of ZI agent used in Wah and Wellman. [2] We note that these agents are
not strictly zero intelligence, because they (1) do observe the current best bid or ask as part of a final rejection process
when placing an order and (2) the same agent arrives at the market multiple times, versus only once in "classic" ZI.
Upon waking, a ZI trader cancels any outstanding order and places a new, single-unit limit order with equal probability
to buy or sell.
The total valuation used by a ZI trader for stock unit q is always as explained in the previous section: rt + θq. If the ZI
agent currently holds q units, then it could contemplate selling unit q with valuation rt + θq or buying unit q + 1 with
valuation rt + θq+1
.
i
To ensure a surplus (gain) from the trade, the agent must shade its limit prices away from this valuation, as pricing at
the valuation would produce zero net gain even if the order is executed. Bids are shaded downward (bid less than your
valuation) and asks are shaded upward (ask more than your valuation).
Call the trader's total valuation for the share in question v. The range [Rmin, Rmax] represents the minimum and
maximum surplus (gain) the agent will demand on a filled order. The agent draws a random number R, called the
requested surplus, from that range. R is the difference between the agent's total valuation of the stock unit and the
"shaded" limit price it actually offers. The agent will gain R surplus if the order trades, or nothing if it does not. Thus
the limit price is v − R if buying, and v + R if selling.
This particular (Wah and Wellman [2]) agent also has a specified experimental parameter η ∈ [0, 1] called the "strategic
threshold". If an agent can immediately secure an executed surplus (gain) of ηR by taking the current best bid or offer,
it will do that instead of placing the order previously described.
Shortcoming: Note that the ZI agent selected a "requested surplus" R at random with some fixed parameter range.
This may well result in its order not being executed at all (too far from the spread). This is the problem solved by
the Heuristic Belief Learner (HBL) agent, which improves overall performance by estimating the likelihood that any
particular limit price will successfully trade, therefore allowing it to bid less or ask more without too badly damaging
the odds that the order will be filled.
5
A PREPRINT - SEPTEMBER 26, 2019
Example of ZI Limit Price Determination
Here we extend the private and total valuations example from the Agent Preferences section, using private value vector:
and total value vector:
Θq = (0.5, 0.3, 0.2, 0.1, −0.2, −0.4)
V q = (100.5, 100.3, 100.2, 100.1, 99.8, 99.6)
for q ∈ [−2, 3], and assuming rt = 100 as before.
An agent holds (long) q = 1 unit of stock. It has been randomly selected to buy on this market arrival. The agent
estimates the final fundamental at 100. Its private valuation for stock unit q + 1 = 2 is −0.20, thus its total valuation
for stock unit 2, buying one unit while already holding one unit, is 99.8.
Assume Rmin = 0.1 and Rmax = 0.5. This means the agent will place orders with limit prices chosen to achieve
at least 10 cents gain (for one unit) but no more than 50 cents gain relative to its total valuation of the stock unit.
Assume the agent randomly selects R to be 0.25 this time. Since the agent is buying, its BID limit price will be
99.8 − 0.25 = 99.55, which will net a gain of 0.25 if the order is filled.
The strategic threshold η could cause the agent not to place the above order. Assume η = 0.5, meaning the agent
will accept one-half of its desired gain if the order can be executed immediately. If the agent can get at least ηR =
0.5(0.25) = 0.125 by accepting the current best offer, it will do that. In this example, if the best offer is at most 99.67,
the agent will place a bid to trade with that offer. Otherwise, it will place the bid with limit price 99.55 as planned.
Heuristic Belief Learning Agent
GD [5], named for its authors Gjerstad and Dickhaut, is at the root of a family tree of strategies that maintain a belief
function representing the probability that a particular auction bid will be accepted based on its price. Heuristic Belief
Learning (HBL) is the name used for a generalized form of GD presented later by Gjerstad. [6] Here we present HBL
as described in Wang and Wellman. [3]
These HBL agents work exactly like the previously described ZI agents (having private preferences, arriving according
to a Poisson distribution, estimating the final fundamental value, etc) except as noted herein.
Using its belief function (likelihood of bid success conditioned on price), an HBL agent places the bid that maximizes
expected surplus, which accounts for both the surplus if the bid is accepted and the likelihood the bid will be accepted.
This belief function is constructed from an accurately observed history of all bids (accepted or rejected) leading
to the last L (memory length) transactions. In the context of a more complex market with an order book that allows
delayed execution and order cancellation, we can instead estimate the probability that a limit order will be successfully
transacted within some maximum time period based on its limit price.
Important: This belief function has nothing to do with price prediction, nothing to do with the "true value" of an item,
and nothing to do with the profitability of a transaction. It simply asks: "If I were to place a bid at this price, what are
the odds I would receive the item I bid on?"
GD and HBL are considered heuristic agents because, as we will see, their estimation of bid success probability uses
an arbitrary (but intuitive) formulation.
Intuitive Formulation of HBL Success Likelihood Estimation
At any given time, the HBL agent can use its memory of observed limit orders and their results to estimate the
probability of a successful transaction given any observed price.
If the HBL wished to buy, we could construct a simple ratio of successful to unsuccessful bids at exactly the proposed
limit price p, but this ratio is unbounded and we want a probability. Normalizing the denominator restricts the range
to [0, 1] and permits our desired interpretation:
P r(p) =
Sp
Sp + Up
6
A PREPRINT - SEPTEMBER 26, 2019
where Sp is the number of successful bid orders at price p, and Up is the number of unsuccessful bid orders at price p.
The agent thus asks: "Of all bids at price p, what proportion were successful?"
This approach has an obvious limitation: data sparsity. If there have been ten unsuccessful bids at p = 5, ten successful
bids at p = 7, and ten successful bids at p = 9, the value P r(8) = 0
0+0 is undefined. If we extend S and U to include
bids equal or less than p:
P r(p) =
S≤p
S≤p + U≤p
the agent now asks: "Of all bids at or below price p, what proportion were successful?". Now P r(8) = 10
which is at least something.
10+10 = 0.5,
Including unsuccessful bids at prices lower than our proposed price is unhelpful, though. If we are offering 8, why
should failed bids at 5 affect our estimation of success? P r(8) = 0.5 seems unrealistic; all bids at p = 7 were
accepted!
Instead, we should evaluate successful bids at or below p, but unsuccessful bids at or above p. The agent now considers
two different factors: "How many bids succeeded offering no more than this?" and "How many bids failed despite
offering at least this much?"
P r(p) =
S≤p
S≤p + U≥p
We now appropriately disregard the failed bids at 5, because we are offering more than that, and estimate P r(8) =
10+0 = 1, because we have never observed the failure of a bid with a price of at least 8.
10
There is still one problem. Orders in our type of auction can be cancelled. Consider the ten unsuccessful bids at p = 5.
Were they unsuccessful because there has never been an ask at p <= 5? Or were they merely unlucky, and there have
been offers that low, but not at the same time as those bids? In the second case, our estimation of P r(5) should be
greater than zero, because although no bid succeeded, one clearly could have.
Our agent also has limited memory. What if it should happen that within this memory, there were no bids at or below
p (successful or not), but there were ask orders below p? Our heuristic might end up zero or undefined, when again,
the chance of fulfillment is near certain due to the volume of unmatched ask orders.
Generally speaking then, the volume of asks in our considered price range (<= p) should be a factor. In addition to
considering bids and their success rate, the more shares that have been available at or below p, the more likely we are
to succeed with a bid at p. The fewer shares that have been available, the less likely we are to succeed.
Thus we arrive at the final formulation for the HBL agent's heuristic estimation of success, with this rough interpreta-
tion: "Seeing ask order volume, or successful bid orders, at prices no more than p increases the likelihood of success.
Seeing unsuccessful bid orders at prices of at least p decreases the likelihood of success." The ratio is formalized and
normalized as:
P r(p) =
A≤p + S≤p
A≤p + S≤p + U≥p
where A≤p is the total number of ask orders at price p or lower, S≤p is the number of successful bid orders at price
p or lower, and U≥p is the number of unsuccessful bid orders at price p or higher. The heuristic calculation is exactly
mirrored when the agent considers selling a unit.
The agent can now estimate the probability of success for any candidate limit order price p. If it is desired to estimate
probability for all possible prices without bound, cubic spline interpolation can be employed as suggested by Wang
and Wellman. [3]
7
A PREPRINT - SEPTEMBER 26, 2019
HBL Considerations and Strategy
In a market with an order book, orders are almost never rejected. They simply enter the order book to await a later
match. The HBL, though, depends on the idea of rejected or unsuccessful orders. A simple approach to this problem
would be to place a time limit on each order.
If the order is not matched within this time limit, it is considered
unsuccessful.
The approach taken by Wang and Wellman is a little more complex. [3] They define a grace period relative to the
frequency of trader arrivals (the Poisson distribution mentioned earlier). An order is considered rejected if it existed in
the order book for a length of time greater than the grace period. All orders not accepted immediately are considered
"partially rejected" and assigned a fractional weight depending on the length of time they spent in the order book.
The HBL agent selects the limit price that maximizes expected surplus, where surplus is simply the difference between
the limit price p and the agent's total valuation of a share (see "Agent Preferences"), and expected surplus is the same,
weighted by the probability of successful transaction computed in the previous section.
For example, an HBL agent contemplating buying unit q + 1 would consider total valuation rt + θq+1
and estimation of successful transaction P r(p), and select its optimal bid p∗ as:
i
, limit price p,
p∗ = argmaxp(rt + θq+1
i − p)P r(p)
Anytime the HBL agent does not have enough information to enact the above strategy, it temporarily behaves as a ZI
agent instead.
Limitation: HBL assumes the agent can observe a complete order stream from all agents, including order execution
(even if delayed), cancellation, and the length of time each order was "alive" in the order book. This may not be
practical in many environments.
HBL Limit Price Example
Here we reuse the private and total valuations example from the Example of ZI Limit Price Determination section,
using private value vector:
and total value vector:
Θq = (0.5, 0.3, 0.2, 0.1, −0.2, −0.4)
V q = (100.5, 100.3, 100.2, 100.1, 99.8, 99.6)
for q ∈ [−2, 3], and assuming rt = 100 as before.
An agent is short one unit of stock (q = −1). It has been randomly selected to buy on this market arrival. The agent
estimates the final fundamental at 100. Its private valuation for stock unit q + 1 = 0 is 0.20, thus its total valuation for
stock unit 0, buying one unit while being short one unit, is 100.2.
Assume the HBL agent has access to the following order memory of length L = 4, with superscripts denoting pairs of
matched orders, and that no orders were cancelled:
Transaction 1
ASK 100.01
99.8
BID
ASK 100.34
99.6
BID
BID 100.01
Transaction 2
ASK 100.22
BID 100.0
ASK 100.3
BID 100.13
BID 100.22
Transaction 3
BID 100.1
99.93
ASK
Transaction 4
BID 100.2
ASK 100.4
BID 100.44
8
The agent can compute the estimated probability of transaction for a proposed limit price p as described in the previous
section:
A PREPRINT - SEPTEMBER 26, 2019
P r(100.5) =
P r(100.4) =
P r(100.3) =
P r(100.2) =
P r(100.1) =
6 + 4
6 + 4 + 0
6 + 4
6 + 4 + 0
5 + 3
5 + 3 + 0
3 + 3
3 + 3 + 1
2 + 2
2 + 2 + 2
= 1
= 1
= 1
P r(100.0) =
P r(99.9) =
P r(99.8) =
= 0.86
P r(99.7) =
= 0.67
P r(99.6) =
2 + 1
2 + 1 + 3
1 + 0
1 + 0 + 3
0 + 0
0 + 0 + 4
0 + 0
0 + 0 + 4
0 + 0
0 + 0 + 5
= 0.50
= 0.25
= 0
= 0
= 0
Now the agent can use its total valuation, the proposed limit price, and the estimated probability of transaction to
compute expected surplus E(s):
E(sp = 100.5) = (100.2 − 100.5)(1) = −0.3
E(sp = 100.4) = (100.2 − 100.4)(1) = −0.2
E(sp = 100.3) = (100.2 − 100.3)(1) = −0.1
E(sp = 100.2) = (100.2 − 100.2)(0.86) = 0
E(sp = 100.1) = (100.2 − 100.1)(0.67) = 0.067
E(sp = 100.0) = (100.2 − 100.0)(0.5) = 0.1
E(sp = 99.9) = (100.2 − 99.9)(0.25) = 0.075
E(sp = 99.8) = (100.2 − 99.8)(0) = 0
E(sp = 99.7) = (100.2 − 99.7)(0) = 0
E(sp = 99.6) = (100.2 − 99.6)(0) = 0
Intuitively, the agent can see that any price above $100.2 would almost certainly be accepted, but the agent would lose
money relative to its valuation of the stock unit. The agent would love to achieve the surpluses associated with prices
less than $99.9, but it can guess that those bids would never transact. Only the prices p ∈ [99.9, 100.1] have a nonzero
chance of transaction and a positive surplus if transacted.
In this example, the agent will place a limit buy order at limit price p = 100.0, the price that produces the maximum
expected surplus of $0.10.
The ABIDES Simulator
Our simulation platform, ABIDES (Agent-Based Interactive Discrete Event Simulation) provides the funda-
including the novel "Megashock OU Fun-
mental value series and agents discussed herein "out of the box",
damental".
[8]
The ABIDES platform is available under a BSD-style license at https://github.com/abides-sim/abides.
The distribution includes relevant example configurations which are further explained in the project's wiki at
https://github.com/abides-sim/abides/wiki.
[7] It also provides the ability to use real market data as the fundamental value series.
The ABIDES simulator uses the OU process (plus megashock events) to provide a "sparse discrete" fundamental pro-
cess that provides extremely fine time resolution (nanoseconds) while still permitting quick simulation of reasonable
time scales (days) due to its ability to completely skip computation of time periods during which nothing happens. The
Megashock OU fundamental additionally produces intraday time series that more closely resemble typical real-world
intraday stock charts.
The ABIDES implementation of Zero Intelligence agents reproduces the strategy described herein, including the
strategic threshold parameter. The implementation of Heuristic Belief Learning agents treats an order as successful if
any part of it is transacted within the observed order stream, and unsuccessful otherwise.
9
A PREPRINT - SEPTEMBER 26, 2019
Acknowledgements
This material is based on research supported in part by the National Science Foundation under Grant no. 1741026, and
by a JPMorgan AI Research Fellowship.
References
[1] Tanmoy Chakraborty and Michael Kearns. Market making and mean reversion. In Proceedings of the 12th ACM
conference on Electronic commerce, pages 307 -- 314. ACM, 2011.
[2] Elaine Wah, Mason Wright, and Michael P Wellman. Welfare effects of market making in continuous double
auctions. Journal of Artificial Intelligence Research, 59:613 -- 650, 2017.
[3] Xintong Wang and Michael P Wellman. Spoofing the limit order book: An agent-based model. In Proceedings
of the 16th Conference on Autonomous Agents and MultiAgent Systems, pages 651 -- 659. International Foundation
for Autonomous Agents and Multiagent Systems, 2017.
[4] Dhananjay K Gode and Shyam Sunder. Allocative efficiency of markets with zero-intelligence traders: Market as
a partial substitute for individual rationality. Journal of political economy, 101(1):119 -- 137, 1993.
[5] Steven Gjerstad and John Dickhaut. Price formation in double auctions. Games and economic behavior, 22(1):1 --
29, 1998.
[6] Steven Gjerstad. The competitive market paradox. Journal of Economic Dynamics and Control, 31(5):1753 -- 1780,
2007.
[7] David Byrd, Maria Hybinette, and Tucker Hybinette Balch. ABIDES: towards high-fidelity market simulation for
AI research. CoRR, abs/1904.12066, 2019.
[8] Tucker Hybinette Balch, Mahmoud Mahfouz, Joshua Lockhart, Maria Hybinette, and David Byrd. How to eval-
arXiv preprint
uate trading strategies: Single agent market replay or multiple agent interactive simulation?
arXiv:1906.12010, 2019.
10
|
1905.13225 | 1 | 1905 | 2019-05-29T18:26:26 | Modeling Theory of Mind in Multi-Agent Games Using Adaptive Feedback Control | [
"cs.MA",
"cs.AI",
"q-bio.NC",
"stat.ML"
] | A major challenge in cognitive science and AI has been to understand how autonomous agents might acquire and predict behavioral and mental states of other agents in the course of complex social interactions. How does such an agent model the goals, beliefs, and actions of other agents it interacts with? What are the computational principles to model a Theory of Mind (ToM)? Deep learning approaches to address these questions fall short of a better understanding of the problem. In part, this is due to the black-box nature of deep networks, wherein computational mechanisms of ToM are not readily revealed. Here, we consider alternative hypotheses seeking to model how the brain might realize a ToM. In particular, we propose embodied and situated agent models based on distributed adaptive control theory to predict actions of other agents in five different game theoretic tasks (Harmony Game, Hawk-Dove, Stag-Hunt, Prisoner's Dilemma and Battle of the Exes). Our multi-layer control models implement top-down predictions from adaptive to reactive layers of control and bottom-up error feedback from reactive to adaptive layers. We test cooperative and competitive strategies among seven different agent models (cooperative, greedy, tit-for-tat, reinforcement-based, rational, predictive and other's-model agents). We show that, compared to pure reinforcement-based strategies, probabilistic learning agents modeled on rational, predictive and other's-model phenotypes perform better in game-theoretic metrics across tasks. Our autonomous multi-agent models capture systems-level processes underlying a ToM and highlight architectural principles of ToM from a control-theoretic perspective. | cs.MA | cs |
Modeling Theory of Mind in Multi-Agent Games Using
Adaptive Feedback Control
Ismael T. Freirea,c,∗, Xerxes D. Arsiwallaa,b,c, Jordi-Ysard Puigb`oa,c, Paul
Verschure a,b,c,d,∗
aInstitute for Bioengineering of Catalonia (IBEC), Barcelona, Spain
bUniversitat Pompeu Fabra (UPF), Barcelona, Spain
cBarcelona Institute of Science and Technology (BIST), Barcelona, Spain
dCatalan Institution for Research and Advanced Studies (ICREA), Barcelona, Spain
Abstract
A major challenge in cognitive science and AI has been to understand how
autonomous agents might acquire and predict behavioral and mental states of
other agents in the course of complex social interactions. How does such an agent
model the goals, beliefs, and actions of other agents it interacts with? What are
the computational principles to model a Theory of Mind (ToM)? Deep learning
approaches to address these questions fall short of a better understanding of
the problem.
In part, this is due to the black-box nature of deep networks,
wherein computational mechanisms of ToM are not readily revealed. Here, we
consider alternative hypotheses seeking to model how the brain might realize a
ToM. In particular, we propose embodied and situated agent models based on
distributed adaptive control theory to predict actions of other agents in five dif-
ferent game theoretic tasks (Harmony Game, Hawk-Dove, Stag-Hunt, Prisoners
Dilemma and Battle of the Exes). Our multi-layer control models implement
top-down predictions from adaptive to reactive layers of control and bottom-
up error feedback from reactive to adaptive layers. We test cooperative and
competitive strategies among seven different agent models (cooperative, greedy,
tit-for-tat, reinforcement-based, rational, predictive and other's-model agents).
∗ITF and XDA equally contributed to this work.
Corresponding authors: ITF, PV
Email addresses: [email protected] (Ismael T. Freire),
[email protected] (Paul Verschure )
Preprint submitted to Artificial Intelligence
June 3, 2019
We show that, compared to pure reinforcement-based strategies, probabilistic
learning agents modeled on rational, predictive and other's-model phenotypes
perform better in game-theoretic metrics across tasks. Our autonomous multi-
agent models capture systems-level processes underlying a ToM and highlight
architectural principles of ToM from a control-theoretic perspective.
Keywords: Multi-Agent Systems, Cognitive Architectures, Theory of Mind,
Game Theory, Reinforcement Learning
1. Introduction
How do autonomous social agents model goals, beliefs and actions of other
agents they interact with in complex social environments? This has long been a
central question in cognitive science and philosophy of mind. It forms the basis
of what is known as Theory of Mind (ToM), a long-standing problem in cognitive
science concerned with how cognitive agents form predictions of mental and
behavioral states of other agents in the course of social interactions [1, 2]. The
precise mechanisms by which the brain achieves this capability is not entirely
understood. Furthermore, how a ToM can be embodied in artificial agents
is very relevant for the future course of artificial intelligence (AI), especially
with respect to human-machine interactions. What can be said concerning
architectural and computational principles necessary for a Theory of Mind? As
a small step in that direction, we propose and validate control-based cognitive
architectures to predict actions and models of other agents in five different game
theoretic tasks.
It is known that humans also use their ToM to attribute mental states,
beliefs, and intentions to inanimate objects [3], suggesting that the same mech-
anisms governing ToM may be integral to other aspects of human cognition. We
argue that in order to understand both, how ToM is implemented in biological
brains and how it may be modeled in artificial agents, one must first identify
its architectural principles and constraints from biology.
Several notable approaches addressing various aspects of this problem al-
2
ready exist, particularly, those based on artificial neural networks [4, 5]. Recent
work in [6] on Machine Theory of Mind is one such example. This work pro-
poses that some degree of theory of mind can be autonomously obtained from
pure Reinforcement Learning (RL). However, the use of Black-Box Optimiza-
tion algorithms, hinders the understanding one can obtain at the mechanistic
level. Because BBO algorithms are able to approximate any complex function
(as suggested for Long-Short Term Memory (LSTM)-based neural networks [7]),
one cannot use that to decipher specific mechanisms that may underlie a ToM
in these systems.
Another interesting approach to ToM follows the work of [8, 9, 10, 11] which
uses hierarchical Bayesian inference. These methods are cognitively-inspired
and suggest the existence of a "psychology engine" in cognitive agents to pro-
cess ToM computations. Nevertheless, the challenge for this approach remains
explaining how the computational cost of running these models might be im-
plemented in biological substrates of embodied and situated agents.
A third approach to this problem follows from work on autonomous multi-
agent models (see [12] for a recent survey). This approach has had its roots in
statistical machine learning theory and robotics. It includes agent models ca-
pable of policy reconstruction, type-classification, planning, recursive reasoning
and group modeling. In a sense, this is closest to the approach we take in this
work, even though, our models are grounded in cognitive control systems. Open
challenges in the field of autonomous multi-agent systems include modeling fully
embodied agents that operate with only partial observability of their environ-
ment and are flexibly able to learn across tasks (including meta-learning), when
interacting with multiple types of other agents.
The cognitive agent models we propose here, advance earlier work on Control-
based Reinforcement Learning (CRL) [13], where we studied the formation of
social conventions in the Battle of the Exes game. The CRL model implements
a feedback control loop handling the agent's reactive behaviors (pre-wired re-
flexes), along with an adaptive layer that uses reinforcement learning to maxi-
mize long-term reward. We showed that this model was able to simulate human
3
data in the above-mentioned coordination game.
The new contribution of the current paper is to advance real-time learning
and control strategies to model agents with specific phenotypes, that can learn to
either model or predict the opposing agents actions. Using top-down predictions
from the adaptive to reactive control layers and bottom-up error feedback from
reactive to adaptive layers, we test cooperative and competitive strategies for
seven different multi-agent behavioral models. The purpose of this study is
to understand how the architectural assumptions behind each of these models
impact agent performance across standard game-theoretic tasks, and what this
implies for the development of an artificial embodied ToM.
2. Methods
2.1. Game Theoretic Tasks
Benchmarks inspired by game theory are becoming standard in the multi-
agent reinforcement learning literature [4, 5, 6, 14]. However, most of the work
developed in this direction presents models that are tested in one single task or
environment [15, 16, 17, 13, 18, 19], at best two [20, 21]. This raises a funda-
mental question about how these models generalize to deal with a more general
or diverse set of problems. Therefore, this approach does not readily enable
one to extract principles and mechanisms or unravel the dynamics underlying
human cooperation and social decision-making.
In this work, we want to go a step further and propose a five-task benchmark
for predictive models based on classic normal-form games extracted from Game
Theory: the Prisoner's dilemma, the Harmony game, the Hawk-dove, the Stag-
hunt and the Battle of the Exes.
In its normal-form, games are depicted in a matrix that contains all the pos-
sible combinations of actions that players can choose, along with its respective
rewards. The most common among them are the dyadic games known as social
dilemmas, in which players have to choose between two actions: cooperate or
defect. One action (cooperate) is more generous and renders a good amount of
4
reward to each player if both choose it, but gives very poor results if the other
player decides to defect. On the other hand, the second action (defect) provides
a significant individual reward if its taken alone, but a very small one if both
choose it.
Although simple in nature, these dyadic games are able to model key ele-
ments of social interaction, such as the tension between the benefit of a coop-
erative action and the risk (or temptation) of free-riding. This feature can be
described in a general form, such as:
Cooperate Defect
Cooperate
Defect
R, R
T, S
S, T
P, P
Table 1: A Social Dilemma in matrix-form
This matrix represents the outcomes of all possible combination of actions
between the row player and the column player. R stands for "reward for mutual
cooperation"; T for "temptation of defecting"; S for the "sucker's payoff for non-
reciprocated cooperation"; and P for "punishment for mutual defection". By
manipulating the relationship between the values of this matrix, many different
situations can be obtained that vary in terms of what would be an optimal
solution (or Nash equilibria [22]).
The Prisoner's Dilemma (table 2) represents the grim situation in which
the temptation of defecting (T) is more rewarding that mutual cooperation
(R), and the punishment for mutual defection (P) is still more beneficial than a
failed attempt of cooperation (S). This relationship among the possible outcomes
can be stated as T > R > P > S. Mutual defection is the only pure Nash
equilibrium in this game since there is no possibility for any player to be better
off by individually changing its own strategy.
On the Stag-Hunt (table 3) we face a context in which mutual cooperation
(R) gives better results and individual defection (T), but at the same time a
failed cooperation (S) is worse than the punishment for mutual defection (P). In
5
this case, R > T > P > S, there are two Nash equilibria: mutual cooperation
and mutual defection.
Hawk-dove (table 4) presents a scenario in which temptation (T) is more
rewarding than cooperation (R), but a mutual defection (P) is less desirable than
non-reciprocated cooperation (S). So, for a relationship of T > R > S > P like
this one, there are three Nash equilibria: two pure anti-coordination equilibrium,
in which each player chooses always the opposite action of its opponent, and
one mixed equilibrium, in which each player probabilistically chooses between
the two pure strategies.
The fourth type of social dilemma is the Harmony game (table 5). In this case
(R > T > S > P ), the game has only one pure equilibrium, pure cooperation,
since mutual cooperation (R) renders a better outcomes that the temptation
to defect (T), and also the penalty for failing to cooperate (S) is less than the
punishment for mutual defection (P).
Finally, as the last game, we will introduce a coordination game, the Battle
of the Exes [23, 24]. In this type of game, the main goal is to achieve coordi-
nation between two players (either congruent coordination on the same action
or incongruent coordination upon choosing different actions), since the failure
to do so is heavily penalized (see table 6). Following the previous nomenclature
applied in the social dilemmas, we could define this game as T > S; R, P = 0.
This game has two pure-dominance equilibria, in which one player chooses the
more rewarding action and the other the low rewarding action; and a turn-
taking equilibrium, in which players alternate over time in choosing the more
rewarding action.
This selection of games provides enough variability to test how learning
agents can perform across different contexts, so we avoid problems derived from
over-fitting on a specific payoff distribution, or related to the possibility of a
model to exploit/capitalize on certain features of a game that could not have
been predicted beforehand. Moreover, a similar selection of games have been
tested in human experiments, proving to be sufficient to classify a reduced set of
behavioral phenotypes across games in human players [25, 26]. Using the above
6
Cooperate Defect
Cooperate Defect
Cooperate
Defect
2, 2
3, 0
0, 3
1, 1
Cooperate
Defect
2, 2
3, 1
1, 3
0, 0
Table 2: prisoner's dilemma
Table 4: hawk-dove
Cooperate Defect
Cooperate Defect
Cooperate
Defect
3, 3
2, 0
0, 2
1, 1
Cooperate
Defect
3, 3
2, 1
1, 2
0, 0
Table 3:
stag-hunt
Table 5: harmony game
A
B
A 0, 0
1, 4
B 4, 1
0, 0
Table 6: battle of the exes
games as benchmarks, below we describe the control-theoretic framework and
seven agent models that capture phenotypes relevant for testing ToM capabili-
ties across games.
2.2. Control-Based Reinforcement Learning
Our starting point is the Control-based Reinforcement Learning (CRL) model
presented in previous work [13]. The CRL is a biologically grounded cognitive
model composed of two control layers (Reactive and Adaptive, see Figure 1),
and based on the principles of the Distributed Adaptive Control (DAC) theory
[27]. The Reactive Layer represents the agent's pre-wired reflexes and serves for
real-time control of sensorimotor contingencies. The Adaptive Layer endows the
agent with learning abilities that maximize long-term reward by choosing which
action to perform in each round of the game. This layered structure allows for
top-down and bottom-up interactions between the two layers [28], resulting in
an optimal control at different time-scales: within each round of play and across
7
rounds [13].
Figure 1: A) Representation of the Control-based Reinforcement Learning (CRL) model. The
top red box depicts the Adaptive Layer or the original CRL model, composed of an Actor-
Critic Temporal-Difference (TD) learning algorithm. The bottom green box represents the
Reactive layer, with its three sets of sensors, one for the 'cooperate' location (sC), one for the
'defect' location (sD), and one for the other agent (sA); the two reactive behaviors, "approach
cooperate location" (fC), "approach defect location" (fD); and the two motors, one for the left
wheel (ml) and one for the right wheel (mr). Between the two layers, the inhibitory function
(i) regulates which reactive behaviors will be active depending on the action received from
the Adaptive layer, while the error monitoring function (pe) manages the mismatch between
the opponent's predicted behavior and the actual observation in real-time.
The Reactive Layer is inspired by Valentino Braitenberg's Vehicles [29] and
presents the intrinsic mechanic behaviors that are caused by a direct mapping
between the sensors and the motors of the agent. The main one, "orienting to-
wards rewards", is based on a cross excitatory connection and a direct inhibitory
connection between the reward sensors and the two motors. This results in an
approaching behavior in which the agents will turn towards a reward location
increasingly faster the closer they detect it. Since the agents are equipped with
two sets of sensors specifically tuned for detecting each reward location, they
also have two instances of this reactive behavior, one for the 'cooperative' and
8
one for the 'individual' reward (see Figure 1 green box).
The Adaptive Layer is in charge of prediction, learning, and decision-making.
Each of the Agent Models described in the next section instantiates a different
Adaptive layer. Its main role is to learn from previous experience and to decide
at the beginning of each round which action the agent will take. When playing
normal-form games, the Adaptive Layer receives as inputs the state of the en-
vironment, which corresponds with the outcome of the last round of the game,
and the reward obtained because of that final state. This information is used
to produce an action as the output. The states S can be either 'R', 'S', 'T' or
'P' (see Table 1), and determines what is the reward that the agent will obtain
based on the specific payoff matrix of the game being played. The actions A
are two: "cooperate" or "defect".
Depending on the action that the Adaptive Layer selects, there is a top-down
selective inhibition that affects the reactive behaviors of the Reactive Layer.
If the action selected is "cooperate", the "approach individual reward" reactive
behavior is inhibited, thus focusing the agent's attention only on the cooperative
reward and on the other agent. Conversely, if the action selected is "defect",
the reactive behavior inhibited will be "approach cooperative reward". This
mechanism aims to mimic how biological systems execute top-down control over
a hierarchy of different control structures [30, 31, 32].
The error monitoring function works in real-time as a bottom-up mechanism
that signals an error function from the Reactive layer sensors to the Adaptive
layer decision-making module. The error signal is only triggered when the agent
detects an inconsistency between its initial prediction of the opponent's action
and the real-time data obtained by its sensors. If a prediction error occurs, the
error monitoring function will update the current state of the agent and this
will make the decision-making module output a new action. Along with the
top-down inhibitory control, this module is inspired by evidence from cognitive
science about the role of bottom-up sensory stimuli in generating prediction
errors [30, 33, 34].
9
2.3. Agent Models
2.3.1. Original Model
Figure 2: Representations of the Original model. Panel A shows a detailed representation of
the algorithm components, as implemented in the Adaptive layer of the original implemen-
tation of the CRL model [13]. Panel B shows a compressed representation showing only the
inputs and outputs of that same model.
The Original CRL model uses an Actor-Critic version of the Temporal-
Difference (TD) learning algorithm [35] for maximizing long-term reward (see
[13] for a detailed description of the implementation).
In brief, the TD-learning algorithm selects an action according to a given
policy P (a = ats = st−1) . Once a round of play is finished, the reward r(st)
obtained by the agent will update the TD-error e signal, following: e(st) =
r(st) + γVΠ(st) − VΠ(st−1) where γ is a discount factor and Vπ(st) = γr(st) is
the Critic. Finally, the policy (or Actor) will be updated following Π(at, st−1) =
Π(at, st−1) + δe(st−1), where δ is a learning rate that is set to 0.15.
2.3.2. Rational Model
The Rational Model is a predictive model that represents an ideal perfectly-
rational and self-interested player provided that its predictions are correct. Its
function is to serve as a benchmark for the other predictive models since once it
learns to predict its opponent's actions accurately, it will always respond auto-
matically with the best response to that predicted action. It is composed of two
main functions: a prediction module and a deterministic utility maximization
function (see Figure 3). The first module tries to predict the next action of the
opponent by using a TD-learning algorithm that uses the opponent's previous
10
Figure 3: Representation of the Rational Model. This model is composed of a predictive
module (RL) that learns to predict the opponent's future action and a utility maximization
function (U) that computes the action that yields the highest reward based on the opponent's
predicted action. At the end of each round, the RL module is updated based on its prediction
error.
state as an input. Once the prediction is made, the second module calculates
the action that will render the highest reward assuming that the opponent has
chosen the predicted action.
This predictive model uses the same TD-learning algorithm as the origi-
nal, but substitutes the explicit reward for an implicit reward that comes from
the error in predicting the other agent's action. We call this reward 'implicit'
because is calculated internally and it is based on whether the action of the
opponent was accurately predicted or not. That way, an accurate prediction
will render a positive reward, while an incorrect one will receive a negative one.
The implicit reward signal is calculated by the function:
1,
if Aot = Aot − 1
−1, otherwise
Ri =
2.3.3. Predictive Model
(1)
The Predictive model has two distinct RL modules: one that learns to predict
and one that chooses an action based on that prediction (see Figure 4). The
predictive RL algorithm is same as the one used by the Rational agent described
above, so it also learns from the implicit reward obtained by Equation 1. The
prediction generated at the beginning of each round by the predictive-RL is sent
as an input to the second RL algorithm. This one, in turn, uses the combined
11
Figure 4: Representation of the Predictive Model. This model is composed of a predictive
module (RL) that learns to predict the opponent's future action and a TD learning module
(RL) that uses that prediction along with the previous state to learn the optimal policy. At the
end of the round, the predictive-RL (green) is updated according to its error in the prediction.
information of the opponent's predicted action and the state of the previous
round to learn the optimal action. For its update function, the second RL
algorithm uses the explicit reward obtained in that round of the game.
2.3.4. Other's Model
Figure 5: Representation of the Other's-Model. This model is composed by two TD learning
algorithms. The first one (left, blue) learns to predict the opponent's future action while the
second (right, red) uses a that prediction to learn the optimal policy. The first algorithm is
updated with the opponent's reward and the second with the agent's own reward.
The Other's Model is also composed by RL algorithms, one predictive and
one for learning the optimal policy (see Figure 5. Technically, the second RL
module is identical to the one used by the Predictive model: it integrates both
the previous state of the game and the opponent's predicted action in order to
learn the optimal policy. However, it differs from the Predictive model in the
way its first module is designed. In this case, the predictive-RL algorithm is
12
updated by the explicit reward that the opponent has obtained, in an attempt
to create an internal model of the other agent's policy. At the functional level
is very similar to the Predictive model, but with an important difference: while
the Predictive model tries predict the opponent's action by focusing on its overt
behavior, the Other's Model does it trying to learn the internal policy of its
opponent.
2.3.5. Deterministic Agent Models
In order to test the correct functioning of the predictive capacities of the
ToM agents described above, we have developed a number of agents have a
fixed behavior or policy. The first two represent two behavioral phenotypes
observed in humans [25] while the third one is a classic benchmark of Game
Theory [36].
2.3.5.1. Greedy Agent Model. This model implements a simple behavioral strat-
egy of pure self-utility maximization: it always chooses the action the renders
the highest reward to itself, without taking into account the opponent's action.
So for the games described in this work, it will always choose to 'defect' with
the exception of the Stag-Hunt and the Harmony Game, where the cooperative
reward can give a higher reward than the temptation to defect (R > T ).
cooperate,
def ect,
πg =
if R > T
otherwise
(2)
2.3.5.2. Cooperative / Nice Agent Model. As a deterministic counterpart of the
Greedy model, the Nice model executes an even simpler deterministic strategy:
it will always choose cooperation.
πn = P (a = cooperatest) = 1
(3)
2.3.5.3. Tit-for-Tat Agent Model. This simple yet powerful strategy became
popular in the famous Axelrod's tournament [37] where it won to all competing
algorithms and strategies in a contest based on the Iterated Prisoner's Dilemma.
13
We have introduced this opponent because we expect that this agent will give
problems to the predictive models. Tit-for-tat starts always by cooperating,
and from then on it always chooses the last action made by its opponent. Since
its capacity to switch actions does not depend on a specific policy, it cannot be
learned by just taking into account variables such as its previous state or its
reward. It can be described by the following equation, where aopponent is the
action made by the opponent at t − 1:
cooperate,
if t = 0
aopponent(t − 1), otherwise
πg =
2.4. Experimental Setup
(4)
First, we perform four different experiments in which we test the four learn-
ing models described above (Original, Rational, Predictive, Other's-Model) in
each of the five games described in section 2.1 (Prisoner's Dilemma, Hawk-Dove,
Stag-Hunt, Harmony game and Battle of the Exes) and against opponents of
different level of complexity (Greedy, Nice, Tit-for-tat and Original), resulting
in a 4x5x4 experimental setup.
This way, in experiment one, all models are tested against the Greedy agent;
in experiment two, against the Nice agent; in experiment three, against the Tit-
for-tat agent; and in experiment four, against the Original agent.
In all experiments, each model plays 50 times each of the five games, during
a total amount of 1000 rounds per iteration (i.e., 5 games, 50 dyads per game,
1000 rounds per dyad). On each iteration, every model starts learning from
scratch, with no previous training period.
After that, we perform a fifth experiment to study in a real-time spatial
version of the games, special cases in which the models encountered problems
in the previous setting. So, in this first four experiments, we use the discrete-
time version of the games, and for the fifth experiment, a real/continuous time
version of the tasks.
In the discrete-time version, both agents simultaneously choose an action at
the beginning of each round. Immediately after that, the outcome of the round
14
is calculated and each agent receives a reward equal to the value stated in the
payoff matrix of the game.
In the continuous-time version, the two agents can choose to go to one of the
two equally distant locations that represent the two actions of a 2-action matrix-
form game (see Figure 6-B). In this version, agents also choose an action at the
beginning of the round, but then they have to navigate towards the selected
location until one of them reaches its destination. The real-time dimension of
this implementation allows the possibility of changing the action during the
course of the round as well as to receive feedback in real-time. Each round of
the game begins with both agents in their initial positions, as shown in Figure
6-B. When the agents reach a reward location, the round ends, the rewards
are distributed accordingly to the payoff matrix and the game restarts with the
agents back in their starting positions.
Figure 6: A) A top-down visualization of the agents used in the continuous version of the
games. The green circles represent the location-specific sensors, and the red circles the agent-
specific ones. The green lines that connect the location sensors with the wheels represent
the Braitenberg-like excitatory and inhibitory connections. B) Image of initial conditions
in the continuous/real-time spatial version of the tasks. The blue circles are the two agents
facing each other (representing two ePuck robots viewed from the top). The big green circle
represents the 'cooperate' reward location; the small green circle, the 'defect' reward. The
white circles around each reward spot mark the threshold of the detection area.
The agents of the continuous version are embodied in virtual ePuck robots
(see 6-A. They are equipped with two motors (for the left and right wheels)
15
and three pairs of proximity sensors; one pair is specialized in detecting the
other agent, and the other two in detecting the two distinct reward locations
('cooperate' or 'defect').
3. Results
In order to study the performance of the four agents (Original, Rational,
Predictive and Other's-Model), we focus on three aspects of the interaction:
Efficacy, Prediction Accuracy, and Stability.
Efficacy tells us how competent the models were in obtaining rewards on
average and in relation to each other. It's computed by calculating the mean
score per game for each model. Stability measures how predictable the behavior
was or, equivalently, whether the models converged to a common strategy or
alternated between non-deterministic states. In other words, stability quanti-
fies how predictable are the outcomes of the following rounds based on previous
results by using the information-theoretic measure of surprisal (also known as
self-information), which can be defined as the negative logarithm of the proba-
bility of an event [38]. By computing the average surprisal value of a model over
time, we can visually observe how much did it take for a model to converge to a
stable strategy and for how long it was able to maintain it. Finally, Prediction
Accuracy give us a better understanding of how well the predictive models were
able to actually predict their opponent's behavior. It's measured by the average
accumulated prediction error of each predictive model in each game.
3.1. Experiment 1: Against a deterministic-greedy agent
The results of this experiment show that on average all models were equally
effective since they obtained a similar amount of rewards (see Figure 7). The
Rational model only achieves a slightly better performance than the rest in the
Stag-hunt and the Harmony game, so this is a sign that the rest of the models
were performing almost optimally. Of course, if we pay attention to the overall
score obtained among the five games, we can observe a significant difference
16
Figure 7: Main results of the four behavioral models (Original, Rational, Predictive and
Other's) against a Greedy agent.
between the performance obtained by all models in the Harmony game and the
Hawk-dove when compared to the other three games. This salient difference
is not due to a malfunctioning of the predictive models, as we can see by the
results in Stability and Prediction Accuracy.
It is caused by the constraints
imposed by the Greedy opponent strategy in those games were T > R, since
are the ones in which it always select to 'defect'. We see that all models learn
to predict or to adapt their strategy to this opponent. In terms of surprisal, we
can observe how Rational Agents outperforms all the other models. This result
17
is not surprising given the simplicity of the opponent's strategy. Against these
simple predictable agents, we consider the Rational agent a reference point of
optimal performance, against which the other models can be compared. In this
line, a more significant result is that overall, both Predictive and Other's-Model
converge to a stable state faster than the original model, who lacks the ability to
predict the opponent's action. In terms of Prediction Accuracy, no significant
difference is observed.
3.2. Experiment 2: Against a deterministic-nice agent
Figure 8: Main results of the four behavioral models (Original, Rational, Predictive and
Other's) against a Nice agent. The legend in the center applies to the left and middle columns.
18
In this experiment, the overall efficacy of all models has increased signifi-
cantly compared to that of the previous experiment (see Figure 8). The ex-
planation was again that all models were able to successfully learn the optimal
policy against a simple deterministic strategy such as the one exhibited by the
Nice agent. But in this case, the best response strategy against a Nice agent
is the one that renders the highest possible benefit in all games, as opposed to
the best response strategy against the Greedy model, that in many occasions
was sub-optimal. In terms of surprisal, we see again how the Predictive and
the Other's-Model perform slightly better than the Original. This time the dif-
ference is smaller, but since predicting the opponent's behavior, in this case, is
almost trivial, we don't find this result unexpected. Again we can also observe
how the Original, the Predictive and the Other's-Model differ from optimal by
comparing their convergence rate with that of the Rational agent.
3.3. Experiment 3: Against a Tit-for-tat agent
The results of the models against the Tit-for-tat agent show a significant
amount of variability in terms of efficacy (see Figure 9). The most remarkable
result in this aspect is the superior efficacy of the non-predictive Original model
in the Battle of the Exes. This model is able to beat the more complex predictive
models in this case precisely because the predictive models are having problems
to accurately predict the TFT agent. This accumulated error in prediction, in
turn, drives their behavior toward a more unstable regime than the one achieved
by the Original model. Moreover, given that this model has a more reduced
state-space than the Predictive models and it only 'cares' about the previous
state of the game, it capitalizes better than the rest on the anti-coordination
structure of the Battle of the Exes.
In terms of suprisal, predictive models perform better than the Original in
the Prisoners Dilemma and in the Harmony game, but this result is reversed
in the Stag-Hunt and the Hawk-Dove and the Battle of the Exes, where the
Original reaches lower levels of surprisal. We can see how in the cases that the
game has one pure Nash equilibrium (the Prisoner's dilemma and the Harmony
19
Figure 9: Main results of the four behavioral models (Original, Rational, Predictive and
Other's) against a TFT agent.
game), the predictive models perform as well in terms of surprisal as against the
simpler deterministic agents studied before. However, in games with a wider
variety of equilibria such as the Stag-hunt or the Hawk-dove, predictive models
have overall more difficulties to fall into a stable state with the TFT opponent
Finally, the most salient feature regarding the resutls of Prediction accuracy
are the elevated accumulation of errors of the Rational model in the Stag-Hunt
and the Battle of the Exes, that almost reach 50%. These results may seem
counter-intuitive, but actually, they highlight one of the weak points of the
20
predictive models and the strength of such a simple strategy. As stated in the
model's description, the TFT agent selects its action based on the last action
performed by its opponent, so attempting to predict a policy of such agent can
become impossible for a predictive agent unless the dyad falls into an equilibrium
state very soon or if the predictive agent was able to integrate it's own action
into its prediction algorithm.
3.4. Experiment 4: Against the Original agent
Figure 10: Mean results of the four behavioral models (Original, Rational, Predictive and
Other's) against the Original agent.
In this experiment, we also observe a similar performance of the models in
21
terms of their efficacy to obtain rewards, with the exception of the Battle of the
Exes, where the Predictive an Other's-Model achieve the best results (see Figure
10. Remarkably, the Rational model is achieving a lower score than the rest.
This is due to the fact that this model has fallen in a pure-dominance equilibrium
with the Original model, where the best response for him is to perform the low-
rewarding action. Again, in this case, the Original model capitalizes in the
fact that is solely driven by its own self-utility maximization function, so it
got the initiative to go to the higher-rewarding action faster than the Rational
model. This, on the other hand, is predicting accurately and responding in the
best optimal way to that accurate prediction. But in this case, this behavioral
strategy makes it fall victim to the pure-dominance equilibrium enforced by its
opponent.
Regarding stability, the results show that overall the Predictive and the
Other's-Model converge faster than the Original towards an equilibrium.
In
all cases, we can observe that these two predictive models fall in between the
'optimal' level of convergence of the Rational model and the more unstable level
of the Original non-predictive model. The results of the Prediction accuracy
show the worse overall performance of the four experiments studied so far. This,
however, was an expected outcome since the predictive agents were facing a
learning non-deterministic agent which is objectively more difficult to predict.
3.5. Experiment 5: Continuous-Time Dyads
Figure 11: Results of a game played in the continuous/real-time version of the game compared
to the same game played in discrete-time.
As an additional control experiment, here we test whether specific instances
22
in experiment 3 above, where the Rational agent shows huge prediction errors
against the TFT agent, are purely a result of the discrete-time nature of ballistic
games, or similar errors against the TFT agent also repeat under continuous-
time interactions involving real-time feedback. As shown in Figure 11, this issue
of prediction errors is in fact solved under real-time conditions. This is because
the Rational agent is now able to change its chosen course of action in real-time
before the round ends. This significantly reduces the prediction error of the
rational agent against the TFT agent and enables the former to choose a better
strategy, which consequently, drives the dyad to a stable equilibrium.
4. Discussion
This work demonstrates a novel control-based cognitive approach to au-
tonomous multi-agent learning models, which combines adaptive feedback con-
trol with reinforcement learning. Based on a layered-control architecture, we
have designed and implemented agent models characterized by seven different
behavioral phenotypes. This includes both, deterministic agents as well as prob-
abilistic learning agents. The former includes cooperative, greedy and tit-for-tat
agents, whereas, the latter includes a reinforcement learning, rational, predic-
tive and other's-model agents. We tested these agent models in dyadic games
against each other across five different game theory settings (Harmony game,
Stag-Hunt, Hawk-Dove, Prisoner's Dilemma and Battle of the Exes). From
human behavioral experiments, it is known that the scope of these games suf-
ficiently encompasses behavioral phenotypes in human social decision-making
[23, 25]. Our proposed agent models were designed and implemented with a
view to simulating and understanding how cognitive agents learn and make de-
cisions when interacting with other agents in social scenarios (in the context of
the game-theoretic tasks stated above).
Additionally, we also implemented our agent models as embodied ePuck
robots. We tested these simulated robots within a spatial continuous version of
the game involving real-time feedback. We showed that under fully embodied
23
and situated conditions, limitations encountered by agents in ballistic scenarios
(such as lack of convergence to an optimal solution) can overcome. Our simu-
lated robots are Braitenberg Vehicles equipped with sensors and fully embodied
controllers (which are the proposed control-based learning modules). These
agents are situated in the sense that they only have partial observability of the
environment and other agents depending on their location and sensors during
rounds.
We found that pure reinforcement learning was inadequate in many tasks (as
shown in performance metrics). On the other hand, agent models that predict
actions or policies of opposing agents showed faster convergence to equilibria
and increased performance. Interestingly, the rational agent (which knows what
is the best action to take in each game and situation if the other's action is
foreseen) outperforms others in many games. Our results show that agents
using pure reinforcement-learning strategies are not optimal in most of the tested
games. Additionally, upon examining human data obtained in the Battle of the
Exes game by [23], we found (not shown in the results) that surprisal profiles
of human players matched closest to our predictive and other's-model agents,
rather than the rational or pure RL agents.
Of particular significance is the result that our predictive and other's-model
agents correctly learn to predict actions and policies respectively of the opposing
agent across all games. This shows how autonomous, embodied and situated
multi-agent systems can be equipped with simple theory of mind capabilities (in
the context of the specified tasks). For future work, it might be interesting to
equip our agents with meta-learning models, so that they may be able to choose
which behavioral phenotype to enact in a given social scenario and flexibly
switch in changing environments.
Even though the different behavioral strategies modeled in this study consist
some of the simplest building blocks of human behavior, it is by no means a
complete list. Also, typical human behavior is not rigidly tied to a given model,
but rather can flexibly shift between behavioral strategies. Model flexibility is
definitely an important consideration for a ToM. Another issue that is relevant
24
for many real-world tasks is the cost or penalty of being engaged too often in less
desirable outcomes such as ties or low rewards. In a sense, this could serve as
an incentive to improve performance in complex tasks and should be taken into
account in a ToM. Yet another typical human feature is trust. Humans tend to
form social conventions by agreeing on simple rules or deals and forming a trust
relation. At the moment, our models currently lack this ability to form teams
that agree on collective goals and work cooperatively to achieve success.
Finally, our control-based reinforcement learning architecture provides an
interesting possibility to implement complex social behaviors on multi-agent
robotic platforms, such as humanoid robots [39] or other socially intelligent ar-
tificial systems. Another scientific domain that may benefit from the type of
agent models described in this paper, is evolutionary dynamics; in particular,
the evolution of cognitive and intelligent agents. Social interactions involving
cooperation and competition are known to play a key role in many evolution-
ary accounts of biological life and consciousness [40], [41]. Lastly, autonomous
multi-agent models, of the type described above, with ToM capabilities, provide
useful tools for modeling the dynamics of global socio-political and cultural
phenomena.
Acknowledgments
The research leading to these results has received funding from the European
Commissions Horizon 2020 socSMC project (socSMC-641321H2020) and by the
European Research Councils CDAC project (ERC-2013-ADG341196).
Additional information
The authors declare that they have no competing interests.
25
References
References
[1] D. Premack, G. Woodruff, Does the chimpanzee have a theory of mind?,
Behavioral and brain sciences 1 (4) (1978) 515 -- 526.
[2] S. Baron-Cohen, A. M. Leslie, U. Frith, Does the autistic child have a
theory of mind?, Cognition 21 (1) (1985) 37 -- 46.
[3] D. Premack, The infant's theory of self-propelled objects, Cognition 36 (1)
(1990) 1 -- 16.
[4] M. Lanctot, V. Zambaldi, A. Gruslys, A. Lazaridou, K. Tuyls, J. P´erolat,
D. Silver, T. Graepel, A unified game-theoretic approach to multiagent re-
inforcement learning, in: Advances in Neural Information Processing Sys-
tems, 2017, pp. 4190 -- 4203.
[5] A. Lerer, A. Peysakhovich, Learning social conventions in markov games,
arXiv preprint arXiv:1806.10071.
[6] N. C. Rabinowitz, F. Perbet, H. F. Song, C. Zhang, S. Eslami, M. Botvinick,
Machine theory of mind, arXiv preprint arXiv:1802.07740.
[7] J. Schmidhuber, Deep learning in neural networks: An overview, Neural
networks 61 (2015) 85 -- 117.
[8] W. Yoshida, R. J. Dolan, K. J. Friston, Game theory of mind, PLoS com-
putational biology 4 (12) (2008) e1000254.
[9] C. Baker, R. Saxe, J. Tenenbaum, Bayesian theory of mind: Modeling
joint belief-desire attribution, in: Proceedings of the annual meeting of the
cognitive science society, Vol. 33, 2011.
[10] C. L. Baker, J. Jara-Ettinger, R. Saxe, J. B. Tenenbaum, Rational quan-
titative attribution of beliefs, desires and percepts in human mentalizing,
Nature Human Behaviour 1 (4) (2017) 0064.
26
[11] J. Tenenbaum, Building machines that learn and think like people, in:
Proceedings of the 17th International Conference on Autonomous Agents
and MultiAgent Systems, International Foundation for Autonomous Agents
and Multiagent Systems, 2018, pp. 5 -- 5.
[12] S. V. Albrecht, P. Stone, Autonomous agents modelling other agents: A
comprehensive survey and open problems, Artificial Intelligence 258 (2018)
66 -- 95.
[13] I. T. Freire, C. Moulin-Frier, M. Sanchez-Fibla, X. D. Arsiwalla, P. Ver-
schure, Modeling the formation of social conventions in multi-agent popu-
lations, arXiv preprint arXiv:1802.06108.
[14] I. T. Freire, J.-Y. Puigb`o, X. D. Arsiwalla, P. F. Verschure, Limits of multi-
agent predictive models in the formation of social conventions, Artificial
Intelligence Research and Development: Current Challenges, New Trends
and Applications 308 (2018) 297.
[15] M. Kleiman-Weiner, M. K. Ho, J. L. Austerweil, M. L. Littman, J. B.
Tenenbaum, Coordinate to cooperate or compete: abstract goals and joint
intentions in social interaction, in: CogSci, 2016.
[16] J. Perolat, J. Z. Leibo, V. Zambaldi, C. Beattie, K. Tuyls, T. Graepel, A
multi-agent reinforcement learning model of common-pool resource appro-
priation, in: Advances in Neural Information Processing Systems, 2017,
pp. 3643 -- 3652.
[17] A. Peysakhovich, A. Lerer, Prosocial learning agents solve generalized stag
hunts better than selfish ones, in: Proceedings of the 17th International
Conference on Autonomous Agents and MultiAgent Systems, International
Foundation for Autonomous Agents and Multiagent Systems, 2018, pp.
2043 -- 2044.
[18] I. T. Freire, J.-Y. Puigb`o, X. D. Arsiwalla, P. F. Verschure, Modeling the
27
opponents action using control-based reinforcement learning, in: Confer-
ence on Biomimetic and Biohybrid Systems, Springer, 2018, pp. 179 -- 186.
[19] M. J. Gaparrini, M. S´anchez-Fibla, Loss aversion fosters coordination in in-
dependent reinforcement learners, Artificial Intelligence Research and De-
velopment: Current Challenges, New Trends and Applications 308 (2018)
307.
[20] J. Z. Leibo, V. Zambaldi, M. Lanctot, J. Marecki, T. Graepel, Multi-agent
reinforcement learning in sequential social dilemmas, in: Proceedings of
the 16th Conference on Autonomous Agents and MultiAgent Systems, In-
ternational Foundation for Autonomous Agents and Multiagent Systems,
2017, pp. 464 -- 473.
[21] A. Peysakhovich, A. Lerer, Consequentialist conditional cooperation in so-
cial dilemmas with imperfect information, arXiv preprint arXiv:1710.06975.
[22] J. F. Nash, et al., Equilibrium points in n-person games, Proceedings of
the national academy of sciences 36 (1) (1950) 48 -- 49.
[23] R. X. Hawkins, R. L. Goldstone, The formation of social conventions in
real-time environments, PloS one 11 (3) (2016) e0151670.
[24] R. X. Hawkins, N. D. Goodman, R. L. Goldstone, The emergence of social
norms and conventions, Trends in cognitive sciences.
[25] J. Poncela-Casasnovas, M. Guti´errez-Roig, C. Gracia-L´azaro, J. Vicens,
J. G´omez-Gardenes, J. Perell´o, Y. Moreno, J. Duch, A. S´anchez, Humans
display a reduced set of consistent behavioral phenotypes in dyadic games,
Science advances 2 (8) (2016) e1600451.
[26] A. G. Sanfey, Social decision-making: insights from game theory and neu-
roscience, Science 318 (5850) (2007) 598 -- 602.
[27] P. F. Verschure, T. Voegtlin, R. J. Douglas, Environmentally mediated syn-
ergy between perception and behaviour in mobile robots, Nature 425 (6958)
(2003) 620.
28
[28] C. Moulin-Frier, X. D. Arsiwalla, J.-Y. Puigb`o, M. Sanchez-Fibla, A. Duff,
P. F. Verschure, Top-down and bottom-up interactions between low-
level reactive control and symbolic rule learning in embodied agents., in:
CoCo@NIPS, 2016.
[29] V. Braitenberg, Vehicles: Experiments in synthetic psychology, MIT press,
1986.
[30] M. Corbetta, G. L. Shulman, Control of goal-directed and stimulus-driven
attention in the brain, Nature reviews neuroscience 3 (3) (2002) 201.
[31] E. Koechlin, C. Ody, F. Kouneiher, The architecture of cognitive control
in the human prefrontal cortex, Science 302 (5648) (2003) 1181 -- 1185.
[32] Y. Munakata, S. A. Herd, C. H. Chatham, B. E. Depue, M. T. Banich, R. C.
OReilly, A unified framework for inhibitory control, Trends in cognitive
sciences 15 (10) (2011) 453 -- 459.
[33] H. E. Den Ouden, P. Kok, F. P. De Lange, How prediction errors shape
perception, attention, and motivation, Frontiers in psychology 3 (2012) 548.
[34] C. Wacongne, E. Labyt, V. van Wassenhove, T. Bekinschtein, L. Naccache,
S. Dehaene, Evidence for a hierarchy of predictions and prediction errors in
human cortex, Proceedings of the National Academy of Sciences 108 (51)
(2011) 20754 -- 20759.
[35] R. S. Sutton, Learning to predict by the methods of temporal differences,
Machine learning 3 (1) (1988) 9 -- 44.
[36] R. Axelrod, W. D. Hamilton, The evolution of cooperation, science
211 (4489) (1981) 1390 -- 1396.
[37] R. Axelrod, Effective choice in the prisoner's dilemma, Journal of conflict
resolution 24 (1) (1980) 3 -- 25.
[38] C. E. Shannon, A mathematical theory of communication, Bell system
technical journal 27 (3) (1948) 379 -- 423.
29
[39] C. Moulin-Frier, J.-Y. Puigbo, X. D. Arsiwalla, M. Sanchez-Fibla, P. Ver-
schure, Embodied artificial intelligence through distributed adaptive con-
trol: An integrated framework, ArXiv e-prints.
[40] X. D. Arsiwalla, I. Herreros, C. Moulin-Frier, M. S´anchez-Fibla, P. F. Ver-
schure, Is consciousness a control process?, in: CCIA, 2016, pp. 233 -- 238.
[41] X. D. Arsiwalla, R. Sole, C. Moulin-Frier, I. Herreros, M. Sanchez-
Fibla, P. Verschure, The morphospace of consciousness, arXiv preprint
arXiv:1705.11190.
30
|
0906.2756 | 6 | 0906 | 2010-11-06T21:47:00 | Norms and Commitment for iOrgs(TM) Information Systems: Direct Logic(TM) and Participatory Grounding Checking | [
"cs.MA",
"cs.LO",
"cs.SE"
] | The fundamental assumption of the Event Calculus is overly simplistic when it comes to organizations in which time-varying properties have to be actively maintained and managed in order to continue to hold and termination by another action is not required for a property to no longer hold. I.e., if active measures are not taken then things will go haywire by default. Similarly extension and revision is required for Grounding Checking properties of systems based on a set of ground inferences. Previously Model Checking as been performed using the model of nondeterministic automata based on states determined by time-points. These nondeterministic automata are not suitable for iOrgs, which are highly structured and operate asynchronously with only loosely bounded nondeterminism. iOrgs Information Systems have been developed as a technology in which organizations have people that are tightly integrated with information technology that enables them to function organizationally. iOrgs formalize existing practices to provide a framework for addressing issues of authority, accountability, scalability, and robustness using methods that are analogous to human organizations. In general
-iOrgs are a natural extension Web Services, which are the standard for distributed computing and software application interoperability in large-scale Organizational Computing.
-iOrgs are structured by Organizational Commitment that is a special case of Physical Commitment that is defined to be information pledged.
iOrgs norms are used to illustrate the following:
-Even a very simple microtheory for normative reasoning can engender inconsistency In practice, it is impossible to verify the consistency of a theory for a practical domain. -Improved Safety in Reasoning. It is not safe to use classical logic and probability theory in practical reasoning. | cs.MA | cs | Published in ArXiv 0906.2756
Norms and Commitment for iOrgsTM Information Systems:
Direct LogicTM
and
Participatory Grounding Checking
Carl Hewitt
http://carlhewitt.info
iOrgsTM Information Systems raise important issues for formalizing norms that require extensions and
revisions of previous foundational work. For example, extension and revision is required of the fundamental
assumption of the Event Calculus:
Time-varying properties hold at particular time-points if they have been initiated by an action at
some earlier time-point, and not terminated by another action in the meantime.
The fundamental assumption of the Event Calculus is overly simplistic when it comes to organizations in
which time-varying properties have to be actively maintained and managed in order to continue to hold and
termination by another action is not required for a property to no longer hold. I.e., if active measures are not
taken then things will go haywire by default.
Similarly extension and revision is required for Grounding Checking properties of systems based on a set
of ground inferences. Previously Model Checking has been performed using the model of nondeterministic
automata based on states determined by time -points. These nondeterministic automata are not suitable for
iOrgs, which are highly structured and operate asynchronously with only loosely bounded nondeterminism.
iOrgs Information Systems have been developed as a technology in which organizations have people that
are tightly integrated with information technology that enables them to function organizationally . iOrgs
formalize existing practices to provide a framework fo r addressing issues of authority, accountability,
scalability, and robustness using methods that are analogous to human organizations. In general
iOrgs are a natural extension Web Services, which are the standard for distributed computing and
software application interoperability in large-scale Organizational Computing.
iOrgs are structured by Organizational Commitment that is a special case of Physica l Commitment
that is defined to be information pledged.
iOrgs norms are used to illustrate the following:
Even a very simple microtheory for normative reasoning can engender inconsistency In practice, it
is impossible to verify the consistency of a theory for a practical domain.
Improved Safety in Reasoning. It is not safe to use classical logic and probability theory in
practical reasoning.
November 6, 2010 Page 1 of 18
Contents
iOrgsTM Information Systems ....................................................................... 3
Participatory Semantics ................................................................................ 3
Contrast between Participatory Semantics and the Event Calculus ............. 4
Commitment ................................................................................................. 6
Participatory Grounding Checking ............................................................... 7
Direct Inference ............................................................................................ 9
Collusion at Santa Cruz FishMarket ........................................................... 10
Inconsistent policy in War .......................................................................... 10
Paradigm shift from Inconsistency Denial to Rapid Recovery .................. 11
Norms at Santa Cruz FishMarket ............................................................... 11
Conclusion .................................................................................................. 12
Acknowledgments ...................................................................................... 12
Bibliography ............................................................................................... 12
Appendix 1. A simple auction procedure in ActorScript ........................... 15
End Notes ................................................................................................... 16
November 6, 2010 Page 2 of 18
iOrgsTM Information Systems
In the organization lies the power.
iOrgs Information Systems have been developed as a technology in which organizations have people
that are tightly integrated with information technology that enables them to function organizationally
[Hewitt and Inman 1991; Hewitt 2008b, 2008d, 2008g].i iOrgs formalize existing practices to provide
a framework for addressing issues of authority, accountability, scalability, and robustness using
methods that are analogous to human organizations.ii. In general
iOrgs are a natural extension Web Services, which are the standard for distributed computing
and software application interoperability in large-scale Organizational Computing.
iOrgs are structured by Organizational Commitment that is a special case of Physical
Commitment [Hewitt 2007 2008b] that is defined to be information pledged.iii
This paper discusses how iOrgs require Direct Logic in reasoning and participatory grounding
checking in systems analysis:
1. The development of iOrgs and the extreme dependence of our society on these systems have
introduced new phenomena. These systems have pervasive inconsistencies among and
within the following:
o Norms that express how systems can be used and tested in practice
o Policies that express over-arching justification for systems and their technologies
o Practices that express implementations of systems
Different parties are responsible for constructing, evolving, justifying and maintaining
practices, norms, and operations for large-scale Organizational Computing. In specific cases
any one consideration can trump the others. Sometimes debates over inconsistencies can
become quite heated, e.g., between sales, engineering and finance.
2. Grounding checkingiv is a fundamental tool in the analysis of iOrgs. However, previous
work on model checking has been performed using the model of nondeterministic automata
based on states determined by time-points. These nondeterministic automata are not suitable
for iOrgs, which are highly structured and operate asynchronously with only loosely
bounded nondeterminism. Instead analysis based on regions of space-time (as in
Participatory Semantics [Hewitt and Manning 1996]) is required.
Participatory Semantics
Participatory Semantics [Hewitt and Manning 1996] is based on regions of space-time called
Participationsv according to the following legend:vi
November 6, 2010 Page 3 of 18
XMLvii
(a message or data structure)
A participation
A reference
A message transmission
arrived at the specified time
An object transport
XML
Participation
Source
Target
Origin
Message@Time
Destination
Origin
Object
Destination
Figure 1. Legend
Contrast between Participatory Semantics and the Event Calculus
Participatory Semantics is based on 4-dimensional regions of space-time whereas the Event
Calculus is based on events (which take 0 time) on a global universal time-line.
This paper does not use the usual Event Calculus formalism [Kowalski and Sergot 1986 , Miller and
Shanahan 1999]. A principle reason for not adopting the Event Calculus is avoidance of its
fundamental assumption:
“Time-varying properties hold at particular time-points if they have been initiated by an action at
some earlier time-point, and not terminated by another action in the meantime.”
The fundamental assumption of the Event Calculus is overly simplistic when it comes to organizations
in which time-varying properties have to be actively managed in order to continue to hold and
termination by another action is not required for a property to no longer hold. I.e., if active measures
are not taken then things will go haywire by default. For example consider the following property:
“Drive safely” It might be said that the property was “terminated” when a driver collides with another
vehicle. However, it is often the case that some “unsafe driving” occurred before the collision!
November 6, 2010 Page 4 of 18
Another problem with the Event Calculus is that it is formulated at the very low level of abstraction of
time-points. As Carlo Rovelli has explained:viii
“We never really see time. We see only clocks. If you say this object moves, what you really mean
is that this object is here when the hand of your clock is here, and so on. We say we measure time
with clocks, but we see only the hands of the clocks, not time itself. And the hands of a clock are a
physical variable like any other. So in a sense we cheat because what we really observe are
physical variables as a function of other physical variables, but we represent that as if everything
is evolving in time.”
Preoccupation with global time-points is a serious problem with the Event Calculus. This problem is
closely related to another problem with the Event Calculus: A time is not bound to a locale but is
instead imagined to be free floating!
Consider an example in which safe driving is followed by unsafe driving leading to a
collision. An important issue is that there may be no event which clearly delineates the
transition from safe driving to unsafe driving.ix The lack of such an event is not material to
Participatory Semantics. However, there is no clear terminating event for the Event
Calculus.
Driving
Safe Driving
Unsafe Driving
Collision
Figure 2. Transition from Safe to Unsafe Driving (with no clipping event)
Next consider an example in which AM transitions to PM on July 31, 2008 in California.
Here the issue is that there is no physical event that occurs throughout California that
marks the transition from AM to PM.
California on July 31, 2008
AM
PM
Figure 3. Transition from AM to PM (with no clipping event)
November 6, 2010 Page 5 of 18
By convention, the AM and PM regions for California are adjacent to each other. However,
this adjacency does not require the existence of any event that occurs throughout California
and the lack of such an event is not material to Participatory Semantics. However, there is
no terminating event as required by the Event Calculus.
Commitment
According to [Hewitt 2007], a Physical Commitment PC is defined to be a pledge that certain
information I holds for a physical system PS for a space-time region R. Note that physical
commitment is defined for whole physical systems; not just a participant or process.
This paper uses a mythical Santa Cruz FishMarketx to illustrate how organizational commitments can
be formulated at a higher level of abstraction. The Santa Cruz FishMarket uses an electronic English
Auction starting with a reserved price in which a certain time is allowed for more bids to come in
before the bidding is closed. As each higher bid is received, the new minimum bid is announced to the
participants. Tie bids are broken by choosing the one which arrived first as the winner. Consequently
the Santa Cruz FishMarket is an organizational commitment with an auction of buyers and sellers.
An implementationxi for a SimpleAuction for the Santa Cruz FishMarket is given in the appendix.xii
The commitments below are made by the implementation pledging the following information:
Bidder1
Request
Bid.[amountBid bidder : Bidder1
]@arrivalTime
Exception
TooLate.[deadline]
SimpleAuction
arrivalTime > deadline
Figure 4. Commitment: Bidding too late causes an exception
November 6, 2010 Page 6 of 18
Request
Alarm.[alarmTime]
SimpleAuction
alarmTime > deadline
ProcessOutcome
CurrentBidding
Figure 5. Commitment: An alarm triggers the processing of the auction outcome
Participatory Grounding Checking
The denotational semantics of concurrency were first developed in [Clinger 1981]. Subsequently
[Hewitt 2006b] developed the TimedDiagrams model with the Computational Representation
Theorem which states:
The denotation DenoteS of a closedxiii system S represents all the possible behaviors of S as
DenoteS = ⊔iω ProgressionS
i(⊥S)
where ProgressionS is an approximation function that takes a set of approximate
behaviors to their next stage and ⊥S is the initial behavior of S.
The denotational semantics exhibits relatively unbounded nondeterminism because in the delivery
of a message can occur a relatively unbounded amount of time after it is sent. This relatively
unbounded nondeterminism can cause trouble with traditional global state-based approaches
[Bianculli, Morzenti, Pradella, San Pietro, Spoletini 2007; Bordini, Fisher, Visser, and Wooldridge
2004; Cliffe, De Vos, and Padget 2006; Desai, Cheng, Chopra, and Singh 2007; Venkatraman and
Singh 1999, etc.] because of the following issues:
State explosion because of the increase in possible interleavings of global states
Modeling failure because the system being modeled is not finite state
Participatory grounding checking makes use of the Representation Theorem to characterize possible
alternative computations. In participatory grounding checking:
Explosion is less of a problem because local groundings are modeled instead of global state.
Also Participatory Semantics can be used to abstract high level properties of denotations in a
way that is similar to how abstraction has been used in global state model checking.
November 6, 2010 Page 7 of 18
Systems are not modeled as nondeterministic state machines, Petri nets, or process calculi
[Aceto and Gordon 2005].xiv
Communication is modeled as being fundamentally one-way and asynchronous. In this way
modeling problems such as occur using Petri Nets and synchronous process calculi are
avoided [Padget and Bradford 1998].
For example consider the system with SimpleAuction (defined in the appendix of this article)
augmented with bidders like the following:xv
SimpleBidder
behavior { xvi serialize the messages received by this bidder
::theAuction auction that bidder is bidding for
Auction
Currency
::maximumBid maximum that bidder will bid for this auction
-----------------------------------------------------------------------------------------------------
implements Bidder the Bidder interface is implemented below
self.newMinimum (amount) →
a newMinimum message with amount has been received
amount ?? {
(< maximumBid) : if amount is less than maximumBid
theAuction.bid(amount for: self) then bid the minimum
catcher { and if it throws the exception
TooLittle.[minimumBid] : that the bid was too small
relay self.newMinimum(minimumBid)}
send yourself the new minimum bid
not (< maximumBid) : void }} else do nothing and return void
Using bidders like the above, execution scenarios of the system can be computed using the
Representation Theorem. These execution scenarios can be checked against norms such as the
following:
Commitment: At Santa Cruz FishMarket, deliveries from seller to buyer are paid at the
agreed price; i.e. Santa Cruz FishMarket pledges the following information: For every
seller, buyer, and delivery,
Delivers[seller, buyer, delivery] ├ FishMarket PaysAgreedPrice[buyer, seller, delivery]xvii
Note that seller, buyer, and delivery are all space-time participations in the above norm.
Consequently, there is enough information to specify that the buyer pays the agreed price to the seller
on delivery of the purchase.xviii
November 6, 2010 Page 8 of 18
Seller1
Goods
Delivery1
Payment
Buyer1
Figure 6. A commitment involving the exchange of Goods for Payment
Direct Inference
Direct inference is used in to directly infer conclusions from premises. For example, suppose that
we have
1) WeekdayAt5PM ├Boston TrafficJam
which says that in Boston
, a weekday at 5PM infers a traffic jam.
2) ├BostonTrafficJam
which says that in Boston
, no traffic jam.
In classical logic, WeekdayAt5PM is inferred in Boston from 1) and 2) above. But
fortunately in Direct Logic:
⊬BostonWeekdayAt5PM
which says in Boston,
there is a particular proposition (WeekdayAt5PM) that cannot be
inferred in Direct Logic from 1) and 2) above.
Consequently, direct inference comes into play even in the absence of overt inconsistency.xix
November 6, 2010 Page 9 of 18
Collusion at Santa Cruz FishMarket
Ethical conduct at the FishMarket is worthy of further study. A classic guide to norms for ethical
conduct appears in The Prince [Machiavelli 1532]. So roughly in this spirit, consider the follow
axiomatizationxx of collusion at FishMarket:xxi
1) p, action: CanResult[Do[p, action], Rich[p]]├ThePrince Do[p, action] axiomxxii
2) ├ThePrince CanResult[Do[Machiavelli, Collude], Rich[Machiavelli]] axiom
2’) Do[Machiavelli, Collude] ├ThePrince CanResult[Rich[Machiavelli]]] axiom
3) ├ ThePrince Do[Machiavelli, Collude] from 1) and 2)
4) p, action: CanResult[Do[p, action], Ruined[p]]├ ThePrinceDo[p, action] axiomxxiii
5) ├ ThePrince CanResult[Do[Machiavelli, Collude], Ruined[Machiavelli]] axiom
5’) Do[Machiavelli, Collude] ├ThePrince CanResult[Ruined[Machiavelli]]] axiom
6) ├ ThePrince Do[Machiavelli, Collude] from 4), and 5)
Note that, in the ThePrince, there is an inconsistency between 3) and 6). However, even though the
ThePrince is inconsistent, it is not meaningless because in some respects it captures some
underlying paradoxical aspects of Machiavelli‟s theory:
“The wish to acquire more is admittedly a very natural and common thing; and when men
succeed in this they are always praised rather than condemned. But when they lack the ability to
do so and yet want to acquire more at all costs, they deserve condemnation for their mistakes.”
Inconsistent policy in War
A classic case of inconsistency occurs in the novel Catch-22 [Heller 1961] which states that a
person “would be crazy to fly more missions and sane if he didn't, but if he was sane he had to fly
them. If he flew them he was crazy and didn't have to; but if he didn't want to he was sane and had to.
Yossarian was moved very deeply by the absolute simplicity of this clause of Catch-22 and let out a
respectful whistle. „That's some catch, that Catch-22,‟ he observed.”
In the spirit of Catch-22, consider the follow axiomatization of the above:
1. ├Catch-22 ℙ(Able[x, Fly] Fly[x]) ≤ ℙ(Sane[x]) axiom
2. ├Catch-22 ℙ(Sane[x]) ≤ ℙ(Obligated[x, Fly]) axiom
3. ├Catch-22 ℙ(Sane[x] Obligated[x, Fly]) ≤ ℙ(Fly[x]) axiom
4. ├Catch-22 ℙ(Fly[x]) ≤ ℙ(Crazy[x]) axiom
5. ├Catch-22 ℙ(Crazy[x]) ≤ ℙ(Obligated[x, Fly]) axiom
6. ├Catch-22 ℙ(Sane[p] Obligated[p, Fly]) ≤ ℙ(Fly[p]) axiom
November 6, 2010 Page 10 of 18
For Yossarian, we have the following axioms:
7. ├Catch-22 ℙ(Able[Yossarian, Fly]) ≅ 1 axiom
8. ├Catch-22 ℙ(Sane[Yossarian]) ≅ 1 axiom
Consequently,
2’. ├Catch-221 ≅ ℙ(Obligated[Yossarian, Fly]) Yossarian using 2 and 8
3’. ├Catch-22 1 ≅ ℙ(Fly[Yossarian]) Yossarian using 3, 2’ and 8
4’. ├Catch-22 1 ≅ ℙ(Crazy[Yossarian]) Yossarian using 4 and 3’
5’. ├Catch-22 1 ≲ 1 - ℙ(Obligated[Yossarian, Fly]) Yossarian using 5 and 4’
5’’. ├Catch-22 ℙ(Obligated[Yossarian, Fly]) ≅ 0 reformulation of 5’
6’. ├Catch-22 1 ≲ 1 - ℙ(Fly[Yossarian]) Yossarian using 6’, 8 and 5’’
6’’. ├Catch-22 ℙ(Fly[Yossarian]) ≅ 0 reformulation of 6’
Thus there is an inconsistency in Catch-22 in that both of the following hold:
3’. ├Catch-22 1 ≅ ℙ(Fly[Yossarian])
6’’. ├Catch-22 ℙ(Fly[Yossarian]) ≅ 0
Paradigm shift from Inconsistency Denial to Rapid Recovery
ThePrince and Catch-22 illustrate the following important points:
Even a very simple microtheory for normative reasoning can engender inconsistency In
practice, it is impossible to verify the consistency of a theory for a practical domain.
Improved Safety in Reasoning. It is not safe to use classical logic and probability theory in
practical reasoning.
Norms at Santa Cruz FishMarket
Norms are commitment supported by communities of practice.
Norm: At Santa Cruz FishMarket, there is no collusion among buyers and sellers.
Formalizing the norm above is the subject of future research.
November 6, 2010 Page 11 of 18
Conclusion
iOrgs raise important issues for inconsistency robustness and participatory grounding checking. This
paper presents some ideas for formalizing these issues. Relationships among these issues are analyzed
using illustrations from FishMarket and The Prince. The following conclusions are proposed.
Extension and revision is required of the fundamental assumption of the Event Calculus:
Time-varying properties hold at particular time-points if they have been initiated by an
action at some earlier time-point, and not terminated by another action in the meantime. The
fundamental assumption of the Event Calculus is overly simplistic when it comes to iOrgs in
which time-varying properties have to be actively maintained and managed in order to
continue to hold and termination by another action is not required for a property to no longer
hold. I.e., if active measures are not taken then things will go haywire by default.
Consequently the Event Calculus approach must evolve into a strongly paraconsistent system
structured around participations in space-time.
Similarly extension and revision is required for Model Checking properties of systems.
Previously Model Checking has been performed using the model of nondeterministic
automata based on states determined by time-points. These nondeterministic automata are not
suitable for iOrgs, which are highly structured and operate asynchronously with only loosely
bounded nondeterminism. Consequently Model Checking needs to evolve in the direction of
verifying participatory behavior in iOrgs.
Acknowledgments
The development of Participatory Semantics was joint work with Carl Manning. Les Gasser, Mike
Huhns, Victor Lessor, Pablo Noriega, Sascha Ossowski, Jaime Sichman, Munindar Singh, etc.
provided valuable suggestions at AAMAS‟07. The reviewers and participants of MALLOW‟07
(including John Lloyd, John-Jules Meyer, Pablo Noriega, Jaime Sichman, Munindar Singh, Rineke
Verbrugge, etc.) provided valuable comments. Afterwards Munindar Singh provided helpful pointers
to the literature. The reviewers for COIN@AAMAS‟08 made helpful suggestions. Conversations with
Jeremy Forth were helpful in developing the comparison of Participatory Semantics and the Event
Calculus. Munindar Singh made extensive comments and suggestions that significantly improved the
presentation. A helpful review was provided by an anonymous referee for the special issue of the
Journal of IGPL on Normative Multi-agent Systems.
Bibliography
Chris Anderson. The End of Theory: The Data Deluge Makes the Scientific Method Obsolete
Wired. June 23, 2009.
Luca Aceto and Andrew D. Gordon (Ed.) Algebraic Process Calculi: The First Twenty Five
Years and Beyond 2005
Marco Alberti, Marco Gavanelli, Evelina Lamma, Paola Mello, and Paolo Torroni “Modeling
interactions using social integrity constraints: A resource sharing case study” DALT. LNAI 2990.
Springer-Verlag, 2004.
Aldo Antonelli. “Non-monotonic Logic” Stanford Encyclopedia of Philosophy. March 2006.
November 6, 2010 Page 12 of 18
Domenico Bianculli, Angelo Morzenti, Matteo Pradella, Pierluigi San Pietro, Paola Spoletini
“Trio2Promela: A Model Checker for Temporal Metric Specifications” ICSE'07.
Brian Blum. Contracts: Examples and Explanations Aspen Publishers. 3rd edition 2004.
Rafael Bordini, Michael Fisher, Willem Visser, and Michael Wooldridge “Verifiable Multi-Agent
Programs” Springer LNAI 3067. 2004.
Geof Bowker, Susan L. Star, W. Turner, and Les Gasser. (Eds.) Social Science Research,
Technical Systems and Cooperative Work. Lawrence Earlbaum. 1997.
Cristiano Castelfranchi “Commitments: From individual intentions to groups and organizations”
AAAI-93 Workshop on AI and Theories of Groups and Organizations: Conceptual and Empirical
Research.
Amit Chopra and Munindar Singh. Multiagent Commitment Alignment. Draft of March 7, 2009.
Owen Cliffe, Marina De Vos, and Julian Padget (2006). “Specifying and Reasoning about
Multiple Institutions” COIN@AAMAS‟06.
Jiangbo Dang, Jingshan Huang and Mike Huhns “Workflow Coordination for Service-Oriented
Multi-agent Systems” AAMAS‟07.
Nirmit Desai, Zhengang Cheng, Amit Chopra, and Munindar Singh “Toward Verification of
Commitment Protocols and their Compositions” AAMAS‟07.
Virginia Dignum A Model for Organizational Interaction. PhD thesis. Utrecht. 2004..
Rogier Van Eijk, Frank De Boer, Wiebe Van Der Hoek and John-Jules Meyer. “A verification
framework for agent communication” AAMAS‟03.
Andrew Farrell, Marek Sergot, Mathias Salle and Claudio Bartolini “Using the event calculus for
tracking the normative state of contracts” International Journal of Cooperative Information
System June-September 2005.
Tim Folger. “Newsflash: Time May Not Exist” Discover. June 12, 2007.
Dorian Gaertner, Pablo Noriega, J.-A. Rodriguez-Aguilar, Wamberto Vasconcelos (2007).
“Distributed Norm Management in Regulated Multi-Agent Systems” AAMAS‟07.
Joseph Heller. Catch-22 Simon and Schuster. 1961.
Carl Hewitt and Jeff Inman (1991). “DAI Betwixt and Between: From „Intelligent Agents‟ to
Open Systems Science” IEEE Transactions on Systems, Man, and Cybernetics. Nov. /Dec. 1991.
Carl Hewitt and Carl Manning (1996). Synthetic Infrastructures for Multi-Agency Systems
ICMAS'96.
Carl Hewitt What is Commitment? Physical, Organizational, and Social (Revised) Pablo Noriega
.et. al. editors. LNAI 4386. Springer-Verlag. 2007.
Carl Hewitt (2008a) Large-scale Organizational Computing requires Unstratified Reflection and
Strong Paraconsistency Coordination, Organizations, Institutions, and Norms in Agent Systems
III. Jaime Sichman, Pablo Noriega, Julian Padget and Sascha Ossowski (Ed.). Springer-Verlag.
2008. http://organizational.carlhewitt.info/
Carl Hewitt (2008b). ORGs for Scalable, Robust, Privacy-Friendly Client Cloud Computing
IEEE Internet Computing September/October 2008.
Carl Hewitt (2008c) Middle History of Logic Programming: Resolution, Planner, Edinburgh Logic for
Computable Functions, Prolog and the Japanese Fifth Generation Project ArXiv 0904.3036
Carl Hewitt (2008d) A historical perspective on developing foundations for client cloud
computing: iConsultTM & iEntertainTM Apps using iInfoTM Information Integration for iOrgsTM
Information Systems (Revised version of “Development of Logic Programming: What went
wrong, What was done about it, and What it might mean for the future” AAAI Workshop on What
Went Wrong. AAAI-08.) ArXiv 0901.4934.
Carl Hewitt (2008e) Common sense for concurrency and inconsistency tolerance using Direct
LogicTM and the Actor Model ArXiv 0812.4852.
November 6, 2010 Page 13 of 18
Carl Hewitt (2008f) Scalable Privacy-Friendly Client Cloud Computing: a gathering Perfect
Disruption Stanford Computer Systems Laboratory Colloquium Oct. 22, 2008. Video recording.
http://stanford-online.stanford.edu/courses/ee380/081022-ee380-300.asx
Carl Hewitt (2008g) Perfect Disruption: Causing the Paradigm Shift from Mental Agents to
ORGs IEEE Internet Computing. Jan/Feb 2009.
Carl Hewitt ActorScript™: Industrial strength integration of local and nonlocal concurrency for
Client-cloud Computing ArXiv. 0907.3330
Bryan Horling and Victor Lesser (2005). “Using ODML to Model Multi-Agent Organizations”
IAT‟05
Nicholas Jennings “Commitments and conventions: The foundation of coordination in multi-
agent systems” Knowledge Engineering Review. 3. 1993.
Frederick Knabe “A Distributed Protocol for Channel-Based Communication with Choice”
PARLE’92
Bill Kornfeld and Carl Hewitt “The Scientific Community Metaphor” IEEE Transactions on
Systems, Man, and Cybernetics. January 1981.
Robert Kowalski and Marek Sergot. “A Logic-based Calculus of Events” New Generation
Computing Volume 4 , Issue 1. 1986.
Robert Kowalski. “Database updates in the Event Calculus” Journal of Logic Programming.
1992.
Niccolò Machiavelli (1532) The Prince (Bantam Classics 1984).
Robin Milner “Elements of interaction: Turing award lecture.” CACM. January 1993.
Pablo Noriega Agent Mediated Auctions: The Fishmarket Metaphor. Ph.D. Barcelona. 1997.
Julian Padget and Russell Bradford (1998). “A π-calculus Model of a Spanish Fish Market”
AMET-98 also in LNAI 1571 Springer Verlag. 1999.
Carl Petri Kommunikation mit. Automate. Ph. D. Thesis. University of Bonn. 1962.
Murray Shanahan and Rob Miller (1999) “The event-calculus in classical logic — alternative
axiomatization,” Electronic Transactions on Artificial Intelligence 3(1).
Jaime Sichman, Virginia Dignum, and Cristiano Castelfranchi (2005). “Agent organizations.”
JBCS, 11(3).
Munindar Singh. “Social and Psychological Commitments in Multi-agent Systems” AAAI Fall
Symposium on Knowledge and Action at Social and Organizational Levels. 1991
Munindar Singh “An ontology for commitments in multi-agent systems: Toward a unification of
normative concepts” Artificial Intelligence and Law 7. 1999.
Feng Wan and Munindar Singh (2005). “Formalizing and achieving multiparty agreements via
commitments” AAMAS‟05.
Michael Winikoff, Wei Liu, and James Harland (2005). “Enhancing commitment machines”
DALT. LNAI 3476. Springer-Verlag. 2005.
Pinar Yolum and Munindar Singh “Enacting Protocols by Commitment Concession” AAMAS‟07.
Mahadevan Venkatraman and Munindar Singh. “Verifying Compliance with Commitment
Protocols: Enabling Open Web-Based Multi-agent Systems” JAAMAS. 1999.
November 6, 2010 Page 14 of 18
Appendix 1. A simple auction procedure in ActorScript
SimpleAuction
behavior {
Bidders::theBidders a collection of those allowed to bid on this auction
Currency::minimumBid current minimum bid for this auction
Time::deadline current deadline by which this auction will end unless
another higher bid is received for this auction
Bidding::currentBidding a recording of the current state of bidding for this auction
------------------------------------------------------------------------------------------------
implements Auction the Auction interface is implemented below
self.bid (amountBid)@Time::arrivalTime →
a message with amount bid and bidder has been received at arrivalTime
arrivalTime ?? { if arrivalTime
(> deadline) : throw TooLate.[deadline]; is after deadline, complain bid is too late
not (> deadline) :
amountBid ?? { else if the amount bid is smaller than the minimum
(< minimumBid) : throw TooLittle .[minimumBid]; then complain the bid is too little
not (< minimumBid) :
{currentBidding.bid (amountBid time : arrivalTime); record the bid in currentBidding
this may throw an exception if the bidder is unqualified
let (Time
::newDeadline = CurrentTime( )+δ,
Amount ::newMinimumBid = amountBid * 110%)
compute the new deadline and new minimum bid
{theBidders.newMinimum(newMinimumBid deadline : newDeadline),
inform the allowed bidders of the new minimum and deadline
self.alarm(newDeadline), set an alarm for this auction with a new deadline
Acknowledgment .[ ] also become (minimumBid = newMinimumBid,
deadline = newDeadline)}}}
return acknowledgment that the bid has been accepted and
also update this auction with the new minimum bid and deadline
alarm.(alarmTime) → an alarm message with alarmTime has been received
alarmTime ?? { if alarmTime
(<deadline) : void; is before the deadline return void
not (<deadline) : currentBidding.processOutcome}}
else return the result of processing the outcome of this auction according to the currentBidding
November 6, 2010 Page 15 of 18
End Notes
i The architecture of an iOrg differs fundamentally from a Mental Agent that cognitively
operates in a human-like fashion. The Mental Agent paradigm [Alberti, Gavanelli, Lamma,
Mello, and Torroni 2004] has had some success in modeling and simulating human-like
behavior. However, computing has changed dramatically from the time of its invention and
we are in the midst of a “perfect disruption” [Hewitt 2008g] brought on by the following:
Hardware. Many-core architecture that will soon support thousands of threads in a
process for widely-used software applications using semantic integration (see below).
Software. Client cloud computing in which information is permanently stored in
servers on the Internet and cached temporarily on clients that range from si ngle chip
sensors, handhelds, notebooks, desktops, and entertainment centers to huge data
centers. (Even data centers are clients that often cache their information to guard
against geographical disaster.) Client cloud computing will provide much needed new
capabilities including the following:
o maintaining the privacy of client information by storing it on servers encrypted
so that it can be decrypted only by using the client‟s private key. (The
information is unencrypted only when cached on clients.)
o providing greater integration of user information obtained from servers of
competing vendors without requiring them to interact with each other.
o providing better advertising relevance and targeting without exposing client
privacy.
Applications. Scalable semantic integration, e.g., integrating the following:
o
calendars and to do lists
o
email archives
o presence information including physical, psychological and social
o documents (including presentations, spread sheets, proposals, job applications,
photos, videos, gift lists, memos, purchasing, contracts, articles, etc.)
o
contacts (including social graphs)
o
search results
o marketing and advertising relevance influenced by the above
This perfect disruption is causing a paradigm shift from Mental Agents to iOrgs as the
foundation for implementing large-scale Internet applications [Hewitt 2008g].
ii For background information on iOrgs see [Bowker, Star, Turner, and Gasser 1977; Dignum
2004, Singh and Huhns 2005; Horling and Lesser 2005].
iii In some previous work, the subject of contracts [Blum 2004, etc.] has been treated using the
(somewhat unfortunate name) “commitment” for contractual obligations [Singh 1991,
Jennings 1993; Noriega 1997; Singh and Huhns 2005; Chopra and Singh 2009]. In this paper,
these obligations are treated as special cases of Physical Commitment. See Hewitt [2007].
iv i.e., checking the behavior of a system against a model
v Note that Participations (being regions of space -time) represent both objects and activities.
November 6, 2010 Page 16 of 18
vi Note that Participatory Semantics is based on space-time as opposed to the more usual
approach of basing semantics just on time, e.g. the Event Calculus [Farrell, Sergot, Salle, and
Bartolini 2005], etc.
vii XML is used be because it is increasingly dominant as the de facto standard for structured
message communication and stands to become the de facto standard for data structures.
viii quoted in [Folger 2007]
ix For example, such an event would have to exist in the Event Calculus formulation in Kowalski
[1992].
x Inspired by the Blanes FishMarket Metaphor [Noriega 1997].
xi appendix in the concurrent programming language ActorScript [Hewitt 2008c]
xii A functional notation is used for XML. For example
PersonName<First<“Kurt”> Last< “Gödel”>> can print as:
<PersonName> <First> Kurt </First> <Last> Gödel </Last> </PersonName>
xiii Closed means that the system does not receive any communications from outside itself.
xiv The Actor Model subsumes other models of concurrency, e.g. Process Calculi [Milner 1993;
Aceto and Gordon 2005] and Petri Nets [Petri 1962] using a two -phase commit protocol
[Knabe 1992].
xv Implemented in ActorScriptTM [Hewitt 2010]
xvi The symbol is used to begin a comment that extends to the end of the line.
xvii Expressed in Direct Logic [Hewitt 2008c] (see discussion later in this paper).
xviii Note that (unlike Venkatraman and Singh [1999]), no assumption is made that the buyers and
sellers are not malicious, e.g., no use is made of time-stamps that can be forged.
xix Statistical probabilistic (fuzzy logic) systems are affected follows: Suppose (as above)
├Boston ℙ(TrafficJam WeekdayAt5PM) = 1xix
├Boston ℙ(TrafficJam) = 0
Then
├Boston ℙ(WeekdayAt5PM) =
ℙ( )
= 0xix
ℙ( )
Thus contraposition is built into probabilistic (fuzzy logic) systems and consequently incorrect
information can be generated.
The above example illustrates that the choice of how to incorporate measurements into
statistics can effectively determine the model being used. In this particular case, the way that
measurements were taken did not happen to take into account things like holidays and severe
winter storms This point was largely missed in [Anderson 2008] which stated
November 6, 2010 Page 17 of 18
“Correlation is enough.” We can stop looking for models. We can analyze the data
without hypotheses about what it might show. We can throw the numbers into the biggest
computing clusters the world has ever seen and let statistical algorithms find patterns
where science cannot.” (emphasis added)
xx The axiomatization makes use of higher order capabilities. For example a predicate like
CanResult can take arguments Do[Machiavelli, Collude] and Rich[Machiavelli] to form the
proposition CanResult[Do[Machiavelli, Collude], Rich[Machiavelli]].
xxi The axiomatization uses a colon (:) to separate universally quantified variable from the following
proposition to which they apply.
Direct Logic supports fine grained reasoning because inference does not necessarily carry argument in
the contrapositive direction. For example, the general principle “A person does anything that can make
them rich”
(i.e., CanResult[Do[p, action], Rich[p]]├ ThePrince Do[p, action] does not support the inference of
CanResult[Do[Machiavelli, Collude], Rich[Machiavelli]] from
Do[Machiavelli, Collude]). E.g., it might be the case that
CanResult[Do[Machiavelli, Collude], Rich[Machiavelli]] even though it infers
Do[Machiavelli, Collude] contradicting Do[Machiavelli, Collude].
xxii “A prince never lacks legitimate reasons to break his promise” [Machiavelli 1532]
xxiii “The one who adapts his policy to the times prospers, and likewise that the one whose policy
clashes with the demands of the times does not.” [Machiavelli 1532]
November 6, 2010 Page 18 of 18
|
1909.08974 | 1 | 1909 | 2019-09-19T13:22:54 | Robust time-varying formation design for multi-agent systems with disturbances: Extended-state-observer method | [
"cs.MA",
"eess.SY",
"eess.SY"
] | Robust time-varying formation design problems for second-order multi-agent systems subjected to external disturbances are investigated. Firstly, by constructing an extended state observer, the disturbance compensation is estimated, which is a critical term in the proposed robust time-varying formation control protocol. Then, an explicit expression of the formation center function is determined and impacts of disturbance compensations on the formation center function are presented. With the formation feasibility conditions, robust time-varying formation design criteria are derived to determine the gain matrix of the formation control protocol by utilizing the algebraic Riccati equation technique. Furthermore, the tracking performance and the robustness property of multi-agent systems are analyzed. Finally, the numerical simulation is provided to illustrate the effectiveness of theoretical results. | cs.MA | cs |
Robust time-varying formation design for multi-agent systems
with disturbances: Extended-state-observer method
Le Wang, Jianxiang Xi∗, Ming He, Guangbin Liu
Rocket Force University of Engineering, Xi'an, 710025, P.R. China
SUMMARY
Robust time-varying formation design problems for second-order multi-agent systems subjected to
external disturbances are investigated. Firstly, by constructing an extended state observer, the disturbance
compensation is estimated, which is a critical term in the proposed robust time-varying formation control
protocol. Then, an explicit expression of the formation center function is determined and impacts of
disturbance compensations on the formation center function are presented. With the formation feasibility
conditions, robust time-varying formation design criteria are derived to determine the gain matrix of the
formation control protocol by utilizing the algebraic Riccati equation technique. Furthermore, the tracking
performance and the robustness property of multi-agent systems are analyzed. Finally, the numerical
simulation is provided to illustrate the effectiveness of theoretical results.
Received . . .
KEY WORDS: Multi-agent system, robust time-varying formation, external disturbance, extended state
observer, algebraic Riccati equation.
1. INTRODUCTION
Recently, the distributed cooperative control of multi-agent systems has aroused extensive
attentions in multiple fields,
including wireless communication,
robotics and distributed
computation as shown in [1]-[8]. As an important research topic on the distributed control, formation
control refers to design a control strategy with the neighboring information such that a group of
autonomous agents reach an expected geometrical shape. In the past two decades, several classical
formation control methodologies were investigated, such as the leader-follower method [9], the
virtual-structure-based strategy [10] and the behavioral approach [11], among many others. Beard
et al. [12] showed that each of the above-mentioned classical methodologies has its corresponding
weakness. Ren [13] addressed formation control problems by implementing the consensus-based
approach and showed that the above-mentioned classical methodologies could be unified in the
framework of the consensus-based formation control. Inspired by the development of the consensus
Contract/grant sponsor: This work was supported by the National Natural Science Foundation of China under Grants
61867005, 61763040, 61703411 and 61374054, Innovation Foundation of High-Tech Institute of Xi'an (2015ZZDJJ03)
and Scientific Research Project under Grant JJ20172B03058, also supported by Innovation Zone Project under Grant
18-163-11-ZT-004-005-01. ; contract/grant number:
2
theory in the control community (e.g., [14]-[19]), the newly developed consensus-based formation
control strategies were reported in many application fields including mobile robots, intelligent
ground vehicles and unmanned aerial vehicles (see [20]-[25] and the references therein).
In many practical circumstances, multi-agent systems may suffer external disturbances due to
environmental uncertainties, which can drive these systems to oscillations or divergences. For
example, in the formation flying of multiple quadrotors, atmospheric disturbances can be regarded
as additional forces and may cause instabilities in both the attitude and position dynamics. It
is significant to address disturbance rejection problems such that multi-agent systems can reach
the asymptotical disturbance rejection while conserving the closed-loop stability. Jafarian et al.
[26] studied the formation keeping control for a group of nonholonomic wheeled robots with
matched input disturbances, where the disturbances were compensated by internal-model-based
controllers. In [27], the time-invariant formation tracking control for a group of quadrotors with
unknown bounded disturbances was achieved by designing an H∞ control controller, where the
disturbance cannot be rejected by the proposed method in the whole desired frequency range. Liu et
al. [28] proposed a robust compensating filter to handle time-invariant formation control problems of
multiple quadrotors with disturbance rejections in the whole frequency domain as much as desired.
Note that the desired formation was time-invariant in [20]-[28]. However, time-varying formation
configurations are required in many applications due to complex external environments and/or
variable mission situations. For example, the formation shape should be changed in the obstacle
avoidance for multiple mobile robots. Several significant results about the time-varying formation
control were obtained in [29]-[33]. Cooperative time-varying formation control methods was studied
in [29], where the formation was characterized by time-varying external parameters. A time-varying
formation of collaborative unmanned aerial vehicles and unmanned ground vehicles was achieved
in [30]. Time-varying formation tracking control was reached when considering the influence of
switching topologies in [31]. Dong et al. [32] investigated the time-varying formation analysis and
design problems for second-order multi-agent systems with directed topologies, where a formation
feasibility condition was proposed to show that not all expected formation could be achieved. Time-
varying group formation control for multi-agent systems with directed topologies was showed in
[33]. However, further investigates on disturbance rejections of the time-varying formation with the
influence of external disturbances were not considered in [29]-[33], and the disturbance rejection
methods for time-invariant formations in [26]-[28] cannot be implemented since the expected
formation is time-varying. To the best of our knowledge, robust time-varying formation design
problems for second-order multi-agent systems with unknown external disturbances have not been
investigated extensively.
Motivated by the above-mentioned facts, the current paper develops an extended-state-observer
method to tackle the robust time-varying formation control problem for multi-agent systems
subjected to external disturbances. With the disturbance compensation, a novel robust time-varying
formation control protocol is proposed, using only relative neighboring information. By regarding
external disturbances as additional states, an extended state observer (ESO) is constructed to
determine the disturbance compensation. Then, the closed-loop dynamics of the whole multi-agent
system is divided into two parts. The first one is the formation agreement dynamics, which is
utilized to derive an explicit expression of the formation center function. The second part, called
the disagreement dynamics, can describe the relative motion among agents. Sufficient conditions of
()
3
the robust time-varying formation design are determined via algebraic Riccati equation techniques,
together with the formation feasibility conditions. Moreover, the tracking performance and the
robustness stability of the closed-loop system are analyzed.
Compared with the existing results about the time-varying formation control of multi-agent
systems, the current paper contains the following three novel features. Firstly, to achieve the
disturbance rejection control objective, a robust time-varying formation control protocol is proposed
with the robust disturbance compensation. Time-varying formation protocols in [29]-[33] cannot
deal with the robust time-varying formation control problems when the influence of external
disturbances is considered. Secondly, an ESO is constructed to determine the robust disturbance
compensation, which can actively compensate the external disturbance in real time. Tracking
performances and robust properties are analyzed with the ESO and the formation feasibility
condition. However, disturbance compensations and robust properties were not considered in [29]-
[33]. Thirdly, an explicit expression of the formation center function is deduced to show the
macroscopic motion of the whole formation under the influence of external disturbances. It is
revealed that that the disturbance compensation has effects on formation center functions. However,
[29]-[31] did not give the formation center function and the formation center functions in [32] and
[33] could not determine the impact of disturbance compensation.
An outline of the current paper is presented as follows. Section 2 gives the problem description.
In Section 3, an explicit expression of the formation center functions is determined. In Section
4, sufficient conditions of the robust time-varying formation design are shown and the tracking
performance and the robustness stability are analyzed. Section 5 illustrates the effectiveness of
theoretical results via a numerical simulation. Conclusions are stated in Section 6.
Notations: Let Rn and Rn×m be the n-dimension real column vector and the n × m-
dimension real matrix, respectively. For simplicity, 0 uniformly represents the zero number,
zero vectors and zero matrices. 1N stands for an N -dimensional column vector with each
entry being 1. P −1, P H and P T denote the inverse matrix,
the Hermitian adjoint matrix
and the transpose matrix of P , respectively. The norm used here is respectively defined
as kh(t)k1 = maxi(cid:16)PjR ∞
0
and kp(t)k∞ =
maxisupt>0 pi(t), where · is the absolute value, h(t) = [hij(t)] ∈ Rm×n and p(t) = [pi(t)] ∈ Rn.
The Kronecker product is represented by the notation ⊗.
hij(t) dt(cid:17), kp(t)k = kp(t)k2 =(cid:16)Pn
i=1 pi(t)2(cid:17)1/2
2. PROBLEM DESCRIPTION
Consider a group of N identical agents with the dynamics of the ith agent described by:
( pi(t) = vi(t),
vi(t) = αppi(t) + αvvi(t) + ui(t) + ωi(t),
(1)
where i = 1, 2, · · · , N , pi(t), vi(t) and ui(t) ∈ Rn represent the position, the velocity and the
control input, respectively, ωi(t) ∈ Rn is the unknown bounded external disturbance and αp and
αv are the damping constants. The interaction topology among agents is described by a digraph G,
where agent i is represented by the ith node, the interaction channel among nodes is denoted by an
edge and the interaction strength is depicted by the edge weight wij . Note that wij > 0 if agent j
()
4
belongs to the neighbor set Ni of agent i and wij = 0 otherwise. For the digraph G, the weighted
adjacency matrix is W = [wij ]N ×N and D = diag{d1, d2, · · · , dN } stands for the in-degree matrix.
Define the Laplacian matrix of G as L = D − W . A directed path from node i to node j is a finite
ordered sequence of edges described as {(vi, vm), (vm, vn), · · · , (vl, vj)}. A digraph is said to have
a spanning tree if a root node i exists such that it at least has a directed path to every other node.
Lemma 1 ([34])
If G has a spanning tree, then 0 is its single eigenvalue with 1N being the related eigenvector and
other N − 1 eigenvalues have positive real parts; that is, 0 = λ1 < Re(λ2) 6 · · · 6 Re(λN ).
Let fi(t) = [f T
iv(t)]T ∈ R2n (i ∈ {1, 2, · · · , N }) be a piecewise continuously differ-
ip(t), f T
then the expected time-varying formation is specified by a vector f (t) =
N (t)]T . By considering the disturbance compensation, a robust time-varying
entiable vector,
[f T
formation control protocol is proposed as follows:
2 (t), · · · , f T
1 (t), f T
ui(t) = Ku Xj∈Ni
wij (xj(t) − xi(t) − fj(t) + fi(t)) − αfi(t) + fiv(t) − zi(t),
(2)
i (t)]T , α = [αp, αv] ⊗ In, Ku ∈ Rn×2n is the gain
where i ∈ {1, 2, · · · , N }, xi(t) = [pT
matrix and zi(t) is the robust disturbance compensation, which is determined by the following ESO:
i (t), vT
( gi(t) = zi(t) + ui(t) + αppi(t) + αvvi(t) − βig (gi(t) − vi(t)) ,
zi(t) = −βiz (gi(t) − vi(t)) ,
(3)
where βig and βiz are bandwidth constants, and gi(t) is the intermediate variable of the ESO.
Let x(t) = [xT
N (t)]T , θ1 = [1, 0]T ⊗ In and θ2 = [0, 1]T ⊗ In, then multi-agent
system (1) with protocol (2) can be rewritten as a global closed-loop system with the following
2 (t), · · · , xT
1 (t), xT
dynamics:
x(t) =(cid:0)IN ⊗(cid:0)θ1θT
2 + θ2α(cid:1) − L ⊗ θ2Ku(cid:1) x(t) − (IN ⊗ θ2α − L ⊗ θ2Ku) f (t)
+(cid:0)IN ⊗ θ2θT
2(cid:1) f (t) + (IN ⊗ θ2) (ω(t) − z(t)) .
Definition 1
(4)
For any given positive constant ε and bounded initial states x(0), multi-agent systems (1) is said to
be robust time-varying formation-reachable by protocol (2) if all states involved in the global closed-
loop system (4) are bounded and there exist a gain matrix Ku, a vector-valued function c(t) and a
finite constant tε such that kxi(t) − fi(t) − c(t)k 6 ε (∀i ∈ {1, 2, · · · , N }), ∀t > tε, where c(t) is
the formation center function and ε is called the time-varying formation error bound, respectively.
The control objective of the current paper is to design the robust time-varying formation control
protocol such that second-order multi-agent systems with external disturbances can reach the
expected robust time-varying formation. The following three problems are focused: (i) Determining
an explicit expression of formation center functions; (ii) Designing the gain matrix Ku of protocol
(2); (iii) Analyzing the tracking performance and the robustness property of the global closed-loop
system.
()
3. FORMATION CENTER FUNCTIONS
This section gives an explicit expression of formation center functions and shows impacts of
the time-varying formation and the disturbance compensation on the formation center function,
respectively.
5
Let ξi(t) = xi(t) − fi(t) (i ∈ {1, 2, · · · , N }) and ξ(t) = [ξT
1 (t), ξT
2 (t), · · · , ξT
N (t)]T , then global
closed-loop system (4) can be transformed into
ξ(t) =(cid:0)IN ⊗(cid:0)θ1θT
2 + θ2α(cid:1) − L ⊗ θ2Ku(cid:1) ξ(t) + (IN ⊗ θ2) (ω(t) − z(t))
+(cid:0)IN ⊗ θ1θT
2(cid:1) f (t) −(cid:0)IN ⊗ θ1θT
1(cid:1) f (t).
Let U = [1N , u] ∈ RN ×N be a nonsingular matrix with u = [u2, u3, · · · , uN ] ∈ RN ×(N −1) such
N ]H ∈ R(N −1)×N and J is
that U −1LU = J , where U −1 = [¯uH
the Jordan canonical form of L.
1 , ¯uH]H with ¯u = [¯uH
3 , · · · , ¯uH
2 , ¯uH
According to Lemma 1 and the structure of U , one can obtain that J = diag{0, J}, where J
consists of the corresponding Jordan blocks of λi (i = 2, 3, · · · , N ). Let ξ(t) = (U −1 ⊗ I2n)ξ(t) =
[κT (t), ϕT (t)]T , in which κ(t) = ξ1(t) and ϕ(t) = [ ξT
N (t)]T , then multi-agent
system (5) can be transformed into
3 (t), · · · , ξT
2 (t), ξT
(5)
(6)
(7)
κ(t) =(cid:0)θ1θT
2 + θ2α(cid:1) κ(t) + (¯u1 ⊗ θ2) (ω(t) − z(t)) + (¯u1 ⊗ θ1)(cid:0)fv(t) − fp(t)(cid:1) ,
2 + θ2α(cid:1) − J ⊗ θ2Ku(cid:1) ϕ(t) + (¯u ⊗ θ2) (ω(t) − z(t))
ϕ(t) =(cid:0)IN −1 ⊗(cid:0)θ1θT
+ (¯u ⊗ θ1)(cid:0)fv(t) − fp(t)(cid:1) ,
where fv(t) = [f T
1v(t), f T
2v(t), · · · , f T
N v(t)]T and fp(t) = [ f T
1p(t), f T
2p(t), · · · , f T
N p(t)]T .
Subsystems (6) and (7) depict the formation agreement and disagreement dynamics of multi-agent
system (1), which describe the absolute movement of the whole system and the relative movement
among agents, respectively. According to subsystem (6), the following theorem determines the
impact of the disturbance compensation on the formation center function and shows an explicit
expression of the formation center function, which describes the macroscopic motion of the whole
formation.
Theorem 1
For any given ε > 0, if multi-agent system (1) reaches the expected robust time-varying formation
f (t), then the formation center function c(t) satisfies that
kc(t) − c0(t) − cz(t) − cf (t)k 6 ε, ∀t > tε,
where tε is a finite constant and
c0(t) = e(θ1θT
2 +θ2α)t(¯u1 ⊗ I2n)x(0),
cz(t) =Z t
0
e(θ1θT
2 +θ2α)(t−ς) (¯u1 ⊗ θ2) (ω(ς) − z(ς))ds,
()
6
cf (t) =Z t
0
e(θ1θT
2 +θ2α)(t−τ ) (¯u1 ⊗ θ2)(cid:0) fv(τ ) − αpfp(τ ) − αvfv(τ )(cid:1)dτ − (¯u1 ⊗ I2n) f (t).
Proof
Let e1 ∈ RN denote a unit vector with its first element being 1. Define the following auxiliary
functions:
ξa(t) = (U ⊗ I2n)[κT (t), 0]T ,
ξd(t) = (U ⊗ I2n)[0, ϕT (t)]T ,
(8)
(9)
with kU ⊗ I2nk = εN . Due to (U −1 ⊗ I2n)ξ(t) = [κT (t), ϕT (t)]T , it can be obtained from (8) and
(9) that
ξ(t) = ξa(t) + ξd(t).
(10)
Since U ⊗ I2n is nonsingular, one can concluded that ξa(t) and ξd(t) are linearly independent. It
follows from (8) and the fact [κT (t), 0]T = e1 ⊗ κ(t) that
ξa(t) = (U ⊗ I2n) (e1 ⊗ κ(t)) = U e1 ⊗ κ(t) = 1N ⊗ κ(t).
From (10) and (11), one can show that
ξd(t) = ξ(t) − 1N ⊗ κ(t).
(11)
(12)
From (9), (10) and (12), one can find that for any given positive constant ε, there exists a finite
constant tε such that kxi(t) − fi(t) − κ(t)k 6 ε (i ∈ {1, 2, · · · , N }), ∀t > tε, if kϕ(t)k 6 ε/εN =
εϕ, ∀t > tε, which means that ϕ(t) represents the time-varying formation error and κ(t) shows one
of the candidates of formation center functions, respectively.
From (8), one can obtain that
κ(0) = (¯u1 ⊗ I2n)(x(0) − f (0)).
(13)
One can show that
0
Z t
=Z t
−Z t
0
0
and
e(θ1θT
2 +θ2α)(t−τ ) (¯u1 ⊗ θ1)(cid:0)fv(τ ) − fp(τ )(cid:1)dτ
2 +θ2α)(t−τ ) (¯u1 ⊗ θ1) fv(τ )dτ + e(θ1θT
e(θ1θT
2 +θ2α)t (¯u1 ⊗ θ1) fp(0)
2 + θ2α)θ1(cid:1) fp(τ )dτ − (¯u1 ⊗ θ1) fp(t),
e(θ1θT
2 +θ2α)(t−τ )(cid:0)¯u1 ⊗ (θ1θT
Z t
0
= (¯u1 ⊗ θ2) fv(t) − e(θ1θT
e(θ1θT
2 +θ2α)(t−τ ) (¯u1 ⊗ θ2) fv(τ )dτ
2 +θ2α)t (¯u1 ⊗ θ2) fv(0)
(14)
()
+Z t
0
e(θ1θT
2 +θ2α)(t−τ )(cid:0)¯u1 ⊗ (θ1θT
2 + θ2α)θ2(cid:1) fv(τ )dτ.
By the structure of f (t), it can be found that
(¯u1 ⊗ θ1) fp(t) + (¯u1 ⊗ θ2) fv(t) = (¯u1 ⊗ I2n) f (t),
e(θ1θT
2 +θ2α)t ((¯u1 ⊗ θ1) fp(0) + (¯u1 ⊗ θ2) fv(0)) = e(θ1θT
2 +θ2α)t (¯u1 ⊗ I2n) f (0).
Then, it follows from (14)-(17) that
e(θ1θT
0
Z t
=Z t
0
2 +θ2α)(t−τ ) (¯u1 ⊗ θ1)(cid:0)fv(τ ) − fp(τ )(cid:1)dτ
2 +θ2α)(t−τ ) (¯u1 ⊗ θ2)(cid:0) fv(τ ) − αpfp(τ ) − αvfv(τ )(cid:1)dτ
e(θ1θT
+e(θ1θT
2 +θ2α)t (¯u1 ⊗ I2n) f (0) − (¯u1 ⊗ I2n) f (t).
7
(15)
(16)
(17)
(18)
In virtue of (6), (13) and (18), the conclusion of Theorem 1 can be obtained.
4. ROBUST TIME-VARYING FORMATION DESIGN
In this section, firstly, an algorithm is presented to show the procedure of designing the robust
time-varying formation control protocol. Then, sufficient conditions of the robust time-varying
formation design are shown and the tracking performance and the robustness stability of multi-agent
systems are analyzed, respectively.
The core idea of designing robust time-varying formation control protocol (2) is to determine the
gain matrix and the robust disturbance compensation. The following algorithm with four steps is
presented to design protocol (2).
Robust Time-Varying Formation Design Algorithm
Step 1: Check the following formation feasibility condition for the expected time-varying
formation.
6 εf , ∀t > tε, ∀i ∈ {1, 2, · · · , N }.
(19)
(cid:13)(cid:13)fiv(t) − fip(t)(cid:13)(cid:13)∞
If condition (19) is satisfied, then go to Step 2; else the expected time-varying formation cannot be
reached by multi-agent system (1) with protocol (2) and the algorithm stops.
Step 2: Solve the following algebraic Riccati equation for a positive definite matrix P
P (θ1θT
2 + θ2α) + (θ1θT
2 + θ2α)T P − P θ2θT
2 P + I = 0.
(20)
Step 3: Set the gain matrix Ku as Ku = Re−1(λ2)θT
Step 4: Choose sufficiently large bandwidth constants βig and βiz (i ∈ {1, 2, · · · , N }) of ESO (3)
2 P .
to effectively estimate the robust disturbance compensation.
()
8
With the robust time-varying formation design algorithm, tracking performances and robustness
stability properties are analyzed in the following theorem.
Theorem 2
For any given bounded initial states, if formation feasibility condition (19) is satisfied, then multi-
agent system (1) reaches the robust time-varying formation by protocol (2) designed in the robust
time-varying formation design algorithm.
Proof
Firstly, consider the stability of the following subsystem:
where ∀k ∈ {2, 3, · · · , N }. Then construct the Lyapunov function candidate as follows:
ηk(t) =(cid:0)θ1θT
2 + θ2α − λkθ2Ku(cid:1) ηk(t),
Vk(t) = ηH
k (t)P ηk(t).
Let Ku = Re−1(λ2)θT
2 P , then differentiating V (t) along the trajectories of (21) yields
Vk(t) = ηH
k (t)(cid:16)(θ1θT
2 + θ2α)
T
P + P (θ1θT
2 + θ2α) − 2 Re(λk)Re−1(λ2)P θ2θT
2 P(cid:17) ηk(t).
Substituting P (θ1θT
2 + θ2α) + (θ1θT
2 + θ2α)T P = P θ2θT
2 P − I into (23) gives
Vk(t) = ηH
k (t)(cid:0)(cid:0)1 − 2 Re(λk)Re−1(λ2)(cid:1) P θ2θT
2 P − I(cid:1) ηk(t).
(21)
(22)
(23)
(24)
Due to 0 < Re(λ2) 6 · · · 6 Re(λN ), one can derive from (24) that
{2, 3, · · · , N }). Therefore, ηk(t) converges to 0 asymptotically, which means that θ1θT
λkθ2Ku is Hurwitz. By the structure of J , one can conclude that IN −1 ⊗ (θ1θT
is Hurwitz.
k (t)ηk(t) (∀k ∈
2 + θ2α −
2 + θ2α) − J ⊗ θ2Ku
Vk(t) 6 −ηH
Then, the tracking performance and the robustness stability is analyzed. Let A = IN −1 ⊗ (θ1θT
2 +
θ2α) − J ⊗ θ2Ku, then subsystem (7) can be rewritten as
ϕ(t) = Aϕ(t) + (¯u ⊗ θ2) (ω(t) − z(t)) + (¯u ⊗ θ1)(cid:0)fv(t) − fp(t)(cid:1) .
By Laplace transform, (3) can be converted to
( zi(s) − ωi(s) + (βig + s) (vi(s) − gi(s)) = 0,
szi(s) + βizgi(s) − βizvi(s) = 0,
where i ∈ {1, 2, · · · , N }. From (26), it can be shown that
zi(s) = Gi(s)ωi(s),
(25)
(26)
(27)
where Gi(s) = βiz(cid:14)(s2 + βigs + βiz), i ∈ {1, 2, · · · , N }. Let βig = 2σi and βiz = σ2
obtain that
i , then one can
Gi(s) =
σ2
i
(s + σi)2 .
(28)
()
From (27), it follows that
ω(s) − z(s) = diag{1 − G1(s), 1 − G2(s), · · · , 1 − GN (s)}ω(s) = ΦN (s)ω(s).
(29)
Define
9
ρω =(cid:13)(cid:13)(cid:13)
ρf =(cid:13)(cid:13)(cid:13)
υϕ(0) =(cid:13)(cid:13)eAtϕ(0)(cid:13)(cid:13)∞
.
,
(sI2n(N −1) − A)−1 (¯uΦN (s) ⊗ θ2)(cid:13)(cid:13)(cid:13)1
(sI2n(N −1) − A)−1 (¯u ⊗ θ1)(cid:13)(cid:13)(cid:13)1
kϕ(t)k∞ 6 υϕ(0) + ρωkω(t)k∞ + ρf(cid:13)(cid:13)fv(t) − fp(t)(cid:13)(cid:13)∞
,
(30)
.
(31)
From (25), (27), (29) and (30), it can be derived that
For agent i (i ∈ {1, 2, · · · , N }), since ω(t) is bounded, there exist two positive constants γϕi and
δωϕi such that
kωi(t)k∞ 6 γϕikui(t)k∞ + δωϕi, i ∈ {1, 2, · · · , N }.
It follows from (32) that positive constants γϕ and δωϕ exist such that
kω(t)k∞ 6 γϕku(t)k∞ + δωϕ.
By (2), (3) and (29), one can show that
ku(t)k∞ = δuϕ1kϕ(t)k∞ + δuϕ2kω(t)k∞ + δuϕ3,
(32)
(33)
(34)
where δuϕ1, δuϕ2 and δuϕ3 are positive constants. Substituting (34) into (33), one can obtain that υϕ
and υe exist such that
kω(t)k∞ 6 υϕkϕ(t)k∞ + υe.
(35)
If k¯u ⊗ θ2k∞ is bounded and σi (i = 1, 2, · · · , N ) are sufficiently large, then it can be deduced from
(31) and (35) that
It follows from (36) that
kω(t)k∞ 6
kϕ(t)k∞ 6
υϕυϕ(0) + υe
1 − υϕρω
υϕ(0) + υeρω
1 − υϕρω
,
.
( kω(t)k∞ 6 υω,
kϕ(t)k∞ 6 υϕ,
(36)
(37)
()
10
where υω and υϕ are positive constants. According to the formation feasibility condition (19), one
has that
6 εf , ∀t > tf .
(38)
From (31), (37) and (38), one can obtain that
(cid:13)(cid:13)fv(t) − fp(t)(cid:13)(cid:13)∞
eAtϕ(0)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)
2n(N −1),i
cT
i
max
i
ϕi(t) 6 max
+ ρω υω + ρf εf , ∀t > tf ,
(39)
where i ∈ {2, 3, · · · , N }, c2n(N −1),i is a 2n(N − 1)-dimensional unit column vector with the ith
element 1 and other elements 0. For the bounded initial states ϕi(0) (i ∈ {2, 3, · · · , N }), one can find
that ϕi(t) is bounded. It can also be obtained that the states of the robust disturbance compensation
zi(t) and the control protocol ui(t) are bounded. It follows that all states involved in the closed-loop
system (4) are bounded. Furthermore, since A is Hurwitz, there exists a finite constant tε > tf such
that kϕ(t)k 6 εϕ, ∀t > tε for any given positive constant εϕ, which means that multi-agent system
(1) is robust time-varying formation-reachable by protocol (2). This completes the proof.
5. NUMERICAL SIMULATIONS
In this section, a simulation example is provided to demonstrate the effectiveness of the theoretical
results obtained in previous sections.
Consider a second-order multi-agent system containing six agents in the XY Z space
(n = 3), where the interaction topology among agents is described as a 0-1 weighted
digraph in Figure 1. The dynamics of each agent can be described by (1) with αp = −0.01
and αv = 0. Let xi(t) = [piX (t), piY (t), piZ (t), viX (t), viY (t), viZ (t)]T
(i ∈ {1, 2, · · · , 6}),
where piX (t), piY (t), piZ(t) and viX (t), viY (t), viZ (t) are positions and velocities
along the X axis, Y axis and Z axis, respectively. The initial states of each agent are
x2(t) = [−1.5, −0.3, 1.8, −1.6, 2.3, 1.1]T ,
x5(t) =
set
x3(t) = [2.1, 0.8, −1.6, 0.3, −1.9, 2.5]T ,
[4.5, 1.9, −1.2, −2.9, 3.5, −1.4]T and x6(t) = [−4.2, 2.9, 3.8, −5.1, −3.5, 2.7]T .
x4(t) = [3.8, 1.7, −2.6, 1.8, −3.3, 1.5]T ,
x1(t) = [0.6, 1.2, 0.5, −1.2, −0.3, 0.8]T ,
as
1
6
2
5
3
4
Figure 1. Interaction topology G.
()
The six agents are required to follow a time-varying formation in the form of
11
fi(t) = 3
sin(t + (i − 1)π/3)
cos(t + (i − 1)π/3)
− sin(t + (i − 1)π/3)
cos(t + (i − 1)π/3)
− sin(t + (i − 1)π/3)
− cos(t + (i − 1)π/3)
, (i = 1, 2, · · · , 6).
It can be found from fi(t) that both positions and velocities of six agents can take shape regular
hexagons with time-varying edges. One can see that formation feasibility condition (19) is satisfied.
The external disturbances are generated by
ωi(t) =
(2.5 + 0.2(i − 1)) sin t + 1.5 + 1.2(i − 1)
(1.5 + 0.2(i − 1)) sin t + 2.5 + 1.2(i − 1)
(2 + 0.2(i − 1)) sin(t + 0.4π) + 3 + 0.2(i − 1)
, (i = 1, 2, · · · , 6).
Choose the bandwidth constants of ESO (3) as βig = 2σi and βiz = σ2
i with σi = 10 (i ∈
{1, 2, · · · , 6}). From Theorem 2, one can determine that Ku = [1.0654, 1.8576] ⊗ I3.
Figures 2 and 3 describe the position and velocity trajectory of six agents and the formation
center at t = 0s, t = 10s, t = 15s and t = 20s, respectively, where the position and velocity states
of agents are represented by asterisks, plus signs, circles, x marks, pentagrams and squares, and the
formation centers are denoted by hexagrams. Figure 4 presents the curves of the formation centers
for positions and velocities within t = 20s, where the initial and final states are depicted by circles
and squares, respectively. Figure 5 shows the trajectory of the formation center within t = 20s.
From Figures 2(a)-(b) and 3(a)-(b), one can find that the multi-agent system can achieve the
regular pentagon formation in both position and velocity states. Figures 2(b)-(d) and 3(b)-(d) present
that the formation keeps rotation in position and velocity states, respectively; that is, the formation
is time-varying. From the simulation results shown in Figures 2-5, one can conclude that second-
order multi-agent system (1) with external disturbance achieves the robust time-varying formation
by protocol (2).
)
t
(
Z
i
p
4
2
0
-2
-4
3
2
0
1
piY (t)
(a) t = 0s
50
48
46
44
42
-2
)
t
(
Z
i
p
5
0
piX (t)
-5
-1
-40
piX (t)
-45
-50-8
-4
-6
piY (t)
(b) t = 10s
()
12
)
t
(
Z
i
p
56
54
52
50
48
0
48
46
44
42
)
t
(
Z
i
p
-2
-6
-4
piY (t)
(c) t = 15s
-45
-50
piX (t)
40
2
-55
-8
-35
-40
piX (t)
-45
-6
0
-4
-2
piY (t)
(d) t = 20s
Figure 2. Position curves of six agents and the formation center at different time.
)
t
(
Z
i
v
)
t
(
Z
i
v
3
2
1
0
-1
-2
4
4
2
0
-2
-4
4
2
2
-2
0
viY (t)
(a) t = 0s
-2
0
viY (t)
(c) t = 15s
6
4
2
0
)
t
(
Z
i
v
10
0
viX (t)
-2
4
-10
-4
2
0
-2
-4
-6
4
)
t
(
Z
i
v
5
0
viX (t)
-5
-4
2
2
5
0
viX (t)
-5
-4
-2
0
viY (t)
(b) t = 10s
10
0
viX (t)
-10
-4
-2
0
viY (t)
(d) t = 20s
Figure 3. Velocity curves of six agents and the formation center at different time.
)
t
(
Z
p
c
60
50
40
30
20
10
0
2
8
6
4
2
0
-2
-4
1
)
t
(
Z
v
c
50
0
cpX (t)
0
-2
-4
-50
-6
cpY (t)
0
-1
cvY (t)
-2
-3
10
0
-10
cvX (t)
(a) Formation center for positions
(b) Formation center for velocities
Figure 4. Curves of the formation center for positions and velocities.
()
13
20
15
10
5
k
)
t
(
ϕ
k
0
0
5
10
t / s
15
20
Figure 5. Trajectory of the time-varying formation error.
6. CONCLUSIONS
In the current paper, robust time-varying formation design problems for second-order multi-
agent systems with external disturbances and directed topologies were studied. A new robust
time-varying formation control protocol was proposed with only relative neighboring information
and an ESO was designed to estimate and compensate the external disturbances. An explicit
expression of the formation center function was derived, where the impacts of the disturbance
compensation and the time-varying formation on the motion mode of the whole formation were
determined. Sufficient conditions of the robust time-varying formation design were presented via
algebraic Riccati equation technique together with the formation feasibility conditions. The tracking
performance and the robustness stability of multi-agent systems were analyzed. It was proven that
multi-agent systems can reach the expected robust time-varying formation if the gain matrix can be
designed and the bandwidth constants of the ESO could be selected properly.
REFERENCES
1. Richert D, Cortes J. Optimal leader allocation in UAV formation pairs ensuring cooperation. Automatica 2013;
49(11): 3189-3198.
2. Tan Z, Cai N, Zhou J, Zhang S. On performance of peer review for academic journals: analysis based on distributed
parallel system. IEEE Access 2019, DOI: 10.1109/ACCESS.2019.2896978.
3. Xi J, Fan Z, Liu H, Zheng T. Guaranteed-cost consensus for multiagent networks with Lipschitz nonlinear dynamics
and switching topologies. International Journal of Robust and Nonlinear Control 2018; 28(7): 2841-2852.
4. Ji Z, Yu H. A new perspective to graphical characterization of multi-agent controllability. IEEE Transactions on
Cybernetics 2017; 47(6): 1471-1483.
5. Ma H, Jia X, Cai N, Xi J. Adaptive guaranteed-performance consensus control for multi-agent systems with an
adjustable convergence speed. Discrete Dynamics in Nature and Society 2019; 5190301.
6. Yang X, Yu W, Wang R, Zhang G, Nie F. Fast spectral clustering learning with hierarchical bipartite graph for
large-scale data. Pattern Recognition Letters 2018, DOI: 10.1016/j.patrec.2018.06.024.
7. Liu X, Ji Z, Hou T. Stabilization of heterogeneous multi-agent systems via harmonic control. Complexity 2018;
8265637.
8. Xi J, Wang C, Liu H, Wang L. Completely distributed guaranteed-performance consensualization for high-order
multiagent systems with switching topologies. IEEE Transactions on Systems, Man, and Cybernetics: Systems
2019; 49(7): 1338-1348.
9. Mahmood A, Kim Y. Leader-following formation control of quadcopters with heading synchronization. Aerospace
Science and Technology 2015; 47(1): 68-74.
()
14
10. Balch T, Arkin R. Behavior-based formation control for multirobot teams. IEEE Transactions on Robotics and
Automation 1998; 14(6): 926-939.
11. Lewis M, Tan K. High precision formation control of mobile robots using virtual structures. Autonomous Robots
1997; 4(4): 387-403.
12. Beard R, Lawton J, Hadaegh F. A coordination architecture for spacecraft formation control. IEEE Transactions on
Control System Technology 2001; 9(6): 777-790.
13. Ren W. Consensus strategies for cooperative control of vehicle formations. IET Control Theory and Application
2007; 1(2): 505-512.
14. Zhang Y, Sun J, Liang H, Li H. Event-triggered adaptive tracking control for multi-agent systems with unknown
disturbances. IEEE Transactions on Cybernetics 2018, DOI: 10.1109/TCYB.2018.2869084.
15. Xi J, He M, Liu H, Zheng J. Admissible output consensualization control for singular multi-agent systems with
time delays. Journal of the Franklin Institute 2016; 353(16): 4074-4090.
16. Li Z, Chen J. Robust consensus of linear feedback protocols over uncertain network graphs. IEEE Transactions on
Automatic Control 2017; 62(8): 4251-4258.
17. Zheng T, Xi J, Yuan M, Liu G. Guaranteed-performance consensus design for Lipschitz nonlinear multi-agent
systems with jointly connected topologies. International Journal of Robust and Nonlinear Control 2019, DOI:
10.1002/rnc.4572.
18. Zhang Y, Li H, Sun J, He W. Cooperative adaptive event-triggered control for multi-agent systems with actuator
failures. IEEE Transactions on Systems, Man and Cybernetics: Systems 2018, DOI: 10.1109/TSMC.2018.2883907.
19. Zuo Z, Wang C, Ding Z. Robust consensus control of uncertain multi-agent systems with input delay: a model
reduction method. International Journal of Robust and Nonlinear Control 2017; 27(11): 1874-1894.
20. Fax J, Murray R. Information flow and cooperative control of vehicle formations. IEEE Transactions on Automatic
Control 2004; 49(9): 1465-1476.
21. Du H, Wen G, Cheng Y, He Y, Jia R. Distributed finite-time cooperative control of multiple high-order
nonholonomic mobile robots. IEEE Transactions on Neural Networks and Learning Systems 2017; 28(12): 2998-
3006.
22. Abdessameud A, Taybi A. Formation control of VTOL unmanned aerial vehicles with communication delays.
Automatica 2011; 47(11): 2383-2394.
23. Du H, Zhu W, Wen G, Duan Z, Lu J. Distributed formation control of multiple quadrotor aircraft based on
nonsmooth consensus algorithms. IEEE Transactions on Cybernetics 2019; 49(1): 342-353.
24. Ou M, Du H, Li S. Finite-time formation control of multiple nonholonomic mobile robots. International Journal of
Robust and Nonlinear Control 2014; 24(1): 140-165.
25. Qin W, Liu Z, Chen Z. A novel observer-based formation for nonlinear multi-agent systems with time delay and
intermittent communication. Nonlinear Dynamics 2015; 79(3): 1651-1664.
26. Jafarian M, Vos E, Persis C, Scherpen J, Schaft A. Disturbance rejection in formation keeping control of
nonholonomic wheeled robots. International Journal of Robust and Nonlinear Control 2016; 26(15): 3344-3362.
27. Jasim W, Gu D. Robust team formation control for quadrotors. IEEE Transactions on Control System Technology
2018; 26(4): 1516-1523.
28. Liu H, Ma T, Lewis F, Wan Y. Robust formation control for multiple quadrotors with nonlinearities and
disturbances. IEEE Transactions on Cybernetics, 2018, DOI: 10.1109/TCYB.2018.2875559.
29. Brin´on-Arranz L, Seuret A, Canudas-de-Wit C. Cooperative control design for time-varying formations of multi-
agent systems. IEEE Transactions on Automatic Control 2014; 59(8): 2283-2288.
30. Rahimi R, Abdollahi F, Naqshi K. Time-varying formation control of a collaborative heterogeneous multi agent
system. Robotics and Autonomous Systems 2014; 62(12): 1799-1805.
31. Dong X, Zhou Y, Ren Z, Zhong Y. Time-varying formation tracking for second-order multi-agent systems subjected
to switching topologies with application to quadrotor formation flying. IEEE Transactions on Industrial Electronics
2017; 64(6): 5014-5024.
32. Dong X, Yu B, Shi Z, Zhong Y. Time-varying formation control for unmanned aerial vehicles: theories and
applications. IEEE Transactions on Control Systems Technology 2015; 23(1): 340-348.
33. Dong X, Li Q, Zhao Q, Ren Z. Time-varying group formation analysis and design for general linear multi-agent
systems with directed topologies. International Journal of Robust and Nonlinear Control 2017; 27(9): 1640-1652.
34. Godsil C, Royal G. Algebraic Graph Theory New York, NY: Springer-Verlag; 2001.
35. Sun B, Gao Z. A DSP-based active disturbance rejection control design for a 1-kW H-bridge DC-DC power
converter. IEEE Transactions on Industrial Electronics 2005; 52(5): 1271-1277.
36. Han J. From PID to active disturbance rejection control. IEEE Transactions on Industrial Electronics 2009; 56(3):
900-906.
()
|
1703.02367 | 1 | 1703 | 2017-03-07T13:18:14 | Vocabulary Alignment in Openly Specified Interactions | [
"cs.MA"
] | The problem of achieving common understanding between agents that use different vocabularies has been mainly addressed by designing techniques that explicitly negotiate mappings between their vocabularies, requiring agents to share a meta-language. In this paper we consider the case of agents that use different vocabularies and have no meta-language in common, but share the knowledge of how to perform a task, given by the specification of an interaction protocol. For this situation, we present a framework that lets agents learn a vocabulary alignment from the experience of interacting. Unlike previous work in this direction, we use open protocols that constrain possible actions instead of defining procedures, making our approach more general. We present two techniques that can be used either to learn an alignment from scratch or to repair an existent one, and we evaluate experimentally their performance. | cs.MA | cs |
Vocabulary Alignment in Openly Specified Interactions
Paula Chocron1,2 and Marco Schorlemmer1
1Artificial Intelligence Research Institute (IIIA-CSIC), Bellaterra, Catalonia, Spain
2Universitat Aut`onoma de Barcelona, Bellaterra, Catalonia, Spain
Abstract
The problem of achieving common understanding be-
tween agents that use different vocabularies has been
mainly addressed by designing techniques that explic-
itly negotiate mappings between their vocabularies,
requiring agents to share a meta-language. In this pa-
per we consider the case of agents that use different
vocabularies and have no meta-language in common,
but share the knowledge of how to perform a task,
given by the specification of an interaction protocol.
For this situation, we present a framework that lets
agents learn a vocabulary alignment from the expe-
rience of interacting. Unlike previous work in this
direction, we use open protocols that constrain pos-
sible actions instead of defining procedures, making
our approach more general. We present two tech-
niques that can be used either to learn an alignment
from scratch or to repair an existent one, and we
evaluate experimentally their performance.
1
Introduction
Addressing the problem of vocabulary heterogeneity
is necessary for the common understanding of agents
that use different languages, and therefore crucial for
the success of multi-agent systems that act jointly by
communicating. This problem has been tackled sev-
eral times in the past two decades, in general from
two different perspectives. Some approaches [20, 8]
consider the existence of external contextual elements
that all agents perceive in common, and explore how
those can be used to explain the meaning of words.
C: beer
W: to drink?
C: wine
1
C: birra
W: da bere?
C: vino
1
0
0
2
3
2
3
W: color?
C: white
7
5
5
C: red
6
C: bianco
7
C: rosso
6
W: tipo?
Figure 1: Protocols for Waiter (W) and Customer
(C)
A second group of techniques [15, 18] (and also [8],
where both approaches are combined) consider the
situation, reasonable in agents that communicate re-
motely, in which this kind of context is not available.
They do so by providing explicit ways of learning or
agreeing on a common vocabulary (or alignment be-
tween vocabularies). These techniques require agents
to share a common meta-language that they can use
to discuss about the meaning of words and their align-
ments. The complex question of how to communi-
cate with heterogeneous interlocutors when neither
a physical context nor a meta-language are available
remains practically unexplored.
The work by Atencia and Schorlemmer [1] ap-
proaches this situation by considering a different kind
of context, given by the interactions in which agents
engage. Agents are assumed to share the knowl-
edge of how to perform a task, or, more concretely,
the specification of an interaction, given by a finite
state automaton. For example, they consider the in-
1
teraction in Figure 1, in which an English speaking
customer and an Italian waiter communicate to or-
der drinks. The authors show how agents can pro-
gressively learn which mappings lead to successful
interactions from the experience of performing the
task. After several interactions, agents converge to
an alignment that they can use to always succeed at
ordering and delivering drinks with that particular
interlocutor. The interesting aspect of this idea is
that the only shared element, namely the interaction
specification, is already necessary to communicate.
However, using finite state automata as specifications
implies that agents need to agree on the exact order in
which messages should be sent, which is unnecessar-
ily restrictive. In addition, agents learn an alignment
that is only useful for one task, and how it can be
extrapolated to further interactions is not explored.
In this paper we present a general version of the
interaction as context approach to vocabulary align-
ment in multi-agent systems. Instead of automata,
we consider agents that specify interactions with
constrained-based open protocols that define rules
about what can be said instead of forcing one par-
ticular execution. In particular, we use ConDec pro-
tocols [14] that use linear temporal logic constraints.
One ConDec protocol may not define completely the
meaning of all words in a vocabulary, so we study
how agents can learn mappings from performing dif-
ferent tasks. For this reason, the learning process is
substantially different to the one in [1], where a map-
ping that works in one task is considered correct. If
enough experiences are available, agents that use the
techniques we present converge to an alignment that
is useful in the general case, even for future interac-
tions that they do not yet know. Since learning is
done gradually as different interaction opportunities
appear, agents can use what they learned early, even
if they still do not know the alignment completely.
After presenting open interaction protocols, we de-
fine a framework for interacting with partners that
use different vocabularies. We later present two dif-
ferent techniques to learn an alignment from the
experience of interacting. The first relies only on
analysing if a message is allowed or not in a particular
moment, and it can be used in any constraint-based
protocol. The second technique uses the semantics
2
of the protocols we present to improve the perfor-
mance of the first one. Both methods can be used
to learn alignments from scratch when there is no in-
formation, as well as to repair alignments obtained
with other methods. We evaluate experimentally the
techniques when used for both purposes, and show
how different factors affect their performance.
2 Open Interaction Protocols
The question of what should be expressed by an in-
teraction protocol has been extensively discussed in
the multi-agent systems community. Traditional ap-
proaches such as finite state automata and Petri Nets
are simple to design and read, but also too rigid.
More flexible alternatives constrain possible actions
instead of defining a fixed procedure. Constraint
Declarative Protocols (commonly known as ConDec
protocols) are an example of these approaches. Con-
Dec protocols were first proposed by Pesic and van
der Aalst [14] as a language to describe business pro-
tocols, and then used as specifications for agent inter-
actions by Montali [13], who presented an extension
of ConDec and tools for its verification, and by Bal-
doni et al. [2, 3], who integrated ConDec constrains
and commitment protocols, a framework to specify
interactions with social semantics first proposed by
Singh [19]. An important advantage of ConDec pro-
tocols is that they use linear temporal logic, a well-
known logic for which many tools are available.
d
Linear temporal logic (LTL from now on) is a nat-
ural choice to express constraints about actions that
occur in time. The syntax of LTL includes the one of
propositional logic, and the additional temporal op-
erators {(cid:3), ♦,
, U}, that can be applied to any LTL
formula. LTL formulae are interpreted over paths in
Kripke structures, which are sequences of states asso-
ciated to a truth-valuation of propositional variables.
Temporal operators are interpreted in these paths as
follows: (cid:3)p means that p must be true in the truth-
valuation of all states, ♦p means that p must be true
eventually,
p means that p must be true in the next
state, and p U q means that p must be true until q is
valid. A set of LTL formulae, called a theory, is satis-
fiable if there exists a path for which all the formulae
d
are true. In that case, the path is a model of the the-
ory. The satisfiability problem in LTL is decidable,
as well as the model checking problem, which consists
in deciding if a given path is a model of a theory.
A ConDec protocol is a tuple (cid:104)M, C(cid:105), where M
is a set of action names and C is a set of constraints
about how actions can be performed. Constraints are
LTL sentences renamed conveniently. We use a min-
imal version1 of the constraints introduced originally
by Pesic and van der Aalst that we show in Table 1,
where n ∈ N and a, b ∈ M . These constraints are
generally divided into three classes. Existential con-
straints (existence and absence) predicate over the
amount of times some action can be performed, rela-
tional constraints describe binary relations between
two actions, and negation constraints (preceded by
a '!' sign) are relations that do not hold. Given a
set M of actions, Cons(M ) is the set of all possible
constraints over M . In a protocol (cid:104)M, C(cid:105), necessarily
C ⊆ Cons(M ).
absence(0, a)
absence(n, a)
correlation(a, b)
!correlation(a, b)
response(a, b)
!response(a, b)
bef ore(a, b)
!bef ore(a, b)
premise(a, b)
!premise(a, b)
imm af ter(a, b)
!imm af ter(a, b)
Constraint
existence(n + 1, a) ♦(a ∧ d
existence(1, a)
LTL meaning
♦a
existence(n, a))
¬existence(1, a)
¬existence(n + 1, a)
♦a =⇒ ♦b
♦a =⇒ ¬♦b
(cid:3)(a =⇒ ♦b)
(cid:3)(a =⇒ ¬♦b)
¬bU a
(cid:3)(♦b =⇒ ¬a)
b =⇒ a)
(cid:3)(
b =⇒ ¬a)
(cid:3)(
b)
d
d
(cid:3)(a =⇒ d
(cid:3)(a =⇒ d¬b)
2.1 Open Protocols as
Interaction
Table 1: LTL definitions of constraints
Protocols
In the rest of this section we present the technical
notions that are necessary to use ConDec protocols
as specifications of interactions between agents that
may use different vocabularies. We first define in-
teraction protocols, that constrain the way in which
agents can utter messages. We introduce the notion
of bound to capture the knowledge of agents about
how many steps they have to finish the interaction.
If this knowledge is not available it can be omitted
and replaced for the fact that the interaction must
finish in finite steps.
Definition 1 Given a set A of agent IDs and a
propositional vocabulary V , an interaction protocol is
a tuple P = (cid:104)M, C, b(cid:105), where the set of actions M =
A × V is a set of messages formed by the ID of the
sender agent and the term it utters, C ⊆ Cons(M ),
and b ∈ N is the bound.
1 Since we are not interested in the usability of the protocols
here, we do not include some constraints that work as syntactic
sugar, such as exactly(n, a), that can be replaced by including
existence(n, a) and absence(n, a).
The semantics of an interaction protocol is defined
over interactions that represent a sequence of uttered
messages.
Definition 2 Given a set of messages M , an in-
teraction i ∈ M∗ is a finite sequence of messages
m ∈ M . The length of an interaction (len(i)) and
the append operation (i . m) are defined in the same
way as for sequences.
An interaction i can be encoded into a Kripke
structure path by including one propositional vari-
able for each message m ∈ M and considering the
following sequence of states with truth-valuations. In
states with index 0 ≤ j < len(i), let the proposi-
tional variable for the j-th message in i be true, and
all other propositional variables be false. For states
after len(i), let all propositional variables be false.
With this construction, we can define the semantics
of interaction protocols.
Definition 3 An interaction i is a model of an in-
teraction protocol P = (cid:104)M, C, b(cid:105) (noted i = P) if
3
len(i) ≤ bound, and the Kripke path that encodes i
is a model of C. An interaction i(cid:48) is a partial model
(noted i(cid:48) =p P) of P if it is a prefix of a model of P.
Definition 3 implies that checking satisfiability of
an interaction protocol is equivalent to checking LTL
satisfiability.
As already mentioned, we are interested in agents
that use different vocabularies, but share the knowl-
edge of how to perform a task.
In the rest of this
section we define more precisely what that means.
From now on, V1 and V2 are two possibly different
vocabularies. Let us start by defining the notion of
vocabulary alignment.
Definition 4 An alignment is a function between
vocabularies α : V2 → V1. Given a set of agents
A, α can be extended homomorphically to a function
between:
- messages M1 = A × V1 and M2 = A × V2 (α :
M2 → M1)
- constraints (α : Cons(M2) → Cons(M1)) and
sets of constraints (α : 2Cons(M2) → 2Cons(M1))
- interactions (α : M∗
1 ) and sets of inter-
2 → M∗
actions (α : 2M∗
2 → 2M∗
1 )
To capture the idea of sharing the knowledge of how
to perform a task, we define notion of compatibility
between protocols, which consists simply on having
the same models modulo an alignment. Given a pro-
tocol P = (cid:104)M, C, b(cid:105), we define the set of its models
as Int(P) = {i ∈ M∗ such that i = P}.
Definition 5 Two interaction protocols P1 =
(cid:104)M1, C1, b(cid:105) and P2 = (cid:104)M2, C2, b(cid:105) with sets of mes-
sages M1 = A× V1 and M2 = A× V2 are compatible
if there exists an alignment α : V2 → V1 such that
Int(P1) = α(Int(P2))
If we know the alignment α for which the condition
holds, we can say they are compatible under α.
An Example Let us extend the example of Figure
1 by considering again a waiter W and a customer
C and the ordering drinks situation. The vocabulary
of the customer is VC = {to drink, beer, wine, water,
size, pint, half pint}, while the Waiter uses VW = {da
bere, birra, vino, acqua, tipo, media, piccola}. Con-
sider the bijective alignment α : VW → VC obtained
by mapping each word in VW with the word in the
same position in VC. We consider the following pro-
tocols to specify the ordering drinks interactions:
PW = (cid:104){W, C} × VW ,
{existence((cid:104)W, da bere(cid:105), 1, 1),
premise((cid:104)C, birra(cid:105),(cid:104)W, da bere(cid:105)),
premise((cid:104)C, vino(cid:105),(cid:104)W, da bere(cid:105)),
premise((cid:104)C, acqua(cid:105),(cid:104)W, da bere(cid:105)),
!correlation((cid:104)C, birra(cid:105),(cid:104)W, vino(cid:105)),
response((cid:104)C, birra(cid:105),(cid:104)W, tipo(cid:105)),
premise((cid:104)C, piccola(cid:105),(cid:104)W, tipo(cid:105)),
premise((cid:104)C, media(cid:105), tipo(cid:105))}, 5(cid:105)
PC = (cid:104){W, C} × VC,
{existence((cid:104)W, to drink(cid:105), 1, 1),
premise((cid:104)C,
beer(cid:105),(cid:104)W, to drink(cid:105)),
premise((cid:104)C, water(cid:105),(cid:104)W, to drink(cid:105)),
premise((cid:104)C, wine(cid:105),(cid:104)W, to drink),
response((cid:104)C, beer(cid:105),(cid:104)W, size(cid:105)),
premise((cid:104)C, halfpint(cid:105),(cid:104)W, size(cid:105)),
premise((cid:104)C, pint(cid:105),(cid:104)W, size(cid:105))}, 5(cid:105)
The two protocols above are not compatible under
any bijective alignment, since the customer protocol
has as model an interaction in which the customer or-
ders wine and beer, while the waiter protocol has as
models interactions in which the waiter only accepts
one beverage. If !correlation(beer, wine) is added to
PC, the resulting protocols are compatible, in par-
ticular under α.
3 Communicating with Hetero-
geneous Partners
In this paper, for simplicity, we will restrict to us-
ing only bijective alignments over vocabularies of the
same size.
We focus on interactions between two agents a1 and
a2 with vocabularies V1 and V2 respectively. During
an interaction, an agent can send messages composed
4
of words in its vocabulary and receive others from
its interlocutor, or finish the communication if cer-
tain conditions hold. We assume messages are never
lost and always arrive in order, and more strongly,
that each message is received before the following
one is uttered. This requires that agents agree in
who speaks at each time, although we do not force
them to follow any particular turn-taking pattern.
Agents interact to perform some task together,
that each of them specifies with an interaction proto-
col. We assume a1 and a2 agree on a set of tasks that
they can perform and on how they are performed:
their respective protocols for each task are compat-
ible. Our agents use their vocabulary consistently
throughout different tasks, i.e., the protocols for all
tasks are compatible under the same alignment.
From now on, we will use sets of messages M1 =
{a1, a2} × V1 and M2 = {a1, a2} × V2. We assume
there exists a bijective alignment α : V2 → V1 such
that, for each task that a1 and a2 can perform, they
have protocols P1 : (cid:104)M1, C1, b(cid:105) and P2 : (cid:104)M2, C2, b(cid:105)
respectively, and P1, P2 are compatible under α. We
assume that whenever agents interact, they use pro-
tocols that correspond to the same task.
We present a general approach to learn the align-
ment α from the experience of interacting to perform
different tasks sequentially. The methods we present
are used by one agent alone, and do not require that
its interlocutor uses them as well. To explain the
techniques, we adopt the perspective of agent a1.
Briefly, we propose to learn α by taking into ac-
count the coherence of messages. When a1 receives
a message, it learns from analysing which interpreta-
tions are allowed by the protocol and which are not.
This information, however, is not definitive. An al-
lowed word can be an incorrect interpretation for a
received message, because protocols do not restrict
completely all possible meanings. Moreover, a for-
bidden message can still be a correct interpretation.
Since some rules express relations between messages,
a previously misinterpreted word (by the agent or by
its interlocutor) can make a correct mapping be im-
possible.
foreign words, and update their values with the ex-
perience of interacting. Formally, for each v2 ∈ V2
that a1 knows about (because it received it in a mes-
sage at some point) and for each v1 ∈ V1, agent a1
has a weight ω(v2, v1) that represents its confidence
in that α(v2) = v1. Using the bijectivity, we will keep
ω(v2, v) = 1, but we do not require the same
in the other direction since the foreign vocabulary is
unknown a priori.
(cid:80)
v∈V1
The techniques we propose can incorporate exist-
ing alignments, in the same way as it is done in [5] for
the case when protocols are automata. These align-
ments represent a priori knowledge about the foreign
vocabulary, that can have been obtained from previ-
ous experience, from a matching tool, or with other
techniques such as a syntactic similarity measure.
Since these techniques are never fully correct, pre-
vious alignments should not be trusted completely,
and the techniques we propose can work as methods
for repairing them. A previous alignment for agent
a1 is a partial function A : V (cid:48)
2 is
a set of terms. Note that A is not necessarily defined
over V2, since the vocabulary that a2 uses may be un-
known a priori. However, the information it provides
can be used even if V (cid:48)
2 is a subset of V2 or they over-
lap. We interpret the confidences always positively,
meaning that a mapping with low confidence has still
more confidence than one that does not exist. If a1
has a previous alignment A, it initializes the inter-
pretation distribution using that information. They
first assign raw values:
2 × V1 → N, where V (cid:48)
if (v2, v1) ∈ dom(A)
otherwise
A(v2, v1)
0
ω(v2, v1) =
Then they normalize the values using an exponen-
tial method such as softmax. This is necessary to
start the technique with the values for all mappings in
the interval (0, 1), since the alignment can be wrong.
If there is no previous alignment, they always initial-
ize the interpretation values with an uniform distri-
bution: ω(v2, v1) =
1
V1 for all v1 ∈ V1.
To handle this uncertainty, we propose a simple
probabilistic learning technique. Agents maintain an
interpretation distribution over possible meanings for
We explain in detail the learning techniques that
we propose in Section 4. In the rest of this section we
focus on the dynamics of the interaction: how agents
5
choose which messages to say, finish the interaction,
and decide interpretations for the messages they re-
ceive.
Choosing messages to send The choice of mes-
sages to utter is internal of each agent and depends on
its interests while interacting. We do not impose any
restriction on this, besides respecting the constraints
in the protocol. Formally, a1 can utter a message v1
only if (cid:104)a1, v1(cid:105) is a possible message as defined below.
Definition 6 Consider an agent a1 following proto-
col P1 = (cid:104)M1, C1, b(cid:105) and suppose interaction i has
happened so far (and i =p P1). The possible mes-
sages at that point are all messages m ∈ M1 such
that when they are performed the interaction remains
a partial model, that is i . m =p P1.
Finishing the interaction An interaction can fin-
ish in three situations. First, if there is a bound,
reaching it implies the automatic end of the conver-
sation. The interaction also finishes if an agent re-
ceives a message that has no possible interpretation.
An agent can also finish the conversation whenever it
wants if it considers that the interaction is successful,
i.e., if it is a model of the protocol when it ends. In
this case, it simply stops talking, producing a time-
out that let the other agent realize the interaction
finished.
Choosing interpretations When agent a1 re-
ceives a message v2 ∈ V2 from a2, it needs to in-
terpret it in V1 to be able to continue the interaction.
To this aim, agents use the information in the in-
terpretation distribution. Since agents assume there
exists a bijective alignment, they always choose the
same interpretation for a word during one interaction
and they do not choose words that have been already
chosen as interpretations.
Consider agent a1 receives word v2 after interaction
i. Let µ : V2 → V1 be the mappings made function,
where µ(v2) = v1 if and only if v1 was chosen as a
mapping for v2 before in the same interaction. The
domain dom(µ) are the v2 ∈ V2 that a1 received since
the current task started, and its image img(µ) are the
v1 ∈ V1 that were chosen as mappings. The set W of
possible interpretations for v2 is the set of words v1
in V1 such that (cid:104)a2, v1(cid:105) is a possible message after i,
and such that v1 (cid:54)∈ img(µ).
If rnd(S) is a function that chooses an element
randomly in the set S, the agent will choose an in-
terpretation as follows:
µ(v2)
µ(v2) =
if v2 ∈ dom(µ) ∧ µ(v2) ∈ W
if v2 (cid:54)∈ dom(µ)
rnd(argmaxv1∈W ω(v2, v1))
If either v2 ∈ dom(µ) but µ(v2) (cid:54)∈ W , or W = ∅,
the interaction is in one of the failure cases, and it
finishes because a1 stops talking.
4 Learning Alignments
from
Interactions
To learn an alignment from the experience of inter-
acting, agents make the following well behaviour as-
sumptions about their interlocutor. Essentially, these
assumptions imply that the dynamics of the interac-
tion described in the previous section are respected.
1. Compliance: An agent will not utter a message
that violates the constraints, or that makes it
impossible to finish successfully the interaction
in the steps determined by the bound (or in finite
steps if there is no bound).
2. Non Abandonment: An agent will not fin-
ish intentionally the conversation unless the con-
straints are fulfilled.
Agents learn when they decide how to interpret a
received word. The overall method is simple:
if a1
receives v2, it updates the value of ω(v2, v1) for all
the words v1 in a set U ⊆ V1. The set U , unlike
the set of possible interpretations W used to decide
the mapping, can have interpretations that are not
possible messages in that particular moment, and the
values in the interpretation distribution are updated
according to that.
A first decision is how to choose the set U . Since
checking satisfiability is computationally expensive,
6
this can impact considerably the overall performance
of the methods. A first option is to update the
value of all possible words (U = V1), which can be
slow for large vocabularies. Another option is to
use only the options that are considered until an in-
terpretation is chosen, or all those words that have
more value than the first possible interpretation. In
this case, if v1 is chosen as an interpretation for v2,
U = {v ∈ V1 such that ω(v2, v) > ω(v2, v1)}. A third
possibility is a midpoint between these two, in which
U = {v ∈ V1 such that ω(v2, v) ≥ ω(v2, v1)}. This
option updates also the words that have the same
value as the chosen interpretation.
If the distribu-
tion is uniform when the learning starts, this updates
many possible interpretations in the beginning and
fewer when there is more information. This is the
approach we choose, and the one used for the evalu-
ation.
We present two methods to update the values in
the interpretation distribution. The first one is a gen-
eral technique that only takes into account whether
messages are possible. This method is not exclusively
designed for protocols that use LTL, and can be used
for constraints specified in any logic for which agents
have a satisfiability procedure. In a second method
we show how the particular semantics of the protocols
we introduced can be taken into account to improve
the learning process.
4.1 Simple Strategy
The first approach consists in updating the value of
interpretations for a received message, according to
whether it is coherent or not with the protocol and
the performed interaction. When a new message v2
is received, the values for all interpretations are ini-
tialized as described in the last section. We use a
simple learning method in which updates are a pun-
ishment or reward that is proportional to the value
to update. We chose this method because it incorpo-
rates naturally previous alignments, it is very simple
to formulate, and the proportional part updated can
be manipulated easily to consider different situations,
something that will be relevant for the second tech-
nique. However, it is not necessarily the most efficient
choice, and other methods, such as one that records a
history, can converge faster. Consider a reward and a
punishment parameter rr, rp ∈ [0, 1]. We use the no-
tion of possible messages in Definition 6 and update
the values for v1 ∈ U as follows:
ω(v2, v1) + rr · ω(v2, v1)
ω(v2, v1) − rp · ω(v2, v1)
if (cid:104)a2, v1(cid:105) is possible
if (cid:104)a2, v1(cid:105) otherwise
ω(v2, v1) :=
(cid:80)
After all the updates for a given message are
made, the values are normalised in such a way that
ω(v2, v) = 1. Either simple sum-based or soft-
v∈V1
max normalization can be used for this.
4.2 Reasoning Strategy
The simple strategy does not make use of some eas-
ily available information regarding which constraints
were violated when a message is considered impos-
sible. To illustrate the importance of this infor-
mation, consider an interaction in which the Cus-
tomer said water but the Waiter interpreted it as
vino.
If the Customer says beer, the Waiter may
think that interpreting it as birra is impossible,
since ordering two alcoholic beverages is not allowed
(!correlation(birra, vino) would be violated). How-
ever, this is only due to misunderstanding water in
the first place.
In the reasoning approach, agents
make use of the semantics of violated rules to perform
a more fine-grained update of the values. Before pre-
senting the method, we need to divide constraints in
two kinds, that will determine when agents can learn
from them.
Definition 7 A constraint c is semantically non-
monotonic if there exist interactions i and i(cid:48) such that
i is a prefix of i(cid:48), and i = c but i(cid:48) (cid:54)= c. A constraint c
is semantically monotonic if this cannot happen, and
i = c implies that also all its extensions are models
of c.
The following proposition divides the constraints
in our protocols between monotonic and non-
monotonic.
Proposition 1 The
existence,
coexistence and response are semantically mono-
tonic. All the rest of the constraints defined are
semantically non-monotonic.
constraints
7
Violated Constraint
-
absence((cid:104)a2, v1(cid:105), n)
!correlation((cid:104)a2, v1(cid:105),(cid:104)a2, v(cid:48)
1(cid:105))
!response((cid:104)a2, v1(cid:105),(cid:104)a2, v(cid:48)
1(cid:105))
!bef ore((cid:104)a2, v1(cid:105),(cid:104)a2, v(cid:48)
1(cid:105))
!premise((cid:104)a2, v1(cid:105),(cid:104)a2, v(cid:48)
1(cid:105))
!imm af ter((cid:104)a2, v1(cid:105),(cid:104)a2, v(cid:48)
1(cid:105))
premise((cid:104)a2, v1(cid:105),(cid:104)a2, v(cid:48)
1(cid:105))
1(cid:105),(cid:104)a2, v1(cid:105))
imm af ter((cid:104)a2, v(cid:48)
bef ore(v(cid:48), v)
relation((cid:104)a1, v1(cid:105),(cid:104)a2, v(cid:48)
1(cid:105))
Action
ω(v2, v1) := ω(v2, v1) − rp · ω(v2, v1)
ω(v2, v1) := 0
ω(v2, v1) := ω(v2, v1) − rp · ω(µ(v(cid:48)
ω(µ(v(cid:48)
1), v(cid:48)
1)
1) − rp · ω(v2, v1)
1) := ω(µ(v(cid:48)
1), v(cid:48)
1), v(cid:48)
ω(v2, v1) := ω(v2, v1) − rp · ω(µ(v(cid:48)
ω(µ(v(cid:48)
1), v(cid:48)
1)
1) − rp · ω(v2, v1)
1) := ω(µ(v(cid:48)
1), v(cid:48)
1), v(cid:48)
ω(v2, v1) := ω(v2, v1) − rp · ω(v2, v1)
1) − rp · ω(v2, v1)
ω(v(cid:48)
1) := ω(v(cid:48)
1, v(cid:48)
1, v(cid:48)
Table 2: Updates for each violated constraint when a message is received
Proof 1 To prove non-monotonicity, it is enough to
show an example that violates the constraint. For
example, consider bef ore(m, m(cid:48)) and suppose an in-
teraction I such that m (cid:54)∈ i and m(cid:48) (cid:54)∈ i. Then i =
(cid:54)= bef ore(m, m(cid:48)). Mono-
bef ore(m, m(cid:48)) and i . m(cid:48)
tonicity can be proven by counterexample, showing
that for a constraint c, if i . m(cid:48) (cid:54)= c then necessarily
i (cid:54)= c. For example, if i . m violates existence(m(cid:48), n),
there are two options, either m = m(cid:48) or m (cid:54)= m(cid:48).
Let #(m, i) and #(m, i . m(cid:48)) be the number of occur-
rences of c in i and i . m(cid:48) respective. In both cases,
#(m, i . m(cid:48)) ≥ #(m, i), so if #(m, i . m(cid:48)) ≤ n, then
necessarily #(m, i) ≤ n.
Non-monotonic constraints can be used to learn
while interacting, using the principle of compliance
(if a constraint is violated, something must be wrong
in the interpretations). Monotonic constraints could
be used when the interaction ends, using the prin-
ciple of non-abandonment (if not all constraints are
satisfied, there must be an error). However, since our
agents cannot communicate in any way, they ignore
why the interaction ended, and therefore they do not
know if their interlocutor considers the constraints
were satisfied or not, making it very difficult to de-
cide how to update values. In this work we focus on
handling non-monotonic constraints, leaving mono-
tonic ones for future work where an ending signal is
introduced. To start, let us define formally when a
non-monotonic constraint is violated.
Definition 8 Given an interaction i and an inter-
pretation v1 for a received message, a constraint c is
violated if i = c but i .(cid:104)a2, v1(cid:105) (cid:54)= c.
If agent a1 decides that v1 is not a possible interpre-
tation for v2, let V iol be all c ∈ C1 that are violated
by v1. It is possible that V iol = ∅. This can happen
when there are no violated constraints because a sub-
set of constraints becomes unsatisfiable, for example,
if the waiter has not said size and a protocol includes
the rules {!response((cid:104)C, wine(cid:105),(cid:104)W, size(cid:105)),
existence((cid:104)W, size(cid:105), 1)},
the customer cannot say
wine even when V iol would be empty. Another situ-
ation in which a message can be impossible without
any constraint being violated is when it makes the in-
teraction impossible to finish successfully in the num-
ber of utterances indicated by the bound, or in finite
time if there is no bound.
To explain the update in the reasoning strategy,
consider again reward and a punishment parameter
rr, rp ∈ [0, 1]. If an interpretation v1 is possible for a
8
received message v2, then ω(v2, v1) := ω(v2, v1) + rr ·
ω(v2, v1). If it is not, it is updated as indicated in to
Table 2, according to which rule was violated. These
actions are performed iteratively for each broken rule.
After the updates, all values are normalized to obtain
In this case, we need to use
a method that maintains the values that are 0, so
softmax is not a possibility.
ω(v2, v) = 1.
(cid:80)
v∈V1
Let us explain the motivation under each update.
Violating the existential non-monotonic constraint is
different to violating a relation constraint, because it
expresses a condition over only one action, so if the
chosen interpretation violates it, it must be wrong.
For this reason, the value of the mapping is set to 0.
Negation constraints express a relation over two
actions, and if their are broken, necessarily both of
them happened in the interaction, and if a constraint
is violated, it could be because either of them was
misinterpreted. To take this into account, we pro-
pose an update inspired in the Jeffrey's rule of con-
ditioning [17], that modifies the Bayes Rule of condi-
tional probability for the cases when evidence is un-
certain. If Q is a probability distribution for a parti-
tion E1 . . . En of possible evidence, then Jeffrey's rule
states that the posterior probability of A is computed
as:
Q(A) =
P (AEi) · Q(Ei)
n(cid:88)
i=0
Back to our problem, let v2 be a received message
such that µ(v2) = v1. Suppose a1 receives a new
message v(cid:48)
2, and it discovers that mapping it with
v(cid:48)
1 violates a constraint relation(v1, v(cid:48)
1). This means
that P (α(v2) = v1 ∧ α(v(cid:48)
2) = v(cid:48)
1) = 0, and therefore
P (α(v2) = v1α(v(cid:48)
2) = v(cid:48)
1) = 0. Then, if P os is the
set of possible mappings, we can write:
ω(v2, v1) := ω(v2, v1) · (cid:88)
ω(α(v(cid:48)
2, v(cid:48)
1)
1∈P os
v(cid:48)
or, what is equivalent, if Impos is the set of incom-
ω(v2, v1) = 1:
patible mappings, since(cid:80)
v∈v1
ω(v2, v1) := ω(v2, v1) − (cid:88)
1∈Impos
v(cid:48)
ω(v(cid:48)
2, v(cid:48)
1)
9
We implement this by subtracting for each map-
ping a value that is proportional to the confidence
in the other mapping. Since agents do not explore
all the possible interpretations, the value is not com-
puted exactly, but approximated considering only the
interpretations that have more confidence.
When the interpretation depends on the interpre-
tation of many other mappings, there are too many
possibilities that difficult reasoning about which one
can be wrong. In this case, the agents use a default
punishment rp. This occurs when the positive ver-
sions of premise and imm af ter are violated, when
V iol = ∅, and when and a violated constraint de-
pends on a message that a1 sent, since they have
no information about how their messages were inter-
preted by v2.
Lastly, let us discuss the value of rp. Choosing
rp = 1V1 , the technique has the effect of subtracting
a larger value when the agent is confident in the cor-
rectness of the previous interpretation. This is the
value that we use.
5 Experimental Evaluation
In this section we present the results of experiments
that show how the methods we propose perform ex-
perimentally when used by agents with different vo-
cabularies. Before presenting the results, let us dis-
cuss the generation of data for experimentation.
Due to the lack of existing datasets, we performed
all the experimentation with randomly generated
data. We created protocols with different sizes of
vocabulary and of set of constraints. To generate a
protocol from a vocabulary, we randomly chose con-
straints and added them to the protocol if it remained
satisfiable with the new rule. In the implementation,
we express constraints over words instead of over mes-
sages, so a constraint are valid for any combination of
senders. We used bounds in our experimentation to
limit the possible constraints and to eventually finish
the interactions between agents, that in all cases had
the value of the vocabulary size. We created pairs
of compatible protocols P1, P2, with vocabularies V1
and V2 respectively by building a bijective alignment
α : V2 → V1 and then using it to translate the con-
Figure 2: Results for a vocabulary of size 10
straints in a protocol P2, obtaining P1. We used the
NuSMV model checker [6] to perform all the neces-
sary satisfiability checks.
To evaluate how agents learn α from the experi-
ence of interacting, we first need to define a mea-
sure of how well an agent knows an alignment. Since
agents always choose the possible mapping with high-
est weight, we can easily extract from the interpre-
tation distribution an alignment that represents the
first interpretation choice for each foreign word in the
following interaction. This alignment, that we call
αi for agent ai, is one of the mappings with highest
weight for each foreign word. Formally, for a1, the do-
main of α1 is v2 ∈ V2 such that ω(v2, v) is defined for
all v ∈ V1, and α1(v2) = rnd(argmaxv∈V1 ω(v2, v)).
Note that there are multiple possibilities with the
same value, α1(v2) takes a random word between
them.
To compare this alignment with α, we used the
standard precision and recall measures, and their har-
monic mean combination, commonly known as F-
score. To use the standard definitions directly, we
only need to consider alignments as relations instead
of functions: given an alignment α : V2 → V1, the
corresponding relation is the set of pairs (cid:104)v2, v1(cid:105) such
that α(v2) = v1.
Definition 9 Given two alignments β and γ ex-
pressed as relations, the precision of β with respect
to γ is the fraction of the mappings in β that are also
in γ:
precision(β, γ) =
β ∩ γ
β
while its recall is the fraction of the mappings in γ
that were found by β:
recall (β, γ) =
β ∩ γ
γ
Given an alignment β and a reference alignment γ,
the F-score of the alignment is computed as follows:
F − score(β, γ) = 2 · precision(β, γ) · recall (β, γ)
precision(β, γ) + recall (β, γ)
We performed experiments parametrized with a
protocol and vocabulary size.
In all experiments,
agents are sequentially given pairs of compatible pro-
tocols, and after each interaction we measure the F-
score of their alignments with respect to α. The same
protocols can appear repeatedly in the sequence, but
10
we consider that eventually new ones appear always
in the interaction, which implies that at some point
the meaning of all words is fixed. Our agents also
used the NuSMV model checker for both checking
satisfiability and finding violated constraints. We fist
intended to perform experiments with the same vo-
cabulary sizes that are used for testing in [1], which
are 5, 10, 15, 40, 80. However, the interactions for size
80 were too slow to be able to perform a reasonable
amount of repetitions for each technique. This prob-
lem is intrinsic to the protocols we use (since agents
have to decide if the messages they want to send are
possible), and should be taken into account in fu-
ture work. Since varying the vocabulary size did not
provide particularly interesting results, we show here
the experiments for a vocabulary of 10 words and
four different protocol sizes. We used punishment
and rewards parameters of rp, rr = 0.3 for the simple
strategy, which were best in a preliminary test. Each
experiment was repeated 10 times.
In a first experiment, we let agents interact 200
times and measured the way they learned the align-
ment for different protocol sizes, shown in Figure 2.
The F-score if computed by averaging the F-score of
each agent that participates in the interaction. The
curves show that the smart strategy is always bet-
ter than the simple one, but this difference is more
dramatic when the protocols are smaller. This is be-
cause there is less information, so using it intelligently
makes a great difference. The curves are sharp in the
beginning, when agents do not have much informa-
tion about α, and they make many mistakes from
which they learn fast. When agents reach a reason-
able level of precision and recall, mistakes are less
frequent, and therefore the pace of the learning slows
considerably. Although this affects the convergence
to an F-score of 1.0, it also means that, after a cer-
tain number of interactions, agents have learn enough
to communicate successfully most of the times. The
sparse mistakes that they will make in future conver-
sations make them learn more slowly the remaining
mappings. Table 3 shows how fast agents reach a rea-
sonable level of F-score, that we chose of 0.8 based
on [10].
In a second experiment, we studied the perfor-
mance of our methods when used to repair existing
6
-
8
10
187
101
simple
smart
139
87
57
12
47
33
Table 3: 0.8 convergence
alignments. To this aim, we created alignments with
different values of precision and recall with respect to
α, that we divided into three categories. Low qual-
ity alignments had precision and recall 0.2, medium
quality 0.5, and high quality 0.8. We only studied
alignments with the same value of precision and re-
call to reduce the number of combinations and have
a clear division of quality levels. We made agents
with this alignments interact, using the simple and
the smart technique. The results can be seen in Fig-
ure 3. The number next to each agent class in the
reference is the precision and recall of the alignment
that was given to that agent. Again, the reasoning
strategy performs particularly better when only lit-
tle information is available; in this case, the repair is
much better than when using the simple strategy for
alignments of low quality.
Figure 3: Results for different alignment qualities
11
6 Related Work
7 Conclusions
and
Future
As we already mentioned, a large portion of the
existent work that tackles the problem of vocabu-
lary alignment for multi-agent communication con-
sists on techniques to let agents discuss and agree
on a common vocabulary or alignment. Some of
these approaches use argumentation or negotiation
techniques [15, 12], performing an offline negotiation
that takes place before actual interactions can start.
Other approaches, such as the three-level schema de-
veloped by van Diggelen [8], discuss the meaning of
works on a by-demand basis, that is, only when it is
necessary to use them to perform some task.
Other approaches use different kinds of grounding
to learn alignments or the meaning of words from in-
teractions. The well established research by Steels
[20] considers a situation in which learners share a
physical environment, and have the ability of point-
ing to things to communicate if they are talking
about the same things. Goldman et al.
[11], inves-
tigated how agents can learn to communicate in a
way that maximises rewards in an environment that
can be modelled as a Markov Decision Process. In
[4], the authors study a version of the multiagent,
multiarmed bandit problem in which agents can com-
municate between each other with a common lan-
guage, but message interpretations are not know.
Closer to our approach is the work by Euzenat on
cultural alignment repair [9], and particularly the al-
ready mentioned work by Atencia and Schorlemmer
[1], where the authors consider agents communicat-
ing with interaction protocols represented with finite
state automata, and use a shared notion of task suc-
cess as grounding. Our work, instead, considers only
the coherence of the interaction as grounding. To
the best of our knowledge, the work presented in this
paper is the first approach in which agents learn lan-
guages dynamically from temporal interaction rules,
taking into account only the coherence of the utter-
ances.
A very well studied problem consists on learning
different structures from experience, such as gram-
mars [7] or norms [16]. Our approach can be seen as
the reverse of these approaches, where some structure
is considered shared and used to learn from.
Work
The techniques we propose allow agents to learn
alignments between their vocabularies only by inter-
acting, assuming that their protocols share the same
set of models, without requiring any common meta-
language or procedure. The assumption of sharing
the models can be removed, but the learning will be
slower because an extra component of uncertainty is
added. Although our methods converge slowly, they
use only very general information, and can be easily
combined with other methods that provide more in-
formation about possible mappings, as we show with
the integration of previous alignments. Our agents
learn dynamically while interacting and the tech-
niques achieve fast a reasonable level of knowledge
of the alignment. Therefore, a possible way of using
our techniques would be to go first through a training
phase in which agents learn many mappings fast, and
then start performing the real interactions, and keep
learning to discover the remaining mappings. The
technique that uses particular semantics of the proto-
cols, as expected, improves the learning. These tech-
niques can be particularly developed for each kind of
protocols. Agents do not need to use the same tech-
niques or even the same logic, as long as they share
the models.
There exist many possible directions of research
derived form this work. First, we could relax the
assumption that agents do not share any meta-
language, considering agents that can in some way
exchange information about the interaction. For ex-
ample, considering only that agents can communi-
cate whether they finished a task successfully would
make possible to reason about monotonic rules when
an interaction ends. An approach like this would re-
late our work to the one by Santos [15] about di-
alogues for meaning negotiation. Another possible
extension consists in considering agents that priori-
tize the language learning to performing the task. In
this case, agents would have to decide what utter-
ances it is more convenient to make in order to get
more information about the alignment. An aspect
that should be improved in future work is the per-
12
formance in terms of runtime per interaction, since it
was very slow for larger vocabularies.
In International Conference on Computer Aided
Verification, pages 359 -- 364. Springer, 2002.
References
[1] M. Atencia and M. Schorlemmer.
An
interaction-based approach to semantic align-
ment. Journal of Web Semantics, 12-13:131 -- 147,
2012.
[2] M. Baldoni, C. Baroglio, and E. Marengo.
Behavior-oriented commitment-based protocols.
In ECAI 2010 - 19th European Conference on
Artificial Intelligence, Lisbon, Portugal, August
16-20, 2010, Proceedings, pages 137 -- 142, 2010.
[3] M. Baldoni, C. Baroglio, E. Marengo, and
V. Patti. Constitutive and regulative specifi-
cations of commitment protocols: A decoupled
approach (extended abstract).
In Proceedings
of the Twenty-Fourth International Joint Con-
ference on Artificial Intelligence, IJCAI 2015,
Buenos Aires, Argentina, July 25-31, 2015,
pages 4143 -- 4147, 2015.
[4] S. Barrett, N. Agmon, N. Hazon, S. Kraus,
and P. Stone. Communicating with unknown
teammates.
In T. Schaub, G. Friedrich, and
B. O'Sullivan, editors, ECAI 2014 -- 21st Euro-
pean Conference on Artificial Intelligence, 18-22
August 2014, Prague, Czech Republic -- Includ-
ing Prestigious Applications of Intelligent Sys-
tems (PAIS 2014), volume 263 of Frontiers in
Artificial Intelligence and Applications, pages
45 -- 50. IOS Press, 2014.
[5] P. Chocron and M. Schorlemmer. Attuning on-
tology alignments to semantically heterogeneous
multi-agent interactions. In ECAI 2016 - 22nd
European Conference on Artificial Intelligence,
The Hague, The Netherlands, pages 871 -- 879,
2016.
[6] A. Cimatti, E. Clarke, E. Giunchiglia,
F. Giunchiglia, M. Pistore, M. Roveri, R. Se-
bastiani, and A. Tacchella. Nusmv 2: An
opensource tool for symbolic model checking.
[7] C. de la Higuera. Grammatical Inference: Learn-
ing Automata and Grammars. Cambridge Uni-
versity Press, New York, NY, USA, 2010.
[8] J. V. Diggelen, R. Beun, F. Dignum, R. M. V.
Eijk, and J. J. Meyer. Ontology negotiation:
Goals, requirements and implementation.
Int.
J. Agent-Oriented Softw. Eng., 1(1):63 -- 90, Apr.
2007.
[9] J. Euzenat. First Experiments in Cultural Align-
ment Repair.
In Semantic Web: ESWC 2014
Satellite Events, volume 8798, pages 115 -- 130,
2014.
[10] J. Euzenat, C. Meilicke, H. Stuckenschmidt,
P. Shvaiko, and C. Trojahn. Ontology align-
ment evaluation initiative: Six years of expe-
rience. Journal on Data Semantics XV, pages
158 -- 192, 2011.
[11] C. V. Goldman, M. Allen, and S. Zilberstein.
Learning to communicate in a decentralized en-
vironment. Autonomous Agents and Multi-Agent
Systems, 15(1):47 -- 90, Aug. 2007.
[12] L. Laera, I. Blacoe, V. A. M. Tamma, T. R.
Payne, J. Euzenat, and T. J. M. Bench-Capon.
Argumentation over ontology correspondences
in MAS.
In E. H. Durfee, M. Yokoo, M. N.
Huhns, and O. Shehory, editors, 6th Interna-
tional Joint Conference on Autonomous Agents
and Multiagent Systems (AAMAS 2007), Hon-
olulu, Hawaii, USA, May 14-18, 2007, page 228,
2007.
[13] M. Montali.
Specification and verification of
declarative open interaction models: A logic-
based approach, volume 56 of Lecture Notes
in Business Information Processing. Springer-
Verlag New York Inc, 2010.
[14] M. Pesic and W. M. P. van der Aalst. A declar-
ative approach for flexible business processes
13
In Proceedings of the Interna-
management.
tional Conference on Business Process Manage-
ment Workshops, BPM, pages 169 -- 180, Berlin,
Heidelberg, 2006. Springer-Verlag.
[15] G. Santos, V. Tamma, T. R. Payne, and
F. Grasso. A dialogue protocol to support mean-
ing negotiation. (extended abstract).
In Pro-
ceedings of the 2016 International Conference
on Autonomous Agents and Multiagent Systems,
AAMAS '16, pages 1367 -- 1368, Richland, SC,
2016. International Foundation for Autonomous
Agents and Multiagent Systems.
[16] S. Sen and S. Airiau. Emergence of norms
through social learning. In IJCAI, volume 1507,
page 1512, 2007.
[17] G. Shafer. Jeffrey's rule of conditioning. Philos-
ophy of Science, 48(3):337 -- 362, 1981.
[18] N. Silva, G. I. Ipp, and P. Maio. An approach to
ontology mapping negotiation. In Proceedings of
the Workshop on Integrating Ontologies, pages
54 -- 60, 2005.
[19] M. P. Singh. A social semantics for agent com-
munication languages. In Issues in Agent Com-
munication, pages 31 -- 45, London, UK, 2000.
Springer-Verlag.
[20] L. Steels. The origins of ontologies and com-
munication conventions in multi-agent systems.
Autonomous Agents and Multi-Agent Systems,
1(2):169 -- 194, Oct. 1998.
14
|
1106.3134 | 1 | 1106 | 2011-06-16T03:16:20 | Communicate only when necessary: Cooperative tasking for multi-agent systems | [
"cs.MA",
"eess.SY",
"math.OC"
] | New advances in large scale distributed systems have amazingly offered complex functionalities through parallelism of simple and rudimentary components. The key issue in cooperative control of multi-agent systems is the synthesis of local control and interaction rules among the agents such that the entire controlled system achieves a desired global behavior. For this purpose, three fundamental problems have to be addressed: (1) task decomposition for top-down design, such that the fulfillment of local tasks guarantees the satisfaction of the global task, by the team; (2) fault-tolerant top-down design, such that the global task remains decomposable and achievable, in spite of some failures, and (3) design of interactions among agents to make an undecomposable task decomposable and achievable in a top-down framework. The first two problems have been addressed in our previous works, by identifying necessary and sufficient conditions for task automaton decomposition, and fault-tolerant task decomposability. This paper deals with the third problem and proposes a procedure to redistribute the events among agents in order to enforce decomposability of an undecomposable task automaton. The decomposability conditions are used to identify the root causes of undecomposability which are found to be due to over-communications that have to be deleted, while respecting the fault-tolerant decomposability conditions; or because of the lack of communications that require new sharing of events, while considering new violations of decomposability conditions. This result provides a sufficient condition to make any undecomposable deterministic task automaton decomposable in order to facilitate cooperative tasking. Illustrative examples are presented to show the concept of task automaton decomposabilization. | cs.MA | cs |
Technical Report: NUS-ACT-11-003-Ver.1:
Communicate only when necessary:
Cooperative tasking for multi-agent systems
Mohammad Karimadini, and Hai Lin
Abstract
New advances in large scale distributed systems have amazingly offered complex functionalities
through parallelism of simple and rudimentary components. The key issue in cooperative control of
multi-agent systems is the synthesis of local control and interaction rules among the agents such that
the entire controlled system achieves a desired global behavior. For this purpose, three fundamental
problems have to be addressed: (1) task decomposition for top-down design, such that the fulfillment of
local tasks guarantees the satisfaction of the global task, by the team; (2) fault-tolerant top-down design,
such that the global task remain decomposable and achievable, in spite of some failures, and (3) design of
interactions among agents to make an undecomposable task decomposable and achievable in a top-down
framework. The first two problems have been addressed in our previous works, by identifying necessary
and sufficient conditions for task automaton decomposition, and fault-tolerant task decomposability,
based on decision making on the orders and selections of transitions, interleaving of synchronized strings
and determinism of bisimulation quotient of local task automata. This paper deals with the third problem
and proposes a procedure to redistribute the events among agents in order to enforce decomposability of
an undecomposable task automaton. The decomposability conditions are used to identify the root causes
of undecomposability which are found to be due to over-communications that have to be deleted, while
respecting the fault-tolerant decomposability conditions; or because of the lack of communications
that require new sharing of events, while considering new violations of decomposability conditions.
This result provides a sufficient condition to make any undecomposable deterministic task automaton
decomposable in order to facilitate cooperative tasking. Illustrative examples are presented to show the
concept of task automaton decomposabilization.
M. Karimadini and H. Lin are both from the Department of Electrical and Computer Engineering, National University of
Singapore, Singapore. Corresponding author, H. Lin [email protected]
I. INTRODUCTION
With new advances in technology and emergence of large scale complex systems [1], [2], there
is an ever-increasing demand for cooperative control of distributed systems with sophisticated
specifications [3], [4], [5], [6] which impose new challenges that fall beyond the traditional
methods [7], [8], [9], [5]. Conventional approaches either consider the team of agents as a
monolithic plant to be controlled by a centralized unit, or design and iteratively adjust local
controllers, in a bottom-up structure, to generate a behavior closed to a desired global behavior.
Although the latter approache offers more flexibility, scalability and functionality with lower
cost, due to local actuation and communications of agents [10], [11], [12], they fail to guarantee
a given global specification [13]. For this purpose, top-down cooperative control aims at formal
design of local controllers in order to collectively achieve the global specification, by design
[14], [15].
To address the top-down cooperative control, three fundamental questions are evoked: The first
question is the task decomposition problem that is interested in understanding of whether all tasks
are decomposable, and if not, what are the conditions for task decomposability. It furthermore
asks that if the task is decomposable and local controllers are designed to satisfy local tasks,
whether the whole closed loop system satisfies the global specification. Subsequently, the second
question refers to the cooperative control under event failures, and would like to know if after
the task decomposition and local controller designs for global satisfaction, some events fail in
some agents, then whether the task still remains decomposable and globally satisfied, in spite of
event failures. As another follow-up direction, the third question investigates the way to make an
undecomposable task decomposable through modification of local agents in order to accomplish
the proposed cooperative control.
For cooperative control of logical behaviors [16], represented in automata [17], [18], the first
question (task decomposability for cooperative tasking) was addressed in our previous work
[19], by decomposing a given global task automaton into two local task automata such that
their parallel composition bisimulates the original task automaton. By using the notion of shared
events, instead of common events and incorporating the concept of global decision making on the
orders and selections between the transitions, the decomposability result was generalized in [20]
to an arbitrary finite number of agents. Given a deterministic task automaton, and a set of local
event sets, necessary and sufficient conditions were identified for task automaton decomposability
based on decision making on the orders and selections of transitions, interleaving of synchronized
strings and determinism of bisimulation quotient of local automata. It was also proven that the
fulfillment of local task automata guarantees the satisfaction of the global specification, by design.
The second question, cooperative tasking under event failure, was investigated in [21], by
introducing a notion of passive events to transform the fault-tolerant task decomposability prob-
lem to the standard automaton decomposability problem in [20]. The passivity was found to
reflect the redundancy of communication links, based on which the necessary and sufficient
conditions have been then introduced under which a previously decomposable task automaton
remains decomposable and achievable, in spite of events failures. The conditions ensure that
after passive failures, the team of agents maintains its capability for global decision making on
the orders and selections between transitions; no illegal behavior is allowed by the team (no
new string emerges in the interleavings of local strings) and no legal behavior is disabled by
the team (any string in the global task automaton appears in the parallel composition of local
automata). These conditions interestingly guarantee the team of agents to still satisfy its global
specification, even if some local agents fail to maintain their local specifications.
This paper deals with the third question to investigate how to make undecomposable task
automata decomposable in order for cooperative tasking of multi-agent systems. For a global
task automaton that is not decomposable with respect to given local event sets, the problem
is particularly interested in finding a way to modify the local task automata such that their
parallel composition bisimulates the original global task automaton, to guarantee its satisfaction
by fulfilling the local task automata.
Decomposition of different formalisms of logical specification have been reported in the
literature. Examples of such methods can be seen for decomposition of a specification given
in CSP [22], decomposition of a LOTOS [23], [24], [25] and decomposition of petri nets [26],
[27]. The problem of automaton decomposabilization has been also studies in computer science
literature. For example, [28] characterized the conditions for decomposition of asynchronous
automata in the sense of isomorphism based on the maximal cliques of the dependency graph.
The isomorphism equivalence used in [28] is however a strong condition, in the sense that
two isomorphic automata are bisimilar but not vise versa [17]. Moreover, [28] considers a set
of events to be attributed to a number of agents, with no predefinition of local event sets.
While event attribution is suitable for parallel computing and synthesis problems in computer
science, control applications typically deal with parallel distributed plants [29] whose events are
predefined by the set of sensors, actuators and communication links across the agents. Therefore,
it would be advantageous to find a way to make an undecomposable automaton decomposable
with respect to predefined local event sets, by modifying local task automata. Since the global task
automaton is fixed, one way to modify the local task automata is through the modification in local
event sets, which is the main theme of this paper. Another related work is [30] that proposes
a method for automaton decomposabilization by adding synchronization events such that the
parallel composition of local automata is observably bisimilar to the original automaton. The
approach in [30], however, allows to add synchronization events to the event set that will enlarge
the size of global event set. Our work deals with those applications with fixed global event sets
and predefined distribution of events among local agents, where enforcing the decomposability
is not allowed by adding the new synchronization events, but instead by redistribution of the
existing events among the agents.
For this purpose, we propose an algorithm that uses previous results on task decomposition
[19], [20] to identify and overcome dissatisfaction of each decomposability condition. The
algorithm first removes all redundant communication links using the fault-tolerant result [21].
As a result, any violation of decomposability conditions, remained after this stage, is not due
to redundant communication links, and hence cannot be removed by means of link deletions.
Instead, the algorithm proceeds by establishing new communication links to provide enough
information to facilitate the task automaton decomposition. Since each new communication link
may overcome several violations of decomposability conditions, the algorithm may offer different
options for link addition, leading to the question of optimal decomposability with minimum
number of communication links. It is found that if link additions impose no new violations
of decomposability conditions, then it is possible to make the automaton decomposable with
minimum number of links. However, it is furthermore shown that, in general, addition of new
communication links may introduce new violations of decomposability conditions that in turn
require establishing new communication links. In such cases, the optimal path depends on the
structure of the automaton and requires a dynamic exhaustive search to find the sequence of link
additions with minimum number of links. Therefore, in case of new violations, a simple sufficient
condition is proposed to provide a feasible suboptimal solution to enforce the decomposability,
without checking of decomposability conditions after each link addition. This approach can
decompose any deterministic task automaton, after which, according to the previous results,
designing local controllers such that local specification are satisfied, guarantees the fulfillment
of the global specification, by design.
The rest of the paper is organized as follows. Preliminary lemmas, notations, definitions and
problem formulation are represented in Section II. This section also establishes the links to pre-
vious works on task automaton decomposition and fault-tolerant decomposition results. Section
III proposes an algorithm to make any undecomposable deterministic automaton decomposable
by modifying its local event sets. Illustrative examples are also given to elaborate the concept of
task automaton decomposabilization. Finally, the paper concludes with remarks and discussions
in Section IV. Proofs of the lemmas are readily given in the Appendix.
II. PROBLEM FORMULATION
A. Definitions and notations
We first recall the definitions and notations used in this paper.
A deterministic automaton is a tuple A := (Q, q0, E, δ) consisting of a set of states Q;
an initial state q0 ∈ Q; a set of events E that causes transitions between the states, and a
transition relation δ ⊆ Q × E × Q, with partial map δ : Q × E → Q, such that (q, e, q′) ∈ δ
if and only if state q is transited to state q′ by event e, denoted by q e→ q′ (or δ(q, e) = q′).
A nondeterministic automaton is a tuple A := (Q, q0, E, δ) with a partial transition map δ :
Q × E → 2Q, and if hidden transitions (ε-moves) are also possible, then a nondeterministic
automaton with hidden moves is defined as A := (Q, q0, E ∪ {ε}, δ) with a partial map δ :
Q × (E ∪ {ε}) → 2Q. For a nondeterministic automaton the initial state can be generally from a
set Q0 ⊆ Q. Given a nondeterministic automaton A, with hidden moves, the ε-closure of q ∈ Q,
A(q).
denoted by ε∗
The transition relation can be extended to a finite string of events, s ∈ E∗, where E∗ stands for
A(q) ⊆ Q, is recursively defined as: q ∈ ε∗
A(q); q′ ∈ ε∗
A(q) ⇒ δ(q′, ε) ⊆ ε∗
hidden moves, ε∗
Kleene−Closure of E (the set of all finite strings over elements of E). For an automaton without
A(q) = {q}, and the transition on string is inductively defined as δ(q, ε) = q
(empty move or silent transition), and δ(q, se) = δ(δ(q, s), e) for s ∈ E∗ and e ∈ E. For
an automaton A, with hidden moves, the extension of transition relation on string, denoted by
δ : Q × E∗ → 2Q, is inductively defined as: ∀q ∈ Q, s ∈ E∗, e ∈ E: δ(q, ε) := ε∗
A(q) and
δ(q, se) = ε∗
A(δ(δ(q, s), e)) = ∪
q′∈δ(q,s)(cid:26) ∪
q′′∈δ(q′,e)
ε∗
A(q′′)(cid:27) [18].
The operator Ac(.) [17] is then defined by excluding the states and their attached transitions
that are not reachable from the initial state as Ac(A) = (Qac, q0, E, δac) with Qac = {q ∈ Q∃s ∈
E∗, q ∈ δ(q0, s)} and δac = δQac × E → Qac, restricting δ to the smaller domain of Qac. Since
Ac(.) has no effect on the behavior of the automaton, from now on we take A = Ac(A).
We focus on deterministic global task automata that are simpler to be characterized, and cover
a wide class of specifications. The qualitative behavior of a deterministic system is described by
the set of all possible sequences of events starting from the initial state. Each such a sequence is
called a string, and the collection of strings represents the language generated by the automaton,
denoted by L(A). The existence of a transition over a string s ∈ E∗ from a state q ∈ Q is
denoted by δ(q, s)!. Considering a language L, by δ(q, L)! we mean that ∀ω ∈ L : δ(q, ω)!. For
e ∈ E, s ∈ E∗, e ∈ s means that ∃t1, t2 ∈ E∗ such that s = t1et2. In this sense, the intersection
of two strings s1, s2 ∈ E∗ is defined as s1 ∩ s2 = {ee ∈ s1 ∧ e ∈ s2}. Likewise, s1\s2 is
defined as s1\s2 = {ee ∈ s1, e /∈ s2}. For s1, s2 ∈ E∗, s1 is called a sub-string of s2, denoted
by s1 6 s2, when ∃t ∈ E∗, s2 = s1t. Two events e1 and e2 are called successive events if
∃q ∈ Q : δ(q, e1)! ∧ δ(δ(q, e1), e2)! or δ(q, e2)! ∧ δ(δ(q, e2), e1)!. Two events e1 and e2 are called
adjacent events if ∃q ∈ Q : δ(q, e1)! ∧ δ(q, e2)!.
2 ∈ Q2 such that δ2(q2, e) = q′
To compare the task automaton and its decomposed automata, we use the bisimulation rela-
i , E, δi), i = 1, 2. A relation R ⊆ Q1 × Q2 is said to
tions. Consider two automata Ai = (Qi, q0
1, then
be a simulation relation from A1 to A2 if (q0
2) ∈ R. If R is defined for all states and all events in A1,
∃q′
then A1 is said to be similar to A2 (or A2 simulates A1), denoted by A1 ≺ A2 [17]. If A1 ≺ A2,
A2 ≺ A1, with a symmetric relation, then A1 and A2 are said to be bisimilar (bisimulate each
∼= A2 [31]. In general, bisimilarity implies languages equivalence but the
other), denoted by A1
converse does not necessarily hold [32].
2) ∈ R, and ∀ (q1, q2) ∈ R, δ1(q1, e) = q′
1, q0
2, (q′
1, q′
In these works natural projection is used to obtain local tasks, as local perspective of agents
from the global task. Consider a global event set E and its local event sets Ei, i = 1, 2, ..., n,
with E =
is inductively defined as pi(ε) = ε,
Ei. Then, the natural projection pi : E∗ → E∗
i
n
∪
i=1
and ∀s ∈ E∗, e ∈ E : pi(se) =
: E∗
p−1
i
i → 2E ∗ is defined on an string t ∈ E∗
pi(s)e if e ∈ Ei;
otherwise.
i as p−1
pi(s)
Accordingly, inverse natural projection
i (t) := {s ∈ E∗pi(s) = t}.
The natural projection is also defined on automata as Pi
: A → A, where, A is the set
of finite automata and Pi(AS) are obtained from AS by replacing its events that belong to
E\Ei by ε-moves, and then, merging the ε-related states. The ε-related states form equivalent
classes defined as follows. Consider an automaton AS = (Q, q0, E, δ) and a local event set
Ei ⊆ E. Then, the relation ∼Ei is the equivalence relation on the set Q of states such that
δ(q, e) = q′ ∧ e /∈ Ei ⇒ q ∼Ei q′, and [q]Ei denotes the equivalence class of q defined on
and defined as
∼Ei. The set of equivalent classes of states over ∼Ei, is denoted by Q/∼Ei
= {[q]Eiq ∈ Q} [28]. The natural projection of AS into Ei is then formally defined as
Q/∼Ei
, [q0]Ei, Ei, δi), with δi([q]Ei, e) = [q′]Ei if there exist states q1 and q′
Pi(AS) = (Qi = Q/∼Ei
1
such that q1 ∼Ei q, q′
1 ∼Ei q′, and δ(q1, e) = q′
1.
To investigate the interactions of transitions between automata, particularly between Pi(AS),
i = 1, . . . , n, the synchronized product of languages is defined as follows. Consider a global
event set E and local event sets Ei, i = 1, . . . , n, such that E =
Ei. For a finite set of
i=1, the synchronized product (language product) of {Li}, denoted by
languages {Li ⊆ E∗
i }n
n
n
Li = {s ∈ E∗∀i ∈ {1, . . . , n} : pi(s) ∈ Li} =
i=1
i=1
Li, is defined as
Then, parallel composition (synchronized product) is used to define the composition of local
p−1
i (Li) [14].
n
∪
i=1
n
∩
i=1
task automata to retrieve the global task automaton, and to model each local closed loop system
by compositions of its local plant and local controller automata. Let Ai = (Qi, q0
i , Ei, δi), i = 1, 2
be automata. The parallel composition (synchronous composition) of A1 and A2 is the automaton
2), E = E1 ∪ E2, δ), with δ defined as ∀(q1, q2) ∈ Q, e ∈ E:
A1A2 = (Q = Q1 × Q2, q0 = (q0
1, q0
δ((q1, q2), e) =
(δ1(q1, e), δ2(q2, e)) ,
(δ1(q1, e), q2) ,
;
e ∈ E1 ∩ E2
δ1(q1, e)!, δ2(q2, e)!
if
if δ1(q1, e)!, e ∈ E1\E2;
if δ2(q2, e)!, e ∈ E2\E1;
otherwise.
(q1, δ2(q2, e)) ,
undefined,
The parallel composition of Ai, i = 1, 2, ..., n is called parallel distributed system (or concur-
rent system), and is defined based on the associativity property of parallel composition [17] as
n
k
i=1
Ai = A1 k ... k An = An k (An−1 k (· · · k (A2 k A1))).
The set of labels of local event sets containing an event e is called the set of locations of e,
denoted by loc(e) and is defined as loc(e) = {i ∈ {1, . . . , n}e ∈ Ei}.
Based on these definitions, a task automaton AS with event set E and local event sets Ei,
Ei, is said to be decomposable with respect to parallel composition and
i = 1, ..., n, E =
n
∪
i=1
natural projections Pi, i = 1, · · · , n, when
B. Problem formulation
n
k
i=1
Pi (AS) ∼= AS.
In [19], we have shown that not all automata are decomposable with respect to parallel
composition and natural projections, and subsequently necessary and sufficient conditions were
proposed for decomposability of a task automaton with respect to parallel composition and
natural projections into two local event sets. These necessary and sufficient conditions were then
generalized to an arbitrary finite number of agents, in [20], as
Ei, δ(cid:19) is
Lemma 1: (Corollary 1 in [20]) A deterministic automaton AS = (cid:18)Q, q0, E =
decomposable with respect to parallel composition and natural projections Pi, i = 1, ..., n such
Pi (AS) if and only if AS satisfies the following decomposability conditions (DC):
that AS
Si=1
n
∼=
n
i=1
• DC1: ∀e1, e2 ∈ E, q ∈ Q: [δ(q, e1)! ∧ δ(q, e2)!]
⇒ [∃Ei ∈ {E1, . . . , En}, {e1, e2} ⊆ Ei] ∨ [δ(q, e1e2)! ∧ δ(q, e2e1)!];
• DC2: ∀e1, e2 ∈ E, q ∈ Q, s ∈ E∗: [δ(q, e1e2s)! ∨ δ(q, e2e1s)!]
⇒ [∃Ei ∈ {E1, . . . , En}, {e1, e2} ⊆ Ei] ∨ [δ(q, e1e2s)! ∧ δ(q, e2e1s)!];
n
i=1
• DC3: δ(q0,
pi (si))!, ∀{s1, · · · , sn} ∈ L (AS), ∃si, sj ∈ {s1, · · · , sn}, si 6= sj, where,
L (AS) ⊆ L (AS) is the largest subset of L (AS) such that ∀s ∈ L (AS) ∃s′ ∈ L (AS) , ∃Ei, Ej ∈
{E1, ..., En} , i 6= j, pEi∩Ej (s) and pEi∩Ej (s′) start with the same event, and
• DC4: ∀i ∈ {1, ..., n}, x, x1, x2 ∈ Qi, x1 6= x2, e ∈ Ei, t ∈ E∗
i , δi(x, e) = x1, δi(x, e) = x2:
δi(x1, t)! ⇔ δi(x2, t)!.
The first two decomposability conditions require the team to be capable of decision on
choice/order of events, by which for any such decision there exists at least one agent that
knows both events, or the decision is not important. Moreover, the third and fourth conditions,
guarantee that the cooperative perspective of agents from the tasks (parallel composition of local
task automata) neither allows a string that is prohibited by the global task automaton, nor disables
a string that is allowed in the global task automaton.
It was furthermore shown that once the task automaton is decomposed into local task automata
and local controllers are designed for local plants to satisfy the local specifications, then the global
specification is guaranteed, by design.
The next question was the reliability of task decomposability to understand whether a pre-
viously decomposable and achievable global task automaton, can still remain decomposable
and achievable by the team, after experiencing some event failures. For this purpose, in [21],
a class of failures was investigated as follows to defined a notion of passivity. Consider an
automaton A = (Q, q0, E, δ). An event e ∈ E is said to be failed in A (or E), if F (A) =
PΣ(A) = PE\e(A) = (Q, q0, Σ = E\e, δF ), where, Σ, δF and F (A) denote the post-failure
event set, post-failure transition relation and post-failure automaton, respectively. A set ¯E ⊆ E
of events is then said to be failed in A, when for ∀e ∈ ¯E, e is failed in A, i.e., F (A) =
PΣ(Ai) = PE\ ¯E(A) = (Q, q0, Σ = E\ ¯E, δF ). Considering a parallel distributed plant A :=
n
0, Ei, δi), i = 1, . . . , n. Failure of e
i=1
in Ei is said to be passive in Ei (or Ai) with respect to
failure in Ai is a passive failure is called a passive event in Ai.
Ei, δ) with local agents Ai = (Qi, qi
n
i=1
Σi. An event whose
Ai = (Z, z0, E =
Ai, if E =
n
∪
i=1
n
∪
i=1
The passivity was found to reflect the redundancy of communication links and shown to
be a necessary condition for preserving the automaton decomposability. It was furthermore
shown that when all failed events are passive in the corresponding local event sets, the problem
of decomposability under event failure can be transformed into the standard decomposability
problem to find the conditions under which AS
∼=
PEi\ ¯Ei(AS), as follows.
Lemma 2: (Theorem 1 in [21]) Consider a deterministic task automaton AS = (Q, q0, E =
n
Pi(AS), and furthermore, assume that
Ei, δ). Assume that AS is decomposable, i.e., AS
∪
i=1
¯Ei = {ai,r} fail in Ei, r ∈ {1, ..., ni}, and ¯Ei are passive for i ∈ {1, . . . , n}. Then, AS remains
decomposable, in spite of event failures, i.e., AS
F (Pi (AS)) if and only if
n
i=1
∼=
∼=
n
i=1
n
i=1
• EF 1: ∀e1, e2 ∈ E, q ∈ Q: [δ(q, e1)! ∧ δ(q, e2)!]
⇒ [∃Ei ∈ {E1, · · · , En}, {e1, e2} ⊆ Ei\ ¯Ei] ∨ [δ(q, e1e2)! ∧ δ(q, e2e1)!];
• EF 2: ∀e1, e2 ∈ E, q ∈ Q, s ∈ E∗: [δ(q, e1e2s)! ∨ δ(q, e2e1s)!]
⇒ [∃Ei ∈ {E1, · · · , En}, {e1, e2} ⊆ Ei\ ¯Ei] ∨ [δ(q, e1e2s)! ∧ δ(q, e2e1s)!];
• EF 3: δ(q0,
pi (si))!, ∀{s1, · · · , sn} ∈ L (AS), ∃si, sj ∈ {s1, · · · , sn}, si 6= sj, where,
L (AS) ⊆ L (AS) is the largest subset of L (AS) such that ∀s ∈ L (AS) , ∃s′ ∈ L (AS) , ∃Σi,
n
i=1
Σj ∈ {Σ1, ..., Σn} , i 6= j, pΣi∩Σj (s) and pΣi∩Σj (s′) start with the same event, and
• EF 4: ∀i ∈ {1, . . . , n}, x, x1, x2 ∈ Qi, x1 6= x2, e ∈ Ei\ ¯Ei, t1, t2 ∈ ¯E∗
2 such that pEi\ ¯Ei(t′
δi(x, t2e) = x2: δi(x1, t′
2)!, for some t′
i , δi(x, t1e) = x1,
1) = pEi\ ¯Ei(t′
2).
1)! ⇔ δi(x2, t′
1, t′
EF 1-EF 4 are respectively the decomposability conditions DC1-DC4, after event failures
with respect to parallel composition and natural projections into refined local event sets Σi =
Ei\ ¯Ei, i ∈ {1, . . . , n}, provided passivity of ¯Ei, i ∈ {1, . . . , n}.
In this paper we are interested in the case that a task automaton is not decomposable and would
like to ask whether it is possible to make it decomposable, and if so, whether the automaton
can be made decomposable with minimum number of communication links. This problem is
formally stated as
Problem 1: Consider a deterministic task automaton AS with event set E =
Ei for n
agents with local event sets Ei, i = 1, . . . , n. If AS is not decomposable, can we modify the
sets of private and shared events between local event sets such that AS becomes decomposable
with respect to parallel composition and natural projections Pi, with the minimum number of
communication links?
n
∪
i=1
One trivial way to make an automaton A decomposable, is to share all events among all
agents, i.e., Ei = E, ∀i = 1, . . . , n. This method , however, is equivalent to centralized control.
In general, in distributed large scale systems, one of the objectives is to sustain the systems
functionalities over as few number of communication links as possible, as will be addressed in
the next section.
III. TASK AUTOMATON DECOMPOSABILIZATION
A. Motivating Examples
This section is devoted to Problem 1 and proposes an approach to redefine the set of private and
shared events among agents in order to make an undecomposable task automaton decomposable.
For more elaboration, let us to start with a motivating examples.
Example 1: Consider two sequential belt conveyors feeding a bin, as depicted in Figure 1. To
avoid the overaccumulation of materials on Belt B, when the bin needs to be charged, at first
Belt B and then (after a few seconds), Belt A should be started. After filling the bin, to stop the
charge, first Belt A and then after a few seconds Belt B is stopped to get completely emptied.
The global task automaton, showing the order of events in this plant, is shown in Figure 2.
Fig. 1. The process of two belt conveyors charging a bin.
AS:
BStart
/ •
/ •
@A
BinF ull
/ •
/ •
AStart
/ •
AStop
BStop
/ •
BCD
Fig. 2. Global task automaton for belt conveyors and bin.
BinEmpty
The local event sets for Belt A and Belt B are EA = {AStart, BinF ull, AStop} and EB =
{BStart, BStop, BinEmpty}, respectively, with AStart:= Belt A start; BinF ull:= Bin full; AStop:=
Belt A stop and wait for 10 Seconds; BStart:= Belt B start and wait for 10 Seconds; BStop:=
Belt B stop, and BinEmpty: Bin empty.
The task automaton is not decomposable with respect to parallel composition and natural pro-
jection Pi, i ∈ {A, B}, due to violation of DC by successive private event pairs {BStart, AStart}
and {AStop, BStop}. To make AS decomposable, (BStart ∨ AStart) ∧ (AStop ∨ BStop) should
become common between EA and EB. Therefore, four options are possible: (BStart ∧ BStop),
(BStart ∧ AStop), (AStart ∧ BStop), or (AStart ∧ AStop) become common. In each of these op-
tions two private events should become common, and hence, all four options are equivalent
in the sense of optimality. Consider for example AStart and AStop to become common. In
this case the new local event sets are formed as EA = {AStart, BinF ull, AStop} and EB =
{BStart, BStop, BinEmpty, AStart, AStop}. The automaton AS will then become decomposable (i.e.,
PA(AS)PB(AS) ∼= AS) with the new local event sets with the corresponding local task automata
as are shown in Figure 3.
In this example, different sets of private events can be chosen to make AS decomposable.
All of these sets have the same cardinality, and hence, no optimality is arisen in this example.
Next example shows a case with different choices of private event sets to be shared, suggesting
/
/
/
/
/
/
O
O
PA(AS):
AStart
/ •
BinF ull
/ •
@A
, PB(AS):
/ •
BCD
BStart
/ •
/ •
AStart
/ •
@A
AStop
/ •
BStop
/ •
BCD
AStop
BinEmpty
Fig. 3.
{BStart, BStop, BinEmpty, AStart, AStop}.
task automata
Local
for belt
conveyors, with EA = {AStart, BinF ull, AStop} and EB =
optimal decomposition by choosing the set with the minimum cardinality.
Example 2: Consider two local event sets E1 = {e1, e3} and E2 = {e2}, with the global task
. This automaton is undecomposable due to violation of
automaton
e2
e3
/ •
/ •
(PPPPPP
e1
•
•
DC by e2 ∈ E2\E1 and {e1, e3} ∈ E1\E2. To make it decomposable, one event among the
set {e1, e2} and another event among the set {e2, e3} (either {e2} or {e1, e3}) should become
common. Therefore, in order for optimal decomposabilization, {e2} is chosen to become common
due to its minimum cardinality. It is obvious that in this case only one event should become
common while if {e1, e3} was chosen, then two events were required to be shared.
Motivated by these examples, the core idea in our decompozabilization approach is to first
check the decomposability of a given task automaton AS, by Lemma 1, and if it is not decom-
posable, i.e., either of DC1-DC4 is violated then the proposed method is intended to make AS
decomposable, by eradicating the reasons of dissatisfying of decomposability conditions. We will
show that violation of decomposability conditions, can be rooted from two different sources: it
can be because of over-communication among agents, that may lead to violation of DC3 or/and
DC4, or due to lack of communication, that may lead to violation of DC1, DC2, DC3 or/and
DC4. Accordingly, decomposability can be enforced using two methods of link deletion and link
addition, subjected to the type of undecomposability. Considering link deletion as an intentional
event failure, according to Lemma 2 a link can be deleted only if it is passive and its deletion
respects EF 1-E4. On the other hand, the second method of enforcing of decomposability, i.e.,
establishing new communication links, may result in new violations of DC3 or DC4, that should
be treated, subsequently.
In order to proceed the approach, we firstly introduce four basic definitions to detect the
components that contribute in violation of each decomposability condition and then propose basic
lemmas through which the communication links, and hence the local event sets are modified to
/
/
/
O
O
/
/
/
/
/
O
O
/
/
/
(
/
resolve the violations of decomposability conditions.
B. Enforcing DC1 and DC2
This part deals with enforcing of DC1 and DC2. For this purpose, the set of events that
violate DC1 or DC2 is defined as follows.
n
∪
i=1
Definition 1: (DC1&2-Violating set) Consider the global task automaton AS with local event
sets Ei for n agents such that E =
Ei. Then, the DC1&2-Violating set operator V : AS →
E × E, indicates the set of event pairs that violate DC1 or DC2 (violating pairs), and is
defined as V (AS) := {{e1, e2}e1, e2 ∈ E, ∀Ei ∈ {E1, . . . , En}, {e1, e2} 6⊂ Ei, ∃q ∈ Q such that
δ(q, e1)!∧δ(q, e2)!∧¬[δ(q, e1e2)!∧δ(q, e2e1)!] or ¬[δ(q, e1e2s)! ⇔ δ(q, e2e1s)!]}, for some s ∈ E∗.
Moreover, W : AS → E is defined as W (AS) := {e ∈ E∃e′ ∈ E such that {e, e′} ∈ V (AS)},
and shows the set of events that contribute in V (AS) (violating events). For a particular event e
and a specific local event set Ei ∈ {E1, . . . , En}, We(AS, Ei) is defined as We(AS, Ei) = {e′ ∈
Ei{e, e′} ∈ V (AS)}. This set captures the collection of events from Ei that pair up with e to
contribute in violation of DC1 or DC2. The cardinality of this set will serve as an index for
optimal addition of communication links to make V (AS) empty.
This definition suggests a way to remove a pair of events {e1, e2} from V (AS), by sharing
e1 with one of the agents in loc(e2) or by sharing e2 with one of the agents in loc(e1). Once
there exist an agent that knows both event, loc(e1) ∩ loc(e2) becomes nonempty and e1 and e2
no longer contribute in violation of DC1 or DC2 since [∃Ei ∈ {E1, . . . , En}, {e1, e2} ⊆ Ei]
becomes true for e1 and e2 in Lemma 1. Therefore,
Lemma 3: The set V (AS) becomes empty, if for any {e, e′} ∈ V (AS), e is included in Ei for
some i ∈ loc(e′), or e′ is included in Ej for some j ∈ loc(e). In this case, {e, Ei} or {e′, Ej} is
called a DC1&2-enforcing pair for DC1&2-violating pair {e, e′}.
Example 3: In Example 2, V (AS) = {{e1, e2}, {e2, e3}}, W (AS) = {e1, e2, e3}. Including e2
in E1 vanishes V (AS) and makes AS decomposable.
However, applying Lemma 3 may offer different options for event sharing, since pairs in
V (AS) may share some events. In this case, the minimum number of event conversions would
be obtained by forming a set of events that are most frequently shared between the violating
pairs. This gives the minimum cardinality for the set of private events to be shared, leading to
minimum number of added communication links. Such choice of events offers a set of events
that span all violating pairs. These pairs are captured by We(AS, Ei) for any event e. In order to
minimize the number of added communication links for vanishing V (AS), one needs to maximize
the number of deletions of pairs from V (AS) per any link addition. For this purpose, for any
event e, We(AS, Ei) is formed to understand the frequency of appearance of e in V (AS) for
any Ei, and then, the event set Ei with maximum We(AS, Ei) is chosen to include e (Here,
. denotes the set's cardinality). In this case, inclusion of e in Ei will delete as many pairs as
possible from V (AS).
Interestingly, these operators can be represented using graph theory as follows. A graph G =
(W, Σ) consists of a node set W and an edge set Σ, where an edge is an unordered pair of
distinct vertices. Two nodes are said to be adjacent if they are connected through an edge, and
an edge is said to be incident to a node if they are connected. The valency of a node is then
defined as the number of its incident edges [33]. Now, since we are interested in removing the
violating pairs by making one of their events to be shared, it is possible to consider the violating
events as nodes of a graph such that two nodes are adjacent in this graph when they form a
violating pair. This graph is formally defined as follows.
Definition 2: (DC1&2-Violating Graph) Consider a deterministic automaton AS. The DC1&2-
Violating graph, corresponding to V (AS), is a graph G(AS) = (W (AS), Σ). Two nodes e1 and
e2 are adjacent in this graph when {e1, e2} ∈ V (AS).
In this formulation, the valency of each node e with respect to a local event set Ei ∈ {E1, . . . , En}
is determined by val(e, Ei) = We(AS, Ei). When e is included into Ei, it means that all
violating pairs containing e and events from Ei are removed from V (AS), and equivalently, all
corresponding incident edges are removed from G(AS). For this purpose, following algorithm
finds the set with the minimum number of private events to be shared, in order to satisfy DC1
and DC2. The algorithm is accomplished on graph G(AS), by finding e and Ei with maximum
We(AS, Ei) and including e in Ei, deleting all edges from e to Ei, updating W (AS), and
continuing until there is not more edges in G(AS) to be deleted.
Algorithm 1:
1) For a deterministic automaton AS, with local event sets Ei, i = 1, . . . , n, violating DC1 or
i = Ei, i = 1, . . . , n; V 0(AS) = V (AS);
DC2, form the DC1&2-Violating graph ; set E0
W 0(AS) = W (AS); G0(AS) = (W (AS), Σ); k=1;
2) Among all events in the nodes in W k−1(AS), find e with the maximum W k−1
e
(AS, Ek−1
i
),
i
for all Ek−1
i = Ek−1
4) update W k
3) Ek
i
∈ {Ek−1
1
, . . . , Ek−1
n };
∪ {e}; and delete all edges from e to Ek
i ;
e (AS, Ei) for all nodes of G(AS);
5) set k = k + 1 and go to step (2);
6) continue, until there exist no edges.
This algorithm successfully terminates due to finite set of edges and nodes in the graph G(AS)
and enforces AS to satisfy DC1 and DC2 as
Lemma 4: Algorithm 1 leads AS to satisfy DC1 and DC2 with minimum addition of com-
munication links. Moreover if AS satisfies DC3 and DC4 and Ek
i ∪ {e} in Step 3 does
not violate DC3 and DC4 in all iterations, then Algorithm 1 makes AS decomposable with
minimum addition of communication links.
i = Ek−1
Proof: See the Appendix for proof.
Remark 1: (Special case: Two agents) For the case of two agents, since they are only two
local event sets, for all {e, e′} ∈ V (AS), e and e′ are from different local event sets, and hence,
for n = 2, We(AS, Ei) is equivalent to val(e), and addition of e into Ei in each step implies
the deletion of all incident edges of e.
Remark 2: Although Algorithm 1 leads AS to satisfy DC1 and DC2, it may cause new
violations of DC3 or/and DC4, due to establishing new communication links.
Example 4: Consider a task automaton AS:
e2
e1
e5
e1
•
•
•
e3
hPPPPPP
e1
•
•
•
b
/ •
vnnnnnn
a
/ •
(PPPPPP
e2
with local event sets E1 = {a, b, e1,
/ •
/ •
/ •
e6
•
e4
e3, e5} and E2 = {a, b, e2, e4, e6}. Both DC1 and DC2 are violated by event pair {e1, e2} when
they require decision on a choice and a decision on their order from the initial state, while none of
the agents knows both of them. To vanish V (AS) = {{e1, e2}}, two enforcing pairs are suggested:
{e1, E2} (e1 to be included in E2) or {e2, E1} (e2 to be included in E1). However, inclusion of e1
in E2, cause a new violation of DC4 since with new E2 = {a, b, e1, e2, e4, e6}, P2(AS) is obtained
, violating DC4, due to new nondeterminism,
as P2(AS):
e1
/ •
/ •
•
e1
hPPPPPP
e1
vnnnnnn
a
•
•
•
b
for which e3 also is required to be included to E2 in order to make AS decomposable. On the
other hand, if instead of including e1 in E2, one included e2 in E1, then besides violation of
o
o
o
o
/
(
v
/
/
o
o
o
o
h
/
/
/
v
/
o
o
o
o
h
DC4 (as there does not exists a deterministic automaton that bisimulates P2(AS)), new violations
of DC3 emerged, as with new event set E1 = {a, b, e1, e2, e3, e5}, the parallel composition of
P1(AS): •
and P2(AS):
/ •
/ •
/ •
•
•
e5
e1
e2
e1
•
b
•
e3
hPPPPPP
e1
•
vnnnnnn
a
(PPPPPP
e2
•
produces string e1e2e4e6 that does not appear
/ •
vnnnnnn
a
e2
(PPPPPP
e2
/ •
•
•
b
•
e4
e6
/ •
/ •
in AS. To make AS decomposable, we also need to include e1 and e3 in E2.
C. Enforcing DC3
Lemma 3 proposes adding communication links to make DC1 and DC2 satisfied. Next step
is to deal with violations of DC3. In contrast to the cases for DC1 and DC2, violation of DC3
can be overcome either by disconnecting one of its communication links to prevent the illegal
synchronization of strings, or by introducing new shared events to fix strings and avoid illegal
interleavings.
To handle violation of DC3, we firstly define the set of tuples that violate DC3 as follows.
Definition 3: (DC3 − violating tuples) Consider a deterministic automaton AS, satisfying
DC1 and DC2 and let L (AS) ⊆ L (AS) be the largest subset of L (AS) such that ∀s ∈
L (AS) ∃s′ ∈ L (AS) , ∃Ei, Ej ∈ {E1, ..., En} , i 6= j, pEi∩Ej (s) and pEi∩Ej (s′) start with the
same event a ∈ Ei ∩ Ej. For any such Ei, Ej and a, if ∃{s1, · · · , sn} ∈ L (AS), ∃si, sj ∈
{s1, · · · , sn}, si 6= sj, si, sj ∈ L (AS), ¬δ(q0,
pi (si))!, then a is called a DC3 − violating
event with respect to s1, s2, Ei and Ej, and (s1, s2, a, Ei, Ej) is called a DC3-violating tuple.
The set of all DC3 − violating tuples is denoted by DC3 − V and defined as DC3 − V =
{(s1, s2, a, Ei, Ej)e is a DC3-violating event with respect to s1, s2, Ei and Ej }.
n
i=1
Any violation in DC3 can be interpreted in two ways: firstly, it can be seen as over-communication
of shared event a that lead to synchronization of s1 and s2 in (s1, s2, a, Ei, Ej) and emerging
illegal interleaving strings from composition of Pi(AS) and Pj(AS). In this case, if event a
is excluded from Ei or Ej, then a will no longer contribute in synchronization to generate
illegal interleavings, and hence, (s1, s2, a, Ei, Ej) will no longer remain a DC3-violating tuple.
However, exclusion of a from Ei or Ej is allowed, only if it is passive (exclusion is considered as
an intentional event failure) and does not violate EF 1-EF 4. The second interpretation reflects
a violation of DC3 as a lack of communication, such that if for any DC3 violating tuple
o
o
o
o
/
(
v
/
/
o
o
o
o
h
/
(
v
/
o
o
/
/
(s1, s2, a, Ei, Ej), one event that appears before a in s1 or s2, is shared between Ej and Ej,
then Pi(AS) and Pj(AS) will have enough information to distinguish s1 and s2 to prevent illegal
interleaving of strings. Two methods for resolving the violation of DC3 can be therefore stated
as the following lemma.
Lemma 5: Consider an automaton AS, satisfying DC1 and DC2. Then any DC3-violating
tuple (s1, s2, a, Ei, Ej) is overcome, when:
1) a is excluded from Ei or Ej (eligible if it respects passivity and EF 1-EF 4), or
2) if ∃b ∈ (Ei ∪ Ej)\(Ei ∩ Ej) that appears before a in only one of s1 and s2, then b is
included in Ei ∩ Ej, otherwise, pick e1 ∈ pEi∪Ej (s1), e2 ∈ pEi∪Ej (s2), such that e1 6= e2,
e1, e2 appear before a in s1 and s2, are included in Ei ∩ Ej.
To handle a violation of DC3, when, b ∈ Ei\Ej is to be included in Ej, then {b, Ej} is
called a DC3-enforcing pair; while, when {e1, e2} ⊆ Ei\Ej has to be included in Ej, then
{{e1, e2}, Ej} is denoted as DC3-enforcing tuple. Finally, when e1 ∈ Ei\Ej and e2 ∈ Ej\Ei
have to be included in Ej and Ei, respectively, then {{e1, Ej}, {e2, Ei}} is called a DC3-
enforcing tuple.
Proof: See the proof in the Appendix.
Remark 3: Applying the first method in Lemma 5, namely, exclusion of a from Ei or Ej in a
DC3-violating tuple (s1, s2, a, Ei, Ej), is only allowed if a is passive in that local event set, and
the exclusion does not violate EF 1-EF 4. The reason is that once a shared event a ∈ Ei ∩ Ej
becomes a private one in for example Ei, then decision makings on the order/selection between
any e ∈ Ei\a and a cannot be accomplished by the i − th agent, and if there is no other agent to
do so, then AS becomes undecomposable. Moreover, deletion of a communication link may also
result in generation of new interleavings in the composition of local automata, that are not legal
in AS (violation of EF 3). In addition, deletion of a from Ei may impose a nondeterminism in
bisimulation quotient of Pi(AS), leading to violation of EF 4. On the other hand, the second
method, namely, establishing new communication link by sharing b with Ei or Ej may lead to
new violations of DC3 or DC4 that have to be avoided or resolved, subsequently.
Both methods in Lemmas 5 present ways to resolve the violation of DC3. They differ however
in the number of added communication links, as the first method deletes links, whereas the second
approach adds communication links to enforce DC3. Therefore, in order to have as few number
of links as possible among the agents, one should start with the link deletion method first, and
if it is not successful due to violation of passivity or any of EF 1-EF 4, then link addition is
used to remove DC3-violating tuples from DC3 − V .
Example 5: This example shows an undecomposable automaton that suffers from a conflict
on a communication link whose existence violates DC3, whereas its deletion dissatisfies EF 1,
EF 2 and EF 4.
Let snde(i) and rcve(i) respectively denote the set of labels that Ai sends e to those agents and
the set of labels that Ai receives e from their agents, defined as snde(i) = {j ∈ {1, ..., n}Ai sends
e to Aj} and rcve(i) = {j ∈ {1, ..., n}i ∈ snde(j)}. Consider the task automaton AS:
•
•
•
b
b
e2
•
•
•
e1
a
eLLLLLL
yrrrrrr
b
•
•
•
e5
/ •
d
\:::::::::
c
•
a
•
e2
e1
/ •
with communication pattern
e1
9rrrrrr
%LLLLLL
e2
•
•
e2
%LLLLLL
9rrrrrr
e1
•
a
/ •
e3
/ •
2 ∈ snda,b,c,d(1), 1 /∈ snda,b,c,d(1) and local event sets E1 = {a, b, c, d, e1, e3, e5}, E2 =
{a, b, c, d, e2}, leading to P1(AS):
e1
e5
e1
,
b
•
/ •
/ •
•
b
•
a
hPPPPPP
•
a
/ •
e3
/ •
d
(PPPPPP
c
•
•
e1
6nnnnnn
a
/ •
•
•
•
•
b
a
e2
e3
9rrrrrr
%LLLLLL
e2
b
e2
e1
P2(AS): •
•
•
•
b
b
•
•
a
eLLLLLL
yrrrrrr
b
•
•
•
•
b
e5
d
\:::::::::
c
a
hPPPPPP
•
•
d
hPPPPPP
c
•
e2
6nnnnnn
a
/ •
/ •
e1
/ •
e1
9rrrrrr
%LLLLLL
e2
•
a
•
•
•
•
a
e2
9rrrrrr
%LLLLLL
9rrrrrr
e1
and P1(AS)P2(AS):
which is not bisimilar to AS. Here,
•
e2
%LLLLLL
9rrrrrr
e3
/ •
/ •
•
•
a
/ •
e3
/ •
e2
e2
•
•
•
•
AS is not decomposable since two strings e1ae2e3 and e1ae3e2 are newly generated from the
interleaving of strings in P1(AS) and P2(AS), while they do not appear in AS, and hence, DC3 is
not fulfilled, due to DC3-violating tuples (e1e2ae3, ae2, a, E1, E2) and (e2e1ae3, ae2, a, E1, E2).
Now, as Lemma 5, one way to fix the violation of DC3 is by excluding a from E2. However,
although a is passive in E2, its exclusion from E2 dissatisfies EF 1( as δ(q0, e2)! ∧ δ(q0, a)! ∧
¬[δ(q0, e2a)! ∧ δ(q0, ae2)!]) and EF 2 (since δ(q0, e1e2a)! ∧ ¬δ(q0, e1ae2)!). In this case, DC4
o
o
o
o
/
/
o
o
%
e
y
o
o
\
9
%
/
/
o
o
o
o
9
o
o
o
o
/
/
o
o
(
/
/
o
o
h
o
o
O
O
6
/
o
o
/
o
o
o
o
h
o
o
h
6
/
/
%
o
o
o
o
/
/
9
%
o
o
%
9
9
e
y
\
o
o
9
%
/
/
o
o
o
o
9
also will be violated as P2(AS) becomes P2(AS) ∼=
bisimulates no deterministic automaton.
•
/ •
vnnnnnn
d
c
(PPPPPP
e2
/ •
•
b
(PPPPPP
b
•
•
e2
/ •
that
Lemma 5 also suggests another method to enforce DC3, by including either e1 in E2 or e2 in
E1. Inclusion of e1 in E2, however, leads to another violation of DC4, as it produces a nonde-
terminism after event d. This in turn will need to include e5 in E2 to make AS decomposable.
Alternatively, instead of inclusion of e1 in E2, one can include e2 in E1, that enforces DC3
and makes AS decomposable. The second method of Lemma 5 is more elaborated in the next
example.
Example 6: This example shows handling of DC3-violating tuples using the second method
in Lemma 5, i.e., by event sharing. Later on, this example will be also used to illustrate the
enforcement of DC4. Now, consider a task automaton AS: •
with local
/ •
/ •
/ •
/ •
e3
e5
e2
a
Z555555
e1
/ •
e5
•
/ •
e3
a
/ •
e6
/ •
/ •
e1
/ •
a
/ •
e4
event sets E1 = {a, e1, e3, e5} and E2 = {a, e2, e4, e6}, and let three branches in AS from top
to bottom to be denoted as s1 := e1e3e5ae2, s3 := ae6 and s2 := e5e3e1ae4. This automaton
does not satisfy DC4 (as P2(AS) has no deterministic bisimilar automaton), as well as DC3, as
the parallel composition of P1(AS): •
have
and P2(AS): •
/ •
/ •
/ •
/ •
e3
e5
e2
a
Z555555
e1
/ •
e5
•
/ •
e3
a
/ •
/ •
e1
/ •
a
•
/ •
a
/ •
e6
Z555555
a
a
/ •
e4
/ •
illegal interleaving strings {e1e3e5ae6, e5e3e1ae2}, e1e3e5ae4 and e5e3e1ae4, corresponding to
DC3-violating tuples (s1, s2, a, E1, E2), (s1, s3, a, E1, E2) and (s2, s3, a, E1, E2), respectively.
For pairs of strings {s1, s3} and {s2, s3}, there exits an event e5 ∈ (E1 ∪ E2)\(E1 ∩ E2) that
appears before a, only in s1 and s2, but not in s3. Therefore, inclusion of e5 in E2, removes
the illegal interleavings between s1 and s2 with s3, but not across s1 and s2, as with new E2 =
/
(
v
/
/
/
(
/
/
/
/
/
/
Z
/
/
/
/
/
/
/
/
/
/
Z
/
/
/
/
/
/
Z
/
/
/
{a, e2, e4, e5, e6} and P2(AS): •
a
/ •
Z555555
e5
/ •
e5
•
/ •
a
e2
/ •
, (s1, s3, a, E1, E2) and (s2, s3, a, E1, E2) are
a
/ •
e6
/ •
/ •
e4
no longer DC3-violating tuples, while (s1, s2, a, E1, E2) still remains a DC3-violating one with
illegal interleavings e1e3e5ae4 and e5e3e1ae2. The reason is that e5 appears before a in both
s1 and s2, and there is no event that appear before a only in one of the strings s1 and s2.
For this case, according to Lemma 5, two different events that appear before "a", one from
pE1∪E2(s1) = s1 and the other from pE1∪E2(s2) = s2, i.e., e1 and e5 have to be attached to E2,
resulting in E2 = {a, e1, e2, e4, e5, e6}, •
and P1(AS)P2(AS) ∼= AS.
/ •
/ •
/ •
e5
e2
a
Z555555
e1
/ •
•
D. Enforcing DC4
e5
/ •
e1
a
/ •
e6
/ •
/ •
a
/ •
e4
Similar to DC1-DC3, a violation of DC4 can be regarded as a lack of communication link
that causes nondeterminism in a local task automaton. Such interpretation calls for establishing
a new communication link to prevent the emergence of local nondeterminism. Moreover, when
this local nondeterminism occurs on a shared event, the corresponding violation of DC4 can be
overcome by excluding the shared event from the respective local event set. It should be noted
however that the event exclusion should respect the passivity and EF 1-EF 4 conditions. When
DC4 is enforced by link additions, similar to what we discussed for DC3, addition of new
communication link may cause new violations of DC3 or/and DC4. To enforce DC4, firstly a
DC4-violating tuple is defined as follows.
Definition 4: (DC4 − violating tuple) Consider a deterministic automaton AS with local
event sets Ei = 1, . . . , n, ∀i ∈ {1, ..., n}, q, q1, q2 ∈ Q, t1, t2 ∈ (E\Ei)∗, e ∈ Ei, δ(q, t1e) =
q1 6= δ(q, t2e) = q2, ∃t ∈ E∗, δ(q1, t)!, but ∄t′ ∈ E∗ such that δ(q2, t′)!, pi(t) = pi(t′). Then,
(q, t1, t2, e, Ei) is called a DC4-violating tuple.
/
/
/
Z
/
/
/
/
/
/
/
/
Z
/
/
/
/
/
This definition suggests the way to overcome the violation of DC4, as stated in the following
lemma.
Lemma 6: Any DC4-violating tuple (q, t1, t2, e, Ei) is overcome, when:
1) e is excluded from Ei, (eligible, if it is passive in Ei and its exclusion respects EF 1−EF 4),
or
2) if ∃e′ ∈ (t1 ∪ t2)\(t1 ∩ t2), e′ is included in Ei; otherwise, e1 ∈ t1 and e2 ∈ t2, such
that e1 6= e2, are included in Ei. In these cases, {e′, Ei} and {{e1, e2}, Ei} are called
DC4-enforcing tuples.
Proof: See the proof in the Appendix.
Following examples illustrate the methods in Lemma 6 to enforce DC4.
Example 7: This example shows an automaton that is undecomposable due to a violation in
DC4, while DC4 can be enforced using both methods: event exclusion as well as event in-
clusion. Consider the task automaton AS:
with E1 =
e1
e2
a
b
/ •
/ •
/ •
/ •
•
(PPPPPP
a
•
e3
/ •
{a, b, e1, e3}, E2 = {a, b, e2}, 2 ∈ snda,b(1), 1 /∈ snda,b(2), leading to P1(AS):
/ •
e1
(PPPPPP
a
•
•
a
e3
/ •
/ •
b
, P2(AS):
/ •
/ •
a
(PPPPPP
a
•
•
b
/ •
e2
, and
/ •
•
b
/ •
e2
/ •
which is not bisimilar to AS, due to
2
i=1
Pi(AS) ∼=
e1
e3
/ •
a
•
a
%LLLLLL
a
•
/ •
•
2(AS) ∼=
violation of DC4 as there does not exist a deterministic automaton P ′
P2(AS). Here, (q0, t1 = e1, t2 = ε, a, E2) is a DC4-violating tuple. Since a is passive in E2 and
its exclusion from E2 keeps EF 1-EF 4 valid, according to Lemma 6, one way to enforce DC4
is exclusion of a from E2, resulting in E2 = {b, e2}, P2(AS):
/ • and
P1(AS)P2(AS) ∼= AS.
2(AS) such that P ′
/ •
/ •
e2
b
Another suggestion of Lemma 6 to overcome the DC4-violating tuple (q0, t1 = e1, t2 =
ε, a, E2) is addition of a communication link to prevent the nondeterminism in P2(AS). Since
there exists e1 that appears before a in t1 only, inclusion of e1 in E2 also enforces DC4 as with
Pi(AS) ∼= AS. For
new E2 = {a, b, e1, e2}, P2(AS):
and
e1
a
b
/ •
/ •
/ •
2
i=1
(PPPPPP
a
•
•
the cases that there does not exist an event b that appears before a in only one of the strings t1
or t2, according to Lemma 6, one needs to attach one event from each of two strings t1 and t2 in
/
/
/
(
/
/
/
/
/
/
/
(
/
/
/
/
/
/
(
/
/
/
/
/
/
/
%
/
/
/
/
/
/
/
/
/
(
/
/
Ei. For instance consider the DC4-violating tuple (t1 = e1e3e5, t2 = e5e3e1, a, E2) in Example
6, with no event that appears before a in (t1 ∪ t2)\(t1 ∩ t2). In that case {e1 ∈ t1, e5 ∈ t2} can
be included in E2 to make AS decomposable, as it was shown in Example 6.
Example 8: Example 7 showed a violation of DC4 that could be overcome using both method
in Lemma 6, namely, by link deletion and link addition. In Example 7, event a was a passive
shared event whose exclusion from E2 respected EF 1-EF 4, otherwise it was not allowed to
with E1 =
be excluded. If the task automaton was
e2
e1
a
b
/ •
/ •
/ •
/ •
•
(PPPPPP
a
•
e3
/ •
{a, b, e1, e3}, E2 = {a, b, e2}, then DC4 could not be enforced by exclusion of a from E2,
as EF 2 was violated since after this exclusion, no agent can handle the decision making on
the order of a and e2. Another constraint for link deletion is the passivity of the event. For
with E1 = {e1, a}, E2 = {e2, e4, a}.
example, consider A′
S:
e4
/ •
/ •
e1
(PPPPPP
e2
•
•
•
•
/ •
a
6nnnnnn
e2
e1
A′
S is not decomposable due to violation of DC4 in P1(AS):
/ •
e1
(PPPPPP
e1
•
•
a
. The
/ •
nondeterminism in P1(AS), and accordingly the DC4-violating tuple (q0, ε, e2, e1, E1), cannot
be removed by event exclusion since it occurs on e1 that is not a shared event. To enforce DC4
according to Lemma 6, e2 is required to be included into E1 that makes A′
S decomposable.
Another important issue for addition of communication link to enforce DC4 is that establishing
new communication link may lead to new violations of DC3 or DC4, as it is shown in the
following example.
Example 9: Assume the task automaton in Example 7 had a part as shown in he left hand
side of the initial state in AS:
•
d
•
e1
c
e5
vnnnnnn
e1
•
•
•
•
e1
(PPPPPP
a
•
•
a
e3
/ •
/ •
b
/ •
e2
/ •
with E1 = {a, b, c, d, e1, e3,
e5}, E2 = {a, b, c, d, e2}. Identical to Example 7, (q0, t1 = e1, t2 = ε, a, E2) is a DC4-violating
tuple and can be overcome by excluding a from E2, removing the nondeterminism on a in
P2(AS). However, unlike Example 7, including e1 into E2 (i.e., E2 = {a, b, c, d, e1, e2}), leads to a
,
new violation of DC4 in P2(AS): •
/ •
/ •
/ •
•
•
•
•
e2
e1
e1
a
d
c
b
vnnnnnn
e1
•
(PPPPPP
a
•
/
/
/
(
/
/
/
/
/
/
/
/
(
/
/
6
/
/
/
/
(
/
o
o
o
o
o
o
v
o
o
/
/
(
/
/
/
O
O
/
o
o
o
o
v
o
o
/
/
(
/
/
/
O
O
with a DC4-violating tuple (δ(q0, c), e5, ε, e1, E2), that in turn requires attachment of e5 to E2,
in order to enforce DC4.
If in this example, the order of e2 and b was reverse, i.e., the task automaton was A′
S:
with E1 = {a, b, c, d, e1, e3,
e1
e2
e1
e5
a
d
c
b
/ •
/ •
•
•
•
•
•
•
vnnnnnn
e1
(PPPPPP
a
•
/ •
•
/ •
e3
e5}, E2 = {a, b, c, d, e2}. Then as it was shown in Example 8, the DC4-violating tuple (q0, e1, ε,
a, E2) could not be dealt with exclusion of a from E2, due to EF 2, neither by inclusion of
e1 into E2 (since as mentioned above, it generates a new violation of DC4 that consequently
requires another inclusion of e5 into E2 to satisfy DC4).
Remark 4: Both Lemmas 5 and 6 provide sufficient conditions for resolving the violations
of DC3 and DC4, respectively. They do not however provide the necessary solutions, neither
the optimal solutions, as illustrated in the following example. We will show that for DC3 and
DC4, in general one requires to search exhaustively to find the optimal sequence of enforcing
tuples, to have minimum number of link additions. In this sense, instead of exhaustive search
for optimal solution, it is reasonable to introduce sufficient conditions to provide a trackable
procedure for a feasible solution to make an automaton decomposable.
Example 10: Consider a task automaton AS:
e6
•
•
b
•
e5
•
e7
•
•
e5
/ •
e3
/ •
a
/ •
e2
/ •
e1
hPPPPPP
vnnnnnn
e7
•
c
•
e1
6nnnnnn
(PPPPPP
e3
e8
b
e1
e5
e5
e1
a
e4
•
•
•
•
•
/ •
/ •
/ •
•
with local event sets E1 = {a, b, c, e1, e3, e5, e7} and E2 = {a, b, c, e2, e4, e6, e8}. AS is undecom-
posable due to DC3-violating tuples (e1e5e3ae2, e3e5e1ae4, a, E1, E2) and (e1e7e5be6, e7e5e1be8,
a, E1, E2) and DC4-violating tuples (q0, e1e5e3, e3e5e1, a, E2) and (δ(q0, c), e1e7e5, e7e5e1, b, E2).
According to Lemmas 5 and 6, two enforcing tuples {{e1, e3}, E2} and {{e1, e7}, E2} remove all
violations of DC3 and DC4. However, this solution is not unique, nor optimal, as the enforcing
tuple {{e1, e5}, E2} enforced DC3 and DC4 with minimum number of added communication
links.
/ •
E. Exhaustive search for optimal decompozabilization
Another difficulty is that enforcing the decomposability conditions using link deletion is limited
to passivity and EF 1-EF 4, and after deletions of redundant links (that are passive and their
o
o
o
o
o
o
v
o
o
/
/
(
/
/
/
O
O
/
o
o
o
o
o
o
o
o
/
/
/
/
h
v
o
o
6
(
o
o
o
o
o
o
o
o
/
/
/
/
deletion respect EF 1-EF 4), the only way to make the automaton decomposable is to establish
new communication links. Addition of new links, on the other hand, may lead to new violations
of DC3 or DC4 (as illustrated in Examples 5 and 9), and in turn may introduce new violations.
It means that, in general, resolution of decomposability conditions can dynamically result in new
violations of decomposability conditions, as it is elaborated in the following example.
Example 11: Consider the task automaton AS:
e10
b
•
•
•
•
d
e2
e12
•
•
•
e6
a
•
c O
•
e2
e4
yrrrrrr
e2
•
•
•
e2
f
%LLLLLL
e1
/ •
•
•
•
•
/ •
e4
9rrrrrr
e8
e2
with local event sets
/ •
/ •
e4
e3
/ •
/ •
g
e5
E1 = {a, b, c, d, f, g, e1, e3, e5} and E2 = {a, b, c, d, f, g, e2, e4, e6, e8, e10, e12}. This automaton
is undecomposable due to DC2-violating event pairs {(e1, e2), (e2, e3)} with the corresponding
enforcing tuples {e1, E2}, {e3, E2} and {e2, E1} and with the following possible sequences:
1) {e1, E2}; {e3, E2}: in this case AS becomes decomposable, without emerging new viola-
tions of decomposability conditions;
2) {e1, E2}; {e2, E1}; {{e4, e6}, E1}; {e8, E1}: if after including e1 in E2, e2 is included in
E1, then two DC4-violating tuples (δ(q0, a), ε, e4, e2, E1) and (δ(q0, c), ε, e6, e2, E1) emerge
that in turn require {e4, e6} to be attached to E1. Inclusion of e4 in E1, on the other hand,
introduces another DC4-violating tuple (δ(q0, f ), ε, e8, e4, E1) that calls for attachment of
e8 to E1; similarly
3) {e3, E2}; {e1, E2};
4) {e3, E2}; {e2, E1}; {{e4, e6}, E1}; {e8, E1}, and
5) {e2, E1}; {{e4, e6}, E1}; {e8, E1}.
In this example, the first and the third sequences, i.e., {{e1, e3}, E2} gives the optimal choice
with minimum number of added communication links, although initially {e2, E1} sought to offer
the optimal solution.
Therefore, in general an optimal solution to Problem 1 will be obtained through an exhaustive
search, using Lemmas 4, 5 and 6, as state in the following algorithm.
Algorithm 2:
1) For any local event set, exclude any passive event whose exclusion respects EF 1-EF 4;
2) identify all DC1&2-violating tuples, DC3-violating tuples and DC4-violating tuples and
o
o
o
o
o
o
o
o
/
o
o
o
o
o
o
o
o
y
o
o
/
/
%
O
/
/
9
/
/
O
O
/
/
/
their respective enforcing tuples;
3) among all enforcing tuples, find the one that corresponds to the most violating tuples;
4) if applying of the enforcing tuples with maximum number of violating tuples, does not
impose new violations of DC3 or DC4, then apply it, go to Step 3 and continue until
there is no violating tuples; otherwise, do the exhaustive search to find the sequence of
link additions with minimum number of added links.
5) end.
n
∪
i=1
Lemma 7: Consider a deterministic task automaton AS with local event sets Ei such that
Ei. If AS is not decomposable with respect to parallel composition and natural projections
E =
Pi, i = 1, ..., n, Algorithm 2 optimally makes AS decomposable, with minimum number of
communication links.
Proof: See the proof in the Attachment.
Remark 5: (Special case: Automata with mutual exclusive branches) When branches of AS
share no events (i.e. ∀q ∈ Q, s, s′ ∈ E∗, δ(q, s)!, δ(q, s′)!, s ≮ s′, s′ ≮ s: s ∩ s′ = ∅), due to
definition of DC3 and DC4 in Lemma 1 DC3 and DC4 are trivially satisfied, and moreover,
since branches from any state share no event, then Algorithm 2 is reduced to Algorithm 1.
F. Feasible solution for task decomposabilization
As Example 11 showed that, in general, additions of communication links may successively
introduce new violations of decomposability conditions, for which new links should be estab-
lished. Therefore, in general an optimal solution to Problem 1 requires an exhaustive search,
using Lemmas 4, 5 and 6. Moreover, checking of DC3 and DC4 is a nontrivial task, while
it has to be accomplished initially as well as upon each link addition. It would be therefore
very tractable if we can define a procedure to make DC3 and DC4 satisfied, without their
examination. Following result takes an automaton whose DC1 and DC2 are made satisfied
using Algorithm 1, and proposes a sufficient condition to fulfill DC3 and DC4.
Lemma 8: Consider a deterministic automaton AS, satisfying DC1 and DC2. AS satisfies
DC3 and DC4 if following steps are accomplished on AS:
1) ∀s1, s2 ∈ E∗, s1 ≮ s2, s2 ≮ s1, q, q1, q2 ∈ Q, δ(q, s1) = q1 6= δ(q, s2) = q2, [∄e1, e2 ∈
E, e1e2 6 s1, e2e1 6 s2, ∀t ∈ E∗, δ(q, e1e2t)! ⇔ δ(q, e2e1t)!], ∃e ∈ s1 ∩ s2,, then
∀i ∈ loc(e), ∀e′ ∈ {e1 6 t1, e2 6 t2}, e′ appears before e, include e′ in Ei.
2) go to Step 1 and continue until ∀s1, s2 ∈ E∗, s1 ≮ s2, s2 ≮ s1, q, q1, q2 ∈ Q, δ(q, s1) =
q1 6= δ(q, s2) = q2, ∃e ∈ s1∩s2, [∄e1, e2 ∈ E, e1e2 6 s1, e2e1 6 s2, ∀t ∈ E∗, δ(q, e1e2t)! ⇔
δ(q, e2e1t)!], then ∀i ∈ loc(e), Ei contains the first events of s1 and s2, that appear before
e.
Proof: See the proof in the Attachment.
Remark 6: The condition in Lemma 8 intuitively means that for any two strings s1, s2 from
any state q, sharing an event e, all agents who know this event e will be able to distinguish two
strings, if they know the first event of each string. The ability of those agents that know this
event e to distinguish strings s1 and s2, prevents illegal interleavings (to enforce DC3) and local
nondeterminism (to satisfy DC4). The significance of this condition is that it does not require to
check DC3 and DC4, instead provides a tractable (but more conservative) procedure to enforce
DC3 and DC4. The expression s1 ≮ s2, s2 ≮ s1 in the lemma, is to exclude the pairs of strings
that one of them is a substring of the other, as their language product does not exceed from
the strings of AS, provided DC1 and DC2. Moreover, the expression [∄e1, e2 ∈ E, e1e2 6 s1,
e2e1 6 s2, ∀t ∈ E∗, δ(q, e1e2t)! ⇔ δ(q, e2e1t)!] in this lemma excludes the pairs of strings e1e2t
and e2e1t from any q ∈ Q that have been already checked using DC1 and DC2 and do not
form illegal interleaving strings, and hance, do not need to include e1 in the local event sets of
e2 and vice versa (see Example 12).
Combination of Lemmas 4 and 8 leads to the following algorithm as a sufficient condition
to make a deterministic task automaton decomposable. Following algorithm uses Lemma 4 to
enforce DC1 and DC2 followed by Lemma 8 to overcome the violations of DC3 and DC4.
Algorithm 3:
1) For a deterministic automaton AS, with local event sets Ei, i = 1, . . . , n, ∀Ei ∈ {E1, . . . , En},
E0
i = Ei\{e ∈ Eie is passive in Ei and exclusion of e from Ei does not violate EF 1-
EF 4};
2) form the DC1&2-Violating graph ; set V 0(AS) = V (AS); W 0(AS) = W (AS); G0(AS) =
(W (AS), Σ); k=1;
3) Among all events in the nodes in W k−1(AS), find e with the maximum W k−1
e
(AS, Ek−1
i
),
for all Ek−1
i = Ek−1
5) update W k
4) Ek
i
i
∈ {Ek−1
1
, . . . , Ek−1
n };
∪ {e}; and delete all edges from e to Ek
i ;
e (AS, Ei) for all nodes of G(AS);
6) set k = k + 1 and go to step (3);
7) continue, until there exist no edges.
8) ∀s1, s2 ∈ E∗, s1 ≮ s2, s2 ≮ s1, q, q1, q2 ∈ Q, δ(q, s1) = q1 6= δ(q, s2) = q2, [∄e1, e2 ∈
E, e1e2 6 s1, e2e1 6 s2, ∀t ∈ E∗, δ(q, e1e2t)! ⇔ δ(q, e2e1t)!], ∃e ∈ s1 ∩ s2, then ∀i ∈
loc(e), ∀e′ ∈ {e1 6 t1, e2 6 t2}, e′ appears before e, include e′ in Ei.
9) go to Step 1 and continue until ∀s1, s2 ∈ E∗, s1 ≮ s2, s2 ≮ s1, q, q1, q2 ∈ Q, δ(q, s1) =
q1 6= δ(q, s2) = q2, ∃e ∈ s1∩s2, [∄e1, e2 ∈ E, e1e2 6 s1, e2e1 6 s2, ∀t ∈ E∗, δ(q, e1e2t)! ⇔
δ(q, e2e1t)!], then ∀i ∈ loc(e), Ei contains the first events of s1 and s2, that appear before
e.
Based on this formulation, a solution to Problem 1 is given as the following theorem.
Theorem 1: Consider a deterministic task automaton AS with local event sets Ei such that E =
n
Ei. If AS is not decomposable with respect to parallel composition and natural projections Pi,
∪
i=1
i = 1, ..., n, Algorithm 3 makes AS decomposable. Moreover, if after Step 7, DC3 and DC4 are
satisfied, then the algorithm makes AS decomposable, with minimum number of communication
links.
Proof: After excluding the redundant shared events in the first step, the algorithm enforces
DC1 and DC2 in Steps 2 to 7, according to Lemma 4 and deals with DC3 and DC4 in Steps
8 and 9, based on Lemma 8.
Remark 7: If after Step 7, no violation of DC3 or DC4 is reported in the automaton, then
AS is made decomposable with minimum number of added communication links; otherwise, the
optimal solution can be obtained through exhaustive search by examining the number of added
links for any possible sequence of enforcing tuples, using Lemmas 5 and 6, as it was presented
in Lemma 7. To avoid the exhaustive search the algorithm provides a sufficient condition to
enforce DC3 and DC4 in Steps 8 and 9, according to Lemma 8. The algorithm terminates, due
to finite number of states and events, and the fact that at the worst case, when all events are
shared among all agents, the task automaton is trivially decomposable.
Example 12: Consider a task automaton
AS:
e6
e7
c
e2
•
•
•
•
•
•
•
•
e9
e12
•
•
e7
\:::::::
e6
e2
e8
•
•
b
a
\:::::::
B
f
•
a
e11
C
8888888
e3
•
•
e2
e4
B
•
d
•
e10
:::::::
e10
/ •
d
/ •
e1
:::::::
e5
•
•
•
•
with local event sets E1 = {a, b, c, d, f, e1, e3, e5, e7, e9, e11} and E2 = {a, b, c, d, f, e2, e4, e6, e8,
e10, e12}, with the communication pattern 2 ∈ snda,b,c,d(1) and no more communication links.
This task automaton is not decomposable, due to the set of DC1&2-violating tuples {e1, e2},
{e1, e4}, {e2, e3}, {e2, e5}, {e3, e4}, {e4, e5}, DC3-violating tuples (e11ade10, ae7e6, a, E1, E2),
(e11ade10, ae6e7, a, E1, E2), (e11ae10d, ae7e6, a, E1, E2), (e11ae10d, ae6e7, a, E1, E2) and DC4-
violating tuple (q0, e11, ε, a, E2). There is also one event d that is redundantly shared with E2 as
d is passive in E2 and its exclusion respects EF 1-EF 4. Therefore, at the first step, the algorithm
excludes d from E2.
Next step is to construct the DC1&2-Violating graph and remove its edges by sharing one node
from each edge. The set of DC1&2-Violating event pair is obtained as V 0(AS) = {{e1, e2}, {e1, e4},
{e2, e3}, {e2, e5}, {e3, e4}, {e4, e5}} with W 0(AS) = {e1, e2, e3, e4, e5}. It can be seen that the
private events d, e6, e7, e8, e9, e10, e11, e12, and shared events a, b, c, f are not included in W 0(AS)
as they have no contribution in violation of DC1 and DC2. The DC1&2-Violating graph is
shown in Figure 4(a).
i
e
(AS, Ek−1
i
(AS, Ek−1
The maximum W k−1
) is formed by {e2, e4} with respect to E1 (here, since the
system has only two local event sets W k−1
) coincides to the valency of e in the
e
1 = {a, b, c, d, e1, e3, e5, e7, e9, e11, e2}) and removing
graph). Marking e2, including it to E1 (E1
i ) (valencies) are shown in Figure 4(b).
its incident edges to E1 and updating the W k
1 = {a, b, c, d, e1, e3, e5, e7, e9, e11, e2, e4}) with the
The next step will include e4 in E1 (E2
highest W k
i ) will
accomplish enforcing of DC1 and DC2 upon Step 7, as it is illustrated in Figure 4 (c). If from
the first stage e4 was chosen instead of e2, the procedure was similarly performed as depicted
in Figures 4 (d) and (e), resulting the same set of private events {e2, e4} to be shared with
E1. Inclusion of e2 in E1, however, introduces a new DC4-violating tuple (δ(q0, b), ε, e8, e2, E1)
i ) and removing its incident edges to E1 and updating the W k
e (AS, Ek
e (AS, Ek
e (AS, Ek
o
o
/
/
/
o
o
o
o
\
O
O
/
o
o
o
o
o
o
o
o
\
/
/
C
/
/
/
/
o
o
o
o
B
B
Fig. 4.
Illustration of enforcing DC 1 and DC 2 in Example 12, using Algorithm 3.
that will be automatically overcome in Step 8 by sharing e8 ∈ s1 = e8e2e12 (as s1 = e8e2e12
together with s2 = e2ce9 evolve from δ(q0, b), sharing e2 ∈ s1 ∩ s2) in all local event sets of
e2, i.e., by including e8 into E1. Similarly, inclusion of e11 in E2 overcomes DC4-violating
tuple (q0, e11, ε, a, E2). It is worth noting that the expression "∄e1, e2 ∈ E, e1e2 6 s1, e2e1 6 s2,
∀t ∈ E∗, δ(q, e1e2t)! ⇔ δ(q, e2e1t)!" in Step 8 prevents unnecessary inclusion of e10 in E1
as well as e7 in E2 and e6 in E1 (e6 and e7 satisfy DC1-DC2 and e10 and d satisfy EF 1-
EF 2). The algorithm terminates in this stages, leading to decomposability of AS, with E3
{a, b, c, d, e1, e3, e5, e7, e9, e11, e2, e4, e8}, E3
2 = {a, b, c, e2, e4, e6, e8, e10, e11, e12}.
2 = E2, E3
1 =
IV. CONCLUSIONS
The paper proposed a method for task automaton decomposabilization, applicable in top-
down cooperative control of distributed discrete event systems. This result is a continuation of
our previous works on task automaton decomposition [19], [20], and fault-tolerant cooperative
tasking [21], and investigates the follow-up question to understand that how an originally un-
decomposable task automaton can be made decomposable, by modifying the event distribution
among the agents.
First, using the decomposability conditions the sources of undecomposability are identified
and then a procedure was proposed to establish new communication links in order to enforce
the decomposability conditions. To avoid the exhaustive search and the difficulty of checking
of decomposability conditions in each step, a feasible solution was proposed as a sufficient
condition that can make any deterministic task automaton decomposable.
V. APPENDIX
A. Proof of Lemma 4
Following lemma will be used during the proof.
Lemma 9: Consider two non-increasing chains ai, bi, i = 1, ..., N, such that a1 ≥ a2 ≥ ... ≥
bi implies that ∃k ∈ {1, ..., N} such that
ai <
aN > 0, b1 ≥ b2 ≥ ... ≥ bN > 0. Then
ak < bk.
N
Σ
i=1
N
Σ
i=1
N
Σ
i=1
ai <
N
Σ
i=1
ai ≥
N
Σ
i=1
Proof: Suppose by contradiction that
bi which contradicts to the hypothesis, and the proof is followed.
bi, but, ∄k ∈ {1, ..., N} such that ak < bk.
Then, ∀k ∈ {1, ..., N} : ak ≥ bk. Therefore, since ak, bk > 0, ∀k ∈ {1, ..., N}, it results in
N
Σ
i=1
Now, we prove Lemma 4 as follows. In each iteration k for the event e and local event set
Ei with maximum W k−1
), all edges from e to Ei are deleted. Denoting the set of
deleted edges in k − th iterations by ∆Σk, in each iteration k, some elements of Σk−1 are moved
(AS, Ek−1
i
e
into ∆Σk until after K iterations, there is no more elements in ΣK to be moved into a new set.
k=1, as {∆Σk} ∩ {∆Σl} = ∅,
This iterative procedure leads to a partitioning of Σ by {∆Σk}K
∀k, l = {1, ..., K}, k 6= l and
K
∪
k=1
∆Σk = Σ. The latter equality leads to
Now, we want to prove that
K
Σ
k=1
∆Σk = Σ
∆Σk = ∆Σkmax, ∀k ∈ {1, ..., K} ⇒ K = Kmin
(1)
(2)
Here, K is the total number of iterations that is also equal to the number of added com-
munication links to remove violations of DC1 and DC2. In this sense, K is desired to be
minimized.
The proof of (2) is by contradiction as follows. Suppose that ∆Σk = ∆Σkmax, ∀k ∈
k=1, with K ′ < K
{1, ..., K}, but, K 6= Kmin, i.e., there exists another partitioning {∆′Σk}K ′
partitions, leading to
K ′
Σ
k=1
∆′Σk = Σ
In this case, from (1) and (3), we have
K
Σ
k=1
∆Σk =
K ′
Σ
k=1
∆Σk +
K
Σ
k=K ′+1
∆Σk =
K ′
Σ
k=1
∆′Σk.
Since ∆Σk > 0, ∀k ∈ {1, ..., K}, then
K
Σ
k=K ′+1
∆Σk > 0, then, (4) results in
K ′
Σ
k=1
∆Σk <
K ′
Σ
k=1
∆′Σk.
(3)
(4)
(5)
Moreover, since ∆Σk > 0, ∆′Σk > 0, ∀k ∈ {1, ..., K}, then (5) together with Lemma 9
imply that ∃k ∈ {1, ..., K ′} ⊆ {1, ..., K}, i.e., ∆Σk < ∆′Σk, i.e., ∃k ∈ {1, ..., K} such that
∆Σk 6= ∆Σkmax, which contradicts to the hypothesis, and hence, (2) is proven. Moreover,
if automaton AS has no violations of DC3 and DC4 before and during the iterations, then
the algorithm make it decomposable with the minimum number of added communication links,
since the problem of making decomposable is reduced to optimal enforcing of DC1 and DC2.
B. Proof for Lemma 5
For any DC3-violating tuple (s1, s2, a, Ei, Ej), exclusion of a from Ei or Ej, excludes a from
Ei ∩ Ej, leading to pEi∩Ej (s1) and pEi∩Ej (s1) do not start with a, and hence (s1, s2, a, Ei, Ej)
will no longer act as a DC3-violating tuple.
For the second method in this lemma, firstly ∀q ∈ Q, s1, s2 ∈ E∗, δ(q, s1)!, δ(q, s2)!, pEi∩Ej (s1)
and pEi∩Ej (s2) start with a, such that (s1, s2, a, Ei, Ej) is a DC3-violating tuple, ∃b ∈ (Ei ∪
Ej)\(Ei ∩ Ej) such that b appears before a in s1 or s2 (since AS is deterministic and pEi∩Ej (s1)
and pEi∩Ej (s2) start with a).
Two cases are possible, here: b appears in only one of the strings s1 or s2; or b appears in
both strings. If b appears before a in only of the strings, then without loss of generality, assume
1 ∈ Qi × Qj, ω1, ω2 ∈ [(Ei ∪ Ej)\(Ei ∩ Ej)]∗,
that b belongs to only s1, and hence, ∃q, q1, q2, q′
1) = q1,
1 ∈ (Ei ∪ Ej)∗, a ∈ Ei ∩ Ej such that δi,j(q, ω1) = q′
ω′
δi,j(q1, a)!, δi,j(q, ω2) = q2, δi,j(q2, a)!, where, δi,j is the transition relation in Pi(AS)Pj(AS).
Now, due to synchronization constraint in parallel composition, inclusion of b in Ei ∩ Ej means
1, b) = q′′
1, δi,j(q′
1 , δi,j(q′′
1 , ω′
1, q′′
1 ], y) and (x, [q′′
1 ]j) are accessible in Pi(AS)Pj(AS) only if y = [q′′
1 ]i,
that ([q′′
respectively. This means that ([q1]i, [q2]j) and ([q2]i, [q1]j) are not accessible in Pi(AS)Pj(AS),
and hence, pi(s1)pj(s2) and pi(s2)pj(s1) cannot evolve after a, and therefore, do not generate
illegal strings out of the original strings, implying that (s1, s2, a, Ei, Ej) will no longer remain
a DC3-violating tuple.
1 ]j and x = [q′′
1, b) = q′′
On the other hand, if b appears before a, in both strings s1 and s2, then ∃q, q1, q2, q′
2, q′′
2 ∈
2 ∈ (Ei∪Ej)∗, a ∈ Ei∩Ej such that δi,j(q, ω1) = q′
1,
Qi×Qj, ω1, ω2 ∈ [(Ei∪Ej)\(Ei∩Ej)]∗, ω′
2) = q2,
δi,j(q′
δi,j(q2, a)!, that leads to accessibility of ([q′
1]j) as well as ([q1]i, [q2]j) and
([q2]i, [q1]j) in Pi(AS)Pj(AS), that means that although (s1, s2, a, Ei, Ej) is no longer a DC3-
violating tuple, (s1, s2, b, Ei, Ej) emerges as a new DC3-violating tuple.
1) = q1, δi,j(q1, a)!, δi,j(q, ω2) = q′
2]j) and ([q′
2, δi,j(q′
2]i, [q′
2, b) = q′′
2, δi,j(q′′
1, δi,j(q′′
1 , ω′
1, q′′
1 , q′
2 , ω′
1, ω′
1]i, [q′
In this case (when ∄b ∈ (Ei ∪ Ej)\(Ei ∩ Ej) that appears before a in only one of the
strings s1 or s2), instead of inclusion of b in Ei ∩ Ej, if two different events that appear before
a in strings pEi∪Ej (s1) and pEi∪Ej (s1) are attached to Ei ∩ Ej, it leads to ∃q, q1, q2, q3, q4 ∈
Qi × Qj, ω1, ω2, ω′
2 ∈ [(Ei ∪ Ej)\(Ei ∩ Ej)]∗, e1, e2, a ∈ Ei ∩ Ej such that δi,j(q, ω1e1) = q1,
2) = q4, δi,j(q4, a)!. Consequently, due
δi,j(q1, ω′
to synchronization constraint in parallel composition, ([q1]i, [q]j), ([q]i, [q1]j), ([q2]i, [q]j) and
([q]i, [q2]j), and hence, ([q3]i, [q4]j) and ([q4]i, [q3]j) are not accessible in Pi(AS)Pj(AS), i.e.,
no more DC3-violating tuples form on strings s1 and s2.
1) = q3, δi,j(q3, a)!, δi,j(q, ω2e2) = q2, δi,j(q2, ω′
1, ω′
C. Proof for Lemma 6
For any DC4-violating tuple (q, t1, t2, e, Ei), with q, q1, q2 ∈ Q, t1, t2 ∈ (E\Ei)∗, e ∈ Ei,
δ(q, t1) = q1 6= δ(q, t2) = q2, exclusion of e from Ei leads to pi(e) = ε, and pi(t1e) = pi(t2e) = ε,
[q]i = [δ(q1, e)]i = [δ(q2, e)]i, and hence, (q, t1, t2, e, Ei) will no longer behave as a DC4-
violating tuple. However, it should be noted that it may cause another nondeterminism on an
event after e, and this event exclusion is allowed only if e is passive in Ei and the exclusion
does not violate EF 1 − EF 4.
For the second method, i.e., event inclusion, if ∃e′ ∈ (t1 ∪ t2)\(t1 ∩ t2), then without loss
1 ∈ Q, t1, t2 ∈ (E\Ei)∗, e ∈ Ei,
1. In this case, inclusion of e′ in Ei leads to pi(t1e) =
6= [q2]i, i.e., in Pi(AS), t1 and t2 will no
of generality, assume that e′ ∈ t1\t2 such that ∃q, q1, q2, q′
δ(q, t1) = q1 6= δ(q, t2) = q2, δ(q′
e′e, while pi(t2e) = e, and therefore, [q1]i = [q′′
1 ]i
1, e′) = q′′
1, q′′
longer cause a nondeterminism on e from q, and accordingly, (q, t1, t2, e, Ei) will not remain a
DC4-violating tuple.
If however ∄e′ ∈ (t1 ∪t2)\(t1 ∩t2), i.e., ∀e′ ∈ (t1 ∪t2), e′ ∈ (t1 ∩t2), then inclusion of any such
e′ generates a DC4-violating tuple (q, t1, t2, e′, Ei). In this case, Lemma 6 suggests to take two
different events that appear before e, one from t1 and the other from t2, and include them into
2 ∈ Q, e1 ∈ t1, e2 ∈ t2, e1 6= e2, δ(q, t1) = q1 6= δ(q, t2) = q2,
Ei such that ∃q, q1, q2, q′
2, q′′
1, q′
2. Thus, including e1 and e2 in Ei results in pi(t1) = e1, pi(t2) = e2,
2, e2) = q′′
δ(q′
6= δi([q]i, t2) = [q2]i, meaning that (q, t1, t2, e, Ei) is not a DC4-violating
1, e1) = q′′
1 , q′′
1, δ(q′
δi([q]i, t1)! = [q1]i
tuple anymore.
D. Proof for Lemma 7
The algorithm starts with excluding events from local event sets in which the events are passive
and their exclusion do not violate EF 1-EF 4. From this stage onwards the decomposability
conditions are no longer allowed to be enforced by link deletion, whereas the algorithm removes
the violations of decomposability conditions by establishing new communication links. Next, the
algorithm applies violating tuples in the order of corresponding number of violating tuples. If
no new violations of decomposability conditions emerge during conducting of enforcing tuples,
then the algorithm decomposes the task automaton with minimum number of communication
links, similar to the proof of Lemma 4, since iterations partition the set of violating tuples, and
applying of enforcing tuples (based on Lemmas 4, 5 and 6) with maximum number of violating
tuples in each iteration gives maximum number of resolutions per link addition that leads to
the minimum number of added communication links. The algorithm will terminate due to finite
number of states and events and at the worst case all events are shared among all agents to make
the automaton decomposable.
E. Proof for Lemma 8
Denoting the expression , "∀Ei, Ej ∈ {E1, . . . , En}, i 6= j, a ∈ Ei ∩ Ej, s = t1at′
2,
1, s′ = t2at′
pEi∩Ej (t1) = pEi∩Ej (t2) = ε" as A, and the expression "δ(q0,
pi (si))! for any ∀{s1, · · · , sn} ⊆
L(AS), s, s′ ∈ {s1, · · · , sn}" as B, the condition DC3 can be written as A ⇒ B. Now, if
∀s1, s2 ∈ E∗, s1 ≮ s2, s2 ≮ s1, q, q1, q2 ∈ Q, δ(q, s1) = q1 6= δ(q, s2) = q2, [∄e1, e2 ∈ E, e1e2 6
s1, e2e1 6 s2, ∀t ∈ E∗, δ(q, e1e2t)! ⇔ δ(q, e2e1t)!], ∃e ∈ s1∩s2, any e′ ∈ {e1 6 t1, e2 6 t2}, such
n
i=1
that e′ appears before e, is included in Ei, ∀i ∈ loc(e), it follows that ∀Ei, Ej ∈ {E1, . . . , En},
i 6= j, a ∈ Ei ∩ Ej, s = t1at′
2, δ(q0, s)! 6= δ(q0, s′)!, a ∈ s ∩ s′, then the first event of
t1 and t2 belong to Ei ∩ Ej, i.e., A (the antecedent of DC3) becomes false, and hence, A ⇒ B
( DC3 ) holds true. Therefore, the procedure in Lemma 8 gives a sufficient conditions to make
1, s′ = t2at′
DC3 always true.
i , δ(x, e) = x1 6= δ(x, e) = x2" and "∀t ∈ E∗
It is similarly a sufficient condition for DC4 as follows. Let the expressions "∀i ∈ {1, . . . , n},
x, x1, x2 ∈ Qi, e ∈ Ei, t ∈ E∗
i : δ(x1, t)! ⇔ δ(x2, t)!"
to be denoted as C and D, respectively. In this case, DC4 can be expressed as C ⇒ D. Then,
for a deterministic automaton AS, if ∀s1, s2 ∈ E∗, s1 ≮ s2, s2 ≮ s1, q, q1, q2 ∈ Q, δ(q, s1) =
q1 6= δ(q, s2) = q2, [∄e1, e2 ∈ E, e1e2 6 s1, e2e1 6 s2, ∀t ∈ E∗, δ(q, e1e2t)! ⇔ δ(q, e2e1t)!],
∃e ∈ s1 ∩ s2, the first event of s1 and s2 are included in all local event sets that contain e,
it results in ¬C(i.e., the antecedent of DC4 becomes false, and consequently, DC4 becomes
always true), since in such case ∀Ei ∈ {E1, . . . , En}, t1, t2 ∈ E∗, q, q1, q2 ∈ Q, e ∈ Ei,
δ(q, t1e) = q1 6= δ(q, t2e) = q2, then ¬[pi(t1) = pi(t2) = ε].
Expression "[∄e1, e2 ∈ E, e1e2 6 s1, e2e1 6 s2, ∀t ∈ E∗, δ(q, e1e2t)! ⇔ δ(q, e2e1t)!]", in
Lemma 8 is to exclude those pairs of strings s1 and s2 that start with e1e2 and e2e1, respectively,
as they have been already checked with DC1 and DC2 and their interleaving does not impose
illegal strings.
REFERENCES
[1] V. R. Lesser, "Cooperative multiagent systems: A personal view of the state of the art," IEEE Transactions on Knowledge
and Data Engineering, vol. 11, pp. 133 -- 142, 1999.
[2] P. U. Lima and L. M. Custdio, Multi-Robot Systems, Book Series Studies in Computational Intelligence, Book Innovations
in Robot, Mobility and Control. Berlin: Springer Berlin / Heidelberg, 2005, vol. 8.
[3] J. Choi, S. Oh, and R. Horowitz, "Distributed learning and cooperative control for multi-agent systems," Automatica,
vol. 45, no. 12, pp. 2802 -- 2814, 2009.
[4] E. Semsar-Kazerooni and K. Khorasani, "Multi-agent team cooperation: A game theory approach," Automatica, vol. 45,
no. 10, pp. 2205 -- 2213, 2009.
[5] G. E. Fainekos, A. Girard, H. Kress-Gazit, and G. J. Pappas, "Temporal logic motion planning for dynamic robots,"
Automatica, vol. 45, no. 2, pp. 343 -- 352, 2009.
[6] Z. Ji, Z. Wang, H. Lin, and Z. Wang, "Brief paper: Interconnection topologies for multi-agent coordination under leader-
follower framework," Automatica, vol. 45, no. 12, pp. 2857 -- 2863, 2009.
[7] P. Tabuada and G. Pappas, "Linear time logic control of discrete-time linear systems," Automatic Control, IEEE Transactions
on, vol. 51, no. 12, pp. 1862 -- 1877, Dec. 2006.
[8] C. Belta, A. Bicchi, M. Egerstedt, E. Frazzoli, E. Klavins, and G. Pappas, "Symbolic planning and control of robot motion
[grand challenges of robotics]," Robotics and Automation Magazine, IEEE, vol. 14, no. 1, pp. 61 -- 70, March 2007.
[9] M. Kloetzer and C. Belta, "Automatic deployment of distributed teams of robots from temporal logic motion specifications,"
Robotics, IEEE Transactions on, vol. 26, no. 1, pp. 48 -- 61, feb. 2010.
[10] M. G. Hinchey, J. L. Rash, W. Truszkowski, C. Rouff, and R. Sterritt, "Autonomous and autonomic swarms," in Software
Engineering Research and Practice, 2005, pp. 36 -- 44.
[11] W. Truszkowski, M. Hinchey, J. Rash, and C. Rouff, "Autonomous and autonomic systems: a paradigm for future space
exploration missions," Systems, Man, and Cybernetics, Part C: Applications and Reviews, IEEE Transactions on, vol. 36,
no. 3, pp. 279 -- 291, May 2006.
[12] M. Kloetzer and C. Belta, "Temporal logic planning and control of robotic swarms by hierarchical abstractions," Robotics,
IEEE Transactions on, vol. 23, no. 2, pp. 320 -- 330, April 2007.
[13] V. Crespi, A. Galstyan, and K. Lerman, "Top-down vs bottom-up methodologies in multi-agent system design," Auton.
Robots, vol. 24, no. 3, pp. 303 -- 313, 2008.
[14] Y. Willner and M. Heymann, "Supervisory control of concurrent discrete-event systems," International Journal of Control,
vol. 54, pp. 1143 -- 1169, 1991.
[15] K. Cai and W. Wonham, "Supervisor localization: a top-down approach to distributed control of discrete-event systems,"
Automatic Control, IEEE Transactions on, vol. 55, no. 3, pp. 605 -- 618, 2010.
[16] X. Koutsoukos, P. Antsaklis, J. Stiver, and M. Lemmon, "Supervisory control of hybrid systems," Proceedings of the IEEE,
vol. 88, no. 7, pp. 1026 -- 1049, Jul 2000.
[17] C. G. Cassandras and S. Lafortune, Introduction to discrete event systems. USA: Springer, 2008.
[18] R. Kumar and V. K. Garg, Modeling and Control of Logical Discrete Event Systems. Norwell, MA, USA: Kluwer
Academic Publishers, 1999.
[19] M. Karimadini and H. Lin, "Guaranteed global performance through local coordinations," Automatica, vol. 47, no. 5, pp.
890 -- 898, 2011.
[20] -- -- , "Necessary and sufficient conditions for task automaton decomposition," submitted for publication, online available
at: http://arxiv.org/abs/1101.2002, 2011.
[21] -- -- , "Fault-tolerant cooperative tasking for multi-agent systems," submitted for publication, online available at:
http://arxiv.org/abs/1101.2003v2, 2011.
[22] A. P. Moore, "The specification and verified decomposition of system requirements using csp," IEEE Transactions on
Software Engineering, vol. 16, pp. 932 -- 948, 1990.
[23] K. Go and N. Shiratori, "A decomposition of a formal specification: An improved constraint-oriented method," IEEE Trans.
Softw. Eng., vol. 25, no. 2, pp. 258 -- 273, 1999.
[24] A. Arora and S. Kulkarni, "Component based design of multitolerant systems," Software Engineering, IEEE Transactions
on, vol. 24, no. 1, pp. 63 -- 78, Jan. 1998.
[25] M. Hultstrom, "Structural decomposition," in PSTV, 1994, pp. 201 -- 216.
[26] T. Cao and A. Sanderson, "Task decomposition and analysis of robotic assembly task plans using petri nets," Industrial
Electronics, IEEE Transactions on, vol. 41, no. 6, pp. 620 -- 630, dec 1994.
[27] D. A. Zaitsev, "Decomposition of petri nets," Cybernetics and Sys. Anal., vol. 40, no. 5, pp. 739 -- 746, 2004.
[28] R. Morin, "Decompositions of asynchronous systems," in CONCUR '98: Proceedings of the 9th International Conference
on Concurrency Theory. London, UK: Springer-Verlag, 1998, pp. 549 -- 564.
[29] M. Mukund, From global specifications to distributed implementations, in B. Caillaud, P. Darondeau, L. Lavagno (Eds.),
Synthesis and Control of Discrete Event Systems, Kluwer. Berlin: Springer Berlin / Heidelberg, 2002.
[30] S. Kiyamura, Y. Takata, and H. Seki, "Process decomposition via synchronization events and its application to counter-
process decomposition," in PPAM, 2003, pp. 298 -- 305.
[31] C. Zhou, R. Kumar, and S. Jiang, "Control of nondeterministic discrete-event systems for bisimulation equivalence,"
Automatic Control, IEEE Transactions on, vol. 51, no. 5, pp. 754 -- 765, may 2006.
[32] R. Alur, T. Henzinger, G. Lafferriere, and G. Pappas, "Discrete abstractions of hybrid systems," Proceedings of the IEEE,
vol. 88, no. 7, pp. 971 -- 984, jul 2000.
[33] C. Godsil and G. Royle, Algebraic Graph Theory. New York: Springer, 2001.
|
1109.2129 | 1 | 1109 | 2011-09-09T20:21:45 | Extremal Behaviour in Multiagent Contract Negotiation | [
"cs.MA"
] | We examine properties of a model of resource allocation in which several agents exchange resources in order to optimise their individual holdings. The schemes discussed relate to well-known negotiation protocols proposed in earlier work and we consider a number of alternative notions of rationality covering both quantitative measures, e.g. cooperative and individual rationality and more qualitative forms, e.g. Pigou-Dalton transfers. While it is known that imposing particular rationality and structural restrictions may result in some reallocations of the resource set becoming unrealisable, in this paper we address the issue of the number of restricted rational deals that may be required to implement a particular reallocation when it is possible to do so. We construct examples showing that this number may be exponential (in the number of resources m), even when all of the agent utility functions are monotonic. We further show that k agents may achieve in a single deal a reallocation requiring exponentially many rational deals if at most k-1 agents can participate, this same reallocation being unrealisable by any sequences of rational deals in which at most k-2 agents are involved. | cs.MA | cs | a f Ai(cid:12)
ia e ige
e Reea
h 23 2005 41{78
S bied 07/04; b ihed 01/05
Reea
h e
Exea Behavi i iage Ca
egiai
a E. D e
ed
. iv.a
. k
Deae f C e S
ie
e
The Uiveiy f ive ive U
Aba
We exaie eie f a de f e
e a
ai i whi
h evea age ex
hage e
e i de iie hei idivid a h dig. The
hee di
ed e
ae we kw egiai
ed i ea ie wk ad we
ide a be
f a eaive i f \aia iy"
veig bh aiaive ea e e.g.
eaive
ad idivid a aia iy ad e a iaive f e.g. ig Da afe. Whi e
i i kw ha iig ai
a aia iy ad
a ei
i ay e i
e ea
ai f he e
e e be
ig ea iab e i hi ae we adde he
i e f he be f ei
ed aia dea ha ay be e ied i ee a a
i
a ea
ai whe i i ib e d . We
exa e hwig ha hi
be ay be exeia i he be f e
e eve whe a f he age
i iy f
i ae i
. We f he hw ha k age ay a
hieve i a ig e
dea a ea
ai e iig exeia y ay aia dea if a k 1 age
a
ai
iae hi ae ea
ai beig ea iab e by ay e e
e f aia dea i
whi
h a k 2 age ae iv ved.
1. d
i
e
hai f egiaig a
ai f e
e wihi a g f age f a i
a bdy f wk wihi he dy f iage ye. Tyi
a aba
de
deive f gae heei
ee
ive i e
i
ad ag he i e ha have bee
addeed ae aegie ha age e bai a ai
a be f he e
e avai
ab e e.g. a 2001; Re
hei Z ki 1994; Sadh 1999 ad
by
whi
h he
e f e ig e a
ai f e
e ag he age iv ved i
ageed e.g. Dig Geave 2000; D e 2003; D e
B ey 2003;
B ey
e a . 2002.
The eig we ae
eed wih i e
a aed i he f wig de(cid:12)ii.
De(cid:12)ii 1 A e
e a
ai eig i de(cid:12)ed by a i e hA; R; U i whee
A = fA
; A
; : : : ; A
g
; R = f
;
; : : : ;
g
1
2
1
2
ae ee
ive y a e f a ea w age ad a
e
i f haeab e e
e.
A i iy f
i i a aig f be f R aia va e. Ea
h age A
2 A
i
ha a
iaed wih i a ai
a i iy f
i
ha U i h
;
; : : : ;
i. A
i
1
2
a
ai f R A i a aii h
;
; : : : ;
i f R. The va e
i
a ed
1
2
i
i
he i iy f he e
e aiged A
. A i iy f
i i e if wheeve
i
S (cid:18) T i h d ha S (cid:20) T i.e. he va e aiged by ay e f e
e T
i eve e ha he va e aa
he ay be S f T .
(cid:13)2005 A A
e F dai ad ga a fa b ihe. A igh eeved.
D e
Tw a j a i
ai i whi
h he aba
view f De(cid:12)ii 1 ha bee ex ied ae
e
e
e ad diib ed ak ea iai. he (cid:12) R eee e
e
i f
diie (cid:11)eed f a e ad idivid a age eek a
ie a be f hee he
\va e" a age aa
he a e
i(cid:12)
e beig de
ibed by ha age i iy f
i.
ak aig he \e
e" e de
ibe a
e
i f b ak be efed i
de ea ie e
ex ak e.g. he \
ex ak" ay be a gd
f a
ea waeh e e e f
iie. hi exa e R de
ibe he
ai
whi
h gd be dia
hed ad a give a
ai de(cid:12)e he a
e whi
h a
age aage de iveie. The i iy f
i i
h
ae de he
a age
a
iae wih
ayig i a ed b ak.
Wihi he vey geea
ex f De(cid:12)ii 1 a be f i e aie eig f
he bevai ha i i ike y ha e iiia a
ai wi be ee a aifa
y
eihe wih ee
he view f a age i he ye wih ee
dive g ba
ideai. Th by ig
hage he iiia aige idivid a age
eek bai a \bee" a
ai. Thi
eai aie w iediae ei: hw
eva ae a give aii ad h have a bai f fig ived ia a
ai;
ad he i e de yig he ai e f hi ae wha ei
i h d be ied
he f ha ed dea ay ake.
We ha be e y eview e f he e wide y died aa
he de(cid:12)ig
dii de whi
h e a
ai ae ee a \bee" ha he. F he e
f hi id
i we i y beve ha
h
ieia ay be eihe aiaive
a iaive i a e. A a exa e f he fe we have he aa
h wheei he
\va e" f a a
ai i i y he f he va e give by he age i iy
f
i he be f R hey have bee aied wihi i.e.
: hi
i =1
i
i
i he
a ed i iaia
ia we fae whi
h avid eeii we wi dee by (cid:27)
.
A a a ai f age wihi a
diy adig
ex i eek a a
ai de
whi
h (cid:27)
i axiied. e exa e f a a iaive
iei i \evy feee": ifa y
a a
ai i evy fee if age aig geae i iy he e
e e
he d
j
by ahe age ha i de wih ee
he e
e e
i ha a
a y bee
i
a
aed i.e. f ea
h dii
ai hi ; j i
(cid:21)
.
i
i
i
j
vey geea e hee ae w aa
he ha have bee
ideed i eaig he
ei f hw a (cid:12)ie
e
i f e
e igh be diib ed ag a e f age
i de iie e
iei f iee: \
a
e" baed ehd e.g. D e
e a . 2003; Edi e a . 2003; Edi a de 2004b; Sadh 1998 1999 deivig
f he wk f Sih 1980; ad \
biaia a
i" e.g. ake Uga 2000a
2000b; Sadh e a . 2001; Sadh 2002; Sadh S i 2003; Teeh z 2000;
Yk e a . 2004 ag he. The igi(cid:12)
a di(cid:11)ee
e bewee hee i i he exe
whi
h a
ea ized
ig age deeie he eve a diib i f e
e
ag age.
e ay view he aegy de yig
biaia a
i a iveig he
a
ia e(cid:11) i a \e
eig" age f wig whi
h a give a
ai i deeied.
Th a
ig age he \a
iee" i ied wih a e f bid { ai hS
;
i
j
j
wheei S
i e be f he avai ab e e
e ad
he i
e age A
i eaed
j
j
j
ay i de a
ie S
. The b e fa
ed by he a
iee i de
ide whi
h bid
j
42
Exea Behavi i iage Ca
egiai
a
e i de axiie he vea (cid:12) b je
he
ai ha ea
h ie
a be baied by a e age.
Wha we ha efe a \
a
e
hee" yi
a y e
hew he e
ai
age ad bdiai a
ig abie e yed i a
i e
hai eekig
iead ea ie a iab e a
ai by a ageed e e
e f dea . The
a
e i
i geea iaiai f
eai f e
e diib ed ag age i
he
ee die
ed gah wih
vei
e ea
h f whi
h i a
iaed wih a dii
a
ai. hi way a ib e dea h ; i i eeeed a a edge die
ed f he
veex abe ed wih ha abe ed . Viewed h ideifyig a e e
e f dea
a
be ieeed a a ea
h
e whi
h i i
i e idivid a age ay
d
i a
a fahi.
Cea ized
hee
a be e(cid:11)e
ive i
ex whee he ai
ia
eae i
he ee f a
eig he a
iee abiai. evie wihi whi
h age
ae high y e f ieeed he exe ha hei ai
(cid:13)i
wih he a
i
e
i whi
h hee i a high degee f \
eaiy" ab he
e i wkig wad a
(cid:12)a a
ai he age iv ved ay y be eaed
eed \
a i y": ha i
a age wi y a
e a ed ea
ai if ai(cid:12)ed ha
h w d e i a
iediae ivee f i w ee
ive.
h
ae he
e f vig f
he iiia a
ai
he eve a ea
ai
i by a e e
e f
a aia
ii
(cid:12)
dea e.g. a age igh ef e a
e dea whi
h ed
ed (cid:27)
be
a e f he ibi iy
ha i (cid:11)e a
eaed i i iy. A key i e hee i he f wig: if he dea
a w y ve i whi
h a ea
h age e age A
(cid:11)e a ig e e
e
j
ahe age A
he he aia ea
ai h
;
i
a a way be i eeed; if
j
ii
(cid:12)
hweve evey ig e ve be \aia " he h
;
i ay be ea iab e.
ii
(cid:12)
We ay ifa y egad he view f
h age a \yi
" i he ee ha hey
ae wi ig a
e a \h e " a dea h ; i de hey igh i
a f
i iy deie he e
f a \ g e gai" a ig (cid:27)
> (cid:27)
h d.
(cid:12)
ii
Thee ae a be f ea why a age ay ad
h view e.g.
ide he
f wig i e
f ageeig a ea
ai.
A ea
ai f e
e i ageed ve a e e
e f age ea
h f whi
h
iv ve
i
ai bewee w age A
ad A
. Thi
i
ai
i
j
i f A
i ig a a A
f he f b y ; ; (cid:11)eig
hae
i
j
f A
f a aye f ; e ; ; (cid:11)eig afe A
i e
j
j
f a aye . The ee f A
i i y a
e f wig whi
h he
j
dea i i eeed eje
.
Thi f
e i a vey i e egiai
e hweve
ide i eai wihi
a w age eig i whi
h e age A
ay wihe big ab a a
ai
1
(cid:12)
ad h
a devie a a { e e
e f dea { ea ie hi f a iiia a
ai
whi e he he age A
de kw
. addii a e ha A
i he y
ii
2
1
(cid:12)
age ha ake a ad ha a (cid:12)a a
ai i (cid:12)xed eihe whe A
i \ai(cid:12)ed"
1
a a A
eje
ay (cid:11)e.
2
Whi e A
d be bee (cid:11) if
i ea ied i ay be he
ae ha he y a
2
(cid:12)
A
wi a
e ae he de whi
h i de e e.g. e age ay be
ei
a
2
ab he ba (cid:12)de f he ad wi a
e y dea f whi
h hey
a e
eive a
43
D e
iediae bee(cid:12). Thee ae evea ea why a age ay eba
e
h ai de
wihi he
hea ied:
e a dea ha bee i eeed A
ay e i iy b
2
f he a ae ade by A
ha he i \eae". We e ha eve if we
1
ei
h he bai
ha A
a de
ibe
A
ay i eje
(cid:11)e de whi
h
1
(cid:12)
2
i (cid:11)e a i
e i i wi ig e y he be e dea ha w d ae iae
i a
a y beig ed. A h gh he ii ake by A
i he eig j
2
de
ibed ay aea d y
a i we w d
ai ha i de e(cid:13)e
\ea " behavi
i
eai
ex. ide he aea f a aed a
ai ad egiai i iage
ye hee ae ay exa e f a
i by idivid a whee ied g e gai
ae i AE
ie egede he a
ea
e f h e . Cide \
hai ee"
hee hei e b e aifeai a \yaid e ig" eeie:
h have
a a a ifeie b ded by he ize f he ai i whi
h hey
i
ae b ay
beak dw befe hi i ea
hed. Fa
ed wih a e e \ed 10 he (cid:12)ve ae a
he head f he i ad fwad he ee e he afe addig y ae" deie he
ibi iy f igi(cid:12)
a gai afe a eay f 50 ige
h b adihe i
ee a ve y
ei
a ad
a i : hee ay be e
a
e a
e ha e wi
eve a y e
eive AE
ie e
ee i e ad i
i ha he ae de ha
bee ai aed.
ay we
a ideify w ia i(cid:13) e
e ha ead
ex i whi
h
age efe ve wad a ea
ai via a e e
e f \aia " dea . Fi y
he age ae e f ieeed b eaig i a ab e evie e.g. i he \
hai
ee" eig a age
a e iab y edi
he exa
i a whi
h he
hai wi fai .
The e
d fa
i ha
aia ei
i ay ii he de
ii a idivid a
age
a ake ab whehe a
e a ed dea . F exa e i eig
whee a dea iv ve e e
e a a ie A
ay eje
a a a
e e
2
e
e i
e i y \ ef " f wig a f he e e
e f dea :
if hi be
f f he dea i \a " he A
d de
ide a
e he ed dea i
e i ha
2
AE
ie
aia we deeie ha hee i a
ex i whi
h i f va e;
if hi be i \ age" hweve he A
ay a
k AE
ie we
a he ea
h
2
a
e f f e ibi iie ha w d a w i a
e . i
e ha i he exee
ae A
ake i de
ii e y whehe i f iediae e i.e. A
i yi
. A
2
2
e wef A
ay be ab e
ide whehe i ef h d k f he dea
2
ake a
e: i hi
ae A
d i ef e a
e i
e a h gh f e A
a
2
2
deeie hi wih a b ded k ahead.
a f he
eai we have de
ibed if A
wihe big ab a a
ai
1
(cid:12)
he fa
ed wih he view aded by A
ad he iiai ied by he dea
2
he y \e(cid:11)e
ive a" ha A
d ad i (cid:12)d a e e
e f aia dea
1
e A
.
2
ai i hi ai
e i hw ha
biig \
a " ei
i e.g. y e
e
e a a ie i iv ved i a
a ea
ai wih aia iy ei
i
a e
i eig i whi
h ay e e
e ea ie a ea
ai h ; i iv ve exeia y
ay i jRj eaae age. We e(cid:12)e hee idea i he ex b e
i.
44
Exea Behavi i iage Ca
egiai
1.1 e iiay De(cid:12)ii
T begi we (cid:12) fa ie he
e f dea ad
a
ah.
De(cid:12)ii 2 e hA; R; U i be a e
e a
ai eig. A dea i a ai h ; i whee
= h
; : : : ;
i ad = h
; : : : ;
i ae dii
aii f R. The e(cid:11)e
f i e
1
1
eig he dea h ; i i ha he a
ai f e
e e
i(cid:12)ed by i e a
ed wih ha
e
i(cid:12)ed by . F wig he ai f Edi a de 2004b f a dea AE = h ; i
we e A
idi
ae he be f A iv ved i.e. A
2 A
if ad y if
6=
.
k
k
k
AE
AE
e AE = h ; i be a dea . A
a
ah ea iig AE i a e e
e f a
ai
= h
;
; : : : ;
;
i
1
2
1
i whi
h =
ad
= . The egh f deed jj i 1 i.e. he be f
1
dea i .
Thee ae w ehd whi
h we
a e ed
e he be f dea ha a ig e
age ay have
ide i eekig ve f e a
ai ahe heeby
avidig he eed
he f exeia y ay a eaive:
a ad aia iy
ai. S
a
ai ii he eied dea he whi
h b d he
be f e
e ad/ he be f age iv ved b ake
ideai f he
view ay age ay have a whehe i a
ai ha ived.
a aia iy
ai ei
dea h ; i he i whi
h \ive" a
dig
ai
a
ieia.
hi ai
e we
ide w
ae f
a
ai:
a
de(cid:12)ed ad
ideed i Sadh 1998 ad wha we ha efe a k
a
.
De(cid:12)ii 3 e AE = h ; i be a dea iv vig a ea
ai f R ag A.
a. AE i a e
a
a
if
1. A
= fi ; j g.
AE
2. Thee i a i e e
e 2
[
f whi
h
=
[ f g ad
=
f g
i
j
i
i
j
j
wih 2
=
[ f g ad
=
f g wih 2
j
j
j
i
i
i
b. F a va e k (cid:21) 2 he dea AE = h ; i i a k
a
if 2 (cid:20) jA
j (cid:20) k ad
AE
[
= [
.
AE
AE
i 2A
i 2A
i
i
Th
a
iv ve he afe f exa
y e e
e f a ai
a age
ahe e ig i he be f dea
aib e wih ay give a
ai beig exa
y
1 : ea
h f he e
e
a be eaiged f i
e we ay f he
he 1 age.
Raia iy
ai aie i a be f di(cid:11)ee way. F exa e f he
adi f a idivid a age A
a give dea h ; i ay have hee di(cid:11)ee
e:
i
<
i.e. A
va e he a
ai
a ei
;
=
i.e.
i
i
i
i
i
i
i
i
i
i
i
A
i idi(cid:11)ee bewee
ad
; ad
>
i.e. A
i we (cid:11) afe he
i
i
i
i
i
i
i
i
dea . Whe g ba ia
h a i iaia
ia we fae ae be axiied hee i
he ei f wha i
eive hee i f ay age a
e a dea h ; i de whi
h i
45
D e
i ef wih a e va ab e e
e h dig. The adad aa
h hi ae ei
i id
e he i f a ay (cid:11) f
i i.e. i de f A
a
e a dea de
i
whi
h i (cid:11)e a ed
i i i iy A
e
eive e aye AE
ie
eae
i
f i . f
e
h
eai be ade by he age i he ye wh
i vidig i d wih ay i ex
e f ay gai. de(cid:12)ig i f ay (cid:11) he
ieeai i ha i ay aa
i ea
h age A
ake a aye (cid:25)
:
if (cid:25)
< 0
i
i
i
he A
i give (cid:25)
i e f a
eig a dea ; if (cid:25)
> 0 he A
ib e (cid:25)
he
i
i
i
i
i
a be diib ed ag he age whe ay (cid:11) i egaive.
Thi i f \eib e afe" i
a ed by he
e f idivid a aia iy
ad i fe de(cid:12)ed i e f a aiae ay (cid:11) ve
exiig. i diAE
hweve hw ha
h de(cid:12)ii ae e iva e he f wig.
De(cid:12)ii 4 A dea h ; i i idivid a y aia R if ad y if (cid:27)
> (cid:27)
.
We ha
ide a eaive bae f aia iy
ai ae: hee ae iai y f
iee wihi
a ed ey fee eig ha
eay aye f a i
i iy i a i.
The
ea i e f iee i hi ae
e he eie f he
a
e
gah whe he a wed dea aify bh a
a ad a aia iy
ai.
Th if we
ide abiay edi
ae dea h ; i { whee he
ae f iee ae
biig a
a ad aia iy
dii { we have
De(cid:12)ii 5 F a edi
ae ve dii
ai f a
ai a
a
ah
1
2
1
h
;
; : : : ;
;
i
ea iig h ; i i a ah if f ea
h 1 (cid:20) i < h
;
i i a dea ha i
i
i 1
i
i 1
;
h d. We ay ha i
ee if ay dea AE ay be ea ied by a ah.
We f he ay ha i
ee wih ee
dea whee i a edi
ae ve
dii
ai f a
ai if ay dea AE f whi
h AE h d ay be ea ied by a ah.
The ai iee i ea ie die f hee idea ha bee i aea
h a ideifyig
e
eay ad/ AE
ie
dii dea be
ee wih ee
ai
a
ieia e.g. Sadh 1998; ad i eab ihig \
vege
e" ad eiai e
ie e.g. Edi e a . 2003 Edi ad a de 2004b
ide dea ye
h
ha evey axia
ah ed i a ae ia a
ai i.e. e i whi
h ay
1
ea
ai de whi
h e age ive i i iy wi ead ahe age (cid:11)eig
a . Sadh 1998 exaie hw ei
i e.g. wih ; = > if ad y if
h ; i i a
a
ay a(cid:11)e
he exie
e f
a
ah ea ie dea . f
ai
a iee f he viewi f he ii
f ex ig he
a
e gah
ae
ae whee ; = > if ad y if he dea h ; i i idivid a y aia . F he
ae f
a
he f wig ae kw:
Thee 1
a.
a
ae
ee.
1. \axia " i he ee ha if h
; : : : ;
i i
h a ah he f evey a
ai
;
1
de h d.
46
Exea Behavi i iage Ca
egiai
b. R
a
ae
ee wih ee
R dea .
he
ideai f a gihi
ad
exiy i e eeed i D e e a . 2003
e diAE
y wih aeig f ae ea
ai a by aia
a
i
a eady aae ha i:
Thee 2 Eve i he
ae = 2 ad wih e i iy f
i he b e f
de
idig if a R
a
ah exi ea ie he R dea h ; i i {had.
Th de
idig if ay aia a i ib e i a eady
aia y had.
hi
ai
e we deae ha eve if a aiae aia a exi i exee
ae
hee ay be igi(cid:12)
a b e: he be f dea e ied
d be exeia i
he be f e
e a(cid:11)e
ig bh he ie i wi ake f he
hea ied
de ad he a
e ha a age wi have dedi
ae ig i. Th i hi f
f Thee 1 b Sadh beve ha whe a R
a
ah exi f a give
R dea i ay be he
ae ha i egh ex
eed i.e. e age ae a e
e
ahe ad he a
e he ae e
e a a ae age.
The yi
a f f he e ha we deive
a be aied a:
F a
a
ai
a
k
a
ad a aia iy
ai e.g. ; h d if h ; i i idivid a y aia hee ae e
e a
ai eig hA
; R
; U i i whi
h hee i a dea h ; i aifyig
a f he f wig.
a. h ; i i a dea .
b. h ; i
a be ea ied by a
a
ah whi
h evey dea ai(cid:12)e he
a
ai ad he aia iy
ai .
. Evey
h
a
ah ha egh a ea g .
F exa e we hw ha hee ae ia
e f whi
h he he R
a
ah ha
egh exeia i .
he ex e
i we wi be ieeed i we b d he
2
va e f he f wig f
i: we id
e hee i geea e avid e
eay
be e eeii.
De(cid:12)ii 6 e hA; R; U i be a e
e a
ai eig. Addiia y e ad be
w edi
ae dea . F a dea AE = h ; i he aia f
i
AE; hA; R; U i;
i he egh f he he
a
ah ea iig h ; i if
h a ah exi ad i
de(cid:12)ed if
h ah i ib e. The aia f
i
hA; R; U i; ; i
ax
ax
hA; R; U i; ; =
ax
AE; hA; R; U i;
dea AE
Fia y he aia f
i (cid:26)
; ; ; i
ax
ax
ax
(cid:26)
; ; ; =
ax
hA
; R
; U i; ;
U =h
;
;:::;
i
1
2
whee
ideai i ei
ed he dea AE = h ; i f whi
h a ea iig ah
exi.
2. Sadh 1998 give a e b d he egh f
h ah whi
h i a exeia i b
de ex i
i y ae ay we b d he ha ha a eady efeed .
47
D e
The hee ea e
ad (cid:26)
diig ih di(cid:11)ee ae
egadig he egh
ax
ax
f
a
ah. The f
i
i
eed wih ah ea iig a ig e dea h ; i
i a give e
e a
ai eig hA; R; U i: he ey f iee beig he be
f dea i he he i.e. ia egh ah. We e ha
i a aia f
i
whe va e i de(cid:12)ed i he eve ha h ; i
a be ea ied by a ah i he
eig hA; R; U i. The f
i
i de(cid:12)ed i e f
agai i he
ex f
ax
a e
i(cid:12)
e
e a
ai eig. The behavi f iee f
hweve i
ax
i y he egh f ah ea iig a e
i(cid:12)
h ; i b he \w
ae" va e f
f dea whi
h ae dea . We e he a i(cid:12)
ai ha
i de(cid:12)ed y f dea
ax
ha ae
aab e f beig ea ied by ah ad h d
ide
ae f whi
h
aiae
a
ah exi. Th if i h d be he
ae ha dea i he eig
hA; R; U i
a be ea ied by a ah he he va e
hA; R; U i; ; i de(cid:12)ed i.e.
ax
ax
ax
i a a aia f
i. We ay iee ay e b d
i he f wig
e:
if
hA; R; U i; ; (cid:20) he ay dea f whi
h a ah exi
a be
ax
ea ied by a ah f egh a .
ai iee wi
ee (cid:26)
whi
h i
eed wih he behavi f
a
ax
ax
a f
i f ad ad age ve a e f i iy f
i h : 2
! i
.
R
aa
h baiig we b d f hi f
i i
ive i.e.
f ea
h h; i
ha i
ideed we hw hw he i iy f
i U ay be de(cid:12)ed i a eig wih
e
e a yie d a we b d (cid:26)
; ; ; .
a he ea e
ax
ad
he f
i (cid:26)
i de
ibed i e f a ig e (cid:12)xed e
e a
ai
ax
ax
eig. i hweve i a aia f
i: deedig h ; ; ; i i ay be he
ae
ha i evey age e
e a
ai eig egad e f whi
h
hi
e f i iy
f
i i ade hee i dea h ; i
aab e f beig ea ied by ah ad f
h
ae he va e f (cid:26)
; ; ; wi be de(cid:12)ed.
ax
3
i ed a hi i ha he de(cid:12)ii f (cid:26)
a w abiay i iy f
i
ax
be e yed i
ig \w
ae" ia
e. Whi e hi i eaab e i e
f geea we b d e a wi be aae f he give
i he i iy
f
i a
a y e yed ae high y ai(cid:12)
ia ad ike y fea e i \ea " a i
ai
eig. We ha ae adde hi b je
i by f he
ideig b d he
f wig vaia f (cid:26)
:
ax
ax
ax
(cid:26)
; ; ; =
ax
hA
; R
; U i; ;
U =h
;
;:::;
i : ea
h
i e
1
2
i
Th (cid:26)
dea wih e
e a
ai eig wihi whi
h a f he i iy f
i
ax
aify a i
iy
ai.
The ai e f hi ai
e ae eeed i he ex e
i. We
ide w
geea
ae f
a
ah:
a
ah de vai aia iy
dii i
3. e
giig he ibi iy ha (cid:26)
; ; ;
d be de(cid:12)ed we ae
aiig ha
h
ax
behavi aie wih ay f he iaiai f h; i
ideed be e y:
i fa
i wi be
ea f he
i ha deig by (cid:26)
; he f
i (cid:26)
; ; ; f a (cid:12)xed
;
ax
ax
iaiai f h; i wih he ei
ed dea ye ad aia iy
dii exaied he f
i
ax
(cid:26)
; i a a f
i. Whehe i i ib e f ae \eib e"
hi
e f h; i wih
;
ax
whi
h (cid:26)
; i de(cid:12)ed f e va e f h ; i ad if deaig exa e f
h i
;
iai y y a ei f
biaia iee whe deve e i
ea he
e f
he
e ai
e.
48
Exea Behavi i iage Ca
egiai
Se
i 2; ad ii a y k
a
ah f abiay va e f k (cid:21) 2 i Se
i 3.
e ae
eed wih he
i f e
e a
ai eig hA; R
; U i
f whi
h give e aia iy e iee e.g. ha dea be idivid a y aia hee
i e dea h ; i ha ai(cid:12)e he aia iy
dii
a be ea ied by a aia
a
ah ee
ive y k
a
ah b wih he be f dea e ied
by
h ah beig exeia i . We addiia y bai igh y weake b i
exeia we b d f aia
a
ah wihi eig f e i iy
f
i i.e.
f he ea e (cid:26)
iig hw ii a e ay be deived f
ax
k
a
ah.
he e
e a
ai eig
ed f deaig hee eie wih
k
a
ah he
ed dea h ; i i ea iab e wih a ig e k 1
a
b ea iab e by ay aia k 1
a
ah. We di
e aed wk
i ai
a he e
e dy f Edi a de 2004a ha addee ii a i e
he
ideed i he ee ai
e i Se
i 4. C
i ad e die
i f
f he wk ae eeed i he (cid:12)a e
i.
2. we B d ah egh {
a
hi e
i we
ide he i e f
a
ah egh whe he
a ei
i
e ie idivid a dea be
a
. We (cid:12) give a veview f he
i
ehd wih he f wig be
i aa yig he
ae f ei
ed i iy f
i
ad be e y e i iy f
i.
2.1 veview
The aegy e yed i vig e iv ve w a: f a give
a f e
i
ed
a
ah we
eed a f w i baiig we b d (cid:26)
; ; ; .
ax
a. F he
a
e gah aiiig e
e ag age
a
ah
= h
;
; : : : ;
i ea iig a dea h
;
i. F he
a
1
2
1
0
ai
i(cid:13) e
ig i i he ved ha:
a1. The
a
ah
i a
ah i.e. f ea
h 1 (cid:20) i < he dea h
;
i
0
i
i 1
ai(cid:12)e he
a
ai
.
0
a2. F ay ai f a
ai
ad
ig i
if j (cid:21) 2 he he
i
i j
dea h
;
i i a
dea .
i
i j
0
Th a1 e e ha
i a iab e
a
ah whi e a2 wi g aaee ha
i 1
hee i exa
y e a
ai
ha
a be ea
hed wihi
f ay give
a
ai
i
by ea f a
dea .
i
0
b. De(cid:12)e i iy f
i U
= h
; : : : ;
i wih he f wig eie
1
b1. The dea h
;
i i a dea .
1
b2. F he aia iy
ai
i(cid:13) e
ig evey dea h
;
i i a
00
i
i 1
00
dea .
49
D e
b3. F evey a
ai
i he
a
ah ad evey a
ai he
i
ha
he dea h
; i i a dea i.e. i vi ae eihe he
a
i 1
i
ai
he aia iy
ai
.
0
00
Th a1 ad b2 e e ha h
;
i ha a de(cid:12)ed va e wih ee
he
1
f
i
f he dea h
;
i i.e. a ah ea iig he dea i ib e.
1
The eie give by a2 ad b3 idi
ae ha wihi he
ed e
e
a
ai eig he ah
i he i e ah ea iig h
;
i. f w
1
ha 1 he egh f hi ah give a we b d he va e f
ad he
e
ax
a we b d (cid:26)
; ; ; .
ax
Befe
i ig i wi be ef (cid:12)x e aia deai .
We e
dee he dieia hye
be. eeed a a die
ed gah
ha 2
vei
e ea
h f whi
h i idei(cid:12)ed wih a dii
bi abe . Uig (cid:11) = a
a
: : : a
1
2
dee a abiay
h abe he edge f
ae fed by
f h(cid:11); (cid:12) i : (cid:11) ad (cid:12) di(cid:11)e i exa
y e bi iig
We ideify bi abe (cid:11) = a
a
: : : a
wih be S
f R
via
2 S
if ad y if
1
2
i
(cid:11)
(cid:11)
a
= 1. Sii a y ay be S f R
a be de
ibed by a biay wd (cid:12) S f egh
i
i.e. (cid:12) S = b
b
: : : b
wih b
= 1 if ad y if
2 S . F a abe (cid:11) we e j(cid:11)j dee
1
2
i
i
he be f bi wih va e 1 ha j(cid:11)j i he ize f he be S
. f (cid:11) ad (cid:12) ae bi
(cid:11)
abe he (cid:11)(cid:12) i a 2 bi abe ha if R
ad T
ae diji e he (cid:11)(cid:12) de
ibe
he i f he be S
f R
wih he be S
f T
. Fia y if (cid:11) = a
a
: : : a
1
2
(cid:11)
(cid:12)
i a bi abe he (cid:11) dee he abe fed by
hagig a 0 va e i (cid:11) 1 ad
vi
e vea. hi way if S
i he be f R
de
ibed by (cid:11) he (cid:11) de
ibe he e
(cid:11)
R
S
. T avid a ex
e f e
i we wi whee abig iy aie e (cid:11) bh
(cid:11)
dee he bi abe ad he be f R
de
ibed by i e.g. we wie (cid:11) (cid:26) (cid:12) ahe
ha S
(cid:26) S
.
(cid:11)
(cid:12)
F = 2 he
a
e gah id
ed by
a
a be viewed a he
dieia hye
be
: he bi abe (cid:11) a
iaed wih a veex f
de
ibig
he a
ai h(cid:11); (cid:11)i hA
; A
i. hi way he e f R
a
de(cid:12)e a bgah
1
2
G
f
wih ay die
ed ah f (cid:12) (cid:12) i G
edig a ib e R
a
ah f he a
ai h ; R i he a
ai h ; R i.
2.2
a
ah { Uei
ed Ui iy F
i
(cid:12) e
ai(cid:12)e e i e i he eeai f Sadh 1998 ii 2:
i hi a e b d ha i exeia i i ved he egh f R
a
ah i.e. i e f ai Sadh 1998 ii 2 eab ihe a e
b d (cid:26)
; ; ; . We w ve a ii a de we b d.
ax
Thee 3 e ; be he edi
ae whi
h h d wheeve h ; i i a R
a
ad ; ha whi
h h d wheeve h ; i i R. F (cid:21) 7
ax
(cid:26)
2; ; ; (cid:21)
2
2
256
(cid:18)
(cid:19)
77
50
Exea Behavi i iage Ca
egiai
f. Cide a ah C = h(cid:11)
; (cid:11)
; : : : ; (cid:11)
i i
wih he f wig ey
1
2
4
8 1 (cid:20) i < j (cid:20) j (cid:21) i 2 (cid:11)
ad (cid:11)
di(cid:11)e i a ea 2 ii
SC
i
j
e.g. if = 4 he
;; f
g; f
;
g; f
;
;
g; f
;
g; f
;
;
g; f
;
g; f
;
;
g
1
1
3
1
2
3
2
3
2
3
4
2
4
1
2
4
i
h a ah a i
ed he e e
e h0000; 1000; 1010; 1110; 0110; 0111; 0101; 1101i.
Che C
be a ge
h ah wih hi ey ha
d be fed i
eig
= h
;
; : : : ;
i be he e e
e f a
ai wih
= h(cid:11)
; (cid:11)
i. We
i
i
1
2
i
w de(cid:12)e he i iy f
i
ad
ha f (cid:13) (cid:18) R
1
2
k
if
(cid:13) = (cid:11)
k
(cid:13)
(cid:13) =
1
2
0
if
(cid:13) 62 f(cid:11)
; (cid:11)
; : : : ; (cid:11)
g
1
2
Wih hi
hi
e he
a
ah
de
ibe he i e R
a
ah ea iig
he R dea h
;
i: ha
i a R
a
ah i iediae i
e
1
i 1
i
(cid:27)
= i 1 > i = (cid:27)
Tha i i i e f w f he fa
ha f a 1 (cid:20) i (cid:20) ad i 2 (cid:20) j (cid:20) he dea
i
j
h
;
i i a
a
he
e hee ae \h
" ib e ad f ea
h
i
i 1
5
hee i exa
y e R
a
ha
a f w i i.e.
.
F he e
edig ag e i f w ha ay we b d he egh f C
i.e. a e e
e aifyig he
dii SC i a we b d (cid:26)
2; ; ; . Thee
ax
ah i
wee igia y died by a z 1958 i he
ex f
dig hey ad
he we b d hei egh f 77=2562
2 eab ihed i Abb a
ha ki
1991.
2
Exa e 1 Uig he ah
4
C
= h0000; 1000; 1010; 1110; 0110; 0111; 0101; 1101i
= h(cid:11)
; (cid:11)
; (cid:11)
; (cid:11)
; (cid:11)
; (cid:11)
; (cid:11)
; (cid:11)
i
1
2
3
4
5
6
7
8
i he e
e a
ai eig hfa
; a
g; f
;
;
;
g; h
;
ii if he i iy f
i
1
2
1
2
3
4
1
2
ae e
i(cid:12)ed a i Tab e 1 be w he (cid:27)
h(cid:11)
; (cid:11)
i = 1 ad (cid:27)
h(cid:11)
; (cid:11)
i = 8. F hee
1
1
8
8
4
C
de
ibe he i e R
a
ah ea iig he ea
ai hh(cid:11)
; (cid:11)
i; h(cid:11)
; (cid:11)
ii
1
1
8
8
Thee ae a be f a eaive f ai f \aia iy" whi
h
a a be
ideed.
F exa e
De(cid:12)ii 7 e AE = h ; i be a dea .
4. Thi de(cid:12)e he
a ed \ake i he bx"
de id
ed i a z 1958.
5. exa e wih = 4 he e e
e h0000; 1000; 1001; 1101i a h gh de(cid:12)ig a
a
ah
give ie a dea whi
h i R ae y ha
edig h1000; 1001i.
51
D e
S
R S
S
R S
(cid:27)
S
R S
S
R S
(cid:27)
1
2
1
2
0000
1111
1
0
1
(cid:11)
1000
0111
1
1
2
(cid:11)
1
2
0001
1110
0
0
0
1001
0110
0
0
0
0010
1101
0
0
0
1010
0101
2
1
3
(cid:11)
3
0011
1100
0
0
0
1011
0100
0
0
0
0100
1011
0
0
0
1100
0011
0
0
0
0101
1010
4
3
7
(cid:11)
1101
0010
4
4
8
(cid:11)
7
8
0110
1001
3
2
5
(cid:11)
1110
0001
2
2
4
(cid:11)
5
4
0111
1000
3
3
6
(cid:11)
1111
0000
0
0
0
6
Tab e 1: Ui iy f
i de(cid:12)ii f = 4 exa e.
a. AE i
eaive y aia if f evey age A
(cid:21)
ad hee i a ea
i
i
i
i
i
e age A
f wh
>
.
j
j
j
j
j
b. AE i e iab e if i
> i
.
AE
AE
i 2A
i 2A
i
i
i
i
. AE i a ig Da dea if A
= fi ; j g
=
ad
i
i
j
j
i
i
j
j
AE
j
j < j
j whee j : : : j i ab e va e.
i
i
j
j
i
i
j
j
Thee ae a be f view we
a ake
eig he aia iy
dii give i Def
iii 7. e haed fea e i ha ike he
e f idivid a aia iy f whi
h
e vii
eae age wh (cid:11)e a i i iy i eeded i.e.
idivid a
aia iy e e a \ey baed" ye he f de(cid:12)ed i De(cid:12)ii 7 a w
e f \aia iy" be give i \ey fee" evie. Th i a
eaive y
aia dea age iv ved (cid:11)e a i i iy ad a ea e i bee (cid:11). ay
be ed ha give he
haa
eiai f De(cid:12)ii 4 i i iediae ha ay
ea
ive y aia dea i ef
e a idivid a y aia ; he
vee hweve
ea y de
h d i geea . e eig a e iab e dea ay be eihe
eaive y
idivid a y aia . e ay iee
h dea a e ehd f ed
ig ie a iy
bewee he va e age a
e hei a
ai: f he iv ved i a e iab e dea
i i e ed ha he age wh a
e ea va e hei
e a
ai wi bai a
e
e e whi
h i va ed e high y. ay f
e be he
ae ha e age
(cid:11)e a f i iy: he
dii f a dea be e iab e ii hw gea
h a
d be. Fia y he
e f ig Da dea igiae f ad ha bee died i
deh wihi he hey f ex
hage e
ie. Thi i e f ay aa
he ha have
bee ed agai i de de
ibe dea whi
h ed
e ie a iy bewee ebe
f a age
iey e.g.
Edi a de 2004b.
e f he de(cid:12)ii give
h dea e
a ae he
a ed ig Da i
i e i e
i
hey: ha ay
afe f i
e f a wea hy idivid a a e e h d ed
e he diaiy
bewee he. We e ha i i
i e we
d de(cid:12)e e aed aia iy
e
baed evea exei f hi i
i e ha have bee ggeed e.g. Aki 1970;
Chaea eaf e a . 2002; 1976.
Uig he ae
a
ah
ed i Thee 3 we eed y vay he
de(cid:12)ii f he i iy f
i e yed i de bai
C ay 1 F ea
h f he
ae be w
52
Exea Behavi i iage Ca
egiai
a. AE h d if ad y if AE i a
eaive y aia
a
.
AE h d if ad y if AE i
eaive y aia .
b. AE h d if ad y if AE i a e iab e
a
.
AE h d if ad y if AE i e iab e.
. AE h d if ad y if AE i a ig Da
a
.
AE h d if ad y if AE i a ig Da dea .
ax
(cid:26)
2; ; ; (cid:21)
2
2
256
(cid:18)
(cid:19)
77
f. We e y exa
y he ae e e
e f a
ai
de
ibed i he f f
Thee 3 b dify he i iy f
i h
;
i f ea
h
ae.
1
2
a. Che h
;
i wih
(cid:13) = 0 f a (cid:13) (cid:18) R ad
1
2
2
(cid:13) =
1
k
if
(cid:13) = (cid:11)
k
0
if
(cid:13) 62 f(cid:11)
; : : : ; (cid:11)
g
1
The e ig
a
ah i
eaive y aia : he i iy ejyed by A
e
2
ai
a whi e ha ejyed by A
i
eae by 1 wih ea
h dea . Ay deviai
1
f hi
a
ah e yig a a eaive
a
wi e i a f
i iy f A
.
1
b. Che h
;
i wih
(cid:13) =
(cid:13) ad
1
2
2
1
(cid:13) =
1
k
if
(cid:13) = (cid:11)
k
0
if
(cid:13) 62 f(cid:11)
; : : : ; (cid:11)
g
1
The
a
ah i e iab e: bh A
ad A
i
eae hei ee
ive i iy
1
2
va e by 1 wih ea
h dea . Agai ay
a
deviaig f hi wi e i
bh age ig e i iy.
. Che h
;
i a
1
2
k
if
(cid:13) = (cid:11)
2
k
if
(cid:13) = (cid:11)
k
k
(cid:13) =
;
(cid:13) =
1
2
0
if
(cid:13) 62 f(cid:11)
; : : : ; (cid:11)
g
2
if
(cid:13) 62 f(cid:11)
; : : : ; (cid:11)
g
1
1
T ee ha he
a
ah
i f ig Da dea i AE
e e ha
(cid:11)
(cid:11)
= 2
f ea
h 1 (cid:20) i (cid:20) . addii j
(cid:11)
(cid:11)
j = 2
2i 2
1
i
2
i
2
i 1
1
i 1
whi
h i i
y e ha j
(cid:11)
(cid:11)
j = 2
2i . Fia y ay
a
h ; i whi
h
2
i
1
i
deviae f hi e e
e wi be a ig Da dea i
e
j
j = 2
> j
j
2
2
1
1
2
2
1
1
whi
h vi ae e f he
dii e ied f ig Da dea .
2
The
i f w age eig eai y exed age be.
53
D e
C ay 2 F ea
h f he
hi
e f h; i
ideed i Thee 3 ad C ay 1
ad a (cid:21) 2
ax
(cid:26)
; ; ; (cid:21)
2
2
256
(cid:18)
(cid:19)
77
f. Fix a
ai i whi
h A
i give (cid:11)
A
a
aed (cid:11)
ad A
aiged ; f ea
h
1
1
2
1
j
3 (cid:20) j (cid:20) . Uig idei
a i iy f
i h
;
i a i ea
h f he evi
ae we
1
2
e y f
:
; = 1
S = 0 wheeve S 6= ; h; i a i Thee 3;
S = 0 f
j
j
j
j
a S C ay 1a;
; = 2
S = 0 wheeve S 6= ; C ay 1b; ad (cid:12)a y
j
j
S = 2
f a S C ay 1
. Cideig a ea iai f he dea h
;
i
j
1
he y
a
ah adiib e i he ah
de(cid:12)ed i he e aed f. Thi give
he we b d aed.
2
We e a hi i e he
e e
e f C ay 1 wih ee
Edi
a de 2004b Thee 1 3 whi
h ae
Fa
1 We e
a ha a ah h
; : : : ;
i i axia if f ea
h a
ai h
; i
1
i a dea .
a. f h
; : : : ;
i i ay axia ah f
eaive y aia dea he
i
1
ae ia .
b. f h
; : : : ;
i i ay axia ah f e iab e dea he
axiie he
1
va e (cid:27)
= i
i.e. he
a ed ega iaia
ia we fae.
e
1(cid:20)i (cid:20)
i
i
The e e
e f
eaive y aia dea i C ay 1a eiae i he ae
ia a
ai
: he a
ai f A
a way ha i iy 0 ad hee i a
ai
2
A
whe i iy
a ex
eed . Sii a y he e e
e f e iab e dea i C ay 1b
1
eiae i he a
ai
f whi
h (cid:27)
= he axi ha
a be aaied
e
f he ia
e de(cid:12)ed. bh
ae hweve he ia ae ea
hed by e e
e f
exeia y ay i dea : h a h gh Fa
1 g aaee
vege
e f ai
a
dea e e
e ia ae i ay be he
ae a i aed i C ay 1a{b ha
he
e f
vege
e ake
ideab e ie.
2.3
a
ah { e Ui iy F
i
We
de e
eig
a
by eeig a we b d (cid:26)
i.e.
ax
he egh f ah whe he i iy f
i ae e ied be e.
i
i e e
d ae
aiae e i iy f
i ha
w d have he deied eie wih ee
he ah ed i Thee 3. i hweve
fa f
ea whehe
h a
i i ib e. We d ae e ve hi
ei hee. Whehe a exa
a ai
d be a
ihed i iae y a ei
f e y
biaia iee: i
e ai i deae ha exeia egh
a
ah ae eeded wih e i iy f
i we ae iai y
eed
wih baiig a ia b d.
54
Exea Behavi i iage Ca
egiai
Thee 4 Wih ; ad ; be de(cid:12)ed a i Thee 3 ad (cid:21) 14
ax
(cid:26)
2; ; ; (cid:21)
8
(cid:16)
(cid:17)
77
=2
>
>
<
128
2
3
if i eve
(cid:16)
(cid:17)
>
>
:
77
1=2
128
2
3
if i dd
f. We de
ibe he deai y f he
ae f beig eve: he e whe i
dd i baied by a i e di(cid:12)
ai whi
h we ha ee y vide i ie.
e = 2 wih (cid:21) 7. F ay ah
= h(cid:11)
; (cid:11)
; : : : ; (cid:11)
i
1
2
i
whee (cid:11)
de
ibe a be f R
by a bi abe he ah d b e
i
i
i
2
de(cid:12)ed by
d b e
= h (cid:11)
(cid:11)
; (cid:11)
(cid:11)
; : : : ; (cid:11)
(cid:11)
; (cid:11)
(cid:11)
; : : : ; (cid:11)
(cid:11)
i
1
1
2
2
i
i
i 1
i 1
= h(cid:12)
; (cid:12)
; : : : ; (cid:12)
; (cid:12)
; : : : ; (cid:12)
i
1
3
2i 1
2i 1
2 1
The ea f
eive idi
e f (cid:12) i
eaig by 2 wi be
e
ea be e y
f
e d b e
de de
ibe a
a
ah
: i i hweve diAE
6
ie ae aiae a
ai (cid:12)
i de
ve i
h a ah. Cide
2i
he be (cid:12)
wih 1 (cid:20) i < de(cid:12)ed a f w:
2i
(cid:12)
=
2i
(cid:11)
(cid:11)
if (cid:11)
(cid:26) (cid:11)
i 1
i
i
i 1
(cid:11)
(cid:11)
if (cid:11)
(cid:27) (cid:11)
i
i 1
i
i 1
f we w
ide he ah ex
wihi
give by
2
ex
= h(cid:12)
; (cid:12)
; (cid:12)
; : : : ; (cid:12)
; (cid:12)
i
1
2
3
2 1
2 1
he hi ai(cid:12)e
a. f
ha ey SC f Thee 3 i
he ex
ha ey SC i
.
2
b. f j i dd he j(cid:12)
j = .
j
. f j i eve he j(cid:12)
j = 1.
j
F a ad he b d ved i Abb a
ha ki 1991 we ded
e ha ex
a be
he ha wih
deig he a
ai h(cid:12)
; (cid:12)
i
i
i
i
d. ex
de
ibe a
a
ah f
.
1
2 1
e. F ea
h ai hi ; j i wih j (cid:21) i 2 he dea h
;
i i a
a
.
i
j
f. f
i
he a i he f f Thee 3 he he be f dea i ex
i
a give i he aee f he ee hee.
6. e f he
ai(cid:12)
ai de
ibed by Sadh 1998 i
ai y wa dea S
a
:
ea
h dea wa exa
y e ie i (cid:12)
wih a ie i (cid:12)
i de give (cid:12)
.
2i 1
2i 1
2i1
55
D e
We heefe (cid:12)x
a he ah f Thee 3 ha i de
ee he f
we eed
i iy f
i h
;
i ha ae e ad wih whi
h ex
1
2
de(cid:12)e he i e R
a
ah ea iig he ea
ai h
;
i.
1
2 1
The
hi
e f
i e aive y i e. Give S (cid:18) R
2
2
8
>
<
0
if
jS j (cid:20) 2
S =
2 1
if
jS j = 1
2
>
:
2 2
if
jS j (cid:21)
hi i he be f a
ai i
. The behavi f
i
ea y e.
2
The
i f
i ahe e
i
aed. ai idea i ake e f
1
he fa
ha he ize f ea
h e (cid:12)
ig i ex
i vey igh y
aied:
j(cid:12)
j
i
i
i eihe 1 a
dig whehe i i dd eve. We (cid:12) deae ha ea
h
e f ize 1
a have a w i
be f ize
ig wihi ex
:
h evey S f ize 1 ha exa
y 2 1 0 be f ize ex
. T ee hi
e he
ay. e (cid:13) (cid:12)
(cid:12)
ad (cid:12)
be
h ha j(cid:13) j = 1 wih
2i 1
2j 1
2k 1
(cid:12)
(cid:26) (cid:13) ; (cid:12)
(cid:26) (cid:13) ; (cid:12)
(cid:26) (cid:13)
2i 1
2j 1
2k 1
ig ha (cid:12)
= (cid:11)
(cid:11)
ad ha
ha he ey SC i be he
ae ha a
2i 1
i
i
ea w f he bi abe f f(cid:11)
; (cid:11)
; (cid:11)
g di(cid:11)e i a ea w ii. Wih
i
j
k
f geea iy e hi i e f (cid:11)
ad (cid:11)
. A a e we ded
e ha he e (cid:12)
i
k
2i 1
ad (cid:12)
have a 2 e ee i
i.e. j(cid:12)
\ (cid:12)
j (cid:20) 2: (cid:12)
= (cid:11)
(cid:11)
2k 1
2i 1
2k 1
2i 1
i
i
ad (cid:12)
= (cid:11)
(cid:11)
i ay ii a whi
h (cid:11)
di(cid:11)e f (cid:11)
(cid:11)
di(cid:11)e f (cid:11)
a
2k 1
k
k
i
k
i
k
exa
y he ae ii. a j(cid:12)
(cid:12)
j (cid:21) 2 i.e. hee ae a ea w e ee
2i 1
2k 1
f (cid:12)
ha d
i (cid:12)
; ad i he ae way j(cid:12)
(cid:12)
j (cid:21) 2 i.e. hee ae
2i 1
2k 1
2k 1
2i 1
a ea w e ee f (cid:12)
ha d
i (cid:12)
. The e (cid:13) hweve ha y
2k 1
2i 1
1 ebe ad
a have bh (cid:12)
ad (cid:12)
a be: hi w d e ie
2i 1
2k 1
(cid:12)
\ (cid:12)
[ (cid:12)
(cid:12)
[ (cid:12)
(cid:12)
(cid:18) (cid:13)
2i 1
2i 1
2i 1
2k 1
2k 1
2k 1
b a we have j ee
j (cid:12)
\ (cid:12)
[ (cid:12)
(cid:12)
[ (cid:12)
(cid:12)
j (cid:21) 2
2i 1
2k 1
2i 1
2k 1
2k 1
2i 1
e iediae
e e
e f he ag e j give i ha f ay e (cid:13) f ize 1 hee
ae exa
y w i
be f (cid:13)
ig ex
if ad y if (cid:13) = (cid:12)
[ (cid:12)
= (cid:12)
2i 1
2i 1
2i
f e va e f i wih 1 (cid:20) i < . We
a w
haa
eie ea
h be f R
f ize 1
2
a fa ig i e f hee
aegie.
C1. Gd e give by f(cid:13) : (cid:13) = (cid:12)
g.
2i
C2. Digei
iig f
f (cid:13) : (cid:12)
(cid:26) (cid:13) (cid:13) 6= (cid:12)
ad i < g
2i 1
2i
C3. a
eib e e
iig f
f (cid:13) : (cid:13) i eihe Gd a Digei g
56
Exea Behavi i iage Ca
egiai
Gd e ae he de
ibig a
ai A
wihi he ah de(cid:12)ed by ex
;
1
Digei ae he a
ai ha
d be ea
hed ig a
a
f a e f
ize ex
i.e. (cid:12)
b di(cid:11)e f he e ha a
a y
i ex
i.e.
2i 1
(cid:12)
. Fia y a
eib e e ae he ha d
ex
ad
a be ea
hed
2i
via a
a
f ay e ex
. We e ha we view ay e f ize 1
ha
d be ea
hed by a
a
f (cid:12)
a beig ia
eib e: i i
i e i i
2 1
ib e exed he
a
ah beyd (cid:12)
hweve we
he
i
ae
2 1
he
i i hi way.
We w de(cid:12)e
a
1
8
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
:
(cid:13) =
1
2i 1
if
(cid:13) = (cid:12)
2i 1
2i 1
if
(cid:13) = (cid:12)
2i
2i
if
j(cid:13) j = 1 ad (cid:13) i a Digei f (cid:12)
2i 1
0
if
j(cid:13) j (cid:20) 1
0
if
j(cid:13) j = ad (cid:13) 62 ex
2 1
if
(cid:13) i a
eib e j(cid:13) j (cid:21) 2
eai y ve f hee
hi
e f h
;
i ha he
a
ah h
; : : : ;
i
1
2
1
2 1
de(cid:12)ed f ex
i he i e R
a
ah ea iig he R dea h
;
i
1
2 1
ad ha
i e.
1
T hw ha h
; : : : ;
i i R we eed deae
1
2 1
8 1 (cid:20) j < 2 1
(cid:12)
(cid:12)
<
(cid:12)
(cid:12)
1
j
2
j
1
j 1
2
j 1
We have via he de(cid:12)ii f h
;
i
1
2
(cid:12)
(cid:12)
= 2 i 1
1
2i 1
2
2i 1
<
(cid:12)
(cid:12)
1
2i
2
2i
= 2 i 2
<
(cid:12)
(cid:12)
1
2i 1
2
2i 1
= 2 i 3
Th via De(cid:12)ii 4 i f w ha ex
give ie a R
a
ah.
T ee ha hi ah i he i e R
a
ah i eeig h
;
i
1
2 1
ide ay ii
= h(cid:12)
; (cid:12)
i ad a
ai he ha
. ay be
j
j
j
j 1
j 1
a ed ha he dea h
; i i a
a
. f j = 2i 1 he (cid:27)
= 2 i 1
j
2i 1
ad j(cid:12)
j = . e
e j
j 2 f 1; 1g. he fe
ae
= 0 ad
= 2 2
j
1
1
1
2
2
f whi
h (cid:27)
= 2 2 ad h h
; i i R. he ae
ae
= 2i
1
1
j
i
e
i a Digei f (cid:12)
ad
= 2 1 givig (cid:27)
= 2 i 1. Agai
1
2i 1
2
2
j
h
; i fai be R i
e fai give ay i
eae i he va e f (cid:27)
. We ae ef wih
he
ae j = 2i ha (cid:27)
= 2 i 2 ad j(cid:12)
j = 1. Si
e h
; i i a ed
j
2i
j
be a
a
hi give j
j 2 f ; 2g. F he (cid:12) ibi iy
d be a
1
1
e ex
: (cid:12)
ad (cid:12)
ae bh be f (cid:12)
ad hee
a be a w
h
2i 1
2i 1
2i
be
ig ex
. f w heefe ha
= 0 givig (cid:27)
= 2 2
1
1
ha h
; i i R. he e
d ibi iy
= 2 1 b
= 0 a
1
1
2
2
j
j
j = 2 he dea w d e i a vea . We ded
e ha f ea
h
he
2
j
y R
a
ie wih i i he dea h
;
i.
j
j 1
57
D e
The (cid:12)a age i ve ha he i iy f
i
i ideed a e f
i.
1
S e S ad T ae be f R
wih S (cid:26) T . We eed hw ha
S (cid:20)
T . We
2
1
1
ay a e ha jS j = ha S
a e e wihi ex
ad ha jT j = 1.
f jS j < jS j = b de
ex
we have
S = 0 ad he e ied
1
ie a iy h d; if jS j (cid:21) 1 he i de f S (cid:26) T be ib e we w d eed
jT j (cid:21) 2 whi
h w d give
T = 2 1 ad hi i he axi va e ha ay
1
be i aiged by
. We ae ef wih y jS j = jT j = 1 ad S ex
1
ide. ha a eady bee hw ha hee ae a w be f T ha
a
ex
. Cide he di(cid:11)ee ibi iie:
a. T = (cid:12)
ha exa
y w be f T
i ex
: (cid:12)
ad (cid:12)
. Si
e
2i
2i 1
2i 1
(cid:12)
= 2i 1 ad hi i a ea axf
(cid:12)
;
(cid:12)
g h d S be eihe f
1
2i
1
2i 1
1
2i 1
(cid:12)
(cid:12)
he
S (cid:20)
T a e ied.
2i 1
2i 1
1
1
b. T i a Digei f S = (cid:12)
ha
T = 2i ad
S = 2i 1 ad agai
2i 1
1
1
S (cid:20)
T .
1
1
We ded
e ha
i e
eig we b d f f (cid:26)
f eve va e
1
ax
f .
We
de by bevig ha a ii a
i
a be ed if = 2 1 i dd:
e he ah ex
de
ibed abve b difyig i ha e e
e
i a way
he d by A
. y i di(cid:12)
ai he i iy f
i de(cid:12)ii ae eeded.
2
2
Exa e 2 F = 3 we
a
he
= h000; 001; 101; 111; 110i ha = 5. Thi
3
give d b e
a
3
h000111; 001110; 101010; 111000; 110001i
wih he
a
ah beig de(cid:12)ed f ex
whi
h i
3
h000111; 001111; 001110; 101110; 101010; 111010; 111000; 111001; 110001i
= h(cid:12)
; (cid:12)
; (cid:12)
; (cid:12)
; (cid:12)
; (cid:12)
; (cid:12)
; (cid:12)
; (cid:12)
i
1
2
3
4
5
6
7
8
9
Cideig he 15 be f ize 1 = 4 give
Gd
= f001111; 101110; 111010; 111001g
Digei
= f010111; 100111; 101011; 011110; 111100g
a
eib e = f011011; 011101; 101101; 110110; 110011; 110101g
i
e ha bh f he e i f110011; 110101g ae a
eib e :
i i
i e we
d
i e f (cid:12)
= 110001 ig eihe hweve i de i ify he
i he
9
ah i ha ed a (cid:12)
.
9
F wig he
i eeed i Thee 4 give he f wig i iy f
i
de(cid:12)ii wih S (cid:18) R = f
;
;
;
;
;
g.
1
2
3
4
5
6
8
>
<
0
if
jS j (cid:20) 1
S =
11
if
jS j = 2
2
>
:
12
if
jS j (cid:21) 3
58
Exea Behavi i iage Ca
egiai
F
we bai
1
8
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
<
0
if
jS j (cid:20) 2
0
if
jS j = 3 ad S 62 f000111; 001110; 101010; 111000; 110001g
1
if
S = 000111 (cid:12)
1
2
if
S = 010111 digei f (cid:12)
1
2
if
S = 100111 digei f (cid:12)
1
3
if
S = 001111 (cid:12)
2
3
if
S = 001110 (cid:12)
3
4
if
S = 011110 digei f (cid:12)
3
S =
5
if
S = 101110 (cid:12)
1
4
>
>
>
>
5
5
if
S = 101010 (cid:12)
>
>
>
>
>
>
>
>
>
>
>
>
6
if
S = 101011 digei f (cid:12)
5
7
if
S = 111010 (cid:12)
6
>
7
if
S = 111000 (cid:12)
7
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
:
8
if
S = 111100 digei f (cid:12)
7
9
if
S = 111001 (cid:12)
8
9
if
S = 110001 (cid:12)
9
9
if
jS j (cid:21) 5 S 2 f011011; 011101; 101101; 110110; 110011; 110101g
The e i iy f
i h
;
i e yed i vig Thee 4 ae de(cid:12)ed ha
1
2
he ah aiig f ex
i R: i he eve f eihe age (cid:11)eig a f i iy he
gai ade by he he i AE
ie vide a
eay aye. A a a ei
ha w aie i whehe he b d baied i Thee 4
a be hw a y whe
he aia iy
dii e
de ay eay aye e.g. f
ae whee he
e
f aia iy i e f he give i De(cid:12)ii 7. ex e hw ha if we e he
aia iy
dii ef
e
eaive y aia e iab e dea he he b d f
Thee 4 i h d.
Thee 5 F ea
h f he
ae be w ad (cid:21) 14
a. AE h d if ad y if AE i a
eaive y aia
a
.
AE h d if ad y if AE i
eaive y aia .
b. AE h d if ad y if AE i a e iab e
a
.
AE h d if ad y if AE i e iab e.
8
(cid:16)
(cid:17)
77
=2
>
>
<
128
2
3
if i eve
ax
(cid:26)
2; ; ; (cid:21)
(cid:16)
(cid:17)
>
>
:
77
1=2
128
2
3
if i dd
f. We agai i ae he
i y f he
ae f beig eve ig he
di(cid:12)
ai dea wih dd va e f ied a he ed f he f f Thee 4.
The ah ex
i ed f bh
ae.
59
D e
F a we e ie h
;
i be de(cid:12)ed a e f
i wih whi
h ex
wi
1
2
be he i e
eaive y aia
a
ah ea ie he
eaive y aia
dea h
;
i whee
= h(cid:12)
; (cid:12)
i. hi
ae we e h
;
i be
j
j
1
2
1
2 1
j
8
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
:
h
(cid:13) ;
(cid:13) i =
1
2
Si
e
hi ; i i
if
(cid:13) = (cid:12)
2i 1
hi 1; i i
if
(cid:13) = (cid:12)
2i
hi ; i 1i
if
j(cid:13) j = 1 ad (cid:13) i a Digei f (cid:12)
2i 1
h0; 2 1i
if
j(cid:13) j (cid:20) 1
h0; 2 1i
if
j(cid:13) j = ad (cid:13) 62 ex
h2 1; 0i
if
(cid:13) i a
eib e j(cid:13) j (cid:21) 2
h
(cid:12)
;
(cid:12)
i = hi ; i i
1
2i 1
2
2i 1
h
(cid:12)
;
(cid:12)
i
= hi 1; i i
1
2i
2
2i
h
(cid:12)
;
(cid:12)
i = hi 1; i 1i
1
2i 1
2
2i 1
i i
eai y he
ae ha h
;
i ad a dea he
a
ah de(cid:12)ed
1
2 1
by ex
ae
eaive y aia . F hee if = h(cid:13) ; (cid:13) i i ay a
ai he
ha
he he dea h
; i wi fai be a
eaive y aia
a
.
j 1
j
F e he
ay eig h
; i wih f geea iy be a
a
j
wih 62 f
;
g { we
a e he fe
ae i
e we have a eady hw
j 1
j 1
h a dea i
eaive y aia . f j = 2i 1 ha h
(cid:12)
;
(cid:12)
i = hi ; i i
1
j
2
j
he j(cid:13) j 2 f 1; 1g: he fe
ae ead a i i iy f A
; he ae
1
i
e (cid:13) i a Digei f (cid:12)
a i i iy f A
. Sii a y if j = 2i ha
2i 1
2
h
(cid:12)
;
(cid:12)
i = hi 1; i i he j(cid:13) j 2 f ; 2g: f he (cid:12) (cid:13) 62 ex
eadig a f
1
j
2
j
i iy f A
; he e
d e i a f i iy f A
. f w ha he ah de(cid:12)ed by
1
2
ex
i he i e
eaive y aia
a
ah ha ea ie h
;
i.
1
2 1
eai y hw ha hee
hi
e f h
;
i de(cid:12)e e i iy f
i.
1
2
Cide
ad e S ad T ae be f R
wih S (cid:26) T .
f jS j (cid:20) 1
1
2
S de
ex
he
S = 0.
f jT j (cid:21) 2 i a
eib e he
1
T = 2 1 whi
h i he axi va e aaiab e by
. S we ay a e ha
1
1
jS j =
ex
i.e. S = (cid:12)
; f e i ad ha jT j = 1 ad i eihe
2i 1
a Gd e a Digei . F he de(cid:12)ii f
S = i : if T 2 f(cid:12)
; (cid:12)
g he
1
1
2i
2i 2
T (cid:21) i =
S ; if T i a Digei f (cid:12)
he
T = i =
S . We ded
e
1
1
2i 1
1
1
ha if S (cid:18) T he
S (cid:20)
T i.e. he i iy f
i i e.
1
1
w
ide
wih S ad T be f R
havig S (cid:26) T .
f jT j (cid:21) 1
2
2
R
T de
i ex
he
T = 2 1 i axia va e. f jS j (cid:20) 2
2
2
R
S i a
eib e he
S = 0. Th we ay a e ha T = (cid:12)
givig
2
2
2i 1
T = i ad jS j = 1 ha R
S i eihe a Digei e f he Gd e
2
2
f(cid:12)
; (cid:12)
g. f R
S i a Digei he
S = i 1; if i i he Gd e (cid:12)
he
2i
2i 2
2
2
2i 2
S = i 1 <
T ; if i i he Gd e (cid:12)
he
S = i =
T . f w ha
2
2
2i
2
2
i e
eig he f f a a.
2
60
Exea Behavi i iage Ca
egiai
F b we e
h
(cid:13) ;
(cid:13) i =
1
2
8
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
:
h2i 1; 2i i
if
(cid:13) = (cid:12)
2i 1
h2i 1; 2i i
if
(cid:13) = (cid:12)
2i
h2i ; 2i 1i
if
j(cid:13) j = 1 ad (cid:13) i a Digei f (cid:12)
2i 1
h0; 2 1i
if
j(cid:13) j (cid:20) 1
h0; 2 1i
if
j(cid:13) j = ad (cid:13) 62 ex
h2 1; 0i
if
(cid:13) i a
eib e j(cid:13) j (cid:21) 2
Thee
hi
e give ex
a he i e e iab e
a
ah ea ie he e iab e
dea h
;
i i
e
1
2 1
if
(cid:12)
;
(cid:12)
g = 2i 1
1
2i 1
2
2i 1
if
(cid:12)
;
(cid:12)
g
= 2i
1
2i
2
2i
if
(cid:12)
;
(cid:12)
g = 2i 1
1
2i 1
2
2i 1
ea
h dea h
;
i i e iab e.
f = h(cid:13) ; (cid:13) i i ay a
ai he ha
j
j 1
j 1
he he dea h
; i i a e iab e
a
. A e ha h
; i i a
j
j
a
ad ha 62 f
;
g.
f j = 2i 1 ha
= h(cid:12)
; (cid:12)
i
2i 1
2i 1
j 1
j 1
j
ad if
(cid:12)
;
(cid:12)
g = 2i 1 he j(cid:13) j 2 f 1; 1g.
he (cid:12) f hee
1
2i 1
2
2i 1
if
(cid:13) ;
(cid:13) g = 0;
i he e
d if
(cid:13) ;
(cid:13) g = 2i 1 i
e (cid:13) be a
1
2
1
2
Digei . Thi eave y j = 2i wih
= h(cid:12)
; (cid:12)
i ad if
(cid:12)
;
(cid:12)
g = 2i .
2i
2i
1
2i
2
2i
j
F hi j(cid:13) j 2 f ; 2g: if j(cid:13) j = he if
(cid:13) ;
(cid:13) g (cid:20) 2i 1 wih e a iy whe
1
2
(cid:13) = (cid:12)
; if j(cid:13) j = 2 he if
(cid:13) ;
(cid:13) g = 0. a hee eab ih ha ex
2i 1
1
2
i he i e e iab e
a
ah ea iig he e iab e dea h
;
i.
1
2 1
Tha he
hi
e f h
;
i de
ibe e i iy f
i
a be hw by a
1
2
ii a ag e ha f a a.
2
Exa e 3 F = 3 ig he ae
a
ah ex
a he evi exa e
3
i.e.
h000111; 001111; 001110; 101110; 101010; 111010; 111000; 111001; 110001i
= h(cid:12)
; (cid:12)
; (cid:12)
; (cid:12)
; (cid:12)
; (cid:12)
; (cid:12)
; (cid:12)
; (cid:12)
i
1
2
3
4
5
6
7
8
9
F h
;
i i a we bai
1
2
8
>
>
>
>
>
>
>
>
>
>
>
>
h0; 9i
if
jS j (cid:20) 2
h0; 9i
if
jS j = 3 ad S 62 f000111; 001110; 101010; 111000; 110001g
h1; 1i
if
S = 000111 (cid:12)
1
h1; 0i
if
S = 010111 digei f (cid:12)
>
h1; 0i
if
S = 100111 digei f (cid:12)
>
>
>
>
>
>
>
>
<
h2; 1i
if
S = 001111 (cid:12)
2
h2; 2i
if
S = 001110 (cid:12)
3
h2; 1i
if
S = 011110 digei f (cid:12)
1
1
h
S ;
R S i =
h3; 2i
if
S = 101110 (cid:12)
1
2
4
>
>
>
>
h3; 3i
if
S = 101010 (cid:12)
5
>
h3; 2i
if
S = 101011 digei f (cid:12)
3
5
>
>
>
>
>
>
>
>
>
>
>
>
h4; 3i
if
S = 111010 (cid:12)
6
h4; 4i
if
S = 111000 (cid:12)
7
h4; 3i
if
S = 111100 digei f (cid:12)
7
>
8
h5; 4i
if
S = 111001 (cid:12)
>
>
>
:
h5; 5i
if
S = 110001 (cid:12)
9
h9; 0i
if
jS j (cid:21) 5 S 2 f011011; 011101; 101101; 110110; 110011; 110101g
61
D e
Sii a y i b
8
>
>
>
>
>
>
h0; 9i
if
jS j (cid:20) 2
h0; 9i
if
jS j = 3 ad S 62 f000111; 001110; 101010; 111000; 110001g
>
1
h1; 2i
if
S = 000111 (cid:12)
>
>
>
>
>
h2; 1i
if
S = 010111 digei f (cid:12)
>
h2; 1i
if
S = 100111 digei f (cid:12)
1
1
>
>
>
>
>
>
>
>
<
h3; 2i
if
S = 001111 (cid:12)
2
h3; 4i
if
S = 001110 (cid:12)
3
h4; 3i
if
S = 011110 digei f (cid:12)
3
h
S ;
R S i =
h5; 4i
if
S = 101110 (cid:12)
1
2
4
>
>
>
>
h5; 6i
if
S = 101010 (cid:12)
5
>
h6; 5i
if
S = 101011 digei f (cid:12)
h7; 6i
if
S = 111010 (cid:12)
6
h7; 8i
if
S = 111000 (cid:12)
7
h8; 7i
if
S = 111100 digei f (cid:12)
5
7
>
>
>
>
>
>
>
>
>
>
>
>
>
8
h9; 8i
if
S = 111001 (cid:12)
>
>
>
:
h9; 10i
if
S = 110001 (cid:12)
9
h9; 0i
if
jS j (cid:21) 5 S 2 f011011; 011101; 101101; 110110; 110011; 110101g
Tha we
a deae ii a exea behavi f
a
ah egh wih
aia iy
ai i bh ey baed idivid a aia iy ad ey fee
eaive aia iy e iab e eig iee
ive f whehe i
iy eie ae
a ed ha e ieeig aa e wih he
ex i whi
h i
iy i e
eva. ai
a we
a beve ha i
wih he
exiy e a eady
ed f D e e a . 2003 { de
idig if a a
ai i ae ia if a a
a
i axiie (cid:27)
if a R
a
ah exi { e iig i iy f
i be
e de e i a eig whi
h i
aia y e a
ab e.
3. k
a
ah
We w ii a i e wih ee
k
a
e
a ig ha i e ee
hee (cid:11)e a f f dea ha de (cid:12) i he
ai(cid:12)
ai f Sadh 1998. Thi
ai(cid:12)
ai de(cid:12)e f f f
a
ye:
a
a
ideed i he evi
e
i; S
a
ha iv ve exa
y 2 age waig ig e e
e; C
a
i whi
h e age afe a ea w f i e
e ahe; ad
a
i whi
h
hee e age ea
ae hei e
e h dig ag hee ve. de(cid:12)ii
f k
a
ei w age ex
hage e
e h ae
a
i
Sadh 1998
hee ad he dea eied ae ei
ed S ad C
a
. e egad hweve k
a
ae a geea a
a
i
e
a ee b d k i e
i(cid:12)ed f he be f age iv ved.
ai e k
a
ah i he f wig deve e f Thee 3.
Thee 6 e
; be he edi
ae whi
h h d wheeve h ; i i a R k
k
(cid:16)
(cid:17)
k
a
. F a k (cid:21) 3 (cid:21) k ad (cid:21)
hee i a e
e a
ai eig
2
hA; R; U i ad a R dea AE = h ; i f whi
h
AE; hA; R; U i;
=
1
a
k
b2 =k k 1
AE; hA; R; U i;
(cid:21) 2
1
b
k 1
AE; hA; R; U i;
i de(cid:12)ed
k 2
62
Exea Behavi i iage Ca
egiai
Befe eeig he f we
e ab he f ai f he hee aee
ad give a veview f he f
e.
We (cid:12) e ha he we b d whee de(cid:12)ed have bee haed i e f
he f
i
a ed (cid:26)
ed i he vai e
a
ah i
ax
Se
i 2.2. i f
e he
ae ha he b d
aied f
AE; hA; R; U i;
k 1
wi a be a we b d (cid:26)
; ;
; whe (cid:21) k ad ; h d wheeve
k 1
ax
he dea h ; i i R. The aee f Thee 6 hweve
ai ahe e ha hi
ae y ha a e
i(cid:12)
e
e a
ai eig hA; R; U i
a be de(cid:12)ed f ea
h (cid:21) k
ad ea
h gehe wih a R dea h ; i i
h a way ha: h ; i
a be a
hieved by
a ig e k
a
ad
a be ea ied by a R k 2
a
ah. Re
a ig
ha
i a aia f
i he ae ey i e iva e he
ai ade i a
f he dea h ; i f he hee aee. F hee hi ae dea a h gh
a
hievab e by a R k 1
a
ah
a be ea ied y by e whe egh i
a give i a b f he hee aee.
Regadig he f ie f hee ae a be f aia
exiie whi
h we have
aeed ae iae by akig e i ifyig a i
eig he e ai
hi bewee { he ize f he e
e e R { ad k { he be f age whi
h ae
eeded ea ie h ; i i a ig e R dea . ai
a we ha a e ha i a
(cid:16)
(cid:17)
k
exa
i e f
. We beve ha by e yig a ii a devi
e ha ed i
2
he f f Thee 4 we
a dea wih
ae f whi
h de have hi ey: if
(cid:16)
(cid:17)
(cid:16)
(cid:17)
k
k
=
f iege va e (cid:21) 1 ad 1 (cid:20) <
we i y e y exa
y he
2
2
ae
i ig e
e wih he \iig" e
e f R
beig a
aed A
ad eve beig ea
aed wihi he k 1
a
ah. Thi aa
h
1
a
f he dig eai b: : :
i he exe e f he we b d. We
ha a a e ha he be f age i A i exa
y k . Wihi he f we e a
ig exa e f whi
h k = 4 ad = 18 = 3 6 i ae e
i(cid:12)
fea e.
We (cid:12) give a ie f i
e.
Give hA; R; U i a e
e a
ai eig iv vig k age ad e
e
ai i de(cid:12)e a R k 1
a
ah
= h
;
; : : : ;
i
1
2
ha ea ie he R k dea h
;
i. We wi e d idex ai
a a
ai
1
wihi ha 1 (cid:20) d (cid:20) .
de i ify he eeai we e y a eig i whi
h he k age ae
(cid:16)
(cid:17)
k
A = fA
; A
; : : : ; A
g. Re
a ig ha =
he e
e e R
i fed by
0
1
k 1
2
(cid:16)
(cid:17)
k
he i f
aiwie diji e f ize . Give dii
va e i ad j wih
2
0 (cid:20) i < j (cid:20) k 1 we e R
dee e f hee be wih f
;
; : : : ;
g
1
2
i ;j
fi ;j g
fi ;j g
fi ;j g
he e
e ha f R
.
fi ;j g
Thee ae w ai idea deiig he
e f ea
h k 1
a
i .
Fi y i he iiia ad be e a
ai he e
e e R
i aiied
fi ;j g
bewee A
ad A
ad ay ea
ai f e
e bewee A
ad A
ha ake a
e
i
j
i
j
wihi he dea h
;
i wi iv ve y e
e i hi e. Th f evey a
d
d 1
ai
ad ea
h ai fi ; j g if h 62 fi ; j g he
\ R
= ;. F hee f
h
d
fi ;j g
d
63
D e
AE = h
;
i h d bh A
ad A
be iv ved i.e. fA
; A
g (cid:18) A
he hi ea
i
j
i
j
d
d 1
AE
ai f R
bewee A
ad A
wi be a
a
. Tha i eihe exa
y e
i
j
fi ;j g
e ee f R
wi be ved f
be
e a ebe f he a
ai
i
j
fi ;j g
d
d 1
exa
y e e ee f R
wi be ved f
be
e a ebe f he a
ai
j
fi ;j g
d
d 1
. a evey k 1
a
AE i
i f a i ae i eeai
i
(cid:16)
(cid:17)
k 1
f
a
: a ig e
a
f ea
h f he dii
ai fA
; A
g f age
i
j
2
f he k 1 age i A
.
AE
The e
d key idea i ex i e we kw ey f he dieia hye
be ewk: f evey (cid:21) 2
ai a ai ia
y
e i.e. a i e die
ed
y
e
fed ig y he edge f
ad
aiig a 2
vei
e.
w e
7
v
0
1
i
2
1
0
S
= v
; v
; : : : ; v
; : : : ; v
; v
i a ai ia
y
e i he hye
be
ad
w
0
1
i
2
1
0
S
= w
; w
; : : : ; w
; : : : ; w
; w
he ai ia
y
e i whi
h w
i baied by
eeig ea
h bi i v
. A we
i
i
have de
ibed i he veview f Se
i 2.1 we
a iee he bi abe v = v
v
: : : v
1
2
a de
ibig a ai
a be f R
i.e. ha be i whi
h
if ad y if
fi ;j g
fi ;j g
fi ;j g
k
v
= 1. Sii a y f ay be f R
we ay de(cid:12)e a i e bi wd. w e
k
d
ha
i he a
ai he d by A
i he a
ai
f . The dea AE = h
;
i
i
i
d
d
d 1
wi a(cid:11)e
\ R
i he f wig way: if i 62 A
j 62 A
he
\ R
=
i
i
d
d 1
fi ;j g
AE
AE
fi ;j g
d
d 1
d
fi ;j g
fi ;j g
fi ;j g
AE
\ R
ad
\ R
=
\ R
. hewie we have fi ; j g (cid:18) A
ad
i
j
j
he
eeay h dig
\ R
ad
\ R
de(cid:12)e
eeay bi
d
d
fi ;j g
fi ;j g
i
j
h
h
abe f vei
e i
: if hee
ed a
e hv
; w
i i he ai ia
y
e
he i
ad
he bi abe de(cid:12)ed f
\ R
ad
\ R
i
j
i
j
d 1
d 1
d 1
d 1
fi ;j g
fi ;j g
d
e he bi abe v
ad w
i.e. he vei
e ha
eed v
ad w
i
h1
h1
h
h
he ai ia
y
e. a f ea
h j A
iiia y h d eihe he be f R
ha
i
fi ;j g
a v
ha a w
ad a he
i f he k 1 ah h d he
0
0
be ha a v
w
. The (cid:12)a deai i ha he gei h gh he
2
1
2
1
ai ia
y
e i
d
ed ve a eie f d ea
h d
iig k k 1
dea .
We have ed ha ea
h k 1
a
h
;
i ha
i hi ah
a
d
d 1
(cid:16)
(cid:17)
k 1
be ieeed a a e f
dii
a
. A ia ey f he i iy
2
f
i e yed i ha e (cid:21) k 1 hee wi be idivid a y aia
a
ah ha ea ie he dea h
;
i i.e. he
a
dea
d
d 1
k 1
(cid:16)
(cid:17)
i ae y i de f he gei f
be R. A h gh
2
d
d 1
he e ied dea
d be ea ied by a e e
e f
a
e geea y ay
iab e k 2
a
ah
h ea iai wi de
ibe a R
a
ah.
7. Thi
a be hw by a eay id
ive ag e. F = 2 he e e
e h00; 01; 11; 10; 00i de(cid:12)e a
ai ia
y
e i
. d
ive y a e ha h(cid:11)
; (cid:11)
; : : : ; (cid:11)
; (cid:11)
i wih = 2
i
h a
y
e i
2
1
2
1
he h0(cid:11)
; 1(cid:11)
; 1(cid:11)
; 1(cid:11)
; : : : ; 1(cid:11)
; 0(cid:11)
: : : ; 0(cid:11)
; 0(cid:11)
i de(cid:12)e a ai ia
y
e i
.
1
1
1
2
2
1
1
64
Exea Behavi i iage Ca
egiai
The
i f i iy f
i g aaee
h behavi vide he i
ia
e i hwig ha he R dea h
;
i
a be ea ied wih a R k 2
1
a
ah:
if i ay a
ai f whi
h h
; i i a k 2
a
he
1
1
h
; i i R.
We w
eed wih he f f Thee 6.
f.
f Thee 6 Fix A = fA
; A
; : : : ; A
g. R
i f
aiwie diji
0
1
k 1
2
(cid:16)
(cid:17)
k
e f e
e
fi ;j g
fi ;j g
fi ;j g
fi ;j g
R
= f
;
; : : : ;
g
1
2
F k = 4 ad = 3 hee yie d A = fA
; A
; A
; A
g ad
0
1
2
3
f0;1g
f0;1g
f0;1g
f0;1g
R
= f
;
;
g
1
2
3
f0;2g
f0;2g
f0;2g
f0;2g
R
= f
;
;
g
1
2
3
f0;3g
f0;3g
f0;3g
f0;3g
R
= f
;
;
g
1
2
3
f1;2g
f1;2g
f1;2g
f1;2g
R
= f
;
;
g
1
2
3
f1;3g
f1;3g
f1;3g
f1;3g
R
= f
;
;
g
1
2
3
f2;3g
f2;3g
f2;3g
f2;3g
R
= f
;
;
g
1
2
3
We e w deig
e i de(cid:12)ig he k 1
a
ah.
a.
b.
v
0
1
i
2
1
0
S
= v
; v
; : : : ; v
; : : : ; v
; v
a ai ia
y
e i
whee wih f geea iy v
= 111 : : : 11.
0
w
0
1
i
2
1
0
S
= w
; w
; : : : ; w
; : : : ; w
; w
he
eeay ai ia
y
e hi ha w
= 000 : : : 00.
0
Th f k = 4 ad = 3 we bai
a.
S
= h111; 110; 010; 011; 001; 000; 100; 101i
v
b.
S
= h000; 001; 101; 100; 110; 111; 011; 010i
w
We
a w de
ibe he k 1
a
ah.
= h
;
; : : : ;
i
1
2
iia A
ai:
.
1
De(cid:12)e he k k B ea aix B = [b
℄ wih 0 (cid:20) i ; j (cid:20) k 1 by
i ;j
8
>
<
?
if
i = j
b
=
:b
if
i > j
i ;j
j ;i
>
:
:b
if
i < j
i ;j 1
65
D e
We he have f ea
h 1 (cid:20) i (cid:20) k
i 1
k 1
[
[
1
fj ;i g
fi ;j g
=
f R
: b
= >g [
f R
: b
= >g
i
i ;j
i ;j
j =0
j =i 1
Th i exa e
2
3
? > ? >
6
7
6
7
? ? > ?
B =
6
7
4
> ? ? >
5
? > ? ?
Yie dig he aig a
ai
1
f0;1g
f0;3g
f0;1g
f0;2g
f0;3g
= R
[ R
= h111; 000; 111i (cid:18) R
[ R
[ R
0
1
= R
= h000; 111; 000i (cid:18) R
[ R
[ R
f1;2g
f0;1g
f1;2g
f1;3g
1
1
= R
[ R
= h111; 000; 111i (cid:18) R
[ R
[ R
f0;2g
f2;3g
f0;2g
f1;2g
f2;3g
2
1
= R
= h000; 111; 000i (cid:18) R
[ R
[ R
3
f1;3g
f0;3g
f1;3g
f2;3g
The hid
i
idi
aig he 3 bi abe
haa
eiig ea
h f he be f
1
fi ;j g
i
R
f he hee va e ha j
a a e.
R d: The iiia a
ai i
haged ve a eie f d
1
2
z
;
; : : : ;
ea
h f whi
h iv ve exa
y k dii
k 1
a
. We e
idi
ae he
x ;
a
ai e ig afe age i d x whee 0 (cid:20) (cid:20) k 1. We e he f wig:
a. The iiia a
ai
wi be deed by
.
1
0;k 1
b.
i baied ig a ig e k 1
a
f
whe x (cid:21) 1.
x ;0
x 1;k 1
x ;
x ; 1
.
i baied ig a ig e k 1
a
f
whe 0 < (cid:20) k 1.
(cid:12)a ie f ai i ha f he
be ii f i wih ee
j i a a
ai
deed (cid:31)i ; j ; . eig be he bi ig de
ibig
\ R
i e a
ai
i
fi ;j g
(cid:31)i ; j ; i he idex f i he ai ia
y
e S
whe R
(cid:18)
he
i
v
fi ;j g
1
ai ia
y
e S
whe R
(cid:18)
. Whe =
f e a
ai i he
j
w
fi ;j g
x ;
1
e e
e de
i we e y he ai (cid:31)i ; j ; x ; ig ha e ivaia
f ah wi be (cid:31)i ; j ; x ; = (cid:31)j ; i ; x ; a ey ha
eai y h d e f
=
1
0;k 1
i
e (cid:31)i ; j ; 0; k 1 = (cid:31)j ; i ; 0; k 1 = 0.
The e e
e f a
ai i i b i a f w. Si
e
i he iediae
e
1;0
f he iiia a
ai
i AE
e de
ibe hw
i fed f
whe
0;k 1
x ;
x ; 1
> 0 ad
f
. e
be he a
ai be fed f
. The
x 1;0
x ;k 1
y ;
x ;
dea AE = h
;
i wi be a k 1
a
i whi
h A
= A fA
g. F ea
h ai
x ;
y ;
AE
fi ; j g (cid:18) A
we have (cid:31)i ; j ; x ; = (cid:31)j ; i ; x ; i he a
ai
. vig
AE
x ;
y ;
exa
y e e ee f R
i ea
aed bewee A
ad A
i
h a way ha i
i
j
fi ;j g
y ;
66
Exea Behavi i iage Ca
egiai
(cid:31)i ; j ; y ; = (cid:31)i ; j ; x ; 1 i
e A
ad A
ae a
ig
eeay ai ia
y
e
i
j
wih ee
R
hi e e ha (cid:31)j ; i ; y ; = (cid:31)j ; i ; x ; 1 heeby aiaiig
fi ;j g
he ivaia ey.
ig ha f ea
h dii
ai hi ; j i we eihe have R
a
aed A
i
i
fi ;j g
1
fi ;j g
1
R
a
aed A
i
he de
ii j ied idi
ae ha he a
ai
j
d
x ;
=
i
ee y e
i(cid:12)ed a f w.
The
be ii (cid:31)i ; j ; x ; ai(cid:12)e
8
>
0
if
x = 0 ad = k 1
>
>
>
>
<
1 (cid:31)i ; j ; x 1; k 1
if
x (cid:21) 1 = 0 ad 62 fi ; j g
(cid:31)i ; j ; x ; =
(cid:31)i ; j ; x 1; k 1
if
x (cid:21) 1 = 0 ad 2 fi ; j g
>
>
>
>
>
:
1 (cid:31)i ; j ; x ; 1
if
1 (cid:20) (cid:20) k 1 ad 62 fi ; j g
(cid:31)i ; j ; x ; 1
if
1 (cid:20) (cid:20) k 1 ad 2 fi ; j g
F ea
h i he be f R
ha i he d by A
i he a
ai
i
i
fi ;j g
x ;
(cid:31)i ;j ;x ;
fi ;j g
1
v
if R
(cid:18)
(cid:31)i ;j ;x ;
fi ;j g
w
if R
(cid:18)
j
i
1
whee we e
a ha bi abe i he hye
be
ae idei(cid:12)ed wih be
fi ;j g
f R
.
The ab e be w i ae hi
e f exa e.
A
A
A
A
0
1
2
3
i j
i j
i j
i j
i j
i j
i j
i j
i j
i j
i j
i j A
d 1
d
h
;
i
d
x
0 1
0 2
0 3
1 0
1 2
1 3
2 0
2 1
2 3
3 0
3 1
3 2
1
0
3
111 000 111 000 111 000 111 000 111 000 111 000
{
2
1
0
111 000 111 000 110 001 111 001 110 000 110 001
fA
; A
; A
g
1
2
3
3
1
1
111 001 110 000 110 001 110 001 010 001 110 101
fA
; A
; A
g
0
2
3
4
1
2
110 001 010 001 110 101 110 001 010 101 010 101
fA
; A
; A
g
0
1
3
5
1
3
010 101 010 101 010 101 010 101 010 101 010 101
fA
; A
; A
g
0
1
2
6
2
0
010 101 011 101 011 100 010 100 011 101 011 100
fA
; A
; A
g
1
2
3
7
2
1
010 100 001 101 011 100 011 100 001 100 011 110
fA
; A
; A
g
0
2
3
8
2
2
011 100 001 100 011 110 011 100 001 110 001 110
fA
; A
; A
g
0
1
3
9
2
3
001 110 001 110 001 110 001 110 001 110 001 110
fA
; A
; A
g
0
1
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
S be f R
he d by A
i
k = 4 = 3
i
fi ;j g
x ;
67
D e
A
A
A
A
0
1
2
3
i j
i j
i j
i j
i j
i j
i j
i j
i j
i j
i j
i j A
d 1
d
h
;
i
d
x
0 1 0 2 0 3 1 0 1 2 1 3 2 0 2 1 2 3 3 0 3 1 3 2
1 0
3
0
0
0
0
0
0
0
0
0
0
0
0
{
2 1
0
0
0
0
0
1
1
0
1
1
0
1
1
fA
; A
; A
g
1
2
3
3 1
1
0
1
1
0
1
1
1
1
2
1
1
2
fA
; A
; A
g
0
2
3
4 1
2
1
1
2
1
1
2
1
1
2
2
2
2
fA
; A
; A
g
0
1
3
5 1
3
2
2
2
2
2
2
2
2
2
2
2
2
fA
; A
; A
g
0
1
2
6 2
0
2
2
2
2
3
3
2
3
3
2
3
3
fA
; A
; A
g
1
2
3
7 2
1
2
3
3
2
3
3
3
3
4
3
3
4
fA
; A
; A
g
0
2
3
8 2
2
3
3
4
3
3
4
3
3
4
4
4
4
fA
; A
; A
g
0
1
3
9 2
3
4
4
4
4
4
4
4
4
4
4
4
4
fA
; A
; A
g
0
1
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
C be ii (cid:31)i ; j ; x ; k = 4 = 3
i
eai y he
ae ha hi
e f a yig
eive d f k dea
d be
i ed hweve we wih d hi y g a i i ib e g f e
a
ai
i he e e
e ahe
f e (cid:21) 2 via a k 1
a
.
d
d
w if
ad
ae dii
a
ai geeaed by he
e abve he he
x ;
y ;
dea AE = h
;
i i a k 1
a
if ad y if f e A
=
.
i
i
i
x ;
y ;
x ;
y ;
f w ha if h
;
i i a k 1
a
f e > 1 he f e i ad
d
d
a j 6= i
\ R
=
\ R
.
i
i
d
d
fi ;j g
fi ;j g
T deeie he ii va e f > 1 wih whi
h
=
we beve ha
i
i
d
d
wih f geea iy we eed
ide y he
ae d = i = 0 i.e. we deeie
he ii be f dea befe
eaea. Fi e ha i ea
h d
if
0
1
x
(cid:31)0; j ; x 1; k 1 = he (cid:31)0; j ; x ; k 1 = k 2 i.e. ea
h d adva
e he
be
ii k 2 a
e: (cid:31)0; j ; x 1; k 1 = (cid:31)0; j ; x ; 0 ad (cid:31)0; j ; x ; j = (cid:31)0; j ; x ; j 1.
We
a a beve ha
=
6=
f ay wih 0 < < k 1 i
e
0
0
0
1
0;k 1
x ;
(cid:31)0; 1; x ; = (cid:31)0; 2; x ; = : : : = (cid:31)0; k 1; x ;
y i he
ae = 0 ad = k 1. f w ha va e > 1 be f he f
k whee be
h ha k 2 i a exa
i e f 2
. F hi bevai we
ee ha
if > 1 :
=
g = if k : k 2 i a i e f 2
g
0
0
1
1
w if k i dd he = 2
i he iia
h va e ha = k 2
. f k i eve he
i ay be i e y wie i he f z 2
2 whee z i dd givig a 1 if (cid:21)
1
2
if (cid:20) ha hee give = k ad = z 2
2
e.g. f k = 4 ad = 3 we
ge k = 1 2
2 ha = 2
2
= 16 ad i exa e
=
ay be
0
0
1
3
3 11
1
17
eai y vei(cid:12)ed. a
8
>
<
k 2
if
k i dd
(cid:21)
k
if
k = z 2
2 z i dd ad (cid:21)
>
:
2
if
k = z 2
2 z i dd ad (cid:20)
68
Exea Behavi i iage Ca
egiai
A f whi
h iediae y give (cid:21) 2
i he e
d
ae k (cid:21) 2
he ie a iy h d
ivia y ad h we
a
i e he
hai f k 1
a
f a ea 2
ve.
Re
a ig ha =
hi give he egh f he k 1
a
ah
2
(cid:16)
(cid:17)
k
wie i e f ad k a a ea
8
= h
;
; : : : ;
i
1
2
(cid:16)
(cid:17)
=
k
2
2
k k 1
2
1 = 2
1
eai de(cid:12)e aiae i iy f
i U = h
; : : : ;
i i de e e ha
0
k 1
i he i e R k 1
a
ah ea iig he R k dea h
;
i.
1
de(cid:12)ig U i wi be
veie dee a he ah
= h
;
;
; : : : ;
; : : : ;
; : : : ;
i
0;k 1
1;0
1;1
1;k 1
x ;
;k 1
ad i
e k (cid:21) 2
we ay wih f geea iy f
he (cid:12) 2
a
ai i
hi
a
ah.
Re
a ig ha (cid:31)i ; j ; x ; i he idex f he bi abe
edig
\ R
i
x ;
fi ;j g
i he e eva ai ia
y
e { i.e. S
if R
(cid:18)
S
if R
(cid:18)
{ we
i
j
v
fi ;j g
w
fi ;j g
0;k
0;k 1
e he f wig eie f he e e
e f a
ai de(cid:12)ed by ha h d f ea
h
dii
i ad j .
1. 8 x ; (cid:31)i ; j ; x ; = (cid:31)j ; i ; x ;
2. f
i he iediae
e f
i he (cid:31)i ; j ; y ; (cid:20) (cid:31)i ; j ; x ; 1
y ;
x ;
wih e a iy if ad y if 62 fi ; j g.
3. 8 i
; j
wih 0 (cid:20) i
; j
(cid:20) k 1 (cid:31)i ; j ; x ; k 1 = (cid:31)i
; j
; x ; k 1.
0
0
0
0
0
0
The (cid:12) w eie have a eady bee eab ihed i de
ii f . The hid
f w f he bevai ha wihi ea
h d
ea
h
be ii i adva
ed by
x
exa
y k 2 i geig f
.
x 1;0
x ;k 1
The i iy f
i
i w give f S (cid:18) R
by
i
x ;
j 6=i
i
(cid:31)i ; j ; x ;
if S =
f e 0 (cid:20) x (cid:20) 0 (cid:20) (cid:20) k 1
S =
i
k
2
hewie
We
ai ha wih hee
hi
e
= h
;
;
; : : : ;
; : : : ;
; : : : ;
i
0;k 1
1;0
1;1
1;k 1
x ;
;k 1
i he i e R k 1
a
ah ea iig he R k dea h
;
i. Ce
0;k 1
;k 1
ai y i a R k 1
a
ah: ea
h dea AE = h
;
i hi ah ha
x ;
y ;
AE
AE
y ;
jA
j = k 1 ad i
e f ea
h age A
i A
= A fA
g he i iy f
ha i
eaed
i
i
8. We i he dig eai b: : :
i he exe whi
h i igi(cid:12)
a y if i a exa
(cid:16)
(cid:17)
k
i e f
i whi
h eve he devi
e de
ibed i veview f he f i a ied.
2
69
D e
by exa
y k 2 i.e. ea
h
be ii f i wih ee
j wheeve 62 fi ; j g ha
i
eaed i f w ha (cid:27)
> (cid:27)
ad he
e h
;
i i R.
y ;
x ;
x ;
y ;
We w hw ha i he i e R k 1
a
ah
i ai f
0;k 1
S e AE = h
; i i a dea ha deviae f he
a
ah havig f wed
x ;
i h gh he a
ai
. Ceai y bh f he f wig h d f :
f
x ;
ea
h i
(cid:18) [
R
; ad hee i a k e f ai hx
;
; : : : ; x
;
i wih
i
j 6=i
0
0
k 1
k 1
fi ;j g
whi
h
=
f if eihe fai be he
ae f e i he
= 2
wih he
i
i
i
i
x
;
i
i
k
e e e(cid:11)e
ha (cid:27)
< 0 ad he
e R. w if
i he a
ai ha
y ;
w d
eed
i he 6=
ad h f a ea e age
6=
.
i
i
x ;
y ;
i
i
x
;
y ;
a be he
ae ha
ed a a
ai
ig i
y ae ha
x
;
i
i
y ;
i
i i
e
h a
ai
d be ea ied by a k 1
a
. addii
i
x
;
i
i
AE
i
e
=
i be he
ae ha jA
j = k 1 i
e exa
y k 1
be ii i
i
i
he h dig f A
hage. f w ha hee ae y w ibi iie f y
;
:
i
i
i
eve he a
ai iediae y e
edig
adva
e he h dig
.
i
i
i
x ;
y ;
w AE
e beve ha a dea i whi
h e age aify he (cid:12) f hee whi e
he eaide
eed i a
da
e wih he e
d eihe de give ie a va id
a
ai
a be ea ied by a k 1
a
. he he had if
ed
he a
ai e
edig
he AE i R. We ded
e heefe ha he y R
x ;
k 1 dea ha i
ie wih
i ha e
ibed by
.
x ;
y ;
Thi
ee he aa yi eeded f he f f a b f he hee.
i
ea ha i
e he ye
ai y k age ay dea h ; i
a be e(cid:11)e
ed wih
a ig e k
a
heeby eab ihig a a. F a
{ ha he R dea
1
h
;
i
a be ea ied ig a idivid a y aia k 2
a
ah i
AE
e beve ha i
e he
a f R k 2
a
ae a be f he
a
f R k 1
a
wee i he
ae ha a R k 2
a
ah exied
i ee h
;
i hi w d i y ha wa he i e R k 1
a
1
ah. We have hweve ved ha i i e ad a
f he hee f w.
2
We bai a ii a deve e f C ay 1 i
(cid:16)
(cid:17)
k
C ay 3 F a k (cid:21) 3 (cid:21) k (cid:21)
ad ea
h f he
ae be w
2
a.
AE h d if ad y if AE i a
eaive y aia k
a
.
k
AE h d if ad y if AE i
eaive y aia .
b.
AE h d if ad y if AE i AE i a e iab e k
a
.
k
AE h d if ad y if AE i i e iab e.
hee i a e
e a
ai eig hA; R; U i ad a dea AE = h ; i f whi
h
AE; hA; R; U i;
=
1
a
k
b2 =k k 1
AE; hA; R; U i;
(cid:21) 2
1
b
k 1
AE; hA; R; U i;
i de(cid:12)ed
k 2
f. A wih he f f C ay 1 i e ai Thee 3 i ea
h
ae we e y
he
a
ah f he f f Thee 6 vayig he de(cid:12)ii f U = h
;
; : : : ;
i
1
2
k
i de eab ih ea
h e . Th e
= h
;
; : : : ;
; : : : ;
i
1
2
= h
;
; : : : ;
; : : : ;
i
0;k 1
1;0
x ;
z ;
70
Exea Behavi i iage Ca
egiai
be he k 1
a
ah ea iig he k dea h
;
i de
ibed i he f
1
f Thee 6 hi ah havig egh (cid:21) 2
1.
b2 =k k 1
a. The i iy f
i U = h
; : : : ;
i f Thee 6 e e ha h
;
i i
0
k 1
1
eaive y aia ad ha
i a
eaive y aia k 1
a
ah ea iig h
;
i: he i iy he d by A
eve de
eae i va e ad hee i
i
1
a ea e age i fa
exa
y k 1 whe i iy i
eae i va e. F hee
i he i e
eaive y aia k 1
a
ah ea iig h
;
i
1
i
e by he ae ag e ed i Thee 6 ay deviai wi e i e
age (cid:11)eig a f i iy.
b. Se he i iy f
i U = h
; : : : ;
i a
0
k 1
8
>
>
>
>
1
if
S 6=
f ay
2
i
x ;
x ;
>
2
xk
k i
if
S =
>
>
>
<
x ;k 1
i
x ;
x 1k
k
if
S =
< k 1 ad i = 0
0
2
S =
i
x ;
2
>
>
>
>
x 1k
k i 1
if
S =
< i 1 ad i 6= 0.
i
x ;i 1
x ;i
>
2
>
>
>
:
xk
1
if
S =
=
ad i 6= 0.
2
xk
1 i
if
S =
> i ad i 6= 0
i
i
i
x ;
T ee ha hee
hi
e adi
a a e iab e k 1
a
ah ea iig
he e iab e dea h
;
i we (cid:12) e ha
0;k 1
z ;
i
f
g > 1 = i
f
g
i
i
i
i
0(cid:20)i (cid:20)k 1
0(cid:20)i (cid:20)k 1
z ;
0;k 1
h h
;
i i ideed e iab e. Cide ay dea AE = h
;
i
ig
0;k 1
z ;
x ;
y ;
wihi
. AE
e hw ha
i
f
g 6=
i
i
0(cid:20)i (cid:20)k 1
x ;
x ;
i
e A
62 A
ad f a he age
>
. We have w ibi iie:
i
i
i
i
AE
y ;
x ;
= 0 i whi
h
ae = k 1 ad y = x 1; > 0 i whi
h
ae = 1.
Cide he (cid:12) f hee:
= xk
k hweve
0
0
x ;k 1
2
if
g = xk
1 =
i
i
k 1
k 1
x ;k 1
x ;k 1
2
ad he
e evey dea h
;
i fig a f
i e iab e.
x ;k 1
x 1;0
he eaiig
ae
= xk
1 ad
x ; 1
2
if
g (cid:20)
i
0
i
0
x ; 1
x ; 1
= x 1k
k 1
2
< xk
k
2k 1
2
2
= xk
k 1
2
2
< xk
1
2
=
x ; 1
ad h he eaiig dea h
;
i wihi
ae e iab e. By a ii a
x ; 1
x ;
ag e ha e yed i Thee 6 i f w ha
i he i e e iab e
k 1
a
ah ea iig h
;
i.
0;k 1
z ;
2
71
D e
e Ui iy F
i ad k
a
ah
The devi
e ed deve Thee 3 bai he ah f Thee 4
a be a ied
he ahe e ii
ae
i f Thee 6 heeby a wig exeia we
b d (cid:26)
; ;
; be deived. We wi ee y ie he aa
h ahe ha
k
ax
ee a deai ed e
hi
a exii. We e
a ha i be
ae e aive y aighfwad
de(cid:12)e iab e e i iy f
i
e i wa e ed ha he be ize f
iee { i.e. he f a
ai aiig i he
a
ah { wee f
ed fa i
a ie ei
ed age. The ai diAE
y ha aie i a yig ii a ehd he
ah f Thee 6 i he f wig: i he f f Thee 4 we
ide w age
ha
veig
f a eig wih e
e i Thee 3 ex
wih 2
e
e i Thee 4 i a
hieved by
biig \
eeay" a
ai i.e. (cid:11) (cid:18) R
wih (cid:11) (cid:18) T
. We
a ex i w fa
hweve deve a ah i f whi
h
e i iy f
i
d be de(cid:12)ed: he e
e e R
i Thee 6
i f
(cid:16)
(cid:17)
k
2
diji e f ize ; ad ay dea AE he ah iv ve a ea
ai f R
fi ;j g
bewee A
ad A
whe fi ; j g (cid:18) A
. Th eig T
be fed by
diji e
i
j
2
AE
(cid:16)
(cid:17)
k
fi ;j g
d
T
ea
h f ize e ha
i de
ibed by
i
d
d
d
d
d
(cid:11)
(cid:11)
(cid:11)
(cid:11)
(cid:11)
i ;0
i ;1
i ;i 1
i ;i 1
i ;k 1
d
fi ;j g
d
wih (cid:11)
he bi abe
edig he be f R
ha i he d by A
i
.
i ;j
i
Cide he e e
e f a
ai
i = hC
; C
; : : : ; C
i
1
2
i a e
e a
ai eig have k age ad 2 e
e { R
[ T
f whi
h C
d
i
i
haa
eied by
d
d
d
d
d
(cid:12)
(cid:12)
(cid:12)
(cid:12)
(cid:12)
i ;0
i ;1
i ;i 1
i ;i 1
i ;k 1
hi (cid:12)
idi
ae he be f R
[ T
de
ibed by he 2 bi abe
i ;j
d
fi ;j g
fi ;j g
d
d
d
(cid:12)
= (cid:11)
(cid:11)
i ;j
i ;j
i ;j
d
d
fi ;j g
fi ;j g
i.e. (cid:11)
e e
a be f R
whi e (cid:11)
a be f T
.
i ;j
i ;j
i iediae f hi
i ha f ea
h a
ai C
i i ad ea
h
d
A
i i a way he
ae ha jC
j = k 1 . f w heefe ha he y be
i
i
d
ha ae e eva he de(cid:12)ii f e i iy f
i wih whi
h a aa g
e Thee 6 f he ah i
d be deived ae he f ize k 1 : if
S (cid:18) R
[ T
ha jS j < k 1 we
a (cid:12)x
S a a a e gh egaive va e; ii a y
i
if jS j > k 1 he
S
a be e a age e gh iive va e.
i
9
de
ii i he e
edig aagah
a be aied i he f wig e
whe f i ied: exedig he ie give abve a fa we b d
9. i wh ig ha he \ie ai" age ed i Thee 4 i eeded i fig i :
he dea hC
; C
i i a k 1
a
. We e
a ha i gig f
f Thee 3 ex
d
d 1
he ieediae age { d b e
{ wa a
a
ah.
72
Exea Behavi i iage Ca
egiai
f i age y a e
hi
a exe
ie e yig
h f he aa yi a eady id
ed ad
i
e hig igif
a y ew i e ied f
h a aa yi we ha give a deai ed
eeai f i.
Thee 7 e
; be he edi
ae whi
h h d wheeve h ; i i a R k
k
(cid:16)
(cid:17)
k
a
. F a k (cid:21) 3 (cid:21) k ad (cid:21) 2
hee i a e
e a
ai eig
2
hA; R; U i i whi
h evey 2 U i e ad a R dea AE = h ; i f whi
h
AE; hA; R; U i;
=
1
a
k
b =k k 1
AE; hA; R; U i;
(cid:21) 2
1
b
k 1
AE; hA; R; U i;
i de(cid:12)ed
k 2
4. Re aed Wk
The i
ia f
f hi ai
e ha
ideed a ey f
a
ah ea iig ai
a ea
ai h ; i whe he
i e dea ae e ied
f a
a
ei
i ad aify a aia iy
ai. Se
i 2 he
a ei
i iied
dea he iv vig a ig e e
e i.e.
a
. F he aia iy
ai
f
ig dea i
y ive i iaia
ia we fae i.e. be idivid a y aia
R we have he f wig eie.
a. Thee ae e
e a
ai eig hA; R; U i wihi whi
h hee ae R ea
ai
h ; i ha
a be ea ied by a e e
e f R
a
. Sadh 1998
ii 2
b. Evey R ea
ai h ; i ha
a be ea ied by a R
a
ah
a be
ea ied by a R
a
ah f egh a
1 . Sadh 1998
ii 2
. Give hA; R; U i gehe wih a R ea
ai h ; i he b e f de
idig if
h ; i
a be i eeed by a R
a
ah i {had eve if jAj = 2 ad
bh i iy f
i ae e. D e e a . 2003 Thee 11.
d. Thee ae e
e a
ai eig hA; R; U i wihi whi
h hee ae R ea
ai
h ; i ha
a be ea ied by a R
a
ah b wih ay
h ah havig
egh exeia i . Thi h d eve i he
ae jAj = 2 ad bh i iy f
i
ae e. Thee 3 ad Thee 4 f Se
i 2
a e
e ai
e Edi ad a de 2004a aa ye
a
ah egh a
ideig
a
wih vai aia iy
ai. A h gh he aa
h i f a ahe
di(cid:11)ee ee
ive he
ea ei addeed { \ w ay aia dea ae e ied
ea
h a ia a
ai?" Edi a de 2004a Tab e 1 . 629 { i
e y
e aed he i e di
ed abve. e igi(cid:12)
a di(cid:11)ee
e i he aa yi f aia
a
f Sadh 1998 eae ad he e i Se
i 2 i ha i Edi
a de 2004a he i iy f
i ae ei
ed ha evey aia ea
ai
h ; i
a be ea ied by a aia
a
ah. The w ai ei
i exaied
ae e iig i iy f
i be addiive i.e.
f evey S (cid:18) R S =
;
2S
73
D e
ad e iig he va e e ed be eihe 0 1
a ed 0 1 i iy f
i.
Addiive i iy f
i ae
ideed i he
ae f R
a
Edi a de
2004a Thee 3 9 wheea 0 1 i iy f
i f
eaive y aia
a
Edi a de 2004a Thee 4 11. Uig (cid:26)
; ; ; ad (cid:26)
; ; ;
add
0 1
ax
ax
dee he f
i id
ed i De(cid:12)ii 6 whee a i iy f
i ae addiive
ee
ive y 0 1
f. he de(cid:12)ii f (cid:26)
he wih
; h dig if h ; i i a R
1
ax
a
;
; h dig if h ; i i a
eaive y aia
a
ad ;
2
e whe h ; i i R we ay f ae Thee 9 ad 11 f Edi a de 2004a
i e f he faewk ed i De(cid:12)ii 6 a
ax
(cid:26)
; ;
; =
Edi a de ; 2004a ; Thee 9
add
1
ax
(cid:26)
; ;
; =
Edi a de ; 2004a ; Thee 11
0 1
2
We
a f
e e a y
h Thee 3 ad 4 f Se
i 2 i e f he \he
ah"
vei aded i Edi a de 2004a vided ha he dai f
i iy ad ea
ai ia
e ae ei
ed he f whi
h a aiae
a
ah exi. Th we
a bai he f wig deve e f Edi a de 2004a
Tab e 1 i he
ae f
a
.
Ui iy F
i Addiive
0 1 Uei
ed e Uei
ed e
Raia iy
R
CR
R
R
CR
CR
She ah
(cid:10)2
(cid:10)2
(cid:10)2
(cid:10)2
=2
=2
C ee
Ye
Ye
Tab e 2: w ay
a
aia dea ae e ied ea
h a a
ai?
Exei f Tab e 1 f Edi a de 2004a . 629
5. C
i ad F he Wk
ai i hi ai
e ha bee deve he ea ie die f Sadh 1998
eig
he
e ad ii f ai
a \a
i
a "
a
f. Whi e Sadh 1998 ha
eab ihed ha iiig idivid a aia iy i addii he
a ei
i
e
ibed by
a
ead
eai whi
h ae i
ee i he ee ha hee
ae idivid a y aia dea ha
a be ea ied by idivid a y aia
a
f
ha bee wih ee
dea whi
h
a be ea ied by ei
ed
a
ah
wih he iei f deeiig wha exe he
biai f
a ad aia
iy
dii i
eae he be f dea e ied. We have hw ha ig a be
f a a de(cid:12)ii f aia iy f eig iv vig e
e aia
a
ah f egh (cid:10)2
ae eeded wheea wih he aia iy ei
i idivid a
dea a
a
AE
e ea ie ay dea . We have a
ideed a
a
f dea { k
a
{ ha wee exaied i Sadh 1998 eab ihig f
hee
ae ha whe ai
a aia iy
dii ae ied k 1
a
ah f egh (cid:10)2
ae eeded ea ie a dea ha
a be a
hieved by a ig e
2 =k
2
k
a
.
We e ha aa ye have iai y bee f
ed w
ae we b d
ah egh whe aiae ah exi ad a
h hee ae evea ei f
74
Exea Behavi i iage Ca
egiai
a
i
a iee ha ei f he di
i. ay be ed ha he ah
e
ad a
iaed i iy f
i ae ahe ai(cid:12)
ia beig die
ed aaiig a ah f a
e
i(cid:12)
egh eeig a give aia iy
iei. We have ee hweve i Thee 4
ad 5 a ied i di
i
dig Se
i 3 ha he i e f exeia egh
a
ah
i e aie eve whe we e ie he i iy f
i aify a
i
iy
dii. We
a ideify w
ae f e ei ha aie f hee
e .
Fi y f
ig R
a
ah i w d be f iee ideify \a a "
ei
i i iy f
i whi
h w d e e ha if a dea h ; i
a be i eeed
by a R
a
ah he i
a be ea ied by e whe egh i yia y
b ded i e.g.
h a addiiviy eied i he e
edig e
i. We
a iee
Thee 4 a idi
aig ha i
iy de g aaee \h" R
a
ah.
We e hweve ha hee ae e ei
i ha AE
e. T e a ahe ivia
exa e if he be f dii
va e ha (cid:27)
a a e i a
f e
a he R
a
ah
a have egh ex
eedig
:
eive dea
i
y i
eae (cid:27)
ad if hi
a ake a di(cid:11)ee va e he R
a
ah
a have egh ex
eedig . A we a beig f a
i
a iee
ae f i iy
f
i wih he ey beig
ideed w d a be f e iee egadig e
exiy i e. The e ved i D e e a . 2003 eab ihig ha de
idig if a
R
a
ah exi i had give a we b d he
aia
exiy
f hi b e. A ee ivia e b d hi b e
exiy
ha bee deaed. e i Thee 3 ad 4 idi
ae ha if hi de
ii
b e i i h i
exiy w d be {
ee ahe ha {had he
he e ied yia egh exie
e
ei(cid:12)
ae ay have be ehig he ha
he ah ie f.
We e ha he f f {hade i D e e a . 2003
10
a ia
e i whi
h (cid:27)
a ake a dii
va e: h f exa e f
a ei
i e ig ha if
h ae ee he R
a
ah ae \h" hi
e f D e e a . 2003 idi
ae ha he ei f de
idig hei exie
e igh
eai
aia y had.
Cideig ei
i he f f i iy f
i i e aa
h ha
d be
ake egadig (cid:12)dig \a
ab e"
ae. A a eaive w d be gai e iigh
i wha he \aveage" ah egh i ike y be. aeig adde hi ei
hweve a be f
ha egig i e aie. The iediae f hee
e
f
e he i f de ig a diib i i iy f
i give de(cid:12)ii
f aia iy i e f he va e age aa
h hei e
e h dig. i
i e
a aveage
ae aa yi f
eai iv vig exa
y w age
d be
aied i
e y gah heei
e i.e. wih he
i
ai f
ideig i iy f
i
die
y. i
ea hweve whehe
h a gah heei
aa yi bviaig he eed
f
ideai f iea i iy f
i
a be exeded beyd eig iv vig
exa
y w age. e diAE
y aiig wih hee e age i ha i iy
10. The e f \ay" ahe ha \ " i eeded be
a e f he
vei f eeeig i iy f
i
e yed i D e e a . 2003.
75
D e
f
i have a
aive exea iie i.e. give a a
ai hX ; Y ; Z i hee age
X i
haged h d Y [ Z be ediib ed ag A
ad A
.
1
2
3
11
A e (cid:12)a e f i e ha ay ei f he dy we aie he f wig.
i he idivid a dea a
a
ah aify bh a
a
dii be a
a
iv ve a k age ad a aia iy
ai.
F
ig
a
we have he f wig exee: f Sadh 1998 a
a
AE
e ea ie ay aia dea ; f e abve (cid:10)2
aia
a
ae eeded ea ie e aia dea . Thee ae a be f e
hai
we
a e y e ax he
dii ha evey ig e dea be a
a
ad be
aia . F exa e a w a ah
ai e be f dea whi
h ae
a
b i be R ii ha a dea ae
a
b a w e
be iaia . Th i he ae
ae if we g he exe f a wig iaia
a
he ay aia dea
a be ea ied eAE
ie y. w d be f e iee
exaie i e
h a he e(cid:11)e
f a wig a
a be f iaia dea
ad ei
h a whehe hee ae i ai i whi
h iaia
a
yie d
a `h
a
ah b 1 f
e e f exeia egh. f ai
a iee
f a a i
ai viewi i he f wig: de(cid:12)e a (cid:13) ; ah a a
a
ah
aiig a (cid:13)
a
whi
h ae idivid a y aia . We kw
ha if (cid:13) = 0 he idivid a y aia 0; ah ae
ee wih ee
idivid a y aia dea ; ii a y if (cid:13) = he ; ah ae
ee wih
ee
idivid a y aia dea . A ei f e iee w d be eab ih
if hee i e (cid:13) = f whi
h (cid:13) ; ah ae
ee wih ee
idivid a y aia dea ad wih he axi egh f
h a
a
ah b ded
by a yia f
i f .
A
kw edgee
The a h hak he eviewe f a ea ie vei f hi ai
e f hei va ab e
e ad ggei whi
h have
ib ed igi(cid:12)
a y i
e ad gaiai.
The wk eed i hi ai
e wa
aied de he f ESRC Ga
GR/R60836/01.
Refee
e
Abb . . a
ha ki . 1991. he
i f ake i he bx
de.
Ui ia aheai
a 40 97{116.
Aki A. 1970. he ea ee f ie a iy. . E
. They 2 244{263.
Chaea eaf A. Ga jd T. Wi hie . . 2002. The i
i e f g diiihig
afe. . E
. They 103 311{333.
11. A vey e iiay iveigai f
exiy heei
ei aiig i eig wih a
aive
exea iie i eeed i D e 2004 whee hee ae efeed a \
ex{deede":
h
i iy f
i aea have bee eg e
ed i he
aia ad a gihi
aa yi f e
e
a
ai b e a h gh he idea i we kw gae heei
de i e
i
f whi
h
he e \a
aive exea iy" igiae.
76
Exea Behavi i iage Ca
egiai
Dig F. Geave . 2000. e i Age C i
ai V . 1916 f CS.
Sige Ve ag.
D e . 2003. evai
ai i di e
.
. ih eaia Cf.
A.. ad aw CA03 . 12{21 Edib gh. AC e.
D e . 2004. Cex deede
e i iage e
e a
ai.
. ECA04
. 1000{01 Va e
ia.
D e .
B ey . 2003. ia ea
e i dia g e
.
. Se
d eaia i Cf. A Age ad iage Sye AA
AS03 . 608{615. AC e.
D e . W didge . a e
e . 2003. The
exiy f
a
egiai.
Te
h. e. UCS 03 002 De. f C e S
ie
e Uiv. f ive . aea
Ai(cid:12)
ia e ige
e.
Edi U. a de . 2004a. he
i
ai
exiy f i aea ad
ig.
. Thid eaia i Cf. A Age ad iage
Sye AAAS04 . 622{629.
Edi U. a de . 2004b. We fae egieeig i iage ye. i
ii
A. ea . i . Ed.
. F h eaia Wkh Egieeig
S
ieie i he Age W d ESAW 2003 V . 3071 f A . 93{106. Sige
Ve ag.
Edi U. a de . Sadi F. Ti F. 2003. ia
e f egiai
ve e
e.
. Se
d eaia i Cf. A Age ad
iage Sye AAAS03 . 177{184. AC e.
a z W. . 1958. Ui dia
e e
he
kig
de. RE Ta. E e
i
C
e 7 179{180.
S. C. 1976. Ue a ie a iie. . E
. They 13 82{111.
a S. 2001. Saegi
egiai i iage evie. T e.
B ey . a S. W didge . 2002. Deideaa f ag eai
.
. Fi eaia . i Cf. A Age ad iage
Sye AAAS02 . 402{409. AC e.
ake D. C. Uga . . 2000a. eaive
biaia a
i: hey ad a
i
e.
. 17h aia Cf. Ai(cid:12)
ia e ige
e AAA 00 . 74{81.
ake D. C. Uga . . 2000b. eveig aegi
ai ai i ieaive
a
i: xy age ad i
e adj e.
. 17h aia Cf. Ai(cid:12)
ia
e ige
e AAA 00 . 82{89.
Re
hei . S. Z ki G. 1994. R e f E
e. T e.
Sadh T. W. 1998. Ca
ye f ai(cid:12)
ig ak a
ai: heei
a e .
AAA Sig Syi : Sai(cid:12)
ig de .
Sadh T. W. 1999. Diib ed aia de
ii akig. Wei(cid:25) G. Ed. ia
ge Sye . 201{258. T e.
77
D e
Sadh T. W. 2002. A gih f ia wie deeiai i
biaia
a
i. Ai(cid:12)
ia e ige
e 135 1{54.
Sadh T. W. S i S. 2003. Bb: ved wie deeiai i
biaia
a
i ad geea izai. Ai(cid:12)
ia e ige
e 145 33{58.
Sadh T. W. S i S. Gi i A. evie D. 2001. Cabb: A fa ia a gih
f
biaia a
i..
. CA 01 . 1102{1108.
Sih R. G. 1980. The
a
e
: high eve
i
ai ad
i a
diib ed b e ve. EEE Ta. C e C 29 12 1104{1113.
Teeh z . 2000. Se a
ab e
biaia a
i.
. 17h aia
Cf. Ai(cid:12)
ia e ige
e AAA 00.
Yk . Sak ai Y. a baa S. 2004. The e(cid:11)e
f fa e ae bid i
bi
aia a
i: ew fa d i iee a
i. Gae ad E
i
Behavi
46 1 174{188.
78
|
1506.07781 | 1 | 1506 | 2015-06-25T15:17:24 | Microsimulations of Arching, Clogging, and Bursty Exit Phenomena in Crowd Dynamics | [
"cs.MA",
"physics.soc-ph"
] | We present in this paper the behavior of an artificial agent who is a member of a crowd. The behavior is based on the social comparison theory, as well as the trajectory mapping towards an agent's goal considering the agent's field of vision. The crowd of artificial agents were able to exhibit arching, clogging, and bursty exit rates. We were also able to observe a new phenomenon we called double arching, which happens towards the end of the simulation, and whose onset is exhibited by a "calm" density graph within the exit passage. The density graph is usually bursty at this area. Because of these exhibited phenomena, we can use these agents with high confidence to perform microsimulation studies for modeling the behavior of humans and objects in very realistic ways. | cs.MA | cs |
Microsimulations of Arching, Clogging, and
Bursty Exit Phenomena in Crowd Dynamics
Francisco Enrique Vicente G. Castro and Jaderick P. Pabico
Institute of Computer Science
University of the Philippines Los Baños
College 4031, Laguna
{fevgcastro, jppabico}@uplb.edu.ph
ABSTRACT
We present in this paper the behavior of an artificial
agent who is a member of a crowd. The behavior is
based on the social comparison theory, as well as the
trajectory mapping towards an agent's goal considering
the agent's field of vision. The crowd of artificial agents
were able to exhibit arching, clogging, and bursty exit
rates. We were also able to observe a new phenomenon
we called double arching, which happens towards the
end of the simulation, and whose onset is exhibited by
a "calm" density graph within the exit passage. The
density graph is usually bursty at this area. Because
of these exhibited phenomena, we can use these agents
with high confidence to perform microsimulation studies
for modeling the behavior of humans and objects in very
realistic ways.
INTRODUCTION
1.
Arching is a rainbow-like structure that naturally forms
at the edge of a pedestrian crowd that jam and clog
at exits. Clogging results as an effect of competition
for space resource among members of a crowd who are
unable to pass each other. Bursty exit rates are a result
of jostling for position which prevents each pedestrian
smooth passage along the exit width [10]. These phe-
nomena are interesting to study and simulate because
they are the most commonly observed behavior in
crowd dynamics. In designing behavior for an agent in
a multi-agent system (MAS), it is very important that
the agents be able to exhibit these phenomena while
in a crowd. When a crowd of artificial agents exhibits
these phenomena, we then say that our microsimula-
tion approach is more akin to modeling humans and
real-world objects in very realistic ways. Thus, our
microsimulation can be used with higher confidence to
perform what-if scenarios to aid decision makers and
researchers.
The growing number of students being admitted to
various higher education institutions (HEI) in the
country in the last 10 years, coupled with the slow-
paced development of
infrastructure to support the
student population growth, particularly among the
state colleges and universities (SUC), has resulted
into crowding along building corridors that connect
classrooms, lecture halls, and laboratories, despite the
wishes of these institutions to provide better services
and safer environments. Nowadays, students do not
only contend for slots in subjects, which happen only
once per semester during registration, but more so with
walk spaces along the building corridors, which happen
every class day. Students contend with corridor spaces
whose nominal widths are effectively reduced because of
the presence of other students, tables, chairs, and other
pedestrian flow-retarding objects. Because the nominal
widths of the corridors have become smaller than their
potential widths, the effective travel times of students'
egress along these corridors are lengthened, specifically
during periodic flash crowd situations, such as during
in-between class hours.
If the walk-time of students
along the building corridors is lengthened, specifically
during normal egress conditions, how much more will
be added to the evacuation time during chaotic panic
situations (i.e., the faster-is-slower phenomenon) such
as when there is a fire, or during an earthquake, or
during riots or fraternity rumbles, or during critical
incidents like chemical or biological spills? One need
not wait for a disastrous event to answer this ques-
tion because similar events can be simulated using a
computer without compromising human lives.
In this paper, we present our design of an artificial
agent whose behavior resembles that of the humans
when walking (or traveling) within a crowd. Our argu-
ment here is that if the crowd of our artificial agents
can exhibit the phenomena that are usually observed in
real-life, then we can use our agents in microsimulation
studies of the dynamics of humans under normal egress
and panic escape conditions. We present in this paper
the behavior of our artificial agent, which is based on
two recently developed theories: The social comparison
theory [4] and a trajectory mapping towards an agent's
goal considering the agent's field of vision. With this
behavior, our crowd of artificial agents was able to
exhibit the arching, clogging, and bursty exit rates
phenomena. The arch exhibited at the edge of the
crowd resembles that of a half ellipse, with the major
Figure 1: An example output frame of the MVS. The upper Left part contains the input image of
student pedestrians with trackers (bounding boxes). The upper Right part contains the trajectory of
the pedestrians. The lower part contains the distance vs. time graph of the trajectory. This figure is
in color in the electronic copy of this paper printed with permission from Ngoho and Pabico [12] and
the Philippine Society of Information Technology Educators (PSITE).
axis coincident along with the direction of crowd flow,
while the minor axis is parallel to the exit width. We
further present the effect of varying exit widths on the
respective lengths of the major and minor axes in the
arching phenomenon.
2. RECENT ADVANCES IN
CROWD DYNAMICS
In the past years, we developed a machine vision
system (MVS) to automatically capture the dynamics
of pedestrian students in a university building under
four different traffic scenarios [12]. We considered the
overhead view of each student as a digital object, where
our MVS processes the image sequences to track the
students. By considering the interactive effect of the
camera lens perspective and the projected area of the
corridor, the distance of each tracked student from
its original position to its current position is approx-
imated every video frame. The quantified motion of
each tracked student are output by our MVS using
2-dimensional graphs of the kinematics of motion as a
visualization (see for example Figure 1).
Because of the importance of crowd dynamics on several
real-world applications, several researchers attempted
to quantify the collective dynamics of the pedestrians
through developed simulations of motion. Helbing
and Molnar [7], borrowing some ideas in gas-kinematic
models,
introduced the social force model (SFM) to
simulate pedestrian flows. In their model, a self-driven
particle (i.e., a pedestrian) that interacts through social
rules and regulations tries to move in its desired speed
and direction while at the same time attempts not to
collide with obstacles, other particles, and surrounding
barriers. In order to reach its destination faster, pedes-
trians take detours even if the route is crowded [9]. The
choice, however, is dependent on the recent memory of
what the traffic was like the last time they took the
route, the figure of which was found by other researchers
to be polygonal in nature [5].
In agreement with the
social force model, Weidmann [14] observed that, as
long as it is not necessary to go faster, such as going
down a ramp, a pedestrian prefers to walk with his
or her desired speed, which corresponds to the most
comfortable walking speed. However, Weidmann [14]
further observed that pedestrians keep a certain dis-
tance from other pedestrians and borders. The distance
between the pedestrians decreases as the density of
the crowd increases. The pedestrians themselves cause
delays and obstructions. Arns [1] observed that the
motion of the crowd is similar to the motion of gases
and fluid, while Helbing et al. [9] suggested that it is
similar to granular flow as well.
Helbing et al. [9], in his extension of the SFM, showed
that many aspects of traffic flow can be reflected by
self-driven many-particle systems.
In this system, he
identified the various factors that govern the dynamics
of the particles such as the specification of the desired
velocities and directions of motion, the geometry of the
boundary profiles, the heterogeneity among the parti-
cles, and the degree of fluctuations. One such observ-
able pattern is the formation of lanes of uniform walking
direction, formed because of the self-organization of the
pedestrians [6]. Aside from the self-organizing behavior
of the crowd, obstacles were also observed to both posi-
tively and negatively contribute to the flow of the traffic.
During escape panic of large crowds, several behavioral
phenomena were observed [8]: build up of pressure, clog-
Figure 2: Snapshots of an example what-if scenario simulations on the effect of exit door configurations
to student egress in a very large lecture hall (students are shown as black circles with pointers to
visualize heading): (a) Current exit door configuration which opens outward but hinders student flow
towards the corridor exit; (b) Effect of moving the door hinge such that the exit door opens behind
the student flow; and (c) Effect of increasing the effective width of the exit door by 50% and replacing
the swinging door by a sliding door and notice the effect on the utilization of the nominal corridor
width. This figure is in color in the electronic copy of this paper. The artwork is licensed under a
Creative Commons Attribution-NonCommercial-NoDerivs 3.0 Unported License.
ging effects at bottlenecks, jamming at room widening
areas, faster-is-slower effect, inefficient use of alterna-
tive exits,
initiation of panics by counter flows, and
impatience. It was observed that the main contributing
behaviors in these situations is a mixture of individual
and grouping behavior.
3. METHODOLOGY
3.1 General Behavior
In our simulation, each agent was provided with
the following general behaviors described originally
by Wooldridge and Jennings [15], and later on extended
and explained further by Epstein [2], Franklin and
Graesser [3], Torrens [13] and Macal and North [11]:
1. Autonomy: We program our agents as autonomous
units (i.e. governed without the influence of cen-
tralized control). Each agent is capable of pro-
cessing information and exchanging this informa-
tion with other agents in order to make indepen-
dent decisions. They are free to interact with
other agents, at least over a limited range of situ-
ations, and this does not (necessarily) affect their
autonomy. In this respect, we say that our agents
are active rather than purely passive (see item
number 3 below).
2. Heterogeneity: We believe that the notion of
mean-individuals is redundant. Instead we believe
that our agent's programming permit the devel-
opment of autonomous individuals. When these
agents are put into a group, such as in a crowd,
the formed group is a heterogenous one composed
of different yet
similarly-programmed agents.
Therefore, groups of agents can exist, but pro-
grammatically the groups are spawned from the
bottom-up, and can be clearly seen as amalga-
mations of
similarly-programmed autonomous
individuals.
3. Active: Our agents are active because they
exert independent influence in a simulation. For
example, an agent can directly affect the decision
made by other agents, or its very presence in the
simulated environment can greatly affect changes
in that environment and thus indirectly affect
other agents. Because of this, we can identify the
following active features of our agents:
(a) Pro-active or goal-directed
(b) Reactive or perceptive
(c) Bounded rationality
(d) Interactive and communicative
Because our agents exhibit these behaviors, our sim-
ulation approach is more akin to modeling humans
and objects in very realistic ways than other modeling
approaches, such as those that aggregate mathemat-
ical equations to explain the dynamics of pedestrian,
say, during panic situations or emergency evacuations.
Thus, our simulation can be used with higher confidence
to perform what-if scenarios to aid, for example, uni-
versity administrators and decision makers with regards
to management policies, as well as infrastructure devel-
opment plans, for safer learning environments for the
students and constituents (see for example Figure 2).
3.2 Agent Behavior
In a two-dimensional flat world, let G be the goal of all
agents with set coordinates {(x1, y1), (x2, y2), . . . , (xn, yn)}.
This set is usually the location of an exit door. Since
all agents aim to reach any one of the exit coordinates,
the agents will face towards the direction of the nearest
the simulation world. Until this distance has not been
achieved, the agent will just repeat the agent's decision
making process described above.
3.3 Effect of Exit Width on
Arching Profile
We conducted experiments using a crowd of 400 agents
to see the effect of various widths of the exit door to the
length of the arch's minor and major axes. Based on our
preliminary simulations, 400 was the minimum number
of agents that can exhibit arching, given each agent's
gait speed and a corridor width of 19. The width is in
terms of agent's width and the various exit widths that
we used were: 1, 3, 5, 7, 9, 11, and 13. The various exit
widths were computed by adding one person width on
each side of the exit along the minor axis.
4. RESULTS AND DISCUSSION
4.1 Arching Effects
Figure 3a shows the profile of a crowd that exhibits
arching near the exit door. We observed that the arch
resembles a half ellipse with the ellipse's major axis run-
ning along the direction of the crowd flow and perpen-
dicular to the exit width, while the minor axis is par-
allel to the exit width. This result confirms that of the
other researchers' [2, 3, 9]. One particular result that
we feel contributed new knowledge on this field is our
observation of the occurrence of what we called "double
arches" that formed towards the end of the simulation
(Figure 3b). As far as we are concerned, none of the
literature that we reviewed described this occurrence.
The double arches appeared when the clogging at the
exit loosened up, and more agents that are near and
along the major axis were able to pass through. We
also observed that the burstiness of the exit rate disap-
peared at the start of the formation of double arches.
We described this disappearance of burstiness as "calm"
in the visualization of the phenomenon using the density
graph at the exit area.
4.2 Clogging
Clogging results when two or more agents compete for
the same corridor space and not one of them are able
to pass each other. Figure 4a shows a snap-shot of the
simulation where clogging occurred. The accompanying
density graph at the exit door (Figure 4b) confirms that
no agent was able to pass the exit. As opposed to the
event of a bursty exit rate where agents temporarily clog
the exit door but due to emergent behavior are able to
unclog themselves, clogging occurred because the agents
were not able to resolve their conflict in grabbing the
space resource just in front of them.
4.3 Bursty Exit Rate
Figure 5 shows the density graph of the crowd at the exit
door. The saw-tooth-like form of the density graph con-
firms that all exit scenario exhibit a bursty rate. The
burstiness can be attributed to agents competing for
position which for a brief moment clogged the exit way.
However, through the agents' emergent behavior, they
were able to unclog themselves, and then pass the exit
door in a burst carrying more agents. This results to
Figure 3: Profile of the crowd before the exit
door: (a) Onset of arching where the minor and
major axes are visible; and (b) Occurrence of
double arches toward the end of the simulation.
Figure 4: (a) Profile of the crowd at a clogged
exit door.
(b) Density graph at the exit door
shows that no agent was able to pass the exit
during the clogging effect.
in this paper,
exit coordinates from their current coordinates. The
agent will scan its field of vision for the closest free
space and move towards that space with its gaiting
speed, which,
is currently set at one
pace per simulation time step. In the future, we wish
to vary this gaiting speed depending on the agent's
height. A free space is a location in the flat world that
is not a wall, another agent, or any other object.
If
other agents are blocking the agent's space within its
field of vision, the agent stops at that coordinates. It
is possible that the agent might move away from the
target exit coordinate if the chosen free space was at
the edge of its field of vision. When that happens,
the agent will still move to the free space, but after
moving, it will redirect its heading towards the possibly
new nearest exit coordinates. When the distance of the
agent to the nearest exit coordinates is < 1, it considers
itself as already exited and will move to the edge of
rate graph. We also described the behavior of the den-
sity graph during clogging, and described clogging to
have occurred when two or more agents compete for
space near the exit. We have also shown that the prop-
erty of bursty exit rate is exhibited by the density graph
with a saw-tooth-like profile. Lastly, we conducted an
experiment on the effect of various exit door widths on
the arching profile of the crowd before the exit door. At
a narrow exit width, the major axis of the arch tends
to be longer, while the minor axis tends to be shorter.
However, at a wide exit width the major axis of the
arch tends to be shorter, while the minor axis tends
to be longer. Because our crowd of agents were able
to exhibit the arching, clogging, and bursty exit rate
phenomena in crowd dynamics, we can conclude that
we can use these agents in microsimulation studies for
modeling the behavior of humans and objects in very
realistic ways. Thus, we can use these microsimulations
with higher confidence to perform what-if scenarios to
aid in decision making.
6. EXTENSION
The following efforts are already underway as extensions
to this research endeavor:
1. As we explained earlier, the arching phenomenon
is an emergent pattern formed by a c-sized crowd
of intelligent, goal-oriented, autonomous, hetero-
geneous individuals moving towards a w-wide exit
along a W -wide corridor, where W > w. We are
currently collecting empirical data from microsim-
ulations to identify the combination effects of c
and w to the time T of the onset of and the size S
of the formation of the arch. We aim to measure
the S with respect to the lengths of the major and
minor axes of the ellipse, respectively.
2. The total time T of egress of large crowds, such
as students exiting a large lecture hall filled to its
maximum capacity, is hindered by the number of
exit doors, as well as the doors' dimensions, posi-
tions, and orientation. In this current endeavour,
we aim to find out the combinatorial effects of the
number of exit doors, their dimensions, and their
orientations using microsimulations of our agents.
For our simulations to be realistic, we will scale
down a real physical structure of a large lecture
hall, including the dimensions of its exit doors,
into a two-dimensional simulation. We plan to
fill the simulated lecture hall with agents corre-
sponding to the full capacity of the real lecture
hall. We will then conduct experiments on various
crowd sizes C and different exit door scenarios D.
We will then measure T under all C × D combi-
nations.
3. We would like to compare the capabilities of the
SFM and our approach in simulating the following
real-world crowd phenomena: "faster-is-slower" in
escape panic, "arching" and "bursty exit" as side
effects to "clogging" on exit ways, "flocking,""bidi-
rectional lane formation," and "roundabout forma-
tion." We believe that our approach is also able
to exhibit two more individual behaviors that the
Figure 5: Density graph at the exit door shows
a bursty exit rate.
abrupt change of density at the exit area, and is visual-
ized as a saw-tooth-like line in the density graph.
4.4 Effect of Exit Width on
Arching Profile
Figure 6 shows the effect of varying the width of the
exit door on arching profile. We can see here that at
narrower exit widths, the major axis is longer, while
the minor axis is shorter. When the exit door is wider,
the major axis is shorter than normal, while the minor
axis is longer than normal.
Figure 6: Arch profile at various door widths:
(a) 1; (b) 3; (c) 5; (d) 7; (e) 9; (f ) 11; and (g)
13.
5. SUMMARY AND CONCLUSION
In this paper, we present the behavior for an artificial
agent used in microsimulation of pedestrian crowd. The
behavior is based on the two recently developed theo-
ries, the social comparison theory [4] and a trajectory
mapping towards an agent's goal considering the agent's
field of vision. With these behaviors, we were able
to show that the crowd with these agents was able to
exhibit arching, clogging, and bursty exit rate. We also
observed a new phenomenon that occurred towards the
end of the simulation that we called Double Arching.
As far as our literature review is concerned, we are
the first to describe the emergent behavior of double
arching in the crowd dynamics, which happen during
the temporary "calm" that occurred in the bursty exit
[12] L.V.A. Ngoho and J.P. Pabico. Capturing the
dynamics of pedestrian traffic using a machine
vision system. Philippine Information Technology
Journal, 2(2):1 -- 11, 2009.
[13] P.M. Torrens.
Simulating Sprawl: A Dynamic
Entity-Based Approach to Modelling North Amer-
ican Suburban Sprawl Using Cellular Automata and
Multi-Agent Systems. PhD thesis, University Col-
lege London, London, 2004.
[14] U. Weidmann. Transportation technique for pedes-
trians, 1993. Zurich, Switzerland.
[15] M. Wooldridge and N.R. Jennings.
Intelligent
agents: Theory and practice. Knowledge Engi-
neering Review, 10(2):115 -- 152, 1995.
SFM can not do: (1) Imitation -- where individ-
uals tend to move into groups whose members they
thought would have the same opinion as theirs;
and (2) Contagion -- where people tend to "adopt"
the behavior of others in the same group.
Acknowledgment
This research effort is under the research program
Multi-agent Simulation of Student Egress from
the ICS Mega Lecture Hall Under Normal, Con-
trolled Emergency, and Panic Situations funded
by the Institute of Computer Science, University of the
Philippines Los Banos.
References
[1] T. Arns. Video films of pedestrian crowds, 1993.
Wannenstr. 22, 70199 Stuttgart, Germany.
[2] J.M. Epstein. Agent-based computational models
and generative social science. Complexity, 4(5):41 --
60, 1999.
[3] S. Franklin and A. Graesser. Is it an agent, or just a
program? A taxonomy for autonomous agents. In
Proceedings of the Third International Workshop
on Agent Theories, Architectures, and Languages.
Springer-Verlag, 1996.
[4] N. Fridman and G.A. Kaminka. Modeling pedes-
trian crowd behavior based on a cognitive model
of Social Comparison Theory.
Computational
and Mathematical Organization Theory, 16:348 --
372, 2010.
[5] J. Ganem. A behavioral demonstration of Fermat's
principle. The Physics Teacher, 36, 1998.
[6] D. Helbing. Safety Management at Large Events:
The Problem of Crowd Panic. Institute for Trans-
port and Economics, Dresden University of Tech-
nology, 2006.
[7] D. Helbing and P. Molnar. Social force model for
pedestrian dynamics. Physical Review E, 51(5):
4282+, 1995.
[8] D. Helbing, I. Farkas, and T. Vicsek. Simulating
dynamical features of escape panic. Nature, 407
(6803):487 -- 490, 2000.
[9] D. Helbing, P. Molnar, I. Farkas, and K. Bolay.
Self-organizing pedestrian movement. Environment
and Planning B: Planning and Design, 28:361 -- 383,
2001.
[10] C.M. Henein and T. White. Agent-based Modeling
of Forces in Crowds, volume 3415, pages 173 -- 184.
2005.
[11] C.M. Macal and M.J. North. Tutorial on agent-
based modelling and simulation.
In M.E. Euhl,
N.M. Steiger, F.B. Armstrong, and J.A. Joines,
editors, Proceedings of the 2005 Winter Simulation
Conference, 2005.
|
1604.04725 | 1 | 1604 | 2016-04-16T11:01:44 | Unanimously acceptable agreements for negotiation teams in unpredictable domains | [
"cs.MA",
"cs.AI",
"cs.GT"
] | A negotiation team is a set of agents with common and possibly also conflicting preferences that forms one of the parties of a negotiation. A negotiation team is involved in two decision making processes simultaneously, a negotiation with the opponents, and an intra-team process to decide on the moves to make in the negotiation. This article focuses on negotiation team decision making for circumstances that require unanimity of team decisions. Existing agent-based approaches only guarantee unanimity in teams negotiating in domains exclusively composed of predictable and compatible issues. This article presents a model for negotiation teams that guarantees unanimous team decisions in domains consisting of predictable and compatible, and also unpredictable issues. Moreover, the article explores the influence of using opponent, and team member models in the proposing strategies that team members use. Experimental results show that the team benefits if team members employ Bayesian learning to model their teammates' preferences. | cs.MA | cs |
Unanimously Acceptable Agreements for Negotiation Teams in
Unpredictable Domains
Victor Sanchez-Anguixa, Reyhan Aydoganb, Vicente Juliana, Catholijn Jonkerb
aDepartamento de Sistemas Inform´aticos y Computaci´on
Universitat Polit`ecnica de Val`encia
Cam´ı de Vera s/n, 46022, Valencia, Spain
bInteractive Intelligence Group
Delft University of Technology
Delft, The Netherlands
Abstract
A negotiation team is a set of agents with common and possibly also conflicting preferences that forms one
of the parties of a negotiation. A negotiation team is involved in two decision making processes simulta-
neously, a negotiation with the opponents, and an intra-team process to decide on the moves to make in
the negotiation. This article focuses on negotiation team decision making for circumstances that require
unanimity of team decisions. Existing agent-based approaches only guarantee unanimity in teams negoti-
ating in domains exclusively composed of predictable and compatible issues. This article presents a model
for negotiation teams that guarantees unanimous team decisions in domains consisting of predictable and
compatible, and alsounpredictable issues. Moreover, the article explores the influence of using opponent, and
team member models in the proposing strategies that team members use. Experimental results show that
the team benefits if team members employ Bayesian learning to model their teammates' preferences.
Keywords: Automated negotiation, Multi-agent systems, Agreement technologies
1. Introduction
In the last decade, there has been an increase
in the profit earned by electronic commerce sys-
tems. This increase has lead to a strong interest
of the academic world in researching problems re-
lated to e-commerce (Ngai and Wat, 2002; Grieger,
2003; Wareham et al., 2005). As of today, most e-
commerce systems rely on users manually browsing
their catalogs and selecting which goods they desire
to buy. This task may end up being time consuming
and suboptimal in terms of users' preferences, es-
pecially as the number of items and services offered
on the Web increases. Therefore, it is necessary
to propose mechanisms that helps costumers take
better decisions while saving their time efforts.
Agent-based electronic commerce has been pro-
posed as a solution to such problems (Guttman
Email addresses: [email protected] (Victor
Sanchez-Anguix), [email protected] (Reyhan
Aydogan), [email protected] (Vicente Julian),
[email protected] (Catholijn Jonker)
et al., 1998; Sierra and Dignum, 2001; Oliveira and
Rocha, 2001; He et al., 2003). In an agent-based e-
commerce system, autonomous agents act on behalf
of their users with the goal of finding and closing
satisfactory deals. Automated negotiation is one
of the most common approaches when implement-
ing these systems since they allow different elec-
tronic parties to reach agreements by exchanging
offers and feedback (Lomuscio et al., 2003; Nguyen
and Jennings, 2005; Buffett and Spencer, 2007; lau,
2007; Chan et al., 2008). The benefits of automated
negotiation and agent-based e-commerce are many.
Being brief, some of the most important include:
• As stated, browsing online catalogs for an op-
timal deal may be time consuming. The state-
of-the-art in automated negotiation can com-
plete complex negotiations for multiple issues
in less than a few minutes (Klein et al., 2003a;
Williams et al., 2011; Baarslag et al., 2012).
• On the one hand, automated negotiation saves
the user from having to browse the entire cata-
Preprint submitted to Elsevier
September 21, 2018
log. Additionally, its personal agent is directed
by the preferences of the user in the negotia-
tion, which should result in deals that are ad-
justed to the personal liking of the individual.
Personalization has been reported to increase
user satisfaction in many computational sys-
tems (Ball et al., 2006; Liang et al., 2007). On
the other hand, a dynamic process like auto-
mated negotiation allows sellers to adapt their
deals to the users' preferences, their current
business needs, and their competitor dynamics
(He et al., 2003).
• Agreements achieved by human negotiators,
suffer from the leaving money on the nego-
tiation table effect (Thompson, 2003). This
means that human negotiators are content
with current agreements, which are usually
suboptimal, when they could have performed
much better. Agents in automated negotiation
have been reported to provide agreements close
to the optimal solution (Lai et al., 2008).
• Compared to centralized and offline ap-
proaches (e.g., preference aggregation, recom-
mendation approaches, etc.), automated nego-
tiation is a dynamic and parallel process. For
instance, some centralized approaches like pref-
erence aggregation are computationally hard
especially if the preference space is combina-
torial (Chevaleyre et al., 2007). On the other
hand, recommendation approaches only filter
prospective deals, but they do not close specific
contracts adapted to business needs. Contrar-
ily, automated negotiation can be adapted to
current business needs (e.g., concede to gain
customers and close fast deals). Additionally,
as stated above, team members are also moti-
vated by their own personal interests. There-
fore, it is possible that some team members
show opportunistic behavior inside the team.
In such cases, preference aggregation may be
manipulated by exaggerating preferences. Ad-
ditionally, each parties' preferences are private,
therefore making it difficult for the other par-
ties to exploit and manipulate. This latter fac-
tor is important, since nowadays most users in
electronic applications care about the informa-
tion they filtrate in systems (Taylor, 2003).
Most negotiation mechanisms proposed for e-
commerce settings have focused on solving bilat-
eral or multiparty negotiations where parties are
2
individual agents (Faratin et al., 1998; Zeng and
Sycara, 1998; Klein et al., 2003b; Nguyen and Jen-
nings, 2005; Coehoorn and Jennings, 2004; Buf-
fett and Spencer, 2007; Lai et al., 2008; Williams
et al., 2011; Sanchez-Anguix et al., 2013; Aydogan
and Yolum, 2012). However, some real life scenar-
ios involve negotiation parties that are not neces-
sarily formed by single individuals. Instead, each
party may be formed by more than a single individ-
ual. For instance, imagine that a group of travelers
wants to go on a holiday together. As a group, they
have to negotiate with several travel agencies to get
the best travel package for the group. Despite shar-
ing a common goal, each member in the multiplayer
party may also be motivated by its own personal
interests Mannix (2005); Halevy (2008). Therefore,
the group not only faces a possibly difficult nego-
tiation with the travel agency, but it also needs to
deal with the conflict present in the group. This
type of multi-individual negotiating party has been
studied in the social sciences under the name of ne-
gotiation team (Thompson et al., 1996; Brodt and
Thompson, 2001).
As far as the authors are concerned, multi-
individual parties have been overlooked in auto-
mated negotiation research. The use of computa-
tional models for negotiation teams opens doors for
new types of interesting and novel applications in
electronic commerce. The inclusion of agent-based
negotiation teams allows for e-commerce systems to
deploy dynamic deal mechanisms for groups, mak-
ing of e-commerce a more social system. Classi-
cally, when purchasing for groups in e-commerce
systems, one representative takes decisions for the
whole group. Either he makes decisions accord-
ing to his own preferences or the group needs to
engage in a human negotiation which is usually a
costly process due to different schedules, logistics,
lack of communication problems or interpersonal
conflict (Behfar et al., 2008). With the inclusion
of agent-based negotiation teams these problems
are eluded since autonomous agents take decisions
jointly while saving time and efforts for their users.
We believe that agent-based negotiation teams
could provide potentially interesting new services :
• Electronic markets for groups of travelers:
Online travel agencies offer their services by
means of online catalogs where users can
browse different products like flights, hotels,
restaurants, activities, etc.
The possibili-
ties for travels are vast, and usually a single
travel operator may offer thousands of possi-
ble trip packages/services. Exhaustively look-
ing through this online catalog for an optimal
deal becomes an unfeasible task for humans.
Additionally, more often than not, travel is a
social activity for groups (e.g., friends, family,
young people, etc.). Users can benefit from
agent-based negotiation teams since they can
exhaustively look for deals while taking the
preferences of the group into account and sav-
ing efforts. Service providers can also bene-
fit from these models since they could adapt
their business strategies in a dynamic way and
add a level of personalization that may help to
retain customers. Moreover, offering the pos-
sibility for groups to close travel deals based
on their preferences is a value-added service,
that as far as we know, is not currently of-
fered by the industry. As an example of its
application, users may indicate to their per-
sonal agents their desire to go on a travel to-
gether. Then, the agents prepare to negotiate
with different travel agencies in order to pro-
vide a complete and satisfactory travel package
for the users. The fact that the negotiation is
carried out automatically by electronic agents
also gives room to looking for several alterna-
tives in parallel. Once several trip packages
have been negotiated, the personal agents may
communicate the agreements to users, who can
validate them in the last instance.
• Electronic support for agricultural coopera-
tives: Agricultural cooperatives are supposed
to be democratic institutions where groups of
farmers join together to save resources for the
distribution of their products. One of the main
problems of agricultural cooperatives is the
principal-agent problem (Ortmann and King,
2007). Basically, despite being democratic in-
stitutions, agricultural cooperatives are man-
aged by a board of directors who take deci-
sions on behalf of the democratic institution.
It has been reported in the literature (Ort-
mann and King, 2007) that dissatisfaction in
cooperatives comes from the fact that the goals
of members are not aligned with those of the
managers. As a novel application for electronic
commerce, agent-based negotiation teams may
provide support for the processes that are car-
ried out by cooperatives. For instance, the
negotiations between agricultural cooperatives
and distributors may be supported by an elec-
tronic market where the agricultural coopera-
tive is modeled as an agent-based negotiation
team. Each member may be represented by an
electronic and personal agent that participates
in the negotiation team according to the pref-
erences of its owner. This way, if the model
is capable of ensuring unanimity with regards
to team decisions, it may be possible to avoid
the principal-agent problem. Of course, agri-
cultural cooperatives are large institutions and
considerable research has still to be done to
provide scalable and fair computational mod-
els. However, research as the one presented in
this article contributes to the obtention of such
models in the long term.
• Groups of energy producers in the smart grid:
The smart grid is addressed to be the next
generation network for electricity distribution
(Farhangi, 2010).
In this network, energy
generation may come from geographically dis-
tributed small generators (e.g., green energy
generators) that have to compete with large
energy producers. Decisions at the smart grid
have to be taken dynamically since energy pro-
duction and consumption may vary or face un-
expected events (Ramchurn et al., 2012). Re-
cently, agent-based electronic commerce has
been proposed as proper paradigm for this sce-
nario due to its dynamic nature and adaptive
response (Brazier et al., 2002; Lamparter et al.,
2010; Morais et al., 2012; Ramchurn et al.,
2012).
If small generators want to compete
with large generators like power plants, they
may need to group together and act together
as a single generator. Agent-based negotiation
teams can give support for the group decision
making of small generators in a dynamic en-
vironment like the smart grid. For instance,
an agent-based negotiation team for the smart
grid may decide on different contract attributes
like energy price for different time slots, con-
tract duration and cancellation fees with dif-
ferent energy consumers.
The applications described above present benefits
for electronic commerce systems. However, there
are still several issues that need to be solved for de-
ploying real applications based on agent-based ne-
gotiation teams due to the novelty of the topic. One
of the main issues that should be addressed when
3
designing agent-based negotiation team models is
unanimity. The authors argue that, whenever it is
possible, it is desirable for the final agreement with
the opponent to be unanimously acceptable for all
of the team members. When the members of the
negotiation team are going to interact in the long
term, the intra-team strategy should avoid one or
some of the team members being clearly at disad-
vantage (e.g., unacceptable deal) with respect to
the other team members.
In the first place, the
aforementioned situation may end up in users per-
ceiving unfairness, which may affect commitment
to the decision, group attachment, and trust (Ko-
rsgaard et al., 1995). And second, but not the least
important, users that are not satisfied with agree-
ments found automatically may end up leaving the
electronic commerce application.
The existing approaches (Sanchez-Anguix et al.,
2011, 2012a,b) have focused on achieving unani-
mously acceptable agreements for negotiation do-
mains exclusively comprised by predictable and
compatible issues among the team members. An
issue is predictable and compatible if the preference
order over issue values is the same for team mem-
bers and this fact is known from the domain (e.g.,
price in a team of buyers). While some e-commerce
domains are exclusively composed by these issues,
many domains also contain issues whose preferen-
tial ordering over issue values is not known from the
domain (i.e., unpredictable issues). For instance, it
is difficult to predict from a set of cities which rank-
ing represents the preferences of a traveler, which
can diverge from the preferences of other travelers.
This article advances the state of the art in agent-
based electronic commerce in two different ways.
Firstly, it introduces a new model for agent-based
negotiation teams, which could support dynamic
negotiations for groups of autonomous agents rep-
resenting their users. Secondly, the present model
is capable of assuring that the final agreement is
unanimously acceptable for all of the team mem-
bers in domains that contain both predictable and
compatible and unpredictable issues. We propose an
intra-team protocol in which a team mediator helps
team members to reach unanimously acceptable de-
cisions. Furthermore, we propose two negotiation
strategies for team members: a basic negotiation
strategy based on concession tactics and a nego-
tiation strategy using Bayesian learning to model
teammates' and opponent's preferences for unpre-
dictable issues. The model is capable of outper-
forming state-of-the-art approaches for agent-based
4
negotiation teams. We describe our general frame-
work in Section 2 and the intra-team protocol that
allows team members to reach unanimity in Section
3. After that, we propose two negotiation strategies
for team members in Section 4 and we explain why
unanimity is guaranteed among team members in
Section 5. After analyzing the experiments in Sec-
tion 6, we relate our work to existing approaches
and discuss future lines of work in Section 7.
2. Overview of the Negotiation Framework
Let A represent a negotiation team consisting of
A = M different team members and a trusted
team mediator medA, and let a ∈ A represent a
team member in negotiation team A. Let op rep-
resent the opponent party of the negotiation team.
The negotiation between team and opponent is car-
ried out in a bilateral fashion, using an alternating-
offers protocol (Rubinstein, 1982).
In this proto-
col, one of the two parties is the initiating party
and sends the first offer to the other party or re-
sponding party. The responding party receives the
offer and decides whether or not he/she accepts the
offer. Accordingly, she or he may accept the cur-
rent offer or send a counter-offer. If the responding
agent sends a counter-offer, the initiating party has
to decide whether he/she accepts the counter-offer
or not. If the counter-offer is rejected, the process
is repeated in a turn-taking fashion until a deal is
mutually accepted (successful negotiation) or one
of the parties decides to quit the negotiation since
its deadline has been reached (failed negotiation).
Concerning inter-party communications, the team
mediator interacts with the opponent by sending
team's proposals and transmitting opponent deci-
sions to team members. The team mediator plays a
key role since it coordinates the team members and
helps them reach unanimously acceptable deals.
Let X be the object under negotiation, j ∈
{1, ..., n} be the issues under negotiation, Dj be the
negotiation domain or valid values for issue j and
xj ∈ Dj represent a valid value for issue j. Each
agent's preferences are represented by means of a
private additive utility function. We assume that
there is no preferential interdependency among ne-
gotiation issues; that is, the valuation given to a
certain issue does not affect preference on the valua-
tion of other negotiation issues. The utility function
for an agent in our framework can be formalized as
follows:
2.2. Types of negotiation issues among team mem-
U (X) = w1V1(x1)+w2V2(x2)+...+wnVn(xn). (1)
that
j=1
n(cid:80)
where wj represents the importance given to issue
j by the agent, and Vj : xj → [0, 1] is a scoring
function for issue j that gives the score that the
It is assumed
agent assigns to an issue value xj.
wa,j = 1 and wa,j ≥ 0, and then U (.) is
a function scaled in [0,1], where 0 represents the
least desirable negotiation deals, and 1 represents
the most desirable negotiation deals. For agents,
RU ∈ [0, 1] represents the reservation utility or the
minimum level of utility to consider an agreement
as acceptable.
In the proposed framework, private information
and bounded rationality are assumed. The former
has been introduced above: information regarding
agents' preferences is private, and so are the strate-
gies and minimum acceptable values of each agent.
This is true even among team members, since prior
to the negotiation they do not know any informa-
tion regarding other teammates' preferences. The
only information available is obtained via interac-
tions in the intra-team protocol. The latter refers
to the fact that given the limited time, informa-
tion privacy, and limited computational resources,
agents cannot calculate the optimal strategy to be
carried out during the negotiation.
Instead, they
employ heuristic strategies that aim to be as good
as possible in terms of the achievable utility.
2.1. Unanimously acceptable agreements
Each team member a ∈ A has a reservation util-
ity RUa ∈ [0, 1] that represents the minimum util-
ity that satisfies the team member's need. Each
outcome whose utility is lower than the reservation
utility is unacceptable for the team member. As
stated along this article, we consider that unanim-
ity in a negotiation team is of extreme importance.
An offer is unanimously acceptable for a team A if
it is acceptable for all of the team members inside
the negotiation team:
∀a ∈ A, Ua(X) ≥ RUa.
(2)
The proposed intra-team strategy will assure
that team members only accept those offers that
are unanimously acceptable for all the team mem-
bers and that offers proposed to the opponent are
over each team members' reservation utilities, thus,
making it unanimously acceptable.
5
bers
Among the different negotiation issues that com-
pose the negotiation domain, we consider that
there are issues that are predictable and compatible
among team members and issues that are unpre-
dictable among team members.
Formally, we can define an issue j with domain
Dj as compatible among team members if for each
possible pair of team members a, b ∈ A and for
each pair of issue values v1, v2 ∈ Dj, the following
expression is true:
Va,j(v2) > Va,j(v1) → Vb,j(v2) ≥ Vb,j(v1).
(3)
Hence, an issue is compatible among team members
if one of the team members can increase its utility
by selecting a certain issue value with respect to
the current assignment, then the rest of team mem-
bers stay at the same utility or they also increase
their utility. Thus, there is no preferential conflict
among issue values between the team members, and
there is full potential for cooperation among team
members with respect to compatible issues. Figure
1 shows two examples of compatible issues among
two agents (top part) and an example of a non com-
patible issue (bottom part). As it can be observed,
in the case of price (top left), both agents obtain a
better valuation when choosing a lower price value
with respect to a high price value. Thus, Equation
3 holds and it is a compatible issue for both agents.
In the case of the city of destination (top right), the
issue is also compatible among the two agents. For
any pair of cities, if one of the agents prefers one of
the cities with respect to other city, the other agent
also holds the same preferential relationship. For
instance, both agents prefer Paris to Berlin, Berlin
to London, and London to Madrid. However, in the
case of the type of room (bottom part), the blue
agent prefers an individual room with respect to an
apartment, whereas the red agent prefers exactly
the opposite. Thus, there is no full potential for co-
operation among team members in that negotiation
issue since conflict is present.
The concept of predictability and unpredictabil-
ity (Hindriks and Tykhonov, 2010; Marsa-Maestre
et al., 2013) is related to vertical and horizon-
tal
issues found in economics literature (Stole,
1995). The definition of predictable issues matches
with vertical
issues, while the definition of un-
predictable issues matches with horizontal issues.
From this point on, we will use the concepts of
Figure 1: Two compatible issues among two agents (top) and a non compatible issue among two agents (bottom).
unpredictable/predictable and we will briefly intro-
duce them. An issue is predictable for an agent if
the preference ordering of issue values is known in
the negotiation domain. Therefore, an issue is com-
patible and predictable among team members if the
preferences regarding issue values are known in the
negotiation domain and increasing the utility of one
of the team members by selecting one specific issue
value results in other team members staying at the
same utility or also increasing their respective util-
ities. For instance, from the examples in Figure 1,
one can consider that inside a team of buyers the
price is a compatible and predictable issue among
team members since it is known that all of the buy-
ers prefer low prices to high prices, and reducing
the price results in all of the buyers increasing their
utility or staying at the same utility. On the other
hand, an issue is unpredictable among team mem-
bers if the preference ordering of the issue values
cannot be accurately predicted and Equation 3 may
not hold for that issue.
In the case of Figure 1,
the city of destination is a compatible issue among
the two agents. Nevertheless, in a travel negotia-
tion domain it is not true that all of the travelers
will hold the same preference ranking over the issue
6
values and without additional knowledge, the pref-
erence ordering may not be predicted accurately.
Hence, it is an unpredictable issue. With respect
to the type of room, the preference ordering over
issue values may vary for the travelers. Moreover,
we cannot predict their preference ordering directly,
thus making the issue unpredictable.
In this framework, PR denotes the set of pre-
dictable and compatible issues among team mem-
bers, while UN denotes the set of unpredictable is-
sues.
2.3. Forbidden unpredictable partial offers among
team members
We define an unpredictable partial offer X
as
a partial offer that has a concrete instantiation
of all the unpredictable issues in UN. The utility
of an unpredictable partial offer is calculated as
Ua(X
) = (cid:80)
(cid:48)
(cid:48)
wa,jVa,j(xj).
j∈UN
(cid:48)
For a team member a ∈ A, an unpredictable par-
tial offer X
will never be part of an acceptable offer
(i.e., it will never be an unanimously acceptable of-
(cid:48)
fer for the team) when the sum of the utility of X
and the maximum utility that can be obtained from
10203040506070809010000.20.40.60.81Price ($)V(v)LondonParisMadridBerlinCity of destinationIndividualDoubleTripleApartmentType of room predictable issues maxP Ra = (cid:80)
j∈PR
wa,j is less than
(cid:48)
its reservation value RUa, since any full offer that
completes X
is below the reservation utility. For a
team member a, we refer to the set of unpredictable
partial offers that will never be part of an accept-
able offer as forbidden unpredictable partial offers,
Fa (see Equation 4).
Fa = {X
(cid:48)Ua(X
(cid:48)
) + maxP Ra < RUa}
(4)
It is worth noting that Fa does not represent the
whole negotiation space that is unacceptable for a,
but just a portion of it. In fact, some unpredictable
partial offers that are not contained in Fa, can be-
come unacceptable when the agent does not get the
value needed from predictable issues. The size of Fa
may grow as the reservation utility increases. Thus,
agents with high reservation utilities are expected
to have larger sets of Fa than agents with low reser-
vation utilities.
2.4. Case of Study
In this article we have employed a case of study
(i.e., a negotiation domain) that is extracted from
a possible tourism electronic market. The case of
study is used to illustrate and test the proposed
negotiation framework.
A group of travelers wants to go on a holiday
together and arrange their accommodation. The
group negotiates with a hotel on the following is-
sues.
• Price (p):
It represents the price per night
that each traveler pays to the hotel for the
booking service. The value goes from 200$,
which is the minimum rate applicable by the
hotel, to 400$, which is the maximum rate
found in the hotel. This negotiation issue is
considered to be predictable and compatible
among team members since all of the travel-
ers obviously prefer low prices to high prices.
Contrarily, the hotel prefers high prices to low
prices.
• Cancellation fee (cf ): This issue represents
the amount of the final price that each friend
pays if the reservation is canceled. Possible
values for this negotiation issue go from 0% to
50%. This is a predictable and compatible is-
sue among team members since all of the trav-
elers prefer low cancellation fees to high can-
cellation fees. On the contrary, the opponent
7
prefers high cancellation fees to low cancella-
tion fees.
• Arranged Foods Included (af ): The hotel
may also offer some meals included in deal with
the travelers. The type of meal plans included
are none, breakfast, breakfast+lunch, break-
fast+dinner, lunch+dinner, and all. In our ne-
gotiation scenario, we have considered that this
negotiation issue is unpredictable among team
members since preferences of team members on
this issue may vary and it cannot be assumed
to be same for each member.
• Type of room (tr ): The four travelers can
be accommodated in different types of room
depending on their preferences. More specif-
ically, the hotel offers 4 individual rooms, 2
twin rooms, 1 triple and 1 individual room, or
1 apartment. The type of room is an unpre-
dictable negotiation issue among team mem-
bers.
• Payment method (pm): The amount of
money paid by the travelers may be paid by
different methods. The hotel allows for the
payment to be made in cash, via credit card,
by bank transfer, in a 3 months deferred pay-
ment through the bank, and in a 6 months de-
ferred payment. This negotiation issue is un-
predictable since team members may prefer to
choose different payment methods and we can-
not predict their preference ordering directly.
• Room orientation (ro): If possible, the team
members can decide upon an orientation for
the balcony of their rooms. The different op-
tions are inner garden, main street, pool, sea,
and outer garden. This issue is also considered
an unpredictable issue among team members.
• Free amenity (fa): As a token of generos-
ity for booking as a group, the hotel offers one
free service to all of the team members. More
specifically, the team members can choose be-
tween gym service, free wi-fi, 1 free drink per
day, 1 free spa session, pool service, cable tv
service, and one free guided tour. Since the
preferences of team members vary for this is-
sue and no assumption about their preferences
can be made, this issue is also considered as
unpredictable.
To sum up, for this case study we have that
PR= {p,cf } and UN = {af ,tr ,pm,ro,fa} with a to-
tal of 4200 different combinations of discrete issue
values (af,tr,pm,ro,fa) and two real issues (p,cf ).
We assume that the team mediator knows which is-
sues are predictable and can apply an operator that
determines the best value for team members from a
given set. For unpredictable issues, team members
can have different types of valuation functions and
the mediator does not know which issue values are
better for team members. Each team member may
assign different weights (i.e., priorities) to negotia-
tion issues and the opponent's valuation functions
and issue weights may be different from those of
team members. The team mediator does not know
the weights given by agents to the different issues.
3. Intra-Team Protocol
In a negotiation involving a negotiation team,
the intra-team protocol defines how and when de-
cisions are taken regarding the negotiation. In this
framework, we propose an intra-team protocol that
is governed by the trusted team mediator medA.
Basically, the team mediator regulates the interac-
tions that can be carried out among team mem-
bers and, accordingly, helps team members reach-
ing unanimous acceptable decisions inside the team
during the negotiation. The proposed protocol is
clearly differentiated into two different phases: Pre-
negotiation and Negotiation. On the one hand, dur-
ing the pre-negotiation, the mediator helps team
members identifying potential offers that are not
unanimously acceptable for every teammate. On
the other hand, during the negotiation the mediator
coordinates the offer proposal mechanism, which is
composed of a voting process for unpredictable is-
sues and an iterated building process for predictable
issues, and the offer acceptance mechanism for of-
fers that come from the opponent. We describe
those phases in a detailed way in the following sec-
tions 3.1 and 3.2. An overview of all of the com-
munications carried out in the negotiation model
are depicted in Figure 2. It specifies the protocols
carried out within the team and the communica-
tions carried out with the opponent by means of
Agent UML (Bauer et al., 2001) sequence diagrams.
More detailed views of the intra-team protocols for
the pre-negotiation, evaluation opponent's propos-
als and proposing offers can be observed in Figures
3, 4, and 5 respectively.
3.1. Pre-negotiation Phase
In the pre-negotiation phase, the mediator co-
ordinates the following intra-team protocol to dis-
cover the set of forbidden unpredictable partial of-
fers FA for the team . The set of forbidden unpre-
dictable partial offers for the team, FA, is defined
as FA = {X(cid:48)∃a ∈ A, X(cid:48) ∈ Fa}. This means that
any unpredictable partial offer in FA is never part
of an acceptable offer for at least one team mem-
ber. Thus, these unpredictable partial offers should
be avoided for the team since the goal of the nego-
tiation model is reaching unanimously acceptable
agreements.
A formal description of the pre-negotiation pro-
tocol is presented in Figure 3. The picture describes
the protocol by means of Agent UML sequence di-
agrams. According to the proposed protocol, the
team mediator initiates the pre-negotiation phase
by asking each team member a to calculate its
own set of forbidden unpredictable partial offers Fa
(message 1 in Figure 3). Each team member builds
its own (forbidden) set as requested, and it is com-
municated to the mediator privately (message 2 in
Figure 3). When the mediator receives the sets from
the team members, it aggregates them in order to
construct the set of forbidden unpredictable partial
Fa. Then, the team
offers for team A, FA = (cid:83)
a∈A
mediator makes public the list of forbidden unpre-
dictable partial offers of the team FA (message 3
in Figure 3).
It should be stated that, since any
unpredictable partial offer in this set will prevent
one of the team members from reaching its reser-
vation utility, the team is not allowed to generate
an offer involving any of these partial offers in FA.
After the team mediator has shared FA with team
members, the negotiation phase starts.
The reader may realize that it is possible that
during this phase, most of the unpredictable par-
tial offers are pruned. In that case, it means that
there is little potential for cooperation among team
members. This issue can be observed by the team
mediator prior to starting any negotiation process.
In that case, the team mediator may suggest the
team not to negotiate and save the computational
resources used in the negotiation.
If the team is
not static and can be dynamically formed, it may
suggest team members to disband the team and
look for other potential partners. However, this
team/coalition formation(Gaston and desJardins,
2005; Rahwan et al., 2009) is outside of the scope
of this work since we focus on studying the perfor-
8
Figure 2: Overview of the communications carried out by the team mediator.
9
Team MediatorTeam Member aOpponent op1. INFORM Start negotiationref Pre-negotiation 2. INFORM Start negotiation3. ACK[deadline] 4.a INFORM withdrawref Offer proposal 5. PROPOSE[deadline]6.a INFORM withdraw[acceptable] 6.b ACCEPT6.c PROPOSE6.b.1 INFORM Accept[deadline]7.a INFORM withdrawref Opponent's Offer Evaluation [acceptable] 8.a ACCEPTtAXtopX Figure 3: Overview of the intra-team protocol carried out during the pre-negotiation
mance of the negotiation model. We consider the
use of the information provided with forbidden un-
predictable partial offers for negotiation team for-
mation as a future line of work. In general, combi-
nations of team members that prune a small portion
of the space should be more similar among them,
and it should be more easy to achieve cooperation.
3.2. Negotiation Phase
In the negotiation, two mechanisms are carried
out at each round: a mechanism for deciding to ac-
cept/reject the opponent's offer (Evaluation of Op-
ponent's Offer), and a mechanism for proposing an
offer to the opponent (Offer Proposal). For the for-
mer, a unanimity voting process is employed, while
for the latter an offer building process is governed
by the team mediator.
3.2.1. Evaluation of Opponent's Offer
This mechanism is carried out each time the team
mediator receives an offer from the opponent. Since
the main goal of the proposed intra-team strategy
is achieving unanimously acceptable agreements for
the team, a unanimity voting is carried out to de-
cide whether or not the opponent's offer is accept-
able for the team. With this mechanism, as long
as one of the team members is not satisfied with
the opponent's offer, the offer is not accepted by
the team, precluding the team from reaching agree-
ments that are not unanimously acceptable. A for-
malization of the protocol followed in this mech-
anism can be observed in Figure 4. The picture
shows the formalization employing sequence dia-
grams from Agent UML. The intra-team protocol
used for this mechanism goes as follows. First, the
team mediator receives the offer X t from the op-
ponent at time t. If X t involves any forbidden un-
predictable partial offer in FA, the opponent offer
is automatically rejected. However, the opponent's
offer is also informed to team members in order to
allow each team member to process the new infor-
mation leaked by the opponent if they see it nec-
essary (message 1 in Figure 4). Otherwise, if the
combination of unpredictable issue values is not in
FA, in order to see whether the offer is unanimously
acceptable for team members, the mediator makes
the opponent's offer public among team members
and starts an anonymous voting process (message
2 in Figure 4). Each team member a ∈ A states
to the mediator whether he is willing to accept X t
(positive vote) or to reject it (negative vote) at that
specific instant (messages 3.a or 3.b in Figure 4).
Since our aim is to guarantee unanimity, the offer
is only accepted if all of the team members emit a
positive vote (message 4.a in Figure 4). Otherwise,
the offer is rejected and a counter-offer is proposed
as explained in Section 3.2.2.
3.2.2. Offer Proposal
Proposing an offer to the opponent is a complex
task, since the space of offers may be huge and the
preferences of the team members should be reflected
in the offer sent to the opponent, and, in our case,
the offer sent should be unanimously acceptable for
team members. The process is divided into two
sub-phases: constructing an unpredictable partial
offer, and setting up predictable issues.
In both
phases, the team mediator acts according to Algo-
rithm 3.1. We include another formal description of
the interactions between the mediator and a team
member during the offer proposal. This informa-
tion can be found in Figure 5, which depicts the
intra-team protocol specified in and Agent UML
sequence diagram.
• Constructing an unpredictable partial
offer: The first step is to propose an un-
10
Team Member aTeam Mediator1. REQUEST2. INFORM3. INFORMAFaFSD Pre-negotiationaF Figure 4: Overview of the intra-team protocol employed to evaluate opponent's offers.
predictable partial offer, a partial offer which
has all of the unpredictable issues instantiated.
Since team members know from FA the list of
unpredictable partial offers that will not re-
sult in unanimously acceptable offers under
any circumstance, any offer proposed by the
team should avoid being constructed from un-
predictable partial offers found in FA. The
method used to propose offers to the oppo-
nent relies on the fact that unpredictable is-
sues are those where intra-team conflict may
be present, whereas there is full potential for
cooperation in predictable and compatible is-
sues. Hence, in order to build an offer to be
sent to the opponent,
it seems more appro-
priate to jointly set unpredictable issue val-
ues first and then, depending on the remaining
needs of team members, allow team members
to set compatible and predictable issues as they
require for reaching their demands. The pro-
posed mechanism for the first part, proposing
an unpredictable partial offer, is based on vot-
ing and social choice. The voting process goes
as follows.
1. The mediator asks each team member to
anonymously propose one unpredictable
(cid:48)t
partial offer X
a (message 1 in Figure 5).
2. Each team member privately sends its
proposal to the mediator, who gathers
all of the proposals in a list that will be
later sent to team members.
If any un-
11
predictable partial offer proposed by a is
contained in FA, the mediator automati-
cally ignores this proposal (message 2 in
Figure 5).
3. Once all of the proposals have been gath-
ered, the mediator makes public the list
(cid:48)t among team mem-
of proposal UPO
bers and opens a Borda scoring process
(Nurmi, 2010) on proposed candidates
(message 3 in Figure 5).
4. Each team member anonymously scores
candidates and sends the scores to the
team mediator (message 4 in Figure 5).
5. The team mediator sums up scores and se-
lects the winner candidate with the high-
(cid:48)t
A, making it public among
est score X
team members (message 5 in Figure 5).
This candidate, an unpredictable partial
offer, will be the base for the full offer that
will be sent to the opponent.
• Setting up predictable and compatible
issue values: Once unpredictable issues have
been set, it is necessary to set predictable and
compatible issues to construct a complete offer.
As it has been stated along this article, there
is full potential for cooperation among team
members in these issues since increasing the
utility of one of the team members by select-
ing one issue value will result in the other team
members staying at the same utility or increas-
Team Mediator2. REQUEST accept opponent offer ?topXTeam Member atopXSD Evaluation of Opponent's Offer3.a ACCEPT3.b REJECT4.a INFORM ACCEPT4.b INFORM REJECTtopX])([AtopFXUN1. INFORMtopX ing their utility. Obviously, the selected unpre-
dictable partial offer will not satisfy equally
the needs of all the team members. Never-
theless, team members can make use of pre-
dictable and compatible issues to satisfy their
remaining needs while not generating conflict
inside the team. To complete the partial offer
(cid:48)t
X
A, an iterative mechanism that we proposed
in (Sanchez-Anguix et al., 2012a) is used to
build the final offer issue per issue. The mecha-
nism follows an order for predictable and com-
patible issues that is constructed by the me-
diator at each round according to the history
of the opponent's concessions. The rationale
used to build this order is that the opponent
would concede less on those predictable issues
more important for him in the first negotia-
tion rounds, whereas it would concede more
on those predictable issues that are less impor-
tant. Thus, the order established by the team
mediator attempts to order predictable and
compatible issues in ascending order of impor-
tance for the opponent. The general idea be-
hind this ordering is attempting to satisfy team
members' demands with those predictable is-
sues less important for the opponent first. The
order is updated as new information becomes
available from the offers sent by the opponent.
Based on this order, the iterative mechanism
goes as follows.
6. The mediator selects the first predictable
issue j and asks team members, given the
(cid:48)t
A, the necessary
current partial offer X
value xj for j to get as close as possible
to their current demands (message 6 in
Figure 5).
7. Accordingly, each team member a informs
the mediator privately about the most
convenient value xa,j for that issue (mes-
sage 7 in Figure 5). To decide the final
value xj for the issue j, the trusted media-
tor aggregates agents' opinions (since the
issue is predictable) by means of a func-
tion that, for team members, returns the
most preferred issue value from a given
set (best(.)). After deciding the value xj,
X
(cid:48)t
A is updated with xj.
8. The mediator asks the team whether or
not the new partial offer is already satis-
factory at round t (message 8 in Figure
19:
20:
12
5).
9. Each team member emits an affirmative
response if the current partial offer cov-
ers its current demands and a negative
response if it still has not covered its de-
mands (message 9.a or 9.b in Figure 5).
Those agents that agree with the current
(cid:48)t
state of X
A leave the iterative mechanism
for this offer since they already are satis-
fied with the current partial offer. The
process steps back to the selection of the
next issue.
10. The process continues until all of the pre-
dictable issues have been set or until all
of the team members have left the it-
erative mechanism.
In the latter case,
the remaining issues are set attempting
to maximize the opponent's preferences.
Once the offer is complete, it is announced
among team members and sent to the op-
ponent (message 10.b in Figure 5).
Algorithm 3.1. Pseudo-algorithm for the offer
construction from the point of view of the medi-
ator. Send (message −→ condition ) means that
message is sent to every agent that fulfills condi-
tion
(cid:48)t = ((cid:83)
1:
2: /*Proposing an unpredictable offer*/
a −→ ∀a ∈ A)
(cid:48)t
3: Send (REQUEST X
a ←− ∀a ∈ A)
(cid:48)t
4: Receive (INFORM X
a ) − FA
(cid:48)t
5: UPO
X
(cid:80)
6: Send (REQUEST Borda on UPO
7: Receive (INFORM scorea ←− ∀a ∈ A)
8: X
a∈A
A)
(cid:48)t
A = argmax
(cid:48) t
X(cid:48)∈UPO
score(a, X(cid:48))
a∈A
A ←− ∀a ∈ A)
(cid:48)t
9: Send (INFORM X
10: order = build predictable order(); A(cid:48) = A
(cid:48)t −→ ∀a ∈
11:
12: /*Setting predictable issues*/
13: for all j ∈ order do
14:
15:
16:
17: X
18:
(cid:83){xj}
Send (REQUEST value for j −→ ∀aa ∈ A(cid:48))
Receive (INFORM xa,j ←− ∀aa ∈ A(cid:48))
xj = best({xa,ja ∈ A(cid:48)})
(cid:48)t
(cid:48)t
A = X
A
Send (REQUEST Satisfied with X
∀a ∈ A(cid:48))
for all a ∈ A(cid:48) do
A? −→
(cid:48)t
Receive (INFORM ac(cid:48)
A) ←− a)
(cid:48)t
a(X
21:
22:
23:
24:
25:
if ac(cid:48)
(cid:48)t
A) = true then
a(X
A(cid:48) = A(cid:48) − {a}
end if
end for
if A(cid:48) = ∅ then
break;
26:
end if
27:
28: end for
29: for all j ∈ order ∧ issue not set(j) do
xj = maximize f or opponent(j)
(cid:83){xj}
30:
(cid:48)t
(cid:48)t
A = X
31: X
A
32: end for
(cid:48)t
33: X t
A = X
A
4. Team Members' Strategies
The team mediator defines the coordination
mechanisms inside the team. However, each team
member's internal strategy has a great effect on
team dynamics.
In this article, we propose two
types of strategy for team members. According
to the first strategy (i.e., our basic team member),
the team member only proposes unpredictable par-
tial offers based on its own utility. In the second
strategy, team members model the preferences of
the team and the opponent on unpredictable par-
tial offers. Then, in the mechanism employed to
set unpredictable partial offers, each team member
selects the candidate that guarantees that it can
reach its current aspirations at time t, and max-
imizes the probability of being acceptable for the
opponent and the team. The learning mechanism
employed by these team members is Bayesian learn-
ing (i.e., Bayesian team member).
4.1. Basic Strategy for Team Members
Since negotiations are time-bounded in our
framework, we consider that team members have
to perform some kind of concession if an agreement
is to be found. For this purpose we have designed
basic team members as agents whose demands are
controlled by an individual and private concession
strategy. More specifically, the concession strategy
for a team member a ∈ A is based on time-based
tactics sa(t) (Faratin et al., 1998; Lai et al., 2008).
It estimates the utility demanded by a at time t
by using the formula in Equation 5, where RUa is
its reservation utility, T is the negotiation deadline,
and βa is the concession speed, which determines
how fast the agent's demands are lowered towards
RUa.
sa(t) = 1 − (1 − RUa) × (
1
βa
)
t
T
(5)
Based on this concession tactic, each team mem-
ber participates in the intra-team protocol with
their demands regulated by his private concession
tactic. Next, we define how team members take
their decisions: evaluating the opponent's offer, and
proposing an offer for the opponent.
4.1.1. Evaluation of Opponent's Offer
Given an offer X t proposed by the opponent at
instant t, the team member emits a positive vote in
the unanimity voting process if it reports a utility
greater than or equal to its current demands sa(t).
Otherwise, a negative vote is emitted.
(cid:26) true
f alse
aca(X t) =
if sa(t) ≤ Ua(X t)
otherwise
(6)
4.1.2. Offer Proposal
As documented in Section 3.2, team members
interact at three points during the offer proposal.
First, they propose an unpredictable partial offer
to the team mediator. Since each team member a
has its demands regulated by a time-based tactic,
when proposing an unpredictable partial offer to the
mediator at instant t, the proposed unpredictable
partial offer X
(cid:48)t
a fulfills:
a /∈ FA ∧ (Ua(X
(cid:48)t
X
a ) + maxP Ra ≥ sa(t))
(cid:48)t
(7)
Hence, agent a selects an unpredictable partial
offer which is not forbidden inside the team (since
it will be ignored by the team mediator) and whose
utility allows him to achieve or surpass its current
demands at time t. This way, the team member as-
sures that if its proposed unpredictable partial offer
is the winner of the Borda voting process, it can
reach its current demands. However, one should be
aware that many unpredictable partial offers may
fulfill Equation 7. Therefore, it is necessary to se-
lect one of them as the proposed candidate. Being
our basic team member, from the set of partial of-
fers that fulfill Equation 7, a team member selects
one of the candidates randomly.
The second time that a team member interacts
with the team mediator is for scoring unpredictable
partial offers that have been proposed by team
members. For scoring candidate partial offers in
13
Figure 5: Overview of the intra-team protocol carried out to propose offers to the other party.
14
Team Mediator1. REQUEST unpredictable partial offer2. INFORMTeam Member ataX'3. REQUEST Borda on4. INFORM5. INFORM winner unpredictable partial offerascoretAX'6. REQUEST value for issue j7. INFORMtjax,8. REQUEST Satisfied with ?taX'},...,{''1'tMttXXUPOSD Offer Proposal9.a INFORM OK9.b INFORM NO[All issues instantiated] 10. b INFORMtAX10.a REQUEST value for next issue j}{'''tjtAtAxXXSD Unpredictable IssuesSD Predictable and Compatible Issues the Borda voting process, a basic team member or-
ders the candidates according to the partial utility
reported by each of the candidates. That is, the
team member assigns the highest score to the par-
tial offer whose utility is the highest for itself, and
the second highest score to the partial offer whose
utility is the second best one, and so forth.
Finally, team members also interact with the me-
diator during the mechanism used to set predictable
and compatible issues. When team members are
asked about a value for issue j, each team mem-
ber communicates anonymously the value xa,j. The
value is the one that, given the current partial offer
(cid:48)t
A, gets the utility of the new partial offer as close
X
as possible to its current aspiration sa(t). Taking
normalized additive utility functions, it can be cal-
culated as:
xa,j =
argmax Va,j(x)
x∈Dj
if Ua(X
A) + wa,j ≤ sa(t)
(cid:48)t
argmin Va,j(x)
x ∈ Dj
∧
wa,j Va,j (x) ≥ (sa(t) − Ua(X
(cid:48) t
A ))
otherwise
(8)
where sa(t) is the utility demanded by the agent
(cid:48)t
A) is the utility reported by
a at round t, Ua(X
the current partial offer, and wa,jVa,j(x) is the
weighted utility reported by the value demanded
by the agent. The value asked for issue j is the
closest one to the current demands of the agent.
On the one hand, if the agent cannot reach its cur-
rent demands by just setting issue j, it asks for the
value that reports the highest utility. On the other
hand, if the agent can reach or surpass its current
demands by setting j, it asks for the value that
makes the new partial offer the closest to the cur-
rent demands. In the same iterative process, team
members still have to declare whether or not they
are satisfied with the different partial offers that are
constructed. Team members follow a similar crite-
rion to the method proposed to determine if an op-
ponent offer is acceptable at t. Basically, a partial
offer is acceptable for an agent a at t if it reports a
utility greater than or equal to the aspiration level
marked by its concession strategy:
(cid:26) true
f alse
ac(cid:48)
a(X
(cid:48)t
A) =
A) ≥ sa(t)
(cid:48)t
if Ua(X
otherwise
(9)
where true indicates that the partial offer is ac-
ceptable at its current state for agent a, and f alse
indicates the opposite.
15
4.2. Bayesian-based Strategy for Team Members
The Bayesian-based negotiation strategy for a
team member is based on modeling the team's (as
a whole) and its opponent's preferences on unpre-
dictable issues, and acting accordingly. For this
purpose, two Bayesian models are employed to pre-
dict if unpredictable partial offers are acceptable
for both teammates and the opponent. One of the
Bayesian models is employed to capture the prefer-
ences of the team on unpredictable issues, whereas
the other is used for capturing the preferences of
the opponent on unpredictable issues. The strategy
used to evaluate the opponent's offer is the same as
the one described in the basic strategy.
4.2.1. Bayesian Learning
Bayesian learning is a probabilistic learning
method based on Bayes' theorem (Russell and
Norvig, 2003). Given a certain set of hypothesis
H and some observation e, Bayesian learning at-
tempts to compute the probability p(he) that a
certain hypothesis h is true after observing e. In our
case, we want to determine whether or not the pro-
posed offer will be acceptable for the opponent (or
the team) (H={acc,¬acc}) given a certain unpre-
(cid:48)t) where acc stands for
dictable partial offer (e = X
"acceptable" and ¬acc stands for "unacceptable".
Since we assume that there is no interdependence
among negotiation issues, we can consider that each
negotiation issue contributes individually to the ac-
ceptability of an offer/unpredictable partial offer.
Thus, applying Bayes' theorem under independence
assumption we have:
p(acc) (cid:81)
(cid:80)
p(H) (cid:81)
j∈UN
p(xjacc)
H∈{acc,¬acc}
j∈UN
p(xjH)
(10)
p(accX
(cid:48)t) =
where p(acc) is the prior probability for an unpre-
dictable partial offer to be acceptable, p(¬acc) is
the prior probability for an unpredictable partial of-
fer to be non-acceptable, and p(xjacc) is the condi-
(cid:48)t as acceptable,
tional probability that assuming X
it has the value of the j issues instantiated to xj.
We consider positive examples Sacc as those ex-
amples that correspond to the acceptable hypothe-
sis (acc) and negative examples S¬acc as those ex-
amples that correspond to the not acceptable hy-
pothesis (¬acc). For the opponent's model, we em-
ploy unpredictable partial offers that have appeared
in opponent's offers as positive examples, and un-
predictable partial offers that appear in offers re-
jected by the opponent as negative samples. For
the team's model, we use FA and those opponent's
offers rejected by team members as the set of neg-
ative examples. Winners in the Borda votings (i.e.,
unpredictable partial offers contained in offers sent
to the opponent) are considered as positive exam-
ples. For computing p(xjh), we use the proportion
between the number of times that xj appears in
hypothesis h (acc or ¬acc) and the total number of
examples for h:
p(xjh) =
#{xj ∈ Sh}
Sh
(11)
The reasons for employing Bayesian learning are
varied. The most important one is that it allows
online updating of the model as new samples be-
come available. This is important in a process
like negotiation, where at each interaction new in-
formation becomes available regarding the oppo-
nent's/teammates' preferences.
If a computation-
ally expensive learning mechanism was used,
it
would not be possible to include the new informa-
tion in the model as it becomes available. Further-
more, the learning mechanism is computationally
cheap since it mainly involves counting. This is
also important in a real application since it allows
for simultaneous negotiation threads to exist, which
should be maintained with different opponents to
look for the best alternatives in an electronic mar-
ketplace.
4.2.2. Offer Proposal
Bayesian models are employed to help in the se-
lection of the unpredictable partial offer that is pro-
posed to the other team members. Bayesian team
members propose at t unpredictable partial offers
in the set defined in Equation 7. Bayesian models
help to select a candidate from that set.
However, it is reasonable to think that in the first
interactions Bayesian model do not accurately rep-
resent other agents' preferences. For that purpose,
a team member invests part of the negotiation time
texp in exploring the negotiation space and collect-
ing information regarding the opponent's and the
team's preferences. As long as the negotiation pro-
cess has not surpassed texp, the team member just
selects randomly one of candidate unpredictable
partial offers as basic team members do. Mean-
while, the Bayesian models are continuously up-
dated with the new information that becomes avail-
able during the negotiation. After reaching the time
threshold, the team member starts to use Bayesian
models in order to select the unpredictable partial
offer to be proposed to the mediator during the offer
proposal phase. The heuristic used in the selection
of the candidate is proposing an unpredictable par-
tial offer that is both acceptable for the team and
the opponent. The model has an additional pa-
rameter named pesc. It represents the probability
of avoiding the Bayesian proposal model and using
the random proposal model as described in the ba-
sic team member model when the negotiation time
has gone beyond texp. This parameter is included
in the model in order to: (i) further explore the
negotiation space; (ii) escape from local optima in-
duced by inaccurate Bayesian models (e.g., wrong
samples, limited number of samples, etc.). We can
formalize the selection as follows:
(cid:80) wbpb(accX)
b∈{A,op}
rand ≤ pesc
∧
t ≥ texp
if
argmax
X∈B
(cid:48)t
a =
X
random partial offer(B) otherwise
(12)
where B is the set of candidate unpredictable par-
tial offers that fulfill Equation 7, rand is a random
number, pA(accX) is the probability for a candi-
date unpredictable partial offer to be acceptable
for the team, popp(accX) is the probability for the
candidate unpredictable partial offer to be accept-
1 represent
able for the opponent, and wA and wop
the weights given to the acceptability of the un-
predictable partial offer for the team and the op-
ponent, respectively (i.e., we will refer to them as
Bayesian weights). Varying these Bayesian weights
allow team members to show different behaviors de-
pending on their inclination to satisfy either the
team or the opponent with the unpredictable par-
tial offer.
5. Provably Unanimously Acceptable Deci-
sions
As stated in the introduction, one of our re-
search goals is proposing a negotiation team model
that is able to guarantee unanimously acceptable
team decisions. Next, we show that under the as-
1wA + wop = 1.
16
sumption of rationality2, team members are able to
achieve unanimously acceptable final agreements, if
an agreement is found. For that matter, let us em-
ploy reductio ad absurdum (reduction to absurdity).
If X is the final agreement, let us suppose that
Equation 2 (unanimously acceptable) is violated in
a negotiation: unanimity is not reached because a
obtained a utility below its reservation utility.
∃a ∈ A, Ua(X
wa,jVa,j(xj) < RUa (13)
(cid:88)
(cid:48)t) +
j∈PR
The final agreement is found when (1) team mem-
bers accept an opponent's offer or (2) the opponent
accepts a team's offer. Next, we show that in both
cases, Equation 13 is never true.
1. When the team members accept an opponent's
offer, a unanimity voting process has been car-
ried to decide whether or not to accept the final
offer. The offer is only accepted if all of the
team members have emitted a positive vote.
Since a rational agent a would never have in-
centive to emit a positive vote of the offer re-
ported a utility below its reservation utility,
this scenario is never true due to the intra-team
mechanism.
2. When the opponent accepts a team's offer X,
this offer has been necessarily proposed by
the intra-team mechanism mentioned in Sec-
tion 3.2.2. The offer can be decomposed into
(cid:48)t and the in-
an unpredictable partial offer X
stantiation of predictable issues. The team
member a is not able to get over its reserva-
(cid:48)t ∈ Fa or when
tion utility if and only if X
(cid:48)t /∈ Fa and a could not get what it de-
X
manded in predictable issues. A rational agent
has no incentive to exclude a forbidden unpre-
(cid:48)t when declaring Fa.
dictable partial offer X
Fa and the mediator ignores
Since FA = (cid:83)
a∈A
unpredictable partial offers in FA, an unpre-
(cid:48)t that forms a team
dictable partial offer X
(cid:48)t /∈ Fa then the
offer is never in Fa.
If X
agent can accomplish to satisfy the following
(cid:48)t) + maxP Ra ≥ RUa. Agent
expression Ua(X
a could not get over its reservation value be-
cause he could not demand the most of pre-
dictable issues. However, when the team me-
diator aggregates predictable issues inside the
2Rational agents seek to improve their current welfare.
Thus, they would not take actions that lead to utilities below
their reservation utilities.
17
team, the team mediator selects the highest
value for team members in the list of values
proposed by them. This makes possible for a to
obtain the maximum utility from predictable
issues. Hence, Equation 13 never holds if the
negotiation ends with agreement.
Since both possible scenarios are never true un-
der our initial assumption, we have shown by re-
duction ad absurdum that, if a final agreement is
found, it is unanimously acceptable among team
members. Another issue is the presence of exagger-
ating agents (i.e., agents that exaggerate to get the
most from the negotiation). In our setting, even if
team members exaggerate and decide to include in
Fa unpredictable partial offers that are acceptable
but report low utility, or demand more than they
need from predictable issues, if a final agreement is
found it will be unanimous among team members.
However, by doing so, they may be pruning negoti-
ation space and lowering the probability of finding
agreements. This is an interesting situation that we
plan to study in the future.
6. Experiments
In this section, we explore the behavior of the
proposed negotiation model
in different scenar-
ios. The proposed framework has been imple-
mented in genius (Lin et al., 2012), a simulation
framework for automated negotiation that allows
researchers to test their frameworks and strate-
gies against state-of-the-art agents designed by
other researchers. Recently, genius has become a
widespread tool that increases its repository of ne-
gotiating agents with the annual negotiation com-
petition (Baarslag et al., 2012).
In order to assess the performance of the pro-
posed negotiation approach, we have performed dif-
ferent experiments. All of the experiments have
been carried out in the negotiation domain (or case
study) introduced in Section 2.4. The first exper-
iment (Section 6.1) studies the performance of the
proposed model when facing single opponent agent.
The comparison is carried out in scenarios with dif-
ferent degrees of team's preference dissimilarity. In
the second experiment, we study the performance
of our negotiation team model when facing another
negotiation team in bilateral negotiations. In the
third experiment (Section 6.3) we study how the
Bayesian weights wA and wop, which control the
importance given to the preferences of the team and
the opponent in the unpredictable partial offer pro-
posed to teammates, impact the performance of the
proposed model when team members employ the
Bayesian strategy. Finally, we conduct an experi-
ment to study the effect of team members' reser-
vation utility on the performance of the proposed
negotiation model (Section 6.4).
6.1. Performance Against a Single Opponent
In this first set of experiments we study the per-
formance of the proposed negotiation team model
when facing a single opponent. The study is carried
out with an emphasis on observing if the perfor-
mance of the team is higher by employing Bayesian
team members rather than basic team members.
Due to the fact that we are interested in open en-
vironments, we study how the team performance
varies with team configurations ranging from sce-
narios where no team members plays the Bayesian
strategy (i.e., all Basic team members) to situa-
tions where all the team members play the Bayesian
strategy. The performance of the team is measured
using the average team joint utility3. As an addi-
tional measure of optimality, we also measure the
distance to the closest Pareto optimal point. In this
case, the Pareto frontier is computed taking into ac-
count the team joint utility and the utility of the
opponent. Our initial hypotheses were:
• H1 As more Bayesian team members form the
team, the team is able to obtain average team
joint utilities that are higher than or equal to
those configurations with less Bayesian team
members.
• H2 As more Bayesian team members form the
team, the team is able to obtain average oppo-
nent utilities that are higher than or equal to
those configurations with less Bayesian team
members.
Since Bayesian team members take the prefer-
ences of the team and the opponent into account
when deciding which offer is sent to the other part,
they should be able guarantee equal or higher av-
erage team joint utility and opponent utility than
basic team members. But in no case, they should
3We consider the joint utility of the team to be the prod-
uct of the utilities of the team members. Since the utility of
an agent is between 0 and 1, the team joint utility tends to
be lower as more team members are present.
18
not obtain lower team joint utility and opponent
utility. As a consequence of both hypothesis, the
distance to the closest Pareto optimal point should
be also equal or lower than that obtained by config-
urations solely composed by basic team members.
In order to compare the proposed model with
other models in the literature, we also included
the Similarity Borda Voting model (i.e., SBV)
(Sanchez-Anguix et al., 2011) in our experiment.
SBV is a mediated intra-team strategy that is able
to guarantee semi-unanimity regarding team deci-
sions. The mediator imposes a unanimity voting
process to decide on whether or not to accept the
opponent's offer, whilst team members propose of-
fers to be sent to the opponent by means of a simi-
larity heuristic that takes into account the last offer
proposed by the opponent, and the last offer pro-
posed by the team. A Borda voting process is used
in order to decide on which offer is sent to the oppo-
nent. The reason to include this intra-team strat-
egy in our study is due to the fact that it has been
documented to achieve similar results to intra-team
strategies that guarantee unanimity under certain
circumstances for domains solely composed by pre-
dictable issues (Sanchez-Anguix et al., 2011).
In
order to adapt this approach for domains with un-
predictable issues, we use a similarity heuristic that
uses Euclidean distance for real/integer issues and
string matching for other types of issues. Due to
the fact that our proposed model guarantees unan-
imously acceptable agreements and SBV does not,
we formulated the following hypothesis:
• H3 Teams exclusively formed by basic team
members and teams exclusively formed by
Bayesian team members obtain equal or higher
average team joint utility than teams following
the Similarity Borda Voting model.
The performance of this first experiment is an-
alyzed in three scenarios with different degrees of
preference dissimilarity among team members: very
similar, average similarity, and very dissimilar pref-
erences. For this reason, we introduce a measure
for measuring team members' preference similar-
ity in different scenarios. The authors proposed a
method for calculating preference dissimilarity in
teams based on the utility difference of offers among
teammates (Sanchez-Anguix et al., 2011). The dis-
similarity between two teammates a, b ∈ A can be
measured as:
D(Ua(.), Ub(.)) =
(cid:80)
∀X
Ua(X) − Ub(X)
# possible offers
(14)
Due to the fact that a team may be composed
of more than two members, it is necessary to pro-
vide a team dissimilarity measure. The team dis-
similarity measure is calculated as the average of
the dissimilarity between all of the possible pairs of
teammates. For this experiment, we decided to ex-
plore teams whose preferences are dissimilar, teams
whose preferences are similar, and teams with an
average degree of similarity/dissimilarity (i.e., aver-
age similarity). For the scenario of dissimilar prefer-
ences, 9 negotiation cases were randomly generated
(i.e., a combination of 3 different negotiation teams
consisting of four team members with 3 different op-
ponents), while 9 negotiation cases were randomly
generated for the similar preferences scenario (i.e.,
a combination of 3 different negotiation teams con-
sisting of four team members with 3 different op-
ponents) and 12 negotiation cases were randomly
generated for the average similarity scenario (i.e.,
a combination of 4 different negotiation teams con-
sisting of four team members with 3 different op-
ponents). Since we consider that in practice it is
less likely to meet extreme cases such as dissimilar
or similar teams, we decided to increase the num-
ber of negotiation cases in the average similarity
scenario.
As for the single opponents, we decided to test
the negotiation team models against different fami-
lies of opponents. More specifically, we followed the
categorization of negotiation strategies proposed by
Baarslag et al. (Baarslag et al., 2011), which di-
vides negotiation strategies into four categories:
• Competitors: They hardly concede, indepen-
dently of opponent behavior. Agent K is
a competitor agent (Kawaguchi et al., 2011;
Baarslag et al., 2011) from the 2010 ANAC
competition(Baarslag et al., 2012) that adjusts
its aspirations (i.e., target utility) in the nego-
tiation process considering to an estimation of
the maximum utility that will be offered by the
other party. More specifically, the agent grad-
ually reduces its target utility based on the av-
erage utility offered by the opponent and its
standard deviation.
• Matchers: They concede when they perceive
that the opponent concedes, and they do not
concede if they perceive that the other party
does concede. Nice Tit-for-Tat is a matcher
agent (Baarslag et al., 2011, 2013) from the
2011 ANAC competition that reciprocates the
other party's moves by means of a Bayesian
model of the other party's preferences. Accord-
ing to the Bayesian model, the Nice Tit-for-Tat
agent attempts to calculate the Nash point and
it reciprocates moves by calculating the dis-
tance of the last opponent offer to the afore-
mentioned point. When the negotiation time
is reaching its deadline, the Nice TFT agent
will wait for an offer that is not expected to
improve in the remaining time and accept it in
order to secure an agreement.
• Conceders: They yield independently of the
opponent behavior. Conceder is an imple-
mentation of the time-based concession tactics
proposed by Faratin et al. (Faratin et al., 1998)
categorized by Baarslag et al. as conceder. For
the Conceder agent βop = 2, which leads to
large concessions towards the reservation util-
ity in the first rounds.
• Inverter: They respond by implementing the
opposite behavior shown by the other party.
Boulware is an implementation of the time-
based concession tactics proposed by Faratin
(Faratin et al., 1998) categorized by
et al.
Baarslag et al.
In the case of
the Boulware agent, the concession speed is set
to βop = 0.2. Hence, the agent concedes very
insignificantly during most of the negotiation
and it concedes very quickly as the deadline
approaches.
as inverter.
The reservation utility of each team member was
set to RUa = 0.5 to represent scenarios where team
members have average aspirations. Additionally,
for each team member (i.e., basic, Bayesian and
SBV) the concession speed was randomly selected
from a uniform distribution βa = U [0.5, 1].
In
the case of Bayesian members, the time of explo-
ration was set to texp = 70% and the probability
of escape after the exploration phase was set to
pesc = 30% 4. Therefore, Bayesian models are not
used unless a 70% of the negotiation time (126 sec-
onds) has passed. Initially, we set Bayesian mem-
4These values were found to be the best ones after carry-
ing out a grid search over values of texp and pesc in a subset
of test negotiation scenarios
19
Agent K
Op.
0.743
0.683
0.690
Agent K
Op.
0.629
0.574
0.583
Agent K
Op.
0.522
0.397
0.457
T.
0.181
0.259
0.263
T. J
0.168
0.211
0.248
T. J
0.07
0.174
0.209
SBV
Basic
Bayesian
SBV
Basic
Bayesian
SBV
Basic
Bayesian
Similar
Nice Tit-for-Tat
Boulware
Conceder
D.
0.070
0.065
0.058
T.
0.150
0.173
0.164
D.
Op.
0.694
0.760
0.746
Average Similarity
0.130
0.067
0.080
T.
0.184
0.223
0.224
Op.
0.755
0.696
0.695
D.
0.064
0.078
0.080
T.
0.552
0.561
0.557
Op.
0.482
0.468
0.472
Nice Tit-for-Tat
Boulware
Conceder
D.
0.065
0.070
0.034
T. J
0.137
0.141
0.158
Op.
0.562
0.691
0.669
D.
0.116
0.050
0.047
Dissimilar
T. J
0.170
0.210
0.224
Op.
0.598
0.585
0.574
D.
0.070
0.060
0.045
T. J
0.324
0.386
0.390
Op.
0.428
0.414
0.414
D
0.037
0.045
0.043
D.
0.074
0.052
0.050
Nice Tit-for-Tat
Boulware
Conceder
D.
0.168
0.180
0.118
T. J
0.160
0.184
0.196
Op.
0.457
0.572
0.559
D.
0.157
0.055
0.60
T. J
0.128
0.254
0.271
Op.
0.547
0.505
0.489
D.
0.110
0.053
0.058
T. J
0.257
0.472
0.475
Op.
0.430
0.367
0.367
D.
0.110
0.046
0.044
Table 1: Average team joint utility (T. ), the average opponent utility (Op.), and the average Euclidean distance to the closest
Pareto optimal point (D.) for the first set of experiments.
bers to care equally about the acceptability of un-
predictable partial for the team and the opponent
wA = wop = 0.5.
Following the type of setting used in the annual
agent competition, the negotiation time was set to
T = 180 seconds. Each opponent strategy was
faced against each negotiation team model in ev-
ery possible negotiation case. A total of 20 repe-
titions were done per negotiation case in order to
capture stochastic variations in negotiation strate-
gies. Therefore, 3 × 3 × 3 × 4 × 20 = 2160 (team
preference profiles × opponent preference profiles ×
team negotiation models × opponent strategies ×
repetitions) negotiations were simulated in the sim-
ilar scenario, 2160 negotiations were simulated in
the dissimilar scenario, and 2880 negotiations were
simulated in the average similarity scenario.
Table 1 shows the average team joint utility
and opponent utility for the cases where all of the
team members either play the Basic strategy, the
Bayesian model, or the team employs the SBV team
negotiation model. It also shows the Euclidean dis-
tance to the closest point in the Pareto frontier. An
ANOVA test (α = 0.05) with a Bonferroni post-
hoc analysis was carried out to assess statistical
differences among the different measures gathered.
Those measures that are statistically the best con-
figurations for each column are highlighted in bold
style. All of the claims in this experimental sec-
tion are supported by the ANOVA test with the
Bonferroni post-hoc analysis. The average negotia-
tion time taken by each method is included in Ta-
ble 2. Figure 7 shows the evolution of the average
team joint utility and the average opponent utility
as more team members play the Bayesian strategy.
We have also included some examples of agreements
obtained in the different negotiation scenarios and
how they relate to the Pareto frontier5. These re-
sults can be observed in Figure 6. Next, we analyze
the results.
6.1.1. Results for the first hypothesis
• H1 As more Bayesian team members form the
team, the team is able to obtain average team
joint utilities that are higher than or equal to
those configurations with less Bayesian team
members.
First, we focus on the situations when all of the
team members either play the basic strategy or the
Bayesian strategy (Tables 1 and 2). It can be ob-
served that when team members' preferences are
similar, both types of teams perform equally in
terms of the average team joint utility. This result
is consistent with H1, since both prove to be statis-
tically equivalent with the ANOVA test with Bon-
ferroni post-hoc analysis. The reason why Bayesian
5The quadratic root of the team joint utility is taken to
convert the results to the same scale (remember that the
team joint utility is the product of for team members' utili-
ties)
20
Figure 6: Examples of agreements and their distance to the Pareto frontier
21
0,30,40,50,60,70,80,90,10,20,30,40,50,60,70,80,91Average similarity scenario vs Nice TFT(Team Joint Utility)^(1/4)Opponent utility00,10,20,30,40,50,60,70,80,9100,10,20,30,40,50,60,70,80,91Average similarity scenario vs Agent K(Team Joint Utility)^(1/4)Opponent utility0,30,40,50,60,70,80,911,10,250,350,450,550,650,750,850,95Similar scenario vs Nice TFT(Team Joint Utility)^(1/4)Opponent utilityParetoBasicBayesianSBVParetoBasicBayesianSBVParetoBasicBayesianSBVParetoBasicBayesianSBV0,30,40,50,60,70,80,911,10,250,350,450,550,650,750,850,95Similar scenario vs Agent K(Team Joint Utility)^(1/4)Opponent utilityParetoBasicBayesianSBV00,20,40,60,811,20,10,20,30,40,50,60,70,80,91Dissimilar scenario vs Nice TFT(Team Joint Utility)^(1/4)Opponent utility00,10,20,30,40,50,60,70,80,9100,10,20,30,40,50,60,70,80,91Dissimilar scenario vs Agent KParetoBasicBayesianSBV(Team Joint Utility)^(1/4)Opponent utilityParetoBasicBayesianSBV models do not give an advantage over the basic
model in the similar scenario can be explained since
team members are similar and there is no neces-
sity to carry out team modeling. The distance to
the closest Pareto optimal point is also very simi-
lar for both team configurations, which can be also
explained due to the fact that team modeling is
not necessary due to team members' similar prefer-
ences.
As conflict is introduced inside the team by mak-
ing team members' preferences more dissimilar (i.e.,
dissimilar and average similarity scenarios, middle
and bottom part of Table 1), it can be observed that
usually the team formed exclusively by Bayesian
team members gets the statistically highest average
team joint utility, which is also coherent with our
hypothesis H1 and refines our hypothesis for these
scenarios. In this case, the teammates' preferences
are no longer similar and some sort of modeling
mechanism is needed in order to guide the intra-
team negotiation towards agreements that are good
for all of the team members. The only exception is
found in the conceder case, where the performance
in terms of the team joint utility was found to be
statistically equivalent among the team exclusively
formed by basic team members and the team ex-
clusively formed by Bayesian members. Taking a
closer look at the negotiation traces, we observed
that, in all of the negotiations, the exploration time
texp was never surpassed. Since the Conceder agent
concedes rapidly in the negotiation, the team's de-
mands are also met early. Therefore, Bayesian mod-
els do not get to be used. In fact, the average nego-
tiation time against Conceder agents was 61.7, 76.3,
and 88 seconds respectively in the similar scenario
(see Table 2), the average similarity scenario (see
Table 2), and the dissimilar scenario (see Table 2).
All of them are below the threshold of 126 seconds
delimited by texp. As a result, the team members
have not used their Bayesian model while generat-
ing their proposals and they are equivalent to the
team formed by basic members. This is also con-
sistent with H1, since in no case the team formed
by Bayesian members gets statistically lower results
than the team formed by basic members.
If we observe the evolution of the average team
joint utility in Figure 7 6, there is a tendency
to increase the average team joint utility as more
Bayesian members are included (triangle shaped
6Results for the Conceder agent are omitted since the
Bayesian models are not employed.
Similar
139.4
145.4
144.0
N. TFT B.
164.3
162.5
165.2
K
148.7
141.5
142.3
Average Similarity
K
155.1
153.1
150.1
N. TFT B.
177.6
174.8
175.0
153.1
154.9
154.7
Dissimilar
K
163.8
162.2
163.7
N. TFT B.
175.7
176.6
177.2
156.6
160.1
160.5
C.
61.0
62.8
61.7
C.
74.1
77.4
76.3
C.
73.4
87.7
88.0
SBV
Basic
Bayesian
SBV
Basic
Bayesian
SBV
Basic
Bayesian
Table 2: Average time (seconds) for negotiations in the first
set of experiments. K (Agent K), N. TFT (Nice Tit-for-Tat),
B. (Boulware), C. (Conceder)
data series on Figure 7) in situations where team
members' preferences have an average similarity or
they are very dissimilar. This tendency is more
pronounced against Agent K (left plot in Figure
7) and Boulware agents (right plot in Figure 7).
However, when team members' preferences are very
similar the team performance remains at statisti-
cally equivalent values (ANOVA test with Bonfer-
roni analysis) for the average team joint utility. The
results of these graphics are coherent with our find-
ings in Table 1 and H1.
In conclusion, we have found that as more
Bayesian team members form the team, the team is
able to obtain average team joint utilities that are
higher than or equal to those configurations with
less Bayesian team members. Being more specific,
we have been able to detect that, as long as there
is preferential conflict among team members (i.e.,
average similarity among team members, and very
dissimilar team members), and the opponent does
not concede early in the negotiation, more Bayesian
team members result in higher team joint utility.
6.1.2. Results for the second hypothesis
• H2 As more Bayesian team members form the
team, the team is able to obtain average oppo-
nent utilities that are higher than or equal to
those configurations with less Bayesian team
members.
For the average opponent utility, the Bayesian
team obtained significantly better results than the
basic team only in the scenario where the team faces
22
Figure 7: Evolution of the average team joint utility and opponent utility as more Bayesian team members are introduced in
the team.
Agent K and team members are dissimilar with
regards to their preferences (see Table 1, bottom
part). The same pattern is found in Figure 7, where
we can observe that the average opponent utility in-
creases as more Bayesian team members are present
in the team (first plot in the second row of Figure
7). In other cases, the Bayesian and the basic team
obtain statistically equivalent results to each other,
and, in some situations, the basic team obtained
significantly better results. More specifically, when
the team faces Nice Tit-for-Tat, the basic team ob-
tains significantly better results than the Bayesian
team. We can also observe this pattern in Figure
7. As more Bayesian members are introduced, the
average opponent utility slightly decreases (middle
plot in the second row of Figure 7).
These findings only support partially our hypoth-
esis H2 since we found a set of scenarios where the
basic team provides better utility to the opponent
(i.e., when facing Nice Tit-for-Tat). We analyzed
the trace of different negotiations against Nice Tit-
for-Tat and Boulware opponents.
In the former
case, we could observe that close to the end of
the negotiation the Nice Tit-for-Tat opponent had
only sent on average 5 different unpredictable par-
tial offers in a domain that has 4200 different un-
predictable partial offers. This behavior results in
scarce information for any learning mechanism. In
the case of negotiations against Boulware agents,
one should consider that the Boulware strategies
concede only towards the end of the negotiation
and, most of the time, the aspirations are high.
Thus, most of the samples gathered by the Bayesian
classifier when facing Boulware agents correspond
to offers with high demands where usually only the
best issue values appear. Other issue values do not
appear in the samples or they have their frequency
misinterpreted with respect to the utility that they
actually report. Therefore, the learning mechanism
misinterprets the preferences of the opponent and
23
None123All0.10.150.20.250.30.35vs. Agent KNumber of Bayesian team membersAverage Team Joint UtilityNone123All0.10.150.20.250.30.35vs. Nice TFTNumber of Bayesian team membersNone123All0.10.150.20.250.30.35vs. BoulwareNumber of Bayesian team membersVery SimilarAverage SimilarityVery DissimilarNone123All0.350.40.450.50.550.60.650.70.75vs. Agent KNumber of Bayesian team membersAverage Opponent UtilityNone123All0.350.40.450.50.550.60.650.70.75vs. Nice TFTNumber of Bayesian team membersNone123All0.350.40.450.50.550.60.650.70.75vs. BoulwareNumber of Bayesian team members the team formed by Bayesian members is not able
to obtain statistically better utility for the oppo-
nent than the team formed by basic members.
The behavior of Nice Tit-for-Tat and Boulware
opponents also has a direct consequence on the dis-
tance to the closest Pareto optimal point obtained
by both team configurations (Bayesian and basic).
Despite the fact that the Bayesian configuration is
capable of obtaining statistically better results for
the team joint utility by learning the preferences
of the team, the utility reported to the opponent
is usually lower than the one reported by the ba-
sic configuration. The only exception to this case
are scenarios against Agent K, where the Bayesian
configuration obtains a statistically higher utility
for the opponent. The inability to model the oppo-
nents' preferences in the case of the Boulware and
Nice Tit-for-Tat opponents, results in a higher team
joint utility (due to team modeling) at the cost
of reducing the utility received by the opponent.
Hence, there is not an improvement in the distance
to the closest Pareto optimal point. These find-
ings can also be observed in some of the examples
included in Figure 6.The agreements found by the
basic team configuration and the Bayesian configu-
ration tend to be found at the same distance to the
Pareto frontier (the Bayesian configuration tending
to populate regions with higher team joint utility).
The exception to this rule are negotiations against
Agent K, where the basic configuration tends to
populate regions of no agreement (close to the axis
origin).
6.1.3. Results for the third hypothesis
• H3 Teams exclusively formed by basic team
members and teams exclusively formed by
Bayesian team members obtain equal or higher
average team joint utility than teams following
the Similarity Borda Voting model.
It can be observed that when team members' pref-
erences are similar (top part of Table 1), both basic
and Bayesian models are statistically equivalent to
each other and better than SBV with respect to the
average team joint utility. As conflict is introduced
inside the team by making team members' prefer-
ences more dissimilar (i.e., dissimilar and average
similarity scenarios, middle and bottom part of Ta-
ble 1), the team get statistically lower average team
joint utility by employing the SBV model. Basic
and Bayesian models outperform SBV with respect
to the average team joint utility since they are able
24
to guarantee unanimously acceptable agreements,
while SBV does not guarantee such condition. This
result supports and refines our hypothesis H3, since,
in general, the team formed exclusively by Bayesian
team members and the team formed by basic team
members obtains statistically higher team joint util-
ities than the SBV model. There is only one ex-
ception to this refinement. The basic team, the
Bayesian team and SBV perform statistically equal
in terms of the average team joint utility only when
the opponent is a conceder and team members' pref-
erences are similar. Since the opponent concedes
rapidly in the first rounds, the three models ob-
tain equivalent team joint utility due to the oppo-
nent concessions and the fact that conflict is almost
nonexistent among team members. However, this
finding is still consistent with our initial guess H3.
Regarding the optimality of the solutions found
by the model proposed in this article with respect
to SBV, it is possible to observe that both Bayesian
and basic configurations obtains statistically lower
distance to the Pareto frontier than SBV as long
as preferential conflict between team members is
present (i.e., average similarity and dissimilar sce-
narios). Only when the team is very similar, SBV
gets statistically equal distances to the model pro-
posed in this article. However, despite the fact that
the distance to the Pareto frontier may be statisti-
cally equal, we can observe in Figure 6 that agree-
ments found by SBV tend to populate areas that
are closer to the lowest team joint utility. This re-
sult can be explained due to the fact that unanimity
cannot be guaranteed in the team, and some team
members end up with low utility agreements.
In conclusion, we found that teams exclusively
formed by basic team members and teams exclu-
sively formed by Bayesian team members obtain
equal or higher average team joint utility than
teams following the Similarity Borda Voting model.
More specifically, we found that, in general, the re-
sults for the team formed by Bayesian team mem-
bers and the team formed by basic team members
obtains statistically higher results than the SBV
model. The distance to the Pareto frontier also
shows higher quality and more optimal agreements
for teams employing the model proposed in this
article. This result is important, since it shows
that the present model, not only guarantees unan-
imously acceptable agreements, but it also ensures
that better results are obtained with respect to
other state-of-the-art team negotiation models.
6.2. Performance Against Another Team
In this experiment we analyze the performance
of the proposed model when two teams face each
other. More specifically, we simulate negotiations
where one negotiation team represents a group of
four travelers, and the other negotiation team rep-
resents the board of managers for a hotel, which
consists also of four managers. The negotiation has
a common deadline of T = 180 seconds. Both par-
ties negotiate with each other by means of the al-
ternating bilateral protocol, but they employ the
negotiation team model proposed in this article to
coordinate and take team decisions. The goal of the
experiment is determining if Bayesian team mem-
ber improve the performance of the team when it
faces another team.
In this setting, we use 4 different team prefer-
ence profiles to represent the group of travelers and
2 different team preference profiles to represent the
board of managers. Since we are interested in open
environments, we consider different team configura-
tions from the perspective of strategy profiles: No
team member plays the Bayesian strategy (0-0),half
of the team members play the Bayesian strategy
in one team (2-0), all of the team members play
the Bayesian strategy in one team (4-0), half of the
team members are Bayesian in both teams (2-2), all
of the team members are Bayesian in one team and
half of them are Bayesian in the other team (4-2),
and all of the team members are Bayesian (4-4).
For each strategy profile, each group of travelers' is
faced against each board of managers 20 times to
capture stochastic variations. Therefore, a total of
4×2×20 = 160 negotiations is carried out per team
strategy profile, giving a total of 960 negotiations
for this experiment.
Our initial guess is that more Bayesian team
members will help to obtain higher team joint util-
ities due to the learning and proposal mechanism
used to take into account the preferences of the
team and the preferences of the opponent. More
specifically, we formulated the following hypothe-
ses:
• H4 As long as only one team includes Bayesian
team members (configurations 2-0 and 4-0),
the average team joint utility for both teams
will be higher than the average team joint util-
ity obtained by negotiations where no Bayesian
team member participates (configuration 0-0).
• H5 Those configurations where both teams in-
clude Bayesian team members (configurations
25
2-2, 4-2, and 4-4) will obtain higher aver-
age team joint utilities for both teams than
configurations where only one team includes
Bayesian team members (configurations 2-0
and 4-0).
We configure the parameters as we did in our
previous experiment. The results of the experiment
can be observed in Figure 8. The blue points repre-
sent the average team joint utility for the group of
travelers, whereas the red points represent the av-
erage team joint utility for the board of managers.
The graphic shows an increasing average team joint
utility as the total of Bayesian team members in
both sides increases. The worst results for both
teams are obtained when all of the team members
act as the basic team member (0-0). This result
is explainable due to the large number of negoti-
ations that finished with no agreement (80 out of
160 negotiations, a 50% of failure). As long as one
of the sides implements the Bayesian strategy (2-0,
4-0), both teams benefit by obtaining higher av-
erage team joint utilities. An ANOVA test with
Bonferroni post-hoc analysis (α = 0.05) confirmed
that both configurations 2-0 and 4-0 obtain statis-
tically different and higher team joint utilities for
both teams than the configuration 0-0. This result
confirmed our initial hypothesis H4.
It can be observed that the next relevant increase
in the average team joint utility of both teams is
present as long as both sides apply the Bayesian
strategy (2-2, 4-2, 4-4). An ANOVA test with Bon-
ferroni post-hoc analysis (α = 0.05) revealed that
the averages obtained by configurations 2-2, 4-2,
and 4-4 are statistically different and higher than
the averages obtained by 2-0 and 4-0. Hence, both
teams obtain higher average team joint utilities as
long as both teams include Bayesian team members,
supporting our hypothesis H5.
In conclusion, we have been able to determine
that when two teams face each other by means of
the proposed model, both teams benefit by obtain-
ing higher team joint utilities when Bayesian team
members participate in the negotiation. This is
especially true when Bayesian team members are
distributed between both teams, which obtains the
highest team joint utilities for both sides.
6.3. Analyzing the impact of Bayesian weights for
the proposal of unpredictable partial offers
Recalling from Section 4.2, there are two weight
parameters that control how important the op-
Ia ∈
[0.33, 0.66] it gives average importance to
unpredictable issues, and when Ia ∈ [0.66, 1.0] the
agent gives high importance to unpredictable issues.
We generated 8 random negotiation cases where
team members give a high importance to unpre-
dictable issues and the opponent gives low (4 cases)
and average (4 cases) importance to unpredictable
issues, 8 different randomly generated negotiation
cases where team members give a low importance
to unpredictable issues and the opponent gives av-
erage (4 cases) and high (4 cases) importance to un-
predictable issues, and 12 negotiation cases where
the opponent and the team give the same impor-
tance to unpredictable issues (4 cases where both
give low importance, 4 cases where both give aver-
age importance, and 4 cases where both give high
importance).
We tested three configurations for Bayesian
teams: standard Bayesian members that give the
same importance to the acceptability of the un-
predictable partial offer by the opponent and the
team wA = wop = 0.5 (Normal), Bayesian mem-
bers that give more importance to the acceptabil-
ity of the unpredictable partial offer by the oppo-
nent wA = 0.25 wop = 0.75 (Opponent Oriented),
and Bayesian members that give more importance
to the acceptability of the unpredictable partial of-
fer by the team wA = 0.75 wop = 0.25 (Team
Oriented). As for the opponent's strategies, we se-
lected Agent K, Nice TFT and Boulware.
In situations where the opponent gives more im-
portance to unpredictable issues than the team, the
team should be able to obtain higher average team
joint utility by playing high values for wop, satisfy-
ing opponent needs with unpredictable issues, and
demanding more on predictable issues. If the team
gives more importance to unpredictable issues, the
team should be able to obtain higher average team
joint utility by playing high values for wA, satis-
fying opponent needs with predictable issues, and
demanding more on unpredictable issues. If both
teams give the same importance to unpredictable
issues, the team should be able to guarantee higher
average team joint utility by giving the same weight
to wop and wA since no party gives more importance
to unpredictable issues. Attending to these initial
guesses, we formulated the following hypotheses:
• H6 When the opponent gives more impor-
tance to unpredictable issues, the team obtains
higher team joint utility by using high values
for wop and demanding more on predictable is-
Figure 8: Evolution of the average team joint utility with the
different strategy profiles (Blue= group of travelers, Red=
board of managers).
ponent's and team's preferences are while gener-
ating the unpredictable partial offer (respectively
wop, wA). wA represents how important it is for
the team members to make an unpredictable par-
tial offer that is acceptable for the team, whereas
wop represents how important it is for us to make
an unpredictable partial offer that is acceptable for
the opponent. The use of these weights is not triv-
ial, since one should consider that, it only refers
to the acceptability of the unpredictable partial of-
fer by one of the two parties. A complete offer is
composed by the predictable and unpredictable is-
sues. Therefore, for instance, using a high value of
wop may not have the desired effect on the oppo-
nent unless unpredictable issues are important for
the opponent. Additionally, one should also con-
sider that the more acceptable an unpredictable is
for the team/opponent, the more utility it should
report.
In this experiment we explore the impact of these
weights in a wide variety of situations. More specif-
ically, we study how different values for wop and wA
affect situations where the team gives more impor-
tance to unpredictable issues than the opponent,
situations where the opponent gives more impor-
tance to unpredictable issues than the team, and
situations where both team and the opponent give
the same importance to unpredictable issues.
To assess the importance given by an agent to
unpredictable partial issues, we consider the sum of
unpredictable issue weights in its utility function.
(cid:88)
Ia =
wi,j
(15)
j∈UN
We consider that when Ia ∈ [0.0, 0.33] the agent a
gives low importance to unpredictable issues, when
26
0 -- 02 -- 04 -- 02 -- 24 -- 24 -- 40.050.070.090.110.130.150.170.19Number of Bayesian team members in both teamsAverage team joint utility sues (wA = 0.25 wop = 0.75).
• H7 When the team gives more importance to
unpredictable issues, the team obtains higher
team joint utility by using high values for
wA and letting the opponent demand on pre-
dictable issues (wA = 0.75 wop = 0.25).
• H8 When both parties give the same impor-
tance to unpredictable issues, the team obtains
higher team joint utility by giving the same
importance to the opponent's and team's pref-
erences on unpredictable issues (wA = wop =
0.5).
For each negotiation case, we repeated the nego-
tiation 20 times in order to capture stochastic vari-
ations in strategies. Therefore, a total of 1680 ne-
gotiations were carried out in this experiment. The
results of this experiment can be observed in Table
3. It shows the average of the joint team utility and
proportion of negotiations that finished with suc-
cess (agreement) in the scenario. An ANOVA test
(α = 0.05) with Bonferroni post-hoc analysis was
carried out to detect statistically different averages.
The best configurations for each of the three sce-
narios are highlighted in bold font style. All of the
claims of this experimental section are supported
with the aforementioned ANOVA test with post-
hoc analysis.
Focusing on the case of negotiations against
Agent K, it can be appreciated that when unpre-
dictable issues are more important for the oppo-
nent (bottom part of Table 3), the best results
are obtained by taking an opponent oriented ap-
proach: proposing unpredictable partial offers that
are likely to be acceptable for the opponent and sat-
isfy remaining members' aspirations by demanding
on predictable and compatible issues, which are less
important for the opponent. This finding supports
H6. As for the scenario where unpredictable is-
sues are more important for the team (middle part
of Table 3),
it is clearly observed that the best
choice for team joint utility is to give a high weight
to wA, thus employing a team oriented approach.
Since unpredictable issues are more important for
the team, they should satisfy their needs as much
as possible with proposed unpredictable partial of-
fers and demand less on predictable issues, which
are more important for the opponent, supporting
our initial hypothesis H7. Finally, the last sce-
nario corresponds to the case where unpredictable
issues have the same importance for both parties
27
(top part of Table 3). In this case, there may be
more conflict between the team and its opponent
since the parties do not have a clear trade-off op-
portunity such as increasing the demand on unpre-
dictable issue while decreasing the demand on pre-
dictable issues as appeared in two previous cases.
One can observe that the best team joint utility is
obtained when using the standard team members
(wA = wop = 0.5). Since both parties give the same
importance to unpredictable issues, it seems natural
to give the same importance to the acceptability of
the unpredictable partial offer by the team and the
opponent, supporting our hypothesis H8. The team
oriented approach is clearly worse than the rest of
approaches since many negotiations (only a 58.7%
were successful. See top part of Table 3 ) ended in
failure due to the team being too demanding and
not satisfying the opponent's preferences.
When negotiating against Nice TFT and Boul-
ware, the results are different. It can be appreciated
that, generally, the team oriented approach always
reports statistically better results from the point of
view of the average team joint utility. This means
that the team should indistinctly select highly ac-
ceptable unpredictable partial offers for the team
when facing opponents like Nice TFT and Boul-
ware agents. These findings do not support our
hypotheses H6, H7, and H8, and drove a more in-
depth analysis and study. There are two factors
that should be taken into account. First, as we de-
tected in Section 6.1, the Bayesian models that are
learnt from both agents are not adequate due to
the lack of learning samples detected in the case of
Nice TFT and the misinterpretation of the impor-
tance of issue values in the case of the Boulware
agent. This factor precludes the agents from find-
ing good agreements for both parties when using an
opponent oriented approach in scenarios where un-
predictable issues are more important for the oppo-
nent. Since the Bayesian models misinterpret the
preferences of Boulware and Nice TFT agents, it
is not possible to create win-win situations. Sec-
ond, by selecting a team oriented approach, the
Bayesian team members are selecting the more ac-
ceptable (i.e., the best) unpredictable partial offers
for the team. These unpredictable partial offers re-
port high utility for the team. Thus, once com-
pleted with predictable issues, these offers are ex-
pected to report high utility for the team. Differ-
ently to Agent K, Nice TFT and Boulware agents
are not competitor agents since they will eventu-
ally accept the team's offer when the deadline is
approaching. In the case of the Boulware agent, it
concedes quickly with respect to its aspirations as
the deadline approaches, eventually meeting the re-
quirements of the demanding team offer. As stated
in (Baarslag et al., 2013), when the negotiation time
is reaching its deadline, the Nice TFT agent will
wait for an offer that is not expected to improve
in the remaining time in order to secure an agree-
ment. Hence, the team is able to exploit the other
party by selecting the team oriented approach. Due
to these circumstances, hypotheses H6, H7, and H8
were not supported for agents Boulware and Nice
Tit-for-Tat.
In conclusion, hypotheses H6, H7 and H7 are only
partially supported.
In general, we have discov-
ered that depending on the type of opponent agent,
the values for weights wop, wA have different effects
on the negotiation. When facing exploitable agents
like Nice TFT and Boulware, the team benefits if
the team members take the team oriented approach,
selecting those unpredictable partial offers that are
more acceptable for the team.
If the team faces
competitors like Agent K, the team should match
the values for wop and wA depending on the impor-
tance given by each party to unpredictable issues.
It is acknowledged that, depending on the opponent
and the desired behavior, the team should select dif-
ferent values for the Bayesian proposal weights. A
mechanism for adjusting Bayesian weights based on
the type of opponent is considered as future work.
6.4. Analyzing the Impact of the Reservation Util-
ity
In this experiment, we investigated the impact
of the reservation utility of team members on the
team joint utility. As explained in Section 3.1,
team members jointly prune a part of the negotia-
tion space (i.e., a set of unpredictable partial offers)
which does not contain, with absolute certainty, any
unanimously acceptable offer. This pruning is re-
lated with the reservation utility of team members,
which represents the minimum acceptable utility
by team members. Any offer with a utility lower
than the reservation utility is not acceptable for the
agent.
Lower reservation utilities make it easier to ob-
tain the needed utility by just setting compatible
and predictable issues. Thus, each team member
needs to prune less negotiation space with the un-
predictable partial offers sent to the team media-
tor. Presumably, a joint list of forbidden unpre-
dictable partial offers (i.e., the negotiation space
RUa = 0.35
RUa = 0.50
RUa = 0.65
Similar Avg. Similarity Dissimilar
0.4%
23.8%
73.7%
35.3%
72.6%
90.8%
11.6%
34.2%
81.8%
Table 4: Average percentage of unpredictable partial offers
pruned in the pre-negotiation.
that is pruned) with lower reservation utilities is
smaller than lists constructed with higher reserva-
tion utilities. This leaves more room for finding
an agreement with the opponent. However, if team
members have low reservation utilities, despite hav-
ing more room for finding an agreement, they may
end up with low utility agreements in the end. On
the contrary, with higher reservation utilities, it is
harder to obtain the needed utility with compat-
ible and predictable issues. Therefore, each team
member may need to prune more negotiation space
and the joint list of forbidden unpredictable partial
offers will be larger than the list constructed with
lower reservation utilities. In fact, if team members
set high aspirations with their reservation utility,
it may end up with all the negotiation space be-
ing pruned. If an agreement is found under these
conditions, it may lead to team members achieving
high levels of utility.
In this experiment, we test the impact of different
levels of reservation utility on the team joint utility.
More specifically, as we did in Section 6.1, we tested
teams employing the Bayesian model against dif-
ferent families of strategies: competitor (i.e., Agent
K), matcher (i.e., Nice Tit-for-Tat), inverter (i.e.,
Boulware), and conceder (i.e., conceder). As an ad-
ditional dimension to our study, we also introduced
preference similarity among team members. There-
fore, teams are tested in the scenario where team
members' preferences are dissimilar, the scenario
where team members' preferences have an average
degree of similarity, and the scenario where team
members' preferences are similar.
We configured three types of Bayesian teams with
different levels of reservation utilities: a team with
a relatively low reservation utility RUa = 0.35, a
team with a moderate reservation utility RUa =
0.5, and a team with a high reservation utility
RUa = 0.65. We expect that playing higher reser-
vation utilities against competitor like Agent K will
result in lower average team joint utility due to
many negotiations ending with no agreement. On
the other hand, we expected that playing high reser-
vation utilities against Conceder agents,
inverter
28
Equal importance on unpredictable issues
vs. Agent K
vs. Nice TFT
vs. Boulware
Normal
Opponent Oriented
Team Oriented
T. Joint % Ag. T. Joint % Ag. T. Joint % Ag.
0.168
0.155
0.116
0.202
0.188
0.206
Unpredictable issues more important for the team
0.137
0.126
0.188
84.2
91.1
58.7
100
100
100
100
100
100
vs. Agent K
vs. Nice TFT
vs. Boulware
T. Joint % Ag. T. Joint % Ag. T. Joint % Ag.
Normal
Opponent Oriented
Team Oriented
0.213
0.200
0.248
100
100
100
0.135
0.154
0.196
100
100
100
0.185
0.189
0.213
100
100
100
Unpredictable issues more important for the opponent
vs. Agent K
vs. Nice TFT
vs. Boulware
T. Joint % Ag. T. Joint % Ag. T. Joint % Ag.
Normal
Opponent Oriented
Team Oriented
0.280
0.296
0.271
100
100
92.0
0.186
0.192
0.259
100
100
100
0.326
0.300
0.340
100
100
100
Table 3: Impact of wA and wop on the average team joint utility in different scenarios and proportion of negotiations that
finished with success (% Ag.).
agents like Boulware, and the special case of Nice
Tit-for-Tat, would result in higher average team
joint utility due to the fact that both agents should
eventually concede towards the other parties' de-
mands.
In the case of Nice Tit-for-Tat, we have
observed that when the deadline is approaching it
attempts to secure a deal, making this agent can-
didate to be exploited by setting a high reservation
utility. Therefore, we formulated the following hy-
potheses:
• H9 Playing a high reservation value for team
members (RUa = 0.65) will result in the lowest
average team joint utility against Agent K.
• H10 Playing a high reservation value for team
members (RUa = 0.65) will result in the high-
est average team joint utility against Conceder,
Boulware, and Nice Tit-for-Tat.
These types of teams (RUa = 0.35, RUa =
0.5, RUa = 0.65) were faced in every scenario and
negotiation case against every type of opponent for
20 repetitions. We gathered information on the
team joint utility and the utility obtained by the
opponent, and an ANOVA (α = 0.05) with Bonfer-
roni post-hoc analysis was carried out to determine
results that are statistically better than the rest.
All of the claims of this experimental section are
supported by the ANOVA test with the Bonferroni
post-hoc analysis.
Table 5 shows the results of this experiment in
terms of the average joint utility. A bold font style
is used to highlight those Bayesian team configura-
tions that are statistically the best option against
each opponent. Additionally, the same table shows
the proportion of negotiations that finished with
success (agreement) in this experiment. Table 4
shows the average percentage of unpredictable par-
tial offers that were pruned in the pre-negotiation
depending on the team configuration and team pref-
erence similarity.
With respect to H10, it can be observed that in-
dependently of the degree of dissimilarity among
team members' preferences, team members ob-
tained statistically better team joint utility by set-
ting high reservation utilities (RUa = 0.65) against
Boulware, Conceder, and Nice Tit-for-Tat oppo-
nents. These results support our initial hypothesis
H10.
Nevertheless, H9 is only partially supported.
When facing Agent K, playing the highest reserva-
tion utility configuration lead to the highest team
joint utilities when team members' preferences are
similar or they have an average similarity (top and
middle part of Table 5), which is opposite to our
initial hypothesis H9. In this case, the demands of
the team are still not high enough to preclude agree-
ments with the competitor agent. However, when
team members' preferences are dissimilar (bottom
29
Agent K
Nice Tit-for-Tat
Boulware
Conceder
Similar
T. Joint % Ag. T. Joint % Ag. T. Joint % Ag. T. Joint % Ag.
0.195
0.263
0.350
0.117
0.164
0.286
0.160
0.224
0.354
0.526
0.557
0.635
100
100
99.98
100
100
100
100
100
100
100
100
100
Agent K
Nice Tit-for-Tat
Boulware
Conceder
Average Similarity
T. Joint % Ag. T. Joint % Ag. T. Joint % Ag. T. Joint % Ag.
0.167
0.248
0.242
0.136
0.224
0.313
0.090
0.158
0.268
0.342
0.390
0.470
100
100
74.2
100
100
100
100
100
100
100
100
100
Agent K
Dissimilar
Nice Tit-for-Tat
Boulware
Conceder
T. Joint % Ag. T. Joint % Ag. T. Joint % Ag. T. Joint % Ag.
0.193
0.209
0.068
0.173
0.271
0.409
0.115
0.196
0.346
0.373
0.475
0.580
100
86.6
18.3
100
100
100
100
100
100
100
100
100
RUa = 0.35
RUa = 0.50
RUa = 0.65
RUa = 0.35
RUa = 0.50
RUa = 0.65
RUa = 0.35
RUa = 0.50
RUa = 0.65
Table 5: Average joint Utility (T. Joint) for teams composed by Bayesian team members with different reservation utilities
and proportion of negotiations that finished with success (% Ag.).
part of Table 5), we can observe how setting a high
reservation utility (i.e., RUa = 0.65) gradually be-
comes the worst possible course of action when fac-
ing Agent K, making H9 true for this situation. The
reason for this behavior is mainly explained due to
the decrease in the number of successful negotia-
tions. If we observe the similar scenario, the num-
ber of successful negotiations when facing Agent K
and RUa = 0.65 is 99.98%. If we change to aver-
age similarity scenarios, the number of successful
negotiations is 74.2% when facing Agent K with
RUa = 0.65. The same measure is decreased to
18.3% in the dissimilar scenario. If we observe Ta-
ble 4, as team dissimilarity increases, the number of
unpredictable partial offers to be pruned is larger.
This leaves less negotiation space to be played with
Agent K. Agent K is a competitor agent that at-
tempts to concede as less as possible by estimat-
ing the maximum utility that can be obtained from
the opponent and employing a limit of compromise
when the opponent takes a hard stance. First of all,
if reservation utilities are high, it can be considered
that team members play a hard stance. Second,
if too much negotiation space is pruned, it may be
feasible that the set of remaining unpredictable par-
tial offers precludes Agent K from reaching its limit
of compromise. Thus, employing high reservation
utilities against a competitor agent like Agent K
may result, as we have observed in this case, in an
increase in the number of failed negotiations and
lower team joint utilities.
In conclusion, H10 is supported by our findings,
but H9 is only partially supported. We have ob-
served that team members may benefit from playing
high reservation utilities against Conceder, Boul-
ware, and Nice Tit-for-Tat. If faced against com-
petitors like Agent K, setting high reservation util-
ities may prune too much negotiation space, espe-
cially when team members are dissimilar. This re-
sults in negotiation spaces that may not contain the
minimum limits established by competitor agents,
thus, ending negotiations with failure. For other
scenarios against agent K, the team can benefit
from setting high reservation utilities since the re-
sulting negotiation space still has room to accom-
modate an agreement with the competitor agent.
In general, team members should be cautious when
setting the reservation utility since it may end up
in more failures.
6.5. Team performance with risk attitudes
One scenario that should be considered in multi-
agent systems is agent's attitude towards risk.
Some team members may be more willing to choose
actions that guarantee a safer agreement, while
other may prefer to go for more profitable but less
probable agreements. Classically, agents can show
a risk seeking, a risk averse, and a risk neutral at-
30
titude. The goal of this experiment is determin-
ing how risk attitudes affect the performance of the
proposed negotiation model. In this experiment, we
decided to test three different patterns of behavior:
• Risk averse team member: This team member
selects from the list of available unpredictable
partial offers (see Equation 7, those that guar-
antee the current aspirations of the team mem-
ber sa(t)) the best unpredictable partial of-
fer according to the acceptance probability for
the opponent by using the Bayesian mecha-
nism proposed in this article. When the team
sets predictable issue, the maximum utility ob-
tainable with other unpredictable partial of-
fers may be higher since they may provide a
higher utility of oneself. However, in order to
secure a deal, the team member selects the un-
predictable partial offer that is supposed to be
more acceptable to the opponent even if the
maximum achievable utility is lower. There-
fore, the team member bases its choices on the
acceptability of the offer by the opponent party
instead of the maximum achievable utility.
• Risk seeking team member: This team member
selects for the list of available unpredictable
partial offers (see Equation 7) the best unpre-
dictable partial offer according to the utility
reported by one's own utility function. Hence,
this unpredictable partial offer represents the
choice that enables the team member to get
the maximum achievable utility in the negoti-
ation domain. By filling predictable issues by
the team, the team member is more likely to
get closer to its maximum achievable utility,
even if this action reduces acceptability of the
proposed offer for the opponent.
• Risk neutral team member: This behavior is
represented by the Bayesian team member pre-
sented in this article.
Our initial hypothesis is that the proposed model
is robust to risk attitudes. By robustness, we mean
that the proposed negotiation model will be able
to obtain a team joint utility higher than or com-
parable to the team joint utility obtained by other
state-of-the-art models like SBV:
• H11 Different configurations of team mem-
bers' risk attitudes will yield a team joint util-
ity that is statistically higher or equal to the
results obtained by SBV.
K
0.241
0.259
0.250
0.169
0.181
Similar
N. TFT B.
0.160
0.164
0.216
0.248
0.150
0.238
0.224
0.267
0.283
0.184
Average Similarity
N. TFT B.
0.147
0.158
0.171
0.162
0.137
K
0.188
0.248
0.231
0.149
0.168
0.186
0.224
0.222
0.224
0.170
Dissimilar
K
0.203
0.209
0.228
0.060
0.07
N. TFT B.
0.167
0.196
0.183
0.247
0.16
0.245
0.271
0.265
0.294
0.128
C.
0.372
0.557
0.520
0.615
0.552
C.
0.308
0.390
0.383
0.477
0.324
C.
0.373
0.475
0.452
0.566
0.257
Averse
Neutral
Mix
Seeker
SBV
Averse
Neutral
Mix
Seeker
SBV
Averse
Neutral
Mix
Seeker
SBV
Table 6: Team joint utility obtained by the different team
risk configurations. K (Agent K), N. TFT (Nice Tit-for-Tat),
B. (Boulware), C. (Conceder)
The experimental parameters of the previous
experiments were repeated (RUa = 0.5, βa =
U [0.5, 1]), and we selected four different team con-
figurations. The first one is composed by four neu-
tral team members (Neutral), the second is com-
posed by four risk seeking team members (Seeker),
the third is composed by three risk averse team
members (Averse), and the last team configura-
tion is composed by two neutral team members,
one risk seeker, and one risk averse (Mix). As in
the previous experiments, the team members faced
different opponent profiles in scenarios where team
members' preference profiles are similar, scenarios
with average similarity of team members' prefer-
ences, and scenarios where team members have dis-
similar preferences. The team faces the same op-
ponent strategies presented in the previous experi-
ments. We gathered information on the team joint
utility.
The results of this experiment can be found on
Table 6. Results highlighted in bold represent the
statistically higher team joint utility configurations
for each negotiation scenario (ANOVA, α = 0.05).
As it can be observed, the risk seeking configura-
tion usually gets a higher team joint utility as long
as the opponent faced is not a competitor. This
configuration is able to obtain one of the best team
31
joint in all negotiation scenarios with Nice Tit-for-
Tat, Boulware, and Conceder. When negotiating
against Agent K , many negotiations end with fail-
ure since the opponent also has high aspirations.
Other opponents like Nice Tit-for-Tat, Boulware,
or Conceder match the high aspirations of the risk
seeking team at one point or another of the negoti-
ation, resulting in a higher team joint utility.
The mixed configuration usually gets the next
highest team joint utility, being able to obtain sta-
tistically equivalent results in some scenarios like
negotiating against Nice Tit-for-Tat in the average
similarity scenario, and statistically higher results
like negotiations against Agent K in dissimilar sce-
narios. This configuration obtains the best team
joint utility in all of the scenarios involving against
Agent K.
The neutral configuration, obtains statistically
good results in some scenarios where team members
have similar preferences and the Bayesian mecha-
nism is able to learn the opponent preferences prop-
erly (e.g., negotiating against Agent K, some sce-
narios against Boulware and Nice Tit-for-Tat). The
risk averse configuration is only able to obtain some
of the top results in very specific scenarios like ne-
gotiations against in similar settings against Agent
K.
However, in no case any of the team configura-
tions proposed in this article is worse than SBV. All
of the results are statistically higher than those ob-
tained by SBV, except for those obtained by the risk
seeking configuration against Agent K, which are
statistically equivalent to those obtained by SBV.
Therefore, despite being affected by team hetero-
geneity, the proposed model is robust and it is able
to obtain results that are at least equal to the state-
of-the-art (i.e., SBV), outperforming it in many sit-
uations. This result supports our initial hypothesis
H11.
7. Related Work
The contributions of this article to the auto-
mated negotiation community can be divided into
two different categories: general contributions to
the field of automated negotiation, and contribu-
tions to the specific field of agent-based negotiation
teams. Next, we analyze the contributions to each
of these fields.
32
7.1. Automated negotiation with single individual
parties
The artificial intelligence community has focused
on bilateral or multi-party negotiations where par-
ties are composed of single individuals. The main
contribution of our work with respect to the general
field of automated negotiation is that we support
negotiation parties composed by multiple individ-
uals. Apart from that, in the next paragraphs we
discuss other contributions of our present negotia-
tion framework with respect to works in automated
negotiation.
Faratin et al. (Faratin et al., 1998) introduced
some of the most widely used families of concession
tactics in negotiation. The authors proposed con-
cession strategies for negotiation issues that are a
mix of different families of concession tactics. The
authors divide these concession tactics into three
different families:
(i) time-dependent concession
tactics; (ii) behavior-dependent concession tactics;
and (iii) resource-dependent tactics. Our negotia-
tion framework also considers time as crucial ele-
ment in negotiation. Therefore, team members em-
ploy time tactics inspired in those introduced by
Faratin et al. However, the authors do not propose
any explicit preference learning mechanism.
In Zeng and Sycara (Zeng and Sycara, 1998), the
authors argue about the benefits of using Bayesian
models in negotiation and they study a bilateral
negotiation case where the buyer attempts to learn
the reservation price of the seller by updating its
beliefs with Bayesian learning. Despite the fact
that it introduces the use of Bayesian learning in
negotiation, the article only focuses on single issue
models. One of our team member models also uses
Bayesian learning as a method for learning other
agents' preferences. The main different resides in
the fact that our Bayesian approach attempts to
model which instantiations of unpredictable issues
are acceptable for the opponent and the team in
multi-issue negotiations.
Ehtamo et al.
(Ehtamo et al., 2001) propose
a mediated multi-party negotiation protocol which
looks for joint gains in an iterated way. The algo-
rithm starts from a tentative agreement and moves
in a direction according to what the agents prefer
regarding some offers' comparison. Results showed
that the algorithm converges quickly to Pareto
optimal points. However, the work proposed by
Ehtamo et al. does not support unpredictable is-
sues and multiple individual parties.
Klein et al. (Klein et al., 2003a) propose a me-
diated negotiation model which can be extended to
multiple parties. Their main goal is to provide so-
lutions for negotiation processes that use complex
utility functions to model agents' preferences. The
negotiation attributes are binary and no longer in-
dependent. Our work supports multiple individual
parties and negotiation issues with unrestricted do-
main (e.g., real, integer, discrete, binary, etc.).
In Coehoorn et al.
(Coehoorn and Jennings,
2004), the authors propose the use of kernel den-
sity estimation for the estimation of the importance
weights of the linear additive utility function. The
agent calculates tuples composed of the difference
between pairs of consecutive offers, the estimated
weight for the issue, and the probability density
of the weight. These tuples form a three dimen-
sional kernel that is used along the other kernels
to calculate an estimation of the real issue weight.
Our proposed model is capable of learning in nego-
tiations where domains are also composed by dis-
crete issues, which is not supported by Coehoorn
et al. Moreover, the learning mechanism proposed
for our team members deals with the information in
one single negotiation, whereas the aforementioned
mechanism learns over several negotiations.
Later, Narayanan et al.(Narayanan and Jennings,
2006) present a negotiation framework where pairs
of agents negotiate over a single issue. The authors
assume that agents' strategies may change over
time. Non-stationary Markov chains and Bayesian
learning are employed to tackle the uncertainty in
this domain, and eventually converge towards the
optimal negotiation strategy.
In our case, we fo-
cus on one single negotiation process, and our team
members learn over the information provided by the
current negotiation. Additionally, we consider ne-
gotiations where multiple issues are involved.
Another example of the use of Bayesian learning
in negotiation is presented by Buffett et al. (Buf-
fett and Spencer, 2007). In the aforementioned ar-
ticle, a bilateral framework is presented in a do-
main where agents negotiate over a set of binary
issues. A Bayesian classifier is employed to classify
opponent's preferences into classes of preference re-
lations. Groups of similar preference relations are
grouped according to the k-means algorithm prior
to the negotiation process. Our model does not re-
quire learning prior to the negotiation, which may
require a costly learning process every negotiation
to avoid domain dependent classifiers. Moreover,
we consider any kind of issue type in the negotia-
33
tion domain.
Carbonneau et al. (Carbonneau et al., 2008) pro-
pose a neural network that takes as input the nego-
tiation history of a bilateral negotiation with con-
tinuous issues and an offer to make an estimation
of the opponent's counter-offer. This approach re-
quires that an artificial neural network is trained
per negotiation case. Similarly, the same authors
propose an improvement over their previous work
in (Carbonneau et al., 2011).
It aims to make a
predictive model that does not depend on the ne-
gotiation case. The model takes pairs of negotiation
issues as inputs of the neural network, where one of
the issues is considered the primary issue (i.e., inde-
pendent variable) and the other issue is considered
the secondary issue (i.e., dependent variable). Dif-
ferently to these works, our proposed model does
not rely on information from past negotiations. It
only employs information gathered in the present
process.
Robu et al.(Robu and La Poutr´e, 2008) introduce
a bilateral negotiation model where agents repre-
sent their preferences by means of utility graphs.
The negotiation domain is formed of bundles of
items that can be either included or excluded in a
final deal. Utility graphs are graphical models that
relate negotiation issues that are dependent. Nodes
represent negotiation issues whereas arcs connect is-
sues that have some joint effect on the utility func-
tion (i.e., positive for complementary issues, and
negative for substitutable issues). Hence, utility
graphs represent binary dependencies between is-
sues. The authors propose a negotiation scenario
where the buyer's preferences and the seller's pref-
erences are modeled through utility graphs. The
seller is the agent that carries out a more thor-
ough exploration of the negotiation space in order
to search for agreements where both parties are sat-
isfied. With this purpose, the seller builds a model
of the buyer's preferences based on historic informa-
tion of past deals and expert knowledge about the
negotiation domain. Our work does not consider
dependencies between issues, however it is capable
of supporting every type of domain for negotiation
issues (i.e., real, binary, discrete, etc.).
In (Aydogan and Yolum, 2012), a concept-based
learning method is proposed for modeling oppo-
nent preferences and generating well-targeted of-
fers. In that method, the preferences of the oppo-
nent are represented via disjunctive and conjunc-
tive constraints. In this article, our aim is also to
find the agreement earlier by means of learning the
other participants' preferences but the preferences
are represented by means of additive utility func-
tions. In our case, an offer that is rejected by the
opponent may become acceptable over time because
the opponent may concede; therefore, we choose a
probabilistic learning method.
Williams et al.
(Williams et al., 2012) present
a negotiation framework for coordinating multi-
ple bilateral negotiations with different opponents.
The agent simultaneously negotiates with differ-
ent opponent in order to acquire a desired good
at the best possible condition. The framework
makes use of optimization techniques and proba-
bilistic information in order to carry out this coor-
dination. Similarly, Mansour et al. (Mansour and
Kowalczyk, 2012) present a meta-strategy for coor-
dination multiple negotiations with different sellers.
The meta-strategy adjusts the concession speed ac-
cording to the current state of the multiple nego-
tiation threads. The problem of agent-based nego-
tiation teams is different since multiple agents col-
laborate in the same party to get a deal from an
opponent, instead of competing between each other
to get an individual deal.
7.2. Agent-based negotiation teams
As far as we are concerned, only our previ-
ous works (Sanchez-Anguix et al., 2011, 2012a,b)
have considered negotiation teams in computa-
tional models.
In (Sanchez-Anguix et al., 2011,
2012a) four different computational models for a
negotiation team negotiating with a single oppo-
nent are presented. These four models attempt
to gather four different minimum levels of una-
nimity regarding team decisions: representative ap-
proach (RE, no unanimity), similarity simple voting
(SSV, majority/plurality), similarity Borda voting
(SBV, semi-unanimity), and full unanimity medi-
ated (FUM, unanimity).
The RE model is based on the selection of team
members as representative of the team. The rep-
resentative acts on behalf of the group by taking
decisions according to its own negotiation strategy
and utility function.
SSV and SBV are models based on the presence
of a mediator that coordinates voting processes. In
the case of SSV, a majority voting is employed to
determine whether or not the opponent's offer is
accepted, and a majority voting is used to select
which offer is sent from those offers proposed by
team members to be sent to the opponent. In the
case of SBV, a unanimity voting is designated as
34
the mechanism to decide if opponent's offers are
accepted and Borda count is used to decide which
offer proposed by team members is sent to the oppo-
nent. In both cases, SSV and SBV team members
decide which offers are sent to the opponent using
similarity heuristics that consider the opponent's
and team's last offers.
FUM is a mediated model where the opponent's
offers are evaluated by means of a unanimity vot-
ing process, and team's offers are built issue per
issue by aggregation rules. These models were in-
troduced at AAMAS 2011 (Sanchez-Anguix et al.,
2011) as the first approach to tackle problems in-
volving negotiation teams.
Later, we studied the special properties of
FUM, given that it was the intra-tem negotiation
model capable of guaranteeing unanimity regarding
team decisions at each negotiation round (Sanchez-
Anguix et al., 2012b). We proved how unanimity
is guaranteed in FUM, how the intra-team strat-
egy us robust against certain types of manipulation
attacks, and how team members did not have incen-
tive to strongly deviate from the proposed model.
We provided a full fledged experimental analysis
of the four intra-team strategies in different negoti-
ation environments (Sanchez-Anguix et al., 2012a).
The results showed that FUM was able to guarantee
better results for the negotiation team in most ne-
gotiation scenarios. Only SBV is able to guarantee
similar results in a limited number of negotiation
cases.
Even though these four models cover different
levels of unanimity with regards to team decisions,
they were initially designed to provide solutions for
negotiation domains that are exclusively composed
by predictable and compatible issues among team
members (e.g., price, quality, and dispatch time in
a team of buyers).
Domains exclusively composed by compatible and
predictable issues among team members cover a
range of feasible negotiation domains. However, a
relatively large number of negotiation domains also
include unpredictable issues among team members.
RE, SSV, and SBV can be easily adapted to do-
mains that include unpredictable issues among team
members by using a different similarity heuristic.
FUM, the model that guarantees unanimity with
regards to team decisions, cannot be used in these
domains since, in the offer proposal mechanism, it
aggregates issue values based on the fact that all of
the negotiation issues are compatible and predictable
among team members.
As stated, whenever it is possible, it is desirable
for the final agreement to be unanimously accept-
able for all of the team members since the oppo-
site situation may end up in users perceiving un-
fairness, which may affect commitment to the deci-
sion, group attachment, and trust (Korsgaard et al.,
1995), and users that are not satisfied with agree-
ments found automatically may end up leaving the
electronic application.
The model proposed in this article advances the
stat of the art in agent-based negotiation teams
by solving both problems:
it guarantees unanim-
ity with regards to team decisions, and it sup-
ports unpredictable negotiation issues, which were
not specifically supported in previous models. For
that matter, the negotiation domains is split be-
tween those issues that are compatible and pre-
dictable among team members, and those issues
that are unpredictable.
In the former case, part
of the mechanism employed in FUM is employed.
By doing so, the model is capable of guarantee-
ing that team members are able to get as much as
they need from predictable and compatible issues.
In the latter case, the team discards those com-
binations of unpredictable partial offers that pre-
clude the team from reaching unanimously accept-
able agreements, even if the most is obtained from
predictable and compatible issues. Then, in the offer
proposal mechanism team members select one of re-
maining unpredictable partial offers, and complete
the predictable and compatible issue with the val-
ues that they need to make it at least unanimously
acceptable. The idea behind this splitting, is coop-
erating as much as possible on those issues that are
predictable and compatible among team members
to create less intra-team conflict in unpredictable
issues.
8. Conclusions & Future Work
In this article we have presented a new me-
diated team negotiation model for a team as a
multi-individual party negotiating with an oppo-
nent in the alternating offers protocol. The present
model is capable of assuring unanimously accept-
able agreements for all of the team members.
It
takes advantage of the categorization of negotia-
tion issues as predictable and compatible, and un-
predictable. The former are those issues whose
preferential order over issue values is known from
the negotiation domain and it is common among
team members (e.g., price in a team of buyers),
35
whereas the latter are those issues whose prefer-
ential order over issues values is not known in the
negotiation domain. In the case of predictable and
compatible issues, there is full potential for coop-
eration among team members since if one of the
team members demands more from the issue, the
other team members are also benefited. Our negoti-
ation model takes advantage of this property. Dur-
ing the pre-negotiation, each team member shares
with a team mediator those unpredictable partial
offers (i.e., partial offers that have all of the unpre-
dictable issues instantiated) that, even if the team
demands the most from predictable issues, preclude
the agent from achieving its reservation value. A
joint list forbidden unpredictable partial offers is
constructed by the team mediator from the lists re-
ceived from the team members. In the negotiation,
the team mediator coordinates a unanimity voting
process to decide whether or not to accept offers
received from the opponent. As for the mechanism
employed to decide on which offer should be sent to
the opponent, the team mediator coordinates two
processes: a proposing and voting process where
each team member suggests an unpredictable par-
tial offer not included in the forbidden list followed
by a Borda voting on candidates received, and an
iterated process where predictable issues are set is-
sue per issue attending to the demands of the team
members.
We have proposed two different types of team
members for the current model: a basic team mem-
ber that proposes unpredictable partial offers dur-
ing the negotiation solely guided by its own utility
function, and a Bayesian team member that sug-
gests unpredictable partial offers based on the pref-
erences of the team and the preferences of the op-
ponent. Results have shown that, as long as pref-
erential conflict is present in the team, team mem-
bers have an incentive to employ the Bayesian strat-
egy over the basic strategy. In any case, we have
shown that both approaches outperform other ex-
isting models for negotiation teams. Moreover, we
have determined that when two negotiation teams
face each other, both teams benefit from includ-
ing Bayesian team members in the negotiation. Fi-
nally, we have shown that team members may ben-
efit from playing higher reservation utilities against
conceders, matchers and inverters. Nevertheless,
setting high reservation utilities may become the
worst option as team members' preferences are
more dissimilar and the opponent plays a competi-
tor strategy.
The topic of agent-based negotiation teams re-
mains largely unexplored compared to negotiation
involving single individual parties. The present
work has focused on agent-based negotiation teams
where team members have different preferences but
they share the same knowledge regarding the nego-
tiation domain. One potential area of research is
modeling negotiation teams where team members
differ in their knowledge of the negotiation domain
and their skills related to the negotiation process.
As another line of future work, one could consider
the problem of forming negotiation teams based on
the individual list of unpredictable partial offers.
Lists that are more similar may suggest team mem-
bers that are more similar, which, if put together
in the same team, may result in more cooperative
teams. Related to team formation, dynamic nego-
tiation teams where team members may join and
leave the team during the negotiation can be con-
sidered an alternative line of research.
Acknowledgements
This research is partially supported by TIN2012-
36586-C03-01 of
the Spanish government and
PROMETEOII/2013/019 of Generalitat Valen-
ciana. Other part of this research is supported
by the Dutch Technology Foundation STW, ap-
plied science division of NWO and the Technology
Program of the Ministry of Economic Affairs; the
Pocket Negotiator project with grant number VICI-
project 08075.
References
E. W. Ngai, F. Wat, A literature review and classification
of electronic commerce research, Information & Manage-
ment 39 (5) (2002) 415 -- 429.
M. Grieger, Electronic marketplaces: A literature review and
a call for supply chain management research, European
journal of operational research 144 (2) (2003) 280 -- 294.
J. Wareham, J. G. Zheng, D. Straub, Critical themes in elec-
tronic commerce research: a meta-analysis, Journal of In-
formation Technology 20 (1) (2005) 1 -- 19.
R. H. Guttman, A. G. Moukas, P. Maes, Agent-mediated
electronic commerce: a survey, Knowledge Engineering
Review 13 (2) (1998) 147 -- 159, ISSN 0269-8889, doi:
http://dx.doi.org/10.1017/S0269888998002082.
C. Sierra, F. Dignum, Agent-mediated electronic commerce:
Scientific and technological roadmap, in: Agent Mediated
Electronic Commerce, Springer, 1 -- 18, 2001.
E. Oliveira, A. Rocha, Agents advanced features for nego-
tiation in electronic commerce and virtual organisations
formation process, in: Agent Mediated Electronic Com-
merce, Springer, 78 -- 97, 2001.
36
M. He, N. R. Jennings, H.-F. Leung, On agent-mediated
electronic commerce, Knowledge and Data Engineering,
IEEE Transactions on 15 (4) (2003) 985 -- 1003.
A. R. Lomuscio, M. Wooldridge, N. R. Jennings, A Classi-
fication Scheme for Negotiation in Electronic Commerce,
Group Decision and Negotiation 12 (2003) 31 -- 56, ISSN
0926-2644.
T. D. Nguyen, N. R. Jennings, Managing commitments in
multiple concurrent negotiations, Electronic Commerce
Research and Applications 4 (4) (2005) 362 -- 376, ISSN
1567-4223, developments in intelligent support for e-
Commerce negotiation applications The 6th International
Conference on Electronic Commerce.
S. Buffett, B. Spencer, A bayesian classifier for learning
opponents' preferences in multi-object automated nego-
tiation, Electronic Commerce Research and Applications
6 (3) (2007) 274 -- 284.
Towards a web services and intelligent agents-based negotia-
tion system for {B2B} eCommerce, Electronic Commerce
Research and Applications 6 (3) (2007) 260 -- 273, ISSN
1567-4223.
C.-C. H. Chan, C.-B. Cheng, C.-H. Hsu, Bargaining strategy
formulation with CRM for an e-commerce agent, Elec-
tronic Commerce Research and Applications 6 (4) (2008)
490 -- 498.
M. Klein, P. Faratin, H. Sayama, Y. Bar-Yam, Protocols
for Negotiating Complex Contracts, IEEE Intelligent Sys-
tems 18 (6) (2003a) 32 -- 38, ISSN 1541-1672.
C. Williams, V. Robu, E. Gerding, N. Jennings, Using gaus-
sian processes to optimise concession in complex negotia-
tions against unknown opponents, in: Proceedings of the
22nd International Joint Conference on Artificial Intelli-
gence (IJCAI'11), 432 -- 438, 2011.
T. Baarslag, K. Hindriks, C. Jonker, S. Kraus, R. Lin,
The First Automated Negotiating Agents Competition
(ANAC 2010), in: New Trends in Agent-Based Complex
Automated Negotiations, vol. 383 of Studies in Compu-
tational Intelligence, Springer Berlin / Heidelberg, ISBN
978-3-642-24695-1, 113 -- 135, 2012.
D. Ball, P. S. Coelho, M. J. Vilares, Service personalization
and loyalty, Journal of Services Marketing 20 (6) (2006)
391 -- 403.
T.-P. Liang, H.-J. Lai, Y.-C. Ku, Personalized content rec-
ommendation and user satisfaction: Theoretical synthesis
and empirical findings, Journal of Management Informa-
tion Systems 23 (3) (2007) 45 -- 70.
L. Thompson, The Mind and heart of the negotiator, Pren-
tice Hall Press, Upper Saddle River, NJ, USA, 2003.
G. Lai, K. Sycara, C. Li, A decentralized model for auto-
mated multi-attribute negotiations with incomplete infor-
mation and general utility functions, Multiagent and Grid
Systems 4 (1) (2008) 45 -- 65, ISSN 1574-1702.
Y. Chevaleyre, U. Endriss, J. Lang, N. Maudet, A Short In-
troduction to Computational Social Choice, in: SOFSEM
2007: Theory and Practice of Computer Science, vol. 4362
of Lecture Notes in Computer Science, Springer Berlin /
Heidelberg, 51 -- 69, 2007.
H. Taylor, Most people are privacy pragmatists who, while
concerned about privacy, will sometimes trade it off for
other benefits, The Harris Poll 17 (2003) 19.
P. Faratin, C. Sierra, N. R. Jennings, Negotiation Decision
Functions for Autonomous Agents, International Journal
of Robotics and Autonomous Systems 24 (3-4) (1998)
159 -- 182.
D. Zeng, K. Sycara, Bayesian learning in negotia-
tion, International Journal of Human-Computer Stud-
ies 48 (1)
ISSN 1071-5819, doi:
http://dx.doi.org/10.1006/ijhc.1997.0164.
(1998) 125 -- 141,
M. Klein, P. Faratin, H. Sayama, Y. Bar-Yam, Negotiating
Complex Contracts, Group Decision and Negotiation 12
(2003b) 111 -- 125.
R. M. Coehoorn, N. R. Jennings, Learning on oppo-
nent's preferences to make effective multi-issue negotia-
tion trade-offs, in: The 6th International Conference on
Electronic Commerce (ICEC'04), 59 -- 68, 2004.
V. Sanchez-Anguix, S. Valero, V. Julian, V. Botti, A. Garc´ıa-
Fornes, Evolutionary-aided negotiation model for bilateral
bargaining in Ambient Intelligence domains with complex
utility functions, Information Sciences 222 (2013) 25 -- 46.
R. Aydogan, P. Yolum, Learning opponent's preferences for
effective negotiation: an approach based on concept learn-
ing, Auton. Agents Multi-Agent Syst. 24 (2012) 104 -- 140.
E. Mannix, Strength in Numbers: Negotiating as a Team,
Negotiation 8 (5) (2005) 3 -- 5.
N. Halevy, Team Negotiation: Social, Epistemic, Economic,
and Psychological Consequences of Subgroup Conflict,
Personality and Social Psychology Bulletin 34 (2008)
1687 -- 1702.
L. Thompson, E. Peterson, S. Brodt, Team negotiation: An
examination of integrative and distributive bargaining,
Journal of Personality and Social Psychology 70 (1) (1996)
66 -- 78.
S. Brodt, L. Thompson, Negotiating Teams: A levels of anal-
ysis, Group Dynamics 5 (3) (2001) 208 -- 219.
K. Behfar, R. A. Friedman, J. M. Brett, The Team Nego-
tiation Challenge: Defining and Managing the Internal
Challenges of Negotiating Teams, in: Proceedings of the
21st Annual Conference for the International Association
for Conflict Management (IACM-2008), 2008.
G. Ortmann, R. King, Agricultural cooperatives I: history,
theory and problems, Agrekon 46 (1) (2007) 18 -- 46.
V. Sanchez-Anguix, V. Julian, V. Botti, A. Garcia-Fornes,
Reaching Unanimous Agreements within Agent-Based
Negotiation Teams with Linear and Monotonic Utility
Functions, IEEE Transactions on Systems, Man and Cy-
bernetics, Part B 42 (3) (2012a) 778 -- 792.
V. Sanchez-Anguix, T. Dai, Z. Semnani-Azad, K. Sycara,
V. Botti, Modeling power distance and individual-
ism/collectivism in negotiation team dynamics,
in: 45
Hawaii
International Conference on System Sciences
(HICSS-45), 628 -- 637, 2012b.
A. Rubinstein, Perfect equilibrium in a bargaining model,
Econometrica 50 (1982) 155 -- 162.
K. V. Hindriks, D. Tykhonov, Towards a quality assessment
method for learning preference profiles in negotiation, in:
Agent-Mediated Electronic Commerce and Trading Agent
Design and Analysis, Springer, 46 -- 59, 2010.
I. Marsa-Maestre, M. Klein, C. M. Jonker, R. Aydo-
to protocols: Towards a ne-
gan, From problems
gotiation handbook, Decision Support Systems doi:
http://dx.doi.org/10.1016/j.dss.2013.05.019.
L. A. Stole, Nonlinear pricing and oligopoly, Journal of Eco-
nomics & Management Strategy 4 (4) (1995) 529 -- 562.
B. Bauer, J. P. Muller, J. Odell, Agent UML: A formalism
for specifying multiagent interaction, in: Agent-oriented
software engineering, vol. 1957, Springer, Berlin, 91 -- 103,
2001.
M. E. Gaston, M. desJardins, Agent-organized networks for
dynamic team formation, in: Proceedings of the fourth
international joint conference on Autonomous agents and
multiagent systems (AAMAS'05), ACM, New York, NY,
USA, ISBN 1-59593-093-0, 230 -- 237, 2005.
T. Rahwan, S. D. Ramchurn, N. R. Jennings, A. Giovan-
nucci, An anytime algorithm for optimal coalition struc-
ture generation, Journal of Artificial Intelligence Research
34 (1) (2009) 521 -- 567, ISSN 1076-9757.
H. Nurmi, Voting systems for social choice, Handbook of
H. Farhangi, The path of the smart grid, Power and Energy
Group Decision and Negotiation (2010) 167 -- 182.
Magazine, IEEE 8 (1) (2010) 18 -- 28.
S. Russell, P. Norvig, Artificial Intelligence: A Modern Ap-
S. D. Ramchurn, P. Vytelingum, A. Rogers, N. R. Jennings,
Putting the'smarts' into the smart grid: a grand challenge
for artificial intelligence, Communications of the ACM
55 (4) (2012) 86 -- 97.
F. M. Brazier, F. Cornelissen, R. Gustavsson, C. M. Jonker,
O. Lindeberg, B. Polak, J. Treur, A multi-agent system
performing one-to-many negotiation for load balancing of
electricity use, Electronic Commerce Research and Appli-
cations 1 (2) (2002) 208 -- 224.
S. Lamparter, S. Becher, J.-G. Fischer, An agent-based mar-
ket platform for smart grids, in: Proceedings of the 9th In-
ternational Conference on Autonomous Agents and Multi-
agent Systems: Industry track, International Foundation
for Autonomous Agents and Multiagent Systems, 1689 --
1696, 2010.
H. Morais, T. Pinto, Z. Vale, I. Pra¸ca, Multilevel negotia-
tion in smart grids for VPP management of distributed
resources, IEEE Intelligent Systems 27 (6) (2012) 8 -- 16.
M. Korsgaard, D. Schweiger, H. Sapienza, Building commit-
ment, attachment, and trust in strategic decision-making
teams: The role of procedural justice, Academy of Man-
agement Journal (1995) 60 -- 84.
V. Sanchez-Anguix, V. Julian, V. Botti, A. Garcia-Fornes,
Analyzing Intra-Team Strategies for Agent-Based Nego-
tiation Teams, in: 10th International Conference on Au-
tonomous Agents and Multiagent Systems (AAMAS'11),
929 -- 936, 2011.
proach, Pearson Education, 2003.
R. Lin, S. Kraus, T. Baarslag, D. Tykhonov, K. Hindriks,
C. M. Jonker, Genius: An Integrated Environment for
Supporting the Design of Generic Automated Negotiators,
Computational Intelligence (2012) In Press.
T. Baarslag, K. V. Hindriks, C. M. Jonker, Towards a Quan-
titative Concession-Based Classification Method of Nego-
tiation Strategies, in: Agents in Principle, Agents in Prac-
tice. Lecture Notes of The 14th International Conference
on Principles and Practice of Multi-Agent Systems, 143 --
158, 2011.
S. Kawaguchi, K. Fujita, T. Ito, Compromising Strategy
Based on Estimated Maximum Utility for Automated Ne-
gotiation Agents Competition (ANAC-10), in: Modern
Approaches in Applied Intelligence, vol. 6704, Springer
Berlin / Heidelberg, 501 -- 510, 2011.
T. Baarslag, K. V. Hindriks, C. M. Jonker, A Tit for Tat
Negotiation Strategy for Real-Time Bilateral Negotia-
tions, vol. 435 of Studies in Computational Intelligence,
Springer Berlin Heidelberg, 229 -- 233, 2013.
H. Ehtamo, E. Kettunen, R. P. Hamalainen, Searching for
joint gains in multi-party negotiations, European Journal
of Operational Research 130 (1) (2001) 54 -- 69.
V. Narayanan, N. R. Jennings, Learning to Negotiate Op-
timally in Non-stationary Environments, in: Cooperative
Information Agents X, 10th International Workshop, CIA
2006, Edinburgh, UK, September 11-13, 2006, Proceed-
37
ings, vol. 4149 of Lecture Notes in Computer Science,
Springer, 288 -- 300, 2006.
R. Carbonneau, G. Kersten, R. Vahidov, Predicting oppo-
nent's moves in electronic negotiations using neural net-
works, Expert Systems with Applications 34 (2) (2008)
1266 -- 1273.
R. A. Carbonneau, G. E. Kersten, R. M. Vahidov, Pair-
wise issue modeling for negotiation counteroffer prediction
using neural networks, Decision Support Systems 50 (2)
(2011) 449 -- 459.
V. Robu, J. A. La Poutr´e, Retrieving the Structure of Utility
Graphs Used in Multi-Item Negotiation through Collab-
orative Filtering of Aggregate Buyer Preferences, in: Ra-
tional, Robust and Secure Negotiations, vol. 89 of Com-
putational Intelligence, Springer, 2008.
C. R. Williams, V. Robu, E. H. Gerding, N. R. Jennings,
Negotiating Concurrently with Unknown Opponents in
Complex, Real-Time Domains, in: 20th European Con-
ference on Artificial Intelligence, vol. 242, 834 -- 839, 2012.
K. Mansour, R. Kowalczyk, A Meta-Strategy for Coordi-
nating of One-to-Many Negotiation over Multiple Issues,
in: Foundations of Intelligent Systems, vol. 122, Springer
Berlin / Heidelberg, ISBN 978-3-642-25663-9, 343 -- 353,
2012.
38
|
cs/9809036 | 1 | 9809 | 1998-09-20T12:48:43 | Document Archiving, Replication and Migration Container for Mobile Web Users | [
"cs.MA",
"cs.MM"
] | With the increasing use of mobile workstations for a wide variety of tasks and associated information needs, and with many variations of available networks, access to data becomes a prime consideration. This paper discusses issues of workstation mobility and proposes a solution wherein the data structures are accessed in an encapsulated form - through the Portable File System (PFS) wrapper. The paper discusses an implementation of the Portable File System, highlighting the architecture and commenting upon performance of an experimental system. Although investigations have been focused upon mobile access of WWW documents, this technique could be applied to any mobile data access situation. | cs.MA | cs | Document Archiving, Replication and Migration Container for
Mobile Web Users
Abstract
With the increasing use of mobile workstations for a wide
variety of tasks and associated information needs, and with
many variations of available networks, access to data
becomes a prime consideration. This paper discusses issues
of workstation mobility and proposes a solution wherein
the data structures are accessed in an encapsulated form -
through the Portable File System (PFS) wrapper. The paper
discusses an implementation of the Portable File System,
the architecture and commenting upon
highlighting
performance of an experimental
system. Although
investigations have been focused upon mobile access of
WWW documents, this technique could be applied to any
mobile data access situation.
file-system, replication, archiving,
Keywords: server,
mobility
1. Introduction
:•
•
•
2. Distributed File Systems Approach
P
e
t
e
r
S
t
a
n
s
k
i
,
S
t
e
p
h
e
n
G
i
l
e
s
a
n
d
A
r
k
a
d
y
Z
a
s
l
a
v
s
k
y
D
e
p
a
r
t
m
e
n
t
o
f
C
o
m
p
u
t
e
r
T
e
c
h
n
o
l
o
g
y
,
M
o
n
a
s
h
U
n
i
v
e
r
s
i
t
y
9
0
0
D
a
n
d
e
n
o
n
g
R
o
a
d
,
C
a
u
l
f
i
e
l
d
E
a
s
t
,
V
i
c
3
1
4
5
,
A
u
s
t
r
a
l
i
a
{
p
s
t
a
n
s
k
i
,
s
g
i
l
e
s
,
a
z
a
s
l
a
v
s
}
@
b
r
o
n
c
h
o
.
c
t
.
m
o
n
a
s
h
.
e
d
u
.
a
u
T
h
e
m
o
b
i
l
e
w
o
r
k
s
t
a
t
i
o
n
s
u
c
h
a
s
P
e
r
s
o
n
a
l
D
i
g
i
t
a
l
A
s
s
i
s
t
a
n
t
(
P
D
A
)
o
r
n
o
t
e
b
o
o
k
,
h
a
v
e
b
e
c
o
m
e
i
n
d
i
s
p
e
n
s
a
b
l
e
t
o
o
l
s
f
o
r
t
h
e
t
r
a
v
e
l
i
n
g
b
u
s
i
n
e
s
s
p
e
r
s
o
n
.
P
e
r
s
o
n
a
l
c
o
m
m
u
n
i
c
a
t
i
o
n
d
e
v
i
c
e
s
s
u
c
h
a
s
t
h
e
A
p
p
l
e
N
e
w
t
o
n
,
U
S
R
o
b
o
t
i
c
s
/
3
C
O
M
P
I
L
O
T
,
N
o
k
i
a
C
o
m
m
u
n
i
c
a
t
o
r
9
0
0
0
,
a
l
o
n
g
w
i
t
h
a
p
l
e
t
h
o
r
a
o
f
P
D
A
s
r
u
n
n
i
n
g
W
i
n
d
o
w
s
C
E
,
o
f
f
e
r
n
e
t
w
o
r
k
c
o
n
n
e
c
t
i
v
i
t
y
w
i
t
h
a
l
i
m
i
t
e
d
b
a
n
d
w
i
d
t
h
.
T
h
e
s
e
d
e
v
i
c
e
s
a
r
e
o
f
t
e
n
u
s
e
d
t
o
c
o
n
n
e
c
t
t
o
t
h
e
I
n
t
e
r
n
e
t
f
o
r
a
v
a
r
i
e
t
y
o
f
a
p
p
l
i
c
a
t
i
o
n
s
r
a
n
g
i
n
g
f
r
o
m
e
m
a
i
l
t
o
W
o
r
l
d
W
i
d
e
W
e
b
a
c
c
e
s
s
.
M
a
n
y
o
r
g
a
n
i
z
a
t
i
o
n
s
a
r
e
a
p
p
e
a
r
i
n
g
o
n
t
h
e
w
e
b
d
a
i
l
y
.
T
h
e
y
o
f
f
e
r
g
e
n
e
r
a
l
u
s
e
r
a
c
c
e
s
s
f
o
r
a
d
v
e
r
t
i
s
i
n
g
a
n
d
m
a
r
k
e
t
i
n
g
,
w
h
i
l
e
a
l
l
o
w
i
n
g
i
n
t
e
r
n
a
l
s
t
a
f
f
t
o
a
c
c
e
s
s
m
o
r
e
c
o
n
f
i
d
e
n
t
i
a
l
i
n
f
o
r
m
a
t
i
o
n
.
T
h
e
t
r
a
v
e
l
i
n
g
b
u
s
i
n
e
s
s
p
e
r
s
o
n
o
f
t
e
n
m
a
k
e
s
u
s
e
o
f
t
h
e
w
e
b
t
o
p
r
o
v
i
d
e
p
r
o
d
u
c
t
i
n
f
o
r
m
a
t
i
o
n
v
i
a
a
w
i
r
e
l
e
s
s
n
e
t
w
o
r
k
.
W
i
r
e
l
e
s
s
m
o
b
i
l
e
c
o
m
p
u
t
i
n
g
i
s
p
r
o
n
e
t
o
d
r
o
p
o
u
t
s
a
n
d
m
a
y
o
f
f
e
r
l
i
m
i
t
e
d
c
o
v
e
r
a
g
e
.
T
h
i
s
m
a
y
o
c
c
a
s
i
o
n
a
l
l
y
l
e
a
v
e
s
t
h
e
u
s
e
r
w
i
t
h
o
u
t
n
e
t
w
o
r
k
a
c
c
e
s
s
,
a
n
d
f
o
r
c
e
t
h
e
b
r
o
w
s
e
r
t
o
o
f
f
e
r
t
h
o
s
e
d
o
c
u
m
e
n
t
s
w
h
i
c
h
a
r
e
i
n
t
h
e
c
a
c
h
e
.
S
o
m
e
u
s
e
r
s
p
l
a
n
a
h
e
a
d
,
a
n
d
w
e
b
-
c
o
p
y
d
o
c
u
m
e
n
t
s
f
r
o
m
t
h
e
i
r
o
r
g
a
n
i
z
a
t
i
o
n
’
s
s
e
r
v
e
r
t
o
t
h
e
i
r
m
o
b
i
l
e
m
a
c
h
i
n
e
.
A
c
o
m
m
o
n
s
o
l
u
t
i
o
n
i
n
v
o
l
v
e
s
r
e
p
l
i
c
a
t
i
n
g
t
h
e
n
e
c
e
s
s
a
r
y
f
i
l
e
s
a
l
o
n
g
w
i
t
h
t
h
e
i
r
s
u
b
-
d
i
r
e
c
t
o
r
i
e
s
.
T
h
i
s
p
r
o
c
e
d
u
r
e
o
f
t
e
n
l
e
a
v
e
s
t
h
e
u
s
e
r
w
i
t
h
a
s
i
g
n
i
f
i
c
a
n
t
n
u
m
b
e
r
o
f
f
i
l
e
s
s
c
a
t
t
e
r
e
d
o
n
l
o
c
a
l
s
t
o
r
a
g
e
.
T
o
a
c
h
i
e
v
e
t
h
i
s
,
t
h
e
u
s
e
r
h
a
s
s
e
v
e
r
a
l
w
a
y
s
t
o
r
e
p
l
i
c
a
t
e
a
s
e
r
v
e
r
’
s
d
i
r
e
c
t
o
r
y
t
r
e
e
I
n
t
e
r
a
c
t
i
v
e
l
y
f
o
l
l
o
w
e
v
e
r
y
l
i
n
k
w
h
i
l
e
s
a
v
i
n
g
t
h
e
m
t
o
l
o
c
a
l
s
t
o
r
a
g
e
(
f
i
l
e
n
a
m
e
s
n
e
e
d
t
o
b
e
k
n
o
w
n
i
n
a
d
v
a
n
c
e
)
;
U
s
e
a
n
a
u
t
o
m
a
t
e
d
w
e
b
-
c
o
p
y
t
o
o
l
t
o
d
u
p
l
i
c
a
t
e
t
h
e
r
e
q
u
i
r
e
d
w
e
b
t
r
e
e
;
F
T
P
r
e
q
u
i
r
e
d
d
o
c
u
m
e
n
t
s
t
o
l
o
c
a
l
s
t
o
r
a
g
e
(
a
c
c
e
s
s
p
e
n
d
i
n
g
)
.
H
T
M
L
’
s
o
p
e
n
a
r
c
h
i
t
e
c
t
u
r
e
a
n
d
d
o
c
u
m
e
n
t
b
u
i
l
d
i
n
g
f
r
o
m
d
i
s
t
r
i
b
u
t
e
d
f
i
l
e
s
i
s
f
l
e
x
i
b
l
e
b
u
t
d
o
e
s
n
o
t
e
x
p
l
i
c
i
t
l
y
s
t
a
t
e
f
i
l
e
r
e
l
a
t
i
o
n
s
h
i
p
s
u
n
t
i
l
t
h
e
y
a
r
e
c
l
o
s
e
l
y
e
x
a
m
i
n
e
d
o
r
v
i
e
w
e
d
.
A
m
i
s
s
i
n
g
r
o
o
t
d
o
c
u
m
e
n
t
m
a
y
i
n
t
r
o
d
u
c
e
a
s
e
v
e
r
e
u
s
e
r
i
n
c
o
n
v
e
n
i
e
n
c
e
i
f
n
o
t
d
u
p
l
i
c
a
t
e
d
.
I
n
t
h
i
s
p
a
p
e
r
w
e
p
r
o
p
o
s
e
a
n
e
w
a
p
p
r
o
a
c
h
t
o
w
e
b
d
o
c
u
m
e
n
t
m
a
n
a
g
e
m
e
n
t
.
W
e
i
n
t
r
o
d
u
c
e
t
h
e
c
o
n
c
e
p
t
o
f
a
d
o
c
u
m
e
n
t
w
r
a
p
p
e
r
w
h
i
c
h
p
a
c
k
a
g
e
s
e
n
t
i
r
e
w
e
b
s
e
r
v
e
r
s
s
i
t
e
s
o
r
p
a
r
t
s
o
f
t
h
e
i
r
d
i
r
e
c
t
o
r
y
t
r
e
e
.
P
a
c
k
a
g
e
d
d
o
c
u
m
e
n
t
s
a
r
e
p
l
a
c
e
d
i
n
a
s
i
n
g
l
e
f
i
l
e
w
h
i
l
e
p
e
r
m
i
t
t
i
n
g
l
i
v
e
a
c
c
e
s
s
t
o
t
h
e
a
r
c
h
i
v
e
d
f
i
l
e
s
.
T
h
e
f
o
l
l
o
w
i
n
g
s
e
c
t
i
o
n
s
p
r
e
s
e
n
t
s
p
o
s
s
i
b
l
e
s
o
l
u
t
i
o
n
s
t
o
t
h
e
r
e
p
l
i
c
a
t
i
o
n
p
r
o
b
l
e
m
a
n
d
a
n
a
l
t
e
r
n
a
t
i
v
e
t
o
d
i
s
c
o
n
n
e
c
t
e
d
o
p
e
r
a
t
i
o
n
s
.
I
n
l
a
t
e
r
s
e
c
t
i
o
n
s
w
e
d
i
s
c
u
s
s
o
u
r
p
r
o
j
e
c
t
i
m
p
l
e
m
e
n
t
a
t
i
o
n
,
f
u
t
u
r
e
e
x
t
e
n
s
i
o
n
s
a
n
d
r
e
l
a
t
e
d
w
o
r
k
.
T
h
e
d
i
s
c
o
n
n
e
c
t
e
d
o
p
e
r
a
t
i
o
n
s
p
r
o
b
l
e
m
h
a
s
a
l
r
e
a
d
y
b
e
e
n
a
d
d
r
e
s
s
e
d
i
n
s
o
m
e
d
i
s
t
r
i
b
u
t
e
d
f
i
l
e
s
y
s
t
e
m
s
.
D
i
s
c
o
n
n
e
c
t
i
o
n
p
r
o
b
l
e
m
s
a
s
s
o
c
i
a
t
e
d
w
i
t
h
S
u
n
’
s
N
F
S
,
t
h
e
A
n
d
r
e
w
F
i
l
e
S
y
s
t
e
m
[
1
]
a
n
d
F
i
c
u
s
[
6
]
l
e
a
d
t
o
t
h
e
d
e
v
e
l
o
p
m
e
n
t
o
f
t
h
e
C
o
d
a
f
i
l
e
s
y
s
t
e
m
a
t
C
a
r
n
e
g
i
e
M
e
l
l
o
n
U
n
i
v
e
r
s
i
t
y
[
3
]
.
F
i
c
u
s
i
s
a
d
i
s
t
r
i
b
u
t
e
d
U
N
I
X
f
i
l
e
s
y
s
t
e
m
d
e
v
e
l
o
p
e
d
a
t
t
h
e
U
n
i
v
e
r
s
i
t
y
o
f
C
a
l
i
f
o
r
n
i
a
,
L
o
s
A
n
g
e
l
e
s
.
I
t
e
m
p
l
o
y
s
a
p
e
e
r
-
t
o
-
p
e
e
r
f
i
l
e
s
h
a
r
i
n
g
m
o
d
e
l
a
n
d
o
f
f
e
r
s
m
u
l
t
i
p
l
e
f
i
l
e
r
e
p
l
i
c
a
t
i
o
n
.
I
n
c
a
s
e
s
o
f
a
n
o
d
e
d
i
s
c
o
n
n
e
c
t
i
o
n
,
t
h
e
c
l
o
s
e
s
t
c
o
p
y
o
f
t
h
e
r
e
q
u
e
s
t
e
d
f
i
l
e
i
s
u
s
e
d
.
F
i
l
e
m
o
d
i
f
i
c
a
t
i
o
n
s
t
o
a
l
o
c
a
l
c
o
p
y
a
r
e
a
d
v
e
r
t
i
s
e
d
t
o
a
l
l
o
t
h
e
r
n
o
d
e
s
o
n
c
e
c
o
m
m
u
n
i
c
a
t
i
o
n
i
s
r
e
-
4. Document Archiving and Storage
2.1 Using Local Replicas
4.1 Web Document Wrapper
3. Link Related Issues
e
s
t
a
b
l
i
s
h
e
d
.
F
i
c
u
s
d
i
f
f
e
r
s
f
r
o
m
C
o
d
a
i
n
t
h
a
t
i
t
i
s
a
p
e
e
r
-
t
o
-
p
e
e
r
s
y
s
t
e
m
a
s
o
p
p
o
s
e
d
t
o
a
c
l
i
e
n
t
-
s
e
r
v
e
r
e
n
v
i
r
o
n
m
e
n
t
.
C
o
d
a
a
l
l
o
w
s
f
o
r
d
i
s
c
o
n
n
e
c
t
e
d
o
p
e
r
a
t
i
o
n
b
y
m
a
i
n
t
a
i
n
i
n
g
a
s
m
a
l
l
(
a
p
p
r
o
x
i
m
a
t
e
l
y
5
0
M
b
)
s
t
o
r
a
g
e
c
a
c
h
e
o
n
e
a
c
h
w
o
r
k
s
t
a
t
i
o
n
.
D
u
r
i
n
g
d
i
s
c
o
n
n
e
c
t
e
d
p
e
r
i
o
d
s
,
C
o
d
a
w
o
u
l
d
e
n
t
e
r
e
m
u
l
a
t
i
o
n
m
o
d
e
a
n
d
l
o
g
l
o
c
a
l
f
i
l
e
a
c
c
e
s
s
.
U
p
o
n
r
e
c
o
n
n
e
c
t
i
o
n
,
C
o
d
a
w
o
u
l
d
e
n
t
e
r
r
e
i
n
t
e
g
r
a
t
i
o
n
m
o
d
e
,
a
n
d
r
e
s
o
l
v
e
c
a
c
h
e
d
f
i
l
e
m
o
d
i
f
i
c
a
t
i
o
n
s
w
i
t
h
t
h
e
p
r
i
m
a
r
y
c
o
p
y
b
y
p
l
a
y
i
n
g
t
h
o
u
g
h
t
h
e
l
o
g
.
C
o
d
a
i
s
a
n
a
w
k
w
a
r
d
s
c
h
e
m
e
t
o
i
m
p
l
e
m
e
n
t
a
s
i
t
r
e
q
u
i
r
e
s
s
i
g
n
i
f
i
c
a
n
t
m
o
d
i
f
i
c
a
t
i
o
n
s
t
o
a
n
e
x
i
s
t
i
n
g
f
i
l
e
s
y
s
t
e
m
.
A
d
d
i
t
i
o
n
a
l
l
y
,
i
m
p
l
e
m
e
n
t
a
t
i
o
n
o
f
a
C
o
d
a
l
i
k
e
f
i
l
e
s
y
s
t
e
m
c
a
c
h
i
n
g
i
n
t
o
e
v
e
r
y
c
l
i
e
n
t
d
e
v
i
c
e
m
a
y
n
o
t
a
l
w
a
y
s
b
e
p
o
s
s
i
b
l
e
o
r
w
o
r
t
h
w
h
i
l
e
.
A
s
i
m
i
l
a
r
f
i
l
e
s
y
s
t
e
m
b
a
s
e
d
o
n
C
o
d
a
,
i
s
t
h
e
H
y
b
r
i
d
F
i
l
e
s
y
s
t
e
m
f
o
r
m
o
b
i
l
e
w
o
r
k
s
t
a
t
i
o
n
s
[
7
]
.
D
e
v
e
l
o
p
e
d
a
t
t
h
e
U
n
i
v
e
r
s
i
t
y
o
f
N
o
t
r
e
D
a
m
e
.
I
t
c
o
m
b
i
n
e
s
i
d
e
a
s
f
r
o
m
F
i
c
u
s
a
n
d
C
o
d
a
,
b
y
a
l
l
o
w
i
n
g
p
e
e
r
-
t
o
-
p
e
e
r
f
i
l
e
s
h
a
r
i
n
g
a
n
d
C
o
d
a
l
i
k
e
o
p
e
r
a
t
i
o
n
w
h
i
l
e
d
i
s
c
o
n
n
e
c
t
e
d
f
r
o
m
a
f
i
l
e
s
e
r
v
e
r
.
I
t
i
s
a
s
u
p
e
r
s
e
t
t
o
C
o
d
a
,
a
s
i
t
e
m
p
l
o
y
s
C
o
d
a
’
s
c
o
d
e
b
a
s
e
a
n
d
e
x
t
e
n
d
s
i
t
t
o
p
r
o
v
i
d
e
p
e
e
r
s
h
a
r
i
n
g
.
E
s
s
e
n
t
i
a
l
l
y
,
o
u
r
n
e
e
d
s
f
o
r
d
i
s
c
o
n
n
e
c
t
e
d
o
p
e
r
a
t
i
o
n
s
a
r
e
a
d
d
r
e
s
s
e
d
b
y
C
o
d
a
a
l
o
n
e
.
U
s
i
n
g
C
o
d
a
m
a
y
a
p
p
e
a
r
a
s
a
n
e
x
c
e
s
s
i
v
e
a
p
p
r
o
a
c
h
.
I
t
r
e
q
u
i
r
e
s
a
s
h
a
r
e
d
f
i
l
e
s
y
s
t
e
m
w
h
i
c
h
d
i
f
f
e
r
s
f
r
o
m
t
h
e
o
p
e
r
a
t
i
o
n
s
o
f
t
h
e
W
e
b
.
W
e
p
r
o
p
o
s
e
t
h
i
s
a
p
p
r
o
a
c
h
m
a
i
n
l
y
b
e
c
a
u
s
e
b
r
o
w
s
e
r
s
a
r
e
c
a
p
a
b
l
e
o
f
o
p
e
n
i
n
g
l
o
c
a
l
l
y
s
t
o
r
e
d
f
i
l
e
s
w
h
i
c
h
m
a
y
e
x
i
s
t
o
n
r
e
m
o
t
e
m
o
u
n
t
s
.
T
h
i
s
t
e
c
h
n
i
q
u
e
e
x
t
e
n
d
s
b
e
y
o
n
d
t
h
e
b
r
o
w
s
e
r
’
s
o
p
e
r
a
t
i
o
n
a
n
d
a
l
l
o
w
s
o
t
h
e
r
a
p
p
l
i
c
a
t
i
o
n
s
t
o
u
s
e
t
h
e
s
e
f
a
c
i
l
i
t
i
e
s
w
h
i
l
e
d
i
s
c
o
n
n
e
c
t
e
d
.
F
i
l
e
s
l
o
c
a
t
e
d
o
n
t
h
e
w
e
b
s
e
r
v
e
r
w
o
u
l
d
a
p
p
e
a
r
a
s
l
o
c
a
l
t
h
e
r
e
f
o
r
e
p
e
r
m
i
t
t
i
n
g
a
p
p
l
i
c
a
t
i
o
n
s
t
o
l
o
a
d
s
e
l
e
c
t
e
d
d
o
c
u
m
e
n
t
s
.
C
u
r
r
e
n
t
l
y
c
o
m
p
a
n
i
e
s
a
n
d
s
o
m
e
u
n
i
v
e
r
s
i
t
i
e
s
d
e
l
i
v
e
r
W
e
b
d
o
c
u
m
e
n
t
s
v
i
a
a
C
D
w
h
i
c
h
i
s
a
m
u
c
h
s
i
m
p
l
e
r
a
p
p
r
o
a
c
h
t
h
e
n
d
i
s
t
r
i
b
u
t
e
d
f
i
l
e
s
y
s
t
e
m
s
.
T
h
e
f
i
r
s
t
g
e
n
e
r
a
t
i
o
n
o
f
C
D
s
p
r
o
v
i
d
e
d
d
o
c
u
m
e
n
t
a
t
i
o
n
w
i
t
h
a
w
e
b
l
i
k
e
i
n
t
e
r
f
a
c
e
w
h
i
l
e
t
h
e
c
u
r
r
e
n
t
r
e
l
e
a
s
e
s
a
r
e
h
y
b
r
i
d
s
.
T
h
e
s
e
a
l
l
o
w
l
o
c
a
l
b
r
o
w
s
i
n
g
w
h
i
l
e
m
a
i
n
t
a
i
n
i
n
g
a
b
s
o
l
u
t
e
U
R
L
l
i
n
k
s
t
o
f
r
e
q
u
e
n
t
l
y
c
h
a
n
g
i
n
g
i
n
f
o
r
m
a
t
i
o
n
.
T
h
e
s
e
s
o
l
u
t
i
o
n
s
a
r
e
o
f
t
e
n
i
d
e
a
l
b
u
t
r
e
f
l
e
c
t
a
s
s
o
c
i
a
t
e
d
c
o
s
t
s
o
f
m
a
n
u
f
a
c
t
u
r
i
n
g
.
A
n
o
t
h
e
r
a
l
t
e
r
n
a
t
i
v
e
i
s
t
o
p
r
o
v
i
d
e
l
o
c
a
l
d
o
c
u
m
e
n
t
c
o
p
i
e
s
i
n
v
o
l
a
t
i
l
e
s
t
o
r
a
g
e
.
T
h
e
s
e
a
r
e
c
h
e
a
p
e
r
t
h
a
n
C
D
-
R
O
M
m
a
s
t
e
r
i
n
g
f
o
r
m
o
d
e
s
t
s
i
z
e
d
d
o
c
u
m
e
n
t
v
o
l
u
m
e
s
.
F
o
r
s
m
a
l
l
s
i
z
e
d
r
e
p
l
i
c
a
s
,
t
h
e
i
r
i
m
p
a
c
t
o
n
l
o
c
a
l
s
t
o
r
a
g
e
m
a
y
n
o
t
b
e
s
i
g
n
i
f
i
c
a
n
t
.
W
h
i
l
e
l
a
r
g
e
d
o
c
u
m
e
n
t
r
e
p
l
i
c
a
s
m
a
y
s
u
b
s
t
a
n
t
i
a
l
l
y
i
m
p
a
c
t
a
f
i
l
e
s
y
s
t
e
m
.
S
i
n
c
e
w
e
b
d
o
c
u
m
e
n
t
s
a
r
e
c
o
m
p
o
s
e
d
o
f
s
e
v
e
r
a
l
s
m
a
l
l
f
i
l
e
s
,
t
h
e
s
e
c
o
u
l
d
p
o
t
e
n
t
i
a
l
l
y
c
o
n
s
u
m
e
m
o
r
e
d
i
s
k
s
p
a
c
e
t
h
e
n
r
e
q
u
i
r
e
d
b
e
c
a
u
s
e
o
f
i
-
n
o
d
e
o
r
c
l
u
s
t
e
r
s
i
z
e
s
.
M
o
s
t
l
i
n
k
s
w
i
t
h
i
n
d
o
c
u
m
e
n
t
s
a
r
e
r
e
l
a
t
i
v
e
,
m
e
a
n
i
n
g
t
h
a
t
i
m
a
g
e
s
e
m
b
e
d
d
e
d
i
n
t
h
e
d
o
c
u
m
e
n
t
a
r
e
l
o
c
a
t
e
d
r
e
l
a
t
i
v
e
t
o
t
h
e
H
T
M
L
’
s
l
o
c
a
t
i
o
n
.
D
u
p
l
i
c
a
t
i
n
g
t
h
e
s
e
d
o
c
u
m
e
n
t
s
d
o
n
o
t
p
r
e
s
e
n
t
a
p
r
o
b
l
e
m
,
p
r
o
v
i
d
e
d
t
h
e
d
i
r
e
c
t
o
r
y
s
t
r
u
c
t
u
r
e
s
a
r
e
m
a
i
n
t
a
i
n
e
d
.
W
h
i
l
e
a
b
s
o
l
u
t
e
l
i
n
k
a
d
d
r
e
s
s
i
n
g
i
s
u
s
e
d
t
o
p
o
i
n
t
a
b
r
o
w
s
e
r
a
t
a
r
e
m
o
t
e
s
i
t
e
a
s
i
t
s
p
e
c
i
f
i
e
s
a
w
h
o
l
e
U
R
L
.
M
a
n
y
w
e
b
a
u
t
h
o
r
s
u
s
e
a
b
s
o
l
u
t
e
l
i
n
k
s
t
o
t
h
e
i
r
o
w
n
s
i
t
e
a
n
d
u
n
k
n
o
w
i
n
g
l
y
c
r
e
a
t
e
d
o
c
u
m
e
n
t
s
w
h
i
c
h
a
r
e
d
i
f
f
i
c
u
l
t
t
o
c
o
p
y
.
S
u
c
h
a
b
s
o
l
u
t
e
l
i
n
k
s
p
r
e
s
e
n
t
a
p
r
o
b
l
e
m
f
o
r
s
i
m
p
l
e
d
o
c
u
m
e
n
t
r
e
p
l
i
c
a
t
i
o
n
.
T
h
e
y
a
l
s
o
f
o
r
c
e
t
h
e
b
r
o
w
s
e
r
t
o
o
b
t
a
i
n
d
o
c
u
m
e
n
t
s
v
i
a
H
T
T
P
w
h
i
c
h
b
r
e
a
k
s
o
p
e
r
a
t
i
o
n
s
w
h
i
l
e
o
p
e
n
i
n
g
l
o
c
a
l
f
i
l
e
s
.
I
t
s
h
o
u
l
d
b
e
n
o
t
e
d
t
h
a
t
m
o
s
t
w
e
b
-
c
o
p
y
t
o
o
l
s
,
a
r
e
i
n
t
e
l
l
i
g
e
n
t
a
n
d
r
e
p
l
a
c
e
a
b
s
o
l
u
t
e
l
i
n
k
s
w
i
t
h
r
e
l
a
t
i
v
e
l
i
n
k
s
r
e
f
l
e
c
t
i
n
g
t
h
e
d
o
c
u
m
e
n
t
s
n
e
w
l
o
c
a
t
i
o
n
.
S
u
c
h
t
o
o
l
s
s
i
g
n
i
f
i
c
a
n
t
l
y
a
i
d
i
n
w
e
b
a
u
t
h
o
r
i
n
g
a
n
d
r
e
p
l
i
c
a
t
i
o
n
.
A
n
o
t
h
e
r
a
l
t
e
r
n
a
t
i
v
e
i
s
t
o
a
d
o
p
t
a
n
a
u
t
h
o
r
i
n
g
s
t
a
n
d
a
r
d
f
o
r
a
s
i
t
e
w
h
e
r
e
a
l
l
d
o
c
u
m
e
n
t
s
u
s
e
r
e
l
a
t
i
v
e
l
i
n
k
s
.
M
a
n
y
u
s
e
r
s
a
r
e
f
a
m
i
l
i
a
r
w
i
t
h
a
r
c
h
i
v
i
n
g
u
t
i
l
i
t
i
e
s
s
u
c
h
a
s
p
k
z
i
p
a
n
d
t
a
r
.
T
h
e
s
e
c
o
l
l
e
c
t
f
i
l
e
s
a
n
d
p
u
t
t
h
e
m
i
n
t
o
a
s
i
n
g
l
e
p
a
c
k
a
g
e
.
C
u
r
r
e
n
t
l
y
t
h
e
r
e
a
r
e
n
o
t
o
o
l
s
w
h
i
c
h
a
l
l
o
w
w
e
b
d
o
c
u
m
e
n
t
a
r
c
h
i
v
a
l
a
n
d
e
a
s
y
r
e
t
r
i
e
v
a
l
.
I
t
i
s
p
o
s
s
i
b
l
e
t
o
u
s
e
t
h
e
a
b
o
v
e
m
e
n
t
i
o
n
e
d
t
o
o
l
s
b
u
t
u
n
t
i
l
n
o
w
,
s
e
r
v
e
r
s
d
i
d
n
o
t
h
a
v
e
h
o
o
k
s
i
n
t
o
t
h
e
s
e
p
r
o
g
r
a
m
s
t
o
p
r
o
v
i
d
e
s
e
a
m
l
e
s
s
i
n
t
e
g
r
a
t
i
o
n
.
T
h
r
o
u
g
h
t
h
e
u
s
a
g
e
o
f
a
p
a
c
k
a
g
i
n
g
t
o
o
l
i
t
w
o
u
l
d
b
e
p
o
s
s
i
b
l
e
t
o
c
o
m
b
i
n
e
a
l
l
t
h
e
n
e
c
e
s
s
a
r
y
w
e
b
d
o
c
u
m
e
n
t
s
i
n
t
o
a
s
i
n
g
l
e
f
i
l
e
.
T
h
i
s
w
o
u
l
d
m
a
k
e
d
o
c
u
m
e
n
t
d
i
s
t
r
i
b
u
t
i
o
n
a
n
d
r
e
p
l
i
c
a
t
i
o
n
m
u
c
h
e
a
s
i
e
r
f
o
r
a
u
s
e
r
a
n
d
p
e
r
m
i
t
a
d
m
i
n
i
s
t
r
a
t
o
r
s
t
o
a
r
c
h
i
v
e
a
g
i
n
g
d
o
c
u
m
e
n
t
s
.
W
e
p
r
o
p
o
s
e
t
h
e
p
a
c
k
a
g
i
n
g
a
p
p
r
o
a
c
h
a
s
a
s
o
l
u
t
i
o
n
t
o
s
i
m
p
l
i
f
y
w
e
b
d
o
c
u
m
e
n
t
r
e
p
l
i
c
a
t
i
o
n
,
m
i
n
i
m
i
z
e
t
h
e
i
m
p
a
c
t
o
n
a
l
o
c
a
l
f
i
l
e
s
y
s
t
e
m
a
n
d
a
s
s
i
s
t
i
n
s
i
t
e
a
r
c
h
i
v
a
l
w
h
i
l
e
m
a
i
n
t
a
i
n
i
n
g
i
t
s
w
e
b
d
o
c
u
m
e
n
t
p
r
e
s
e
n
c
e
.
T
h
e
w
e
b
d
o
c
u
m
e
n
t
w
r
a
p
p
e
r
n
a
m
e
d
t
h
e
P
o
r
t
a
b
l
e
F
i
l
e
S
y
s
t
e
m
(
P
F
S
)
,
h
a
s
b
e
e
n
d
e
v
e
l
o
p
e
d
a
s
a
p
a
r
t
o
f
t
h
e
o
n
g
o
i
n
g
P
E
S
O
S
p
r
o
j
e
c
t
a
t
M
o
n
a
s
h
U
n
i
v
e
r
s
i
t
y
[
8
]
.
P
F
S
o
r
i
g
i
n
a
t
e
d
a
s
a
c
o
n
t
a
i
n
e
r
f
o
r
c
o
l
l
e
c
t
i
n
g
m
o
b
i
l
e
a
g
e
n
t
c
o
m
p
o
n
e
n
t
f
i
l
e
s
i
n
t
o
a
s
i
n
g
l
e
m
i
g
r
a
t
i
o
n
a
r
c
h
i
v
e
.
C
o
m
b
i
n
i
n
g
c
o
m
p
o
n
e
n
t
i
n
t
o
o
n
e
f
i
l
e
r
e
d
u
c
e
d
m
i
g
r
a
t
i
o
n
c
o
m
p
l
e
x
i
t
y
a
n
d
a
l
l
o
w
e
d
f
o
r
g
r
e
a
t
e
r
m
a
n
a
g
e
a
b
i
l
i
t
y
.
I
n
t
h
e
c
a
s
e
o
f
H
T
M
L
d
o
c
u
m
e
n
t
s
,
t
h
e
P
F
S
c
o
n
t
e
n
t
s
a
r
e
t
h
o
s
e
o
f
w
e
b
d
o
c
u
m
e
n
t
s
d
u
p
l
i
c
a
t
e
d
f
r
o
m
a
n
e
x
i
s
t
i
n
g
w
e
b
s
e
r
v
e
r
.
F
i
g
u
r
e
1
b
e
l
o
w
,
s
h
o
w
s
t
h
e
P
F
S
c
o
n
t
e
n
t
s
f
o
r
a
s
i
m
p
l
e
d
i
r
e
c
t
o
r
y
t
r
e
e
.
T
h
r
o
u
g
h
t
h
e
u
s
e
o
f
P
F
S
w
r
a
p
p
e
r
s
,
w
h
o
l
e
o
r
p
a
r
t
i
a
l
d
i
r
e
c
t
o
r
y
t
r
e
e
s
m
a
y
b
e
a
r
c
h
i
v
e
d
,
b
e
c
o
m
e
o
r
g
a
n
i
z
e
d
b
y
c
o
l
l
a
p
s
i
n
g
s
u
b
d
i
r
e
c
t
o
r
i
e
s
w
h
i
l
e
m
a
i
n
t
a
i
n
i
n
g
t
h
e
i
r
c
o
n
t
e
n
t
s
i
n
t
a
c
t
.
I
n
t
h
e
c
a
s
e
o
f
F
i
g
u
r
e
1
,
a
s
i
n
g
l
e
P
F
S
f
i
l
e
c
o
u
l
d
r
e
p
l
a
c
e
a
l
l
e
x
i
s
t
i
n
g
f
i
l
e
s
i
n
a
g
i
v
e
n
A
c
c
o
u
n
t
s
d
i
r
e
c
t
o
r
y
a
l
o
n
g
w
i
t
h
I
m
a
g
e
s
,
D
e
p
t
1
a
n
d
D
e
p
t
2
s
u
b
d
i
r
e
c
t
o
r
i
e
s
.
T
h
i
s
e
x
a
m
p
l
e
r
e
f
l
e
c
t
s
t
h
e
e
m
b
e
d
d
a
b
l
e
a
n
d
r
e
m
o
t
e
r
e
f
e
r
e
n
c
i
n
g
f
a
c
i
l
i
t
i
e
s
o
f
f
e
r
e
d
b
y
a
P
F
S
w
r
a
p
p
e
r
.
T
r
a
d
i
t
i
o
n
a
l
f
i
l
e
a
r
c
h
i
v
i
n
g
t
o
o
l
s
h
a
v
e
r
e
l
i
e
d
o
n
e
m
b
e
d
d
i
n
g
a
l
l
f
i
l
e
s
w
i
t
h
i
n
t
h
e
a
r
c
h
i
v
e
f
i
l
e
i
t
s
e
l
f
.
O
u
r
e
x
t
e
n
d
e
d
a
p
p
r
o
a
c
h
t
o
a
r
c
h
i
v
i
n
g
a
l
l
o
w
s
f
o
r
d
o
c
u
m
e
n
t
s
t
o
b
e
r
e
m
o
t
e
l
y
r
e
f
e
r
e
n
c
e
d
u
s
i
n
g
s
t
a
n
d
a
r
d
I
n
t
e
r
n
e
t
p
r
o
t
o
c
o
l
s
.
F
i
g
u
r
e
1
d
e
p
i
c
t
s
t
w
o
i
m
a
g
e
f
i
l
e
s
a
s
b
e
i
n
g
s
t
o
r
e
d
a
t
t
h
e
A
B
C
.
C
O
M
s
i
t
e
,
w
h
i
c
h
m
a
y
b
e
o
b
t
a
i
n
e
d
u
s
i
n
g
H
T
T
P
o
r
F
T
P
p
r
o
t
o
c
o
l
s
.
F
o
r
m
a
j
o
r
i
t
y
o
f
u
s
e
r
s
r
e
m
o
t
e
f
i
l
e
r
e
f
e
r
e
n
c
i
n
g
w
i
l
l
n
o
t
b
e
n
e
c
e
s
s
a
r
y
a
s
t
h
e
y
s
h
a
l
l
r
e
q
u
i
r
e
r
e
m
o
t
e
f
i
l
e
s
t
o
a
l
w
a
y
s
b
e
a
v
a
i
l
a
b
l
e
.
I
n
c
a
s
e
s
w
h
e
r
e
a
u
s
e
r
w
i
s
h
e
s
t
o
s
a
c
r
i
f
i
c
e
c
o
n
t
e
n
t
/
Accounts/
ACCOUNT-SITE.PFS
Public/
Other/
index.html
/IMAGES
logo.gif
graph.jpg
/Dept1
……….
/Dept2
……….
Figure 1: Directory structure within a PFS wrapper
4.1 Internal PFS Structure
5. Accessing PFS Contents
get
Figure 2: Format of a GET URL with a PFS reference
P F S f i l e n a m e
p a t h / f i l e n a m e
“index.html”
“account-site.pfs”
5.1 Obtaining PFS Packages
Figure 3: URL to obtain the PFS wrapper file
P F S f i l e n a m e
6. Internal PFS Access and Implementation
As an Archive:
As a File System:
Within a Proxy:
f
o
r
a
v
a
i
l
a
b
i
l
i
t
y
a
n
d
s
i
z
e
,
i
t
w
o
u
l
d
b
e
p
o
s
s
i
b
l
e
f
o
r
t
h
e
a
r
c
h
i
v
e
t
o
c
o
n
t
a
i
n
e
m
b
e
d
d
e
d
t
e
x
t
u
a
l
i
n
f
o
r
m
a
t
i
o
n
w
h
i
l
e
a
l
l
o
p
t
i
o
n
a
l
f
i
l
e
s
(
i
m
a
g
e
s
)
w
o
u
l
d
b
e
r
e
m
o
t
e
l
y
t
a
g
g
e
d
w
i
t
h
i
n
t
h
e
P
F
S
.
I
n
t
h
i
s
s
c
e
n
a
r
i
o
,
u
s
e
r
s
e
x
p
e
r
i
e
n
c
i
n
g
n
e
t
w
o
r
k
d
r
o
p
o
u
t
s
w
o
u
l
d
o
n
l
y
a
c
c
e
s
s
l
o
c
a
l
l
y
a
v
a
i
l
a
b
l
e
H
T
M
L
d
o
c
u
m
e
n
t
s
w
i
t
h
o
u
t
o
b
t
a
i
n
i
n
g
t
h
e
r
e
m
o
t
e
g
r
a
p
h
i
c
a
l
c
o
n
t
e
n
t
.
A
B
C
.
C
O
M
H
T
T
P
F
T
P
T
o
m
a
i
n
t
a
i
n
c
o
m
p
a
t
i
b
i
l
i
t
y
w
i
t
h
f
u
t
u
r
e
P
F
S
r
e
v
i
s
i
o
n
s
,
t
h
e
i
n
t
e
r
n
a
l
s
t
r
u
c
t
u
r
e
h
a
s
b
e
e
n
k
e
p
t
o
p
e
n
.
T
h
i
s
p
e
r
m
i
t
s
f
u
t
u
r
e
w
r
a
p
p
e
r
i
m
p
r
o
v
e
m
e
n
t
s
w
h
i
l
e
m
a
i
n
t
a
i
n
i
n
g
b
a
c
k
w
a
r
d
c
o
m
p
a
t
i
b
i
l
i
t
y
.
T
o
a
c
h
i
e
v
e
o
p
e
n
s
t
r
u
c
t
u
r
e
t
h
e
f
i
l
e
d
e
f
i
n
i
t
i
o
n
s
a
n
d
s
t
o
r
a
g
e
a
r
e
a
h
y
b
r
i
d
A
S
C
I
I
-
B
i
n
a
r
y
f
o
r
m
a
t
.
A
l
l
e
m
b
e
d
d
e
d
f
i
l
e
s
m
a
y
b
e
s
t
o
r
e
d
a
s
u
u
e
n
c
o
d
e
d
/
r
a
w
-
b
i
n
a
r
y
c
o
n
t
e
n
t
s
,
w
h
i
l
e
f
i
l
e
d
e
s
c
r
i
p
t
i
o
n
s
a
r
e
d
e
s
c
r
i
b
e
d
i
n
p
l
a
i
n
t
e
x
t
.
T
e
x
t
u
a
l
d
e
s
c
r
i
p
t
i
o
n
s
a
l
l
o
w
e
n
c
o
d
i
n
g
/
d
e
c
o
d
i
n
g
c
o
d
e
t
o
i
n
t
e
r
p
r
e
t
t
a
g
s
w
h
i
c
h
h
a
v
e
s
e
m
a
n
t
i
c
s
r
e
l
e
v
a
n
t
t
o
a
s
o
f
t
w
a
r
e
r
e
v
i
s
i
o
n
.
A
s
a
m
p
l
e
P
F
S
d
e
s
c
r
i
p
t
i
o
n
f
o
r
a
r
e
m
o
t
e
l
y
t
a
g
g
e
d
f
i
l
e
i
s
g
i
v
e
n
b
e
l
o
w
a
l
o
n
g
w
i
t
h
t
h
e
w
r
a
p
p
e
r
h
e
a
d
e
r
:
P
F
S
!
v
e
r
s
i
o
n
=
1
.
0
d
a
t
e
=
2
5
-
0
6
-
9
7
.
.
.
.
.
.
[
E
N
T
I
T
Y
]
o
r
i
g
i
n
a
l
n
a
m
e
=
C
:
\
P
r
o
g
r
a
m
F
i
l
e
s
\
W
I
N
Z
I
P
\
V
e
n
d
o
r
.
t
x
t
l
o
n
g
n
a
m
e
=
V
e
n
d
o
r
.
t
x
t
s
h
o
r
t
n
a
m
e
=
V
e
n
d
o
r
.
t
x
t
d
i
r
n
a
m
e
=
;
P
F
S
d
i
r
e
c
t
o
r
y
(
r
o
o
t
)
c
r
e
a
t
e
d
=
1
4
-
0
8
-
9
5
6
:
0
0
:
0
0
A
M
l
e
n
g
t
h
=
2
9
5
2
;
p
h
y
s
i
c
a
l
f
i
l
e
l
e
n
g
t
h
o
r
i
g
i
n
=
W
i
n
d
o
w
s
9
5
;
c
r
e
a
t
e
d
W
i
n
-
P
F
S
u
t
i
l
v
e
r
1
.
0
d
e
s
c
r
i
p
t
i
o
n
=
T
h
i
s
i
s
t
h
e
V
e
n
d
o
r
s
m
e
s
s
a
g
e
f
i
l
e
r
e
m
o
t
e
r
e
a
d
h
o
s
t
=
h
t
t
p
:
/
/
a
s
t
r
a
l
.
c
t
.
m
o
n
a
s
h
.
e
d
u
.
a
u
/
~
f
i
l
e
s
/
v
e
n
d
o
r
.
t
x
t
m
o
d
e
=
R
O
;
r
e
a
d
-
o
n
l
y
a
c
c
e
s
s
s
t
o
r
a
g
e
=
r
e
m
o
t
e
;
s
t
o
r
e
d
o
n
r
e
m
o
t
e
h
o
s
t
.
.
.
.
.
.
[
E
N
T
I
T
I
Y
]
I
n
o
r
d
e
r
t
o
m
a
k
e
u
s
e
o
f
P
F
S
w
r
a
p
p
e
r
s
a
n
d
d
i
r
e
c
t
o
r
y
c
o
l
l
a
p
s
i
n
g
,
s
m
a
l
l
c
h
a
n
g
e
s
h
a
v
e
t
o
b
e
i
n
t
r
o
d
u
c
e
d
t
o
t
h
e
w
a
y
w
e
b
s
e
r
v
e
r
s
r
e
s
o
l
v
e
a
U
R
L
s
p
e
c
i
f
y
i
n
g
a
P
F
S
p
a
t
h
.
A
H
T
T
P
c
l
i
e
n
t
i
s
s
u
i
n
g
a
r
e
q
u
e
s
t
,
s
p
e
c
i
f
i
e
s
t
h
e
f
i
l
e
t
o
b
e
t
r
a
n
s
f
e
r
r
e
d
.
I
n
t
h
e
c
a
s
e
o
f
F
i
g
u
r
e
2
,
t
h
e
U
R
L
i
n
c
o
r
p
o
r
a
t
e
s
a
H
T
M
L
f
i
l
e
t
o
b
e
f
o
u
n
d
w
i
t
h
i
n
a
P
F
S
.
U
n
d
e
r
n
o
r
m
a
l
c
i
r
c
u
m
s
t
a
n
c
e
s
,
a
s
e
r
v
e
r
i
s
r
e
q
u
i
r
e
d
t
o
e
s
t
a
b
l
i
s
h
t
h
e
p
h
y
s
i
c
a
l
p
a
t
h
f
r
o
m
t
h
e
U
R
L
.
I
n
t
h
i
s
c
a
s
e
i
t
w
o
u
l
d
r
e
s
u
l
t
i
n
a
“
4
0
4
E
r
r
o
r
:
n
o
t
f
o
u
n
d
”
w
h
e
n
p
r
o
c
e
s
s
e
d
b
y
a
s
t
a
n
d
a
r
d
s
e
r
v
e
r
.
T
h
i
s
i
s
a
s
s
u
m
i
n
g
t
h
a
t
a
d
i
r
e
c
t
o
r
y
d
o
e
s
n
o
t
e
x
i
s
t
w
i
t
h
a
n
i
d
e
n
t
i
c
a
l
n
a
m
e
a
s
t
h
e
P
F
S
.
h
t
t
p
:
/
/
b
r
o
n
c
h
o
.
c
t
.
m
o
n
a
s
h
.
e
d
u
.
a
u
/
~
p
s
t
a
n
s
k
i
/
a
c
c
o
u
n
t
-
s
i
t
e
.
p
f
s
/
i
n
d
e
x
.
h
t
m
l
O
u
r
p
r
o
t
o
t
y
p
e
s
e
r
v
e
r
p
a
r
s
e
s
t
h
e
U
R
L
a
n
d
e
s
t
a
b
l
i
s
h
e
s
i
f
t
h
e
r
e
q
u
e
s
t
e
d
f
i
l
e
i
s
a
p
a
r
t
o
f
a
P
F
S
p
a
c
k
a
g
e
.
I
n
t
h
e
c
a
s
e
o
f
t
h
e
a
b
o
v
e
U
R
L
,
w
o
u
l
d
b
e
s
e
a
r
c
h
e
d
i
n
s
i
d
e
t
h
e
f
i
l
e
.
I
f
f
o
u
n
d
,
t
h
e
f
i
l
e
w
o
u
l
d
b
e
s
e
n
t
t
o
t
h
e
c
l
i
e
n
t
,
o
r
f
a
i
l
w
i
t
h
a
n
e
r
r
o
r
s
u
c
h
a
s
f
i
l
e
n
o
t
f
o
u
n
d
.
T
h
e
p
a
r
s
i
n
g
m
o
d
i
f
i
c
a
t
i
o
n
s
m
a
k
e
o
u
r
i
m
p
l
e
m
e
n
t
a
t
i
o
n
t
r
a
n
s
p
a
r
e
n
t
t
o
t
h
e
c
l
i
e
n
t
u
s
e
r
.
I
n
o
r
d
e
r
t
o
o
b
t
a
i
n
P
F
S
p
a
c
k
a
g
e
s
w
h
i
c
h
c
o
n
t
a
i
n
a
w
e
b
s
e
r
v
e
r
’
s
d
i
r
e
c
t
o
r
y
t
r
e
e
,
a
c
l
i
e
n
t
m
a
y
r
e
q
u
e
s
t
t
h
e
w
r
a
p
p
e
r
f
i
l
e
a
s
s
p
e
c
i
f
i
e
d
i
n
F
i
g
u
r
e
3
.
T
h
e
a
p
p
r
o
p
r
i
a
t
e
M
I
M
E
t
y
p
e
i
s
s
e
t
t
o
a
l
l
o
w
f
o
r
a
b
i
n
a
r
y
f
i
l
e
t
r
a
n
s
f
e
r
t
o
t
h
e
c
l
i
e
n
t
’
s
b
r
o
w
s
e
r
.
h
t
t
p
:
/
/
b
r
o
n
c
h
o
.
c
t
.
m
o
n
a
s
h
.
e
d
u
.
a
u
/
~
p
s
t
a
n
s
k
i
/
a
c
c
o
u
n
t
-
s
i
t
e
.
p
f
s
P
r
i
o
r
t
o
o
u
r
P
F
S
s
e
r
v
e
r
e
x
t
e
n
s
i
o
n
s
,
u
s
e
r
s
h
a
d
t
w
o
w
a
y
s
o
f
o
b
t
a
i
n
i
n
g
w
e
b
d
o
c
u
m
e
n
t
s
v
i
a
H
T
T
P
.
T
h
e
y
c
o
u
l
d
d
o
w
n
l
o
a
d
e
a
c
h
p
a
g
e
a
n
d
i
t
s
e
m
b
e
d
d
e
d
i
m
a
g
e
s
b
y
h
a
n
d
,
o
r
u
s
e
a
n
a
u
t
o
m
a
t
e
d
t
o
o
l
w
h
i
c
h
w
o
u
l
d
d
o
i
t
f
o
r
t
h
e
m
.
P
F
S
n
o
w
a
l
l
o
w
s
f
o
r
o
b
t
a
i
n
i
n
g
a
r
c
h
i
v
e
d
d
o
c
u
m
e
n
t
s
a
n
d
t
o
v
i
e
w
t
h
e
m
s
e
a
m
l
e
s
s
l
y
.
A
c
c
e
s
s
w
i
t
h
i
n
a
P
F
S
p
a
c
k
a
g
e
m
a
y
b
e
i
m
p
l
e
m
e
n
t
e
d
i
n
a
v
a
r
i
e
t
y
o
f
w
a
y
s
.
T
h
e
f
o
l
l
o
w
i
n
g
b
r
i
e
f
l
y
d
e
s
c
r
i
b
e
s
s
o
m
e
p
o
s
s
i
b
l
e
a
l
t
e
r
n
a
t
i
v
e
s
:
T
h
e
p
a
c
k
a
g
e
c
o
u
l
d
b
e
v
i
e
w
e
d
a
n
d
h
a
v
e
t
h
e
d
o
c
u
m
e
n
t
s
e
x
t
r
a
c
t
e
d
b
y
a
P
F
S
m
a
n
a
g
e
m
e
n
t
u
t
i
l
i
t
y
(
p
f
s
u
t
i
l
)
.
T
h
i
s
i
s
i
d
e
n
t
i
c
a
l
t
o
a
c
c
e
s
s
i
n
g
a
t
a
r
o
r
z
i
p
a
r
c
h
i
v
e
.
A
r
e
c
e
i
v
e
d
P
F
S
c
o
u
l
d
b
e
e
x
t
r
a
c
t
e
d
w
i
t
h
d
i
r
e
c
t
o
r
y
e
n
t
r
i
e
s
,
a
n
d
b
r
o
w
s
e
d
l
o
c
a
l
l
y
.
F
i
l
e
s
t
a
g
g
e
d
a
s
r
e
m
o
t
e
a
n
d
m
a
i
n
t
a
i
n
e
d
o
n
a
c
e
n
t
r
a
l
s
e
r
v
e
r
,
c
o
u
l
d
s
e
a
m
l
e
s
s
l
y
b
e
c
o
m
e
a
v
a
i
l
a
b
l
e
f
o
r
u
s
e
r
s
a
s
s
o
o
n
a
s
t
h
e
y
b
e
c
a
m
e
a
v
a
i
l
a
b
l
e
o
n
t
h
e
s
e
r
v
e
r
.
T
h
r
o
u
g
h
t
h
e
d
e
v
e
l
o
p
m
e
n
t
o
f
a
n
i
n
s
t
a
l
l
a
b
l
e
f
i
l
e
s
y
s
t
e
m
,
a
P
F
S
f
i
l
e
c
o
u
l
d
b
e
m
o
u
n
t
e
d
a
s
a
v
o
l
u
m
e
i
n
a
U
N
I
X
e
n
v
i
r
o
n
m
e
n
t
.
I
n
t
h
e
c
a
s
e
o
f
D
O
S
/
W
i
n
d
o
w
s
,
a
R
A
M
d
r
i
v
e
c
o
u
l
d
p
r
o
v
i
d
e
a
c
c
e
s
s
.
T
h
i
s
i
s
p
e
r
h
a
p
s
t
h
e
m
o
s
t
f
l
e
x
i
b
l
e
a
p
p
r
o
a
c
h
,
a
l
l
o
w
i
n
g
a
n
y
a
p
p
l
i
c
a
t
i
o
n
t
o
b
r
o
w
s
e
a
m
o
u
n
t
e
d
d
o
c
u
m
e
n
t
p
a
c
k
a
g
e
.
T
h
e
r
e
m
o
t
e
t
a
g
g
i
n
g
o
f
f
i
l
e
s
w
o
u
l
d
c
r
e
a
t
e
a
r
e
a
d
-
o
n
l
y
d
i
s
t
r
i
b
u
t
e
d
f
i
l
e
s
y
s
t
e
m
.
A
H
T
T
P
p
r
o
x
y
c
o
u
l
d
b
e
e
x
t
e
n
d
e
d
t
o
p
r
o
v
i
d
e
a
d
v
a
n
c
e
d
c
a
c
h
i
n
g
f
a
c
i
l
i
t
i
e
s
f
o
r
a
n
e
n
t
i
r
e
w
e
b
s
i
t
e
.
P
F
S
c
r
e
a
t
o
r
s
w
o
u
l
d
h
a
v
e
t
h
e
p
o
s
s
i
b
i
l
i
t
y
o
f
m
a
i
n
t
a
i
n
i
n
g
s
t
a
t
i
c
i
n
f
o
r
m
a
t
i
o
n
a
s
e
m
b
e
d
d
e
d
f
i
l
e
s
w
h
i
l
e
f
r
e
q
u
e
n
t
l
y
c
h
a
n
g
i
n
g
f
i
l
e
c
o
n
t
e
n
t
s
c
o
u
l
d
b
e
t
a
g
g
e
d
a
s
r
e
m
o
t
e
f
i
l
e
s
.
W
e
b
c
l
i
e
n
t
s
u
s
i
n
g
a
p
r
o
x
y
w
o
u
l
d
o
n
l
y
n
e
e
d
t
o
o
b
t
a
i
n
r
e
m
o
t
e
Within a Server:
HTTP
HTTP
Fig.4: Local implementation of PFS handling is done by
Boa WWW server
6.1 Server Performance
6.2 Future Extensions
Figure 5. Sample “DIET.PFS” wrapper used for
performance testing
Original Server
Access to a
normal page
2.5 sec
Access to page
within PFS
---
7. Related Work
2.5 sec
PFS Compliant
Server
Table 1: Access times for web page- 3 files using 26180
bytes in total
2.5 sec
f
i
l
e
s
a
s
n
e
e
d
e
d
,
w
i
t
h
e
m
b
e
d
d
e
d
f
i
l
e
s
b
e
i
n
g
r
e
t
u
r
n
e
d
b
y
t
h
e
l
o
c
a
l
p
r
o
x
y
.
P
r
o
v
i
d
e
b
r
o
w
s
i
n
g
a
n
d
d
o
c
u
m
e
n
t
r
e
t
r
i
e
v
a
l
a
s
d
e
s
c
r
i
b
e
d
i
n
t
h
e
p
r
e
v
i
o
u
s
s
e
c
t
i
o
n
s
.
I
n
o
u
r
i
m
p
l
e
m
e
n
t
a
t
i
o
n
w
e
o
p
t
e
d
f
o
r
t
h
e
s
e
r
v
e
r
o
p
t
i
o
n
.
W
e
u
s
e
d
a
f
r
e
e
l
y
a
v
a
i
l
a
b
l
e
w
e
b
s
e
r
v
e
r
c
a
l
l
e
d
B
o
a
(
v
0
.
9
2
)
.
O
r
i
g
i
n
a
l
l
y
d
e
v
e
l
o
p
e
d
b
y
P
a
u
l
P
h
i
l
i
p
s
a
s
i
t
a
p
p
e
a
r
e
d
s
u
i
t
a
b
l
e
a
n
d
e
a
s
i
l
y
e
x
t
e
n
d
i
b
l
e
.
B
o
a
+
P
F
S
a
s
i
t
i
s
n
o
w
c
a
l
l
e
d
,
r
u
n
s
u
n
d
e
r
t
h
e
L
i
n
u
x
2
.
0
o
p
e
r
a
t
i
n
g
s
y
s
t
e
m
.
F
i
g
u
r
e
4
.
0
,
p
r
o
v
i
d
e
s
a
n
o
v
e
r
v
i
e
w
o
f
o
u
r
i
m
p
l
e
m
e
n
t
e
d
a
r
c
h
i
t
e
c
t
u
r
e
.
W
W
W
B
r
o
w
s
e
r
W
W
W
S
e
r
v
e
r
P
F
S
W
r
a
p
p
e
r
L
o
c
a
l
H
T
M
L
F
i
l
e
s
L
o
c
a
l
H
T
M
L
W
W
W
S
e
r
v
e
r
P
F
S
H
T
M
L
P
F
S
f
i
l
e
s
h
e
l
d
l
o
c
a
l
l
y
a
r
e
a
c
c
e
s
s
e
d
t
h
r
o
u
g
h
a
b
r
o
w
s
e
r
b
y
s
p
e
c
i
f
y
i
n
g
t
h
e
i
r
l
o
c
a
l
U
R
L
.
T
h
e
s
e
r
e
q
u
i
r
e
a
p
p
r
o
p
r
i
a
t
e
p
e
r
m
i
s
s
i
o
n
s
t
o
b
e
s
e
t
i
n
o
r
d
e
r
t
o
b
e
a
c
c
e
s
s
i
b
l
e
b
y
B
o
a
.
T
h
e
l
o
c
a
l
B
o
a
s
e
r
v
e
r
,
b
r
o
w
s
e
s
t
h
e
s
p
e
c
i
f
i
e
d
f
i
l
e
a
n
d
r
e
t
u
r
n
s
d
o
c
u
m
e
n
t
s
h
e
l
d
w
i
t
h
i
n
t
h
e
m
.
A
d
d
i
t
i
o
n
a
l
l
y
,
a
r
e
m
o
t
e
B
o
a
s
e
r
v
e
r
m
a
y
b
e
u
s
e
d
t
o
p
r
o
v
i
d
e
t
h
e
s
a
m
e
f
u
n
c
t
i
o
n
a
l
i
t
y
.
N
e
w
a
d
d
i
t
i
o
n
s
t
o
s
e
r
v
e
r
s
o
f
t
e
n
i
m
p
a
c
t
o
n
t
h
e
i
r
p
e
r
f
o
r
m
a
n
c
e
.
T
h
e
r
e
f
o
r
e
,
m
o
d
i
f
i
c
a
t
i
o
n
s
t
o
f
i
l
e
h
a
n
d
l
i
n
g
a
n
d
r
e
t
r
i
e
v
a
l
a
l
g
o
r
i
t
h
m
s
a
r
e
l
i
k
e
l
y
t
o
i
n
c
r
e
a
s
e
t
h
e
c
u
r
r
e
n
t
s
e
r
v
e
r
’
s
p
r
o
c
e
s
s
i
n
g
o
v
e
r
h
e
a
d
s
.
I
n
o
u
r
c
a
s
e
w
e
t
e
s
t
e
d
s
e
r
v
e
r
p
e
r
f
o
r
m
a
n
c
e
o
n
a
n
C
S
M
A
/
C
D
(
e
t
h
e
r
n
e
t
)
L
A
N
,
w
i
t
h
t
h
e
s
e
r
v
e
r
a
n
d
c
l
i
e
n
t
w
o
r
k
s
t
a
t
i
o
n
.
T
h
e
s
e
r
v
e
r
w
a
s
a
P
e
n
t
i
u
m
1
2
0
M
h
z
n
o
t
e
b
o
o
k
a
n
d
t
h
e
c
l
i
e
n
t
a
4
8
6
D
X
/
1
0
0
M
h
z
d
e
s
k
t
o
p
s
y
s
t
e
m
.
T
h
e
f
o
l
l
o
w
i
n
g
r
e
s
u
l
t
s
w
e
r
e
o
b
t
a
i
n
e
d
f
o
r
t
h
e
o
r
i
g
i
n
a
l
v
e
r
s
i
o
n
o
f
t
h
e
s
e
r
v
e
r
(
v
0
.
9
2
)
a
n
d
t
h
e
P
F
S
c
o
m
p
l
i
a
n
t
s
e
r
v
e
r
(
v
0
.
9
3
)
:
W
e
r
a
n
t
h
e
t
e
s
t
1
0
t
i
m
e
s
a
n
d
o
b
t
a
i
n
e
d
t
h
e
f
o
l
l
o
w
i
n
g
a
v
e
r
a
g
e
v
a
l
u
e
s
.
I
n
o
u
r
e
x
p
e
r
i
e
n
c
e
,
w
e
f
o
u
n
d
n
o
r
e
d
u
c
t
i
o
n
i
n
t
h
e
s
e
r
v
e
r
’
s
p
e
r
f
o
r
m
a
n
c
e
.
A
c
c
e
s
s
i
n
g
d
o
c
u
m
e
n
t
s
w
i
t
h
i
n
t
h
e
P
F
S
w
r
a
p
p
e
r
a
s
o
p
p
o
s
e
d
t
o
n
o
n
-
w
r
a
p
p
e
d
d
o
c
u
m
e
n
t
s
y
i
e
l
d
s
t
h
e
s
a
m
e
r
e
s
u
l
t
s
.
T
h
e
r
e
i
s
i
n
f
a
c
t
a
n
o
v
e
r
h
e
a
d
i
n
c
o
d
e
,
b
u
t
i
t
d
o
e
s
n
o
t
i
m
p
a
c
t
o
n
s
e
r
v
e
r
p
e
r
f
o
r
m
a
n
c
e
.
I
n
o
u
r
i
m
p
l
e
m
e
n
t
a
t
i
o
n
,
w
e
u
s
e
a
l
i
n
e
a
r
f
i
l
e
s
e
a
r
c
h
i
n
g
a
l
g
o
r
i
t
h
m
w
h
i
c
h
i
s
n
o
t
n
o
t
i
c
e
a
b
l
e
f
o
r
s
m
a
l
l
P
F
S
c
o
n
t
a
i
n
e
r
s
.
W
e
s
u
s
p
e
c
t
t
h
a
t
t
h
e
s
e
r
v
e
r
m
a
y
s
l
o
w
d
o
w
n
a
n
d
w
r
a
p
p
e
r
a
c
c
e
s
s
m
a
y
b
e
c
o
m
e
n
o
t
i
c
e
a
b
l
e
f
o
r
l
a
r
g
e
p
a
c
k
a
g
e
s
c
o
n
t
a
i
n
i
n
g
h
u
n
d
r
e
d
s
o
f
f
i
l
e
s
.
F
i
g
u
r
e
5
,
s
h
o
w
s
t
h
e
t
e
s
t
p
a
g
e
u
p
o
n
w
h
i
c
h
w
e
b
a
s
e
d
o
u
r
t
i
m
i
n
g
s
.
I
t
s
h
o
w
s
a
b
r
o
w
s
e
r
l
o
o
k
i
n
g
i
n
s
i
d
e
a
s
a
m
p
l
e
p
a
g
e
w
h
i
c
h
i
s
h
e
l
d
w
i
t
h
i
n
a
r
e
m
o
t
e
“
d
i
e
t
.
p
f
s
”
p
a
c
k
a
g
e
.
F
u
t
u
r
e
i
m
p
l
e
m
e
n
t
a
t
i
o
n
s
a
n
d
e
x
t
e
n
s
i
o
n
s
,
m
a
y
p
r
o
v
i
d
e
a
f
a
s
t
e
r
d
i
r
e
c
t
o
r
y
l
o
o
k
u
p
s
t
r
u
c
t
u
r
e
i
n
s
t
e
a
d
o
f
t
h
e
c
u
r
r
e
n
t
l
i
n
e
a
r
s
e
a
r
c
h
.
W
e
a
r
e
e
x
t
e
n
d
i
n
g
t
h
e
i
n
t
e
r
n
a
l
P
F
S
f
i
l
e
f
o
r
m
a
t
t
o
a
l
l
o
w
f
o
r
f
u
t
u
r
e
f
o
r
m
a
t
e
x
t
e
n
s
i
o
n
s
w
h
i
l
e
m
a
i
n
t
a
i
n
i
n
g
b
a
c
k
w
a
r
d
s
c
o
m
p
a
t
i
b
i
l
i
t
y
.
T
h
i
s
i
s
l
i
k
e
l
y
t
o
r
e
s
u
l
t
s
i
n
f
u
r
t
h
e
r
B
O
A
+
P
F
S
s
e
r
v
e
r
m
o
d
i
f
i
c
a
t
i
o
n
s
.
A
d
d
i
t
i
o
n
a
l
w
o
r
k
i
n
d
e
v
e
l
o
p
i
n
g
a
n
i
n
s
t
a
l
l
a
b
l
e
f
i
l
e
s
y
s
t
e
m
f
o
r
L
i
n
u
x
/
W
i
n
d
o
w
s
9
5
i
s
a
l
s
o
b
e
i
n
g
e
x
a
m
i
n
e
d
.
D
i
s
t
r
i
b
u
t
e
d
w
i
r
e
l
e
s
s
m
u
l
t
i
m
e
d
i
a
a
c
c
e
s
s
h
a
s
g
a
i
n
e
d
s
i
g
n
i
f
i
c
a
n
t
a
t
t
e
n
t
i
o
n
o
v
e
r
t
h
e
l
a
s
t
s
e
v
e
r
a
l
y
e
a
r
s
.
R
e
l
e
v
a
n
t
p
r
o
j
e
c
t
s
a
r
e
l
i
s
t
e
d
b
e
l
o
w
t
o
p
r
o
v
i
d
e
t
h
e
r
e
a
d
e
r
w
i
t
h
s
o
m
e
f
u
r
t
h
e
r
i
n
s
i
g
h
t
i
n
t
o
t
h
e
f
i
e
l
d
.
T
h
e
D
e
f
e
n
s
e
A
d
v
a
n
c
e
d
R
e
s
e
a
r
c
h
P
r
o
j
e
c
t
A
g
e
n
c
y
(
D
A
R
P
A
)
h
a
s
i
n
i
t
i
a
t
e
d
t
h
e
G
l
o
b
a
l
M
o
b
i
l
e
I
n
f
o
r
m
a
t
i
o
n
S
y
s
t
e
m
s
(
G
l
o
M
o
)
p
r
o
g
r
a
m
i
n
1
9
9
4
[
5
]
.
I
t
i
s
a
t
e
s
t
b
e
d
i
n
i
t
i
a
t
i
v
e
l
i
n
k
i
n
g
s
e
v
e
r
a
l
r
e
s
e
a
r
c
h
p
r
o
j
e
c
t
s
w
h
i
c
h
d
e
p
e
n
d
o
n
a
w
i
r
e
l
e
s
s
c
o
m
m
u
n
i
c
a
t
i
o
n
s
i
n
f
r
a
s
t
r
u
c
t
u
r
e
.
O
n
e
s
u
c
h
p
r
o
j
e
c
t
a
t
t
h
e
U
n
i
v
e
r
s
i
t
y
o
f
C
a
l
i
f
o
r
n
i
a
,
B
e
r
k
e
l
e
y
,
i
s
t
h
e
l
i
g
h
t
w
e
i
g
h
t
,
p
o
r
t
a
b
l
e
m
u
l
t
i
m
e
d
i
a
n
o
t
e
b
o
o
k
c
a
l
l
e
d
I
n
f
o
P
a
d
[
4
]
.
I
t
a
i
m
s
t
o
p
r
o
v
i
d
e
m
u
l
t
i
m
e
d
i
a
a
c
c
e
s
s
i
n
a
d
i
s
t
r
i
b
u
t
e
d
w
i
r
e
l
e
s
s
e
n
v
i
r
o
n
m
e
n
t
.
I
n
f
o
P
a
d
p
e
r
f
o
r
m
s
a
s
a
m
o
b
i
l
e
t
e
r
m
i
n
a
l
w
h
i
c
h
m
a
i
n
t
a
i
n
s
a
p
p
l
i
c
a
t
i
o
n
s
o
n
d
e
d
i
c
a
t
e
d
s
e
r
v
e
r
s
.
A
l
l
u
s
e
r
Acknowledgments
References
8. Conclusion
a
p
p
l
i
c
a
t
i
o
n
s
e
x
e
c
u
t
e
r
e
m
o
t
e
l
y
,
w
h
i
l
e
t
h
e
i
r
r
e
s
u
l
t
s
a
r
e
s
e
n
t
t
o
t
h
e
I
n
f
o
P
a
d
t
e
r
m
i
n
a
l
.
W
e
b
o
r
i
e
n
t
e
d
p
r
o
j
e
c
t
s
s
u
c
h
a
s
M
o
b
i
s
a
i
c
a
n
d
M
o
w
s
e
r
a
r
e
i
n
d
i
r
e
c
t
l
y
r
e
l
a
t
e
d
t
o
o
u
r
w
r
a
p
p
e
r
p
r
o
j
e
c
t
.
T
h
e
M
o
b
i
s
a
i
c
[
9
]
p
r
o
j
e
c
t
a
t
t
h
e
U
n
i
v
e
r
i
s
t
y
o
f
W
a
s
h
i
n
g
t
o
n
,
e
m
p
l
o
y
s
c
o
n
t
e
x
t
s
e
n
s
i
t
i
v
e
U
R
L
s
i
n
a
w
i
r
e
l
e
s
s
l
o
c
a
l
a
r
e
a
n
e
t
w
o
r
k
t
o
p
r
o
v
i
d
e
l
o
c
a
t
i
o
n
d
e
p
e
n
d
e
n
t
w
e
b
d
o
c
u
m
e
n
t
s
.
W
e
a
l
s
o
e
m
p
l
o
y
m
o
d
i
f
i
e
d
U
R
L
s
w
h
i
l
e
l
e
a
v
i
n
g
t
h
e
s
e
r
v
e
r
t
o
r
e
s
o
l
v
e
t
h
e
m
.
A
m
o
b
i
l
e
M
o
b
i
s
a
i
c
u
s
e
r
c
o
m
m
u
n
i
c
a
t
e
s
w
i
t
h
p
h
y
s
i
c
a
l
l
y
v
i
s
i
b
l
e
d
e
v
i
c
e
s
v
i
a
i
n
f
r
a
-
r
e
d
t
o
o
b
t
a
i
n
e
n
v
i
r
o
n
m
e
n
t
a
l
v
a
r
i
a
b
l
e
s
.
T
h
e
s
e
a
r
e
t
h
e
n
i
n
c
o
r
p
o
r
a
t
e
d
w
i
t
h
i
n
d
y
n
a
m
i
c
U
R
L
s
w
h
i
c
h
p
r
o
v
i
d
e
t
h
e
u
s
e
r
’
s
l
o
c
a
t
i
o
n
b
y
r
e
t
r
i
e
v
i
n
g
t
h
e
c
o
r
r
e
s
p
o
n
d
i
n
g
w
e
b
p
a
g
e
s
.
M
o
b
i
s
a
i
c
p
r
o
p
o
s
e
s
t
h
e
c
o
n
c
e
p
t
o
f
b
r
o
w
s
e
r
r
e
g
i
s
t
r
a
t
i
o
n
a
n
d
p
e
r
m
i
t
s
s
e
r
v
e
r
c
a
l
l
b
a
c
k
s
t
o
c
l
i
e
n
t
b
r
o
w
s
e
r
s
.
B
r
o
w
s
e
r
c
a
l
l
b
a
c
k
s
a
s
s
i
s
t
i
n
d
o
c
u
m
e
n
t
u
p
d
a
t
e
s
b
y
n
o
t
i
f
y
i
n
g
r
e
g
i
s
t
e
r
e
d
b
r
o
w
s
e
r
s
t
o
o
b
t
a
i
n
n
e
w
l
y
c
h
a
n
g
e
d
d
o
c
u
m
e
n
t
s
.
A
n
o
t
h
e
r
m
u
l
t
i
m
e
d
i
a
p
r
o
j
e
c
t
c
a
l
l
e
d
M
o
w
s
e
r
,
h
a
s
b
e
e
n
d
e
v
e
l
o
p
e
d
a
t
P
u
r
d
u
e
U
n
i
v
e
r
s
i
t
y
[
2
]
.
T
h
e
M
o
w
s
e
r
a
r
c
h
i
t
e
c
t
u
r
e
e
m
p
l
o
y
s
d
e
s
i
g
n
a
t
e
d
p
r
o
x
y
s
e
r
v
e
r
s
w
h
i
c
h
s
e
r
v
i
c
e
m
o
b
i
l
e
u
s
e
r
s
.
T
h
e
s
e
e
x
a
m
i
n
e
i
n
c
o
m
i
n
g
d
o
c
u
m
e
n
t
s
w
h
i
l
e
c
a
c
h
i
n
g
t
h
e
m
.
A
M
o
w
s
e
r
p
r
o
x
y
m
o
d
i
f
i
e
s
r
e
t
r
i
e
v
e
d
d
o
c
u
m
e
n
t
s
a
n
d
f
o
r
w
a
r
d
s
r
e
s
i
z
e
d
i
m
a
g
e
s
t
o
m
o
b
i
l
e
b
r
o
w
s
e
r
s
.
T
h
i
s
r
e
d
u
c
e
s
t
h
e
n
e
c
e
s
s
a
r
y
b
a
n
d
w
i
d
t
h
r
e
q
u
i
r
e
m
e
n
t
s
w
h
i
l
e
d
e
l
i
v
e
r
i
n
g
t
h
e
d
o
c
u
m
e
n
t
c
o
n
t
e
n
t
.
T
h
e
U
n
i
v
e
r
s
i
t
y
o
f
W
a
s
h
i
n
g
t
o
n
,
d
e
v
e
l
o
p
e
d
t
h
e
W
*
w
i
r
e
l
e
s
s
w
e
b
b
r
o
w
s
e
r
[
1
0
]
.
T
h
e
i
r
s
y
s
t
e
m
e
m
p
l
o
y
e
d
a
l
o
w
p
o
w
e
r
e
d
C
P
U
i
n
a
w
i
r
e
l
e
s
s
P
e
r
s
o
n
a
l
D
i
g
i
t
a
l
A
s
s
i
s
t
a
n
t
(
P
D
A
)
.
T
h
i
s
a
p
p
e
a
r
s
t
o
b
e
s
i
m
i
l
a
r
t
o
M
o
w
s
e
r
a
s
b
o
t
h
p
r
o
j
e
c
t
s
e
m
p
l
o
y
p
r
o
x
y
s
e
r
v
e
r
s
t
o
a
s
s
i
s
t
d
o
c
u
m
e
n
t
m
o
d
i
f
i
c
a
t
i
o
n
a
n
d
p
r
e
p
r
o
c
e
s
s
i
n
g
.
T
h
e
m
o
b
i
l
e
P
D
A
r
u
n
n
i
n
g
W
*
c
o
m
m
u
n
i
c
a
t
e
s
w
i
t
h
a
s
p
e
c
i
a
l
i
z
e
d
p
r
o
x
y
s
e
r
v
e
r
v
i
a
a
c
o
m
p
r
e
s
s
e
d
p
r
o
t
o
c
o
l
.
T
h
e
p
r
o
x
y
c
a
c
h
e
s
p
r
e
v
i
o
u
s
l
y
a
c
c
e
s
s
e
d
d
o
c
u
m
e
n
t
s
,
p
r
e
p
r
o
c
e
s
s
e
s
b
r
o
w
s
e
r
d
i
s
p
l
a
y
s
,
p
r
e
f
e
t
c
h
e
s
t
h
e
f
i
r
s
t
p
a
g
e
o
f
p
o
t
e
n
t
i
a
l
n
e
x
t
l
i
n
k
s
,
a
n
d
p
a
s
s
e
s
c
o
m
p
r
e
s
s
e
d
d
o
c
u
m
e
n
t
s
v
i
a
t
h
e
w
i
r
e
l
e
s
s
c
o
n
n
e
c
t
i
o
n
.
T
o
t
h
e
b
e
s
t
o
f
o
u
r
k
n
o
w
l
e
d
g
e
,
d
o
c
u
m
e
n
t
m
a
n
a
g
e
m
e
n
t
s
u
c
h
a
s
t
h
e
P
F
S
w
r
a
p
p
e
r
a
p
p
r
o
a
c
h
h
a
s
n
o
t
b
e
e
n
a
t
t
e
m
p
t
e
d
a
n
d
a
p
p
e
a
r
s
t
o
b
e
a
n
e
w
c
o
n
t
r
i
b
u
t
i
o
n
t
o
t
h
i
s
f
i
e
l
d
o
f
w
o
r
k
.
T
h
e
P
F
S
w
r
a
p
p
e
r
a
p
p
r
o
a
c
h
t
o
a
r
c
h
i
v
i
n
g
s
i
t
e
s
w
h
i
l
e
m
a
i
n
t
a
i
n
i
n
g
t
h
e
i
r
w
e
b
p
r
e
s
e
n
c
e
i
s
a
n
e
w
s
e
r
v
e
r
f
e
a
t
u
r
e
.
I
t
p
e
r
m
i
t
s
s
i
t
e
a
d
m
i
n
i
s
t
r
a
t
o
r
s
t
o
c
o
l
l
a
p
s
e
d
i
r
e
c
t
o
r
i
e
s
a
n
d
o
f
f
e
r
s
p
a
c
k
a
g
e
d
f
i
l
e
s
t
o
u
s
e
r
s
w
i
s
h
i
n
g
t
o
a
c
c
e
s
s
p
o
r
t
i
o
n
s
o
f
a
w
e
b
s
e
r
v
e
r
t
r
e
e
.
A
d
d
i
t
i
o
n
a
l
l
y
,
t
h
e
s
u
p
p
o
r
t
f
o
r
r
e
m
o
t
e
f
i
l
e
r
e
f
e
r
e
n
c
i
n
g
w
i
t
h
o
u
t
p
h
y
s
i
c
a
l
l
y
e
m
b
e
d
d
i
n
g
f
i
l
e
s
h
a
v
e
b
e
e
n
u
s
e
d
i
n
d
i
s
t
r
i
b
u
t
e
d
f
i
l
e
c
a
c
h
i
n
g
s
y
s
t
e
m
s
b
u
t
a
p
p
e
a
r
s
t
o
b
e
a
n
o
v
e
l
a
p
p
r
o
a
c
h
f
o
r
a
r
c
h
i
v
e
s
.
T
h
e
P
F
S
c
o
d
e
a
d
d
i
t
i
o
n
w
i
t
h
i
n
o
u
r
s
e
r
v
e
r
,
p
r
o
v
i
d
e
s
m
i
n
i
m
a
l
o
v
e
r
h
e
a
d
f
o
r
d
o
c
u
m
e
n
t
a
c
c
e
s
s
a
n
d
d
o
e
s
n
o
t
d
e
g
r
a
d
e
s
e
r
v
e
r
p
e
r
f
o
r
m
a
n
c
e
.
M
o
b
i
l
e
u
s
e
r
s
w
h
i
c
h
a
r
e
c
a
p
a
b
l
e
o
f
r
u
n
n
i
n
g
a
s
m
a
l
l
l
o
c
a
l
w
e
b
s
e
r
v
e
r
,
m
a
y
a
c
c
e
s
s
P
F
S
f
i
l
e
s
a
n
d
r
e
t
r
i
e
v
e
n
e
c
e
s
s
a
r
y
d
o
c
u
m
e
n
t
s
w
h
i
l
e
d
i
s
c
o
n
n
e
c
t
e
d
f
r
o
m
a
n
e
t
w
o
r
k
.
T
h
e
u
s
e
o
f
a
s
i
n
g
l
e
P
F
S
w
r
a
p
p
e
r
m
i
n
i
m
i
z
e
s
t
h
e
n
e
e
d
t
o
p
o
t
e
n
t
i
a
l
l
y
m
a
i
n
t
a
i
n
h
u
n
d
r
e
d
s
o
f
f
i
l
e
s
o
n
l
o
c
a
l
s
t
o
r
a
g
e
.
T
h
e
p
o
r
t
a
b
l
e
f
i
l
e
s
y
s
t
e
m
a
l
s
o
s
i
m
p
l
i
f
i
e
s
m
u
l
t
i
-
d
o
c
u
m
e
n
t
d
i
s
t
r
i
b
u
t
i
o
n
.
I
t
s
e
a
m
l
e
s
s
l
y
i
n
t
e
g
r
a
t
e
s
i
n
t
o
a
w
e
b
s
e
r
v
e
r
a
n
d
a
p
p
e
a
r
s
t
r
a
n
s
p
a
r
e
n
t
t
o
i
t
s
u
s
e
r
s
.
T
h
e
a
u
t
h
o
r
s
w
o
u
l
d
l
i
k
e
t
o
t
h
a
n
k
J
o
h
n
N
g
f
o
r
o
b
t
a
i
n
i
n
g
t
h
e
o
r
i
g
i
n
a
l
B
o
a
a
r
c
h
i
v
e
a
n
d
t
h
e
i
n
s
p
i
r
a
t
i
o
n
w
h
i
c
h
l
e
d
t
o
t
h
e
e
x
t
e
n
s
i
o
n
o
f
B
o
a
.
[
1
]
J
.
H
o
w
a
r
d
e
t
a
l
.
,
"
S
c
a
l
e
a
n
d
P
e
r
f
o
r
m
a
n
c
e
i
n
a
D
i
s
t
r
i
b
u
t
e
d
F
i
l
e
S
y
s
t
e
m
"
,
A
C
M
T
r
a
n
s
a
c
t
i
o
n
s
o
n
C
o
m
p
u
t
e
r
S
y
s
t
e
m
s
,
F
e
b
r
u
a
r
y
1
9
8
8
,
p
p
.
5
1
-
8
1
[
2
]
A
.
J
o
s
h
i
,
R
.
W
e
e
r
a
s
i
n
g
h
e
,
S
.
M
c
D
e
r
m
o
t
t
e
t
a
l
.
,
1
9
9
6
,
"
M
o
w
s
e
r
:
M
o
b
i
l
e
P
l
a
t
f
o
r
m
s
a
n
d
W
e
b
B
r
o
w
s
e
r
s
"
,
T
e
c
h
n
i
c
a
l
C
o
m
m
i
t
t
e
e
o
n
O
p
e
r
a
t
i
n
g
S
y
s
t
e
m
s
a
n
d
A
p
p
l
i
c
a
t
i
o
n
E
n
v
i
r
o
n
m
e
n
t
s
(
T
C
O
S
)
,
S
p
r
i
n
g
1
9
9
6
,
N
o
.
1
,
V
o
l
.
8
,
p
p
.
1
3
-
1
6
[
3
]
J
.
J
.
K
i
s
t
l
e
r
a
n
d
M
.
S
a
t
y
a
n
a
r
a
y
a
n
a
n
,
1
9
9
2
,
"
D
i
s
c
o
n
n
e
c
t
e
d
O
p
e
r
a
t
i
o
n
i
n
t
h
e
C
O
D
A
F
i
l
e
S
y
s
t
e
m
"
,
A
C
M
T
r
a
n
s
a
c
t
i
o
n
s
o
n
C
o
m
p
u
t
e
r
S
y
s
t
e
m
s
,
V
o
l
.
1
0
,
N
o
.
1
,
F
e
b
r
u
a
r
y
1
9
9
2
[
4
]
M
.
T
.
L
e
e
t
a
l
.
,
1
9
9
5
,
"
I
n
f
o
N
e
t
:
t
h
e
N
e
t
w
o
r
k
i
n
g
I
n
f
r
a
s
t
r
u
c
t
u
r
e
o
f
I
n
f
o
P
a
d
"
,
P
r
o
c
e
e
d
i
n
g
s
o
f
C
o
m
p
c
o
n
,
S
a
n
F
r
a
n
c
i
s
c
o
,
M
a
r
c
h
1
9
9
5
[
5
]
B
.
M
.
L
e
i
n
e
r
,
R
.
J
.
R
u
t
h
a
n
d
A
.
R
.
S
a
s
t
r
y
,
1
9
9
6
,
"
G
o
a
l
s
a
n
d
C
h
a
l
l
e
n
g
e
s
o
f
t
h
e
D
A
R
P
A
G
l
o
M
o
P
r
o
g
r
a
m
"
,
D
e
c
e
m
b
e
r
1
9
9
6
,
I
E
E
E
P
e
r
s
o
n
a
l
C
o
m
m
u
n
i
c
a
t
i
o
n
s
,
p
p
.
3
4
-
4
3
[
6
]
T
.
P
a
g
e
,
1
9
9
1
,
"
A
r
c
h
i
t
e
c
t
u
r
e
o
f
t
h
e
F
i
c
u
s
S
c
a
l
a
b
l
e
R
e
p
l
i
c
a
t
e
d
F
i
l
e
S
y
s
t
e
m
"
,
T
e
c
h
n
i
c
a
l
R
e
p
o
r
t
C
S
D
-
9
1
0
0
0
5
,
U
n
i
v
e
r
s
i
t
y
o
f
C
a
l
i
f
o
r
n
i
a
,
L
o
s
A
n
g
e
l
e
s
,
M
a
r
c
h
1
9
9
1
[
7
]
J
.
S
a
l
d
a
n
h
a
,
1
9
9
6
,
"
A
N
e
w
F
i
l
e
S
y
s
t
e
m
F
o
r
M
o
b
i
l
e
C
o
m
p
u
t
i
n
g
"
-
T
h
e
s
i
s
,
D
e
p
a
r
t
m
e
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
U
n
i
v
e
r
s
i
t
y
o
f
N
o
t
r
e
D
a
m
e
,
I
n
d
i
a
n
a
[
8
]
P
.
S
t
a
n
s
k
i
a
n
d
A
.
Z
a
s
l
a
v
s
k
y
,
1
9
9
7
,
"
D
i
s
t
r
i
b
u
t
e
d
a
n
d
P
e
r
s
i
s
t
e
n
t
M
o
b
i
l
e
A
g
e
n
t
s
f
o
r
H
e
t
e
r
o
g
e
n
e
o
u
s
P
e
r
s
o
n
a
l
C
o
m
m
u
n
i
c
a
t
i
o
n
s
S
y
s
t
e
m
s
"
,
P
r
o
c
e
e
d
i
n
g
s
:
6
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
C
o
m
p
u
t
e
r
C
o
m
m
u
n
i
c
a
t
i
o
n
s
a
n
d
N
e
t
w
o
r
k
s
(
I
C
3
N
)
,
S
e
p
t
e
m
b
e
r
1
9
9
7
[
9
]
G
.
M
.
V
o
e
l
k
e
r
a
n
d
B
.
N
.
B
e
r
s
h
a
d
,
1
9
9
4
,
"
M
O
B
I
S
A
I
C
:
A
n
I
n
f
o
r
m
a
t
i
o
n
S
y
s
t
e
m
f
o
r
a
M
o
b
i
l
e
W
i
r
e
l
e
s
s
C
o
m
p
u
t
i
n
g
E
n
v
i
r
o
n
m
e
n
t
"
,
P
r
o
c
e
e
d
i
n
g
s
:
W
o
r
k
s
h
o
p
o
n
M
o
b
i
l
e
C
o
m
p
u
t
i
n
g
S
y
s
t
e
m
s
a
n
d
A
p
p
l
i
c
a
t
i
o
n
s
,
D
e
c
e
m
b
e
r
1
9
9
4
[
1
0
]
T
.
W
a
t
s
o
n
,
1
9
9
4
,
"
A
p
p
l
i
c
a
t
i
o
n
D
e
s
i
g
n
f
o
r
W
i
r
e
l
e
s
s
C
o
m
p
u
t
i
n
g
"
,
P
r
o
c
e
e
d
i
n
g
s
:
M
o
b
i
l
e
C
o
m
p
u
t
i
n
g
S
y
s
t
e
m
s
a
n
d
A
p
p
l
i
c
a
t
i
o
n
s
1
9
9
4
|
1805.08320 | 1 | 1805 | 2018-05-21T23:18:58 | The Swarmathon: An Autonomous Swarm Robotics Competition | [
"cs.MA",
"cs.RO"
] | The Swarmathon is a swarm robotics programming challenge that engages college students from minority-serving institutions in NASA's Journey to Mars. Teams compete by programming a group of robots to search for, pick up, and drop off resources in a collection zone. The Swarmathon produces prototypes for robot swarms that would collect resources on the surface of Mars. Robots operate completely autonomously with no global map, and each team's algorithm must be sufficiently flexible to effectively find resources from a variety of unknown distributions. The Swarmathon includes Physical and Virtual Competitions. Physical competitors test their algorithms on robots they build at their schools; they then upload their code to run autonomously on identical robots during the three day competition in an outdoor arena at Kennedy Space Center. Virtual competitors complete an identical challenge in simulation. Participants mentor local teams to compete in a separate High School Division. In the first 2 years, over 1,100 students participated. 63% of students were from underrepresented ethnic and racial groups. Participants had significant gains in both interest and core robotic competencies that were equivalent across gender and racial groups, suggesting that the Swarmathon is effectively educating a diverse population of future roboticists. | cs.MA | cs | The Swarmathon: An Autonomous Swarm Robotics Competition
Sarah M. Ackerman∗1, G. Matthew Fricke1, Joshua P. Hecker1, Kastro M. Hamed2, Samantha R. Fowler2,
Antonio D. Griego1, Jarett C. Jones1, J. Jake Nichol1, Kurt W. Leucht3, and Melanie E. Moses1,4,5
8
1
0
2
y
a
M
1
2
]
A
M
.
s
c
[
1
v
0
2
3
8
0
.
5
0
8
1
:
v
i
X
r
a
Abstract- The Swarmathon is a swarm robotics program-
ming challenge that engages college students from minority-
serving institutions in NASA's Journey to Mars. Teams compete
by programming a group of robots to search for, pick up,
and drop off resources in a collection zone. The Swarmathon
produces prototypes for robot swarms that would collect
resources on the surface of Mars. Robots operate completely
autonomously with no global map, and each team's algorithm
must be sufficiently flexible to effectively find resources from
a variety of unknown distributions. The Swarmathon includes
Physical and Virtual Competitions. Physical competitors test
their algorithms on robots they build at their schools; they
then upload their code to run autonomously on identical
robots during the three day competition in an outdoor arena
at Kennedy Space Center. Virtual competitors complete an
identical challenge in simulation. Participants mentor local
teams to compete in a separate High School Division. In
the first 2 years, over 1,100 students participated. 63% of
students were from underrepresented ethnic and racial groups.
Participants had significant gains in both interest and core
robotic competencies that were equivalent across gender and
racial groups, suggesting that the Swarmathon is effectively
educating a diverse population of future roboticists.
I. INTRODUCTION
A. The Journey to Mars
The National Aeronautics and Space Administration
(NASA) Journey to Mars program has set the ambitious goal
of sending manned missions to Mars by the 2030s.1 These
surface missions may last months or years, and because
transporting sufficient materials from Earth is not practical,
astronauts will need to utilize the resources already available
on Mars. For example, small pockets of ice can be melted for
water and converted to oxygen and hydrogen for fuel. This
approach is known as in-situ resource utilization (ISRU).
B. Meeting the Challenge with Robot Swarms
Autonomous robot swarms could locate, collect, and store
resources in advance of human arrival. A swarm of small,
1Computer Science Department, The University of New Mexico, Albu-
querque, NM, USA, 2Education and Interdisciplinary Studies, Florida
Institute of Technology, Melbourne, FL, USA, 3NASA Kennedy Space
Center, FL, USA, 4Department of Biology, The University of New Mexico,
5Santa Fe Institute, Santa Fe, NM, USA, This work was supported by
NASA Minority University and Research Education Program (MUREP)
#NNX15AM14A and a generous donation from Google. We thank Theresa
Martinez, MUREP STEM Engagement Manager; Paul Secor and Mary
Baker of Secor Strategies, LLC; Kate Cunningham and Beatriz Palacios
Abad of UNM ADVANCE; Elizabeth Esterly and the students of the
Moses Biological Computation Lab, Vanessa Svihla, Aeron Haynie, NASA
volunteers, UNM support staff, KSC Visitor Complex, and the Swarmathon
students and faculty mentors.
∗Correspondence: [email protected]
1https://www.nasa.gov/content/nasas-journey-to-mars
inexpensive robots acting autonomously provides several
advantages over a few large, expensive, manually controlled
rovers [1], [2] because they are robust, scalable and able to
explore unmapped environments [3].
ISRU is closely related to central-place foraging (CPF),
in which resources are transported to a central collection
zone. Inspired by the success of social insects gathering
resources for their colonies, an early study of robot foraging
computationally modeled a swarm collecting rock samples
on a distant planet [4]. CPF is a key robot swarm application
[5], [6], and research continues to improve algorithms and
engineering approaches for foraging swarms [7], [8].
Our team has designed and built a swarm of foraging
robots called Swarmies. Swarmies are rugged enough to
operate outdoors for hours, and feature sensors and grippers
that enable them to complete CPF tasks. We also developed
custom software packages and a graphical user interface
(GUI), enabling rapid development and testing of CPF al-
gorithms in simulation and physical hardware.
Swarm foraging provides an ideal educational challenge
because successful strategies require foundational robotics
skills such as grasping, localization, navigation, exploration,
and decentralized communication and coordination [9].
C. The Swarmathon Competition
In the Swarmathon annual competition, teams of college
students develop algorithms for autonomous swarm foraging.
Teams in the Physical Competition program groups of 3 - 6
Swarmies to collect the most of 128−256 resources placed in
225 − 484 m2 outdoor arenas. The Virtual Competition runs
in a Gazebo simulation of the same arenas and robots. Robots
must search for, collect, and return resources to a central
collection zone completely autonomously. The competition
runs in a series of rounds, and in each round resources are
placed in different distributions (i.e., uniformly at random, or
in various sizes and shapes of clusters), and the same code
is run in each round. Score is determined by the number
of resources in the collection zone at the end of the round
(resources pushed out of the collection zone by robots do not
count toward the score). Robots have limited view and no
global map, making it difficult to find resources. In a third
High School Competition, younger students are mentored by
Swarmathon college students to complete a similar challenge
in a simplified simulation. Winners of the Virtual competition
advance to the Physical competition the following year.
D. Diversifying the STEM/Robotics Pipeline
Tackling NASA's real-world ISRU problem provides an
opportunity to engage students in robotics and computer
science (CS), helping meet the growing need for science,
technology, engineering, and mathematics (STEM) talent
in the workforce. While women and minorities account
for a growing proportion of US college graduates and US
workers, they are underrepresented among STEM college
graduates and professionals, especially in CS and engineer-
ing [10]–[13]. Their inclusion in STEM helps to address
an anticipated shortfall in these fields over the next few
decades [12], [13]. Additionally "workers from a variety
of backgrounds enhance the quality of science insofar as
they are likely to bring a variety of new perspectives. . . in
terms of both research and application" [11]. For these
reasons, the Swarmathon recruits competitors from 2- and
4-year colleges and universities that are minority-serving
institutions (MSIs) including Historically Black Colleges and
Universities (HBCUs), Hispanic-Serving Institutions (HSIs),
and Tribal Colleges and Universities (TCUs). This provides
an educational opportunity that may not otherwise exist for
participating students:
"The Swarmathon is an opportunity for students
to work on a real-life engineering problem that's
interdisciplinary and hard. They know that I don't
know the answer which makes it fun for all of us.
I teach at a small community college and it simply
wouldn't be possible for us do this level of work
from scratch." -Swarmathon Faculty Mentor
The Swarmathon has engaged a diverse student population
with 81% of students from minority groups (including Asian
students) and 63% of the 818 college participants identifying
as belonging to an underrepresented racial or ethnic group
(Black/African American, American Indian/Alaskan Native,
or Hispanic/Latino), substantially more diverse than US
undergraduate CS majors, of which only 15.8% are from
these underrepresented groups [14].
College students mentored K-12 students in their commu-
nities, including 300 high school students who compteed in
a simplified High School Swarmathon in Netlogo. Addition-
ally, the Swarmathon supported 60 students to attend work-
shops and the 2016 and 2017 Robotics Science and Systems
(RSS) conferences and 17 students participated in research
internships with mentors as other Swarmathon schools. All
of these activities further developed pathways into STEM. As
an example, 65% of RSS workshop participants indicated a
desire to pursue a Ph.D. in robotics.
robots, require permanent infrastructure such as buried guide-
wires or visual markers to operate [16].
Many attempts have been made to develop swarm algo-
rithms and robot systems that address various aspects of
the swarm foraging problem including scalability, energy
efficiency, task allocation, and collection speed [17]–[22].
However, autonomous swarm foraging remains a challenging
open problem with no generally recognized best solution.
Projects using physical robot swarms include the Rob-
otarium, a swarm robotics testbed providing remotely ac-
cessible robots and an arena [23]. This innovative project
allows virtual experimentation with physical robot swarms,
but
it uses an overhead camera for localization and the
robots themselves have no on-board sensors. Kilobots are
simple, small robots intended to be integrated into large
swarms. They have sensors and operate autonomously and
collaboratively to push items through a maze or into specific
configurations, but they have relatively limited mobility and
operate in controlled laboratory environments [24]. The
Swarmanoid project demonstrated successful implementation
of a heterogeneous swarm whose robots collaborate in order
to solve tasks like object retrieval in a highly specialized en-
vironment [25]. The challenges of operating physical swarms
result in robots that are usually semi-autonomous in prac-
tice; they require frequent human support [26]. The reality
gap between performance in simulation and performance in
physical hardware [27] makes implementation in real robots
particularly difficult. The hardware and software platforms
for the Swarmies address these challenges so that CPF can
be developed and subsequently tested in a system that is
completely autonomous, operates without global knowledge
or control, and can function outdoors.
B. Educational Robotics Environments and Competitions
Educational robotics is a growing field that is particularly
effective at improving student performance in and attitudes
towards STEM disciplines, and reducing gender and ethnic
achievement gaps [12], [28]–[31]. Examples for younger
students include the distributed Robot Garden [32] and low-
cost AERobots [33].
The NASA Robotic Mining Competition (RMC) is an
annual engineering competition to design and build robots
capable of mining the Martian surface as part of the ISRU
mission. Students focus on hardware design rather than
autonomous control. Only 16% of the 900 RMC participants
in 2017 identified as members of underrepresented groups2.
The Swarmathon was designed to emulate the success and
popularity of RMC while emphasizing robot autonomy and
engagement of students from MSIs.
II. RELATED WORK
C. Hardware
A. Foraging Robot Swarms
Though swarm robot foraging has been studied analyti-
cally for decades [4], hardware implementations are rare [9].
For example, studies of CPF often only simulate the collec-
tion and transport of objects [15], and recent advances in
hardware implementations, such as collaborative warehouse
Teams selected for the Physical Competition are sent
parts and instructions3 for building 3 Swarmies. Each robot
has a front mounted gripper with actuated wrist and finger
2https://www.nasa.gov/offices/education/centers/
kennedy/technology/nasarmc.html
3https://github.com/BCLab-UNM/Swarmathon-Robot
robot pose is estimated using two extended Kalman filters
(EKFs) [36] that fuse encoder, IMU, and GPS data.
This base code includes: Gazebo simulation of the robots,
target cubes, and competition arena (Figure 2); a GUI (Figure
2b) showing output of each hardware sensor, a map of the
robots' estimated positions (Figure 2c); customizable ROS
packages to control behaviors (for search, communication,
obstacle avoidance, and target pick up and drop off) and
diagnostics (for sensors, actuators, wireless quality and the
microcontroller); a ROS package that interfaces with the
Arduino to communicate with actuators and and sensors;
open source packages including AprilTag and EKF packages.
The simple base code algorithm (modeled after the sub-
sumption architecture [37]) is designed to be extended by
teams to improve foraging. It implements simple collision
avoidance, a random walk for search, and when a target
appears in the camera view, the robot picks up the target
and returns to the central collection zone to drop it off.
III. RESULTS
In the first 2 years, 29 of the 31 Physical teams and 20
of the 27 Virtual Teams successfully uploaded code to com-
pete. The skills these students required for the competition
included programming proficiency in ROS, swarm algorithm
design, and experimental design and testing.
Several teams developed effective new search algorithms
that significantly improved the base code. The 2017 Phys-
ical Competition winner, Southwestern Indian Polytechnic
Institute (SIPI), experimented with different search strategies
inspired by snail shells, lawnmower paths, and spirograph-
style geometric designs, ultimately deciding on a strategy
inspired by the spokes of a bicycle wheel. SIPI's algorithm
displayed the greatest flexibility, performing at a high level
across a variety of target distributions.
Figure 2c shows the Montgomery College search algo-
rithm in an environment with one large cluster of target
cubes. The zig-zag pattern shows the robots' search path
through the arena. Once the large cluster of resources
was located, some robots continued searching while others
focused on collecting from the cluster. This strategy suc-
cessfully balanced the explore versus exploit trade-off and
dramatically outperformed the other teams on nearly every
distribution of resources to win the 2017 Virtual Competition.
Engineering innovations (i.e.,
improving localization and
resource pickup and drop-off) improved scores more than
new search algorithms.
A. Assessed Student Learning Outcomes
263 students completed a comprehensive self-assessment
(29%). 44 Faculty Mentors (76%) completed the survey with
the same questions to evaluate their students' performance.
Student growth was rated on 9 Accreditation Board for En-
gineering and Technology (ABET) Criteria for Accrediting
Computing Programs, and 2 ABET Criteria for Accrediting
Engineering Programs: "An ability to design and conduct
experiments, as well as to analyze and interpret data," and,
Fig. 1: Swarmie Architecture. Software packages (shaded)
interface with hardware through a USB connection, or with
the Gazebo simulation.
joints for grasping resources. Robots sense objects in the
environment using three ultrasound range finders and a web
camera. The ultrasounds have a 3 m range. The camera has
a narrow field of view with a 1 rad arc and range of 1 m.
Orientation and positional (pose) data are provided by a
global positioning system (GPS) unit and inertial measure-
ment unit (IMU) with a magnetometer, along with wheel
encoder odometry. Computation is provided by a small on-
board computer (an Intel NUC). An Arduino Leonardo mi-
crocontroller provides the hardware interface between the on-
board computer and both the wheel and gripper actuators and
the encoders that measure wheel velocity. The robot body is
built from laser cut and 3D printed components. The battery
pack allows robots to run for 8 hours between charges, which
with a default speed of speed of 0.2 m/s results in a range
of 5.75 km. This allows a team of 6 Swarmies to search
a linear distance of 34.5 km on a single charge, nearly the
distance of a marathon that the remote-controlled Mars Rover
Opportunity took 11 years to complete.
The resources that Swarmies locate and collect are cubes
marked on all sides with AprilTags [34]. The perimeter of
the central collection zone is also marked with AprilTags.
D. Software
The basic robot software, based on the robot operating
system (ROS) [35], is publicly available4 and provided to
all teams. So teams may focus on the ISRU challenge, we
implement the low-level code that interfaces with physical
and simulated actuators and sensors, performs necessary
calibration and pre-configuration, and defines the simulated
world for competition, including Swarmies, resources, and
the search environment.
The software base is designed so that students can imple-
ment the same algorithms to control simulated or physical
robots (Figure 1). The GUI and ROS master either connect
to the physical robots through a wireless network, or run
in a Gazebo simulation. The diagnostic package monitors
hardware components and alerts the user to problems. The
4https://github.com/BCLab-UNM/Swarmathon-ROS
(a) Gazebo Simulation
(b) User Interface
(c) Example Search Pattern
Fig. 2: Gazebo and GUI. (a) Two robots bring target cubes marked with AprilTags to the collection zone. (b) Gazebo
interface showing robot status, location tracing, and camera view. (c) Trace of the Montgomery College winning swarm of
6 robots; the dense activity shows robots traveling between a large target cluster and the collection zone.
ethical, legal, security and social issues and responsibilities,
and (2) An ability to analyze the local and global impact of
computing on individuals, organizations, and society.
TABLE I: Changes in Student Abilities and Attitudes
2016
2017
ABET Outcomes
+0.35 +0.59
Higher Degree Interest +0.21 +0.53
+0.19 +0.38
Career Motivation
Across both competition years, students also reported
significant gains in their "desire or motivation to pursue
a higher degree beyond the one I am currently pursuing,"
and their "desire or motivation to pursue a career in CS, or
Robotics, or other technical discipline in STEM" as seen in
Table I. Gains were significant even among students who
already possessed a moderate to high degree of interest in
pursuing a more advanced degree.
The Swarmathon is effectively inspiring the next gener-
ation of engineers to tackle major scientific problems like
the NASA ISRU mission and to bring their diversity of
background and experience into the STEM workforce.
ACKNOWLEDGMENT
We thank Theresa Martinez, MUREP STEM Engagement
Manager; Paul Secor and Mary Baker of Secor Strategies,
LLC; Kate Cunningham and Beatriz Palacios Abad of UNM
ADVANCE; Elizabeth Esterly and the students of the Moses
Biological Computation Lab, Vanessa Svihla, Aeron Haynie,
the NASA event volunteers and UNM support staff, KSC
Visitor Complex, and the Swarmathon students and Faculty
Mentors.
REFERENCES
[1] M. Brambilla, E. Ferrante, M. Birattari, and M. Dorigo, "Swarm
robotics: a review from the swarm engineering perspective," Swarm
Intelligence, vol. 7, no. 1, pp. 1–41, 2013.
Fig. 3: Physical Competition. Students compare strategies
and cheer on Swarmies from the University of the District
of Columbia as they drop cubes in the collection zone.
"An ability to function on multidisciplinary teams."5 Students
and faculty rated student degree of competency before and
after the Swarmathon on a 4-point scale: none (1), low degree
(2), moderate degree (3), and high degree (4).
Data for each year was analyzed with a dependent-samples
t-test. Differences between the 2016 and 2017 cohorts were
analyzed with an independent samples t-test. Table I shows
the significant gains in self-assessed performance each year.
2017 gains were greater than 2016 gains in all categories and
all results are statistically significant with p < 0.05.
There were no statistically significant differences in our
assessment or survey data based on race, ethnicity, or gender,
indicating that
the Swarmathon provided a level playing
field for competition, eliminating the achievement gap noted
in other studies [30], [31]. Faculty mentor surveys closely
mirrored student self-assessments on all but
two ABET
standards. Mentors ratings were more than 0.50 points lower
than student ratings on (1) An understanding of professional,
5www.abet.org/accreditation/accreditation-criteria/
collective behaviors," Robotics and Autonomous Systems, vol. 62,
no. 7, pp. 966–975, 2014.
[25] M. Dorigo, D. Floreano, L. M. Gambardella, F. Mondada, S. Nolfi,
T. Baaboura, M. Birattari, M. Bonani, M. Brambilla, and A. Brutschy,
"Swarmanoid: a novel concept for the study of heterogeneous robotic
swarms," IEEE Robotics & Automation Magazine, vol. 20, no. 4, pp.
60–71, 2013.
[26] A. Rosenfeld, N. Agmon, O. Maksimov, and S. Kraus, "Intelligent
team collaboration," Artificial
agent supporting human-multi-robot
Intelligence, 2017.
[27] N. Jakobi, P. Husbands, and I. Harvey, "Noise and the reality gap: The
use of simulation in evolutionary robotics," in European Conference
on Artificial Life. Springer, 1995, pp. 704–720.
[28] S. Freeman, S. L. Eddy, M. McDonough, M. K. Smith, N. Okoroafor,
H. Jordt, and M. P. Wenderoth, "Active learning increases student
performance in science, engineering, and mathematics," Proceedings
of the National Academy of Sciences, vol. 111, no. 23, pp. 8410–8415,
2014.
[29] A. G. Welch, "Using the TOSRA to Assess High School Students'
Attitudes toward Science after Competing In the FIRST Robotics
Competition: An Exploratory Study." Eurasia Journal of Mathematics,
Science & Technology Education, vol. 6, no. 3, 2010.
[30] A. S. Richmond, G. A. Boysen, and R. A. R. Gurung, An Evidence-
based Guide to College and University Teaching: Developing the
Model Teacher. Routledge, 2016.
[31] C. M. Steele, Whistling Vivaldi: How Stereotypes Affect Us and What
We Can Do. New York, NY, US: WW Norton & Co, 2010.
[32] L. Sanneman, D. Ajilo, J. DelPreto, A. Mehta, S. Miyashita, N. A.
Poorheravi, C. Ramirez, S. Yim, S. Kim, and D. Rus, "A Distributed
Robot Garden System," in Robotics and Automation (ICRA), 2015
IEEE International Conference on.
IEEE, 2015, pp. 6120–6127.
[33] M. Rubenstein, B. Cimino, R. Nagpal, and J. Werfel, "AERobot: An
affordable one-robot-per-student system for early robotics education,"
in Robotics and Automation (ICRA), 2015 IEEE International Confer-
ence on.
IEEE, 2015, pp. 6107–6113.
[34] E. Olson, "AprilTag: A robust and flexible visual fiducial system," in
Robotics and Automation (ICRA), 2011 IEEE International Conference
on.
IEEE, 2011, pp. 3400–3407.
[35] M. Quigley, K. Conley, B. Gerkey, J. Faust, T. Foote, J. Leibs,
R. Wheeler, and A. Y. Ng, "ROS: an open-source Robot Operating
System," in ICRA workshop on open source software, vol. 3, no. 3.2.
Kobe, 2009, p. 5.
[36] T. Moore and D. Stouch, "A Generalized Extended Kalman Filter
Implementation for the Robot Operating System," in Proceedings of
the 13th International Conference on Intelligent Autonomous Systems
(IAS-13). Springer, 2014, pp. 335–348.
[37] R. Brooks, "A robust layered control system for a mobile robot," IEEE
journal on robotics and automation, vol. 2, no. 1, pp. 14–23, 1986.
[2] O. Zedadra, N. Jouandeau, H. Seridi, and G. Fortino, "Multi-Agent
Foraging: state-of-the-art and research challenges," Complex Adaptive
Systems Modeling, vol. 5, no. 1, p. 3, 2017.
[3] J. P. Hecker, K. Letendre, K. Stolleis, D. Washington, and M. E.
Moses, "Formica ex machina: ant swarm foraging from physical to
virtual and back again," in Swarm Intelligence. Springer, 2012, pp.
252–259.
[4] L. Steels, "Cooperation between distributed agents through self-
organisation," in Intelligent Robots and Systems' 90.'Towards a New
Frontier of Applications', Proceedings. IROS'90. IEEE International
Workshop on.
IEEE, 1990, pp. 8–14.
[5] A. Drogoul and J. Ferber, "From Tom Thumb to the Dockers: Some
experiments with foraging robots," From Animals to Animats, vol. 2,
p. 451, 1993.
[6] A. Campo and M. Dorigo, "Efficient multi-foraging in swarm
Springer, 2007, pp. 696–
robotics," in Advances in Artificial Life.
705.
[7] A. F. T. Winfield, "Towards an engineering science of robot foraging,"
Distributed autonomous robotic systems, vol. 8, pp. 185–192, 2009.
[8] M. Dorigo, M. Birattari, and M. Brambilla, "Swarm robotics," Schol-
arpedia, vol. 9, no. 1, p. 1463, 2014.
[9] A. F. T. Winfield, "Foraging Robots," in Encyclopedia of Complexity
Springer, 2009, vol. 6, ch. Foraging R, pp.
and Systems Science.
3682–3700.
[10] A. L. Griffith, "Persistence of women and minorities in STEM field
majors: Is it the school that matters?" Economics of Education Review,
vol. 29, no. 6, pp. 911–922, 2010.
[11] C. B. Leggon, "Women in science: Racial and ethnic differences
and the differences they make," The Journal of Technology Transfer,
vol. 31, no. 3, pp. 325–333, 2006.
[12] S. D. Museus, R. T. Palmer, R. J. Davis, and D. Maramba, Racial and
Ethnic Minority Student Success in STEM Education: ASHE Higher
Education Report, Volume 36, Number 6.
John Wiley & Sons, 2011.
[13] US Department of Education, "FACT SHEET: Spurring African-
American STEM Degree Completion [Press
release]," 2016.
[Online]. Available: https://www.ed.gov/news/press-releases/fact-
sheet-spurring-african-american-stem-degree-completion
[14] S. Zweben and B. Bizot, "2016 Taulbee Survey: Generation CS
Continues to Produce Record Undergrad Enrollment; Graduate Degree
Production Rises at both Masters and Doctoral Levels," Computing
Research News, vol. 29, no. 5, 2017.
[15] A. Brutschy, L. Garattoni, M. Brambilla, G. Francesca, G. Pini,
M. Dorigo, and M. Birattari, "The TAM: abstracting complex tasks in
swarm robotics research," Swarm Intelligence, vol. 9, no. 1, pp. 1–22,
2015.
[16] J. Enright and P. R. Wurman, "Optimization and Coordinated Auton-
omy in Mobile Fulfillment Systems." in Automated action planning
for autonomous mobile robots, 2011, pp. 33–38.
[17] W. Liu, A. Winfield, J. Sa, J. Chen, and L. Dou, "Strategies for energy
optimisation in a swarm of foraging robots," in International Workshop
on Swarm Robotics. Springer, 2006, pp. 14–26.
[18] N. R. Hoff, A. Sagoff, R. J. Wood, and R. Nagpal, "Two foraging algo-
rithms for robot swarms using only local communication," in Robotics
and Biomimetics (ROBIO), 2010 IEEE International Conference on.
IEEE, 2010, pp. 123–130.
[19] J. P. Hecker and M. E. Moses, "Beyond pheromones: evolving error-
tolerant, flexible, and scalable ant-inspired robot swarms," Swarm
Intelligence, vol. 9, no. 1, pp. 43–70, 2015.
[20] E. Ferrante, A. E. Turgut, E. Du´enez-Guzm´an, M. Dorigo, and
T. Wenseleers, "Evolution of self-organized task specialization in robot
swarms," PLoS computational biology, vol. 11, no. 8, p. e1004273,
2015.
[21] Q. Lu, J. P. Hecker, and M. E. Moses, "The MPFA: A Multiple-
Place Foraging Algorithm for Biologically-Inspired Robot Swarms,"
IEEE/RSJ International Conference on Intelligent Robots and Systems,
2016.
[22] G. Fricke, J. Hecker, A. Griego, L. Tran, and M. Moses, "A dis-
tributed deterministic spiral search algorithm for swarms," in IEEE
International Conference on Intelligent Robots and Systems, 2016.
[23] D. Pickem, P. Glotfelter, L. Wang, M. Mote, A. Ames, E. Feron, and
M. Egerstedt, "The Robotarium: A remotely accessible swarm robotics
research testbed," in Robotics and Automation (ICRA), 2017 IEEE
International Conference on.
IEEE, 2017, pp. 1699–1706.
[24] M. Rubenstein, C. Ahler, N. Hoff, A. Cabrera, and R. Nagpal,
"Kilobot: A low cost robot with scalable operations designed for
|
1805.08547 | 2 | 1805 | 2019-11-15T12:05:38 | Learning over Multitask Graphs -- Part II: Performance Analysis | [
"cs.MA"
] | Part I of this paper formulated a multitask optimization problem where agents in the network have individual objectives to meet, or individual parameter vectors to estimate, subject to a smoothness condition over the graph. A diffusion strategy was devised that responds to streaming data and employs stochastic approximations in place of actual gradient vectors, which are generally unavailable. The approach relied on minimizing a global cost consisting of the aggregate sum of individual costs regularized by a term that promotes smoothness. We examined the first-order, the second-order, and the fourth-order stability of the multitask learning algorithm. The results identified conditions on the step-size parameter, regularization strength, and data characteristics in order to ensure stability. This Part II examines steady-state performance of the strategy. The results reveal explicitly the influence of the network topology and the regularization strength on the network performance and provide insights into the design of effective multitask strategies for distributed inference over networks. | cs.MA | cs | Learning over Multitask Graphs --
Part II: Performance Analysis
Roula Nassif, Member, IEEE, Stefan Vlaski, Member, IEEE,
C´edric Richard, Senior Member, IEEE, Ali H. Sayed, Fellow Member, IEEE
Abstract
Part I of this paper formulated a multitask optimization problem where agents in the network have individual
objectives to meet, or individual parameter vectors to estimate, subject to a smoothness condition over the graph. A
diffusion strategy was devised that responds to streaming data and employs stochastic approximations in place of
actual gradient vectors, which are generally unavailable. The approach relied on minimizing a global cost consisting
of the aggregate sum of individual costs regularized by a term that promotes smoothness. We examined the first-
order, the second-order, and the fourth-order stability of the multitask learning algorithm. The results identified
conditions on the step-size parameter, regularization strength, and data characteristics in order to ensure stability.
This Part II examines steady-state performance of the strategy. The results reveal explicitly the influence of the
network topology and the regularization strength on the network performance and provide insights into the design
of effective multitask strategies for distributed inference over networks.
Index Terms
Multitask distributed inference, diffusion strategy, smoothness prior, graph Laplacian regularization, gradient
noise, steady-state performance.
9
1
0
2
v
o
N
5
1
]
A
M
.
s
c
[
2
v
7
4
5
8
0
.
5
0
8
1
:
v
i
X
r
a
The work of A. H. Sayed was supported in part by NSF grants CCF-1524250 and ECCS-1407712. A short version of this work appeared
in the conference publication [1].
This work was done while R. Nassif was a post-doc at EPFL. She is now with the American University of Beirut, Lebanon (e-
mail: [email protected]). S. Vlaski and A. H. Sayed are with Institute of Electrical Engineering, EPFL, Switzerland (e-mail:
stefan.vlaski,[email protected]). C. Richard is with Universit´e de Nice Sophia-Antipolis, France (e-mail: [email protected]).
2
I. INTRODUCTION
As pointed out in Part I [2] of this work, most prior literature on distributed inference over networks focuses
on single-task problems, where agents with separable objective functions need to agree on a common parameter
vector corresponding to the minimizer of an aggregate sum of individual costs [3] -- [13]. In this paper, and its
accompanying Part I [2], we focus instead on multitask networks where the agents may need to estimate and track
multiple objectives simultaneously [14] -- [26]. Although agents may generally have distinct though related tasks
to perform, they may still be able to capitalize on inductive transfer between them to improve their performance.
Based on the type of prior information that may be available about how the tasks are related to each other, multitask
learning algorithms can be derived by translating the prior information into constraints on the parameter vectors to
be inferred.
In Part I [2], we considered multitask inference problems where each agent in the network seeks to minimize
an individual cost expressed as the expectation of some loss function. The minimizers of the individual costs are
assumed to vary smoothly over the topology, as dictated by the graph Laplacian matrix. The smoothness property
softens the transitions in the tasks among adjacent nodes and allows incorporating information about the graph
structure into the solution of the inference problem. In order to exploit the smoothness prior, we formulated the
inference problem as the minimization of the aggregate sum of individual costs regularized by a term promoting
smoothness, known as the graph-Laplacian regularizer [27], [28]. A diffusion strategy was proposed that responds
to streaming data and employs stochastic approximations in place of actual gradient vectors, which are generally
unavailable.
The analysis from Part I [2] revealed how the regularization strength η can steer the convergence point of the
network toward many modes of operation starting from the non-cooperative mode (η = 0) where each agent
converges to the minimizer of its individual cost and ending with the single-task mode (η → ∞) where all agents
converge to a common parameter vector corresponding to the minimizer of the aggregate sum of individual costs.
For any values of η in the range 0 < η < ∞, the network behaves in a multitask mode where agents seek their
individual models while at the same time ensuring that these models satisfy certain smoothness and closeness
conditions dictated by the value of η. We carried out in Part I [2] a detailed stability analysis of the proposed
strategy. We showed, under conditions on the step-size learning parameter µ, that the adaptive strategy induces a
contraction mapping and that despite gradient noise, it is able to converge in the mean-square-error sense within
O(µ) from the solution of the regularized problem, for sufficiently small µ. We also established the first and fourth-
order moments stability of the network error process and showed that they tend asymptotically to bounded region
on the order of O(µ) and O(µ2), respectively.
Based on the results established in Part I [2], we shall derive in this paper a closed-form expression for the
steady-state network mean-square-error relative to the minimizer of the regularized cost. This closed form expres-
sion will reveal explicitly the influence of the regularization strength, network topology (through the eigenvalues
3
and eigenvectors of the Laplacian matrix), gradient noise, and data characteristics, on the network performance.
Additionally, a closed-form expression for the steady-state network mean-square-error relative to the minimizers
of the individual costs is also derived. This expression will provide insights on the design of effective multitask
strategies for distributed inference over networks.
Notation. We adopt the same notation from Part I [2]. All vectors are column vectors. Random quantities are
denoted in boldface. Matrices are denoted in capital letters while vectors and scalars are denoted in lower-case
letters. The operator (cid:22) denotes an element-wise inequality; i.e., a (cid:22) b implies that each entry of the vector a is
less than or equal to the corresponding entry of b. The symbol diag{·} forms a matrix from block arguments by
placing each block immediately below and to the right of its predecessor. The operator col{·} stacks the column
vector entries on top of each other. The symbols ⊗ and ⊗b denote the Kronecker product and the block Kronecker
product, respectively. The symbol vec(·) refers to the standard vectorization operator that stacks the columns of a
matrix on top of each other and the symbol bvec(·) refers to the block vectorization operation that vectorizes each
block and stacks the vectors on top of each other.
II. DISTRIBUTED INFERENCE UNDER SMOOTHNESS PRIORS
A. Problem formulation and adaptive strategy
degree matrix D is a diagonal matrix with k-th entry [D]kk =(cid:80)N
Consider a connected network (or graph) G = {N ,E, A}, where N is a set of N agents (nodes), E is a set of
edges connecting agents with particular relations, and A is a symmetric, weighted adjacency matrix. If there is an
edge connecting agents k and (cid:96), then [A]k(cid:96) = ak(cid:96) > 0 reflects the strength of the relation between k and (cid:96); otherwise,
[A]k(cid:96) = 0. We introduce the graph Laplacian L, which is a differential operator defined as L = D − A, where the
(cid:96)=1 ak(cid:96). Since L is symmetric positive semi-definite,
it possesses a complete set of orthonormal eigenvectors. We denote them by {v1, . . . , vN}. For convenience, we
order the set of real, non-negative eigenvalues of L as 0 = λ1 < λ2 ≤ . . . ≤ λN = λmax(L), where, since the
network is connected, there is only one zero eigenvalue with corresponding eigenvector v1 = 1√
1N [29]. Thus,
the Laplacian can be decomposed as:
N
L = V ΛV (cid:62),
(1)
where Λ = diag{λ1, . . . , λN} and V = [v1, . . . , vN ].
Let wk ∈ RM denote some parameter vector at agent k and let W = col{w1, . . . , wN} denote the collection of
parameter vectors from across the network. We associate with each agent k a risk function Jk(wk) : RM → R
assumed to be strongly convex. In most learning and adaptation problems, the risk function is expressed as the
expectation of a loss function Qk(·) and is written as Jk(wk) = E Qk(wk; xk), where xk denotes the random data.
The expectation is computed over the distribution of this data. We denote the unique minimizer of Jk(wk) by wo
k.
Let us recall the assumption on the risks {Jk(wk)} used in Part I [2].
4
Assumption 1. (Strong convexity) It is assumed that the individual costs Jk(wk) are each twice differentiable and
strongly convex such that the Hessian matrix function Hk(wk) = ∇2
Jk(wk) is uniformly bounded from below
and above, say, as:
wk
0 < λk,minIM ≤ Hk(wk) ≤ λk,maxIM ,
(2)
where λk,min > 0 for k = 1, . . . , N.
In many situations, there is prior information available about Wo = col{wo
N}. In the current Part II, and
its accompanying Part I [2], the prior belief we want to enforce is that the target signal Wo is smooth with respect
to the underlying weighted graph. References [16], [17] provide variations for such problems for the special case
of mean-square-error costs. Here we treat general convex costs. Let L = L ⊗ IM . The smoothness of W can be
measured in terms of a quadratic form of the graph Laplacian [27], [28], [30] -- [32]:
1, . . . , wo
S(W) = W(cid:62)LW =
1
2
ak(cid:96)(cid:107)wk − w(cid:96)(cid:107)2,
(3)
where Nk is the set of neighbors of k, i.e., the set of nodes connected to agent k by an edge. The smaller S(W)
is, the smoother the signal W on the graph is. Intuitively, given that the weights are non-negative, S(W) shows that
W is considered to be smooth if nodes with a large ak(cid:96) on the edge connecting them have similar weight values
{wk, w(cid:96)}. Our objective is to devise and study a strategy that solves the following regularized problem:
N(cid:88)
(cid:88)
k=1
(cid:96)∈Nk
N(cid:88)
Jk(wk) +
Wo
η = arg minW J glob(W) =
(4)
N,η}.
in a distributed manner where each agent is interested in estimating the k-th sub-vector of Wo
The tuning parameter η ≥ 0 controls the trade-off between the two components of the objective function. We are
particularly interested in solving the problem in the stochastic setting when the distribution of the data xk in
Jk(wk) = E Qk(wk; xk) is generally unknown. This means that the risks Jk(wk) and their gradients ∇wkJk(wk)
are unknown. As such, approximate gradient vectors need to be employed. A common construction in stochastic
η = col{wo
1,η, . . . , wo
k=1
W(cid:62)LW,
η
2
approximation theory is to employ the following approximation at iteration i:
(cid:92)∇wkJk(wk) = ∇wkQk(wk; xk,i),
(5)
where xk,i represents the data observed at iteration i. The difference between the true gradient and its approximation
is called the gradient noise sk,i(·):
sk,i(w) (cid:44) ∇wkJk(w) − (cid:92)∇wkJk(w).
(6)
Let wk,i denote the estimate of wo
adaptive manner, we proposed in Part I [2] the following diffusion-type algorithm:
k,η at iteration i and node k. In order to solve (4) in a fully distributed and
ψk,i = wk,i−1 − µ (cid:92)∇wkJk(wk,i−1)
wk,i = ψk,i − µη
(cid:88)
ak(cid:96)(ψk,i − ψ(cid:96),i),
(cid:96)∈Nk
(7)
where µ > 0 is a small step-size parameter and ψk,i is an intermediate variable.
5
B. Summary of main results
One key insight that followed from the analysis in Part I [2] is that the smoothing parameter η can be regarded
as an effective tuning parameter that controls the nature of the learning process. The value of η can vary from
η = 0 to η → ∞. We showed that at one end, when η = 0, the learning algorithm reduces to a non-cooperative
mode of operation where each agent acts individually and estimates its own local model, wo
k. On the other hand,
when η → ∞, the learning algorithm moves to a single-mode of operation where all agents cooperate to estimate
a single parameter (namely, the Pareto solution of the aggregate cost function). For any values of η in the range
0 < η < ∞, the network behaves in a multitask mode where agents seek their individual models while at the same
time ensuring that these models satisfy certain smoothness and closeness conditions dictated by the value of η.
In Part I [2], we carried out a detailed stability analysis of the proposed strategy (7). We showed, under some
conditions on the step-size parameter µ, that:
(cid:107)E(Wo
E(cid:107)Wo
E(cid:107)Wo
η − Wi)(cid:107) = O(µ),
η − Wi(cid:107)2 = O(µ),
η − Wi(cid:107)4 = O(µ2),
lim sup
i→∞
lim sup
i→∞
lim sup
i→∞
(see [2, Theorem 4])
(see [2, Theorem 2])
(see [2, Theorem 3])
(8)
(9)
(10)
η is the solution of the regularized problem (4) and Wi = col{w1,i, . . . , wN,i} denotes the network block
where Wo
weight vector at iteration i. Expression (9) indicates that the mean-square error E(cid:107)Wo
η − Wi(cid:107)2 is on the order of µ.
However, in this Part II, we are interested in characterizing how close the Wi gets to the network limit point Wo
η.
In particular, we will be able to characterize the network mean-square deviation (MSD) (defined below in (51))
value in terms of the step-size µ, the regularization strength η, the network topology (captured by the eigenvalues
λm and eigenvectors vm of the Laplacian L), and the data characteristics (captured by the second-order properties
of the costs Hk,η and second-order moments of the gradient noise Rs,k,η) as follows:
Tr
[vm]2
kHk,η + ηλmI
[vm]2
kRs,k,η
m=1
k=1
k=1
O(µ)
(O(1) + O(η))
,
(11)
where [vm]k denotes the k-th entry of the eigenvector vm. The interpretation of (11) is explained in more detail
(O(1)+O(η)) will
in Section IV where it is shown, by coupling η and µ in an appropriate manner, that the term
be a strictly higher order term of µ. As we will explain later in Sections IV and V, by properly setting the
O(µ)
parameters, expression (11) allows us to recover the mean-square-deviation of stand-alone adaptive agents (η = 0)
and single-task diffusion networks (η → ∞).
Recall that the objective of the multitask strategy (7) is to exploit similarities among neighboring agents in an
attempt to improve the overall network performance in approaching the collection of individual minimizer Wo by
means of local communications. Section V in this paper is devoted to quantify the benefit of cooperation, namely,
the objective of improving the mean-square deviation around the limiting point of the algorithm. In particular, we
N(cid:88)
(cid:32) N(cid:88)
MSD =
µ
2N
(cid:33)−1(cid:32) N(cid:88)
(cid:33) +
6
will be able to characterize the mean-square-deviation (MSD) value relative to the multitask objective Wo in terms
of the MSD in (11) and the mismatch Wo
MSD = MSD(cid:124)(cid:123)(cid:122)(cid:125)
O(µ),η
η − Wo as follows:
(cid:123)(cid:122)
(cid:125)
η − Wo(cid:107)2
+(cid:107)Wo
(cid:124)
smoothness,η
+2(Wo
η − Wo)(cid:62) bias(cid:124)(cid:123)(cid:122)(cid:125)
O(µ),η
,
(12)
where "bias" is the bias of algorithm (7) relative to Wo
η given in future expression (40). By increasing η, the MSD
η−Wo(cid:107)2 is more
in (11) is more likely to decrease. However, by increasing η, from expression (31) in Part I [2], (cid:107)Wo
likely to increase and the size of this increase is determined by the smoothness of Wo. From future Lemma 2, it
turns out that the third term on the RHS in (12) is a function of µ, η, and the smoothness of the multitask objective
Wo. By increasing η, this term is more likely to increase. The key conclusion will be that, while the second and
third terms on the RHS in (12) will in general increase as the regularization strength η increases, the size of this
increase is determined by the smoothness of Wo which is in turn function of the network topology captured by L.
The more similar the tasks at neighboring agents are, the smaller these terms will be. This implies that as long
as Wo is sufficiently smooth, moderate regularization strengths η in the range ]0,∞[ exist such that MSD at these
values of η will be less than MSD at η = 0 which corresponds to the non-cooperative mode of operation. The best
choice for η would be the one minimizing MSD in (12). We refer the reader to Fig. 2 in [2, Section II-B] for an
illustration of this concept of multitask learning benefit. This example will be considered further in the numerical
experiments section.
C. Modeling Assumptions from Part I [2]
In this section, we recall the assumptions used in Part I [2] to establish the network mean-square error stability (9).
Assumption 2. (Gradient noise process) The gradient noise process defined in (6) satisfies for any w ∈ F i−1 and
for all k, (cid:96) = 1, 2, . . . , N:
E[sk,i(w)F i−1] = 0,
E[(cid:107)sk,i(w)(cid:107)2F i−1] ≤ β2
(cid:96),i(w)F i−1] = 0,
E[sk,i(w)s(cid:62)
k(cid:107)w(cid:107)2 + σ2
s,k,
k (cid:54)= (cid:96),
k ≥ 0, σ2
(15)
s,k ≥ 0, and where F i−1 denotes the filtration generated by the random processes {w(cid:96),j} for
for some β2
all (cid:96) = 1, . . . , N and j ≤ i − 1.
Let us introduce the network block vector Wi = col{w1,i, . . . , wN,i}. Recall from Part I [2] that at each iteration,
we can view (7) as a mapping from Wi−1 to Wi:
(13)
(14)
(16)
Wi = (IM N − µηL)
Wi−1 − µ col
(cid:18)
(cid:111)N
(cid:110) (cid:92)∇wkJk(wk,i−1)
k=1
(cid:19)
We introduced the following condition on the combination matrix (IM N − µηL).
Assumption 3. (Combination matrix) The symmetric combination matrix (IM N − µηL) has nonnegative entries
and its spectral radius is equal to one. Since L has an eigenvalue at zero, these conditions are satisfied when the
step-size µ > 0 and the regularization strength η ≥ 0 satisfy:
7
0 ≤ µη ≤
2
λmax(L)
0 ≤ µη ≤ min
1≤k≤N
,
(cid:40)
1(cid:80)N
(cid:96)=1 ak(cid:96)
(cid:41)
,
(17)
(18)
where condition (17) ensures stability and condition (18) ensures non-negative entries.
The results in Part I [2] established that the iterates wk,i converge in the mean-square-error sense to a small
O(µ)− neighborhood around the regularized solution wo
k,η. In this part of the work, we will be more precise and
determine the size of this neighborhood, i.e., assess the size of the constant multiplying µ in the O(µ)−term. To do
so, we shall derive an accurate first-order expression for the mean-square error (9); the expression will be accurate
to first-order in µ.
To arrive at the desired expression, we first need to introduce a long-term approximation model and assess how
close it is to the actual model. We then derive the performance for the long-term model and use this closeness to
transform this result into an accurate expression for the performance of the original learning algorithm. To derive
the long-term model, we follow the approach developed in [9]. The first step is to establish the asymptotic stability
η − Wi(cid:107)4, which has already been done in Part I [2]. This
of the fourth-order moment of the error vector, E(cid:107)Wo
property is needed to justify the validity of the long-term approximate model. Recall that to establish the fourth-
order stability, we replaced condition (14) on the gradient noise process by the following condition on its fourth
order moment:
E(cid:2)(cid:107)sk,i(wk)(cid:107)4F i−1
(cid:3) ≤ β
4
k(cid:107)wk(cid:107)4 + σ4
s,k,
(19)
for some β
4
k ≥ 0, and σ4
s,k ≥ 0. As explained in [9], condition (19) implies (14).
To establish the mean-stability (8), we introduced a smoothness condition on the Hessian matrices of the individual
costs. This smoothness condition will be adopted in the next section when we study the long term behavior of the
network.
Assumption 4. (Smoothness condition on individual cost functions). It is assumed that each Jk(wk) satisfies a
k,η, in that the corresponding Hessian matrix is Lipchitz continuous in the proximity
smoothness condition close to wo
of wo
k,η with some parameter κd ≥ 0, i.e.,
(cid:107)∇2
wk
for small perturbations (cid:107)∆wk(cid:107) ≤ .
Jk(wo
k,η + ∆wk) − ∇2
wk
Jk(wo
k,η)(cid:107) ≤ κd(cid:107)∆wk(cid:107),
(20)
8
Let (cid:101)Wi = Wo
we obtain:
III. LONG-TERM NETWORK DYNAMICS
η − Wi. Subtracting the vector (IM N − µηL)Wo
(cid:101)Wi − µηLWo
(cid:16)(cid:101)Wi−1 + µ col{∇wkJk(wk,i−1) − sk,i(wk,i−1)}N
η = (IM N − µηL)
(cid:17)
η from both sides of recursion (16), and using (6),
,
(21)
k=1
From the mean-value theorem [33, pp. 24], [9, Appendix D], we have:
where
∇wkJk(wk,i−1) = ∇wkJk(wo
k,η) − H k,i−1(wo
k,η − wk,i−1),
H k,i−1 (cid:44)
Jk(wo
k,η − t(wo
k,η − wk,i−1))dt,
∇2
(cid:90) 1
col(cid:8)∇wkJk(wo
wk
0
k,η)(cid:9)N
and from the optimality condition of (4), we have:
Replacing (22) into (21) and using (24), we arrive at the following recursion for (cid:101)Wi:
(cid:101)Wi = (IM N − µηL)(IM N − µHi−1)(cid:101)Wi−1 − µ(IM N − µηL)si(Wi−1) + µ2η2L2Wo
k=1
η,
= −ηLWo
η.
(22)
(23)
(24)
(25)
(26)
(27)
where
si(Wi−1) (cid:44) col{sk,i(wk,i−1)}N
k=1,
Hi−1 (cid:44) diag{H k,i−1}N
k=1.
We move on to motivate a long-term model for the evolution of the network error dynamics, (cid:101)Wi, after sufficient
iterations, i.e., for i (cid:29) 1. We examine the stability property of the model, the proximity of its trajectory to that of
the original network dynamics (25), and subsequently employ the model to assess network performance.
A. Long-term error model
We introduce the error matrix (cid:101)Hi−1 (cid:44) Hη − Hi−1, which measures the deviation of Hi−1 from the constant
matrix:
Hη (cid:44) diag{Hk,η}N
k=1,
with each Hk,η given by the value of the Hessian matrix at the regularized solution, namely, Hk,η (cid:44) ∇2
Let
wk
Then, we can write:
Bi−1 (cid:44) (IM N − µηL)(IM N − µHi−1),
Bη (cid:44) (IM N − µηL)(IM N − µHη).
Bi−1 = Bη + µ(IM N − µηL)(cid:101)Hi−1.
(28)
Jk(wo
k,η).
(29)
(30)
(31)
Using (31), we can rewrite the error recursion (25) as:
(cid:101)Wi = Bη(cid:101)Wi−1 − µ(IM N − µηL)si(Wi−1) + µ2η2L2Wo
η + µ(IM N − µηL)ci−1,
in terms of the random perturbation sequence:
ci−1 (cid:44) (cid:101)Hi−1(cid:101)Wi−1.
9
(32)
(33)
Under Assumptions 1 and 4, and for small µ, it can be shown that lim supi→∞ E(cid:107)ci−1(cid:107) = O(µ), and that (cid:107)ci−1(cid:107) =
O(µ) asymptotically with high probability (see Appendix A). Motivated by this result, we introduce the following
approximate model, where the last term involving ci−1 in (32), which is O(µ2), is removed:
(cid:101)W(cid:48)
i = Bη(cid:101)W(cid:48)
i−1 − µ(IM N − µηL)si(Wi−1) + µ2η2L2Wo
η,
i (cid:29) 1.
(34)
Obviously, the iterates that are generated by (34) are generally different from the iterates generated by the original
recursion (25). To highlight this fact, we are using the prime notation for the state of the long-term model. Note
that the driving process si(Wi−1) in (34) is the same gradient noise process from the original recursion (25) and
is evaluated at Wi−1. In the following, we show that, after sufficient iterations i (cid:29) 1, the error dynamics of the
network relative to the solution Wo
η is well-approximated by the model (34).
B. Size of Approximation Error
We start by showing that the mean-square difference between the trajectories {(cid:101)Wi, (cid:101)W(cid:48)
i} is asymptotically bounded
2 ) from the
3
by O(µ2) and that the mean-square error performance of the long term model (34) is within O(µ
performance of the original recursion (25). Working with recursion (34) is much more tractable for performance
analysis because its dynamics is driven by the constant matrix Bη as opposed to the random matrix Bi−1 in the
original error recursion (25). Therefore, we shall work with the long-term model (34) and evaluate its performance,
which will provide an accurate representation for the performance of the original distributed strategy (7) to first
order in the step-size µ.
Lemma 1. (Size of approximation error) Under Assumptions 1, 2, 3, and 4, and condition (19), it holds that:
E(cid:107)(cid:101)Wi − (cid:101)W(cid:48)
lim sup
i→∞
i(cid:107)2 = O(µ2),
E(cid:107)(cid:101)Wi(cid:107)2 = lim sup
i→∞
E(cid:107)(cid:101)W(cid:48)
i(cid:107)2 + O(µ
3
2 ).
(35)
(36)
lim sup
i→∞
Proof. See Appendix B.
We shall discuss now the mean and mean-square error stability of the long-term approximate model (34).
10
C. Stability of First-Order Error Moment
Conditioning both sides of (34), invoking the conditions on the gradient noise from Assumption 2, and computing
the conditional expectations, we obtain:
Taking expectation again, we arrive at:
iF i−1] = Bη(cid:101)W(cid:48)
E[(cid:101)W(cid:48)
E(cid:101)W(cid:48)
i = BηE(cid:101)W(cid:48)
i−1 + µ2η2L2Wo
η.
i−1 + µ2η2L2Wo
η.
(37)
(38)
(39)
(40)
The above recursion is stable if the matrix Bη in (30) is stable. This matrix has a similar form to the matrix
(IM N − µηL)(IM N − µH∞) encountered in Part I [2, Section III-A2]. Similarly, it can be verified that Bη is stable
when condition (17) and condition
(cid:26) 2
(cid:27)
.
0 < µ < min
1≤k≤N
λk,max
are satisfied. In this case, we obtain
(cid:101)W(cid:48)
∞ (cid:44) lim
i→∞
E(cid:101)W(cid:48)
i = µ2η2(IM N − (IM N − µηL)(IM N − µHη))−1L2Wo
η,
where the RHS in the above expression is similar to the RHS in equation (49) [2] with H∞ replaced by Hη.
Lemma 2. (Mean stability of long-term model) Under Assumptions 1, 2, 3, and 4, and for sufficiently small µ, the
steady-state bias (cid:101)W(cid:48)
∞ = limi→∞ E(cid:101)W(cid:48)
i of the long-term model (34) given by (40) satisfies:
(cid:18) 1
µ
(cid:19)
i→∞(cid:107)E(cid:101)W(cid:48)
lim
i(cid:107)
µ lim
µ→0
≤ µ
O(η2)
(O(1) + O(η))2 .
(41)
Proof. The proof is similar to the proof of Theorem 1 in Part I [2] with H∞ replaced by Hη.
D. Stability of Second-Order Error Moment
that E(cid:107)(cid:101)w
In the following, we show that the long term approximate model (34) is also mean-square stable in the sense
(cid:48)
k,i(cid:107)2 tends asymptotically to a region that is bounded by O(µ). We follow the same line of reasoning as
in Part I [2, Section III-A] where we studied the mean-square stability of the original model (25). Based on the
inequality:
lim sup
i→∞
η − W(cid:48)∞ = (cid:101)W(cid:48)
where Wo
the recursion:
E(cid:107)Wo
η − W(cid:48)
i(cid:107)2 = lim sup
i→∞
≤ 2(cid:107)Wo
η − W(cid:48)
E(cid:107)Wo
η − W(cid:48)
∞ + W(cid:48)
∞(cid:107)2 + 2 lim sup
i→∞
∞ − W(cid:48)
E(cid:107)W(cid:48)
i(cid:107)2
∞ − W(cid:48)
i(cid:107)2,
(42)
∞ is the steady-state bias of the long term model given by (40) and where W(cid:48)∞ − W(cid:48)
i follows
∞ − W(cid:48)
W(cid:48)
i = Bη(W(cid:48)
∞ − W(cid:48)
i−1) − µ(IM N − µηL)si(Wi−1),
(43)
11
(cid:44) col(cid:8)E(cid:107)w(cid:48)
MSP(cid:48)
k,i(cid:107)2(cid:9)N
k,∞ − w(cid:48)
and from Theorems 1 and 2 in Part I [2] and previous Lemma 2, we can establish the mean-square stability of (34).
Let us introduce the mean-square perturbation vector (MSP(cid:48)) at time i relative to W(cid:48)∞:
(44)
Lemma 3. (Mean-square stability of the long-term model) Under Assumptions 1, 2, 3, and 4, the MSP(cid:48) at time i
can be recursively bounded as:
k=1
.
i
MSP(cid:48)
i (cid:22) (IN − µηL)(G(cid:48)(cid:48))2MSP(cid:48)
i−1 + 3µ2(IN − µηL)diag{β2
k}N
k=1MSPi−1 + µ2(IN − µηL)b,
where:
G(cid:48)(cid:48) (cid:44) diag{γk}N
b (cid:44) col(cid:8)σ2
k=1 ,
k(cid:107)wo
γk (cid:44) max{1 − µλk,min,1 − µλk,max}.
k,η(cid:107)2 + 3β2
s,k + 3β2
k(cid:107)wo
k,η − wk,∞(cid:107)2(cid:9)N
,
k=1
(45)
(46)
(47)
(48)
and MSPi is the mean-square perturbation vector at time i relative to the fixed point W∞ = col{wk,∞}N
k=1 of
algorithm (7) in the absence of gradient noise (see [2, Section III-A3]). A sufficiently small µ ensures the stability
of the above recursion. It follows that
(cid:107) lim sup
i→∞
MSP(cid:48)
i(cid:107)∞ = O(µ),
E(cid:107)(cid:101)W(cid:48)
i(cid:107)2 = O(µ) +
lim sup
i→∞
O(µ2η4)
(O(1) + O(η))4 = O(µ).
(49)
(50)
and that
Proof. See Appendix C.
IV. MEAN-SQUARE-ERROR PERFORMANCE
We established in Theorem 2 in Part I [2] that a network running strategy (7) is mean-square-error stable for
η − Wi(cid:107)2 = O(µ). In the following, we assess
sufficiently small µ. Specifically, we showed that lim supi→∞ E(cid:107)Wo
the size of the network mean-square-deviation (MSD) using the definition [9, Chapter 11]:
(cid:18)
(cid:18) 1
N
(cid:19)(cid:19)
MSD (cid:44) µ lim
µ→0
lim sup
i→∞
E
1
µ
(cid:107)Wo
η − Wi(cid:107)2
.
(51)
In addition to Assumptions 1, 2, 3, 4, and condition (19) on the individual costs, Jk(wk), the gradient noise process,
sk,i(wk), and the combination matrix, IN − µηL, we introduce a smoothness condition on the noise covariance
matrices.
For any wk ∈ F i−1, we let
Rs,k,i(wk) (cid:44) E[sk,i(wk)s(cid:62)
k,i(wk)F i−1]
(52)
denote the conditional second-order moment of the gradient noise process, which generally depends on i because
the statistical distribution of sk,i(wk) can be iteration-dependent, and is random since it depends on the random
12
iterate wk. We assume that, in the limit, this covariance matrix tends to a constant value when evaluated at wo
k,η
and we denote the limit by:
Rs,k,η (cid:44) lim
i→∞
E[sk,i(wo
k,η)s(cid:62)
k,i(wo
k,η)F i−1].
(53)
Assumption 5. (Smoothness condition on the noise covariance) It is assumed that the conditional second-order
moment of the noise process is locally Lipschitz continuous in a small neighborhood around wo
k,η, namely,
(cid:107)Rs,k,i(wo
k,η + ∆wk) − Rs,k,i(wo
k,η)(cid:107) ≤ κd(cid:107)∆wk(cid:107)θ,
(54)
for small perturbations (cid:107)∆wk(cid:107) ≤ , and for some constant κd ≥ 0 and exponent 0 < θ ≤ 4.
One useful conclusion that follows from Assumption 5 is that, after sufficient iterations, we can express the
covariance matrix of the gradient noise process, sk,i(wk), in terms of the limiting matrix Rs,k,η defined in (53).
Specifically, following the same proof used to establish Lemma 11.1 in [9], we can show that under the smoothness
condition and for small step-size, the covariance matrix of the gradient noise process, sk,i(wk,i−1), at each agent
k satisfies for i (cid:29) 1:
(cid:16)
2}(cid:17)
µmin{1, θ
Esk,i(wk,i−1)s(cid:62)
k,i(wk,i−1) = Rs,k,η + O
.
(55)
For clarity of presentation, we sketch the proof in Appendix D where we used results from Theorems 2 and 3 in
Part I [2].
Before studying the steady-state network performance, we establish some properties of the matrix:
Fη (cid:44) B(cid:62)
η ⊗b B(cid:62)
η ,
(56)
which is defined in terms of the block Kronecker operation using blocks of size M × M. In the derivation that
follows, we shall use the block Kronecker product ⊗b operator [34] and the block vectorization operator bvec(·).
As explained in [9], these operations preserve the locality of the blocks in the original matrix arguments. Since
ρ(Fη) = (ρ(Bη))2, the matrix Fη is stable under conditions (17) and (39). This matrix plays a critical role in
characterizing the performance of the distributed multitask algorithm. In our derivations, the matrix Fη will also
appear transformed under the orthonormal transformation:
F η (cid:44) (V ⊗b V)(cid:62)Fη(V ⊗b V),
where V (cid:44) V ⊗ IM .
Lemma 4. (Coefficient matrix Fη) For sufficiently small step-size, it holds that
(I − Fη)−1 = O(µ−1),
and
(I − F η)−1 = X−1 + W
O(1)
0
+ µ−1 ·
(O(1) + O(η))−1
(O(1) + O(η))−1
(O(1) + O(η))−1
(O(1) + O(η))−2
0
(O(1) + O(η))−1
= µ−1 ·
(57)
(58)
(59)
(60)
where X is an N 2 × N 2 block diagonal matrix with each block of dimension M 2 × M 2:
X = µ · diag
diag{(1 − µηλm)(1 − µηλp)(Hmm ⊕ Hpp) + η(λm + λp − µηλmλp)IM 2}N
p=1
(cid:110)
with ⊕ denoting the Kronecker sum operator [35]:
Hmm ⊕ Hpp = Hmm ⊗ IM + IM ⊗ Hpp,
m ⊗ IM )Hη(vn ⊗ IM ).
Hmn (cid:44) (v(cid:62)
13
(cid:111)N
m=1
,
(61)
(62)
(63)
The matrix W is an N 2 × N 2 block matrix arising from the matrices {Hmnm (cid:54)= n}. Moreover, we can also write:
(I − Fη)−1 = (V ⊗b V)X−1(V ⊗b V)(cid:62) + µ−1(O(1) + O(η))−1.
(64)
Proof. See Appendix E.
As we shall see in Theorem 1, it turns out that the decomposition in (59) is very useful to highlight some important
facts arising in the steady-state performance of the multitask algorithm.
Lemma 5. (Steady-state network performance) Under Assumptions 1, 2, 3, 4, and 5, and condition (19), it holds
that:
lim sup
i→∞
1
N
E(cid:107)Wo
η − Wi(cid:107)2 =
=
1
N
1
N
∞(cid:88)
(cid:16)Bn
Tr
ηY(B(cid:62)
η )n(cid:17)
+ O(µ1+θm)
n=0
(bvec(Y(cid:62)))(cid:62)(I − Fη)−1bvec (IM N ) + O(µ1+θm),
where θm = 1
2 min{1, θ}, Bη and Fη are defined in (30) and (56), and
Y (cid:44) µ2(IM N − µηL)Sη(IM N − µηL),
Sη (cid:44) diag{Rs,k,η}N
k=1 ,
(65)
(66)
(67)
Proof. The proof is a direct extension of the arguments used to establish Theorem 11.2 in [9] for single-task
diffusion adaptation. See Appendix F.
Theorem 1. (Network MSD performance) Under the same conditions of Lemma 5, it holds from Lemma 4 that
the steady-state network MSD defined in (51) can be written as:
N(cid:88)
(cid:32) N(cid:88)
(cid:33)−1(cid:32) N(cid:88)
(cid:33) +
MSD =
µ
2N
Tr
[vm]2
kHk,η + ηλmI
[vm]2
kRs,k,η
m=1
k=1
k=1
O(µ)
(O(1) + O(η))
.
(68)
Proof. See Appendix G.
As the derivation in Appendix G reveals, the second term on the RHS of (68) results from the matrix W in (59)
which is zero when {Hmn = 0, m (cid:54)= n}. When the Hessian matrices are uniform across the agents:
Hk,η ≡ Hη,
k = 1, . . . , N,
(69)
14
we have Hmn = 0 for m (cid:54)= n. In this case, the network MSD in (68) simplifies to:
(cid:32)
N(cid:88)
(cid:32) N(cid:88)
−1
(cid:33)(cid:33)
MSD =
µ
2N
Tr
(Hη + ηλmI)
[vm]2
kRs,k,η
.
(70)
m=1
k=1
Moreover, in the non-uniform Hessian matrices scenario, by letting η = µ− with > 0 chosen such that
Assumption 3 is satisfied, we obtain:
O(µ)
(O(1) + O(η))
= O(µ1+).
(71)
In this case, the first term on the RHS of (68) dominates the factor O(µ1+) and when we evaluate the network
MSD according to definition (51), the last term on the RHS of (68) disappears when computing the limit as µ → 0.
As we will see by simulations, the first term on the RHS of (68) provides a good approximation for the network
MSD for any η ≥ 0.
The first term on the RHS of (68) reveals explicitly the influence of the step-size µ, regularization strength η,
network topology (through the eigenvalues λm and eigenvectors vm of the Laplacian), gradient noise (through the
covariance matrices Rs,k,η), and data characteristics (through the Hessian matrices Hk,η) on the network MSD
performance. Observe that this term consists of the sum of N individual terms, each associated with an eigenvalue
λm of the Laplacian matrix, and given by:
(cid:16)(cid:0)H m,η + ηλmI(cid:1)−1
(cid:17)
Rm,η
,
(72)
µ
2N
Tr
N(cid:88)
N(cid:88)
k=1
k=1
where H m,η and Rm,η are transformed versions of Hη in (28) and Sη in (67) at the m-th eigenvalue λm, respectively:
H m,η (cid:44)
Rm,η (cid:44)
[vm]2
kHk,η = (v(cid:62)
m ⊗ IM )Hη(vm ⊗ IM ),
[vm]2
kRs,k,η = (v(cid:62)
m ⊗ IM )Sη(vm ⊗ IM ).
(73)
(74)
As shown in Section VI-A, under some assumptions on the data and noise characteristics, the individual terms
in (72) are decaying functions of η or λm.
Before proceeding, we note that expression (68) can be written alternatively as:
Tr
(cid:32) N(cid:88)
(cid:32) N(cid:88)
N(cid:88)
Hk,η
Tr
k=1
(cid:33)−1(cid:32) N(cid:88)
Rs,k,η
k=1
(cid:33) +
(cid:33)−1(cid:32) N(cid:88)
[vm]2
kHk,η + ηλmI
[vm]2
kRs,k,η
m=2
k=1
k=1
MSD =
µ
2N
µ
2N
(cid:33) +
(75)
O(µ)
(O(1) + O(η))
,
where we used the fact that λ1 = 0 and v1 = 1√
performance of the single-task diffusion adaptation employed to estimate w(cid:63) given by:
N
1N . Expression (75) allows us to recover the network MSD
N(cid:88)
k=1
w(cid:63) (cid:44) arg min
w
Jk(w).
(76)
To see this, we recall the expression for the network MSD performance of single-task diffusion adaptation derived
in [9, pp. 594]:
MSD =
µ
2N
Tr
(cid:32) N(cid:88)
(cid:33)−1(cid:32) N(cid:88)
Hk,(cid:63)
Rs,k,(cid:63)
k=1
k=1
(cid:33) ,
15
(77)
wk
where Hk,(cid:63) (cid:44) ∇2
Jk(w(cid:63)) and Rs,k,(cid:63) is the covariance of the gradient noise in (53) at w(cid:63). In order to estimate
w(cid:63) using the multitask strategy (7), a very large η needs to be chosen. In this case, we have Hk,η = Hk,(cid:63) and
Rs,k,η = Rs,k,(cid:63). Moreover, the second and third terms on the RHS of expression (75) will be O(µ/η) which are
negligible for a very large η. Thus, we obtain (77).
V. MULTITASK LEARNING BENEFIT
Now that we have studied in detail the mean-square performance of the multitask strategy (7) relative to Wo
η, the
minimizer of the regularized cost (4), we will use the results to examine the benefit of multitask learning compared
to the non-cooperative solution under the smoothness assumption.
Since each cost is strongly convex, each agent k is able to estimate wo
k on its own, if desired, in a non-cooperative
manner by running strategy (7) with η = 0. We know from previous established results on single-agent adaptation [9,
pp. 390] that the network MSD in that case will be given by:
MSDncop =
µ
2N
Tr
H−1
k,o Rs,k,o
,
(78)
wk
Jk(wo
k) and Rs,k,o is the covariance of the gradient noise in (53) at wo
where Hk,o (cid:44) ∇2
k. Note that expression (68)
allows us to recover the mean-square-error performance of stand-alone adaptive agents. In particular, it can be easily
verified that (68) reduces to expression µ
k,o Rs,k,o) for the mean-square deviation of single agent learner [9,
pp. 390] when the network size is set to N = 1 and the topology is removed.
2 Tr(H−1
When η > 0 is used, the graph Laplacian regularizer (4) induces a bias relative to Wo. However, when Wo is
smooth, we expect that promoting the smoothness of Wo through regularization can improve the network MSD
performance despite the bias induced in the estimation.
(cid:33)
(cid:32) N(cid:88)
k=1
A. Induced bias relative to Wo
The bias of the strategy (7) is given by:
From the triangle inequality we have:
where lim supi→∞ (cid:107)E(cid:0)Wo
lim supi→∞ (cid:107)E(cid:0)Wo
η − Wi
lim sup
i→∞
η − Wi
(cid:107)E (Wo − Wi)(cid:107) ≤ (cid:107)Wo − Wo
(cid:1)(cid:107) can be replaced by (cid:107)(cid:101)W(cid:48)
η(cid:107) + lim sup
(80)
i→∞
∞(cid:107) in (40). From expression (31) in Part I [2] we know
η(cid:107) will be. Furthermore, we know from Theorem 4 in Part I [2] that
(cid:1)(cid:107) is O(µ). Thus, for a smooth signal Wo and a small µ, the bias in (79) will be small.
η − Wi
that the smoother Wo is, the smaller (cid:107)Wo − Wo
E (Wo − Wi) = (Wo − Wo
(79)
η) + E(cid:0)Wo
η − Wi
(cid:1) .
(cid:107)E(cid:0)Wo
(cid:1)(cid:107),
16
Fig. 1.
Illustrative example. (Left) Network topology. (Right) Data and noise variances.
B. Network MSD relative to Wo
The mean-square-error performance of the strategy (7) relative to Wo is given by:
E(cid:107)Wo − Wi(cid:107)2 = (cid:107)Wo − Wo
η(cid:107)2 + E(cid:13)(cid:13)Wo
η − Wi
(cid:13)(cid:13)2 + 2(Wo − Wo
η)(cid:62)E(cid:0)Wo
η − Wi
(cid:1) ,
E(cid:0)Wo
(81)
(82)
(cid:1) ,
and the network MSD can be expressed as:
MSD =
1
N
lim sup
i→∞
E(cid:107)Wo − Wi(cid:107)2 = MSD +
(cid:107)Wo − Wo
η(cid:107)2 +
1
N
2
N
(Wo − Wo
η)(cid:62) lim sup
i→∞
η − Wi
or (68). Note that lim supi→∞ E(cid:0)Wo
where we used the bar notation to denote the network MSD relative to Wo and where MSD is given in (65)
∞ in (40). In order to improve the network MSD
(cid:1) can be replaced by (cid:101)W(cid:48)
η − Wi
compared to the non-cooperative case (78), the regularization strength η must be chosen such that:
MSD ≤ MSDncop,
(83)
and the optimal choice of η is the one minimizing MSD in (82) subject to η ≥ 0.
VI. SIMULATION RESULTS
We consider a connected network of N = 15 nodes and M = 5 with the topology shown in Fig. 1 (left). Each
agent is subjected to streaming data {dk(i), uk,i} that are assumed to satisfy a linear regression model [9], [16]:
dk(i) = uk,iwo
k + vk(i),
k = 1, . . . , N,
(84)
for some unknown M × 1 vector wo
agent k:
k with vk(i) a measurement noise. A mean-square-error cost is associated with
Jk(wk) =
(85)
The processes {dk(i), uk,i, vk(i)} are assumed to represent zero-mean jointly wide-sense stationary random pro-
cesses satisfying: i) Eu(cid:62)
k,iu(cid:96),j = Ru,kδk,(cid:96)δi,j where Ru,k > 0 and the Kronecker delta δm,n = 1 if m = n and
k = 1, . . . , N.
Edk(i) − uk,iwk2,
1
2
-3-2-10123-2.5-2-1.5-1-0.500.511.522.5 Node,k2468101214<2u;k0.511.52Node,k2468101214<2v;k0.10.20.30.40.50.6 Recall from Part I [2] that S(W) in (3) can be written as:
wo
m =
· col
1√
M
(cid:111)M
m}N
e−τjλm
.
j=1
(cid:110)
N(cid:88)
17
(86)
(87)
v,kδk,(cid:96)δi,j; iii) the regression and noise processes {u(cid:96),j, vk(i)} are independent
zero otherwise; ii) Evk(i)v(cid:96)(j) = σ2
of each other. We set ak(cid:96) = 0.1 if (cid:96) ∈ Nk and 0 otherwise. It turns out that the Laplacian matrix has 15 distinct
eigenvalues. We generate Wo according to Wo = VWo = col{wo
m=1 with:
S(W) = W(cid:62)LW =
λm(cid:107)wm(cid:107)2,
where wm = (v(cid:62)
is. Note that, for MSE networks, it holds that Hk(wk) = ∇2
m ⊗ IM )W. From (87), we observe that the larger {τj ≥ 0} are, the smoother the signal Wo
Jk(wk) = Ru,k ∀wk. Thus, the fixed point bias
∞ in (40). Furthermore, from
η − W∞ given by (49) in Part I [2] is equal to the long-term model bias (cid:101)W(cid:48)
(cid:101)W∞ = Wo
wk
m=2
the definition (6), the gradient noise process at agent k is given by:
sk,i(wk) = (u(cid:62)
k,iuk,i − Ru,k)(wo
k − wk) + u(cid:62)
k,ivk(i).
Consider the covariance Rs,k,η defined in (53). From (88), we have:
Rs,k,η = E[(u(cid:62)
k,iuk,i − Ru,k)Wk,η(u(cid:62)
k,iuk,i − Ru,k)] + σ2
v,kRu,k,
where
Wk,η = (wo
k − wo
k,η)(wo
k − wo
k,η)(cid:62).
(88)
(89)
(90)
To evaluate (89), we need the fourth order moment of the regressors. Let us assume that the regressors are zero-mean
real Gaussian. In this case, using the fact that Wk,η is symmetric, we obtain [36]:
E[u(cid:62)
k,iuk,iWk,ηu(cid:62)
k,iuk,i] = 2Ru,kWk,ηRu,k + Ru,kTr(Ru,kWk,η).
Replacing the above expression in (89), we obtain:
Rs,k,η = Ru,kWk,ηRu,k + Ru,kTr(Ru,kWk,η) + σ2
v,kRu,k.
(91)
(92)
A. Uniform data profile
In this setting, we assume uniform covariance matrices scenario where Ru,k = Ru ∀k. This scenario is encountered
for example in distributed denoising problems in wireless sensor networks (or image denoising) [28], [37]. In this
case, Ru is equal to one. In such problems, the N sensors in the network are observing N-dimensional signal, with
each entry of the signal corresponding to one sensor. Using the prior knowledge that the signal is smooth w.r.t.
the underlying topology, the sensor task is to denoise the corresponding entry of the signal by performing local
computations and cooperating with its neighbor in order to improve the error variance.
It turns out that in this scenario, the output of the network Wo
η can be interpreted as the output of a low-pass graph
filter applied to the graph signal Wo [1], [28], [38]. To see this, let us first recall the notion of graph frequencies,
18
graph Fourier transform, and graph filters [28], [37] -- [39]. Consider the connected graph G = {N ,E, A} equipped
with a Laplacian matrix L, which can be decomposed as L = V ΛV (cid:62). A graph signal supported on the set N
is defined as a vector x ∈ RN whose k-th component xk ∈ R represents the value of the signal at the k-th
node. By analogy to the classical Fourier analysis, the eigenvectors of the Laplacian matrix L are used to define a
attenuating its spectrum as: Φ =(cid:80)N
graph Fourier basis V and the eigenvalues are considered as the graph frequencies. The graph Fourier transform
(GFT) transforms a graph signal x into the graph frequency domain according to x = V (cid:62)x where {x1, . . . , xN}
are called the spectrum of x. A graph filter Φ is an operator that acts upon a graph signal x by amplifying or
m=1 Φ(λm)xmvm. The frequency response of the filter Φ(λ) controls how much
Φ amplifies the signal spectrum. Low frequencies correspond to small eigenvalues, and low-pass or smooth filters
correspond to decaying functions Φ(λ). Since we are dealing with vectors wk ∈ RM instead of scalars xk, the
graph transformation x = V (cid:62)x becomes W = (V (cid:62) ⊗ IM )W. In the uniform covariance matrices scenario, we have
Ho
η = IN ⊗ Ru and relation (24) in Part I [2] reduces to:
(cid:16)
−1(cid:17)
wo
m,η =
IM − ηλm (ηλmIM + Ru)
wo
m, m = 1, . . . , N.
(93)
m ⊗ IM )Ho
since (v(cid:62)
it holds that the m-th subvector corresponding to the m-th eigenvalue (or graph frequency) of Wo
satisfies:
η(vn ⊗ IM ) = Ru if m = n and zero otherwise. By applying the matrix inversion identity [40],
η = (V (cid:62) ⊗ IM )Wo
wo
m,η = (ηλmIM + Ru)
m,η(cid:107) ≤
(cid:107)wo
1
λm
1 + η
λmax(Ru)
−1 Ruwo
m,
m(cid:107),
(cid:107)wo
(94)
(95)
If η = 0, we are in the case of an all-pass graph filter since the frequency content of the output signal Wo
η, is the
same as the frequency content of the input signal Wo. For η > 0, we are in the case of a low-pass graph filter
m,η(cid:107), is less than or equal to
since the norm of the m-th frequency content of the output signal Wo
the norm of the m-th frequency content of the input signal Wo, namely, (cid:107)wo
m(cid:107). For fixed η, as m increases, the
ratio in (95) decreases. This validates the low-pass filter interpretation. The regularization parameter η controls the
sharpness of the low-pass filter. We illustrate this in Fig. 2 (left). In the experiment, we set Ru = IM and τj = 0
in (86) so that (cid:107)wo
m,η for different values
of η from (94). In order to visualize the frequency response of the graph filter, we also plot in dashed lines the
m(cid:107)2 = 1 ∀m and we illustrate in Fig. 2 (left) the squared (cid:96)2-norm of wo
η, namely, (cid:107)wo
ratio
(cid:107)wo
m,η(cid:107)2
(cid:107)wo
m(cid:107)2
from (94) for λm ∈ [0, 1.2] with (cid:107)wo
m(cid:107)2 = 1 ∀m.
We observe that a similar behavior arises when studying the network MSD for smooth signal Wo. When Wo
is smooth, Wk,η in (90) is small and in this case, the covariance matrix Rs,k,η in (92) can be approximated by
Rs,k,η ≈ σ2
v,kRu. In this case, the network MSD expression in (70) can be approximated as:
(cid:33)
(cid:32) N(cid:88)
k=1
(cid:32) N(cid:88)
N(cid:88)
m=2
k=1
MSD ≈ µM
2
1
N 2
σ2
v,k
+
µ
2N
(cid:33)
(cid:16)(cid:0)IM + ηλmR−1
u
(cid:1)−1(cid:17)
[vm]2
kσ2
v,k
Tr
.
(96)
19
Fig. 2. Uniform data profile scenario. (Left) Spectral content of Wo
for λm ∈ [0, 1.2]. (Right) Network MSD at λm from (98) (relative to Wo
η from (94) for different η. Dashed lines correspond to the ratio
η with Wo a smooth signal) for different η at µ = 0.005.
(cid:107)wo
(cid:107)wo
m,η(cid:107)2
m(cid:107)2
Since the trace of a matrix is equal to the sum of its eigenvalues, we have:
(cid:16)(cid:0)IM + ηλmR−1
u
(cid:1)−1(cid:17)
Tr
M(cid:88)
=
1
1 + ηλm
λq(Ru)
q=1
.
(97)
The above expression shows that for a fixed λm, as η increases, the above trace decreases. We conclude that, when
v ∀k,
η increases, the sum on the RHS of (96) decreases. By further assuming uniform noise profile, i.e., σ2
v,k = σ2
(cid:16)(cid:0)IM + ηλmR−1
u
(cid:1)−1(cid:17)
.
(98)
expression (96) reduces to:
N(cid:88)
m=1
MSD =
MSD(λm)
with
MSD(λm) ≈ µ
2N
σ2
vTr
From (97), we conclude that, for a fixed η, as λm increases, the corresponding trace term in (98) decreases. This
case is illustrated numerically in Fig. 2 (right). In the experiment, we set Ru = IM , σ2
v = 0.1, µ = 0.005, and
τj = 7 + j in (86) so that Wo is smooth and we illustrate MSD(λm) in (98) for different values of η .
B. Varying data profile
We assume that Ru,k = σ2
u,kIM for all k. In this case, expression (92) reduces to:
Rs,k,η = σ4
u,k(Wk,η + (cid:107)wo
k − wo
k,η(cid:107)2IM ) + σ2
v,kσ2
u,kIM .
(99)
The variances σ2
u,k and σ2
v,k are given in Fig. 1 (right). We set τj = j in (86).
of the algorithm relative to Wo
η, we report in Fig. 3 (left) the squared norm of the fixed point bias (cid:101)W∞ = Wo
given by (49) in Part I [2] (which is equal to the long-term model bias (cid:101)W(cid:48)
In order to characterize the influence of the step-size µ and the regularization parameter η on the performance
η − W∞
∞ in (40)) as a function of η where
η = µ− with ≥ −1 for µ = {10−3, 10−4, 10−5}. We observe that for small η, the squared norm of the bias
increases 40 dB per decade (when η goes from η1 to 10η1). This means that the squared norm of the bias is on the
20
Fig. 3. Network performance relative to Wo
regularization strength η = 5, varying step-size µ. (Right) Evolution of the learning curves for fixed µ = 0.001, varying η.
η. (Left) Squared (cid:96)2-norm of the bias (40). (Middle) Evolution of the learning curves for fixed
order of η4 for fixed µ. For large η, the bias becomes constant and we see that, when µ goes from µ1 to 10µ1, it
increases 20 dB. This means that the squared norm of the bias is on the order of µ2. Finally, we observe that for
fixed η, the squared norm of the bias is on the order of µ2. In Fig. 3 (middle), we report the network MSD learning
curves relative to Wo
η obtained by running strategy (7) for η = 5 and for two different values of µ. The curves are
η − Wi(cid:107)2} over 200 repeated experiments. In the simulations, we
obtained by averaging the trajectories { 1
use the following approximation for the network MSD expression in (68):
E(cid:107)Wo
N
(cid:32) N(cid:88)
N(cid:88)
MSDapp =
µ
2N
(cid:33)−1
(cid:32) N(cid:88)
(cid:33)
[vm]2
kσ2
u,k + ηλm
Tr
[vm]2
kRs,k,η
,
(100)
m=1
k=1
k=1
where we use the subscript "app" to indicate that it is an approximation. Compared with (68), we see that the
(O(1)+O(η)) has been removed in (100). It is observed that the learning curves tend to the same MSD value
predicted by the theoretical expression (65) (with O(µ1+θm) removed). Furthermore, we observe that the MSD
term
O(µ)
predicted by expression (100) provides a good approximation for the performance of strategy (7). Finally, it can be
observed that the MSD is on the order of µ. In Fig. 1 (right), we report the MSD learning curves relative to Wo
η
obtained by running strategy (7) for µ = 10−3 and for six different values of η. As it can be seen, by increasing
η, the network MSD decreases. Furthermore, it is observed that expression (100) provides a good approximation
for the network MSD for any η ≥ 0.
In Fig. 4, we characterize the influence of η on the performance of strategy (7) relative to Wo with Wo a smooth
signal generated according to (86) with τj = 7 + j. We set µ = 5 · 10−3. In Fig. 4 (left), we plot MSD for
η ∈ [0, 350]. To generate MSD we use (82) with MSD replaced by MSDapp in (100) which has a low computational
complexity. As it can be seen from this plot, η = 4 gives the best MSD. In Fig. 4 (right), we report the network
E(cid:107)Wo − Wi(cid:107)2 for η = {0, 4, 350}. It is observed that the learning curves tend to the same MSD
value predicted by the theoretical expression (82) with MSD replaced by (65) (with O(µ1+θm) removed) or (100).
learning curves 1
N
210-610-410-2100102104106kWo2!W1k2indB-350-300-250-200-150-100-5007=0:0017=0:00017=0:00001X: 0.01Y: -148.7X: 0.001Y: -188.6X: 0.001Y: -208.6X: 0.01Y: -168.7X: 0.0001Y: -268.6X: 1e-05Y: -308.6X: 1.11e+04Y: -74.64X: 1.11e+05Y: -94.63X: 1110Y: -54.71Iteration,i010002000300040005000600070008000900010000MSDindB-35-30-25-20-15-10SimulatedMSDSteady-stateMSD(65)Steady-stateMSD(100)7=0:017=0:001Iteration,i010002000300040005000600070008000900010000MSDindB-45-40-35-30-25-20-15-10-5SimulatedMSDSteady-stateMSD(65)Steady-stateMSD(100)2=02=52=102=1002=10002=1 21
Fig. 4. Network performance relative to Wo with a smooth signal Wo at µ = 0.005. (Left) Network MSD as a function of the regularization
strength η ∈ [0, 350]. (Right) Evolution of the network MSD learning curve for three different values of η.
VII. CONCLUSION
In this paper, and its accompanying Part I [2], we considered multitask inference problems where agents in the
network have individual parameter vectors to estimate subject to a smoothness condition over the graph. Based
on diffusion adaptation, we proposed a strategy that allows the network to minimize a global cost consisting
of the aggregate sum of the individual costs regularized by a term promoting smoothness. We showed that, for
small step-size parameter, the network is able to approach the minimizer of the regularized problem to arbitrarily
good accuracy levels. Furthermore, we showed how the regularization strength influences the limit point and the
steady-state mean-square-error (MSE) performance of the algorithm. Analytical expressions illustrating these effects
are derived. These expressions revealed explicitly the influence of the network topology, data settings, step-size
parameter, and regularization strength on the network MSE performance and provided insights into the design of
effective multitask strategies for distributed inference over networks. Illustrative examples were considered and
links to spectral graph filtering were also provided.
APPENDIX A
SIZE OF THE PERTURBATION SEQUENCE ci−1
The argument is similar to the one presented in [9, Sec. 10.1]. For ease of reference, we provide a sketch of the
proof here. Let (cid:101)wk,i−1 = wo
k,η − wk,i−1. For each agent k, we have from eq. (174) in Part I [2]:
(cid:107)(cid:102)H k,i−1(cid:107) (cid:44) (cid:107)Hk,η − H k,i−1(cid:107)
k,η) − ∇2
d(cid:107)t(cid:101)wk,i−1(cid:107)dt =
κ(cid:48)
1
2
k,η − t(cid:101)wk,i−1)(cid:107)dt
(cid:90) 1
(cid:90) 1
Jk(wo
d(cid:107)(cid:101)wk,i−1(cid:107),
(cid:107)∇2
wk
Jk(wo
≤
≤
wk
0
0
κ(cid:48)
(101)
210-310-210-1100101102103MSDindB-26.5-26-25.5-25-24.5-24-23.5-23-22.5-22X: 0.0035Y: -23.42X: 4.001Y: -25.27X: 350Y: -22.66Non-cooperationSingle-tasklearningIteration,i0500100015002000250030003500400045005000MSDindB-28-26-24-22-20-18-16-14-12-10SimulatedMSDSteady-stateMSD(82),(65)2=3502=02=4X: 4500Y: -22.66X: 4500Y: -23.48X: 4501Y: -25.36 22
so that
and, consequently,
(cid:107)(cid:101)Hi−1(cid:107) = max
1≤k≤N
(cid:107)(cid:102)H k,i−1(cid:107) ≤ 1
2
lim sup
i→∞
E(cid:107)ci−1(cid:107) ≤ lim sup
i→∞
κ(cid:48)
d lim sup
i→∞
≤ 1
2
d(cid:107)(cid:101)Wi−1(cid:107),
κ(cid:48)
d max
1≤k≤N
2
κ(cid:48)
(cid:107)(cid:101)wk,i−1(cid:107) ≤ 1
E(cid:107)(cid:101)Hi−1(cid:107)(cid:107)(cid:101)Wi−1(cid:107)
E(cid:107)(cid:101)Wi−1(cid:107)2 = O(µ),
(102)
(103)
from Theorem 2 in Part I [2].
In the following, we argue that (cid:107)ci−1(cid:107) = O(µ) asymptotically with high probability. Let us introduce the
nonnegative random variable u = (cid:107)ci−1(cid:107) and let us recall Markov's inequality which states that for any nonnegative
random variable u and ξ > 0, it holds that:
Prob(u ≥ ξ) ≤ (Eu) /ξ.
(104)
Let rc = nµ, for any constant integer n ≥ 1 that we are free to choose. We then conclude from (103) and (104)
that for i (cid:29) 1:
Prob((cid:107)ci−1(cid:107) < rc) = 1 − Prob((cid:107)ci−1(cid:107) ≥ rc)
≥ 1 − E(cid:107)ci−1(cid:107)/rc
≥ 1 − O(1/n)
(105)
where the term O(1/n) is independent of µ. This result shows that the probability of having (cid:107)ci−1(cid:107) bounded by
rc can be made arbitrarily close to one by selecting a large enough value for n. Once the value for n has been
fixed to meet a desired confidence level, then rc = O(µ).
APPENDIX B
PROOF OF LEMMA 1
To simplify the notation, we introduce the difference:
zi (cid:44) (cid:101)Wi − (cid:101)W(cid:48)
i.
Subtracting recursions (32) and (34), we get:
zi = Bηzi−1 + µ(IM N − µηL)ci−1.
in terms of the random perturbation sequence ci−1 given in (33). For each agent k, we have from (101):
(cid:107)(cid:102)H k,i−1(cid:107) ≤ 1
2
d(cid:107)wo
κ(cid:48)
k,η − wk,i−1(cid:107).
(106)
(107)
(108)
It follows that
and
(cid:107)ck,i−1(cid:107) ≤ (cid:107)(cid:102)H k,i−1(cid:107)(cid:107)wo
k,η − wk,i−1(cid:107)
≤ 1
2
≤ 1
2
d(cid:107)wo
κ(cid:48)
d(cid:107)Wo
κ(cid:48)
k,η − wk,i−1(cid:107)2
η − Wi−1(cid:107)2.
(cid:107)ck,i−1(cid:107)2 ≤ 1
4
d)2(cid:107)Wo
(κ(cid:48)
η − Wi−1(cid:107)4.
Applying Jensen's inequality [41, pp. 77] to the convex function (cid:107) · (cid:107)2, we obtain from (107):
E(cid:107)zk,i(cid:107)2 ≤ (cid:88)
(cid:96)∈Nk
[C]k,(cid:96)E(cid:107)(I − µH(cid:96),η)z(cid:96),i−1 + µc(cid:96),i−1(cid:107)2.
C (cid:44) IN − µηL.
where C is given by:
Next note that
(cid:13)(cid:13)(cid:13)(cid:13)t
E(cid:107)(I − µHk,η)zk,i−1 + µck,i−1(cid:107)2
= E
≤ 1
t
(IM − µHk,η)zk,i−1 + µ(1 − t)
E(cid:107)(IM − µHk,η)zk,i−1(cid:107)2 + µ2 1
1 − t
1
t
(cid:13)(cid:13)(cid:13)(cid:13)2
1
ck,i−1
1 − t
E(cid:107)ck,i−1(cid:107)2,
for any arbitrary positive number t ∈ (0, 1). We select t = γk with γk defined in (48). This gives
E(cid:107)(I − µHk,η)zk,i−1 + µck,i−1(cid:107)2 ≤ γkE(cid:107)zk,i−1(cid:107)2 + µ2
1
1 − γk
E(cid:107)ck,i−1(cid:107)2,
Let us introduce the mean-square perturbation vector at time i:
MSPz,i (cid:44) col(cid:8)E(cid:107)zk,i(cid:107)2(cid:9)N
k=1 .
Replacing (114) into (111), and using (110), we obtain:
MSPz,i (cid:22) CG(cid:48)(cid:48)MSPz,i−1 + µ2 1
4
with G(cid:48)(cid:48) = diag{γk}N
k=1. Iterating the above recursion starting from i = 1, we obtain:
η − Wi−1(cid:107)4),
d)2C(IN − G(cid:48)(cid:48))−1(1N ⊗ E(cid:107)Wo
(κ(cid:48)
i−1(cid:88)
(CG(cid:48)(cid:48))jC(IN − G(cid:48)(cid:48))−1(1N ⊗ E(cid:107)Wo
η − Wi−1−j(cid:107)4).
MSPz,i (cid:22) (CG(cid:48)(cid:48))iMSPz,0 + µ2 1
4
(κ(cid:48)
d)2
j=0
23
(109)
(110)
(111)
(112)
(113)
(114)
(115)
(116)
(117)
Under Assumption 3 and condition (39), the matrix CG(cid:48)(cid:48) is guaranteed to be stable. Following similar arguments
to the ones used to establish eq. (68) in Part I (Appendix E) [2], and from Theorem 3 in Part I [2], we conclude
that:
(cid:107) lim sup
i→∞
MSPz,i(cid:107)∞ = O(µ2),
(118)
24
where we used the fact that (cid:107)(IN − G(cid:48)(cid:48))−1(cid:107)∞ ≤ O(µ−1). It follows that
i(cid:107)2 = O(µ2).
lim sup
i→∞
Finally note that
E(cid:107)(cid:101)W(cid:48)
i(cid:107)2 = E(cid:107)(cid:101)W(cid:48)
i − (cid:101)Wi + (cid:101)Wi(cid:107)2 ≤ E(cid:107)(cid:101)W(cid:48)
≤ E(cid:107)(cid:101)W(cid:48)
E(cid:107)(cid:101)Wi − (cid:101)W(cid:48)
i − (cid:101)Wi)(cid:62)(cid:101)Wi
i − (cid:101)Wi(cid:107)2 + E(cid:107)(cid:101)Wi(cid:107)2 + 2E((cid:101)W(cid:48)
(cid:113)
i − (cid:101)Wi(cid:107)2 + E(cid:107)(cid:101)Wi(cid:107)2 + 2
E(cid:107)(cid:101)W(cid:48)
i − (cid:101)Wi(cid:107)2E(cid:107)(cid:101)Wi(cid:107)2
(119)
(120)
where we used Ex ≤ Ex from Jensen's inequality and where we applied Holder's inequality:
Ex(cid:62)y ≤ (Exp)
1
p (Exq)
1
q , when 1/p + 1/q = 1.
Hence we get:
lim sup
i→∞
2 for small µ (cid:28) 1.
3
since µ2 < µ
(cid:0)E(cid:107)(cid:101)W(cid:48)
i(cid:107)2 − E(cid:107)(cid:101)Wi(cid:107)2(cid:1) ≤ O(µ2) +
(cid:112)
APPENDIX C
O(µ3) = O(µ
3
2 ),
(121)
From (43), we have:
k,∞ − w(cid:48)
w(cid:48)
k,i =
where C is defined in (112) and where φ(cid:48)
k,i is given by:
(cid:96)=1
PROOF OF LEMMA 3
N(cid:88)
[C]k(cid:96)φ(cid:48)
(cid:96),i,
k,i−1) − µsk,i(wk,i−1).
Applying Jensen's inequality [41, pp. 77] to the convex function (cid:107) · (cid:107)2, we obtain:
k,i = (IM − µHk,η)(w(cid:48)
φ(cid:48)
k,∞ − w(cid:48)
k,i(cid:107)2 ≤ N(cid:88)
E(cid:107)w(cid:48)
k,∞ − w(cid:48)
[C]k(cid:96)E(cid:107)φ(cid:48)
(cid:96),i(cid:107)2.
Under Assumption 2, we have:
(cid:96)=1
E[(cid:107)φ(cid:48)
k,i(cid:107)2F i−1] = (cid:107)w(cid:48)
k,∞ − w(cid:48)
k,i−1(cid:107)2
Σk
+ µ2E[(cid:107)sk,i(wk,i−1)(cid:107)2F i−1].
where Σk (cid:44) (IM − µHk,η)2, which due to Assumption 1, can be bounded as follows:
(122)
(123)
(124)
(125)
0 < Σk ≤ γ2
kIM ,
(126)
where γk is given by (48). Taking expectation again in (125), and using the bound (136) on E[(cid:107)sk,i(wk,i−1)(cid:107)2F i−1]
from Part I [2], we obtain:
k,∞ − w(cid:48)
E(cid:107)φ(cid:48)
E(cid:107)w(cid:48)
E(cid:107)wk,∞ − wk,i−1(cid:107)2 + µ2(cid:0)3β2
+ µ2E(cid:107)sk,i(wk,i−1)(cid:107)2
k,η − wk,∞(cid:107)2 + 3β2
k,i−1(cid:107)2 + 3µ2β2
k,i(cid:107)2 = E(cid:107)w(cid:48)
k(cid:107)wo
k,η(cid:107)2 + σ2
k,∞ − w(cid:48)
k,i−1(cid:107)2
k(cid:107)wo
≤ γ2
k
Σk
k
(cid:1) .
s,k
(127)
Now, combining (127) and (124), we obtain (45).
Iterating (45) starting from i = 1, we get:
MSP(cid:48)
i (cid:22) (C(G(cid:48)(cid:48))2)iMSP(cid:48)
0 + µ2
25
i−1(cid:88)
j=0
(C(G(cid:48)(cid:48))2)jC(3diag{β2
k}MSPi−1−j + b).
(128)
Under Assumption 3 and condition (39), the matrix C(G(cid:48)(cid:48))2 is guaranteed to be stable. Using the fact that b = O(1)
from Theorem 1 in Part I [2], 1 − (cid:107)(G(cid:48)(cid:48))2(cid:107)∞ = O(µ), and that (cid:107) limi→∞ MSPi(cid:107)∞ = O(µ) from Theorem 2 in
Part I [2], and following similar arguments as the one used to establish eq. (68) in Part I (Appendix E) [2], we
conclude (49).
From (42), Lemma 2, and (49), we conclude (50).
APPENDIX D
LIMITING SECOND-ORDER MOMENT OF GRADIENT NOISE
It is shown in [9, Sec. 4.1] that the conditional noise covariance matrix satisfying Assumptions 2 and 5 satisfies
more globally a condition of the following form for all wk ∈ F i−1:
(cid:107)Rs,k,i(wk) − Rs,k,i(wo
k,η)(cid:107) ≤ κd(cid:107)wo
k,η − wk(cid:107)θ + κ(cid:48)
d(cid:107)wo
k,η − wk(cid:107)2,
for some nonnegative constant κ(cid:48)
E[sk,i(wk,i−1)s(cid:62)
k,i(wk,i−1)F i−1] =E[sk,i(wo
d. By adding and subtracting the same term we have:
k,η)F i−1] + E[sk,i(wk,i−1)s(cid:62)
k,i(wo
k,η)s(cid:62)
E[sk,i(wo
k,i(wo
k,η)s(cid:62)
k,η)F i−1]
k,i(wk,i−1)F i−1]−
which using definition (52) can be rewritten as:
E[sk,i(wk,i−1)s(cid:62)
k,i(wk,i−1)F i−1] = E[sk,i(wo
k,η)s(cid:62)
k,i(wo
k,η)F i−1] + Rs,k,i(wk,i−1) − Rs,k,i(wo
k,η).
(129)
(130)
(131)
Subtracting Rs,k,η defined by (53) from both sides and computing expectations, we get:
Esk,i(wk,i−1)s(cid:62)
k,i(wk,i−1)−Rs,k,η = E(cid:16)E[sk,i(wo
(cid:17)
+E(cid:0)Rs,k,i(wk,i−1) − Rs,k,i(wo
k,η)(cid:1) .
k,η)F i−1] − Rs,k,η
k,η)s(cid:62)
k,i(wo
It then follows from the triangle inequality of norms, and from Jensen's inequality that:
(cid:107)Esk,i(wk,i−1)s(cid:62)
≤(cid:13)(cid:13)(cid:13)E(cid:16)E[sk,i(wo
≤ E(cid:107)E[sk,i(wo
k,i(wk,i−1) − Rs,k,η(cid:107)
k,η)s(cid:62)
k,η)s(cid:62)
k,i(wo
k,i(wo
(cid:17)(cid:13)(cid:13)(cid:13) + (cid:107)E(cid:0)Rs,k,i(wk,i−1) − Rs,k,i(wo
k,η)(cid:1)(cid:107)
k,η)F i−1] − Rs,k,η
k,η)F i−1] − Rs,k,η(cid:107) + E(cid:107)Rs,k,i(wk,i−1) − Rs,k,i(wo
k,η)(cid:107).
(132)
(133)
Computing the limit superior of both sides and using (53) to annihilate the limit of the first term on the RHS, we
conclude that:
lim sup
i→∞
(cid:107)Esk,i(wk,i−1)s(cid:62)
k,i(wk,i−1) − Rs,k,η(cid:107) ≤ lim sup
i→∞
E(cid:107)Rs,k,i(wk,i−1) − Rs,k,i(wo
k,η)(cid:107).
(134)
26
We next use the smoothness condition (129) to bound the right-most term as follows:
(cid:107)Rs,k,i(wk,i−1) − Rs,k,i(wo
(135)
Under expectation and in the limit, we have:
E(cid:107)Rs,k,i(wk,i−1) − Rs,k,i(wo
lim sup
i→∞
(cid:0)(cid:107)(cid:101)wk,i−1(cid:107)4(cid:1) θ
k,η)(cid:107) ≤ κd
(cid:110)
κdE(cid:0)(cid:107)(cid:101)wk,i−1(cid:107)4(cid:1) θ
(cid:110)
(cid:0)E(cid:107)(cid:101)wk,i−1(cid:107)4(cid:1) θ
κd
4 + κ(cid:48)
d(cid:107)(cid:101)wk,i−1(cid:107)2.
E(cid:107)(cid:101)wk,i−1(cid:107)2(cid:111)
E(cid:107)(cid:101)wk,i−1(cid:107)2(cid:111)
d
d
4 + κ(cid:48)
4 + κ(cid:48)
k,η)(cid:107) ≤ lim sup
i→∞
≤ lim sup
i→∞
where we applied Jensen's inequality to the function f (x) = x
(136)
4 ; this function is concave over x ≥ 0 for θ ∈ (0, 4].
Moreover, in the last step we called upon the results in Theorems 2 and 3 in Part I [2] where it is shown that
the second and fourth order moments of (cid:101)wk,i−1 are asymptotically bounded by O(µ) and O(µ2), respectively.
Accordingly, the exponent θ(cid:48) (cid:44) min{θ, 2} since O(µθ/2) dominates O(µ) for values of θ ∈ (0, 2] and O(µ)
dominates O(µθ/2) for values of θ ∈ [2, 4]. Substituting into (134) gives:
= O(µ
θ
θ(cid:48)
2 ),
(cid:107)Esk,i(wk,i−1)s(cid:62)
k,i(wk,i−1) − Rs,k,η(cid:107) = O(µ
θ(cid:48)
2 ),
lim sup
i→∞
which leads to (55).
Consider the matrix F η in (57). Using the block Kronecker product property:
APPENDIX E
PROOF OF LEMMA 4
it can be verified that:
(A ⊗b B)(C ⊗b D) = (AC ⊗b BD),
F η = B(cid:62)
η ⊗b B(cid:62)
η ,
where
so that:
B(cid:62)
η =
Bη = V(cid:62)BηV = (IM N − µηJ )(IM N − µV(cid:62)HηV)
I − µH11
−µH12
...
−µ(1 − µηλ2)H21
(1 − µηλ2)(I − µH22)
...
. . .
. . .
−µ(1 − µηλN )HN 1
−µ(1 − µηλN )HN 2
...
−µH1N
−µ(1 − µηλ2)H2N
. . .
(1 − µηλN )(I − µHN N )
,
(137)
(138)
(139)
(140)
(141)
where Hmn is defined in (63). It can be verified that the matrix Z = I − F η is N × N blocks Zmn with each
block of size M 2N × M 2N:
Z11 Z12
. . . Z1N
Z21 Z22
...
...
ZN 1 ZN 2
. . . Z2N
...
. . . ZN N
,
Z = I − F η =
(142)
We denote by [Zmn]pq the M 2 × M 2 (p, q)-th block of Zmn. We have:
I − (1 − µηλm)(1 − µηλp)[(I − µHmm) ⊗ (I − µHpp)],
µ(1 − µηλm)(1 − µηλq)[(I − µHmm) ⊗ Hqp],
µ(1 − µηλn)(1 − µηλp)[Hnm ⊗ (I − µHpp)],
−µ2(1 − µηλn)(1 − µηλq)[Hnm ⊗ Hqp],
if m = n, p = q
if m = n, p (cid:54)= q
if m (cid:54)= n, p = q
if m (cid:54)= n, p (cid:54)= q
(I − µHmm) ⊗ (I − µHpp) = I − µHmm ⊕ Hpp + µ2Hmm ⊗ Hpp,
where Hmm ⊕ Hpp is given by (62) and
1 − (1 − µηλm)(1 − µηλp) = µη(λm + λp − µηλmλp).
[Zmn]pq =
We have:
Thus,
[Zmn]pq = µ ·
27
(143)
(144)
(145)
(147)
(148)
(149)
(150)
(151)
(152)
(1 − µηλm)(1 − µηλp)(Hmm ⊕ Hpp) + η(λm + λp − µηλmλp)I + O(µ),
(1 − µηλm)(1 − µηλq)(I ⊗ Hqp) + O(µ),
(1 − µηλn)(1 − µηλp)(Hnm ⊗ I) + O(µ),
−µ(1 − µηλn)(1 − µηλq)[Hnm ⊗ Hqp] = O(µ),
if m = n, p = q
if m = n, p (cid:54)= q
if m (cid:54)= n, p = q
if m (cid:54)= n, p (cid:54)= q
(146)
Before proceeding, we recall the following useful properties of the Kronecker and Kronecker sum products [35].
Let {λi(A), i = 1, . . . , M} and {λj(B), j = 1, . . . , M} denote the eigenvalues of any two M × M matrices A and
B, respectively. Then,
{λ(A ⊗ B)} = {λi(A)λj(B)}M,M
i=1,j=1,
{λ(A ⊕ B)} = {λi(A) + λj(B)}M,M
i=1,j=1.
From (142), (143), and (146), it can be verified that the matrix Z can be written as:
Z = X + Y.
The matrix X is N 2 × N 2 block diagonal defined as:
(cid:110)
X (cid:44) µ · diag
diag{[Zmm]pp}N
p=1
O(1)
0
= µ ·
,
0
O(1) + O(η)
(cid:111)N
(cid:80)N
m=1
where we used the fact that for m = 1 and p = 1, we have [Z11]11 = µ· H11⊕ H11 + O(µ2) which is µ· O(1). This
k=1 Hk,η > 0. For the remaining blocks of X, from (145),
is due to property (148) and the fact that H11 = 1
N
k=1[vm]2
property (148), and the fact that Hmm =(cid:80)N
kHk,η > 0, it can be verified that the matrix:
[Zmm]pp = µ(1 − µηλm)(1 − µηλp)(Hmm ⊕ Hpp) + µη(λm + λp − µηλmλp)I
is also positive definite when:
0 < (1 − µηλm)(1 − µηλp) ≤ 1.
28
Furthermore, in this case, we have [Zmm]pp = µ · (O(1) + O(η)). The matrix Y = Z − X in (149) is an N 2 × N 2
block matrix where each block is M 2 × M 2 given by:
[Ymn]pq = µ ·
0,
(1 − µηλm)(1 − µηλq)(I ⊗ Hqp) + O(µ) ≤ O(1),
(1 − µηλn)(1 − µηλp)(Hnm ⊗ I) + O(µ) ≤ O(1),
−µ(1 − µηλn)(1 − µηλq)[Hnm ⊗ Hqp] ≤ O(µ),
if m = n, p = q
if m = n, p (cid:54)= q
if m (cid:54)= n, p = q
if m (cid:54)= n, p (cid:54)= q
Applying the matrix inversion identity [40], we obtain:
(X + Y )−1 = X−1 − X−1Y (I + X−1Y )−1X−1,
From (150), we have:
X−1 = µ−1 · diag
(cid:110)
diag(cid:8)([Zmm]pp)−1(cid:9)N
p=1
(cid:111)N
m=1
= µ−1 ·
O(1)
0
.
0
(O(1) + O(η))−1
(153)
(154)
(155)
(156)
(157)
(158)
(159)
From (155) and (153), we have:
X−1Y =
and
0
O(1)
(O(1) + O(η))−1
(O(1) + O(η))−1
I + X−1Y =
O(1)
O(1)
(O(1) + O(η))−1 O(1)
Applying the block inversion formula to I + X−1Y , we obtain:
(I + X−1Y )−1 =
O(1)
O(1)
(O(1) + O(η))−1 O(1)
,
.
.
Finally, from (155), (156), and (158), we conclude that:
X−1Y (I + X−1Y )−1X−1 = µ−1 ·
(O(1) + O(η))−1
(O(1) + O(η))−1
.
(O(1) + O(η))−1
(O(1) + O(η))−2
Consider now the matrix (I − Fη)−1 = (V ⊗b V)(I − F η)−1(V ⊗b V)(cid:62). It can be verified that:
(I − Fη)−1 = (V ⊗b V)X−1(V ⊗b V)(cid:62) + (V ⊗b V)X−1Y (I + X−1Y )−1X−1(V ⊗b V)(cid:62)
N(cid:88)
N(cid:88)
=
[(vm ⊗ IM ) ⊗b (vp ⊗ IM )]([Zmm]pp)−1[(v(cid:62)
m ⊗ IM ) ⊗b (v(cid:62)
p ⊗ IM )] + µ−1(O(1) + O(η))−1.
m=1
p=1
(160)
APPENDIX F
PROOF OF LEMMA 5
29
zi (cid:44) (cid:101)W(cid:48)
i − E (cid:101)W(cid:48)
i.
Recursion (34) for the long term model includes a constant driving term on the RHS represented by µ2η2L2Wo
η.
To facilitate the variance analysis, we introduce the centered variable:
(161)
(162)
Subtracting (38) from (34), we obtain:
zi = Bηzi−1 − µ(IM N − µηL)si(Wi−1),
where the deterministic driving terms are removed. Although we are interested in evaluating the asymptotic size of
i(cid:107)2, we can rely on the centered variable zi for this purpose, since, from Lemma 2, it holds for i (cid:29) 1:
E(cid:107)(cid:101)W(cid:48)
i(cid:107)2 + O(µ2),
for fixed η. Furthermore, we established in Lemma 1, that the error variances E(cid:107)(cid:101)W(cid:48)
E(cid:107)zi(cid:107)2 = E(cid:107)(cid:101)W(cid:48)
i(cid:107)2 = E(cid:107)(cid:101)W(cid:48)
i(cid:107)2 − (cid:107)E(cid:101)W(cid:48)
i(cid:107)2 and E(cid:107)(cid:101)Wi(cid:107)2 are within
(163)
O(µ
3
2 ) from each other. Therefore, we may evaluate the mean-square error in terms of the mean-square value of
the variable zi by employing the correction:
E(cid:107)(cid:101)Wi(cid:107)2 = lim sup
i→∞
lim sup
i→∞
1
N
E(cid:107)zi(cid:107)2 + O(µ
3
2 ).
1
N
(164)
We therefore continue with recursion (162) and proceed to examine how the mean-square value of zi evolves over
time by relying on the energy conservation arguments.
Let Σ denote an arbitrary symmetric positive semi-definite matrix that we are free to choose. Equating the squared
weighted values of both sides of (162) and taking expectations conditioned on the past history gives:
Taking expectation again removes the conditioning and we get:
E(cid:2)(cid:107)zi(cid:107)2
ΣF i−1
E(cid:107)zi(cid:107)2
η ΣBη
(cid:3) = (cid:107)zi−1(cid:107)2B(cid:62)
Σ = E(cid:16)(cid:107)zi−1(cid:107)2B(cid:62)
(cid:104)
= µ2Tr
(cid:105)
+ µ2E(cid:104)(cid:107)si(cid:107)2
+ µ2E(cid:16)(cid:107)si(cid:107)2
(IM N−µηL)Σ(IM N−µηL)F i−1
(cid:17)
(IM N−µηL)Σ(IM N−µηL)
.
η ΣBη
(cid:17)
(IM N − µηL)Σ(IM N − µηL)E(cid:16)
(cid:17) − Sη(cid:107) = O(µmin{1, θ
i (Wi−1)
2})
(cid:17)
(cid:107)E(cid:16)
lim sup
i→∞
si(Wi−1)s(cid:62)
si(Wi−1)s(cid:62)
i (Wi−1)
Consider the right-most term. We have:
µ2E(cid:16)(cid:107)si(cid:107)2
(IM N−µηL)Σ(IM N−µηL)
Using the sub-multiplicative property of the 2−induced norm, we conclude that:
(cid:16)E(si(Wi−1)s(cid:62)
i (Wi−1) − Sη
(cid:17)(cid:107) = Tr(Σ) · O(µ2+min{1, θ
2}),
µ2(cid:107)(IM N − µηL)Σ(IM N − µηL)
lim sup
i→∞
(165)
(166)
.
(167)
(cid:17)(cid:105)
(168)
(169)
Using (55) and the fact that the gradient noises across the agents are uncorrelated under condition (15), we obtain:
30
where we used the fact that (cid:107)Σ(cid:107) ≤ Tr(Σ) for any positive semi-definite Σ. Using the fact that Tr(X) ≤ c(cid:107)X(cid:107)
for any square matrix X, we obtain:
lim sup
i→∞
(IM N−µηL)Σ(IM N−µηL) − Tr(ΣY) = Tr(Σ) · O(µ2+min{1, θ
µ2E(cid:107)si(cid:107)2
(170)
2}) ≥ 0. The above relation then implies that, given > 0, there exists an io large enough
2}) = b1,
b1 = Tr(Σ)· O(µ2+min{1, θ
such that for all i > io it holds that
µ2E(cid:107)si(cid:107)2
(IM N−µηL)Σ(IM N−µηL) − Tr(ΣY) ≤ b1 + .
(171)
If we select = Tr(Σ) · O(µ2+min{1, θ
2}) and introduce the sum bo = b1 + , then we arrive at:
Tr(ΣY) − bo ≤ µ2E(cid:107)si(cid:107)2
(IM N−µηL)Σ(IM N−µηL) ≤ Tr(ΣY) + bo,
(172)
2}). Substituting (172) into (166) we obtain for i (cid:29) 1:
for some non-negative constant bo = Tr(Σ) · O(µ2+min{1, θ
Σ ≤ E(cid:107)zi−1(cid:107)2B(cid:62)
E(cid:107)zi(cid:107)2
+ Tr(ΣY) + bo.
η ΣBη
Using the sub-additivity property of the limit superior, we obtain:
E(cid:107)zi−1(cid:107)2B(cid:62)
E(cid:107)zi(cid:107)2
Σ ≤ lim sup
i→∞
lim sup
i→∞
η ΣBη
+ Tr(ΣY) + bo.
Grouping terms we get:
lim sup
i→∞
E(cid:107)zi(cid:107)2
Σ−B(cid:62)
η ΣBη
≤ Tr(ΣY) + bo,
We conclude that the limit superior of the error variance satisfies:
E(cid:107)zi(cid:107)2
Σ−B(cid:62)
η ΣBη
lim sup
i→∞
= Tr(ΣY) + Tr(Σ)O(µ2+min{1, θ
2}),
(173)
(174)
(175)
(176)
In order to obtain identity as a weighting matrix on the mean-square value of zi in (176), we select Σ as the
solution to the following discrete time Lyapunov equation:
(177)
We know that Bη is stable under conditions (17) and (39). Accordingly, we are guaranteed that the above Lyapunov
equation has a unique solution Σ, and moreover, this solution is symmetric and non-negative definite as desired.
Σ − B(cid:62)
η ΣBη = IM N .
We can then focus on evaluating the RHS of (176).
For this purpose, we start by applying the block vectorization operation to both sides of (177) to find that:
bvec(Σ) − (B(cid:62)
η ⊗b B(cid:62)
η )bvec(Σ) = bvec(IM N ),
(178)
so that in terms of the matrix Fη defined in (56), we can write:
(179)
Now, substituting this Σ into (176), we obtain E(cid:107)zi(cid:107)2 on the left-hand side while the term Tr(ΣY) on the RHS
becomes:
bvec(Σ) = (I − Fη)−1bvec(IM N ).
Tr(ΣY) = [bvec(Y(cid:62))](cid:62)(I − Fη)−1bvec(IM N ).
(180)
31
Likewise the second term on the RHS of (176) becomes:
2}) · Tr(Σ) = O(µ2+min{1, θ
O(µ2+min{1, θ
2}) · [bvec(IM N )](cid:62)(I − Fη)−1bvec(IM N ).
(181)
We now verify that [bvec(IM N )](cid:62)(I − Fη)−1bvec(IM N ) = O(1/µ). This result will permit us to assess the size
of the second term on the RHS of (176). We have:
[bvec(I)](cid:62)(I − Fη)−1bvec(I) ≤ (cid:107)bvec(I)(cid:107)2 · (cid:107)(I − Fη)−1(cid:107) ≤ r · (cid:107)(I − Fη)−1(cid:107)1(cid:107)bvec(I)(cid:107)2 = O(µ−1),
(182)
where we used a positive constant r to account for the fact that matrix norms are equivalent.
Returning to (176), and using (177), (180), (181), and (182), we conclude that:
E(cid:107)zi(cid:107)2 = [bvec(Y(cid:62))](cid:62)(I − Fη)−1bvec(IM N ) + O(µ1+min{1, θ
2}),
lim sup
i→∞
with θ ∈ (0, 4]. But since Fη is a stable matrix, we can employ the expansion:
η )n ⊗b (B(cid:62)
∞(cid:88)
(I − Fη)−1 =
[bvec(Y(cid:62))](cid:62)(I − Fη)−1bvec(IM N ) =
∞(cid:88)
∞(cid:88)
and write:
F n
η =
(B(cid:62)
n=0
n=0
η )n,
Tr(Bn
ηY(B(cid:62)
η )n),
This series converges to the trace value of the unique solution of the following Lyapunov equation:
n=0
where
X − BηXB(cid:62)
η = Y,
∞(cid:88)
n=0
X =
Bn
ηY(B(cid:62)
η )n.
Consequently, using (164), we obtain:
1
N
lim sup
i→∞
E(cid:107)(cid:101)Wi(cid:107)2 =
=
1
N
1
N
∞(cid:88)
n=0
Tr(X ) + O(µ1+θm)
Tr(Bn
ηY(B(cid:62)
η )n) + O(µ1+θm),
(183)
(184)
(185)
(186)
(187)
(188)
2 min{1, θ} and where Tr(X ) = [bvec(Y(cid:62))](cid:62)(I−Fη)−1bvec(IM N ) which is O(µ) since (cid:107)Y(cid:107) = O(µ2)
where θm = 1
and (cid:107)(I − Fη)−1(cid:107) = O(µ−1). Therefore the value of Tr(X ) is O(µ), which dominates the factor O(µ1+θm).
APPENDIX G
PROOF OF THEOREM 1
From (65) and (160), we have:
(bvec(Y(cid:62)))(cid:62)(I − Fη)−1bvec(IM N ) = O(µ)(O(1) + O(η))−1+
(bvec(Y(cid:62)))(cid:62)[(vm ⊗ IM ) ⊗b (vp ⊗ IM )]([Zmm]pp)−1[(v(cid:62)
m ⊗ IM ) ⊗b (v(cid:62)
p ⊗ IM )]bvec(IM N )
(189)
N(cid:88)
N(cid:88)
m=1
p=1
32
where the bvec operation is relative to blocks of size M × M. Using the property bvec(ACB) = (B(cid:62)⊗bA)bvec(C),
we obtain:
[(v(cid:62)
m ⊗ IM ) ⊗b (v(cid:62)
p ⊗ IM )]bvec(IM N ) =
and we conclude that:
if m = p
if m (cid:54)= p
(190)
bvec(IM ) = vec(IM ),
N(cid:88)
0,
(bvec(Y(cid:62)))(cid:62)(I − Fη)−1bvec(IM N ) = O(µ)(O(1) + O(η))−1 +
(bvec(Y(cid:62)))(cid:62)[(um ⊗ IM ) ⊗b (um ⊗ IM )]xm
where
m=1
xm (cid:44) ([Zmm]mm)−1vec(IM ).
This vector is the unique solution to the linear system of equations:
[Zmm]mmxm = vec(IM ),
or, equivalently, by using (151):
(cid:104)
(cid:104)
(cid:105)
xm
λm(2 − µηλm)
η
2
(1 − µηλm)2(Hmm ⊗ I) +
µ
+ µ
(1 − µηλm)2(I ⊗ Hmm) +
λm(2 − µηλm)I
η
2
(cid:105)
xm = vec(IM ),
Let Xm = unvec(xm). Applying the property vec(ACB) = (B(cid:62) ⊗ A)vec(C), we obtain:
where
vec(XmTm) + vec(TmXm) = vec(IM )
Tm (cid:44) µ(1 − µηλm)2Hmm +
λm(2 − µηλm)I.
µη
2
(191)
(192)
(193)
(194)
(195)
(196)
We conclude from the above equation that Xm is the unique solution to the continuous time Lyapunov equation:
whose solution is given by:
Xm =
T −1
m =
1
2
1
2µ
XmTm + TmXm = IM ,
(cid:16)
(1 − µηλm)2Hmm +
(cid:17)−1
.
λm(2 − µηλm)I
η
2
Using the definitions (66), (67), and applying properties
bvec(ACB) = (B(cid:62) ⊗b A)bvec(C),
and Tr(AB) = (bvec(B(cid:62)))(cid:62)bvec(A),
(197)
(198)
(199)
we get:
(cid:16)
m=1
=
=
m=1
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
= µ2
µ
2
m=1
m=1
Tr
=
=
µ
2
m=1
[bvec(Y(cid:62))](cid:62)[(vm ⊗ IM ) ⊗b (vm ⊗ IM )]vec(Xm)
Tr (unbvec{(vm ⊗ IM ) ⊗b (vm ⊗ IM )bvec(Xm)}Y)
(vm ⊗ IM )Xm(v(cid:62)
(1 − µηλm)2Tr
m ⊗ IM )Sη(vm ⊗ IM )Xm
(v(cid:62)
m ⊗ IM )Y(cid:17)
(cid:16)
(cid:32) N(cid:88)
[vm]2
k=1
kRs,k,η
(cid:33)(cid:32)
(1 − µηλm)2Tr
(cid:32) N(cid:88)
Tr
[vm]2
kHk,η +
m=1
k=1
33
(cid:33)−1
(200)
(201)
(cid:33)
[vm]2
kHk,η
λm(2 − µηλm)I
(cid:17)
(cid:32) N(cid:88)
k=1
(1 − µηλm)2
(cid:33)−1(cid:32) N(cid:88)
[vm]2
kRs,k,η
k=1
+
η
2
(cid:33)
(cid:33) .
(cid:33)−1(cid:32) N(cid:88)
ηλm(2 − µηλm)
2(1 − µηλm)2 I
N(cid:88)
(cid:32) N(cid:88)
Tr
[vm]2
kHk,η +
m=1
+
k=1
O(µ)
(O(1) + O(η))
+ O(µ1+θm).
ηλm(2 − µηλm)
2(1 − µηλm)2 I
[vm]2
kRs,k,η
k=1
Substituting into (191) and (65), we conclude:
lim sup
i→∞
1
N
E(cid:107)Wo
η − Wi(cid:107)2 =
µ
2N
Now, according to definition (51), dividing (201) by µ and computing the limit as µ → 0, we arrive at expression (68)
for the network MSD.
REFERENCES
[1] R. Nassif, S. Vlaski, and A. H. Sayed, "Distributed inference over multitask graphs under smoothness," in Proc. IEEE International
Workshop on Signal Processing Advances in Wireless Communications, Kalamata, Greece, Jun. 2018.
[2] R. Nassif, S. Vlaski, C. Richard, and A. H. Sayed, "Learning over multitask graphs -- Part I: Stability analysis," Submitted for
publication, Nov. 2019.
[3] D. P. Bertsekas, "A new class of incremental gradient methods for least squares problems," SIAM J. Optim., vol. 7, no. 4, pp. 913 -- 926,
1997.
[4] R. Olfati-Saber, J. A. Fax, and R. M. Murray, "Consensus and cooperation in networked multi-agent systems," Proc. IEEE, vol. 95,
no. 1, pp. 215 -- 233, 2007.
[5] A. Nedic and A. Ozdaglar, "Distributed subgradient methods for multi-agent optimization," IEEE Trans. Autom. Control, vol. 54, no.
1, pp. 48 -- 61, 2009.
[6] A. G. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and A. Scaglione, "Gossip algorithms for distributed signal processing," Proc.
IEEE, vol. 98, no. 11, pp. 1847 -- 1864, 2010.
[7] S. S. Ram, A. Nedi´c, and V. V. Veeravalli, "Distributed stochastic subgradient projection algorithms for convex optimization," J. Optim.
Theory Appl., vol. 147, no. 3, pp. 516 -- 545, 2010.
[8] J. Chen and A. H. Sayed, "Distributed Pareto optimization via diffusion strategies," IEEE J. Sel. Topics Signal Process., vol. 7, no. 2,
pp. 205 -- 220, 2013.
34
[9] A. H. Sayed, "Adaptation, learning, and optimization over networks," Foundations and Trends in Machine Learning, vol. 7, no. 4-5,
pp. 311 -- 801, 2014.
[10] J. Chen and A. H. Sayed, "On the learning behavior of adaptive networks -- Part I: Transient analysis," IEEE Trans. Inf. Theory, vol.
61, no. 6, pp. 3487 -- 3517, Jun. 2015.
[11] J. Chen and A. H. Sayed, "On the learning behavior of adaptive networks -- Part II: Performance analysis," IEEE Trans. Inf. Theory,
vol. 61, no. 6, pp. 3518 -- 3548, Jun. 2015.
[12] A. H. Sayed, "Adaptive networks," Proc. IEEE, vol. 102, no. 4, pp. 460 -- 497, Apr. 2014.
[13] S. Vlaski, L. Vandenberghe, and A. H. Sayed, "Diffusion stochastic optimization with non-smooth regularizers," in Proc. Int. Conf.
Acoust., Speech, Signal Process., Shanghai, China, Mar. 2016, pp. 4149 -- 4153.
[14] J. Plata-Chaves, A. Bertrand, M. Moonen, S. Theodoridis, and A. M. Zoubir, "Heterogeneous and multitask wireless sensor networks
-- Algorithms, applications, and challenges," IEEE J. Sel. Topics Signal Process., vol. 11, no. 3, pp. 450 -- 465, Apr. 2017.
[15] A. Hassani, J. Plata-Chaves, M. H. Bahari, M. Moonen, and A. Bertrand, "Multi-task wireless sensor network for joint distributed
node-specific signal enhancement, LCMV beamforming and DOA estimation," IEEE J. Sel. Topics Signal Process., vol. 11, no. 3, pp.
518 -- 533, 2017.
[16] J. Chen, C. Richard, and A. H. Sayed, "Multitask diffusion adaptation over networks," IEEE Trans. Signal Process., vol. 62, no. 16,
pp. 4129 -- 4144, 2014.
[17] R. Nassif, C. Richard, A. Ferrari, and A. H. Sayed, "Proximal multitask learning over networks with sparsity-inducing coregularization,"
IEEE Trans. Signal Process., vol. 64, no. 23, pp. 6329 -- 6344, 2016.
[18] C. Eksin and A. Ribeiro, "Distributed network optimization with heuristic rational agents," IEEE Trans. Signal Process., vol. 60, no.
10, pp. 5396 -- 5411, Oct. 2012.
[19] D. Hallac, J. Leskovec, and S. Boyd, "Network Lasso: Clustering and optimization in large graphs," in Proc. ACM SIGKDD, Sydney,
Australia, Aug. 2015, pp. 387 -- 396.
[20] J. Szurley, A. Bertrand, and M. Moonen, "Distributed adaptive node-specific signal estimation in heterogeneous and mixed-topology
wireless sensor networks," Signal Processing, vol. 117, pp. 44 -- 60, 2015.
[21] J. Plata-Chaves, N. Bogdanovi´c, and K. Berberidis, "Distributed diffusion-based LMS for node-specific adaptive parameter estimation,"
IEEE Trans. Signal Process., vol. 63, no. 13, pp. 3448 -- 3460, 2015.
[22] S. A. Alghunaim, K. Yuan, and A. H. Sayed,
"Decentralized exact coupled optimization,"
in Proc. Ann. Allerton Conf. on
Communication, Control, and Computing, Illinois, USA, 2017, pp. 338 -- 345.
[23] R. Nassif, C. Richard, A. Ferrari, and A. H. Sayed, "Diffusion LMS for multitask problems with local linear equality constraints,"
IEEE Trans. Signal Process., vol. 65, no. 19, pp. 4979 -- 4993, 2017.
[24] J. Chen, C. Richard, and A. H. Sayed, "Diffusion LMS over multitask networks," IEEE Trans. Signal Process., vol. 63, no. 11, pp.
2733 -- 2748, 2015.
[25] J. Chen, C. Richard, A. O. Hero, and A. H. Sayed, "Diffusion LMS for multitask problems with overlapping hypothesis subspaces,"
in Proc. IEEE Int. Workshop Mach. Learn. Signal Process., Reims, France, Sep. 2014, IEEE, pp. 1 -- 6.
[26] V. Kekatos and G. B. Giannakis, "Distributed robust power system state estimation," IEEE Trans. Signal Process., vol. 28, no. 2, pp.
1617 -- 1626, 2013.
[27] D. Zhou and B. Scholkopf, "A regularization framework for learning from graph data," in Proc. ICML Workshop on Statistical
Relational Learning and Its Connections to Other Fields, 2004, vol. 15, pp. 67 -- 68.
[28] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, "The emerging field of signal processing on graphs:
Extending high-dimensional data analysis to networks and other irregular domains," IEEE Signal Process. Mag., vol. 30, no. 3, pp.
83 -- 98, May 2013.
[29] F. R. K. Chung, Spectral Graph Theory, American Mathematical Society, 1997.
[30] R. K. Ando and T. Zhang, "Learning on graph with Laplacian regularization," in Proc. Advances in neural information processing
systems, Canada, Dec. 2006, pp. 25 -- 32.
35
[31] X. Dong, D. Thanou, P. Frossard, and P. Vandergheynst, "Learning Laplacian matrix in smooth graph signal representations," IEEE
Trans. Signal Process., vol. 64, no. 23, pp. 6160 -- 6173, 2016.
[32] P.-Y. Chen and S. Liu, "Bias-variance tradeoff of graph Laplacian regularizer," IEEE Signal Process. Lett., vol. 24, no. 8, pp. 1118 -- 1122,
2017.
[33] B. T. Polyak, "Introduction to Optimization," Optimization Software, New York, 1987.
[34] R. H. Koning, H. Neudecker, and T. Wansbeek, "Block Kronecker products and the vecb operator," Linear algebra and its applications,
vol. 149, pp. 165 -- 184, Apr. 1991.
[35] D. S. Bernstein, Matrix Mathematics: Theory, Facts, and Formulas with Application to Linear Systems Theory, Princeton University
Press, 2005.
[36] L. Isserlis, "On a formula for the product-moment coefficient of any order of a normal frequency distribution in any number of
variables," Biometrika, vol. 12, no. 1/2, pp. 134 -- 139, Nov. 1918.
[37] D. I. Shuman, P. Vandergheynst, D. Kressner, and P. Frossard, "Distributed signal processing via Chebyshev polynomial approximation,"
IEEE Trans. Signal Inf. Process. Netw., 2018.
[38] A. Sandryhaila and J. M. F. Moura, "Discrete signal processing on graphs," IEEE Trans. Signal Process., vol. 61, no. 7, pp. 1644 -- 1656,
Apr. 2013.
[39] A. Ortega, P. Frossard, J. Kovacevi´c, J. M. F. Moura, and P. Vandergheynst, "Graph signal processing: Overview, challenges, and
applications," Proc. IEEE, vol. 106, no. 5, pp. 808 -- 828, 2018.
[40] T. Kailath, Linear Systems, Prentice-Hall, Englewood Cliffs, NJ, 1980.
[41] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, NY, 2004.
|
1610.08168 | 1 | 1610 | 2016-10-26T05:00:16 | Location Aggregation of Spatial Population CTMC Models | [
"cs.MA",
"cs.PF"
] | In this paper we focus on spatial Markov population models, describing the stochastic evolution of populations of agents, explicitly modelling their spatial distribution, representing space as a discrete, finite graph. More specifically, we present a heuristic approach to aggregating spatial locations, which is designed to preserve the dynamical behaviour of the model whilst reducing the computational cost of analysis. Our approach combines stochastic approximation ideas (moment closure, linear noise), with computational statistics (spectral clustering) to obtain an efficient aggregation, which is experimentally shown to be reasonably accurate on two case studies: an instance of epidemic spreading and a London bike sharing scenario. | cs.MA | cs | Location Aggregation of Spatial Population CTMC Models
Luca Bortolussi
University of Trieste, Trieste, Italy
CNR-ISTI, Pisa, Italy
[email protected]
Cheng Feng
University of Edinburgh
Edinburgh, UK
[email protected]
In this paper we focus on spatial Markov population models, describing the stochastic evolution of
populations of agents, explicitly modelling their spatial distribution, representing space as a discrete,
finite graph. More specifically, we present a heuristic approach to aggregating spatial locations,
which is designed to preserve the dynamical behaviour of the model whilst reducing the computa-
tional cost of analysis. Our approach combines stochastic approximation ideas (moment closure,
linear noise), with computational statistics (spectral clustering) to obtain an efficient aggregation,
which is experimentally shown to be reasonably accurate on two case studies: an instance of epi-
demic spreading and a London bike sharing scenario.
1 Introduction
Population processes are widely used to describe a large variety of systems, including systems in bio-
logical [17], ecology [1], performance engineering [26], smart cities facilities like bike sharing [16, 18],
and the spreading of epidemics [2]. Most of these systems are spatially distributed, and space has to be
explicitly modelled to properly capture the relevant features of their dynamics. For example, interactions
may only be allowed for entities which are co-located or within a certain physical distance of each other,
or space may be segmented in such a way that even physically close entities are unable to communicate
or interact. Furthermore, movement in space can be a crucial aspect of the behaviour of entities within
the system. Epidemic spreading, in which infection can pass only by physical contact, is a clear example
in this sense [2]. Bike sharing is another one, as the geographical location of bike stations influences
the travelling time, and different stations have different demands for acquisition or deposit of bikes at
different times of the day [16, 18].
Population processes are often modelled as Markovian stochastic processes, mostly as a subclass
of continuous-time Markov chains (CTMCs) known as Population CTMCs (PCTMCs). These models
have a very large or even infinite state space, a fact resulting in a lot of work in the computer science
community to craft efficient algorithms for their analysis, ranging from specialised stochastic simulation
to the use of mean-field [6] and moment closure techniques [20, 19]. These methods, in particular,
approximate the large or infinite set of linear Kolmogorov equations by a much smaller set of non-linear
differential equations, capturing the mean, variance, and possibly other higher order moments.
Spatial extensions of PCTMCs typically introduce discrete representations of space in terms of lo-
cality, connecting them in a general topology represented by a (weighted) graph [7]. One of the effects
of explicit modelling of space is the increase of the computational cost of analysis of the system. For
instance, a model with l localities will increase the number of equations for the variance in any moment
closure approach by a factor of O(l2).
However, although space shapes the behaviour of the system, our interest is often in space-free prop-
erties, like the total number of infected individuals in an epidemic scenario, or in localised properties,
M. Tribastone and H. Wiklicky (Eds.): International Workshop on
Quantitative Aspects of Programming Languages and Systems (QAPL'16)
EPTCS 227, 2016, pp. 30 -- 43, doi:10.4204/EPTCS.227.3
c(cid:13) L. Bortolussi & C. Feng
This work is licensed under the
Creative Commons Attribution License.
L. Bortolussi & C. Feng
31
like the number of available bikes in a given station or geographic area. In these cases, a full represen-
tation of space may not be needed to compute such quantities with a reasonable accuracy, and model
simplification and abstraction can be a viable strategy.
In this paper, we pursue this direction by developing an efficient approach to aggregate locations
showing a similar dynamical profile. This operation will preserve the dynamical behaviour of the model,
but lower the computational cost: for instance, if the number of locations decreases from l to k, then the
number of differential equations for the variance is reduced by a factor O(k2/l2). In particular, here we
will focus on the reduced cost of the analysis of the full stochastic model by using standard simulation
algorithms [17], as this is the most expensive but also most informative computational technique (ex-
cluding numerical integration of the Kolmogorov equations, which for PCTMC is unfeasible due to state
space explosion).
Our approach is based on aggregating locations using state-of-the-art spectral clustering approaches,
based on metrics between locations that take the steady state mean or distribution into account. Rather
than working with exact solution or estimations of these quantities by simulation, which would be com-
putationally expensive, we obtain them applying stochastic approximation ideas, either by solving mean-
field equations for the mean or using a linear-noise-like approximation of the distribution, obtaining mean
and variance from moment closure equations. The feasibility, effectiveness, and accuracy of our method
is discussed on two case studies: an epidemic model and a London bike sharing scenario.
The paper is organised as follows: Section 2 introduces the formalism of spatial PCTMCs we will
use afterwards and briefly introduces exact and approximate analysis techniques. Section 3 discusses in
detail our aggregation approach, and Section 4 presents the two case studies. Conclusions are drawn in
Section 5.
2 Spatial PCTMC Models
A Population Continuous Time Markov Chain (PCTMC) is a Markovian stochastic process evolving in
continuous time. It consists of a number of individually indistinguishable interacting agents of different
types which can be in different internal states, so that the state of the system can be described by counting
how many agents of each kind are in the system. Agents' interactions are described by a set of transitions,
which will change the internal state of one or more agents. In this paper, we specifically consider spatial
PCTMC models in which agents are distributed in a finite set of discrete locations, with interactions
typically happening either in the same location or in neighbouring ones. Formally, a spatial PCTMC can
be expressed as a tuple P = (L ,X, T ):
• L = ((cid:96)1, ..., (cid:96)l) is the set of discrete locations in the model, where L = l denotes the total number
of locations.
• X = (X@(cid:96)1, ...,X@(cid:96)l) ∈ ZN≥0 is an integer vector representing the agent populations distributed
over all the locations in the model, where X = N = l × n is the total number of distinct agent
populations in the model with n representing the number of agent types. X@(cid:96)i is an sub-vector
with the jth component, Xj@(cid:96)i representing the current population of the agent type j at location
i. We use X0 to denote the initial state of the model.
• T = {τ1, ...,τm} is the set of transitions with size T = m, of the form τ = (rτ (X),Dτ ), where:
1. rτ (X) ∈ R ≥ 0 is the rate function, associating with each transition the rate of an exponential
distribution, depending on the global state of the model.
32
Aggregation of Spatial PCTMC Models
2. Dτ ∈ ZN is the update vector which gives the net change for each element of X caused by
transition τ. Similarly, we let Dτ@(cid:96)i denote the update vector of agent populations at location
i.
Transition rules can be easily visualised in the chemical reaction style, as
τ@(cid:96)1 + . . . + un
u1
τ@(cid:96)l −→ v1
τ@(cid:96)1 + . . . + vn
τ@(cid:96)l
at rate rτ (X)
τ@(cid:96) j −
where the net change of agents of type i in location (cid:96) j due to transition τ is given by Di
τ@(cid:96) j (1 ≤ i ≤ n, 1 ≤ j ≤ l). This is a general format, encompassing local interactions, interactions
ui
between agents in neighbouring locations, and movement in space. The tuple P contains all the infor-
mation that is needed to build a CTMC on the state space ZN≥0: its infinitesimal generator Q is given by
Q(x,y) = ∑{rτ (x)Dτ = y− x}, for x (cid:54)= y. Spatial PCTMC can be simulated using standard stochastic
simulation algorithms [17].
τ@(cid:96) j = vi
2.1 Moment Closure, Mean Field, and Linear Noise
The analysis of the stochastic model underlying a spatial PCTMC is not an easy task. Numerical methods
are hindered by the state space explosion, and even stochastic simulation suffers from the presence of
localities and large populations. Furthermore, it is known that for large populations the behaviour of
the stochastic model becomes deterministic and converges to the solution of a the mean field differential
equation [6, 22], which in the formalism defined above takes the form
X = F(X),
with F(X) = ∑
τ
Dτrτ (X)
If populations in each location are in the order of tens or hundreds of thousands, mean-field equations are
generally very accurate [6]. However, when populations are smaller, in the order of hundreds, stochastic
fluctuations still have an important role, and the linear-noise (central limit) approximation performs better
[22, 21]. The idea behind linear noise is to approximate the original stochastic model by a linearized
Markov process in continuous states, whose solution is a Gaussian process. The distribution of this
process at a given time is thus characterized by solving equations for the mean (incidentally, the mean-
field equations above) and equations for the covariance.
An alternative strategy for the analysis of the spatial PCTMC is to derive equations for the moments
of the population variables [20, 19], up to order k. Due to non-linearities in the rates, there is no exact
closed form of these equations, and the differential equations for moments of order k depend on higher-
order ones. Hence, equations are closed by relying to some heuristic [20, 19]. Typically, moment closure
equations give a better estimation of mean and variance than linear noise, as knowledge of higher-order
moments introduce correction terms in the equations of lower order ones. The first two moments can still
be used to build a Gaussian approximation to the true distribution at a given time t, formally invoking a
maximum entropy argument [3, 9].
3 Aggregation of Locations of Spatial PCTMC
In this section we present the computational methodology to reduce a spatial model by aggregating
locations. The main motivation underlying this approach, as discussed in the introduction, is to reduce
the computational effort in the analysis of the model. This effort, in fact, is proportional to the number
of locations. This is the case for stochastic models, which need to simulate and keep track of the state of
L. Bortolussi & C. Feng
33
populations in each location, and for approximate analysis techniques like moment closure. In fact, even
when considering only equations for the second order moments, the number of such equations grows
quadratically with the number of locations. However, if our final goal is to capture the behaviour of the
system at the global level (i.e. to know how many agents of each kind are in the system), then dealing with
the full set of locations may incur excessive work which can be reduced by grouping together locations
having a similar overall behaviour. To achieve this goal while keeping the error low, we propose the
following heuristic scheme, sketched here and detailed in the following subsections.
1. Define distance metrics between locations taking into account the dynamical behaviour of the sys-
tem (Section 3.1). More specifically, we want locations showing a similar steady state behaviour
to be clustered together. We will define two distances, described below in increasing order of
precision:
(a) Mean field distance. We will consider the distance between the mean of populations at steady
state, approximated by first order moment closure, which corresponds to the mean field ab-
straction of the PCTMC model (see Section 2.1).
(b) Linear noise distance. We will consider the distance between a Gaussian approximation of
the steady state distribution, computed from moment closure equations for mean and vari-
ance.
2. Cluster the locations using the previously computed distance (Section 3.2). We will use a spectral
clustering algorithm on graphs, exploiting the eigengap heuristic to identify the number of clusters.
3. Given a clustering of locations, construct the reduced model by suitably aggregating together the
PCTMC transitions and variables (Section 3.3).
3.1 Distance between Locations
We will consider distances between populations of different locations at a given time. The choice of this
time is important, and should not be taken too small, in order to minimise the effect of the initial state
of the model. In fact, if we took information about the whole trajectory into account, different initial
conditions between the populations of two locations would contribute to the distance, often resulting in
a separation between the two locations. However, especially if we consider aggregated quantities, like
the total number of agents of a certain kind across all locations, this difference is not very relevant, and
better results can be obtained by comparing the behaviour after a finite but large time. We will see the
experimental validation of this choice in the next Section. Note that we do not consider steady state
behaviour (though for a very large time steady state would be reached), as we will use mean-field or
linear noise approximations, which do not necessarily converge at steady state. We will consider two
distances between each pair of locations (cid:96)i and (cid:96) j, with increasing levels of accuracy:
1. Mean field distance dE((cid:96)i, (cid:96) j). This is just the Eucliden distance between the average value of
populations of locations (cid:96)i and (cid:96) j at steady state, i.e. dE((cid:96)i, (cid:96) j) = (cid:107)µX@(cid:96)i − µX@(cid:96) j(cid:107), where µX@(cid:96) is
the mean of the population (vector) X@(cid:96) at steady state. To compute this distance efficiently, we
resort to the mean-field approximation or to any first order moment closure, see Section 2.1. This
reduces the problem to the numerical integration of N differential equations (N = n× l), until they
reach equilibrium (or until their temporal average stabilise in case they oscillate).1
1We are implicitly assuming that the mean field or moment closure equations will not show chaotic behaviour, which is
typically the case for models of interacting agents.
34
Aggregation of Spatial PCTMC Models
2. Linear noise distance dL((cid:96)i, (cid:96) j). To capture more accurately the steady state behaviour, we can
consider a distance between the full distributions. To this end, we will resort to the Bhattacharyya
distance dB(Y1,Y2) [4] which measures the distance of the distribution of two one-dimensional vari-
ables Y1,Y2. more specifically, we will compute the Bhattacharyya distance between each popou-
lation k in locations (cid:96)i and (cid:96) j, denoted by dB(Xk@(cid:96)i,Xk@(cid:96) j), and aggregate over all populations
by taking the average:
dL((cid:96)i, (cid:96) j) =
1
n (dB(X1@(cid:96)i,X1@(cid:96) j) +··· + dB(Xn@(cid:96)i,Xn@(cid:96) j))
In order to compute dB(Xk@(cid:96)i,Xk@(cid:96) j) without the need to estimate the full distribution, we make
a linear noise assumption, i.e.
that the steady state distribution is approximately Gaussian, see
Section 2.1. Under this hypothesis, dB(Xk@(cid:96)i,Xk@(cid:96) j) can be calculated by the following equation:
(cid:33)(cid:33)
(cid:32)
(cid:33)
(cid:32)
(cid:32) σ 2
DB(Xk@(cid:96)i,Xk@(cid:96) j) =
1
4
ln
1
4
Xk@(cid:96)i
σ 2
Xk@(cid:96) j
+
σ 2
Xk@(cid:96) j
σ 2
Xk@(cid:96)i
+ 2
+
1
4
(µXk@(cid:96)i − µXk@(cid:96) j )2
σ 2
Xk@(cid:96)i
+ σ 2
Xk@(cid:96) j
where µX@(cid:96) and σX@(cid:96) denote the mean and variance of the population variable X@(cid:96) at the steady
state. To numerically compute the values of µX@(cid:96) and σX@(cid:96), we resort to the normal moment-
closure approximation of [19], which can be obtained at a much lower computational cost than by
simulating the PCTMC, by integrating O(N2) differential equations.
We observe that the cost of computing dL((cid:96)i, (cid:96) j) is significantly higher than the cost of computing
dE((cid:96)i, (cid:96) j), as we need to integrate O(N2) differential equations rather than O(N). However, this cost
is balanced by a higher accuracy in the reduced system, though for certain models (essentially those hav-
ing similar variance for the same agent kind at different locations in a cluster) accuracy is comparable. In
practice, dE should be used when the cost of solving moment closure equations for the variance is very
high due to the very large number of locations. Hence, aggregation with respect to dE can be seen as a
reduction of the number of moment closure equations. When the goal is to reduce the cost of stochastic
simulation, and solving moment closure equations for the variance is cheap, it is better to rely on the
metric dL. In general, our method should be applied when stochastic simulation of the spatial PCTMC
model requires excessive computational time, and computing dE or dL is much cheaper than the cost of
stochastic simulation.
3.2 Spectral Clustering
Spectral clustering methods are common graph-based approaches to (unsupervised) clustering of data
[28]. The dataset is composed of n objects S = s1, . . . ,sn, among which some local symmetric and non-
negative similarity measure Ai, j is defined. Ai j is often obtained from a distance or difference measure
d(si,s j) between the objects. This information is then arranged in a weighted graph G = (S,A). Within
this framework, clustering is translated into a graph partitioning problem. The most common class
of spectral approaches for graph partitioning (in k subsets) is to map the original data into the first k
eigenvectors of some normalized version of the similarity matrix A and then apply a standard clustering
algorithm such as k-means on these new coordinates.
Among the most commonly used spectral clustering algorithms are the unnormalized spectral clus-
tering [28], normalized spectral clustering according to Shi and Malik [25], and normalized spectral
clustering according to Ng, Jordan, and Weiss [24]. The three algorithms are very similar; their main
distinguishing feature is the fact that they use three different graph Laplacians. As an illustration, the
L. Bortolussi & C. Feng
35
normalized spectral clustering algorithm according to Ng, Jordan, and Weiss is given as follows. Given
a set of objects S = s1, . . . ,sn that we want to cluster into k subsets:
1. Form the similarity matrix A ∈ Rn×n defined by Ai j = exp(− d(si−s j)2
2. Define D to be the diagonal matrix whose (i,i)-element is the sum of A's i-th row, and construct
) if i (cid:54)= j, and Aii = 0.
2σ 2
the normalized Laplacian matrix L = I − D−1/2AD−1/2
3. Find U1,U2, ...,Uk , the first k eigenvectors of L with k smallest eigenvalues, and form the matrix
U = [U1,U2, ...,Uk] ∈ Rn×k by stacking the eigenvectors in columns.
4. Form the matrix Z from U by renormalizing each of U's rows to have unit length (i.e. Zi j =
Ui j/(∑ j U 2
i j)1/2).
5. Treating each row of Z as a point in Rk , cluster them into k clusters via k-means.
6. Finally, assign the original object si to cluster j if and only if row i of the matrix Z was assigned to
cluster j.
We will use the above algorithm hereafter for the clustering of locations in spatial PCMTCs .
3.2.1 Application to Aggregation of Locations of spatial PCTMC
Given a spatial PCTMC model we wish to aggregate its locations L = ((cid:96)1, ..., (cid:96)l). The first step is to
compute one of the two distances dE or dL of the previous section. From this metric, we can derive the
similarity matrix A ∈ Rl×l for the locations L by using Gaussian kernel with width σ:
(cid:18)
−d(cid:63)((cid:96)i, (cid:96) j)2
(cid:19)
Ai j = exp
2σ 2
Then, the standard spectral clustering algorithms can be applied to cluster the locations with a specific
choice of the number of clusters k. We will refer to the locations' clusters by L = ( (cid:96)1, ..., (cid:96)k).
In order to select k, we rely on the eigengap heuristic [23]. Specifically, we choose the number of
clusters k such that all eigenvalues λ1, . . . ,λk are very small, but λk+1 is relatively large. A realisation of
this heuristic will be shown while discussing case studies.
3.3 Model Reduction
In this section, we show how to generate a reduced version of a spatial PCTMC model once the locations
L . Suppose we have clustered
L in the original model have been clustered into the aggregated locations
locations in a spatial PCTMC into k subsets. Formally, we want to map the original spatial PCTMC
L = k < l = L , X < X, and
model P = (L ,X, T ) to a reduced one
T ≤ T . Hence, we need to construct both the aggregated vector of the populations of agents and the
reduced set of transitions.
T ), where
P = ( L , X,
3.3.1 Generating the aggregated vector of agent populations
The algorithm for generating the aggregated vector of agent populations is fairly straightforward. The
basic idea is to treat agents of the same type in the same cluster of locations as identical agents. Thus,
we only need to sum up the populations of those identical agents. Algorithm 1 gives the corresponding
pseudo code for generating the aggregated vector of agent populations. Note that in the pseudo code, we
use the same notations for the original spatial PCTMC model as in Section 2.
36
Aggregation of Spatial PCTMC Models
L
for j = 1 to c do
Algorithm 1 The Algorithm for generating the aggregated vector of agent populations
Require: X, X, L ,
Ensure: X = n× k∧ Xi@ (cid:96) j = 0 ∀ Xi@ (cid:96) j ∈ X {all elements in X are set to zero initially}
1: for i = 1 to l do
2:
3:
4:
5:
6:
7: end for
8: return X
if (cid:96)i belongs to cluster j then
let X@ (cid:96) j = X@ (cid:96) j + X@(cid:96)i
end if
end for
3.3.2 Generating the reduced set of transitions
Three steps are taken to generate the reduced set of transitions. The first step is to copy the transitions in
the original spatial PCTMC model to T . Meanwhile, the update of agent populations in the transition
should be replaced by update of corresponding aggregated populations, the agent populations appearing
in the rate function should also be replaced accordingly. The first step may generate many redundant
transitions in which there is no agent population being updated. Thus, the second step is to remove
redundant transitions from T . The last step is de-duplication, in which we combine all transitions with
the same update vector into one transition. Algorithm 2 gives the pseudo code for the three steps, in which
we use (cid:96)i to denote the number of locations belonging to cluster (cid:96)i. Note that this construction will
produce an approximate model with respect to the original one, the reason being the treatment of rates.
Approximation stems from the fact that in the rates of the aggregated model we replace all occurrences
of each X@(cid:96) with X@ (cid:96)/ (cid:96), where (cid:96) ∈ (cid:96). Hence we assume that for each (cid:96)i, (cid:96) j ∈ (cid:96), X@(cid:96)i = X@(cid:96) j.
However, the heuristics we use to construct the aggregated model guarantee that this condition should be
roughly satisfied (at steady state).
4 Case Studies
In this section we test our method on two case studies: a benchmark model of epidemic spreading,
Section 4.1, and a realistic model of a portion of the London bike-sharing system, Section 4.2.
4.1 Spatial Epidemic Spreading Model
We first consider a classical epidemiological SIS model of individuals partitioned into m communities,
where individuals move between communities but infections only take place within communities. Each
individual is considered to be susceptible (S) or infected (I) with respect to the disease. A continuous-
time SIS epidemiological model is then applied to this population as follows: each individual, regardless
of susceptible or infected, can move to his/her connected communities with a specific rate (ri j). Each
community is connected with three other randomly chosen communities. Each infected individual can
randomly make contact with a susceptible individual in the same community, and infect her with rate βi
in community i. Finally, infected individuals independently recover to the susceptible state at rate µ.
L. Bortolussi & C. Feng
37
L , T ,
T
{all elements in D τ are set to zero initially}
for j = 1 to k do
end if
end for
end for
for all X@(cid:96) appears in r τ do
for i = 1 to k do
if (cid:96)i belongs to cluster j then
let D τ@ (cid:96) j = D τ@ (cid:96) j + Dτ@(cid:96)i
if (cid:96) belongs to cluster i then
replace X@(cid:96) with X@ (cid:96)i/ (cid:96)i
create a new τ, set r τ = rτ, D τ = n× k
for i = 1 to l do
Algorithm 2 The Algorithm for generating the reduced set of transitions
Require: L ,
Ensure: T = 0
1: for all τ in T do
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17: end for
18: for all τ in T do
19:
20:
end if
21:
22: end for
23: for all τi, τ j (i (cid:54)= j) in T do
24:
25:
26:
end if
27:
28: end for
29: return T
create a new τ, set D τ = D τi and r τ = r τi + r τ j
remove τi, τ j from T , add τ to T
if All elements in D τ equal zero then
end if
end for
end for
remove τ from T
if D τi = D τ j then
The model can be studied by a spatial PCTMC containing the following transitions:
S((cid:96)i) → I((cid:96)i)
I((cid:96)i) → S((cid:96)i)
S((cid:96)i) → S((cid:96) j)
I((cid:96)i) → I((cid:96) j)
at βi #(S((cid:96)i)) #(I((cid:96)i))
at µ #(I((cid:96)i))
at ri j #(S((cid:96)i))
at ri j #(I((cid:96)i))
∀i ∈ m
∀i ∈ m
∀i, j ∈ m, (cid:96)i and (cid:96) j are connected
∀i, j ∈ m, (cid:96)i and (cid:96) j are connected
where S((cid:96)i), I((cid:96)i) represent a susceptible, infected individual in Community i; βi represents the contact
rate in community i; ri j denotes the rate for an individual to travel from Community i to Community j.
In our experiment, we consider a model with m = 30, βi and ri j to be random values between zero
and one, µ = 0.1. We first report the analysis using the Linear-Noise metric dL. Computing the distance
and running the spectral clustering algorithm, according to the previous section, we obtained the spectre
of the normalized Laplacian matrix shown in Figure 1 (left), for the smallest 10 eigenvalues. As can be
seen from the figure, the first four eigenvalues are very close to zero, and there is a large gap between
the 4th and 5th eigenvalues. Thus, we set the number of clusters to four. Figure 2 (left) shows the
38
Aggregation of Spatial PCTMC Models
Figure 1: (left) The smallest 10 eigenvalues of the normalized Laplacian matrix of the SIS model. (right)
The smallest 10 eigenvalues of the normalized Laplacian matrix of the bike-sharing model.
Figure 2: (left) Comparison of the infected populations before and after aggregation for the SIS model,
using the dL distance. (right) Comparison of the infected populations before and after aggregation for
the SIS model, using the dE distance.
trajectories of the infected population generated by stochastic simulation before and after aggregation.
Table 1 compares the number of transitions, simulation time of 1000 runs of the SIS model before and
after aggregation, as well as the average error ratio of the trajectories in Figure 2 (left) after aggregation
compared with the counterpart before aggregation. As can be readily seen, our method considerably
reduced the simulation cost of the model, at the price of a reasonably small loss of accuracy, the relative
error being less than 10% (the average error ratio is computed as the average along the trajectory). In
Figure 2 (right), instead, we show the result of the aggregation when using the mean-field distance dE.
In this case, the spectral clustering identifies 5 clusters, with a comparable overall accuracy with respect
to the linear-noise distance as can be seen in Table 1.
4.2 Public Bike-sharing Model
The second example is a spatial PCTMC which models a public bike-sharing system. Bike-sharing
systems are becoming more and more important for urban transportation. In such systems, users arrive
L. Bortolussi & C. Feng
39
SIS model
No. of transitions
simulation time (1000 runs) Avg error ratio
Before aggregation
After aggregation (dL)
After aggregation (dE)
240
28
34
8.81 mins
59 secs
3.67 mins
N/A
8.67%
10.44%
Table 1: Size, simulation cost (including the aggregation cost) of the SIS model before and after aggre-
gation, and error introduced by the aggregation.
Figure 3: The map of bike-sharing stations near Russell Square in London in which red circles repre-
senting bike stations
at a station, pick up a bike, use it for a while, and then return it to another station of their choice.
Recently, PCTMCs have been used to model bike-sharing systems [16, 18]. Here, we consider a map
which consists of N zones. There is one public bike station in each zone. Each station has several bike
slots. The pickup rate of bikes in a station is governed by an exponential distribution. When a user
picks up a bike, the available number of bikes in the station will decrease by one whereas the available
number of slots in that station will increase by one. The user will choose another zone in the city as their
destination. When the user arrives at the destination zone, they will return their bike to the bike station
in that zone. We use a spatial PCTMC containing the following transitions to represent the model:
Bike((cid:96)i) → Slot((cid:96)i) + BikeTo j((cid:96)i)
BikeTo j((cid:96)i) + Slot((cid:96) j) → Bike((cid:96) j)
at λi pi
j
at #(BikeTo j((cid:96)i)) µi
j
where Bike((cid:96)i) and Slot((cid:96)i) denote an available bike or slot in the bike station in Zone i, respectively;
BikeTo j((cid:96)i) denotes a bike in transit from Zone i to Zone j. λi is the pickup rate of bikes in the bike
station in Zone i, pi
j is the probability to choose Zone j as the destination of a trip when picking up a
bike from Zone i. 1/µi
j is the mean trip time from Zone i to Zone j.
Specifically, we use the above spatial PCTMC to model 30 bike stations near Russell Square in
London which is illustrated in Figure 3. All the rates in the model are calculated by journey data which
is available online 2.
2https://tfl.gov.uk/info-for/open-data-users/our-feeds?intcmp=3671#on-this-page-4
40
Aggregation of Spatial PCTMC Models
Bike model
No. of transitions
simulation time (1000 runs) Avg error ratio
Before aggregation
After aggregation (dL)
After aggregation (dE)
1800
300
180
15.6 mins
2.5 mins
1.6 mins
N/A
11.94%
23.06%
Table 2: Size, simulation cost (including the aggregation cost) of the bike-sharing model before and after
aggregation, and error introduced by the aggregation.
We apply our method with the linear noise distance to cluster the 30 bike stations. The smallest 10
eigenvalues of the normalized Laplacian matrix, computed according to the spectral clustering recipe,
are shown in Figure 1 (right). According to the eigengap heuristic, there are 5 well separated clusters.
Figure 4 (left) shows the trajectories of the number of available bikes in the 5 clusters generated by
stochastic simulation before and after aggregation. Table 2 shows the number of transitions, simulation
time of 1000 runs of the bike-sharing model before and after aggregation, as well as the average error
ratio of the trajectories in Figure 4 (left) after aggregation compared with the counterpart before aggre-
gation. In Figure 4 (right), instead, we compare the trajectories of the original model and the reduced
one according to the mean-field metric dE. In this case, we have only three clusters and the accuracy
decreases considerably, as can be numerically seen in Table 2.
Figure 4: (left) Comparison of the available number of bikes before and after aggregation using the dL
distance. (right) Comparison of the available number of bikes before and after aggregation using the dE
distance.
In a bike-sharing scenario, we are often interested in tracking the number of bikes in some specific
locations. This can be achieved in our framework by forcing some locations to be a singleton cluster.
In order to understand the influence of the aggregation of remaining stations on some isolated ones, we
choose, in each experiment, one station from one of the five clusters, and treat that station as a single
cluster. Figure 5 shows the trajectories of the number of available bikes of the five chosen stations in
five different experiments, comparing results of stochastic simulation of the original and the aggregated
model. We can see the population dynamics of available bikes in the five stations still achieve good
accuracy (10.84% average error ratio) even if other bike stations are aggregated.
L. Bortolussi & C. Feng
41
Figure 5: Comparison of the trajectories of the number of available bikes in the five chosen stations from
the original and the aggregated model.
Figure 6: Comparison of the available number of bikes before and after aggregation according to physical
position of stations
Additionally, we also apply a naive approach in which we use spectral clustering to cluster the 30
bike stations according to their physical positions on the map. Figure 6 shows the trajectories of the
number of available bikes generated by stochastic simulation before and after aggregation in this case.
The average error ratio for the trajectories is 24.51%. Thus, location clustering according to agents'
population distribution outperforms clustering according to physical positions.
5 Related Work and Conclusion
Spatiality has been shown to be an important factor for modelling different classes of complex systems
[10, 11]. As a result, there has been some effort to study the effects of spatial aggregation on stochastic
models in order to achieve a balance point between model complexity and accuracy. Most of this work
42
Aggregation of Spatial PCTMC Models
focuses on how to discretise continuous space or on the effect of compartment size to represent the space,
see for example [13, 12, 5]. Our work can be thought of as above theses approaches, as our goal is to
reduce the number of patches (locations). We show that by clustering locations in a spatial PCTMC
according to linear noise or mean field distance, the size of the model can be significantly reduced
but still retain reasonable accuracy. The experiment on the bike-sharing model shows that our method
outperforms aggregation of locations by their physical distance. Moreover, our recent work shows that
the cost of deriving higher moments based on fluid approximation can be significantly reduced using
correlation heuristics [15]. This means that computing linear noise distance can also be largely reduced,
which makes our method more scalable. In the future, we plan to apply this approach to models in spatial
process algebras, such as PALOMA [14], CARMA [8] and MELA [27].
6 Acknowledgments
The authors would like to thank Jane Hillston for her helpful comments on an earlier draft of this work.
This work is supported by the EU project QUANTICOL, 600708.
References
[1] Linda JS Allen (2010): An introduction to stochastic processes with applications to biology. CRC Press.
[2] H. Andersson & T. Britton (2000): Stochastic Epidemic Models and Their Statistical Analysis. Springer-
Verlag, doi:10.1007/978-1-4612-1158-7.
[3] A. Andreychenko, L. Mikeev & C. Wolf (2015): Model Reconstruction for Moment-Based Stochastic
Chemical Kinetics. ACM Transactions on Modeling and Computer Simulation (TOMACS) 25(2), p. 12,
doi:10.1145/2699712.
[4] A Bhattacharyya (1943): On a measure of divergence between two statistical population defined by their
population distributions. Bulletin Calcutta Mathematical Society 35, pp. 99 -- 109.
[5] Juan A Bonachela, Miguel A Munoz & Simon A Levin (2012): Patchiness and demographic noise in three
ecological examples. Journal of Statistical Physics 148(4), pp. 724 -- 740, doi:10.1007/s10955-012-0506-x.
[6] L. Bortolussi, J. Hillston, D. Latella & M. Massink (2013): Continuous approximation of collective system
behaviour: A tutorial. Performance Evaluation 70(5), pp. 317 -- 349, doi:10.1016/j.peva.2013.01.001.
[7] L. Bortolussi & L. Nenzi (2014): Specifying and monitoring properties of stochastic spatio-temporal systems
in signal temporal logic. In: Proceedings of the 8th International Conference on Performance Evaluation
Methodologies and Tools, pp. 66 -- 73, doi:10.4108/icst.Valuetools.2014.258183.
[8] Luca Bortolussi, Rocco De Nicola, Vashti Galpin, Stephen Gilmore, Jane Hillston, Diego Latella, Michele
Loreti & Mieke Massink (2015): CARMA: Collective Adaptive Resource-sharing Markovian Agents.
In:
Proceedings Thirteenth Workshop on Quantitative Aspects of Programming Languages and Systems, QAPL
2015, London, UK., pp. 16 -- 31, doi:10.4204/EPTCS.194.2.
[9] Luca Bortolussi & Roberta Lanciani (2014): Stochastic approximation of global reachability probabil-
In: Computer Performance Engineering, Springer, pp. 224 -- 239,
ities of Markov population models.
doi:10.1007/978-3-319-10885-8 16.
[10] Richard Durrett & Simon Levin (1994): The importance of being discrete (and spatial). Theoretical popula-
tion biology 46(3), pp. 363 -- 394, doi:10.1006/tpbi.1994.1032.
[11] Rick Durrett
(1999):
doi:10.1137/S0036144599354707.
Stochastic
spatial models.
SIAM review 41(4),
pp.
677 -- 718,
[12] Radek Erban & S Jonathan Chapman (2009): Stochastic modelling of reaction -- diffusion processes: algo-
rithms for bimolecular reactions. Physical biology 6(4), p. 046001, doi:10.1088/1478-3975/6/4/046001.
L. Bortolussi & C. Feng
43
[13] Cheng Feng (2014): Patch-based Hybrid Modelling of Spatially Distributed Systems by Using Stochas-
tic HYPE - ZebraNet as an Example.
In: Proceedings Twelfth International Workshop on Quanti-
tative Aspects of Programming Languages and Systems, QAPL 2014, Grenoble, France., pp. 64 -- 77,
doi:10.4204/EPTCS.154.5.
[14] Cheng Feng & Jane Hillston (2014): PALOMA: A process algebra for located markovian agents. In: Quan-
titative Evaluation of Systems, Springer, pp. 265 -- 280, doi:10.1007/978-3-319-10696-0 22.
[15] Cheng Feng, Jane Hillston & Vashti Galpin (2016): Automatic Moment-Closure Approximation of Spa-
tially Distributed Collective Adaptive Systems. ACM Transactions on Modeling and Computer Simulation
(TOMACS) 26(4), doi:10.1145/2883608.
[16] Christine Fricker & Nicolas Gast (2014):
systems with stations of finite capacity.
doi:10.1007/s13676-014-0053-5.
Incentives and redistribution in homogeneous bike-sharing
EURO Journal on Transportation and Logistics, pp. 1 -- 31,
[17] Daniel T Gillespie (1977): Exact stochastic simulation of coupled chemical reactions. The Journal of Physical
Chemistry 81(25), pp. 2340 -- 2361, doi:10.1021/j100540a008.
[18] Marcel C Guenther & Jeremy T Bradley (2013): Journey data based arrival forecasting for bicycle hire
In: Analytical and Stochastic Modeling Techniques and Applications, Springer, pp. 214 -- 231,
schemes.
doi:10.1007/978-3-642-39408-9 16.
[19] Marcel C Guenther, Anton Stefanek & Jeremy T Bradley (2013): Moment closures for performance models
with highly non-linear rates. In: Computer Performance Engineering, Springer, pp. 32 -- 47, doi:10.1007/978-
3-642-36781-6 3.
[20] J. Hasenauer, V. Wolf, A. Kazeroonian & F. J. Theis (2013): Method of conditional moments (MCM) for
the Chemical Master Equation: A unified framework for the method of moments and hybrid stochastic-
deterministic models. Journal of Mathematical Biology, doi:10.1007/s00285-013-0711-5.
[21] N. G. van Kampen (2007): Stochastic processes in physics and chemistry. Elsevier.
[22] T. G. Kurtz (1981): Approximation of population processes. SIAM, doi:10.1137/1.9781611970333.
[23] Bojan Mohar (1997): Some applications of Laplace eigenvalues of graphs. In: Graph Symmetry, NATO ASI
Series 497, Springer Netherlands, pp. 225 -- 275, doi:10.1007/978-94-015-8937-6 6.
[24] Andrew Y Ng, Michael I Jordan, Yair Weiss et al. (2002): On spectral clustering: Analysis and an algorithm.
Advances in neural information processing systems 2, pp. 849 -- 856, doi:10.1.1.19.8100.
[25] Jianbo Shi & Jitendra Malik (2000): Normalized cuts and image segmentation. IEEE Transactions on Pattern
Analysis and Machine Intelligence 22(8), pp. 888 -- 905, doi:10.1109/34.868688.
[26] Anton Stefanek, Richard A Hayden & Jeremy T Bradley (2011): Fluid computation of the performance:
energy tradeoff in large scale Markov models. ACM SIGMETRICS Performance Evaluation Review 39(3),
pp. 104 -- 106, doi:10.1145/2160803.2160817.
[27] Ludovica Luisa Vissat, Jane Hillston, Glenn Marion & Matthew J Smith (to appear): MELA: Modelling
in Ecology with Location Attributes. Proceedings of the fourth international symposium on Modelling and
Knowledge Management applications: Systems and Domains, MoKMaSD 2015, York, UK.
[28] Ulrike Von Luxburg (2007): A tutorial on spectral clustering. Statistics and computing 17(4), pp. 395 -- 416,
doi:10.1007/s11222-007-9033-z.
|
1803.01106 | 2 | 1803 | 2018-08-11T15:04:31 | Model-Based Stochastic Search for Large Scale Optimization of Multi-Agent UAV Swarms | [
"cs.MA"
] | Recent work from the reinforcement learning community has shown that Evolution Strategies are a fast and scalable alternative to other reinforcement learning methods. In this paper we show that Evolution Strategies are a special case of model-based stochastic search methods. This class of algorithms has nice asymptotic convergence properties and known convergence rates. We show how these methods can be used to solve both cooperative and competitive multi-agent problems in an efficient manner. We demonstrate the effectiveness of this approach on two complex multi-agent UAV swarm combat scenarios: where a team of fixed wing aircraft must attack a well-defended base, and where two teams of agents go head to head to defeat each other. | cs.MA | cs | Model-Based Stochastic Search for Large Scale Optimization of
Multi-Agent UAV Swarms
David D. Fan1, Evangelos Theodorou2, and John Reeder3
8
1
0
2
g
u
A
1
1
]
A
M
.
s
c
[
2
v
6
0
1
1
0
.
3
0
8
1
:
v
i
X
r
a
Abstract -- Recent work from the reinforcement
learning
community has shown that Evolution Strategies are a fast and
scalable alternative to other reinforcement learning methods.
In this paper we show that Evolution Strategies are a special
case of model-based stochastic search methods. This class of
algorithms has nice asymptotic convergence properties and
known convergence rates. We show how these methods can
be used to solve both cooperative and competitive multi-
agent problems in an efficient manner. We demonstrate the
effectiveness of this approach on two complex multi-agent UAV
swarm combat scenarios: where a team of fixed wing aircraft
must attack a well-defended base, and where two teams of
agents go head to head to defeat each other†.
I. INTRODUCTION
Reinforcement Learning is concerned with maximizing
rewards from an environment through repeated interactions
and trial and error. Such methods often rely on various
approximations of the Bellman equation and include value
function approximation, policy gradient methods, and more
[1]. The Evolutionary Computation community, on the other
hand, have developed a suite of methods for black box
optimization and heuristic search [2]. Such methods have
been used to optimize the structure of neural networks for
vision tasks, for instance [3].
Recently, Salimans et al. have shown that a particular
variant of evolutionary computation methods, termed Evo-
lution Strategies (ES) are a fast and scalable alternative to
other reinforcement learning approaches, solving the difficult
humanoid MuJoCo task in 10 minutes [4]. The authors
argue that ES has several benefits over other reinforcement
learning methods: 1) The need to backpropagate gradients
through a policy is avoided, which opens up a wider class
of policy parameterizations; 2) ES methods are massively
parallelizable, which allows for scaling up learning to larger,
more complex problems; 3) ES often finds policies which
are more robust than other reinforcement learning methods;
and 4) ES are better at assigning credit to changes in the
policy over longer timescales, which enables solving tasks
with longer time horizons and sparse rewards. In this work
we leverage all four of these advantages by using ES to
solve a problem with: 1) a more complex and decipherable
policy architecture which allows for safety considerations; 2)
a large-scale simulated environment with many interacting
1 Inst. for Robotics and Intelligent Machines, Georgia Institute of
Technology. [email protected]
2 Dept. of Aerospace Engineering, Georgia Institute of Technology.
[email protected]
3 SPAWAR Systems Center Pacific, San Diego, CA, USA.
[email protected]
† Video at http://goo.gl/dWvQi7
Fig. 1: The SCRIMMAGE multi-agent simulation environ-
ment. In this scenario, blue team fixed-wing agents attack
red team quadcopter defenders. White lines indicate missed
shots.
elements; 3) multiple sources of stochasticity including vari-
ations in intial conditions, disturbances, etc.; and 4) sparse
rewards which only occur at the very end of a long episode.
A common critique of evolutionary computation algo-
rithms is a lack of convergence analysis or guarantees.
Of course, for problems with non-differentiable and non-
convex objective functions, analysis will always be diffi-
cult. Nevertheless, we show that the Evolution Strategies
algorithm proposed by [4] is a special case of a class of
model-based stochastic search methods known as Gradient-
Based Adaptive Stochastic Search (GASS) [5]. This class of
methods generalizes many stochastic search methods such as
the well-known Cross Entropy Method (CEM) [6], CMA-
ES [7], etc. By casting a non-differentiable, non-convex
optimization problem as a gradient descent problem, one can
arrive at nice asymptotic convergence properties and known
convergence rates [8].
With more confidence in the convergence of Evolution
Strategies, we demonstrate how ES can be used to efficiently
solve both cooperative and competitive large-scale multi-
agent problems. Many approaches to solving multi-agent
problems rely on hand-designed and hand-tuned algorithms
(see [9] for a review). One such example, distributed Model
Predictive Control, relies on independent MPC controllers
on each agent with some level of coordination between
them [10], [11]. These controllers require hand-designing
dynamics models, cost functions, feedback gains, etc. and
require expert domain knowledge. Additionally, scaling these
methods up to more complex problems continues to be
an issue. Evolutionary algorithms have also been tried as
a solution to multi-agent problems; usually with smaller,
simpler environments, and policies with low complexity [12],
[13]. Recently, a hybrid approach combining MPC and the
use of genetic algorithms to evolve the cost function for a
hand-tuned MPC controller has been demonstrated for a UAV
swarm combat scenario [14].
In this work we demonstrate the effectiveness of our
approach on two complex multi-agent UAV swarm combat
scenarios: where a team of fixed wing aircraft must attack
a well-defended base, and where two teams of agents go
head to head to defeat each other. Such scenarios have been
previously considered in simulated environments with less
fidelity and complexity [15], [14]. We leverage the compu-
tational efficiency and flexibility of the recently developed
SCRIMMAGE multi-agent simulator for our experiments
(Figure 1) [16]. We compare the performance of ES against
the Cross Entropy Method. We also show for the competitive
scenario how the policy learns over time to coordinate a
strategy in response to an enemy learning to do the same.
We make our code freely available for use (https://
github.com/ddfan/swarm_evolve).
II. PROBLEM FORMULATION
We can pose our problem as the non-differentiable, non-
convex optimization problem
θ∗ = argmax
θ∈Θ
J(θ )
(1)
where Θ ⊂ Rn, a nonempty compact set,
is the space
of solutions, and J(θ ) is a non-differentiable, non-convex
real-valued objective function J : Θ → R. θ could be any
combination of decision variables of our problem, including
neural network weights, PID gains, hardware design param-
eters, etc. which affect the outcome of the returns J. For
reinforcement learning problems θ usually represents the
parameters of the policy and J is an implicit function of the
sequential application of the policy to the environment. We
first review how this problem can be solved using Gradient-
Based Adaptive Stochastic Search methods and then show
how the ES algorithm is a special case of these methods.
A. Gradient-Based Adaptive Stochastic Search
The goal of model-based stochastic search methods is
to cast the non-differentiable optimization problem (1) as
a differentiable one by specifying a probabilistic model
(hence "model-based") from which to sample [8]. Let this
model be p(θω) = f (θ;ω),ω ∈ Ω, where ω is a parameter
which defines the probability distribution (e.g. for Gaussian
distributions, the distribution is fully parameterized by the
mean and variance ω = [µ,σ ]). Then the expectation of J(θ )
over the distribution f (θ;ω) will always be less than the
optimal value of J, i.e.
J(θ ) f (θ;ω)dθ ≤ J(θ∗)
(2)
(cid:90)
Θ
The idea of Gradient-based Adaptive Stochastic Search
(GASS) is that one can perform a search in the space
of parameters of the distribution Ω rather than Θ, for a
distribution which maximizes the expectation in (2):
(cid:90)
ω∗ = argmax
ω∈Ω
Θ
J(θ ) f (θ;ω)dθ
(3)
Maximizing this expectation corresponds to finding a distri-
bution which is maximally distributed around the optimal θ.
However, unlike maximizing (1), this objective function can
now be made continuous and differentiable with respect to
ω. With some assumptions on the form of the distribution,
the gradient with respect to ω can be pushed inside the
expectation.
The GASS algorithm presented by [8] is applicable to the
exponential family of probability densities:
where φ (θ ) = ln(cid:82) exp(ω(cid:124)T (θ ))dθ, and T (θ ) is the vector
f (θ;ω) = exp{ω(cid:124)T (θ )− φ (θ )}
(4)
of sufficient statistics. Since we are concerned with showing
the connection with ES which uses parameter perturbations
f (θ;ω) is
sampled with Gaussian noise, we assume that
Gaussian. Furthermore, since we are concerned with learning
a large number of parameters (i.e. weights in a neural
network), we assume an independent Gaussian distribution
(cid:124) ∈ R2n and ω =
over each parameter. Then, T (θ ) = [θ ,θ 2]
[µ/σ 2,−1/nσ 2]
(cid:124) ∈ R2n, where µ and σ are vectors of the
mean and standard deviation corresponding to the distribu-
tion of each parameter, respectively.
Algorithm 1 Gradient-Based Adaptive Stochastic Search
Require: Learning rate αk, sample size Nk, initial policy
(cid:124), smoothing function S(), small
parameters ω0 = [µ0,σ 2
0 ]
constant γ > 0.
1: for k = 0,1,··· do
2:
3:
4:
5:
6:
iid∼ f (θ;ωk),i = 1,2,··· ,Nk.
Sample θ i
k
k)) for i = 1,··· ,Nk.
k = S(J(θ i
Compute returns wi
Compute variance terms Vk = Vk + γI, eq (5),(6)
Calculate normalizer η = ∑Nk
i=1 wi
k.
Update ωk+1:
ωk+1 ← ωk + αk
(cid:18)(cid:20) θ i
(cid:21)
(cid:20)
−
µ
1
η V−1
k
Nk∑
i=1
wi
k
7:
8: end for
k
(θ i
k)2
σ 2 + µ2
(cid:21)(cid:19)
We present the GASS algorithm for this specific set of
probability models (Algorithm 1), although the analysis for
convergence holds for the more general exponential family
of distributions. For each iteration k, The GASS algorithm
iid∼ f (θ;ωk),i =
involves drawing Nk samples of parameters θ i
k
1,2,··· ,Nk. These parameters are then used to sample the
return function J(θ i
k). The returns are fed through a shaping
function S(·) : R → R+ and then used to calculate an update
on the model parameters ωk+1.
The shaping function S(·) is required to be nondecreasing
and bounded from above and below for bounded inputs,
with the lower bound away from 0. Additionally, the set
{argmaxθ∈Θ S(J(θ ))} must be a nonempty subset of the
set of solutions of the original problem {argmaxθ∈Θ J(θ )}.
The shaping function can be used to adjust
the explo-
ration/exploitation trade-off or help deal with outliers when
sampling. The original analysis of GASS assumes a more
general form of Sk(·) where S can change at each iteration.
For simplicity we assume here it is deterministic and un-
changing per iteration.
GASS can be considered a second-order gradient method
and requires estimating the variance of the sampled param-
eters:
Vk =
1
Nk − 1
T (θ i
(cid:124)
k)T (θ i
k)
Nk∑
i=1
−
(cid:32) Nk∑
i=1
(cid:33)(cid:32) Nk∑
i=1
T (θ i
k)
(cid:33)(cid:124)
T (θ i
k)
.
(5)
1
k − Nk
N2
In practice if the size of the parameter space Θ is large, as is
the case in neural networks, this variance matrix will be of
size 2n× 2n and will be costly to compute. In our work we
approximate Vk with independent calculations of the variance
on the parameters of each independent Gaussian. With a
slight abuse of notation, consider θ i
k as a scalar element of
k a 2×2 variance
k. We then have, for each scalar element θ i
θ i
matrix:
Vk =
(cid:21)(cid:2) θ i
(cid:20) θ i
(cid:32) Nk∑
(cid:20) θ i
k
( θ i
k)2
k
i=1
k
( θ i
k)2
( θ i
k)2(cid:3)
(cid:21)(cid:33)(cid:32) Nk∑
i=1
1
Nk∑
Nk − 1
−
i=1
1
k − Nk
N2
k)2(cid:3)(cid:33)
.
( θ i
(cid:2) θ i
k
(6)
Theorem 1 shows that GASS produces a sequence of
ωk that converges to a limit set which specifies a set of
distributions that maximize (3). Distributions in this set will
specify how to choose θ∗ to ultimately maximize (1). As
with most non-convex optimization algorithms, we are not
guaranteed to arrive at the global maximum, but using prob-
abilistic models and careful choice of the shaping function
should help avoid early convergence into suboptimal local
maximum. The proof relies on casting the update rule in the
form of a generalized Robbins-Monro algorithm (see [8],
Thms 1 and 2). Theorem 1 also specifies convergence rates
in terms of the number of iterations k, the number of samples
per iteration Nk, and the learning rate αk. In practice Theorem
1 implies the need to carefully balance the increase in the
number of samples per iteration and the decrease in learning
rate as iterations progress.
Assumption 1
i) The learning rate αk > 0, αk → 0 as k → ∞, and
∑∞
k=0 αk = ∞.
ii) The sample size Nk = N0kξ , where ξ > 0; also αk and
√
Nk = O(k−β ).
Nk jointly satisfy α/
(cid:82)
iii) T (θ ) is bounded on Θ.
iv) If ω∗ is a local maximum of (3),
the Hessian of
Θ J(θ ) f (θ;ω)dθ is continuous and symmetric negative
definite in a neighborhood of ω∗.
Theorem 1
Assume that Assumption 1 holds. Let αk = α0/kα for 0 <
α < 1. Let Nk = N0kτ−α where τ > 2α is a constant. Then
the sequence {ωk} generated by Algorithm 1 converges to a
√
kτ ).
limit set w.p.1. with rate O(1/
B. Evolutionary Strategies
We now review the ES algorithm proposed by [4] and
show how it is a first-order approximation of the GASS
algorithm. The ES algorithm consists of the same two
phases as GASS: 1) Randomly perturb parameters with noise
sampled from a Gaussian distribution. 2) Calculate returns
and calculate an update to the parameters. The algorithm is
outlined in Algorithm 2. Once returns are calculated, they are
sent through a function S(·) which performs fitness shaping
[17]. Salimans et al. used a rank transformation function for
S(·) which they argue reduced the influence of outliers at
each iteration and helped to avoid local optima. It is clear that
policy parameters θ0, smoothing function S().
Algorithm 2 Evolution Strategies
Require: Learning rate αk, noise standard deviation γ, initial
1: for k = 0,1,··· do
2:
3:
Sample ε1,··· ,εn ∼ N (0,In×n)
Compute returns wi
k = S(J(θk +γεi)) for i = 1,··· ,Nk
Update θk+1:
θk+1 ← θk + αk
4:
5:
6: end for
1
Nkγ ∑Nk
i=1 wi
kεi
the ES algorithm is a sub-case of the GASS algorithm when
the sampling distribution is a point distribution. We can also
recover the ES algorithm by ignoring the variance terms on
line 7 in Algorithm 1. Instead of the normalizing term η, ES
uses the number of samples Nk. The small constant in GASS
γ becomes the variance term in the ES algorithm. The update
rule in Algorithm 2 involves multiplying the scaled returns by
k−µ in Algorithm 1. We see that
the noise, which is exactly θ i
ES enjoys the same asymptotic convergence rates offered by
the analysis of GASS. While GASS is a second-order method
and ES is only a first-order method, in practice ES uses
approximate second-order gradient descent methods which
adapt the learning rate in order to speed up and stabilize
learning. Examples of these methods include ADAM, RM-
SProp, SGD with momentum, etc., which have been shown
to perform very well for neural networks. Therefore we can
treat ES a first-order approximation of the full second-order
variance updates which GASS uses. In our experiments we
use ADAM [18] to adapt the learning rate for each parameter.
As similarly reported in [4], when using adaptive learning
rates we found little improvement over adapting the variance
of the sampling distribution. We hypothesize that a first order
method with adaptive learning rates is sufficient for achieving
good performance when optimizing neural networks. For
other types of policy parameterizations however, the full
second-order treatment of GASS may be more useful. It is
also possible to mix and match which parameters require a
Fig. 2: Diagram of each agent's policy. Nearby ally states and sensed enemies, base locations, etc. along with the agent's
own state are fed into a neural network which produces a reference target in relative xyz coordinates. The target is fed into
the safety logic block which checks for collisions with neighbors or the ground. It produces a reference target which is fed
to the PID controller, which in turn provides low-level controls for the agent (thrust, aileron, elevator, rudder).
full variance update and which can be updated with a first-
order approximate method. We use the rank transformation
function for S(·) and keep Nk constant.
C. Learning Structured Policies for Multi-Agent Problems
Now that we are more confident about the convergence
of the ES/GASS method, we show how ES can be used to
optimize a complex policy in a large-scale multi-agent envi-
ronment. We use the SCRIMMAGE multi-agent simulation
environment [16] as it allows us to quickly and in parallel
simulate complex multi-agent scenarios. We populate our
simulation with 6DoF fixed-wing aircraft and quadcopters
with dynamics models having 10 and 12 states, respectively.
These dynamcis models allow for full ranges of motion
within realistic operating regimes. Stochastic disturbances
in the form of wind and control noise are modeled as
additive Gaussian noise. Ground and mid-air collisions can
occur which result in the aircraft being destroyed. We also
incorporate a weapons module which allows for targeting
and firing at an enemy within a fixed cone projecting from
the aircraft's nose. The probability of a hit depends on the
distance to the target and the total area presented by the target
to the attacker. This area is based on the wireframe model
of the aircraft and its relative pose. For more details, see our
code and the SCRIMMAGE simulator documentation.
We consider the case where each agent uses its own policy
to compute its own controls, but where the parameters of the
policies are the same for all agents. This allows each agent
to control itself in a decentralized manner, while allowing
for beneficial group behaviors to emerge. Furthermore, we
assume that friendly agents can communicate to share states
with each other (see Figure 2). Because we have a large
number of agents (up to 50 per team), to keep communication
costs lower we only allow agents to share information locally,
i.e. agents close to each other have access to each other's
states. In our experiments we allow each agent to sense the
states of the closest 5 friendly agents for a total of 5∗10 = 50
incoming state messages.
Additionally, each agent is equipped with sensors to detect
enemy agents. Full state observability is not available here,
instead we assume that sensors are capable of sensing an
enemy's relative position and velocity. In our experiments
we assumed that each agent is able to sense the nearest
5 enemies for a total of 5 ∗ 7 = 35 dimensions of enemy
data (7 states = [relative xyz position, distance, and relative
xyz velocities]). The sensors also provide information about
home and enemy base relative headings and distances (an
additional 8 states). With the addition of the agent's own state
(9 states), the policy's observation input (cid:126)o(t) has a dimension
of 102. These input states are fed into the agent's policy: a
neural network f ((cid:126)o(t);θ ) with 3 fully connected layers with
sizes 200, 200, and 50, which outputs 3 numbers representing
a desired relative heading [xre f ,yre f ,zre f ]. Each agent's neural
network has more than 70,000 parameters. Each agent uses
the same neural network parameters as its teammates, but
since each agent encounters a different observation at each
timestep, the output of each agent's neural network policy
will be unique. It may also be possible to learn unique
policies for each agent; we leave this for future work.
With safety being a large concern in UAV flight, we
design the policy to take into account safety and control
considerations. The relative heading output from the neural
network policy is intended to be used by a PID controller
to track the heading. The PID controller provides low-
level control commands u(t) to the aircraft (thrust, aileron,
elevator, rudder). However, to prevent cases where the neural
network policy guides the aircraft into crashing into the
ground or allies, etc., we override the neural network heading
with an avoidance heading if the aircraft is about to collide
with something. This helps to focus the learning process
on how to intelligently interact with the environment and
allies rather than learning how to avoid obvious mistakes.
Furthermore, by designing the policy in a structured and
interpretable way, it will be easier to take the learned policy
-+PID𝑥𝑟𝑒𝑓𝑦𝑟𝑒𝑓𝑧𝑟𝑒𝑓𝑢(𝑡)𝑓(Ԧ𝑜(𝑡);𝜃)Safety LogicԦ𝑜(𝑡)Ally statesSensingNeural NetworkԦ𝑥(𝑡) directly from simulation into the real world. Since the
neural network component of the policy does not produce
low-level commands, it is invariant to different low-level
controllers, dynamics, PID gains, etc. This aids in learning
more transferrable policies for real-world applications.
III. EXPERIMENTS
We consider two scenarios: a base attack scenario where
a team of 50 fixed wing aircraft must attack an enemy base
defended by 20 quadcopters, and a team competitive task
where two teams concurrently learn to defeat each other. In
both tasks we use the following reward function:
J = 10× (#kills) + 50× (#collisions with enemy base)
− 1e− 5× (distance from enemy base at end of episode)
(7)
The reward function encourages air-to-air combat, as well
as suicide attacks against the enemy base (e.g. a swarm of
cheap, disposable drones carrying payloads). The last term
encourages the aircraft to move towards the enemy during
the initial phases of learning.
A. Base Attack Task
Fig. 3: Snapshot of base attack task. The goal of the blue
fixed wing team (lower left) is to attack the red base (red
dot, upper right) while avoiding or attacking red quadcopter
guards.
In this scenario a team of 50 fixed-wing aircraft must
attack an enemy base defended by 20 quadcopters (Figure
3). The quadcopters use a hand-crafted policy where in the
absence of an enemy, they spread themselves out evenly to
cover the base. In the presence of an enemy they target
the closest enemy, match that enemy's altitude, and fire
repeatedly. We used Nk = 300,γ = 0.02, a time step of 0.1
seconds, and total episode length of 200 seconds. Initial
positions of both teams were randomized in a fixed area at
opposide ends of the arena. Training took two days with full
parallelization on a machine equipped with a Xeon Phi CPU
(244 threads).
We found that over the course of training the fixed-
wing team learned a policy where they quickly form a V-
formation and approach the base. Some aircraft suicide-
attack the enemy base while others begin dog-fighting (see
(a) Training
(b) Testing
Fig. 4: Scores per iteration for the base attack task. Top:
Scores earned by perturbed policies during training. Scores
are on average lower because they result from policies which
are parameterized by randomly peturbed values. Bottom:
Scores during the course of training earned by the updated
policy parameters. Red curve is Evolution Strategies algo-
rithm, blue is Cross Entropy Method. Bold line is the median,
shaded areas are 25/75 quartile bounds.
supplementary video1). We also compared our implementa-
tion of the ES method against the well-known cross-entropy
method (CEM). CEM performs significantly worse than ES
(Figure 4). We hypothesize this is because CEM throws out
a significant fraction of sampled parameters and therefore
obtains a worse estimate of the gradient of (3). Comparison
against other full second-order methods such as CMA-ES or
the full second-order GASS algorithm is unrealistic due to
the large number of parameters in the neural network and
the prohibitive computational difficulties with computing the
covariances of those parameters.
B. Two Team Competitive Match
The second scenario we consider is where two teams each
equipped with their own unique policies for their agents
learn concurrently to defeat their opponent (Figure 5). At
each iteration, Nk = 300 simulations are spawned, each
with a different random perturbation, and with each team
having a different perturbation. The updates for each policy
are calculated based on the scores received from playing
the opponent's perturbed policies. The result is that each
1http://https://goo.gl/dWvQi7
05001000150020002500Iteration025050075010001250ScoreESCEM05001000150020002500Iteration025050075010001250ScoreESCEM Fig. 5: Snapshot of two team competitive match. The goal
of both teams is to defeat all enemy planes while suffering
minimum losses, or to attack the opponent's base. Red lines
indicate successful firing hit.
team learns to defeat a wide range of opponent behaviors
at each iteration. We observed that
the behavior of the
two teams quickly approached a Nash equilibrium where
both sides try to defeat the maximum number of opponent
aircraft in order to prevent higher-scoring suicide attacks (see
supplementary video). The end result is a stalemate with
both teams annihilating each other, ending with tied scores
(Figure 6). We hypothesize that more varied behavior could
be learned by having each team compete against some past
enemy team behaviors or by building a library of policies
from which to select from, as frequently discussed by the
evolutionary computation community [19].
IV. CONCLUSION
We have shown that Evolution Strategies are applicable
for learning policies with many thousands of parameters
for a wide range of complex tasks in both the competi-
tive and cooperative multi-agent setting. By showing the
connection between ES and more well-understood model-
based stochastic search methods, we are able to gain in-
sight into future algorithm design. Future work will include
experiments with optimizing mixed parameterizations, e.g.
optimizing both neural network weights and PID gains. In
this case, the second-order treatment on non-neural network
parameters may be more beneficial, since the behavior of the
system may be more sensitive to perturbations of non-neural
network parameters. Another direction of investigation could
be optimizing unique policies for each agent in the team.
Yet another direction would be comparing other evolutionary
computation strategies for training neural networks, includ-
ing methods which use a more diverse population [20], or
more genetic algorithm-type heuristics [21].
REFERENCES
[1] Y. Li, "Deep Reinforcement Learning: An Overview," ArXiv e-prints,
[2] K. Stanley and B. Bryant, "Real-time neuroevolution in the NERO
[Online]. Available:
video game," IEEE transactions on, 2005.
http://ieeexplore.ieee.org/xpls/abs{ }all.jsp?arnumber=1545941
[3] O. J. Coleman, "Evolving Neural Networks for Visual Processing,"
Jan. 2017.
Thesis, 2010.
(a) Training
(b) Testing
Fig. 6: Scores per iteration for the team competition task.
Top: Scores earned by each team with policies parameterized
by randomly peturbed values during training. Bottom: Scores
earned when testing the trained architectures using updated
policy parameters. Red and blue curves show scores for team
1 and 2 respectively.
[4] T. Salimans, J. Ho, X. Chen, S. Sidor, and I. Sutskever, "Evolution
Strategies as a Scalable Alternative to Reinforcement Learning," ArXiv
e-prints, Mar. 2017.
[5] J. Hu, "Model-based stochastic search methods," in Handbook of
Simulation Optimization. Springer, 2015, pp. 319 -- 340.
[6] S. Mannor, R. Rubinstein, and Y. Gat, "The cross entropy method for
fast policy search," in Machine Learning-International Workshop Then
Conference-, vol. 20, no. 2, 2003, Conference Proceedings, p. 512.
[7] N. Hansen, "The CMA evolution strategy: A tutorial," CoRR, vol.
abs/1604.00772, 2016. [Online]. Available: http://arxiv.org/abs/1604.
00772
[8] E. Zhou and J. Hu, "Gradient-based adaptive stochastic search for non-
differentiable optimization," IEEE Transactions on Automatic Control,
vol. 59, no. 7, pp. 1818 -- 1832, 2014.
[9] L. Panait and S. Luke, "Cooperative multi-agent
learning: The
the art," Autonomous Agents and Multi-Agent Systems,
[Online]. Available: http:
state of
vol. 11, no. 3, pp. 387 -- 434, 2005.
//link.springer.com/10.1007/s10458-005-2631-2
[10] J. B. Rawlings and B. T. Stewart, "Coordinating multiple optimization-
based controllers: New opportunities and challenges," Journal of
Process Control, vol. 18, no. 9, pp. 839 -- 845, 2008.
[11] W. Al-Gherwi, H. Budman, and A. Elkamel, "Robust distributed model
predictive control: A review and recent developments," The Canadian
Journal of Chemical Engineering, vol. 89, no. 5, pp. 1176 -- 1190,
2011. [Online]. Available: http://doi.wiley.com/10.1002/cjce.20555
[12] G. B. Lamont, J. N. Slear, and K. Melendez, "UAV swarm mission
planning and routing using multi-objective evolutionary algorithms," in
IEEE Symposium Computational Intelligence in Multicriteria Decision
Making, no. Mcdm, 2007, Conference Proceedings, pp. 10 -- 20.
[13] A. R. Yu, B. B. Thompson, and R. J. Marks, "Competitive evolution of
tactical multiswarm dynamics," IEEE Transactions on Systems, Man,
0100200300400500600700Iteration0100200300400500ScoreTeam 1Team 20100200300400500600700Iteration0100200300400500ScoreTeam 1Team 2 and Cybernetics Part A:Systems and Humans, vol. 43, no. 3, pp. 563 --
569, 2013.
[14] D. D. Fan, E. Theodorou, and J. Reeder, "Evolving cost functions
for model predictive control of multi-agent uav combat swarms," in
Proceedings of the Genetic and Evolutionary Computation Conference
Companion, ser. GECCO '17. New York, NY, USA: ACM, 2017,
pp. 55 -- 56. [Online]. Available: http://doi.acm.org/10.1145/3067695.
3076019
[15] U. Gaerther, "UAV swarm tactics: an agent-based simulation
[Online]. Available: https:
and Markov process analysis," 2015.
//calhoun.nps.edu/handle/10945/34665
[16] K. J. DeMarco. (2018) SCRIMMAGE multi-agent robotics simulator.
[Online]. Available: http://www.scrimmagesim.org/
[17] D. Wierstra, T. Schaul, T. Glasmachers, Y. Sun, J. Peters, and
J. Schmidhuber, "Natural evolution strategies." Journal of Machine
Learning Research, vol. 15, no. 1, pp. 949 -- 980, 2014.
[18] D. P. Kingma and J. Ba, "Adam: A method for stochastic optimiza-
tion," arXiv preprint arXiv:1412.6980, 2014.
[19] K. O. Stanley and R. Miikkulainen, "Competitive coevolution through
evolutionary complexification," Journal of Artificial Intelligence Re-
search, vol. 21, pp. 63 -- 100, 2004.
[20] E. Conti, V. Madhavan, F. Petroski Such, J. Lehman, K. O. Stanley,
and J. Clune, "Improving Exploration in Evolution Strategies for Deep
Reinforcement Learning via a Population of Novelty-Seeking Agents,"
ArXiv e-prints, Dec. 2017.
[21] F. Petroski Such, V. Madhavan, E. Conti, J. Lehman, K. O. Stanley, and
J. Clune, "Deep Neuroevolution: Genetic Algorithms Are a Competi-
tive Alternative for Training Deep Neural Networks for Reinforcement
Learning," ArXiv e-prints, Dec. 2017.
|
cs/0607133 | 1 | 0607 | 2006-07-27T17:55:16 | Self-Replication and Self-Assembly for Manufacturing | [
"cs.MA",
"cs.CE"
] | It has been argued that a central objective of nanotechnology is to make products inexpensively, and that self-replication is an effective approach to very low-cost manufacturing. The research presented here is intended to be a step towards this vision. We describe a computational simulation of nanoscale machines floating in a virtual liquid. The machines can bond together to form strands (chains) that self-replicate and self-assemble into user-specified meshes. There are four types of machines and the sequence of machine types in a strand determines the shape of the mesh they will build. A strand may be in an unfolded state, in which the bonds are straight, or in a folded state, in which the bond angles depend on the types of machines. By choosing the sequence of machine types in a strand, the user can specify a variety of polygonal shapes. A simulation typically begins with an initial unfolded seed strand in a soup of unbonded machines. The seed strand replicates by bonding with free machines in the soup. The child strands fold into the encoded polygonal shape, and then the polygons drift together and bond to form a mesh. We demonstrate that a variety of polygonal meshes can be manufactured in the simulation, by simply changing the sequence of machine types in the seed. | cs.MA | cs | Self-Replication and Self-Assembly for Manufacturing
Robert Ewaschuk* and Peter D. Turney† (corresponding author)
Submitted to: Artificial Life; Category: Article; Date: April 26, 2005
Abstract
It has been argued that a central objective of nanotechnology is to make products
inexpensively, and that self-replication is an effective approach to very low-cost
manufacturing. The research presented here is intended to be a step towards this vision.
We describe a computational simulation of nanoscale machines floating in a virtual
liquid. The machines can bond together to form strands (chains) that self-replicate and
self-assemble into user-specified meshes. There are four types of machines and the
sequence of machine types in a strand determines the shape of the mesh they will build.
A strand may be in an unfolded state, in which the bonds are straight, or in a folded
state, in which the bond angles depend on the types of machines. By choosing the
sequence of machine types in a strand, the user can specify a variety of polygonal
shapes. A simulation typically begins with an initial unfolded seed strand in a soup of
unbonded machines. The seed strand replicates by bonding with free machines in the
soup. The child strands fold into the encoded polygonal shape, and then the polygons
drift together and bond to form a mesh. We demonstrate that a variety of polygonal
meshes can be manufactured in the simulation, by simply changing the sequence of
machine types in the seed.
Keywords: self-replication, self-assembly, nanotechnology, virtual physics, continuous
space automata, manufacturing, polygonal meshes.
* Institute for Information Technology, National Research Council of Canada, Ottawa, Ontario,
Canada, K1A 0R6, [email protected]
† Institute for Information Technology, National Research Council of Canada, Ottawa, Ontario,
Canada, K1A 0R6, [email protected], (613) 993-8564 (corresponding author)
Self-Replication and Self-Assembly for Manufacturing
1 Introduction
Researchers have argued that one of the main objectives of nanotechnology is to
manufacture products inexpensively, and that this goal can be effectively achieved by
self-replication [2], [8], [9]. We believe that it is useful to develop computational
simulations of self-replicating nanotechnology, as engineering tools to assist in the
design of actual self-replicating machines.
In our earlier work with JohnnyVon 1.0, we developed a computational simulation of
machines that join to form self-replicating strands (chains of machines linked by flexible
bonds) [17]. These machines drifted about in a virtual liquid, simulated as a two-
dimensional continuous space with Brownian motion and viscosity. There were two
types of machines, which enabled a strand to encode an arbitrary bit string, by
designating one type of machine as representing 0 and the other as 1. Although strand
replication faithfully preserved the encoded bit strings, the information in the strings
played no functional role in JohnnyVon 1.0. In effect, the simulation had genotypes
(genetic code) without phenotypes (bodies). From
the perspective of potential
applications in manufacturing, the absence of phenotypes was a major limitation of
JohnnyVon 1.0.
This paper introduces JohnnyVon 2.0, which builds on its predecessor by adding
phenotypes to the simulation. The design of JohnnyVon 2.0 was partly inspired by the
work of Seeman on building nanometer-scale structures with DNA [14], [15]. In living
organisms, replication is based on DNA (the genotype) and the information encoded in
DNA is used to build proteins (the major structural material of the phenotype). Seeman
has shown that DNA can serve both as a device for self-replication (genotype) and
(surprisingly) as a building material for nanoscopic structures and tools (phenotype). By
choosing the appropriate sequence of codons, DNA can be programmed to self-
assemble into a wide variety of structures, such as cubes, octahedra, one-dimensional
strands, two-dimensional meshes, and three-dimensional arrays. Seeman discusses a
variety of potential nanotechnological applications for these structures. For example, a
three-dimensional DNA array could facilitate x-ray crystallography, by serving as a
scaffolding for holding molecular samples in a regular lattice [15].
In JohnnyVon 2.0, there are four types of machines, drifting in a two-dimensional
continuous space with Brownian motion and viscosity (i.e., a simulated liquid). The
2
Self-Replication and Self-Assembly for Manufacturing
machines exert spring-like attractive and repulsive forces on each other, but internally
they are finite state machines. The input to each state machine is based on the presence
or absence of bonds with neighbouring machines and on the internal states of bonded
neighbours. The internal states govern when bonds are formed or broken and the angles
at which bonded machines are joined, and thus determine whether a strand will form a
straight line or fold into a specific polygonal shape.
Following the hint of Seeman’s work, a strand in John nyVon 2.0 serves as both a
genotype and a phenotype, at different stages in its career [14], [15]. Like living
organisms (but unlike von Neumann’s strategy [21]), John nyVon 2.0 takes a template-
based approach to self-replication. A strand begins its career as a genotype. While
acting as a genotype, the strand is approximately straight, so that it can provide a good
template for replication. Brownian motion and interactions with other machines will cause
the strand to bend slightly, because the bonds between the machines are flexible, but
the system is designed so that forces will tend to straighten the strand. Later in its
career, the strand may become a phenotype. When this happens, the bonding forces
change, causing the bonding angles to alter, and the strand folds. This folding is
(approximately) analogous to the way that proteins fold. A folded strand acts as a
structural element and is no longer capable of replication.
A typical run of a JohnnyVon 2.0 simulation begins with a soup of unbonded machines
and an initial unfolded seed strand of bonded machines. Free (unbonded) machines
connect to the seed strand, eventually forming a double strand (two parallel strands).
When the new strand is complete, the two strands break apart, and thus we have self-
replication. A strand will continue to self-replicate until unbonded machines become
scarce. When a strand has not encountered an unbonded machine for a relatively long
period of time, the strand stops replicating and folds. The shape that it folds into
depends on the types of machines in the strand and their sequential ordering. Folded
strands drift in the virtual liquid and bond with each other, forming a mesh. The user can
specify the shape of the holes in the final mesh by selecting the sequence of machine
types that compose the initial seed strand.
In Section 2, we discuss related work with von Neumann’ s universal constructor, self-
replicating loops, and artificial chemistry. The details of JohnnyVon 2.0’s design are
explained in Section 3, including the changes that have been made from JohnnyVon 1.0
[17]. We present our experiments in Section 4. Each experiment is a run of the
3
Self-Replication and Self-Assembly for Manufacturing
simulation with an initial seed strand. We demonstrate that a variety of polygonal
meshes can be manufactured by varying the initial seed. Section 5 examines limitations
of JohnnyVon 2.0 and discusses problems and projects for future work. Potential
applications are suggested in Section 6 and we conclude in Section 7.
2 Related Work
Sipper provides a good survey of research on self-replication [16]. Some of the research
involves actual mechanical devices and some is based on organic chemistry, but we
restrict our discussion here to computer simulations of self-replication. We briefly review
von Neumann’s universal constructor [21], self-replicatin g loops [6], [10], [11], [12], [13],
[18], [19], and artificial chemistry [4], [5].
2.1
Universal Constructor
Von Neumann’s approach to self-replication was to design a universal constructor, which
could build anything, and therefore could build itself as a special case [21]. He described
five different models (i.e., five different kinds of simulations), with varying levels of
realism and concreteness. The design of the universal constructor was only worked out
in detail for the cellular automata model, which was the most abstract of the five models.
In von Neumann’s cellular automata model, the universa l constructor was composed of
a group of several thousand cells that begin in a specific configuration of initial states.
Another line of cells acts as a kind of tape, which is read by the universal constructor.
For any given finite configuration of cell states, there is a tape that can cause the
universal constructor to build the given configuration. As a special case, there is a tape
that can cause the universal constructor to build a copy of itself, thereby self-replicating.
2.2
Self-Replicating Loops
Langton demonstrated self-replication in a cellular automata model that was much
simpler than von Neumann’s model [6]. He achieved this simplification by designing a
constructor that could construct only itself, instead of trying to make a universal
constructor. His cellular automata model had eight states instead of twenty-nine and his
constructor was composed of a group of about a hundred cells in a specific initial
configuration, instead of a group of several thousand cells.
4
Self-Replication and Self-Assembly for Manufacturing
In Langton’s model, the cells of the constructor are arra nged in a loop. The states of the
cells in the loop go through a cycle, periodically creating a copy of the original loop.
Starting from the initial loop, increasing numbers of copies spread across the grid.
The idea of self-replicating loops in cellular automata models has since been developed
further by many researchers [11], [12], [13], [19]. Self-replicating loops have exhibited
many interesting behaviours, including evolution [11], [13] and self-repair [19]. Tempesti
has described a self-replicating cellular automata model that can perform computations
and build structures [18]. Morita and Imai have shown how a self-replicating cellular
automata model can encode a variety of shapes of loops, beyond the usual square loops
[10].
2.3
Artificial Chemistry
Hutton introduced self-replication in an artificial chemistry simulation, using a template-
based approach [4]. A chain of molecules forms a template against which other
molecules bond, similar in concept to JohnnyVon 1.0 [17]. A run of the simulation begins
with a seed chain in a soup of free molecules. By a series of chemical reactions, a
parallel chain of molecules forms next to the seed chain. When the parallel chain is
complete, it separates from the seed chain and the process repeats.
Hutton’s first approach was a cellular automata model [ 4], but the discrete space
constrained the mobility of the simulated molecules, hence Hutton’s second approach
used a continuous space [5], like JohnnyVon 1.0 [17]. In Hutton’s second model,
molecules move in a continuous two-dimensional space, following linear trajectories until
an obstacle (e.g., the container wall or another molecule) is encountered (i.e., the motion
is a billiard ball model). When molecules make contact with each other, they undergo a
chemical reaction that bonds them together, according to the rules of the artificial
chemistry.
In Hutton’s first model [4], the molecules only replica te, but in his second model [5], they
also build a circular barrier, suggestive of a cell wall. Each time a chain replicates, the
new chain builds a wall around itself.
The basic objects in Hutton’s system ( “molecules”) are simpl er than the basic objects in
JohnnyVon ( “machines”), which makes Hutton’s system more com putationally efficient.
However, JohnnyVon has a richer virtual physics, including Brownian motion, attractive
and repulsive fields, momentum, and viscosity. Hutton’s m olecules are points, with no
5
Self-Replication and Self-Assembly for Manufacturing
directional orientation, whereas JohnnyVon ’s machines are shapes with angular
orientation, which can rotate, experience twisting forces, and have angular momentum.
The richer virtual physics in JohnnyVon may be useful for simulations of manufacturing.
3 JohnnyVon 2.0
We first give an overview of JohnnyVon 2.0 and then describe the system in detail. We
encourage the reader to begin by viewing Figure 5 in Section 4.1. This figure should
make it easier to understand the following discussion.
3.1
Overview
The basic objects in the JohnnyVon simulation are called machines. There are four
types of machines, numbered 1 through 4. All four types are shaped like a plus sign ‘+ ’
and appear visually identical in the simulation (see Figure 5). Each machine has five
arms, but two of the arms overlap, so the figures seem to show four arms. In the figures,
the arms are represented by black lines.
The machines are mobile and can rotate at any angle, but it is convenient to describe
them when they are rotated into a standard reference position, which we call the
canonical position. In the canonical position, the shortest arm points down, the two
longest arms point right and left, and the medium-length arm points up. Another short
arm points up in canonical position, but it is hidden by the medium-length arm that points
up. (The three machines labeled B, D, and F in Figure 1 are in canonical position. The
fourth machine is upside-down.)
The two longest arms, pointing left and right when the machine is in canonical position,
are called the left and right arms. When machines bond to form a strand, adjacent
machines in the strand are bonded to each other at the tips of their left and right arms.
The up arm is the longer of the two arms that point up when the machine is in canonical
position. When a strand replicates by forming a mirror strand, the machines in the mirror
strand are bonded to their neighbours in the original strand at the tips of their up arms.
Also, when the strands fold into polygons and join to form a mesh, the polygons bond to
each other at the tips of their up arms.
The shorter of the two arms that point up, when the machine is in canonical position, is
the repellor arm. This arm overlaps the up arm in the figures, so it is not visible. When a
strand has completely replicated, repulsive fields are briefly activated at the tips of the
6
Self-Replication and Self-Assembly for Manufacturing
repellor arms. This splits the original strand from the mirror strand and pushes the two
strands apart.
The short arm that points down in canonical position is the overlap detector arm. This
arm is used to detect when two folded strands (e.g., polygons) overlap in a mesh.
Machines move about in a two-dimensional continuous space, bounded by a grey box.
The centers of the machines are confined to the interior of the grey box. This space is
called the container. A virtual liquid fills the container. The trajectory of a machine is
determined by Brownian motion (random drift due to the liquid) and by interaction with
other machines and the walls of the container. The liquid has a viscosity that dampens
the momentum of the machines. When there are machines suspended in the liquid, we
call it a soup.
The arms of the machines have attractive or repulsive fields. The range of a field is
bounded by a circle. In addition to attracting or repelling, a field can also exert a bending
force, which twists the machines to form a particular angle. A field’s interaction (attract,
repel, twist, or ignore) with another field is determined by many factors, including the
type and state of each machine. The fields behave somewhat like springs. The center of
every field is always at the tip (the outer end) of an arm.
The point at which the five arms meet is called the middle of the machine. This is not the
machine’s geometrical center, but it is treated as the ce nter of mass in the simulation.
For example, a rotational force will cause a machine to rotate about its middle.
Although the space is two-dimensional, machines can slide over one another, as if they
were floating in a thin film of liquid. Machines interact with each other through their fields,
rather than by direct contact. They do not experience direct collisions, but their fields can
collide.
Machines can bond together when the field of one machine intersects the field of
another. Not all fields can bond. This is described in detail later. The machine that is
bonded to the up (left, right) arm of a given machine is called the up (left, right)
neighbour. A chain of machines joined by left arm to right arm bonds forms a strand.
During replication, the bond angles are such that the replicating strand tends to be
straight. Brownian motion and other forces perturb the strand, so it cannot be perfectly
straight, but twisting forces in the bonding fields tend to straighten the strand, so it is
7
Self-Replication and Self-Assembly for Manufacturing
rarely far from being straight. We call these approximately straight strands unfolded
strands or genes (a strand in its genotype state).
Under specific circumstances (described in Section 3.4.1), each bond in a strand will
change its angle, causing the strand to fold. Such strands are called folded strands or
phenes (a strand in its phenotype state). A phene will fold to form a closed loop. A group
of phenes can bond together, forming a mesh.
A machine with no bonds is called a free machine. A typical simulation begins with a
soup of free machines and a single seed gene. The seed gene is an unfolded strand that
initiates the process of self-replication. The first child of the seed gene forms the seed
phene, which acts as a starting point for the growth of the mesh. This ensures that one
seed phene will yield only one mesh.
A left-neighbour-to-right-neighbour bond is a sideways bond. Machines in a strand (both
phenes and genes) are joined by sideways bonds. An up-neighbour-to-up-neighbour
bond is an up bond. Phenes in a mesh are joined by up bonds. During replication, a
parent gene is joined to its partially constructed child gene by up bonds.
The JohnnyVon simulation proceeds in a sequence of discrete time steps. The initial
configuration is called step 0 or time 0. The state of a machine is the combination of
internal
information (counters, bonds, and other state variables) and external
relationships (position, rotation, and velocity) that determines the behaviour of a
machine. A counter is a special piece of information stored in each machine that
normally increments during each time step. Each machine has several counters. States
are described in more detail in Section 3.3.
Each bond has a desired angle (which changes when a strand folds). The two machines
that participate in each bond have a tolerance for the difference between the current
angle and the desired angle. Forces can push bonds out of tolerance. Bonds that are
consistently out of tolerance can break.
One of our main design objectives was to make JohnnyVon 2.0 programmable by the
user, by specifying the initial configuration of the simulation, without making any changes
to the rules of the virtual physics or the behaviours of the machines. By selecting the
types of machines in the seed gene and specifying their sequential order, the user can
program JohnnyVon to make a variety of polygonal meshes. In a manufacturing
application, we envision the user selecting the desired machines from four bins, joining
8
Self-Replication and Self-Assembly for Manufacturing
them together to make a seed gene, and then dropping the seed into a vat of free
machines. The user would be able to manufacture a variety of products without making
any modifications to the individual machines.
W ith only one type of machine, the user would be able to program the simulation only by
specifying the length of the seed gene. W ith two types of machines (as in JohnnyVon 1.0
[17]), programs can use more efficient binary coding. In principle, two types of machines
would be sufficient for programming mesh construction, but we found that four types
provide a good balance of simplicity and programmability. Four types are enough to
encode a variety of shapes (see Table 6 in Section 3.4.2 and the experiments in Section
4), yet four types are not so many that the system is unwieldy. The four types of
machines are somewhat analogous to the four amino acids in DNA (adenine, thymine,
guanine, and cytosine), which encode programs for building proteins.
3.2
Basic Modifications
This subsection describes the core changes in the simulation that were made from
JohnnyVon 1.0 to JohnnyVon 2.0. (Sections 3.3 and 3.4 discuss new features, as
opposed to modified features.)
3.2.1 Variable Field Sizes
Bonds between machines are formed by spring-like attractive fields. Part of the
mechanism that was in place to support mutation in JohnnyVon 1.0 was a variable field
size. In certain circumstances, the field would be small, permitting rare accidental bonds,
while other times it would be large, to strengthen intentional bonds. The accidental
bonds were a cause of mutation (replication errors), whereas the intentional bonds were
part of faithful replication.
In JohnnyVon 2.0, field sizes do not change. Fields attract, repel, or ignore other fields,
but they have a constant circle of influence. In situations where the original version has
small fields, the new version has inactive fields. This modification substantially reduces
the likelihood of mutations.1
1 Mutations are desirable when modeling biological life, but may be undesirable when modeling
manufacturing processes.
9
Self-Replication and Self-Assembly for Manufacturing
3.2.2 Physical Constants
The physical constants for viscosity, Brownian motion, and motion dampening were
changed
to suit
the new requirements. The values of
these constants were
experimentally tuned to achieve our design objectives while maximizing the speed
(computation efficiency) of the simulation.
3.2.3 Arms and the Machine
In JohnnyVon 1.0, each machine was shaped like a capital letter ‘T ’. Each machine had
four arms, but two of the arms overlapped (along the vertical bar of the T), so the figures
in the paper seem to show three arms [17]. Each arm had an attractive or repulsive field
with a circular shape, centered on the tips of the arms. The fields were colour coded,
and we named the arms according to the colours of their associated fields.
In JohnnyVon 2.0, each machine is shaped like a plus sign ‘+ ’ (see Figure 5 in Section
4.1). Each machine now has five arms, but two of the arms overlap, so the figures seem
to show four arms. It is no longer convenient to refer to the arms by the colours of their
fields. We now refer to the arms by their relative positions (up, left, right), when the
machine is rotated into a canonical position. The lengths of the arms have been modified
to facilitate building meshes.
3.3 Machine States
The state of a machine is represented by a vector. The vector elements that represent
internal aspects of the machine are all discrete. The vector elements that represent
external relationships between machines are mostly continuous. The discrete, internal
elements are governed by state transition rules that are applied in discrete timesteps.
The continuous, external elements are governed by the laws of the virtual physics. The
physical laws are inherently continuous, but they are necessarily approximated
discretely in any computational simulation.
Internal state information includes various flags, counters, and state variables, as
summarized in Table 1. External state information includes spatial location and
orientation, angular velocity, linear velocity, the presence or absence of bonds with other
machines, and bonding angles, as given in Table 2. Some derived variables are shown
in Table 3. The derived variables are calculated from state variables.
Insert Table 1 here.
10
Self-Replication and Self-Assembly for Manufacturing
Insert Table 2 here.
Insert Table 3 here.
Machines that are directly bonded together can sense each other ’s states. This is
analogous to cells in cellular automata, which can sense the states of their immediate
neighbours in the grid. The state transition rules and the virtual physics are local, in the
sense that there is no global control structure. No machine can directly sense the state
of another machine unless they are directly connected, although state information can
be passed neighbour-to-neighbour along a strand. No machine can directly exert a force
on another machine unless the circular boundaries of their fields overlap, although
forces can be passed neighbour-to-neighbour along a strand.
Most of the state transition rules and physical laws in JohnnyVon 2.0 are carried over
from JohnnyVon 1.0 without change. The details of JohnnyVon 1.0 are fully described
elsewhere [17]. The changes we made in JohnnyVon 2.0 were outlined above, in
Section 3.2.
3.4
New Rules
The following subsections describe the rules that are new in JohnnyVon 2.0.
3.4.1 Folding
The leftmost machine in an unfolded strand determines when the strand will fold. A
machine knows it is leftmost when it has a right-neighbour but no left-neighbour. As part
of its internal state, each machine maintains a fold-counter. After a strand has replicated
and split, the fold-counter in each machine in the newly formed strand starts counting.
When a machine gains an up-neighbour, it triggers a reset-counter signal. This signal is
passed to the left-neighbour, down the strand, until it reaches the leftmost machine.
When the leftmost machine receives the signal, it resets its fold-counter to 0. This is
illustrated in Figure 1. In this way, as long as a replicating strand continues to receive
new up-neighbours, it will not fold up. If the strand successfully replicates, the leftmost
fold-counter in each of the two strands (the child strand and the parent strand) is set to
0.
11
Self-Replication and Self-Assembly for Manufacturing
Insert Figure 1 here.
Once the fold-counter in the leftmost machine hits a fixed upper limit, that machine
triggers a fold-now signal. This signal is passed down the strand, setting the folded flag
as it goes. This causes the strand to fold up according to the types of each of the two
machines involved in a sideways bond.
When a strand folds, there are typically some up-neighbours attached to the folding
strand, as the folding strand has usually partially replicated itself. Machines with an up-
neighbour, but having a false (0) replicated flag, monitor their up-neighbour ’s folded flag.
If this flag becomes true (1), then such a machine will set its own shatter flag to true, and
thus release all its bonds. (Shattering is described in detail in 3.4.7.)
Machines with the seed-gene flag set to 1 never fold, thus the initial seed strand is
always available, to continue replicating as soon as there is a supply of free machines of
the right types. This is a safeguard against situations in which a temporary scarcity of
free machines persists for longer than the fixed upper limit on fold-counter.
3.4.2 Angles and Bonds
In contrast to JohnnyVon 1.0, we now have four types of machines instead of two types,
and the type of a machine affects its behaviour in both genes and phenes. In genes, the
types govern bonding during replication, where the rule is simply, “Likes attract, others
are ignored. ” More formally, the bonding rule for u p bonds in genes is given in Table 4.
Insert Table 4 here.
This rule implies that each replicated strand is a mirror of its parent, rather than an exact
copy. For example, a strand of types 1-2-3, reading left to right, in canonical position, will
replicate as 3-2-1. The phenes that we demonstrate here (in Section 4) have this simple
symmetry (i.e., the strands and their mirror images both fold into the same polygonal
shapes), so this is not a problem.
The rules for up bonds in genes are different from the rules for up bonds in phenes. In
phenes, only certain combinations of types will bond on their respective up arms, as
given in Table 5.
12
Self-Replication and Self-Assembly for Manufacturing
Insert Table 5 here.
Any type of machine can form a sideways bond with any other type, in both genes and
phenes, but a left arm must bond with a right arm (i.e., no left-left nor right-right bonding
is allowed). In phenes, the types on each side of a sideways bond govern the angle the
bond will take when the strands have folded, as specified in Table 6. This means that the
types involved in each bond control the shape that the folded strand will take.
Insert Table 6 here.
Table 6 shows the sideways bond angle formed by each pair of machine types, when a
strand is in its phenotype state; that is, when folded is set to 1 (true). When a strand is in
its genotype state (folded is 0 for all machines in the strand), the sideways bond angles
are all 0° (straight). Up bond angles are always 0° (
ignoring random perturbations, from
Brownian motion, for example).
The blank cells in Table 6 represent combinations of types for which we have not yet
found a use. From Table 5, it can be seen that a type-3 machine cannot form an up bond
with another type-3 machine when they are in phenes. Therefore we could control the
shape of a 3-3-…-3 phene by specifying any desired value for the angle of 3-3 bonds in
Table 6, but the resulting phenes would not be able to form a mesh. Similarly, we could
control the shape of a 3-4-3-4-…-3-4 phene by giving a ny desired value for the angles of
3-4 and 4-3 bonds in Table 6, but the resulting phenes can form up bonds in multiple
ways (3-4, 4-3, and 4-4; see Table 5). Thus we have limited control over the shape of
the mesh that the phenes will form.
In Table 6, it can be seen that all sideways bonds involving type-1 machines are straight.
This allows us to expand the size of a phene, without changing its shape, by inserting a
sequence of type-1 machines along each edge (see Figure 2). However, it does not
allow polygons with exactly two machines on each side, since expansion requires at
least one type-1 machine inserted between two other machines.
Insert Figure 2 here.
13
Self-Replication and Self-Assembly for Manufacturing
Given the angles that are available to us in Table 6, some polygons (e.g., octagons and
squares) require two types of machines, while others (e.g., triangles and hexagons)
involve only one type. We chose to restrict the JohnnyVon 2.0 to four types of machines,
in order to demonstrate that a small number of components can be combined to build a
variety of structures (like Lego blocks). The angles that we chose make it easy to build
triangular meshes. It may seem inconvenient to require two types of machines to build
octagonal meshes, but in fact, having two types of machines is helpful with octagons. An
octagonal mesh has both octagonal holes and square holes (see Image 4 in Figure 6 in
Section 4.2). W ith two types of machines, we can prevent octagonal phenes from filling
in the square holes in the mesh.
A single type of machine would be sufficient to create squares, but using two types
permits rectangles that will mesh correctly. W ith two types of machines, the long and
short sides of the rectangle can be distinguished by type, so that two sides will bond
together only if they have the same machine type, and thus the same length (see Figure
3).
Insert Figure 3 here.
Though they could use two types, hexagons will form a mesh faster with only one type.
Furthermore, in order to get the desired behaviour with four machine types, exactly one
of the squares or hexagons had to use only one machine type, and the other had to use
two. While it would be possible to have an irregular hexagon, this seems much less
natural than a rectangle.2
Table 5 shows how machine types control phene bonding in meshes. The bonding rules
in Table 5 were designed so that octagonal meshes will form correctly. They prevent
octagonal phenes from filling square holes in the mesh.
Additional rules were created to support expansion of the size of a phene, without
changing its shape, by inserting a sequence of type-1 machines along each edge. These
rules are based on the derived variable bend-location. In a phene, a given machine can
have either a straight (type 1) or bending (types 2, 3 or 4) machine bonded to its left or
right arm. By looking at its neighbours, it can determine where it is in the context of the
2 In a regular polygon, all sides have the same length. Rectangles are irregular.
14
Self-Replication and Self-Assembly for Manufacturing
phene, and hence it can calculate the value of bend-location. Certain values of bend-
location can override the rules in Table 5 and disallow a bond that would otherwise be
permitted. Table 7 explains the meaning of the different values of bend-location. Table 8
shows how bend-location affects bonding.
Insert Table 7 here.
Insert Table 8 here.
When the up field of a machine in one phene overlaps with the up field of a machine in
another phene, the rules in both Table 5 and Table 8 must be satisfied before the up
fields can bond. If there are no type 1 machines in the phene (i.e., all sideways bonds
are bent; the shape is not expanded), then bend-location must have the value 3 for all
machines, and thus (by Table 8) up bonds depend only on the machine types (Table 5).
In Table 7, we are assuming that the phene forms a closed loop. The error correction
system will destroy open loops (Section 3.4.4). When a machine has no left or right
neighbour (because it is at the end of an open strand), we treat the missing neighbour as
if it were a bending type.
3.4.3 Overlap Detection
Because of flexibility in the mesh, and in individual phenes, two phenes can sometimes
join a mesh in such a way that their desired positions overlap. This can be most easily
seen by imagining a mesh of hexagons that is complete except for a single gap.
Suppose that two hexagons jostle (by Brownian motion) into the gap, with slightly
different alignments. At roughly the same time, one forms a bond with the phene above
the gap, and one with the phene below the gap. As they straighten towards their ideal
position (due to twisting forces on their up bonds), each phene may pick up new bonds
around the edge of the former gap (see Figure 4).
Insert Figure 4 here.
If ignored, this problem spreads, since there are now unbonded up arms on each of the
two overlapping hexagons, which permit new hexagons to join the mesh, overlapping
15
Self-Replication and Self-Assembly for Manufacturing
those around the former gap, and each of these hexagons in turn can bring in more
overlapping hexagons.
To address this problem, an arm was added to the machines (new since JohnnyVon
1.0), called the overlap detector arm. In a folded strand, it points towards the center, and
will only bond with other overlap detector arms. Both machines must be oriented in the
same direction (up to a fixed degree of tolerance), and both machines must have their
in-mesh flag set to true.
When an overlap detector bond is formed between two machines (necessarily in two
different phenes), one machine (chosen arbitrarily) sets its unfold signal to true. This
signal propagates to its left and right neighbours, setting folded to false as it goes. It also
breaks the overlap bond that triggered it. The resulting unfolded strand behaves exactly
like a newly replicated strand. It tries to replicate until fold-counter exceeds its limit, and
then it folds up again.
In summary, when two phenes compete for the same gap in a mesh, one of them is
forced to become a gene. Converting one of the phenes to a gene, instead of leaving it
as a detached phene, allows time for it to drift away from the problematic area, or for the
remaining phene to fill up the open bonds in the mesh.
3.4.4 Stress Detection
Another type of error can occur in a mesh, again due to the flexibility of the mesh. In this
case, we can imagine five triangles bonding to form a pie shape, missing only one more
triangle to form a hexagon. However, instead of a new triangle coming in to fill the gap,
the two triangles on either side of the gap jostle together, forming a stressed pentagon
rather than a hexagon. This pentagonal mesh may be part of a larger mesh, and thus
some of the stress may be distributed through the larger mesh. This problem can be
partially addressed by increasing the strength of some of the fields and decreasing the
tolerance of some of the bonding angles, but it becomes increasingly hard to prevent as
the mesh grows.
To detect this kind of problem, each machine maintains a stress-counter, which
increments each time interval when the machine is not in-tolerance, and is reset
whenever the machine is in-tolerance. This counter can be used to detect cases in which
the mesh is stressed because phenes have bonded incorrectly. When the counter
exceeds a fixed maximum, it causes the stressed phene to unfold, by setting its unfold
16
Self-Replication and Self-Assembly for Manufacturing
signal to true. The signal propagates to left and right neighbours, setting folded to false
and dropping up bonds as it spreads through the neighbours.
The in-tolerance variable is purely local. Its state depends only on the relation between a
machine and its immediate neighbours. Likewise, the unfold variable is purely local, but
the information that it conveys can spread through each machine in a phene. When one
machine’s unfold variable becomes true, its immediate neighbours sense this, and then
set their unfold variables to true.
3.4.5 Seeding the Mesh
In the initial seed gene, seed-gene is set to true, but it will be set to false for all of the
child genes. If a machine has a true seed-gene, it will never trigger the fold-now signal.
Since the initial seed gene will never fold, there will always be a strand that can continue
replicating whenever free machines become available.3
The first child of the seed gene, and only the first child of the seed gene, becomes the
seed phene. When the seed-phene flag is set to true in a strand, it does not mean that
the given strand is the seed phene; it means that the next child of the given strand will
become the seed phene. In the initial seed gene, seed-phene is set to true. When the
seed gene first replicates, its child examines its parent ’s seed-phene flag and observes
that it is set to true. The child then sets it in-mesh flag and its folded flag to true and it
immediately folds to become the seed phene. The parent (the initial seed gene) then
sets its seed-phene flag to false, so that its future children cannot become seed phenes.
(When we say that a strand sets a flag to a value, it is a shorthand way of saying that
every machine in the strand sets the flag to the value. Strands do not have flags of their
own, other than the flags of their component machines.)
Every other gene, created after the first child, will begin its career with its in-mesh flag
set to false. If two phenes meet with their in-mesh flags set to false, they cannot bond
3 When the seed gene replicates, all of the information that it encodes, by the sequential order of
the machine types in the strand, is also replicated. However, the child strand will have its seed-
gene flags set to false, so it m ight be argued that the seed gene has not fully self-replicated, in a
very strict sense. The purpose of the seed-gene flag is merely to ensure that replication will
continue after a hiatus in the supply of free machines. This is not essential to the simulation. We
could remove this flag without causing any major problems. Therefore full self-replication, in a
very strict sense, is readily attainable, if desired, by a m inor modification to JohnnyVon.
17
Self-Replication and Self-Assembly for Manufacturing
together. A phene can only bond to another phene if the other phene has its in-mesh flag
set to true. When a machine (in a phene) with a false in-mesh flag meets a machine (in a
second phene) with a true in-mesh flag, they bond (assuming they meet all the
conditions in Section 3.4.2), and a signal propagates through the first phene, setting all
of the in-mesh flags to true (but the signal only propagates from one machine to its
sideways neighbour when their bond is in-tolerance; see Section 3.4.4). This ensures
that the mesh can only grow from the seed phene.
3.4.6 Tolerances
Each machine will only form up bonds if all existing bond angles are within a certain
tolerance. That is, if a machine’s sideways bonds are at angles significantly different
from the desired angles (i.e., the angles given by the rules in Section 3.4.2), then no up
bonds will form during the current timestep. This prevents unintended up bonds during
vulnerable times, such as during splitting or folding.
3.4.7 Shattering
There are a number of ways that a gene or phene can break. For example, during
splitting, the phase of self-replication when two genes are pushed apart by their repellor
arms, if one of them hits the wall of the container at an angle, it puts significant strain on
the whole strand. As another example, an error in a mesh can eventually lead to enough
strain to pull a phene apart (see Section 3.4.4).
If a machine loses a bond unexpectedly (which is any time other than when splitting or
unfolding), or if it notices that its neighbour has folded, then the shatter flag is set to true.
When a machine observes that its neighbour ’s shatter flag is true, the machine may
respond by setting its own shatter flag to true. We say that the first machine is the
source of a shatter signal that was received by the second machine.
The shatter signal always propagates through sideways bonds, setting the shatter flag to
true in left-neighbours and right-neighbours. The shatter signal may also propagate to an
up-neighbour, but only if the source machine has replicated but not folded. If the
neighbouring machine has replicated, something went wrong with the split (two
replicated machines should not be bonded before they’re both fol ded); on the other
hand, if the neighbour ’s replicated flag is false, then it may be part of an incomplete
copy, and thus should be abandoned.
18
Self-Replication and Self-Assembly for Manufacturing
When a machine’s shatter flag is true, it drops all of its bonds (the discrete timesteps
ensure that the state is propagated, even if the bonds are broken during that time step,
since machines consult their neighbours’ states as they were at the beginning of the
step). When the bonds have been dropped, it then sets folded, seed-gene, and
replicated flags to false and becomes a free machine.
The shatter mechanism is not a subtle way to handle errors, but we have found it to be
effective. In our simulations, shattering is relatively rare. This error correction mechanism
is similar to Sayama’s method for handling errors in se lf-replicating loops [12], [13].
3.5
Implementation
JohnnyVon 2.0 builds directly on the original JohnnyVon 1.0. Both systems are written in
Java and their source code is available under the GNU General Public License (GPL) at
http://purl.org/net/johnnyvon/.
4 Experiments and Discussion
In our first experiment, we demonstrate the construction of a small mesh of triangles,
highlighting several important points in the replication and assembly. In the next set of
experiments, we demonstrate replication and assembly of meshes built from each of the
supported polygons, with one machine per side. We then show a mesh of polygons with
more than one machine per side, a 3×1 rectangle and a
triangle with three machines per
side. Finally, to demonstrate scalability, we show a large mesh of triangles.
In the following figures, the inner grey square represents the container. The middle of a
machine must stay inside the grey square. (It takes less computation to check whether
the middles are within bounds than to check all of the arms.)
4.1
Self-Replication and Self-Assembly
In Figure 5, the images show a typical run of JohnnyVon 2.0. The run starts with a soup
of 54 free type-2 machines and a seed gene of the form 2-2-2, and it ends with a
triangular mesh.
Image 1: This shows the initial configuration. Each of the free machines is in a random
position and the seed gene is in the center (it is the strand of three machines, forming a
straight line).
19
Self-Replication and Self-Assembly for Manufacturing
Image 2: After 2,385 steps, the first replication is complete. We see two genes,
immediately after they have split and their repellor arms have pushed them apart.
Image 3: The first child of the seed gene is folding up, to become the seed phene. The
seed gene has already begun a second copy.
Image 4: By time 44,235, nearly all of the free machines are now attached to genes.
Since there are so few free machines left, most of these genes cannot complete self-
replication. As some of the incomplete strands’ fold-counters hit their upper limit, they
will fold and release free machines, allowing other genes to complete self-replication.
Image 5: We can see the second phene forming. In this image, it has not completely
folded; the triangle has a small gap at the top.
Image 6: Slightly more than 3,000 timesteps later, the new phene has bonded with the
seed phene.
Image 7: Now many more triangles have folded and joined the mesh. Two triangles
have not yet joined (one is in the lower left corner and the other is near the center).
Image 8: The mesh is almost complete. In the bottom on the right, there is a pentagonal
arrangement of five triangles. This would eventually be corrected (by one of the triangles
releasing and unfolding; see Section 3.4.4), although the container is just barely large
enough to hold a mesh that includes all of the machines, and thus errors may continue
to form even as they are corrected. In a situation where the container constrains the
mesh, it is possible for a machine to get attached to a mesh in such a way that it can
never reach an equilibrium where all of its bonds are in tolerance, since the conditions
for accepting new bonds are much looser than the conditions for detecting stress.
Insert Figure 5 here.
In this simulation, the container is relatively small, and therefore Brownian motion is
relatively strong. W ith strong Brownian motion, free machines are quickly distributed
throughout the container, thus a replicating strand has a steady supply of free machines.
The small container also means that the phenes never have far to go to join the mesh,
and will quickly be bumped into the right position. In a larger container, replicating
strands will consume the machines in their local area, and then replication slows until
diffusion replenishes the supply. It also takes longer for phenes to find a place where
20
Self-Replication and Self-Assembly for Manufacturing
they can join the mesh. We could speed up the action in a larger container by increasing
the Brownian motion (i.e., turning up the heat), but that could damage the mesh.
4.2
Simple Polygonal Meshes
In Figure 6, we show assembled mesh structures. Four different regular polygonal
meshes are shown, all with sides that are one machine in length. Each of these four
simulations was started with a single seed strand and was executed until the mesh was
well developed. The scale of the images in Figure 6 is different from the scale of the
images in Figure 5. These simulations use a container about nine times larger in area
than the simulations in Figure 5. Table 9 summarizes the four simulations in Figure 6.
Insert Figure 6 here.
Insert Table 9 here.
4.3
Fancy Meshes
In Figure 7, Image 1 (timestep 691,900) shows a mesh built of rectangles, rather than
regular polygons. Because squares and rectangles use two types of machines (see
Section 3.4.2), the rectangles only join the mesh if they are correctly oriented. The seed
for Image 1 was 2-4-2-1-2-4-2-1.
Image 2 (timestep 78,800) shows large triangles. The seed was 2-1-2-2-1-2-2-1-2. The
bonds between type-2 machines fold to form the corners, while the type-1 machines
provide the extension to make these triangles larger. In principle, each phene can be
made arbitrarily large using this approach.
Insert Figure 7 here.
4.4
Large Mesh
The image in Figure 8 demonstrates that meshes can grow correctly beyond a small
number of triangles. The seed was 2-2-2. The mesh contains 234 triangles.
Insert Figure 8 here.
21
Self-Replication and Self-Assembly for Manufacturing
4.5
Discussion
The above experiments demonstrate that four types of machines are sufficient to enable
the user to program a variety of meshes. By specifying the seed gene, the user can
create
triangular, square, hexagonal, or octagonal meshes (Section 4.2). The
experiments in Section 4.3 show that the sizes of the holes in the mesh are infinitely
adjustable (in principle; in practice, there will be computational limits). For rectangular
meshes, an infinite number of shapes are possible, by adjusting the ratios of the lengths
of the sides (Section 4.3).
To program JohnnyVon to build a mesh, the most important information is in Table 6.
Looking at this table, the user can easily predict the shape that will be produced by any
given seed. For example, consider the seed 2-4-2-1-2-4-2-1. Take eight sticks of equal
length and put them on a flat surface. Reading the seed from left to right, let the first
stick be “2”. Place the second stick end-to-end with the
first stick, and let it be “4”.
Arrange the two sticks so that their angle is 90°, as gi ven in Table 6. Continue in this
manner until all eight sticks are arranged. The resulting shape will be a rectangle (see
Figure 3).
When programming a mesh, note that some seeds will not form a closed shape, which
will trigger the stress detection (Section 3.4.4). This may be viewed as a bug in the
user ’s program or as a limitation of JohnnyVon 2.0.
5 Limitations and Future Work
JohnnyVon 2.0 has several minor limitations. For example, phenes must be closed for
the system to work correctly. Although a hexagon composed of five machines and a gap
in the sixth side can form a mesh, the error correction system would destroy the resulting
mesh. Because closure of the phenes increases their rigidity, a mesh built of open
phenes would be more flexible, and there may be other interesting effects.
The variety of phenes in JohnnyVon 2.0 is also somewhat limited. Our original goal, to
support all regular polygons that tile the plane (triangles, squares and hexagons), is
satisfied. JohnnyVon 2.0 also supports partial tiling with octagons (square gaps are left
in the mesh), and full tiling with rectangles. However, we would now like to support
concave shapes (e.g., stars), as well as more general polygons. It would be interesting
to enable Penrose tilings and Kepler tilings [3].
22
Self-Replication and Self-Assembly for Manufacturing
The replication phase takes much longer with two (or more) types of machines than it
does with one, since each free machine has fewer places to bond correctly (equivalently,
each machine in the strand has fewer free machines available with which it can bond).
Supplying two (or more) times as many machines increases the computation per
timestep (roughly quadratically). However, JohnnyVon 2.0 should be parallelizable. This
is another area for future work.
Like JohnnyVon 1.0, version 2.0 still runs on a standard desktop computer, thanks in
part to improvements in hardware since the development of version 1.0. However, there
were many experiments we wanted to try (e.g., polygons with 4 or 5 machines per side)
that were not practical, given our available hardware and our patience. This problem can
be addressed by improving the efficiency of our implementation, converting the code to a
more efficient language than Java (which is likely to make it much less portable),
parallelizing the code, or obtaining better hardware.
The computational complexity of the simulation increases with the size of the phenes,
since each phene must be jostled to a place near where it belongs, and larger phenes
move more slowly. Meshes of large phenes require many timesteps to be constructed.
The problem may be alleviated by increasing the Brownian motion or decreasing the
viscosity of the simulated liquid, but each of these solutions presents new problems. We
have tuned the physical parameters, in an effort to balance these conflicting concerns.
The current settings of the physical parameters appear to strike a good balance, but
there is likely room for further improvements.
JohnnyVon would benefit from increased realism and increased programmability.
Although JohnnyVon 2.0 provides a moderate level of programmability, it is not as
programmable as we would like. One problem is that the mesh grows without control.
Sometimes the mesh is relatively dense (as in Image 2 of Figure 6) while at other times
it has many gaps (as in Figure 8). We would like to add a programmable mechanism for
controlling the final size and shape of the mesh, and for avoiding meshes with large
gaps (or for deliberately creating gaps, which may be useful for some applications).
The growth of a mesh from a seed phene may be viewed as a type of dendrite growth,
such as is observed in Diffusion-Limited Aggregation (DLA) models (a class of fractal
growth models) [22]. The literature on fractal growth models may suggest some useful
methods for preventing dendrite growth and encouraging the formation of a filled mesh
[20]. It may be possible to modify JohnnyVon so that the user can specify the desired
23
Self-Replication and Self-Assembly for Manufacturing
fractal dimension of the mesh. A filled mesh (in the current two-dimensional simulation)
would have a fractal dimension of 2, and a dendritic mesh would have a fractal
dimension somewhere between 1 and 2.
In the context of JohnnyVon’s virtual physics, it may be m eaningful to define a universal
constructor (see Section 2.1). For example, we might say that a universal constructor
would be capable of building any two-dimensional structure that can be constructed from
a finite number of machines, such that up arms are bonded to up arms and left arms are
bonded to right arms. The design for the structure should be encoded in a seed gene.
Ideally, the seed would contain many fewer machines than the final structure, although
this may not be possible when the final structure lacks a regular pattern. Much further
work is required to make a universal constructor in the JohnnyVon model.
JohnnyVon 2.0 also provides a moderate level of realism, but again it is not as realistic
as we would like. Our attractive and repulsive forces are somewhat unlike electrical or
magnetic attraction and repulsion. The JohnnyVon simulation also does not attempt to
model conservation of energy. Arbesman has recently done some interesting work on
computational simulation of artificial life with conservation of energy [1].
Other steps towards increased realism would be to extend the simulation to three
dimensions and to model the physics of the internal operations of the machines.
Currently the external relations between machines are governed by a simple virtual
physics, but the internal operations are described by abstract finite automata. However,
both of these steps to realism would involve a significant increase in computational
complexity.
6 Applications
In our previous work, we suggested that JohnnyVon 1.0 provided a plausible mechanism
for nanoscale manufacturing [17]. A vat of liquid containing free machines would be
seeded with a single strand, soon resulting in a vat full of copies of the seed strand.
JohnnyVon 2.0 takes this application one step further, beyond self-replication to
programmable construction of meshes. Since the user has some control over the size
and shape of the holes in the mesh, we can imagine these meshes being produced for
filtration, insulation, or simply as a kind of cloth.
24
Self-Replication and Self-Assembly for Manufacturing
If we can create a mechanism for controlling the size and shape of the mesh, more
applications become possible. Since the system is accurate and self-correcting, pieces
of cloth could be created exactly to specification, down to the size of a single machine.
The design of the individual machines in JohnnyVon is too complex to be implemented
at the nanoscale, given the current state of the art of nanotechnology. However, the
machines appear to be less complex than typical virsuses, which range in size from
about 20 to 250 nanometers. Nanoscale is generally defined as about 1 to 100
nanometers. As mentioned earlier, JohnnyVon 2.0 was partly inspired by Seeman’s work
with DNA [14], [15]. A single DNA molecule is about 2 nanometers wide. JohnnyVon’s
design is somewhat different from Seeman’s work, so it w ould not be accurate to
describe his work as an implementation of JohnnyVon, but it does seem reasonable to
say that JohnnyVon could be implemented at the nanoscale, given improvements in the
state of the art.
On the other hand, it would not be difficult to implement JohnnyVon at the human scale,
with current technology, using small robots (such as the Khepera robot) to implement the
four types of machines.4 This might be useful for research purposes, to test designs at a
level between software simulation and nanoscale hardware.
7 Conclusion
JohnnyVon 1.0 demonstrated self-replication in a continuous two-dimensional space
with virtual physics. JohnnyVon 2.0 goes beyond its predecessor by introducing a user-
programmable phenotype, consisting of a variety of meshes. JohnnyVon 2.0 is more
realistic than cellular automata models [6], [10], [11], [12], [13], [18], [19], more
programmable than artificial chemistry models [4], [5], and more computationally
tractable than von Neumann’s universal constructor [21]. However, there is still much
room for improvement in the degree of physical realism of the simulation and in the
degree of programmability of the phenotype.
Like its predecessor, JohnnyVon 2.0 is a local model. There is no global data structure
that represents strands or meshes; these are emergent entities that arise from the
interactions of the basic elements (the machines). Each machine is autonomous and can
only sense its immediate neighbours. Control is local, distributed, and parallel.
4 http://www.k-team.com/robots/khepera/khepera.htm l
25
Self-Replication and Self-Assembly for Manufacturing
From four different types of machines, JohnnyVon can produce four different polygonal
meshes, with an infinite number of possible sizes (as per Section 4.3). The user can
specify the mesh that will be produced by encoding the desired size and shape in the
initial seed, without making any changes to the physics of the simulation. Errors in
replication and in mesh formation are automatically detected and corrected, using purely
local mechanisms.
JohnnyVon 2.0 also avoids the “grey goo” scenario of self
-replicating nanobots run
amok. Replication and assembly are inherently limited by the supply of machines; when
the free machines have all bonded, the process stops.
Acknowledgements
Thanks to Arnold Smith for starting us down this path, with JohnnyVon 1.0. Thanks to
the anonymous reviewers of Artificial Life for their helpful comments.
References
[1] Arbesman, S. (2004). Erg: A Computational Energetics as an Approach to the Study
of the Origins of Life. Senior Thesis, Computer Science Department, Brandeis
University.
[2] Drexler, K.E. (1992). Nanosystems: Molecular Machinery, Manufacturing, and
Computation. New York: W iley.
[3] Grunbaum, B., and Shephard, G.C. (1986). Tilings and Patterns. W.H. Freeman.
[4] Hutton, T.J. (2002). Evolvable self-replicating molecules in an artificial chemistry.
Artificial Life, 8, 341-356.
[5] Hutton, T.J. (2004). A functional self-reproducing cell in a two-dimensional artificial
chemistry. In J. Pollack et al., eds., Proceedings of the Ninth International
Conference on the Simulation and Synthesis of Living Systems (ALIFE9), 444-449.
[6] Langton, C.G. (1984). Self-reproduction in cellular automata. Physica D, 10, 134-
144.
[7] Maynard Smith, J., and Szathmary, E. (1995). The Major Transitions in Evolution.
Oxford University Press.
26
Self-Replication and Self-Assembly for Manufacturing
[8] Merkle, R.C. (1992). Self replicating systems and molecular manufacturing. Journal
of the British Interplanetary Society, 45, 407-413.
[9] Merkle, R.C. (1994). Self replicating systems and low cost manufacturing. In The
Ultimate Limits of Fabrication and Measurement, M.E. Welland, J.K. Gimzewski,
eds., Dordrecht: Kluwer, pp. 25-32.
[10] Morita, K., and Imai, K. (1997). A simple self-reproducing cellular automaton with
shape-encoding mechanism. In C.G. Langton and K. Shimohara, eds., Artificial Life
V: Proceedings of the Fifth International Workshop on the Synthesis and Simulation
of Living Systems, 489-496. Cambridge, MA: MIT Press.
[11] Reggia, J.A., Lohn, J.D., and Chou, H.-H. (1998). Self-replicating structures:
Evolution, emergence and computation. Artificial Life, 4, 283-302.
[12] Sayama, H. (1998). Introduction of structural dissolution into Langton's self-
reproducing loop. In C. Adami, R.K. Belew, H. Kitano, and C.E. Taylor, eds.,
Artificial Life VI: Proceedings of the Sixth International Conference on Artificial Life,
114-122. Los Angeles, CA: MIT Press.
[13] Sayama, H. (1999). A new structurally dissolvable self-reproducing loop evolving in
a simple cellular automata space. Artificial Life, 5, 343-365.
[14] Seeman, N.C. (2003). DNA in a material world. Nature, 421 (January 23), 427-431.
[15] Seeman, N.C. (2004). Nanotechnology and the double helix. Scientific American,
290 (6) (June), 65-75.
[16] Sipper, M. (1998). Fifty years of research on self-replication: An overview. Artificial
Life, 4 (3), 237-257.
[17] Smith, A., Turney, P., and Ewaschuk, R. (2003). Self-replicating machines in
continuous space with virtual physics. Artificial Life, 9, 21-40.
[18] Tempesti, G. (1995). A new self-reproducing cellular automaton capable of
construction and computation. In F. Moran, A. Moreno, J.J. Morelo, and P. Chacon,
eds., Advances in Artificial Life: Proceedings of the Third European Conference on
Artificial Life (ECAL95), 555-563. Berlin: Springer-Verlag.
[19] Tempesti, G., Mange, D., and Stauffer, A. (1998). Self-replicating and self-repairing
multicellular automata. Artificial Life, 4, 259-282.
27
Self-Replication and Self-Assembly for Manufacturing
[20] Vicsek, T. (1992). Fractal Growth Phenomena. Second Edition. Singapore: World
Scientific.
[21] von Neumann, J. (1966). Theory of Self-Reproducing Automata. Edited and
completed by A.W. Burks. Urbana, IL: University of Illinois Press.
[22] W itten, T.A., and Sander, L.M. (1981). Diffusion-limited aggregation, a kinetic critical
phenomenon, Physical Review Letters, 47, (Issue 19-9), 1400-1403.
28
Self-Replication and Self-Assembly for Manufacturing
Table 1. Variables for elements of the state vector that represent internal aspects of the machine.
Variable name
type
Range
{1, 2, 3, 4}
id
fold-counter
repel-counter
{0, 1, 2, …}
{0, 1, 2, …}
{0, 1, 2, …}
stress-counter
strand-position
{0, 1, 2, …}
{1, 2, 3}
split-state
{1, 2, 3, 4}
reset-counter
fold-now
unfold
seed-gene
seed-phene
in-mesh
replicated
shatter
folded
{0, 1}
{0, 1}
{0, 1}
{0, 1}
{0, 1}
{0, 1}
{0, 1}
{0, 1}
{0, 1}
Description
• type of machine
• static for a given machine
• unique identifier for machine
• static for a given machine
• used to decide when the strand should fold
• during splitting, controls how long the repellor arms of a
strand are active
• counts the time since a machine was last in-tolerance
• used to decide where a machine is in a replicating strand
• described in detail elsewhere [17]
• used to determ ine when to split
• described in detail elsewhere [17], except that split-state
now has a fourth value, indicating that shatter should be
set to 1 (true)
• indicates that the fold-counter should be reset
• indicates that each machine in the strand should set its
folded flag
• indicates that a phene should unfold
• this occurs when one phene overlaps another in a mesh
• flag for identifying the seed gene
• the seed gene never folds, in case more free machines
become available for replication
• flag for making the seed phene
• indicates whether this machine is connected to the mesh
• 1 (true) if and only if the machine has been through a
successful replication.
• in particular, machines in the seed gene have not
replicated at the start of a simulation
• indicates that the machine should break all bonds and
return to being a free machine.
• if 1, the machine tries to form angular bonds with its left
and right neighbours
• if 0, the machine tries to form straight bonds with its left
and right neighbours
29
Self-Replication and Self-Assembly for Manufacturing
Table 2. Variables for elements of the state vector that represent external relations.
Variable name
x-position
y-position
angle
x-velocity
y-velocity
angular-velocity
left-neighbour
right-neighbour
up-neighbour
Range
´
´
[0, 2p]
´
´
[0, 2p]
{0, 1, 2, …}
{0, 1, 2, …}
{0, 1, 2, …}
Description
• state of the machine with respect to the container
• vary due to Brownian motion, viscosity, and forces from
interactions between fields
• identifier of the machine (if any) bonded to the named
arm
30
Self-Replication and Self-Assembly for Manufacturing
Table 3. Derived variables that are calculated from state variables.
Variable name Range
in-tolerance
{0, 1}
bend-location
{1, 2, 3, 4}
Description
• indicates whether each existing bond is within a certain
(fixed) tolerance of the desired angle
• used to help avoid making bonds when the machine is in
a potentially unstable situation.
• indicates where a machine is located in a phene
31
Self-Replication and Self-Assembly for Manufacturing
Table 4. Pairs of machine types that will perm it an up bond when in genes.
Bond
1
2
3
4
1 2 3 4
+
+
+
+
32
Self-Replication and Self-Assembly for Manufacturing
Table 5. Pairs of machine types that will perm it an up bond when in phenes.
Bond
1
2
3
4
1 2 3 4
+
+
+ +
33
Self-Replication and Self-Assembly for Manufacturing
Table 6. Folding angles for sideways bonds between two machine types when in phenes.
Angle
1
2
3
4
1
0°
0°
0°
0°
2
3
4
0°
120°
45°
90°
0°
45°
0°
90°
60°
34
Self-Replication and Self-Assembly for Manufacturing
Table 7. The meaning of the different values of bend-location.
Value Meaning of value
right of bend: right neighbour is straight (1) and left neighbour is bending (2, 3, 4)
1
left of bend: left neighbour is straight (1) and right neighbour is bending (2, 3, 4)
2
in bend: both left and right neighbours are bending types (2, 3, 4)
3
extender: both left and right neighbours are straight types (1)
4
35
Self-Replication and Self-Assembly for Manufacturing
Table 8. Pairs of bend-location values that will perm it an up bond when in phenes.
bend-location
1
2
3
4
1 2 3 4
+
+
+
36
Self-Replication and Self-Assembly for Manufacturing
Table 9. Some basic observations about each image in Figure 6.
Image Phenes
Seed gene
Timestep
1
2
3
4
Triangles
Squares
Hexagons
Octagons
2-2-2
4-2-4-2
4-4-4-4-4-4
2-3-2-3-2-3-2-3
246,000
498,600
448,600
1,107,400
Initial free
machines
201
200
160
120
Phenes in
mesh
50
24
20
9
Genes
remaining
5
1
1
2
37
Self-Replication and Self-Assembly for Manufacturing
Figure 1. This diagram shows how fold-counter resets when a new machine joins a partially
replicated gene. A free machine attaches to a machine in the strand (A), triggering the reset-
counter signal and resetting fold-counter in the machine to which it attaches (B). The signal
propagates (C) to the left neighbour (D), triggering its reset-counter signal and causing it to reset
its fold-counter. In turn, the signal passes (E) to the next machine in the strand (F). The
propagation stops at the leftmost machine (F).
38
Self-Replication and Self-Assembly for Manufacturing
Figure 2. This diagram demonstrates the mechanism used to expand the size of a polygon.
Machines of type 1 are inserted along the edges.
39
Self-Replication and Self-Assembly for Manufacturing
Figure 3. In this diagram, we can see why it is useful to have two types of machines for
rectangles (types 2 and 4), in addition to type 1, for expanding size. The top phene is oriented
incorrectly, relative to the mesh formed by the bottom two phenes. However, the top phene will
not join at (A) or (B), because the types do not match (see Table 5). It will only join the mesh
when it is oriented correctly.
40
Self-Replication and Self-Assembly for Manufacturing
Figure 4. In this diagram, two phenes join a mesh, both trying to fill the same gap. Phene 1 has
joined at (A) and (C). Phene 2 has joined at (B). Because these two phenes both have their in-
mesh flags set to 1 (true), their overlap detector arms become active, and the two machines at
(D) briefly bond their overlap detector arms. The machines then decide arbitrarily which one will
set its unfold flag to true, causing the phene to which it belongs to unfold. The remaining phene
will then pick up the bonds that were released by the unfolded phene.
41
Self-Replication and Self-Assembly for Manufacturing
Image:
1
Time:
0
Image:
3
Time:
3,295
Image:
5
Time:
46,340
Image:
7
Time:
85,810
Image:
2
Time:
2,385
Image:
4
Time:
44,235
Image:
6
Time:
49,780
Image:
8
Time:
294,075
Figure 5. These images illustrate the experiment described in Section 4.1. The eight images are
consecutive screenshots from a typical run of JohnnyVon, selected to show interesting stages in
the simulation. The first image shows a seed gene in a soup of free machines and the last image
presents a nearly completed mesh of triangles.
42
Self-Replication and Self-Assembly for Manufacturing
1
3
2
4
Figure 6. These screenshots illustrate the resulting meshes for each of the four supported regular
polygons. The images show the final stages of four separate simulations, each using a different
seed gene, as described in Section 4.2 and Table 9.
43
Self-Replication and Self-Assembly for Manufacturing
1
2
Figure 7. Image 1 shows a rectangular mesh and Image 2 is a mesh of expanded triangles.
These are screenshots from the final stages of two separate simulations (see Section 4.3). They
demonstrate that it is possible to vary the lengths of the sides of the polygons, by encoding the
desired lengths in the seed gene.
44
Self-Replication and Self-Assembly for Manufacturing
Figure 8. This screenshot shows a mesh of 234 triangles, as discussed in Section 4.4. This
demonstrates that JohnnyVon can scale up to larger meshes.
45
|
1204.6090 | 1 | 1204 | 2012-04-27T00:34:18 | Towards a Formal Model of Privacy-Sensitive Dynamic Coalitions | [
"cs.MA",
"cs.DC"
] | The concept of dynamic coalitions (also virtual organizations) describes the temporary interconnection of autonomous agents, who share information or resources in order to achieve a common goal. Through modern technologies these coalitions may form across company, organization and system borders. Therefor questions of access control and security are of vital significance for the architectures supporting these coalitions.
In this paper, we present our first steps to reach a formal framework for modeling and verifying the design of privacy-sensitive dynamic coalition infrastructures and their processes. In order to do so we extend existing dynamic coalition modeling approaches with an access-control-concept, which manages access to information through policies. Furthermore we regard the processes underlying these coalitions and present first works in formalizing these processes. As a result of the present paper we illustrate the usefulness of the Abstract State Machine (ASM) method for this task. We demonstrate a formal treatment of privacy-sensitive dynamic coalitions by two example ASMs which model certain access control situations. A logical consideration of these ASMs can lead to a better understanding and a verification of the ASMs according to the aspired specification. | cs.MA | cs | Towards a Formal Model of Privacy-Sensitive Dynamic
Coalitions
Sebastian Bab
Nadim Sarrouh
Technische Universitat Berlin, Germany
Technische Universitat Berlin, Germany
[email protected]
SOAMED Graduate College
[email protected]
The concept of dynamic coalitions (also virtual organizations) describes the temporary interconnec-
tion of autonomous agents, who share information or resources in order to achieve a common goal.
Through modern technologies these coalitions may form across company, organization and system
borders. Therefor questions of access control and security are of vital significance for the architec-
tures supporting these coalitions.
In this paper, we present our first steps to reach a formal framework for modeling and verifying
the design of privacy-sensitive dynamic coalition infrastructures and their processes. In order to do
so we extend existing dynamic coalition modeling approaches with an access-control-concept, which
manages access to information through policies. Furthermore we regard the processes underlying
these coalitions and present first works in formalizing these processes. As a result of the present
paper we illustrate the usefulness of the Abstract State Machine (ASM) method for this task. We
demonstrate a formal treatment of privacy-sensitive dynamic coalitions by two example ASMs which
model certain access control situations. A logical consideration of these ASMs can lead to a better
understanding and a verification of the ASMs according to the aspired specification.
1 Introduction
Modern technologies like service-oriented architectures, online social networks, web 2.0, etc. allow for
companies and organizations to form dynamic coalitions across their own system borders. A dynamic
coalition first off is a group of through network technologies (temporary) interconnected autonomous
agents, who share their resources or information in order to reach a common goal. Although each coali-
tion exhibits unique features there exist certain common features like dynamically changing membership,
data-transfer mechanisms and authorization structures. Companies cooperating in order to exploit a tem-
porary market chance or the cooperation of emergency services, military and civilian institutions during
a crisis are popular examples for dynamic coalitions.
In most scenarios different institutions have to share and transfer critical information, which raises
significance for questions of information flow, security, privacy and trust. In security and privacy critical
fields like in the health sector the solving of the question may determine the success of complete software
engineering projects. Nevertheless software architects today still lack formal frameworks and methods to
evaluate policy decisions at design time and thereby examine their consequences for the to-be-deployed
architecture.
In [5] Bryans et al. present formal modeling technologies in order to approach a formal framework
for dynamic coalitions. With the specification language VDM1 they create a basic formal model of
dynamic coalitions, consisting of various dimensions, where each dimension represents a certain per-
spective which could be considered by a software architect. According to object oriented paradigms the
1Vienna Development Method: http://www.vdmportal.org/twiki/bin/view
Jeremy Bryans and John Fitzgerald (Eds.):
Formal Aspects of Virtual Organisations 2011 (FAVO 2011)
EPTCS 83, 2012, pp. 10 -- 21, doi:10.4204/EPTCS.83.2
c(cid:13) Sebastian Bab & Nadim Sarrouh
This work is licensed under the
Creative Commons Attribution License.
Sebastian Bab & Nadim Sarrouh
11
models consist of data types as well as operations and invariants over these types. In [8] the authors
demonstrate the practicability of their modeling approach.
In other works independent of dynamic coalition researches the same authors present concepts for
modeling Access-Control-Policies in VDM. Therefor they translate core components of the OASIS2-
standard XACML (eXtensible Access Control Markup Language into a VDM-Notation. In an outlook
they propose the combination of this approach with the modeling of dynamic coalitions [7].
Modeling and testing of proposed software architectures with respect to their access control policies
is of great importance, especially in dynamic coalitions: Different access control policies of certain
agents may result in processes that do not run smoothly or prevent the processes from running at all.
therefor present a concept of evolution of access control policies in dynamic coalitions
Bryans et al.
(see figure 1). In this concept a process with certain access control requirements is generated through
coalition consensus. With the supposed formal framework agents may investigate these access control
requirements and even simulate the impact on their own architecture, thereby being able to detect and
avert errors or contradictions.
Figure 1: Evolution of access control policies in dynamic coalitions taken from [7]
Since one of the authors is affiliated with the SOAMED Graduate College3, a graduate college dedi-
cated to "Service-oriented Architectures for the Integration of Software-based Processes exemplified by
Health Care Systems and Medical Technology" medical scenarios are of great interest to the authors.
In health care systems dynamic factors of coalitions are of great importance. Possibilities of modern
networking technologies call for new organizational concepts apart from traditional static concepts like
clinical pathways. For example, patient care itself could be seen as a dynamic coalition:
During the treatment in one or several hospitals the patient has to pass through various
station and organizational units. These departments are often not only spatially separated,
2OASIS is a consortium for advancing open standards for information society. See: http://www.oasis-open.org/.
3For more Information on SOAMED see:
http://www.informatik.hu-berlin.de/forschung/gebiete/
soamed-en/willkommen-en.
12
Formal Modeling of Privacy-Sensitive Dynamic Coalitions
but also technology wise, acting as separate, autonomous instances, which create there own
reports and diagnoses which are transferred to the other instances (up to this day still mostly
on paper).
In this paper we present our approach of integrating the above-described modeling concepts for ac-
cess control and dynamic coalitions. Speaking in terms of Bryans et al. we thereby examine the dimen-
sion access control. The models of Bryans et al. solely address structural aspects of dynamic coalitions.
Exceeding these structural features the consideration of coalition processes is of high significance in
dynamic coalitions. In the core of this paper we discuss how Abstract State Machines (ASMs) may be
utilized to enrich the models of Bryans et al. with formal process aspects. The modified VDM-model
is considered as the underlying state and processes may be represented as state transitions in the ASM.
We provide two examples of access control situations in dynamic coalitions in the formal method of
ASMs which shall illustrate the usefulness of a formal treatment and consideration of access control
situations. By the verification of certain liveness and safety properties we can furthermore guarantee or
check specific properties of the ASMs in the first place.
2 Related Work
In recent years several formalizations of single aspects of dynamic coalitions have been presented. Haidar
et al. propose a formal model for PKI-based authentication in dynamic coalitions on the basis of a process
calculus and the formal description language Z [13]. Bocchi et al. present formal description approaches
for breeding environments in virtual organizations, which may be notably relevant in grid-computing.
As mentioned above, Bryans et al. contributed various models for dynamic coalitions by means of
the specification language VDM (see [8, 7, 4, 5]) . However those approaches only cover structural
aspects of dynamic coalition. Process properties and workflows have to be simulated using external (for
example java-based) tools to modify a state that is based on a VDM-model structure. Our work extends
this approach in so far as in that we pursue a formal modeling framework in which both structure and
process properties can be formalized in a single formalism (e.g. Abstract State Machines) and thereby
create the means to formally analyze dynamic coalitions with the underlying processes.
Other than the taxonomies of Bryans et al. only few holistic modeling approaches for dynamic
coalition exist. The work most alike to ours can be found in [18]. Here McGinnis et al. describe a
formal framework for virtual organizations in service-grid-environments. The framework catches formal
descriptions of agents, services, roles and work flows. The used modeling techniques are not based on
standards that we know of. Tool-support in creation and evaluation of dynamic coalitions design, as
proposed by us, is therefor only hardly imaginable. Koshutanski et al. model HDC (highly dynamic
coalitions), a subclass of dynamic coalitions, defined by them as coalitions with extremely short life
time. They also do not make use of any modeling standard, which might create problems in critical
application fields like crisis management or the health sector. Furthermore their model assumes a central
management of the coalition in form of a coalition platform. Both limitations are too restrictive for the
application scenarios we are envisioning.
Other approaches present informal models from the view of economic sciences. [14, 17].
3 Basics
In the following we want to describe the basic concepts which are essential for the understanding of
the model as to be presented in this paper. In the first subsection we give a short introduction into the
Sebastian Bab & Nadim Sarrouh
13
modeling language VDM. In the second subsection we provide the model for dynamic coaliation as
described in [8]. In the third part we give a brief introduction to the OASIS XACML standard, for which
there also exists a VDM specification.
3.1 VDM and VDM++
The model to be introduced in the present paper is specified in the object-oriented extension of the Vienna
Development Method (VDM), called VDM++. This language specifies data, states and functionality and
is well suited for the modeling of the structures and functions of dynamic coalitions.[11, 9].
A VDM-Model consists of definitions of classes in which instance variables, types and functionality
are specified. Here instance variables are the local variables of an object. Types can be defined as being
free of structure (called token) or as complex. Functionality can be specified through functions over
the instance variables or through helping functions, which leave the local variables untouched. The use
of operations and functions can be limited through the statement of pre- and post-conditions. VDM is
widely used method in the modeling of computer-related systems [16, 10] and adequately supported by
tools like for example the VDMTools which were used for the present paper.
3.2 Dynamic coalitions in VDM
The VDM model for dynamic coalitions as introduced by Bryans et al. consists basically of agents, which
can enter or leave certain groups of agents, which are named coalitions. Agents carry a set of information
which can be shared with other agents or coalitions. Based on these conceptions [5] introduces certain
different dimensions of dynamic coalitions for the modeling of seven further aspects like for example
exchanging of information, memberships of coalitions and structures of authorization. The practical
usefulness of these models was proven by certain industrial ([8]) and military ([4]) case studies.
In [1] we extended this basis model of dynamic coalitions by methods for a handling of access-
control. In this respect our approach can be seen as another new dimension of dynamic coalitions: the
dimension of access-control.
3.3 XACML in VDM
XACML is an OASIS standard based on XML for the description and definition of access-control-
policies in distributed systems. A simplified visualization of the functioning of XACML policies can
be found in figure 2. Requests for access to certain resources are picked up by the Policy Enforcement
Point (PEP) and send to the Policy Decision Point (PDP) as a standardized XACML request. As holder of
the policies it is up to the PDP who check whether an access is to be permitted or to be denied and sends
this information to the PEP which is responsible for the granting or denying of access to the resource.
3.4 Abstract State Machines
The concept of abstract state machines (ASMs) (see for example [3, 12]) allows for a formal representa-
tion of computable and non-computable algorithms. Here the concept of ASMs is rather free as it allows
for a direct use of specification languages (like for example VDM) as underlying languages of the ASMs.
ASMs offer a mathematical framework which allows for a reliable formal modeling of processes. Any
ASM is based on the concept of states in which the ASM can occur. Here any state is modeled as a
specification of the underlying specification language. State transitions are modeled over rules which
14
Formal Modeling of Privacy-Sensitive Dynamic Coalitions
Figure 2: Simplified presentation of XACML-architecture.
change the definitions of the functions included in the specification. This change takes place via an up-
date operation, which overwrites the result of the function for a certain set of arguments. Thus a state is
determined by the definition of its included functions.
ASMs allow for a logical handling in the sense that one defines a logic for AMSs in which formulas
describe certain states and conditions of the ASM. As a result of the present work we will show in
the next sections how two examples for a formal representation of access-control questions in dynamic
coalitions can be modeled and studied using the ASM method and its logical considerations.
4 Dimension Access Control: Our VDM state
In the following we present our basic state which is used for the ASM modeling approach in the next
chapter. The models are loosely based on the work of Bryans et al. [6] although it has been modified
and extended on several locations. At first we introduce the basic types of the model, separated in types
for the modeling of dynamic coalitions and access control respectively. In the last subsection we explain
some example operations on the basis of which simple access controlled information transfer within a
coalition becomes possible. 4
4.1 Dynamic coalition types
According to the approaches of Bryans et al. our basis state signature is based on agents which carry
information and may join forces to form dynamic coalitions and thereby sharing information at their
demand. Both, agents and coalitions have unique IDs (Aid- and Cid-token). The model abstracts from
the actual content and structure of the information and considers information as unstructured data (token-
type). Furthermore we add PDPs, both at agent and at coalition level, which host the access control
policies for later evaluation.
public Agent
:: info
aac
: set of Information
: PDP;
public Coalition
:: agents : set of Aid
info
cac
: set of Information
: PDP;
4The full model
ACDimension.vpp
including a test scenario can be found under http://www.user.tu-berlin.de/nsarrouh/
Sebastian Bab & Nadim Sarrouh
15
public Information :: item
: token;
instance variables
coals : map Cid to Coalition := {->};
agents: map Aid to Agent := {->};
inv forall c in set dom coals &
(coals(c).agents subset dom agents)
The invariant (inv) states that only known agents may join coalitions. The PDP of a coalition and the
PDPs of the agents are related as follows: As soon as an agent shares information with the coalition, the
related access control policies are attached to it and stored in the coalitions PDP. This way access control
requirements of the agents may be enforced on coalition level.
4.2 Access control component types
We integrate the approach from [6]: According to the OASIS-standard XACML policies consist of rules
which in turn consist of an optional target) and an effect. Targets consist of subjects, in our case agents,
which want to access certain resources, in our case information. Actions define the type of access in
question, which we limit to read and write for sake of simplicity. When an access (request) matches a
rule target, the effect is being returned which may be permit, deny or not applicable in case the target of
the rule does not match the request. If a rule doesn't contain a target it will be evaluated for each request.
public Rule
:: target
effect
: [Target]
: Effect;
public Target :: subjects : set of Aid
resources : set of Information
actions
: set of Action;
public Request :: target
: Target;
public Action = (<WRITE> <READ>);
public Effect = (<PERMIT> <DENY> <NOTAPPLICABLE>);
Policies are sets of rules including a combination algorithm which combines different effects in
case that more than one rule matches a request. Here we only consider the basic XACML combining
algorithms deny overrides) and permit overrides. Policies have obligatory targets, so that only matching
requests are evaluated. The effect of a policy is returned to the Policy Decision Point (PDP), which
combines the different effects to only one effect, using the above-mentioned combining algorithm. Each
request for access on information is evaluated through the PDP which ensures policy compliance.
public PDP
:: policies
: set of Policy
policyCombAlg : CombAlg;
public Policy :: target
rules
ruleCombAlg
: Target
: set of Rule
: CombAlg;
public CombAlg = (<DENYOVERRIDES> <PERMITAOVERRIDES>);
16
Formal Modeling of Privacy-Sensitive Dynamic Coalitions
4.3 Operations
The model operations CreateEmptyCoalition, CreateNewAgent and Join create empty coalitions, new
agents carrying information, and join agents to coalitions. Agents in coalitions may share information
(method shareInfo) in which case the matching policies are added to the coalitions PDP.
Figure 3: Visualization of shareInfo operation.
public shareInfo : Aid * Cid * set of Information ==> ()
ShareInfo(a,c,i_set)==
(
for all i in set i_set do
coals:= coals ++ {c -> mu(coals(c),
info -> coals(c).info union i_set,
cac -> mk_PDP(coals(c).cac.policies union
GetMatchingPolicies(agents(a).aac, i),
<DENYOVERRIDES>))};
)
pre a in set dom agents and c in set dom coals and
i_set subset agents(a).info
;
If an agent wants to access information, the PDP will at first evaluate the request and then permit or
deny the access.
public RequestInfo : Aid * Cid * Action * set of
Information ==> Effect
RequestInfo(a,c,act,i_set)==
(
evaluatePDP(mk_Request(mk_Target({a},i_set,{act})),
coals(c).cac)
)
pre a in set dom agents and c in set dom coals and
i_set subset coals(c).info
;
Sebastian Bab & Nadim Sarrouh
17
The function evaluatePDP checks the combining algorithm for the policies and calls the according
functions. The single policies and rules are evaluated through evaluatePolicy and evaluateRule which in
turn use the according function depending of the combining algorithm to evaluate the effects. For the
sake of space we refrain from an explicit representation of all the functions and operations in this paper
and refer to the complete model descriptions in the internet.
5 Abstract State Machines
Abstract state machines (ASMs) offer a formalism for the modeling of arbitrary algorithms (see for ex-
ample [12, 3]). ASMs come with different forms of presentation of algorithms, including technical and
visualized presentation forms. Furthermore there exists a tool support which allows for the actual pro-
gramming with ASMs (see for example the CoreASM project). ASMs offer methods for a representation
of algorithms which are not necessarily limited to computer based algorithms, but instead of that allow
for the representation of other algorithmic procedures of the real world, too. The essential advantage
of ASMs for the purposes of this work lies in the freedom of using nearly any modeling language and
formalism as underlying formalism of a certain ASM. Thus it is possible to extend formalisms like VDM
by a meta-level consideration which allows for example to reason about a VDM model from a logical
perspective.
5.1 ASM Example
We now try to illustrate the applicability of the ASM method through a simple example for access control
in a dynamic coalition taken from [7]. This example is chosen due to the fact that it includes the essen-
tial questions and aspects of an access control in a dynamic coalition, while still being of an adequate
simplicity. Assume that a document management tool administers production plan and hazard analysis
documents of a chemical plant named compA. Production orders for chemicals are stored in an order
database and each order has to undergo identical production processes. Due to market opportunities
compA outsources the hazard analysis to compB, thereby forming a kind of dynamic coalition.
Figure 4 visualizes this process in the standard ASM terminology as introduced in [3, 12]. The
underlying structure is the above-mentioned VDM++-Model. According to common ASM-illustrations
we refrain from defining all functions and variables explicitly in this paper and count on the expres-
sive power of function and variable names. Round entities represent the abstract states, i.e. the whole
VDM++-structure, including instances for the agents, coalitions, etc. whereas rectangles illustrate up-
dates of these states. hexagons are symbolic representations of conditions. If-then-else-conditions have
two possible outputs "yes" and "no" depending on the Boolean expression inside the hexagon. The pro-
cess for the pre-production of a new chemical will start with the creation of a hazard analysis. According
to the hazard analysis and its recommendations the production plan is developed. Suppose that legal
requirements demand that compB signs off that all security recommendations of the hazard analysis are
properly implemented in the production plan.
Now suppose that company legislation and therefore the access control policy of compA forbids any
other agent but compA to access the production plan, which is reflected by the following two rules5:
ruleA1:= ( Target( {compA} ,
{PP} , {<WRITE>,<READ>}
)
,
<PERMIT> );
5In order to let ruleA1 be effective the policy will have to use the PermitOverrides combining algorithm.
18
Formal Modeling of Privacy-Sensitive Dynamic Coalitions
Figure 4: ASM visualization of chemical plant process with possible deadlock after i6.
ruleA2:= ( Target( {}
, {PP}
, {<WRITE>,<READ>}
)
,
<DENY>
);
Because all shared information is stored in the coalition together with the according access control
policies, the depicted process will run into a deadlock as soon as compB tries to access the production plan
in the first condition (we suppose that signoff() will internally make use of the requestInfo()-function).
An ASM run easily identifies this deadlock. The resulting error will now have to be reported to the
coalition members who will now have to adjust their process or their security policies in order to make
the process executable. In the following process we insert another condition, checking if access to the
production plan is granted. If not an error message to the participating members will result and thereby
we eliminate the previous deadlock.
6 Logical considerations -- Liveness and safety
The general method of ASMs allows for a logic based analysis of properties of the abstract state machines
(compare with [3, 12]). Based on the example ASMs of the previous section we now want to argue that
certain important properties in the modeling of processes like for example liveness and safety can be
described by logical formulas over the underlying specifications of the ASMs. Here liveness refers to
the property that the ASM is free of deadlocks while safety refers to the general absence of non-desired
states in the ASM. A logical treatment of questions of liveness and safety can be seen as an indicator for
the correctness of a designated ASM.
Consider for example the first ASM example of the previous section (as represented in figure 5.1).
As mentioned before this ASM cannot be considered as free from deadlocks as a rejection of the request
of compB for the production plan of compA does have not a solution in the ASM. The resulting deadlock
Sebastian Bab & Nadim Sarrouh
19
Figure 5: ASM visualization of chemical plant process with new choice to avoid deadlock after i6.
however can be described by the following formula over the VDM++ specification which is underlying
in the ASM:
ϕ := ∃ord.∃HA.∃PP.(createHA(compB,ord) = HA
∧createPP(ord,HA) = PP∧ requestInfo(compB,coal,PP) = DENY
Due to the general concept of AMSs with their freedom of choice in their underlying specifications and
due to the expressiveness of the describle formulas there cannot exist a general algorithm for the test of
satisfiability of formulas in given ASMs. However, this does not imply that such a test is impossible
to define in any case. The structure of our example allows for a testing of the above formula ϕ in the
ASM as the number of states to consider is finite. It is obvious that the only situation in which ϕ can be
considered as a true formula is when calling the signoff()-function after states i6 and i10. Thus a logical
consideration could reveal the possible deadlock situation before an explicit test run of the ASM has to
take place. Doing that the ASM can be considered as not fulfilling the property of liveness by logical
considerations.
As mentioned before the second ASM example (as represented in figure 5.1) of the previous section
can be considered as free of deadlocks. Thus when attempting to test the validity of the above formula
ϕ in this second ASM the formula occurs to be false in all states and thus the ASM can be considered
as satisfying the liveness property. Logical considerations of safety properties include all the testing for
states of the ASM which are non-desired in the aspired specification of the ASM. Consider for example
the following formula ψ:
ψ := ∃ord.∃HA.∃PP.(createHA(compB,ord) = HA
20
Formal Modeling of Privacy-Sensitive Dynamic Coalitions
∧createPP(ord,HA) = PP∧ requestInfo(compB,coal,PP) = PERMIT
→ sendErr(compB,coal,PP))
This formula describes a situation in which the entire process is working as intended, but the system
nevertheless sends an error to compA. It is obvious that this formula ψ describes an error state of the
ASM which shall not occur as a possible state. Thus when testing the validity of ψ in the ASM and find
ψ to be false in any state we can guarantee a priori that this non-desired state cannot occur in the ASM
run.
Besides the mentioned aspects on liveness and safety there are further benefits of a logical consider-
ation of ASMs (like for example consequence relations or other meta-concepts) which lead to more deep
analysis of their structure and logical conditions.
7 Conclusion and open questions
In this paper we have presented a model in the formal specification language VDM, which may describe
dynamic coalitions together with privacy and access control aspects. Our contribution here is the inte-
gration of two existing modeling approaches for dynamic coalitions and XACML access control policies
respectively. Our next steps in this area will consider context-sensitive XACML as well as role based
access control (RBAC) for both of which OASIS-profiles exist. Both extensions to the basic XACML-
concept are of high significance in most application fields, i.e. in the medical sector as the main field of
interest of the authors.
Furthermore we have argued by our two examples of ASMs that the ASM method with its logical
considerations can be of great benefit for a formal treatment of access-control-sensitive processes in a
dynamic coalition. The VDM-structures serve as the basic states on which the ASMs operate. Here the
examples are meant as a first indicator for the usefulness of a formal treatment as provided by the ASM
method. However, it remains to define and study a general theory for a formal treatment of access control
in dynamic coalitions using the ASM method which is general enough to cover a wide range of access
control situations. These studies are part of ongoing and soon to be published work especially of the
second author of the present paper.
References
[1] Sebastian Bab & Nadim Sarrouh (2011): Formale Modellierung von Access-Control-Policies in Dynamischen
Koalitionen. In: GI Proceedings, Informatik 2011 - Informatik schafft Communities, p. 402.
[2] Laura Bocchi, Jos´e Luiz Fiadeiro, Noor Rajper & Stephan Reiff-Marganiec (2009): Structure and Behaviour
of Virtual Organisation Breeding Environments. In: FAVO, pp. 26 -- 40. Available at http://dx.doi.org/
10.4204/EPTCS.16.3.
[3] Egon Borger & Robert Stark (2003): Abstract State Machines: A Method for High-Level System Design and
Analysis. Springer Verlag.
[4] J. W. Bryans, J. S. Fitzgerald, D. Greathead, C. B. Jones & R. J. Payne (2008): A Dynamic Coalitions
Workbench: Final Report. Technical Report, Newcastle University.
[5] J. W. Bryans, J. S. Fitzgerald, C. B. Jones, I. Mozolevsky, Jeremy W. Bryans, John S. Fitzgerald, Cliff B.
Jones & Igor Mozolevsky (2006): Dimensions of dynamic coalitions. Technical Report.
[6] J. W. Bryans, J. S. Fitzgerald & P. Periorellis (2006): Model Based Analysis and Validation of Access Control
Polcies. Technical Report, Newcastle University, School of Computing Science.
Sebastian Bab & Nadim Sarrouh
21
[7] Jeremy Bryans & John Fitzgerald (2007): Formal Engineering of XACML Access Control Policies in
VDM++.
In Michael Butler, Michael Hinchey & Mar´ıa Larrondo-Petrie, editors: Formal Methods and
Software Engineering, Lecture Notes in Computer Science 4789, Springer Berlin / Heidelberg, pp. 37 -- 56.
Available at http://dx.doi.org/10.1007/978-3-540-76650-6_4.
[8] Jeremy Bryans, John S. Fitzgerald, Cliff B. Jones & Igor Mozolevsky (2006): Formal Modelling of Dynamic
Coalitions, with an Application in Chemical Engineering. Technical Report. Available at http://dx.doi.
org/10.1109/ISoLA.2006.21.
[9] John Fitzgerald, Peter Gorm Larsen, Paul Mukherjee, Nico Plat & Marcel Verhoef (2005): Validated Designs
For Object-oriented Systems. Springer-Verlag TELOS, Santa Clara, CA, USA.
[10] John S. Fitzgerald & Peter Gorm Larsen (2006): Triumphs and Challenges for Model-Oriented Formal Meth-
ods: The VDM++ Experience (Abstract). In: ISoLA, pp. 1 -- 4. Available at http://dx.doi.org/10.1109/
ISoLA.2006.33.
[11] John S. Fitzgerald & Peter Gorm Larsen (2009): Modelling Systems - Practical Tools and Techniques in
Software Development (2. ed.). Cambridge University Press.
[12] Yuri Gurevich, Philipp W. Kutter, Martin Odersky & Lothar Thiele, editors (2000): Abstract State Machines,
Theory and Applications, International Workshop, ASM 2000, Monte Verit`a, Switzerland, March 19-24, 2000,
Proceedings. Lecture Notes in Computer Science 1912, Springer.
[13] Ali Nasrat Haidar, P. V. Coveney, Ali E. Abdallah, Peter Y. A. Ryan, B. Beckles, J. M. Brooke & M. A. S.
Jones (2009): Formal Modelling of a Usable Identity Management Solution for Virtual Organisations. In:
FAVO, pp. 41 -- 50. Available at http://dx.doi.org/10.4204/EPTCS.16.4.
[14] R. Klueber (1998): A framework for virtual Organizing. In: VoNet Workshop.
[15] Hristo Koshutanski & Antonio Mana (2010): Interoperable semantic access control for highly dynamic coali-
tions. Security and Communication Networks 3(6), pp. 565 -- 594. Available at http://dx.doi.org/10.
1002/sec.148.
[16] Peter Gorm Larsen, John Fitzgerald & Tom Brookes (1996): Applying Formal Specification in Industry. IEEE
Softw. 13, pp. 48 -- 56. Available at http://doi.ieeecomputersociety.org/10.1109/52.493020.
[17] Nick Lethbridge (2001): An I-Based Taxonomy of Virtual Organisations and the Implications for Effective
Management. Informing Science 4 No 1.
[18] Jarred McGinnis, Kostas Stathis & Francesca Toni (2009): A Formal Framework of Virtual Organisations as
Agent Societies. In: FAVO, pp. 1 -- 14. Available at http://dx.doi.org/10.4204/EPTCS.16.1.
[19] Wolfgang Reisig (2010): Petrinetze: Modellierungstechnik, Analysemethoden, Fallstudien. Leitfaden der
Informatik, Vieweg+Teubner. 248 pages; ISBN 978-3-8348-1290-2.
|
1203.1570 | 1 | 1203 | 2012-03-07T19:14:32 | In-network Sparsity-regularized Rank Minimization: Algorithms and Applications | [
"cs.MA",
"cs.IT",
"cs.NI",
"cs.IT",
"stat.ML"
] | Given a limited number of entries from the superposition of a low-rank matrix plus the product of a known fat compression matrix times a sparse matrix, recovery of the low-rank and sparse components is a fundamental task subsuming compressed sensing, matrix completion, and principal components pursuit. This paper develops algorithms for distributed sparsity-regularized rank minimization over networks, when the nuclear- and $\ell_1$-norm are used as surrogates to the rank and nonzero entry counts of the sought matrices, respectively. While nuclear-norm minimization has well-documented merits when centralized processing is viable, non-separability of the singular-value sum challenges its distributed minimization. To overcome this limitation, an alternative characterization of the nuclear norm is adopted which leads to a separable, yet non-convex cost minimized via the alternating-direction method of multipliers. The novel distributed iterations entail reduced-complexity per-node tasks, and affordable message passing among single-hop neighbors. Interestingly, upon convergence the distributed (non-convex) estimator provably attains the global optimum of its centralized counterpart, regardless of initialization. Several application domains are outlined to highlight the generality and impact of the proposed framework. These include unveiling traffic anomalies in backbone networks, predicting networkwide path latencies, and mapping the RF ambiance using wireless cognitive radios. Simulations with synthetic and real network data corroborate the convergence of the novel distributed algorithm, and its centralized performance guarantees. | cs.MA | cs | In-network Sparsity-regularized Rank Minimization:
Algorithms and Applications†
Morteza Mardani, Gonzalo Mateos, and Georgios B. Giannakis (contact author)∗
Submitted: January 6, 2014
Abstract
Given a limited number of entries from the superposition of a low-rank matrix plus the product of a
known fat compression matrix times a sparse matrix, recovery of the low-rank and sparse components is
a fundamental task subsuming compressed sensing, matrix completion, and principal components pursuit.
This paper develops algorithms for distributed sparsity-regularized rank minimization over networks, when
the nuclear- and ℓ1 -norm are used as surrogates to the rank and nonzero entry counts of the sought
matrices, respectively. While nuclear-norm minimization has well-documented merits when centralized
processing is viable, non-separability of the singular-value sum challenges its distributed minimization.
To overcome this limitation, an alternative characterization of the nuclear norm is adopted which leads to
a separable, yet non-convex cost minimized via the alternating-direction method of multipliers. The novel
distributed iterations entail reduced-complexity per-node tasks, and affordable message passing among
single-hop neighbors. Interestingly, upon convergence the distributed (non-convex) estimator provably
attains the global optimum of its centralized counterpart, regardless of initialization. Several application
domains are outlined to highlight the generality and impact of the proposed framework. These include
unveiling traffic anomalies in backbone networks, predicti ng networkwide path latencies, and mapping the
RF ambiance using wireless cognitive radios. Simulations with synthetic and real network data corroborate
the convergence of the novel distributed algorithm, and its centralized performance guarantees.
Index Terms
Distributed optimization, sparsity, nuclear norm, low rank, networks, Lasso, matrix completion.
EDICS Category: SEN-DIST, SPC-APPL, SEN-COLB.
2
1
0
2
r
a
M
7
]
A
M
.
s
c
[
1
v
0
7
5
1
.
3
0
2
1
:
v
i
X
r
a
† Work in this paper was supported by the MURI Grant No. AFOSR FA9550-10-1-0567. Parts of the paper appeared in
the Proc. of the 45th Asilomar Conference on Signals, Systems, and Computers, Paci fic Grove, CA, Nov. 6-9, 2011, and were
submitted to the 13th Workshop on Signal Processing Advances in Wireless Communications, Cesme, Turkey, Jun. 17-20, 2012.
∗ The authors are with the Dept. of Electrical and Computer Engineering, University of Minnesota, 200 Union Street SE,
Minneapolis, MN 55455. Tel/fax: (612)626-7781/625-4583; Emails: {morteza,mate0058,georgios}@umn.edu
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
1
I . INTRODUCT ION
Let X := [xl,t ] ∈ RL×T be a low-rank matrix [rank(X) ≪ min(L, T )], and A := [af ,t ] ∈ RF ×T be a
sparse matrix with support size considerably smaller than F T . Consider also a matrix R := [rl,t ] ∈ RL×F
and a set Ω ⊆ {1, . . . , L} × {1, . . . , T } of index pairs (l, t) that de fine a sampling of the entries of X.
Given R and a number of (possibly) noise corrupted measurements
F
Xf =1
the goal is to estimate low-rank X and sparse A, by denoising the observed entries and imputing the
rl,f af ,t + vl,t ,
(1)
yl,t = xl,t +
(l, t) ∈ Ω
missing ones. Introducing the sampling operator PΩ (·) which sets the entries of its matrix argument not
in Ω to zero and leaves the rest unchanged, the data model can be compactly written in matrix form as
PΩ (Y) = PΩ (X + RA + V).
(2)
A natural estimator accounting for the low rank of X and the sparsity of A will be sought to fit the
data PΩ (Y) in the least-squares (LS) error sense, as well as minimize the rank of X, and the number of
nonzero entries of A measured by its ℓ0 -(pseudo) norm; see e.g. [9], [21], [8], [12] for related problems
subsumed by the one described here. Unfortunately, both rank and ℓ0 -norm minimization are in general
NP-hard problems [13], [24]. Typically, the nuclear norm kXk∗ := Pk σk (X) (σk (X) denotes the k-th
singular value of X) and the ℓ1 -norm kAk1 := Pf ,t af ,t are adopted as surrogates, since they are the
closest convex approximants to rank(X) and kAk0 , respectively [11], [25], [32]. Accordingly, one solves
1
kPΩ (Y − X − RA)k2
(P1)
F + λ∗kXk∗ + λ1kAk1
min
2
{X,A}
where λ∗ , λ1 ≥ 0 are rank- and sparsity-controlling parameters. Being convex (P1) is appealing, and some
of its special instances are known to attain good performance in theory and practice. For instance, when
no data are missing (P1) can be used to unveil traffic anomalie s in networks [21]. Results in [21] show
that X and A can be exactly recovered in the absence of noise, even when R is a fat (compression)
operator. When R equals the identity matrix, (P1) reduces to the so-termed robust principal component
analysis (PCA), for which exact recovery results are available in [8] and [12]. Moreover, for the special
case R ≡ 0L×F , (P1) offers a low-rank matrix completion alternative with well-documented merits; see
e.g., [10] and [9]. Stable recovery results in the presence of noise are also available for matrix completion
and robust PCA [9], [35]. Earlier efforts dealing with the recovery of sparse vectors in noise led to similar
performance guarantees; see e.g., [6].
In all these works, the samples PΩ (Y) and matrix R are assumed centrally available, so that they can
be jointly processed to estimate X and A by e.g., solving (P1). Collecting all this information can be
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
2
challenging though in various applications of interest, or may be even impossible in e.g., wireless sensor
networks (WSNs) operating under stringent power budget constraints. In other cases such as the Internet or
collaborative marketing studies, agents providing private data for e.g., fitting a low-rank preference model,
may not be willing to share their training data but only the learning results. Performing the optimization
in a centralized fashion raises robustness concerns as well, since the central processor represents an
isolated point of failure. Scalability is yet another driving force motivating distributed solutions. Several
customized iterative algorithms have been proposed to solve instances of (P1), and have been shown
effective in tackling low- to medium-size problems; see e.g., [21], [10], [25]. However, most algorithms
require computation of singular values per iteration and become prohibitively expensive when dealing
with high-dimensional data [26]. All in all, the aforementioned reasons motivate the reduced-complexity
distributed algorithm for nuclear and ℓ1 -norm minimization developed in this paper.
In a similar vein, stochastic gradient algorithms were recently developed for large-scale problems
entailing regularization with the nuclear norm [26]. Even though iterations in [26] are highly paralellizable,
they are not applicable to networks of arbitrary topology. There are also several studies on distributed
estimation of sparse signals via ℓ1 -norm regularized regression; see e.g., [14], [17], [22]. Different from
the treatment here, the data model of [22] is devoid of a low-rank component and all the observations Y
are assumed available (but distributed across several interconnected agents). Formally, the model therein
is a special case of (2) with T = 1, X = 0L×T , and Ω = {1, . . . , L} × {1, . . . , T }, in which case (P1)
boils down to finding the least-absolute shrinkage and selec tion operator (Lasso) [31].
Building on the general model (2) and the centralized estimator (P1), this paper develops decentralized
algorithms to estimate low-rank and sparse matrices, based on in-network processing of a small subset of
noise-corrupted and spatially-distributed measurements (Section III). This is a challenging task however,
since the non-separable nuclear-norm present in (P1) is not amenable to distributed minimization. To
overcome this limitation, an alternative characterization of the nuclear norm is adopted in Section III-A,
which leads to a separable yet non-convex cost that is minimized via the alternating-direction method
of multipliers (AD-MoM) [5]. The novel distributed iterations entail reduced-complexity optimization
subtasks per agent, and affordable message passing only between single-hop neighbors (Section III-C).
Interestingly, the distributed (non-convex) estimator upon convergence provably attains the global optimum
of its centralized counterpart (P1), regardless of initialization. To demonstrate the generality of the proposed
estimator and its algorithmic framework, four networking-related application domains are outlined in
Section IV, namely: i) unveiling traffic volume anomalies fo r large-scale networks [19], [21]; ii) robust
PCA [8], [12]; iii) low-rank matrix completion for networkwide path latency prediction [20], and iv)
spectrum sensing for cognitive radio (CR) networks [4], [22]. Numerical tests with synthetic and real
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
3
network data drawn from these application domains corroborate the effectiveness and convergence of the
novel distributed algorithms, as well as their centralized performance benchmarks (Section V).
Section VI concludes the paper, while several technical details are deferred to the Appendix.
Notation: Bold uppercase (lowercase) letters will denote matrices (column vectors), and calligraphic letters
will be used for sets. Operators (·)′ , tr(·), σmax (·), ⊙ and ⊗, will denote transposition, matrix trace,
maximum singular value, Hadamard product, and Kronecker product, respectively; · will be used for
the cardinality of a set, and the magnitude of a scalar. The matrix vectorization operator vec(Z) stacks
the columns of matrix Z on top of each other to return a supervector, and its inverse is unvec(z). The
diagonal matrix diag(v) has the entries of v on its diagonal, and the positive semide finite matrix M
will be denoted by M (cid:23) 0. The ℓp -norm of x ∈ Rn is kxkp := (Pn
i=1 xi p )1/p for p ≥ 1. For two
matrices M, U ∈ Rn×n , hM, Ui := tr(M′U) denotes their trace inner product. The Frobenious norm of
matrix M = [mi,j ] ∈ Rn×p is kMkF := ptr(MM′ ), kMk := maxkxk2=1 kMxk2 is the spectral norm,
kMk1 := Pi,j mi,j is the ℓ1 -norm, kMk∞ := maxi,j mi,j is the ℓ∞ -norm, and kMk∗ := Pi σi (M) is
the nuclear norm, where σi (M) denotes the i-th singular value of M. The n × n identity matrix will be
represented by In , while 0n will stand for n × 1 vector of all zeros, and 0n×p := 0n0′
p . Similar notations
will be adopted for vectors (matrices) of all ones.
I I . PREL IM INAR IE S AND PROBLEM STATEMENT
Consider N networked agents capable of performing some local computations, as well as exchanging
messages among directly connected neighbors. An agent should be understood as an abstract entity, e.g.,
a sensor in a WSN, a router monitoring Internet traffic; or a se nsing CR from a next-generation commu-
nications technology. The network is modeled as an undirected graph G(N , L), where the set of nodes
N := {1, . . . , N } corresponds to the network agents, and the edges (links) in L := {1, . . . , L} represent
pairs of agents that can communicate. Agent n ∈ N communicates with its single-hop neighboring peers
in Jn , and the size of the neighborhood will be henceforth denoted by Jn . To ensure that the data from
an arbitrary agent can eventually percolate through the entire network, it is assumed that:
(a1) Graph G is connected; i.e., there exists a (possibly) multi-hop path connecting any two agents.
With reference to the low-rank and sparse matrix recovery problem outlined in Section I, in the network
setting envisioned here each agent n ∈ N acquires a few incomplete and noise-corrupted rows of matrix Y .
Specifically, the local data available to agent n is matrix PΩn (Yn ), where Yn ∈ RLn×T , PN
n=1 Ln = L,
N ]′ = X + RA + V. The index pairs in Ωn are those in Ω for which the row
and Y := [Y ′
1 , . . . , Y ′
index matches the rows of Y observed by agent n. Additionally, suppose that agent n has available the
Ln×F , containing a row subset of R associated with the observed rows in Yn , i.e,
local matrix Rn ∈ R
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
4
N ]′ . Agents collaborate to form the wanted estimator (P1) in a distributed fashion, which
R := [R′
1 , . . . , R′
can be equivalently rewritten as
N
kAk1 (cid:21) .
Xn=1 (cid:20) 1
2
The objective of this paper is to develop a distributed algorithm for sparsity-regularized rank minimization
via (P1), based on in-network processing of the locally available data. The described setup naturally
kPΩn (Yn − Xn − RnA)k2
F +
min
{X,A}
kXk∗ +
λ∗
N
λ1
N
suggests three characteristics that the algorithm should exhibit: c1) agent n ∈ N should obtain an estimate
of Xn and A, which coincides with the corresponding solution of the centralized estimator (P1) that uses
the entire data PΩ (Y); c2) processing per agent should be kept as simple as possible; and c3) the overhead
for inter-agent communications should be affordable and con fined to single-hop neighborhoods.
I I I . D ISTR IBUTED ALGOR ITHM FOR IN -NETWORK O PERAT ION
To facilitate reducing the computational complexity and memory storage requirements of the distributed
algorithm sought, it is henceforth assumed that:
(a2)
An upper bound ρ ≥ rank( X) on the rank of matrix X obtained via (P1) is available.
As argued next, the smaller the value of ρ, the more efficient the algorithm becomes. A small value of ρ
is well motivated in various applications. For example, the Internet traffic analysis of backbone networks
in [19] demonstrates that origin-to-destination flows have a very low intrinsic dimensionality, which renders
the traffic matrix low rank; see also Section IV-A. In additio n, recall that rank( X) is controlled by the
choice of λ∗ in (P1), and the rank of the solution can be made small enough, for sufficiently large λ∗ .
Because rank( X) ≤ ρ, (P1)’s search space is effectively reduced and one can factorize the decision variable
as X = LQ′ , where L and Q are L × ρ and T × ρ matrices, respectively. Adopting this reparametrization
(P2)
of X in (P1), one obtains the following equivalent optimization problem
N
Xn=1 (cid:20) 1
kAk1 (cid:21)
2
N ]′ . The number of variables is
1 , . . . , L′
which is non-convex due to the bilinear terms LnQ′ , and L := [L′
reduced from LT + F T in (P1), to ρ(L + T ) + F T in (P2). The savings can be significant when ρ is in the
kPΩn (Yn − LnQ′ − RnA)k2
F +
kLQ′k∗ +
min
{L,Q,A}
λ∗
N
λ1
N
order of a few dozens, and both L and T are large. The dominant F T -term in the variable count of (P3)
is due to A, which is sparse and can be efficiently handled even when both F and T are large. Problem
(P3) is still not amenable to distributed implementation due to: (i) the non-separable nuclear norm present
in the cost function; and (ii) the global variables Q and A coupling the per-agent summands.
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
5
A. A separable nuclear norm regularization
To address (i), consider the following neat characterization of the nuclear norm [25], [26]
1
2 (cid:8)kLk2
F + kQk2
F (cid:9) ,
1/2
For an arbitrary matrix X with SVD X = UX ΣX V′
X , the minimum in (3) is attained for L = UX Σ
X
1/2
X . The optimization (3) is over all possible bilinear factorizations of X, so that the
and Q = VX Σ
number of columns of L and Q is also a variable. Leveraging (3), the following reformulation of (P2)
kXk∗ := min
{L,Q}
s. to X = LQ′ .
(3)
(P3)
provides an important first step towards obtaining a distrib uted estimator:
N
Xn=1 (cid:20) 1
kAk1 (cid:21) .
λ∗
F + kQk2
2N (cid:8)N kLnk2
F (cid:9) +
2
As asserted in the following lemma, adopting the separable Frobenius-norm regularization in (P3) comes
with no loss of optimality, provided the upper bound ρ is chosen large enough.
kPΩn (Yn − LnQ′ − RnA)k2
F +
min
{L,Q,A}
λ1
N
Lemma 1: Under (a2), (P3) is equivalent to (P1).
Proof: Let { X, A} denote the minimizer of (P1). Clearly, rank( X) ≤ ρ, implies that (P2) is equivalent
to (P1). From (3) one can also infer that for every feasible solution {L, Q, A}, the cost in (P3) is no
smaller than that of (P2). The gap between the globally minimum costs of (P2) and (P3) vanishes at
¯A := A, ¯L := U Σ1/2 , and ¯Q := V Σ1/2 , where X = U Σ V′ . Therefore, the cost functions of (P1) and
(P3) are identical at the minimum.
Lemma 1 ensures that by finding the global minimum of (P3) [whi ch could have significantly less
variables than (P1)], one can recover the optimal solution of (P1). However, since (P3) is non-convex, it
may have stationary points which need not be globally optimum. Interestingly, the next proposition shows
that under relatively mild assumptions on rank( X) and the noise variance, every stationary point of (P3)
is globally optimum for (P1). For a proof, see Appendix A.
Proposition 1: Let { ¯L, ¯Q, ¯A} be a stationary point of (P3). If kPΩ (Y − ¯L ¯Q′ − R ¯A)k ≤ λ∗ (no subscript
in k.k signifies spectral norm), then { X = ¯L ¯Q′ , A = ¯A} is the globally optimal solution of (P1).
Condition kPΩ (Y − ¯L ¯Q′ − R ¯A)k ≤ λ∗ captures tacitly the role of ρ, the number of columns of L and
Q in the postulated model. In particular, for sufficiently sma ll ρ the residual kPΩ (Y − ¯L ¯Q′ − R ¯A)k
becomes large and consequently the condition is violated (unless λ∗ is large enough, in which case a
sufficiently low-rank solution to (P1) is expected). This is manifested through the fact that the extra
condition rank( X) ≤ ρ in Lemma 1 is no longer needed. In addition, note that the noise variance certainly
affects the value of kPΩ (Y − ¯L ¯Q′ − R ¯A)k, and thus satisfaction of the said condition.
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
6
B. Local variables and consensus constraints
To decompose the cost function in (P3), in which summands are coupled through the global variables Q
and A [cf. (ii) at the beginning of this section], introduce auxiliary variables {Qn , An}N
n=1 representing
local estimates of {Q, A} per agent n. These local estimates are utilized to form the separable constrained
minimization problem
(P4)
min
{L,Qn ,An ,Bn}
kPΩn (Yn − LnQ′
n − RnBn )k2
F +
N
Xn=1 (cid:20) 1
2
s. to Bn = An , n ∈ N
λ∗
2N (cid:8)N kLnk2
F + kQnk2
F (cid:9) +
λ1
N
kAnk1(cid:21)
Qn = Qm , An = Am , m ∈ Jn , n ∈ N .
For reasons that will become clear later on, additional variables {Bn }N
n=1 were introduced to split the
ℓ1−norm regularization on the {An}N
n=1 (cf.
ℓ2−norm fitting-error part of the cost of (P4), from the
Remark 3). These extra variables are not needed if R′R = IF . The set of additional constraints Bn = An
ensures that, in this sense, nothing changes in going from (P3) to (P4). Most importantly, (P3) and (P4)
are equivalent optimization problems under (a1). The equivalence should be understood in the sense that
Q1 = Q2 = . . . = QN = Q and likewise for A, where { Qn , An}n∈N and { Q, A} are the optimal
solutions of (P4) and (P3), respectively. Of course, the corresponding estimates of L will coincide as well.
Even though consensus is a fortiori imposed within neighborhoods, it extends to the whole (connected)
network and local estimates agree on the global solution of (P3). To arrive at the desired distributed
algorithm, it is convenient to reparametrize the consensus constraints in (P4) as
n = Fm
n , and ¯Fm
n , Qm = Fm
Qn = ¯Fm
n , m ∈ Jn , n ∈ N
n = Gm
n , and ¯Gm
n , Am = Gm
An = ¯Gm
n , m ∈ Jn , n ∈ N
(4)
(5)
n }m∈Jn
where { ¯Fm
n , Fm
n , ¯Gm
n , Gm
n∈N are auxiliary optimization variables that will be eventually eliminated.
C. The alternating-direction method of multipliers
To tackle the constrained minimization problem (P4), associate Lagrange multipliers Mn with the
splitting constraints Bn = An , n ∈ N . Likewise, associate additional dual variables ¯Cm
n ( ¯Dm
n and Cm
n and
Dm
n ) with the first pair of consensus constraints in (4) [respect ively (5)]. Next introduce the quadratically
augmented Lagrangian function
N
Xn=1 (cid:20) 1
2
λ∗
2N (cid:8)N kLn k2
F + kQnk2
F (cid:9) +
kPΩn (Yn − LnQ′
n − RnBn )k2
F +
λ1
N
kAnk1(cid:21)
Lc (V1 , V2 , V3 , M) =
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
7
+
+
+
c
2
kBn − Ank2
F
N
N
Xn=1
Xn=1
hMn , Bn − An i +
N
Xn=1 Xm∈Jn nh ¯Cm
n io
n i + h Dm
n , An − ¯Gm
n i + h ¯Dm
n , Qm − Fm
n i + h Cm
n , Qn − ¯Fm
n , Am − Gm
N
c
F o
Xn=1 Xm∈Jn nkQn − ¯Fm
F + kAm − Gm
F + kAn − ¯Gm
F + kQm − Fm
n k2
n k2
n k2
n k2
(6)
2
where c is a positive penalty coefficient, and the primal variables a re split into three groups V1 :=
n=1 , and V3 := {Bn , ¯Fm
n , Fm
n , ¯Gm
n , Gm
n }m∈Jn
n=1 , V2 := {Ln}N
{Qn , An}N
n∈N . For notational convenience,
n }m∈Jn
n , ¯Dm
n , Cm
n , Dm
collect all multipliers in M := {Mn , ¯Cm
n∈N . Note that the remaining constraints in (4)
and (5), namely CV := { ¯Fm
n = Fm
n , ¯Gm
n = Gm
n , m ∈ Jn , n ∈ N }, have not been dualized.
To minimize (P4) in a distributed fashion, a variation of the alternating-direction method of multipliers
(AD-MoM) will be adopted here. The AD-MoM is an iterative augmented Lagrangian method especially
well-suited for parallel processing [5], which has been proven successful to tackle the optimization tasks
encountered e.g., with distributed estimation problems [22], [28]. The proposed solver entails an iterative
procedure comprising four steps per iteration k = 1, 2, . . .
[S1] Update dual variables:
Mn [k ] = Mn [k − 1] + µ(Bn [k ] − An [k ]), n ∈ N
n [k − 1] + µ(Qn [k ] − ¯Fm
n [k ] = ¯Cm
¯Cm
n [k ]), n ∈ N , m ∈ Jn
n [k − 1] + µ(Qm [k ] − Fm
n [k ] = Cm
Cm
n [k ]), n ∈ N , m ∈ Jn
n [k − 1] + µ(An [k ] − ¯Gm
n [k ] = ¯Dm
¯Dm
n [k ]), n ∈ N , m ∈ Jn
n [k − 1] + µ(Am [k ] − Gm
n [k ] = Dm
Dm
n [k ]), n ∈ N , m ∈ Jn
[S2] Update first group of primal variables:
V1 [k + 1] = arg min
V1
Lc (V1 , V2 [k ], V3 [k ], M[k ]) .
[S3] Update second group of primal variables:
V2 [k + 1] = arg min
V2
Lc (V1 [k + 1], V2 , V3 [k ], M[k ]) .
[S4] Update auxiliary primal variables:
V3 [k + 1] = arg min
V3∈CV
Lc (V1 [k + 1], V2 [k + 1], V3 , M[k ]) .
(7)
(8)
(9)
(10)
(11)
(12)
(13)
(14)
This four-step procedure implements a block-coordinate descent method with dual variable updates. At
each step while minimizing the augmented Lagrangian, the variables not being updated are treated as
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
8
Algorithm 1 : AD-MoM solver per agent n ∈ N
input Yn , Rn , λ∗ , λ1 , c, µ
initialize Mn [0] = Pn [0] = An [1] = Bn [1] = 0F ×T , O[0] = 0T ×ρ , and Ln [1], Qn [1] at random
for k = 1, 2,. . . do
Receive {Qm [k ], Am [k ]} from neighbors m ∈ Jn
[S1] Update local dual variables:
(Qn [k ] − Qm [k ])
Mn [k ] = Mn [k − 1] + µ(Bn [k ] − An [k ])
On [k ] = On [k − 1] + µ Pm∈Jn
Pn [k ] = Pn [k − 1] + µ Pm∈Jn
(An [k ] − Am [k ])
[S2] Update first group of local primal variables:
Qn [k + 1] = arg minQ (cid:26)rn (Ln [k ], Q, Bn [k ]) + λ∗
F + hOn [k ], Qi + c Pm∈Jn (cid:13)(cid:13)(cid:13)
Q − Qn [k]+Qm [k]
2N kQk2
2
An [k + 1] = [c(1 + 2Jn )]−1Sλ1 /N (cid:0)Mn [k ] + cBn [k ] − Pn [k ] + c Pm∈Jm
(An [k ] + Am [k ])(cid:1)
[S3] Update second group of local primal variables:
Ln [k + 1] = arg minL (cid:8)rn (L, Qn [k + 1], Bn [k ]) + λ∗
2 kLk2
F (cid:9)
[S4] Update auxiliary local primal variables:
Bn [k + 1] = arg minB (cid:8)rn (Ln [k + 1], Qn [k + 1], B) + hMn [k ], Bi + c
2 kB − An [k + 1]k2
F (cid:9)
Broadcast {Qn [k + 1], An [k + 1]} to neighbors m ∈ Jn
end for
return An , Qn , Ln
F (cid:27)
2
(cid:13)(cid:13)(cid:13)
fixed and are substituted with their most up to date values. Di fferent from AD-MoM, the alternating-
minimization step here generally cycles over three groups of primal variables V1 -V3 (cf. two groups in
AD-MoM [5]). In some special instances detailed in Sections IV-C and IV-D, cycling over two groups of
variables only is sufficient. In [S1], µ > 0 is the step size of the subgradient ascent iterations (7)-(11).
While it is common in AD-MoM implementations to select µ = c, a distinction between the step size and
the penalty parameter is made explicit here in the interest of generality.
Reformulating the estimator (P1) to its equivalent form (P4) renders the augmented Lagrangian in (6)
highly decomposable. The separability comes in two flavors, both with respect to the variable groups V1 ,
V2 , and V3 , as well as across the network agents n ∈ N . This in turn leads to highly parallelized, simplified
recursions corresponding to the aforementioned four steps. Specifically, it is shown in Appendix B that
if the multipliers are initialized to zero, [S1]-[S4] constitute the distributed algorithm tabulated under
Algorithm 1. For conciseness in presenting the algorithm, de fine the local residuals
rn (Ln , Qn , Bn ) :=
1
2 kPΩn (Yn − LnQ′
n − RnBn )k2
F . In addition, de fine the soft-thresholding matrix Sτ (M) with (i, j )-th
entry given by sign(mi,j ) max{mi,j − τ , 0}, where mi,j denotes the (i, j )-th entry of M.
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
9
Remark 1 (Simplification of redundant variables): Careful inspection of Algorithm 1 reveals that the
inherently redundant auxiliary variables and multipliers { ¯Fm
n , Dm
n , Cm
n , Gm
n , ¯Gm
n , Fm
n } have been elimi-
nated. Agent n does not need to separately keep track of all its non-redundant multipliers { ¯Cm
n , ¯Dm
n }m∈Jn ,
¯Dm
¯Cm
n [k ].
n [k ] and Pn [k ] := 2 Pm∈Jn
but only to update their respective (scaled) sums On [k ] := 2 Pm∈Jn
Remark 2 (Computational and communication cost): The main computational burden of the algorithm
stems from solving unconstrained quadratic programs locally to update Qn , Ln , Bn , and to carry out simple
soft-thresholding operations to update An . On a per iteration basis, network agents communicate their
updated local estimates {Qn [k ], An [k ]} with their neighbors, to carry out the updates of the primal and
dual variables during the next iteration. Regarding communication cost, Qn [k ] is a T × ρ matrix and its
transmission does not incur significant overhead when ρ is small. In addition, the F × T matrix An [k ] is
sparse, and can be communicated efficiently. Observe that th e dual variables need not be exchanged.
Remark 3 (General sparsity-promoting regularization): Even though λ1kAk1 was adopted in (P1) to
encourage sparsity in the entries of A, the algorithmic framework here can accommodate more general
structured sparsity-promoting penalties ψ(A). To maintain the per-agent computational complexity at
affordable levels, the minimum requirement on the admissible penalties is that the proximal operator
A (cid:20) 1
F + ψ(A)(cid:21)
proxψ ( Y) := arg min
k Y − Ak2
2
is given in terms of vector or (and) scalar soft-thresholding operators. In addition to the ℓ1 -norm (Lasso
penalty), this holds for the sum of row-wise ℓ2 -norms (group Lasso penalty [33]), or, a linear combination
of the aforementioned two – the so-termed hierarchical Lass o penalty that encourages sparsity across and
(15)
within the rows of A [29]. All this is possible since by introducing the cost-splitting variables Bn , the
local sparse matrix updates are An [k + 1] = proxψ ( Yn [k ]) for suitable Yn [k ] (see Appendix B).
When employed to solve non-convex problems such as (P4), AD-MoM (or its variant used here) offers
no convergence guarantees. However, there is ample experimental evidence in the literature which supports
convergence of AD-MoM, especially when the non-convex problem at hand exhibits “favorable ” structure.
For instance, (P4) is bi-convex and gives rise to the strictly convex optimization subproblems (12)-(14),
which admit unique closed-form solutions per iteration. This observation and the linearity of the constraints
endow Algorithm 1 with good convergence properties – extens ive numerical tests including those presented
in Section V demonstrate that this is indeed the case. While a formal convergence proof goes beyond
the scope of this paper, the following proposition proved in Appendix C asserts that upon convergence,
Algorithm 1 attains consensus and global optimality.
Proposition 2: If the sequence of iterates {Qn [k ], Ln [k ], An [k ]}n∈N generated by Algorithm 1 converge
to { ¯Qn , ¯Ln , ¯An}n∈N , and (a1) holds, then: i) ¯Qn = ¯Qm , ¯An = ¯Am , n, m ∈ N ; and ii) if kPΩ (Y −
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
10
1 − R ¯A1 )k ≤ λ∗ , then X = ¯L ¯Q′
¯L ¯Q′
1 and A = ¯A1 , where { A, X} is the global optimum of (P1).
IV. A P PL ICAT ION S
This section outlines a few applications that could bene fit f rom the distributed sparsity-regularized rank
minimization framework described so far. In each case, the problem statement calls for estimating low-
rank X and (or) sparse A, given distributed data adhering to an application-dependent model subsumed
by (2). Customized algorithms are thus obtained as special cases of the general iterations in Algorithm 1.
A. Unveiling traffic anomalies in backbone networks
In the backbone of large-scale networks, origin-to-destination (OD) traffic flows experience abrupt
changes which can result in congestion, and limit the quality of service provisioning of the end users.
These so-termed traffic volume anomalies could be due to external sources such as network failures, denial
of service attacks, or, intruders which hijack the network services [30] [19]. Unveiling such anomalies is
a crucial task towards engineering network traffic. This is a
challenging task however, since the available
data are usually high-dimensional noisy link-load measurements, which comprise the superposition of
unobservable OD flows as explained next.
The network is modeled as in Section II, and transports a set of end-to-end flows F (with F := F )
associated with specific source-destinations pairs. For ba ckbone networks, the number of network layer
flows is typically much larger than the number of physical lin ks (F ≫ L). Single-path routing is considered
here to send the traffic flow from a source to its intended desti
nation. Accordingly, for a particular flow
multiple links connecting the corresponding source-destination pair are chosen to carry the traffic. Sparing
details that can be found in [21], the traffic Y := [yl,t ] ∈ RL×T carried over links l ∈ L and measured at
time instants t ∈ [1, T ] can be compactly expressed as
Y = R (Z + A) + V
(16)
where the fat routing matrix R := [rℓ,f ] ∈ {0, 1}L×F is fixed and given, Z := [zf ,t ] denotes the unknown
“clean ” traffic flows over the time horizon of interest, and
A := [af ,t ] collects the traffic volume anomalies.
These data are distributed. Agent n acquires a few rows of Y corresponding to the link-load traffic
measurements Yn ∈ RLn×T from its outgoing links, and has available its local routing table Rn which
indicates the OD flows routed through n. Assuming a suitable ordering of links, the per agent quantities
N ]′ .
N ]′ and R := [R′
1 , . . . , R′
1 , . . . , Y ′
relate to their global counterparts in (16) through Y := [Y ′
Common temporal patterns among the traffic flows in addition t
o their periodic behavior, render most
rows (respectively columns) of Z linearly dependent, and thus Z typically has low rank [19]. Anomalies
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
11
are expected to occur sporadically over time, and only last for short periods relative to the (possibly long)
measurement interval [1, T ]. In addition, only a small fraction of the flows are supposed t o be anomalous
at any given time instant. This renders the anomaly matrix A sparse across rows and columns. Given
local measurements {Yn}n∈N and the routing tables {Rn }n∈N , the goal is to estimate A in a distributed
fashion, by capitalizing on the sparsity of A and the low-rank property of Z. Since the primary goal is
to recover A, de fine X := RZ which inherits the low-rank property from Z, and consider [cf. (16)]
Y = X + RA + V.
(17)
Model (17) is a special case of (2), when all the entries of Y are observed, i.e., Ω = {1, . . . , L} ×
{1, . . . , T }. Note that RA is not sparse even though A is itself sparse, hence principal components pursuit
is not applicable here [35]. Instead, the following estimator is adopted to unveil network anomalies [21]
N
Xn=1 (cid:20) 1
kAk1 (cid:21)
2
which is subsumed by (P1). Accordingly, a distributed algorithm can be readily obtained by simplifying
the general iterations under Algorithm 1, the subject dealt with next.
{ X, A} = arg min
{X,A}
λ∗
N
kYn − Xn − RnAk2
F +
kXk∗ +
λ1
N
Distributed Algorithm for Unveiling Network Anomalies (DUNA). For the specific case here in which
Ω = {1, . . . , L} × {1, . . . , T }, the residuals in Algorithm 1 reduce to rn (Ln , Qn , Bn ) := 1
2 kYn −
LnQ′
n − RnBnk2
F . Accordingly, to update the primal variables Qn [k + 1], Ln [k + 1] and Bn [k + 1]
as per Algorithm 1, one needs to solve respective unconstrained strictly convex quadratic optimization
problems. These admit closed-form solutions detailed under Algorithm 2. The DUNA updates of the local
anomaly matrices An [k + 1] are given in terms of soft-thresholding operations, exactly as in Algorithm 1.
Conceivably, the number of flows F can be quite large, thus inverting the F × F matrix R′
nRn + cIF
to update Bn [k + 1] could be complex computationally. Fortunately, the inversion needs to be carried
out once, and can be performed and cached off-line. In addition, to reduce the inversion cost, the
SVD of the local routing matrices Rn = URn ΣRn V′
can be obtained first, and the matrix inversion
Rn
lemma can be subsequently employed to obtain [R′
nRn + cIF ]−1 = (1/c) (cid:2)Ip − VRn CV′
Rn (cid:3), where
p (cid:17) and p = rank(Rn ) ≪ F . This computational shortcut is commonly adopted
C := diag (cid:16) σ2
σ2
p
1
c+σ2
c+σ2
1
in statistical learning algorithms when ridge regression estimates are sought, and the number of variables
is much larger than the number of elements in the training set [15, Ch. 18]. During the operational phase
, ...,
of the algorithm, the main computational burden of DUNA comes from repeated inversions of (small)
ρ × ρ matrices, and parallel soft-thresholding operations. The communication overhead is identical to the
one incurred by Algorithm 1 (cf. Remark 2).
Remark 4 (Incomplete link traffic measurements):
In general, one can allow for missing traffic data
and the DUNA updates are still expressible in closed form.
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
12
Algorithm 2 : DUNA per agent n ∈ N
input Yn , Rn , λ∗ , λ1 , c, µ
initialize Mn [0] = Pn [0] = An [1] = Bn [1] = 0F ×T , O[0] = 0T ×ρ , and Ln [1], Qn [1] at random
for k = 1, 2,. . . do
Receive {Qm [k ], Am [k ]} from neighbors m ∈ Jn
[S1] Update local dual variables:
(Qn [k ] − Qm [k ])
Mn [k ] = Mn [k − 1] + µ(Bn [k ] − An [k ])
On [k ] = On [k − 1] + µ Pm∈Jn
Pn [k ] = Pn [k − 1] + µ Pm∈Jn
(An [k ] − Am [k ])
[S2] Update first group of local primal variables:
n [k ]Ln [k ] + (λ∗ /N + 2cJn )Iρ ]−1
Qn [k+1] = (cid:8)Y′
(Qn [k ] + Qm [k ])(cid:9) [L′
n [k ]R′
nLn [k ] − B′
nLn [k ] − On [k ] + c Pm∈Jn
An [k + 1] = [c(1 + 2Jn )]−1Sλ1 /N (cid:0)Mn [k ] + cBn [k ] − Pn [k ] + c Pm∈Jm
(An [k ] + Am [k ])(cid:1)
[S3] Update second group of local primal variables:
n [k + 1]Qn [k + 1] + λ∗ Iρ ]−1
Ln [k + 1] = (Yn − RnBn [k ]) Qn [k + 1] [Q′
[S4] Update auxiliary local primal variables:
nRn + cIF ]−1 {R′
Bn [k + 1] = [R′
n(Yn − Ln [k + 1]Q′
n [k + 1]) − Mn [k ] + cAn [k + 1]}
Broadcast {Qn [k + 1], An [k + 1]} to neighbors m ∈ Jn
end for
return An , Qn , Ln
B. In-network robust principal component analysis
Principal component analysis (PCA) is the workhorse of high-dimensional data analysis and dimension-
ality reduction, with numerous applications in statistics, networking, engineering, and the biobehavioral
sciences; see, e.g., [18]. Nowadays ubiquitous e-commerce sites, complex networks such as the Web, and
urban traffic surveillance systems generate massive volume s of data. As a result, extracting the most in-
formative, yet low-dimensional structure from high-dimensional datasets is of paramount importance [15].
Data obeying postulated low-rank models include also outliers, which are samples not adhering to those
nominal models. Unfortunately, similar to LS estimates PCA is very sensitive to the outliers [18]. While
robust approaches to PCA are available, recently polynomial-time algorithms with remarkable performance
guarantees have emerged for low-rank matrix recovery in the presence of sparse – but otherwise arbitrarily
large – errors [8], [12], [35]. Robust PCA is of great interes t in networking-related applications. One
can think of distributed estimation using reduced-dimensionality sensor observations [27], and unveiling
anomalous flows in backbone networks from Netflow data [2]; se
e also Section V-B.
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
13
In the network setting of Section II, each agent n ∈ N acquires Fn outlier-plus-noise corrupted rows
N ]′ , where PN
1 , . . . , Y ′
of matrix Y := [Y ′
n=1 Fn = F . Local data can thus be modeled as Yn = Xn +
1 , . . . , X′
N ]′ has low rank. Agents want to estimate Xn (and the outliers An )
An + Vn , where X := [X′
in a distributed fashion by forming the global estimator [35]
N
Xn=1 (cid:20) 1
2
which is once more a special case of (P1) when R = IF .
Distributed Robust Principal Component Analysis (DRPCA) Algorithm. Regarding the general dis-
kXk∗ + λ1 kAnk1(cid:21)
{ X, A} = arg min
{X,A}
kYn − Xn − Ank2
F +
λ∗
N
(18)
tributed formulation in (P4), the first constraint is no long er needed since R = IF [cf. the discussion after
(P4)]. As agent n is interested in estimating An and kAk1 is separable over the rows of A, the only
required constraints are Qn = Qm , m ∈ Jn , n ∈ N . These are associated with the dual variables On
per agent, and are updated according to Algorithm 3. All in all, each agent stores and recursively updates
the primal variables {Qn , Ln}, along with the Fn × T matrix An .
Mimicking the procedure that led to Algorithm 1, one finds tha t primal variable updates in DRPCA are
expressible in closed form. In particular, the local outlier matrix An [k + 1] minimizes the Lasso cost
{An } (cid:26) 1
F + λ1 kAnk1(cid:27)
kYn − Ln [k + 1]Q′
n [k + 1] − Ank2
An [k + 1] = arg min
2
and is given in terms of soft-thresholding operations as seen in Algorithm 3 [observe that An [k + 1] =
proxk·k1 (Yn − Ln [k + 1]Q′
n [k + 1]), where proxψ (·) is de fined in (15)]. DRPCA iterations are simple
with small ρ × ρ matrices inverted per iteration to update Ln and Qn (see Algorithm 3). Regarding
communication cost, each agent only broadcasts a T × ρ matrix Qn to its neighbors.
C. Distributed low-rank matrix completion
The ability to recover a low-rank matrix from a subset of its entries is the leitmotif of recent advances
for localization of wireless sensors [23], Internet traffic
analysis [20], [34], and preference modeling for
recommender systems [3]. In the low-rank matrix completion problem, given a limited number of (possibly)
noise corrupted entries of a low-rank matrix X, the goal is to recover the entire matrix while denoising
the observed entries, and accurately imputing the missing ones.
In the network setting envisioned here, agent n ∈ N has available Ln incomplete and noise-corrupted
1 , . . . , Y ′
N ]′ . Local data can thus be modeled as PΩn (Yn ) = PΩn (Xn + Vn ). Relying
rows of Y := [Y ′
on in-network processing, agents aim at completing their own rows by forming the global estimator
N
Xn=1 (cid:20) 1
kXk∗ (cid:21)
2
kPΩn (Yn − Xn )k2
F +
X = arg min
X
λ∗
N
(19)
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
14
Algorithm 3 : DRPCA algorithm per agent n ∈ N
input Yn , λ∗ , λ1 , c, µ
initialize An [1] = 0Fn×T , O[0] = 0T ×ρ , and Ln [1], Qn [1] at random.
for k = 1, 2,. . . do
Receive {Qm [k ]} from neighbors m ∈ Jn
[S1] Update local dual variables:
On [k ] = On [k − 1] + µ Pm∈Jn
(Qn [k ] − Qm [k ])
[S2] Update first group of local primal variables:
n [k ]Ln [k ] + (λ∗ /N + 2cJn )Iρ ]−1
Qn [k+1] = (cid:8)Y′
(Qn [k ] + Qm [k ])(cid:9) [L′
nLn [k ] − A′
n [k ]Ln [k ] − On [k ] + c Pm∈Jn
[S2] Update second group of local primal variables:
n [k + 1]Qn [k + 1] + λ∗ Iρ ]−1
Ln [k + 1] = (Yn − An [k ]) Qn [k + 1] [Q′
[S3] Update third group of local primal variables:
An [k + 1] = Sλ1 (Yn − Ln [k + 1]Q′
n [k + 1])
Broadcast {Qn [k + 1]} to neighbors m ∈ Jn
end for
return An , Qn , Ln
which exploits the low-rank property of X through nuclear-norm regularization. Estimator (19) was
proposed in [9], and solved centrally whereby all data PΩn (Yn ) is available to feed an e.g., off-the-
shelf semide finite programming (SDP) solver. The general es timator in (P1) reduces to (19) upon setting
R = 0L×F and λ1 = 0. Hence, it is possible to derive a distributed algorithm for low-rank matrix
completion by specializing Algorithm 1 to the setting here.
Before dwelling into the algorithmic details, a brief parenthesis is in order to touch upon properties
of local sampling operators. Operator PΩn is a linear orthogonal projector, since it projects its matrix
argument onto the subspace Ψn := {Z ∈ RLn×T : supp(Z) ∈ Ωn} of matrices with support contained
in Ωn . Linearity of PΩn implies that vec(PΩn (Z)) = AΩn vec(Z), where AΩn ∈ RLn×T is a symmetric
and idempotent projection matrix that will prove handy later on. To characterize AΩn , introduce an
Ln × T masking matrix Ωn whose (l, t)-th entry equals one when (l, t) ∈ Ωn , and zero otherwise. Since
PΩn (Z) = Ωn ⊙ Z, from standard properties of the vec(·) operator it follows that AΩn = diag(vec(Ωn)).
Distributed Matrix Completion (DMC) Algorithm. Going back to the general distributed formulation
in (P4), since there is no sparse component A in the matrix completion problem (19), the only constraints
that remain are Qn = Qm , m ∈ Jn , n ∈ N . These correspond to the dual variables On [k ] per agent, and
are updated as shown in Algorithm 4.
In the absence of {An }n∈N and the auxiliary variables {Bn }n∈N , it suffices to cycle over two groups
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
15
of primal variables to arrive at the DMC iterations. The primal variable updates can be readily obtained by
λ∗
2N
kQk2
F
Qn [k ] + Qm [k ]
2
capitalizing on the properties of the vec(·) operator. In particular, Algorithm 1 indicates that the recursions
for Qn are given by [let q := vec(Q′ )]
Q (cid:26) 1
kPΩn (Yn − Ln [k ]Q′ )k2
Qn [k + 1] = arg min
F +
2
+hOn [k ], Qi + c Xm∈Jn (cid:13)(cid:13)(cid:13)(cid:13)
Q −
= unvec(cid:18) arg min
q (cid:26) 1
kAΩn vec(Yn ) − AΩn (I ⊗ Ln [k ])qk2 +
2
2) (cid:19).
′ [k ]), qi + c Xm∈Jn (cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)
+hvec(On
q −
Likewise, Ln is updated by solving the following subproblem per iteration (let l := vec(L))
F (cid:27)
L (cid:26) 1
λ∗
kPΩn (Yn − LQ′
kLk2
n [k + 1])k2
F +
Ln [k + 1] = arg min
2
2
= unvec (cid:18)arg min
l (cid:26) 1
klk2(cid:27)(cid:19) .
λ∗
kAΩn vec(Yn ) − AΩn (Qn [k + 1] ⊗ ILn )lk2 +
2
2
Both (20) and (21) are unconstrained convex quadratic problems, which admit the closed-form solutions
(cid:13)(cid:13)(cid:13)(cid:13)
λ∗
2N
′ [k ] + Qm
2
F )
2
kqk2
vec(Qn
′ [k ])
(20)
(21)
tabulated under Algorithm 4.
The per-agent computational complexity of the DMC algorithm is dominated by repeated inversions of
ρ × ρ and ρLn × ρLn matrices to obtain En [k + 1] and Dn [k + 1], respectively (see Algorithm 4). Notice
that En [k + 1] ∈ RρT ×ρT has block-diagonal structure with blocks of size ρ× ρ. Inversion of ρ× ρ matrices
is affordable in practice since ρ is typically small for a number of applications of interest (cf. the low-rank
assumption). In addition, Ln is the number of row vectors acquired per agent which can be controlled by
the designer to accommodate a prescribed maximum computational complexity. On a per iteration basis,
network agents communicate their updated local estimates Qn [k ] only with their neighbors, in order to
carry out the updates of primal and dual variables during the next iteration. In terms of communication
cost, Qn [k ] is a T × ρ matrix and its transmission does not incur significant overh ead for small values
of ρ. Observe that the dual variables On [k ] need not be exchanged, and the overall communication cost
does not depend on the network size N .
D. Distributed sparse linear regression for spectrum cartography
In a classical linear regression setting, training data {Y , R} related through Y = RA + V are given
and the goal is to estimate the regression coefficient matrix A based on the LS criterion1 . However, LS
1Even though vector responses and model parameters are typically considered, i.e., y = Ra + v, the matrix notation is retained
here for consistency.
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
16
Algorithm 4 : DMC algorithm per agent n ∈ N
Input Yn , AΩn , λ∗ , c, µ
Initialize O[0] = 0T ×ρ , and Ln [1], Qn [1] at random
for k = 1, 2,. . . do
Receive {Qm [k ]} from neighbors m ∈ Jn
[S1] Update local dual variables:
On [k ] = On [k − 1] + µ Pm∈Jn
(Qn [k ] − Qm [k ])
[S2] Update first group of local primal variables:
n [k ])AΩn (IT ⊗ Ln [k ]) + (λ∗ /N + 2cJn )IρT }−1
En [k + 1] = {(IT ⊗ L′
n [k ]) + cvec(Pm∈Jn
n [k ])AΩn vec(Yn ) − vec(O′
n [k+1] = unvec (cid:0)En [k + 1] (cid:8)(IT ⊗ L′
Q′
[S3] Update second group of local primal variables:
n [k + 1] ⊗ ILn )AΩn (Qn [k + 1] ⊗ ILn ) + λ∗ IρLn }−1
Dn [k + 1] = {(Q′
n [k + 1] ⊗ ILn )AΩn vec(Yn ))
Ln [k + 1] = unvec (Dn [k + 1] (Q′
Broadcast {Qn [k + 1]} to neighbors m ∈ Jn
end for
Return Qn , Ln
m [k ]))(cid:9)(cid:1)
(Q′
n [k ] + Q′
often yields unsatisfactory prediction accuracy and fails to provide parsimonious estimates; see e.g., [15].
To deal with such limitations, one can adopt a regularization technique known as Lasso [31].
The general framework in this paper can accommodate distributed estimators of the regression coeffi-
cients via Lasso, when the training data are scattered across different agents, and their communication to
a central processing unit is prohibited for e.g., communication cost or privacy reasons. With reference to
the setup described in Section II, each agent has available the training data {Yn , Rn } and wishes to find
N
Xn=1 (cid:20) 1
kAk1(cid:21)
2
in a distributed fashion. The Lasso estimator (22) is a special case of (P1) in the absence of a low-rank
component; that is when X = 0L×T , and Ω = {1, . . . , L} × {1, . . . , T }.
An application domain where this problem arises is spectrum sensing in cognitive radio (CR) networks
ALasso := arg min
A
(22)
kYn − RnAk2
F +
λ1
N
whereby sensing CRs collaborate to estimate the radio-frequency power spectrum density maps Φ(x, f )
across space x ∈ R2 and frequency f ; see [4], [22]. These maps enable identification of opportun isti-
cally available spectrum bands for re-use and handoff operation; as well as localization, transmit-power
estimation, and tracking of primary user activities.
A cooperative approach to spectrum cartography is introduced in [4], based on a basis expansion model
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
17
of Φ(x, f ). Spatially-distributed CRs collect smoothed periodogram samples Yn of the received signal at
given sampling frequencies, based on which they want to determine the unknown expansion coefficients in
A. The sensing scheme capitalizes on two forms of sparsity: the first one introduced by the narrow-band
nature of transmit-PSDs relative to the broad swaths of usable spectrum; and the second one emerging
from sparsely located active radios in the operational space. All in all, locating the active transmitters boils
down to a variable selection problem, which motivates well employment of the Lasso in (22). Since data
are collected by cooperating CRs at different locations, estimation of A amounts to solving a distributed
parameter estimation problem which demands taking into account the network topology, and devising a
protocol to share the data. All these are accomplished by the algorithm outlined next.
Distributed Lasso (DLasso) Algorithm [22]. In the Lasso setting here, (P4) reduces to
N
Xn=1 (cid:20) 1
kAn k1(cid:21)
2
s. to Bn = An , n ∈ N
kYn − RnBn k2
F +
min
{An ,Bn}
λ1
N
An = Am , m ∈ Jn , n ∈ N
which is a convex optimization problem equivalent to (22). Accordingly, the DLasso algorithm (tabulated
under Algorithm 5) is readily obtained from Algorithm 1, after retaining only the update recursions for
{An , Bn } and the multipliers {Mn , Pn }. A distributed protocol to select λ1 via K -fold cross valida-
tion [15, Ch. 7] is also available in [22], along with numerical tests to assess its performance.
Since Lasso in (22) and its separable constrained reformulation are convex optimization problems, the
general convergence results for AD-MoM iterations can be invoked to establish convergence of DLasso
as well. A detailed proof of the following proposition can be found in [22, App. E].
Proposition 3: Under (a1) and for any value of the penalty coefficient c > 0, the iterates An [k ] converge
to the Lasso solution [cf. (22)] as k → ∞, i.e., limk→∞ An [k ] = ALasso , ∀ n ∈ N .
V. NUMER ICAL TE ST S
This section corroborates convergence and gauges performance of the proposed algorithms, when tested
on the applications of Section IV using synthetic and real network data.
Synthetic network data. A network of N = 20 agents is considered as a realization of the random
geometric graph model, that is, agents are randomly placed on the unit square and two agents communicate
with each other if their Euclidean distance is less than a prescribed communication range of 0.35; see Fig. 1.
The network graph is bidirectional and comprises L = 106 links, and F = N (N − 1) = 380 OD flows.
The entries of V are independent and identically distributed (i.i.d.), zero-mean, Gaussian with variance
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
18
Algorithm 5 : DLasso per agent n ∈ N
input Yn , Rn , λ1 , c, µ
initialize Mn [0] = Pn [0] = An [1] = Bn [1] = 0F ×T
for k = 1, 2,. . . do
Receive {Am [k ]} from neighbors m ∈ Jn
[S1] Update local primal variables:
(An [k ] − Am [k ])
Mn [k ] = Mn [k − 1] + c(Bn [k ] − An [k ])
Pn [k ] = Pn [k − 1] + c Pm∈Jn
[S2] Update primal variables:
An [k + 1] = [c(1 + 2Jn )]−1Sλ1 /N (cid:0)Mn [k ] + cBn [k ] − Pn [k ] + c Pm∈Jm
[S3] Update auxiliary local primal variables:
nRn + cIF ]−1 {R′
Bn [k + 1] = [R′
nYn − Mn [k ] + cAn [k + 1]}
Broadcast An [k + 1] to neighbors m ∈ Jn
end for
return An
(An [k ] + Am [k ])(cid:1)
σ2 ; i.e., vl,t ∼ N (0, σ2 ). Low-rank matrices with rank r are generated from the bilinear factorization
model X0 = WZ′ , where W and Z are L × r and T × r matrices with i.i.d. entries drawn from Gaussian
distributions N (0, 100/F ) and N (0, 100/T ), respectively. Every entry of A0 is randomly drawn from the
set {−1, 0, 1} with Pr(ai,j = −1) = Pr(ai,j = 1) = π/2. Unless otherwise stated, r = 3, ρ = 3 and
T = F = 380 are used throughout. Different values of σ , and π are examined.
Internet2 network data. Real data including OD flow traffic levels and end-to-end late
ncies are collected
from the operation of the Internet2 network (Internet backbone network across USA) [1]. Both versions
of the Internet2 network, referred as v1 and v2, are considered. OD flow traffic levels are recorded for a
three-week operation of Internet2-v1 during Dec. 8 –28, 200 8 [19], and are used to assess performance of
DUNA and DRPCA (see Sections V-A and V-B next). Internet2-v1 contains N = 11 agents, L = 41 links,
and F = 121 flows. To test the DMC algorithm, end-to-end flow latencies ar
e collected from the operation
of Internet2-v2 during Aug. 18 –22, 2011 [1]. The Internet2- v2 network comprises N = 9 agents, L = 26
links, and F = 81 flows.
Selection of tuning parameters. The sparsity- and rank-controlling parameters λ1 and λ∗ are tuned to
optimize performance. The optimality conditions for (P1) indicate that for λ1 > kR′Yk∞ and λ∗ > kYk,
{X0 = 0L×T , A0 = 0F ×T } is the unique optimal solution. This in turn con fines the sear ch space for
λ1 and λ∗ to the intervals (0, kR′Yk∞ ] and (0, kYk], respectively. In addition, for the case of matrix
completion and robust PCA one can use the heuristic rules proposed in e.g., [9] and [8].
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
19
A. Unveiling network anomalies
Data is generated from Y = R(X0 +A0)+V, where the routing matrix R is obtained after determining
shortest-path routes of the OD flows. For µ = c = 0.1, DUNA is run until convergence is attained. These
values were experimentally chosen to obtain the fastest convergence rate. The time evolution of consensus
among agents is depicted in Fig. 2 (left), for representative agents in the network. The metric of interest
here is the relative error kQn [k ] − ¯Q[k ]kF /k ¯Q[k ]kF per agent n, which compares the corresponding local
N PN
estimate with the network-wide average ¯Q[k ] := 1
n=1 Qn [k ]; and likewise for the An [k ]. Fig. 2 (left)
shows that DUNA converges and agents consent on the global matrices {Q, A} as k → ∞.
To corroborate that DUNA attains the centralized performance, the accelerated proximal gradient al-
gorithm of [21] is employed to solve (P1) after collecting all the per-agent data in a central processing
unit. For both the distributed and centralized schemes, Fig. 2 (right) depicts the evolution of the relative
error k A[k ] − A0 kF /kA0 kF for various sparsity levels, where A[k ] := ¯A[k ] for DUNA. It is apparent
that the distributed estimator approaches the performance of its centralized counterpart, thus corroborating
convergence and global optimality as per Proposition 2.
Unveiling Internet2-v1 network anomalies from SNMP measurements. Given the OD flow traffic mea-
surements discussed at the beginning of Section V, the link loads in Y are obtained through multiplication
with the Internet2-v1 routing matrix [1]. Even though Y is “constructed ” here from flow measurements,
link loads can be typically acquired from simple network management protocol (SNMP) traces [30]. The
available OD flows are a superposition of “clean ” and anomalo
us traffic, i.e., the sum of unknown “ground-
truth ” low-rank and a sparse matrices X0 + A0 adhering to (16) when R = IF . Therefore, the proposed
algorithms are applied first to obtain a reasonably precise e stimate of the “ground-truth ”
{X0 , A0 }. The
estimated X0 exhibits three dominant singular values, con firming the low -rank property of X0 .
The receiver operation characteristic (ROC) curves in Fig. 3 (left) highlight the merits of DUNA when
used to identify Internet2-v1 network anomalies. Even at low false alarm rates of e.g., PFA = 0.04, the
anomalies are accurately detected (PD = 0.93). In addition, DUNA consistently outperforms the landmark
PCA-based method of [19], and can identify anomalous flows. F ig. 3 (right) illustrates the magnitude of
the true and estimated anomalies across flows and time.
B. Robust PCA
Next, the convergence and effectiveness of the proposed DRPCA algorithm is corroborated with the aid
of computer simulations. An F × T data matrix is generated as Y = X0 + A0 + V, and the centralized
estimator (18) is obtained using the AD-MoM method proposed in [8]. In the network setting, each agent
has available Ln = 19 rows of Y . Fig. 2 (right) is replicated as Fig. 4 (left) for the robust PCA problem
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
20
dealt with here, and for different values of ρ [the assumed upper bound on rank(X0 )]. It is again apparent
that DRPCA converges and approaches the performance of (18) as k → ∞.
Unveiling Internet2-v1 network anomalies from Netflow meas urements. Suppose a router n ∈ N
monitors the traffic volume of OD flows to unveil anomalies usi
ng e.g., the Netflow protocol [2], [30].
Collect the time-series of all OD flows as the rows of the F × T matrix Y = X0 + A0 + V, where A0 and
V account for anomalies and noise, respectively. As elaborated in Section IV-A, the common temporal
patterns across flows renders the traffic matrix X0 low-rank. Owing to the difficulties of measuring the
large number of OD flows, in practice only a few entries of Y are typically available [30], or, link traffic
measurements are utilized as in Section V-A (recall that L ≪ F ). In this example, because the Internet2-v1
network data comprises only F = 121 flows, it is assumed that Ω = {1, ..., F } × {1, ..., T }.
To better assess performance, large spikes are injected into 1% randomly selected entries of the ground
truth-traffic matrix X0 estimated in Section V-A. The DRPCA algorithm is run on this Internet2-v1 Netflow
data to identify the anomalies. The results are depicted in Fig. 4 (right). DRPCA accurately identifies the
anomalies, achieving PD = 0.98 when PFA = 10−3 .
C. Low-rank matrix completion
In addition to the synthetic data specifications outlined at
the beginning of this section, the sampling
set Ω is picked uniformly at random, where each entry of the matrix Ω is a Bernoulli random variable
taking the value one with probability p. Data for the matrix completion problem is thus generated as
PΩ (Y) = PΩ (X0 + V) = Ω ⊙ (X0 + V), where Y is an L × T matrix with L = T = 106. The data
available to agent n is PΩn (Yn ), which corresponds to a row subset of PΩ (Y).
As with the previous test cases, it is shown first that the DMC a lgorithm converges to the (centralized)
solution of (19). To this end, the centralized singular value thresholding algorithm is used to solve (19) [10],
when all data PΩ (Y) is available for processing. For both the distributed and centralized schemes, Fig. 5
(left) depicts the evolution of the relative error k X[k ] − X0 kF /kX0 kF for different values of σ (noise
strength), and percentage of missing entries (controlled by p). Regardless of the values of σ and p, the
error trends clearly show the convergent behavior of the DMC algorithm and corroborate Proposition 2.
Interestingly, for small noise levels where the estimation error approaches zero, the distributed estimator
recovers X0 almost exactly.
Internet2-v2 network latency prediction. End-to-end network latency information is critical towards
enforcing quality-of-service constraints in many Internet applications. However, probing all pairwise delays
becomes infeasible in large-scale networks. If one collects the end-to-end latencies of source-sink pairs
(i, j ) in a delay matrix X := [xi,j ] ∈ RN ×N , strong dependencies among path delays render X low-
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
21
rank [20]. This is mainly because the paths with nearby end nodes often overlap and share common
bottleneck links. This property of X along with the distributed-processing requirements of large-scale
networks, motivates well the adoption of the DMC algorithm for networkwide path latency prediction.
Given the n-th row of X is partially available to agent n, the goal is to impute the missing delays through
agent collaboration.
The DMC algorithm is tested here using the real path latency data collected from the operation of
Internet2-v2. Spectral analysis of X0 reveals that the first four singular values are markedly domi nant,
demonstrating that X0 is low rank. A fraction of the entries in X0 are purposely dropped to yield an
incomplete delay matrix PΩ (X0 ). After running the DMC algorithm till convergence, the true and predicted
latencies are depicted in Fig. 5 (right) (for 20% missing data). The relative prediction error is around 10%.
V I . CONCLUD ING SUMMARY
A framework for distributed sparsity-regularized rank minimization is developed in this paper, that
is suitable for (un)supervised inference tasks in networks. Fundamental problems such as in-network
compressed sensing, matrix completion, and principal components pursuit are all captured under the same
umbrella. The novel distributed algorithms can be utilized to unveil traffic volume anomalies from SNMP
and Netflow traces, to predict networkwide path latencies in large-scale IP networks, and to map-out the
RF ambiance using periodogram samples collected by spatially-distributed cognitive radios.
A P PEND IX
A. Proof of Proposition 1. Recall the cost function of (P3) de fined as
1
λ∗
F + kQk2
2 (cid:0)kLk2
F (cid:1) + λ1kAk1 .
2
The stationary points { ¯L, ¯Q, ¯A} of (P3) are obtained by setting to zero the (sub)gradients, and solving [7]
kPΩ (Y − LQ′ − RA)k2
F +
f (L, Q, A) :=
(23)
∂Af ( ¯L, ¯Q, ¯A) = R′PΩ (Y − ¯L ¯Q′ − R ¯A) − λ1 sign( ¯A) = 0F ×T
∇Lf ( ¯L, ¯Q, ¯A) = PΩ (Y − ¯L ¯Q′ − R ¯A) ¯Q − λ∗ ¯L = 0L×ρ
∇Q′ f ( ¯L, ¯Q, ¯A) = ¯L′PΩ (Y − ¯L ¯Q′ − R ¯A) − λ∗ ¯Q′ = 0ρ×T .
(24)
(25)
(26)
Clearly, every stationary point satis fies ∇Lf ( ¯L, ¯Q, ¯A) ¯L′ = 0L×L and ¯Q∇Q′ f ( ¯L, ¯Q, ¯A) = 0T ×T . Using
the aforementioned identities, the optimality conditions (24)-(26) can be rewritten as
R′PΩ (Y − ¯L ¯Q′ − R ¯A) = λ1 sign( ¯A)
tr(PΩ (Y − ¯L ¯Q′ − R ¯A) ¯Q ¯L′ ) = λ∗ tr{ ¯Q ¯Q′} = λ∗ tr{ ¯L ¯L′}.
(27)
(28)
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
22
(P5)
Consider now the following convex optimization problem
{X,A,W1 ,W2} (cid:20) 1
kPΩ (Y − X − RA)k2
F +
min
2
s. to W :=
W2
which is equivalent to (P1). The equivalence can be readily inferred by minimizing (P5) with respect to
{W1 , W2} first, and taking advantage of the following alternative cha racterization of the nuclear norm
(see e.g., [25])
{tr{W1 } + tr{W2 }} + λ1kAk1(cid:21)
W1 X
λ∗
2
(29)
(cid:23) 0
′
X
s. to
W2
In what follows, the optimality conditions for the conic program (P5) are explored. To this end, the
kXk∗ = min
{W1 ,W2 }
{tr(W1 ) + tr(W2 )} ,
W1 X
(cid:23) 0.
1
2
′
X
Lagrangian is first formed as
L(X, W1 , W2 , A, M) =
1
λ∗
2
2
where M denotes the dual variables associated with the conic constraint (29). For notational convenience,
{tr(W1 ) + tr(W2 )} − hM, Wi + kAk1
kPΩ (Y − X − RA)k2
F +
partition M in four blocks M1 := [M]11 , M2 := [M]12 , M3 := [M]22 , and M4 := [M]21 , in accordance
with the block structure of W in (29), where M1 and M3 are L × L and T × T matrices. The optimal
solution to (P5) must: (i) null the (sub)gradients
∇XL(X, W1 , W2 , A, M) = −PΩ (Y − X − RA) − M2 − M4
′
(30)
∂AL(X, W1 , W2 , A, M) = −R′PΩ (Y − X − RA) − λ1 sign(A)
λ∗
2
λ∗
2
and satisfy (ii) the complementary slackness condition hM, Wi = 0; (iii) primal feasibility W (cid:23) 0; and
∇W1 L(X, W1 , W2 , A, M) =
∇W2 L(X, W1 , W2 , A, M) =
IT − M3
IL − M1
(31)
(33)
(32)
(iv) dual feasibility M (cid:23) 0.
Recall the stationary point of (P3), and introduce candidate primal variables X := ¯L ¯Q′ , A := ¯A,
2 IL , M3 := λ∗
2 IT , M2 := −(1/2)PΩ (Y −
W1 := ¯L ¯L′ and W2 := ¯Q ¯Q′ ; and the dual variables M1 := λ∗
¯L ¯Q′ − R ¯A), and M4 := M′
2 . Then, (i) holds since after plugging the candidate primal and dual variables
in (30)-(33), the subgradients vanish. Moreover, (ii) holds since
2 , X′ i + h M3 , W2 i
h M, Wi = h M1 , W1 i + h M2 , Xi + h M′
λ∗
λ∗
2
2
λ∗
F − λ∗k ¯Lk2
k ¯Qk2
F = 0
2
hIL , ¯L ¯L′ i +
k ¯Lk2
F +
λ∗
2
=
hLT , ¯Q ¯Q′ i − hPΩ (Y − ¯L ¯Q′ − R ¯A), ¯L ¯Q′ i
=
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
23
where the last two equalities follow from (28). Condition (iii) is also met since
′
=
¯Q ¯Q′
¯Q
¯Q
¯L ¯Q′
¯L
¯L
To satisfy (iv), based on a Schur complement argument [16] it suffices to enforce σmax ( M2 ) ≤ λ∗ . (cid:4)
B. Derivation of Algorithm 1. It is shown here that [S1]-[S4] in Section III-C give rise to the set of
¯L ¯L′
¯Q ¯L′
(cid:23) 0.
(34)
recursions tabulated under Algorithm 1. To this end, recall the augmented Lagrangian function in (6) and
focus first on [S4]. From the decomposable structure of Lc , (14) decouples into simpler strictly convex
sub-problems for n ∈ N and m ∈ Jn , namely
c
Bn nrn (Ln [k + 1], Qn [k + 1], Bn ) + hMn [k ], Bn i +
F o
kBn − An [k + 1]k2
(35)
Bn [k + 1] = arg min
2
n n c
n io
2 (cid:0)kQn [k + 1] − ¯Fm
F + kQm [k + 1] − ¯Fm
F (cid:1) − h ¯Cm
n [k ], ¯Fm
n [k ] + Cm
¯Fm
n k2
n k2
n [k + 1] = arg min
¯Fm
(36)
n n c
n io .
F + kAm [k + 1] − ¯Gm
2 (cid:0)kAn [k + 1] − ¯Gm
¯Gm
n [k ] + Dm
F (cid:1) − h ¯Dm
n [k ], ¯Gm
n k2
n k2
n [k + 1] = arg min
¯Gm
(37)
n and Gm
Note that in formulating (36) and (37), the auxiliary variables Fm
n were eliminated using the
constraints ¯Fm
n = Fm
n and ¯Gm
n = Gm
n , respectively. The unconstrained quadratic problems (36) and (37)
admit the closed-form solutions
n [k ] + Dm
( ¯Dm
n [k ]) +
(An [k + 1] + Am [k + 1]) .
(Qn [k + 1] + Qm [k + 1])
n [k + 1] = Fm
¯Fm
n [k + 1] =
n [k ] + Cm
( ¯Cm
n [k ]) +
n [k + 1] = Gm
¯Gm
n [k + 1] =
1
1
2c
2
1
1
2c
2
n [k ] and Fm
Using (38) to eliminate ¯Fm
n [k ] from (8) and (9) respectively, a simple induction argument
n [k ] = − Cm
n [0] = 0T ×ρ , then ¯Cm
n [0] = − Cm
establishes that if the initial Lagrange multipliers obey ¯Cm
n [k ]
n [k ] and Dm
for all k ≥ 0, where n ∈ N and m ∈ Jn . Likewise, the same holds true for ¯Dm
n [k ]. The
n [k ], Dm
collection of multipliers { Cm
n [k ]}m∈Jn
n∈N is thus redundant, and (38)-(39) simplify to
1
2
1
2
n [k ] = ¯Gn
Observe that ¯Fm
n [k ] = ¯Fn
m [k ] and ¯Gm
m [k ] for all k ≥ 0, identities that will be used later on. By
plugging (40) and (41) into (8) and (10) respectively, the non-redundant multiplier updates become
(An [k + 1] + Am [k + 1]) , n ∈ N , m ∈ Jn .
(Qn [k + 1] + Qm [k + 1]) , n ∈ N , m ∈ Jn
n [k + 1] = Gm
¯Gm
n [k + 1] =
n [k + 1] = Fm
¯Fm
n [k + 1] =
(38)
(39)
(40)
(41)
n [k ] = ¯Cm
¯Cm
n [k − 1] +
n [k ] = ¯Dm
¯Dm
n [k − 1] +
µ
2
µ
2
(Qn [k ] − Qm [k ]) , n ∈ N , m ∈ Jn
(An [k ] − Am [k ]) , n ∈ N , m ∈ Jn .
(42)
(43)
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
24
If ¯Cm
n [0] = − ¯Cn
m [0] = 0T ×ρ , then the structure of (42) reveals that ¯Cm
n [k ] = − ¯Cn
m [k ] for all k ≥ 0,
where n ∈ N and m ∈ Jn . Clearly, the same holds true for ¯Dm
n [k ], and these identities will become
handy in the sequel.
Moving on to [S3], (13) decouples into N unconstrained quadratic sub-problems
Ln (cid:26)rn (Ln , Qn [k + 1], Bn [k ]) +
F (cid:27) .
kLnk2
Ln [k + 1] = arg min
The minimization (12) in [S2] also decomposes into simpler sub-problems, both across agents and across
λ∗
2
+
λ∗
2N
λ∗
2N
the variables {Qn}n∈N and {An }n∈N , which are decoupled in the augmented Lagrangian when all other
variables are fixed. Specifically, the per agent updates of Qn are given by
Qn (rn (Ln [k ], Qn , Bn [k ]) +
F + Xm∈Jn
h ¯Cm
n [k ] + Cn
kQnk2
m [k ], Qn i
Qn [k + 1] = arg min
F (cid:17))
c
2 Xm∈Jn (cid:16)kQn − ¯Fm
F + kQn − Fn
m [k ]k2
n [k ]k2
Qn (cid:26)rn (Ln [k ], Qn , Bn [k ]) +
kQnk2
F + hOn [k ], Qn i
= arg min
F )
2
+c Xm∈Jn (cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)
Qn [k ] + Qm [k ]
Qn −
2
n [k ] = Cn
where the second equality was obtained after using: i) ¯Cm
m [k ] which follows from the identities
¯Cm
n [k ] = − ¯Cn
n [k ] and ¯Cm
n [k ] = − Cm
¯Cm
n [k ];
m [k ] established earlier; ii) the de finition On(k) := 2 Pm∈Jn
n [k ] = Fn
and iii) the identity ¯Fm
m [k ], which allows to merge the identical quadratic penalty terms and
eliminate both ¯Fm
n [k ] and Fn
m [k ] using (40).
¯Dm
n [k ] and following similar steps as the ones that led to the second
Upon de fining Pn (k) := 2 Pm∈Jn
equality in (44), one arrives at
An (cid:26) λ1
kAnk1 − hMn [k ], An i +
An [k + 1] = arg min
N
+c Xm∈Jn (cid:13)(cid:13)(cid:13)(cid:13)
An ( N
2 (cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)
An −
= arg min
where the second equality follows by completing the squares. Problem (45) is a separable instance of the
Lasso (also related to the proximal operator of the ℓ1 -norm); hence, its solution is expressible in terms of
the soft-thresholding operator as in Algorithm 1.
kBn [k ] − Ank2
F + hPn [k ], An i
F )
2
(cid:13)(cid:13)(cid:13)(cid:13)
An [k ] + Am [k ]
2
Mn [k ] + cBn [k ] − Pn [k ] + c Pm∈Jn
c(1 + 2Jn )
(An [k ] + Am [k ])
c
2
An −
(44)
2
F
+ λ1 kAnk1)
(45)
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
25
C. Proof of Proposition 2. Let ¯Qn := limk→∞ Qn [k ], and likewise for all other convergent sequences in
[ ¯Qn −
Algorithm 1. Examination of the recursion for On [k ] in the limit as k → ∞, reveals that Pm∈Jn
¯Qm ] = 0T ×ρ , ∀ n ∈ N . Upon vectorizing the matrix quantities involved, this system of equations implies
that the supervector ¯q := [vec[ ¯Q1 ]′ , . . . , vec[ ¯QN ]′ ]′ belongs to the nullspace of L ⊗ IT ρ , where L is the
Laplacian of the network graph G(N , L). Under (a1), this guarantees that ¯Q1 = ¯Q2 = . . . = ¯QN . From
the analysis of the limiting behavior of Pn [k ], the same argument leads to ¯A1 = ¯A2 = . . . = ¯AN , which
establishes the consensus results in the statement of Proposition 2. Hence, one can go ahead and de fine
¯Q := ¯Qn and ¯A := ¯An . Before moving on, note that convergence of Mn [k ] implies that ¯Bn = ¯An = ¯A,
n ∈ N . These observations guarantee that the limiting solution is feasible for (P4).
To prove the optimality claim it suffices to show that upon con vergence, the fixed point { ¯L, ¯Q, ¯A, ¯B} of
the iterations comprising Algorithm 1 satis fies the Karush- Kuhn-Tucker (KKT) optimality conditions for
(P4). Proposition 1 asserts that if kPΩ (Y − LQ′ − RA)k ≤ λ∗ , { ¯L, ¯Q, ¯A} is indeed an optimal solution
to (P1). To this end, consider the updates of the primal variables in Algorithm 1, which satisfy
λ∗
∇Qn rn (Qn [k + 1], Ln [k ], Bn [k ]) +
Qn [k + 1] + On [k + 1]
N
+ 2c Xm∈Jn (cid:18)Qn [k + 1] −
∇Ln rn (Qn [k + 1], Ln [k + 1], Bn [k ]) + λ∗Ln [k + 1] = 0L×ρ
Qn [k ] + Qm [k ]
2
(cid:19) = 0T ×ρ
(47)
(46)
∇Bn rn (Qn [k + 1], Ln [k + 1], Bn [k + 1]) + Mn [k ] + c(Bn [k + 1] − An [k + 1]) = 0F ×T .
Taking the limit from both sides of (46) –(48), and summing up over all n ∈ N yields
∇Qr( ¯Q, ¯L, ¯A) + λ∗ ¯Q = 0T ×ρ
(48)
(49)
(50)
(51)
¯Mn = 0F ×T
∇Lr( ¯Q, ¯L, ¯A) + λ∗ ¯L = 0L×ρ
∇Br( ¯Q, ¯L, ¯A) + Xn∈N
F . To arrive at (49), the assumption that ¯Cm
where r(L, Q, B) := 1
2 kPΩ (Y − LQ′ − RB)k2
n [1] = 0, ∀m ∈
¯Cm
Jn , n ∈ N is used, and thus ¯Cm
n [k ] = − ¯Cn
n [k ] = 0.
m [k ] which leads to Pn∈N On [k ] = Pn∈N Pm∈Jn
Next, consider the auxiliary matrices Θn := ¯Mn − ¯Pn + c(1 + 2Jn ) ¯A, n ∈ N . In the limit as k → ∞,
the update recursion for An [k + 1] in Algorithm 1 can be written as c(1 + 2Jn ) ¯A = S (Θn , λ1 /N ) .
Proceed by de fining Ψn := Θn − c(1 + 2Jn ) ¯A, and observe that the input-output relationship of the
soft-thresholding operator S yields
[Ψn ]f ,t =
−λ1/N ,
f ,t : ξ (n)
ξ (n)
f ,t ≤ λ1 /N ,
[ ¯A]f ,t > 0,
[ ¯A]f ,t < 0,
[ ¯A]f ,t = 0.
λ1 /N ,
(52)
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
26
n=1 Ψn(cid:17) and Γ2 := 1
n=1 Ψn(cid:17),
2 (cid:16)λ11F 1′
2 (cid:16)λ11F 1′
T − PN
T + PN
Given (52), de fine the matrices Γ1 := 1
and show that they satisfy the following properties: (i) Γ1 , Γ2 ≥ 0 (entrywise); (ii) [Γ1 ]f ,t = 0, if
¯Mn .
[ ¯A]f ,t < 0; (iii) [Γ2 ]f ,t = 0, if [ ¯A]f ,t > 0; (iv) Γ1 + Γ2 = λ11F 1′
T ; and (v) Γ1 − Γ2 = Pn∈N
Properties (i)-(iii) follow after adding up the result in (52) for n = 1, 2, . . . , N . Property (iv) is readily
checked from the de finitions of Γ1 and Γ2 . Finally, (v) follows since
N
N
N
N
N
Xn=1 (cid:0)Θn − c(1 + 2Jn ) ¯A(cid:1) =
Xn=1
Xn=1
Xn=1
Xn=1
Γ1 − Γ2 =
Ψn =
¯Pn = 0 (from the identity PN
where PN
n=1 Pn [k ] = 0) is used to obtain the last equality.
n=1
The proof is concluded by noticing that properties (i)-(v) along with (49)-(51) comprise the KKT
conditions for the following optimization problem
¯Mn −
¯Pn =
¯Mn
min
{L,Q,A,T}
1
2
kPΩ (Y − LQ′ − RA)k2
F +
s. to − T ≤ A ≤ T (entrywise)
λ∗
F (cid:9) + λ11′
2 (cid:8)kLk2
F + kQk2
F T1T
where { ¯L, ¯Q, ¯A} and {Γ1 , Γ2 } play the role of the optimal primal and dual variables, respectively. This
last problem is clearly equivalent to (P4).
(cid:4)
RE FERENCE S
[1] [Online]. Available: http://internet2.edu/observatory/archive/data- collections.html
[2] A. Abdelkefi, Y. Jiang, W. Wang, A. Aslebo, and O. Kvittem,
“Robust traffic anomaly detection with principal component
pursuit,” in Proc. of ACM CoNEXT Student Wkshp., Philadelphia, USA, Nov. 2010.
[3] G. Adomavicius and A. Tuzhilin, “Toward the next generat ion of recommender systems: A survey of the state-of-the-art
and possible extensions,”
IEEE Trans. Knowledge and Data Engineering, vol. 17, no. 6, pp. 734–749, Jun. 2005.
[4] J. A. Bazerque and G. B. Giannakis, “Distributed spectru m sensing for cognitive radio networks by exploiting sparsity,”
IEEE Trans. Signal Process., vol. 58, pp. 1847–1862, Mar. 2010.
[5] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computation: Numerical Methods. Athena-Scienti fic, 1999.
[6] P. J. Bickel, Y. Ritov, and A. Tsybakov, “Simultaneous an alysis of Lasso and Dantzig selector,” Ann. Statist., vol. 37, pp.
1705–1732, Apr. 2009.
[7] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press, 2004.
[8] E. J. Candes, X. Li, Y. Ma, and J. Wright, “Robust principa l component analysis?” Journal of the ACM, vol. 58, no. 1, pp.
1–37, May 2011.
[9] E. J. Candes and Y. Plan, “Matrix completion with noise,”
Proceedings of the IEEE, vol. 98, pp. 925–936, Jun. 2010.
[10] E. J. Candes and B. Recht, “Exact matrix completion via c onvex optimization,” Found. Comput. Math., vol. 9, no. 6, pp.
717–722, Dec. 2009.
[11] E. J. Candes and T. Tao, “Decoding by linear programming ,”
IEEE Trans. Info. Theory, vol. 51, no. 12, pp. 4203–4215,
Dec. 2005.
[12] V. Chandrasekaran, S. Sanghavi, P. R. Parrilo, and A. S. Willsky, “Rank-sparsity incoherence for matrix decomposi tion,”
SIAM J. Optim., vol. 21, no. 2, pp. 572–596, 2011.
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
27
[13] A. Chistov and D. Grigorev, “Complexity of quanti fier el
imination in the theory of algebraically closed fields,” in Math.
Found. of Computer Science, ser. Lecture Notes in Computer Science. Springer Berlin / Heidelberg, 1984, vol. 176, pp.
17–31.
[14] S. Chouvardas, K. Slavakis, Y. Kopsinis, and S. Theodoridis, “A sparsity-aware adaptive algorithm for distribute d learning,”
IEEE Trans. Signal Process., 2011, see also arXiv:1112.5716v1 [cs.IT].
[15] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning, 2nd ed. Springer, 2009.
[16] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge University Press, 1985.
[17] D. Jakovetic, J. Xavier, and J. Moura, “Cooperative con vex optimization in networked systems: Augmented Lagrangian
algorithms with directed gossip communication,”
IEEE Trans. Signal Process., vol. 59, pp. 3889 –3902, Aug. 2011.
[18] I. T. Jolliffe, Principal Component Analysis. New York: Springer, 2002.
[19] A. Lakhina, M. Crovella, and C. Diot, “Diagnosing netwo rk-wide traffic anomalies,” in
Proc. of ACM SIGCOMM, Portland,
USA, Aug. 2004.
[20] Y. Liao, W. Du, P. Geurts, and G. Leduc, “DMFSGD: A decent ralized matrix factorization algorithm for network distance
prediction,”
IEEE/ACM Trans. Network., 2011, see also arXiv:1201.1174v1 [cs.NI].
[21] M. Mardani, G. Mateos, and G. B. Giannakis, “Unveiling n etwork anomalies across flows and time via sparsity and low
rank,”
IEEE Trans. Info. Theory, Feb. 2012 (submitted).
[22] G. Mateos, J. A. Bazerque, and G. B. Giannakis, “Distrib uted sparse linear regression,”
IEEE Trans. Signal Process., vol. 58,
pp. 5262–5276, Oct. 2010.
[23] A. Montanari and S. Oh, “On positioning via distributed matrix completion,” in Sensor Array and Multichannel Signal
Processing Workshop, Jerusalem, Oct. 2010, pp. 197–200.
[24] B. K. Natarajan, “Sparse approximate solutions to line ar systems,” SIAM J. Comput., vol. 24, pp. 227–234, 1995.
[25] B. Recht, M. Fazel, and P. A. Parrilo, “Guaranteed minim um-rank solutions of linear matrix equations via nuclear norm
minimization,” SIAM Rev., vol. 52, no. 3, pp. 471–501, 2010.
[26] B. Recht and C. Re, “Parallel stochastic gradient algor ithms for large-scale matrix completion,” 2011 (submitted ).
[27] I. D. Schizas, G. B. Giannakis, and Z. Q. Luo, “Distribut ed estimation using reduced-dimensionality sensor observations,”
IEEE Trans. Signal Process., vol. 55, pp. 4284–4299, Aug. 2007.
[28] I. D. Schizas, A. Ribeiro, and G. B. Giannakis, “Consens us in ad hoc WSNs with noisy links - part I: Distributed estimation
of deterministic signals,”
IEEE Trans. Signal Process., vol. 56, pp. 350–364, Jan. 2008.
[29] P. Sprechmann, I. Ramirez, G. Sapiro, and Y. Eldar, “C-H iLasso: A collaborative hierarchical sparse modeling framework,”
IEEE Trans. Signal Process., vol. 59, no. 9, pp. 4183–4198, Sep. 2011.
[30] M. Thottan and C. Ji, “Anomaly detection in IP networks, ”
IEEE Trans. Signal Process., vol. 51, pp. 2191–2204, 2003.
[31] R. Tibshirani, “Regression shrinkage and selection vi a the Lasso,” J. Royal. Statist. Soc B, vol. 58, pp. 267–288, 1996.
[32] J. Tropp, “Just relax: Convex programming methods for i dentifying sparse signals,”
IEEE Trans. Info. Theory, vol. 51, pp.
1030–1051, Mar. 2006.
[33] M. Yuan and Y. Lin, “Model selection and estimation in re gression with grouped variables,” J. Royal. Statist. Soc B, vol. 68,
pp. 49–67, 2006.
[34] Y. Zhang, M. Roughan, W. Willinger, and L. Qiu, “Spatio- temporal compressive sensing and internet traffic matrices ,” in
Proc. of ACM SIGCOM Conf. on Data Commun., New York, USA, Oct. 2009.
[35] Z. Zhou, X. Li, J. Wright, E. Candes, and Y. Ma, “Stable pr incipal component pursuit,” in Proc. of Intl. Symp. on Info.
Theory, Austin, USA, Jun. 2010, pp. 1518–1522.
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
28
s
i
x
a
−
y
1
0.8
0.6
0.4
0.2
0
Fig. 1. A network of N = 20 agents.
0
0.2
0.4
x−axis
0.6
0.8
1
102
100
10−2
10−4
r
o
r
r
e
s
u
s
n
e
s
n
o
c
e
v
i
t
a
l
e
R
10−6
0
Q A
1000
2000
3000
5000
4000
Iteration index
6000
7000
8000
101
100
10−1
r
o
r
r
e
n
o
i
t
a
m
i
t
s
e
e
v
i
t
a
l
e
R
10−2
0
Distributed, π=0.01
Centralized, π=0.01
Distributed, π=0.002
Centralized, π=0.002
2000
4000
6000
Iteration index
8000
10000
Fig. 2. Performance of DUNA. (left) Relative consensus error for representative network agents with σ = 0.01 and π = 0.01.
(right) Relative estimation error for distributed and centralized algorithms under various sparsity levels.
1
0.8
0.6
0.4
0.2
y
t
i
l
i
b
a
b
o
r
p
n
o
i
t
c
e
t
e
D
0
0
1
PCA−based method, rank=1
PCA−based method, rank=2
PCA−based method, rank=3
PCA−based method, rank=4
Proposed method, per time and flow
0.2
0.6
0.4
False alarm probability
0.8
x 108
True
Estimated
4
3
2
e
d
u
t
i
l
p
m
a
y
l
a
m
o
n
0
A
120
1
110
Flow index (f)
100
50
150
100
Time index (t)
Fig. 3. Unveiling anomalies from Internet2-v1 SNMP data. (left) ROC curves of the proposed versus the PCA-based method.
(right) Amplitude of the true and estimated anomalies for ρ = 5, PFA = 0.04 and PD = 0.93.
IEEE TRANSACTIONS ON SIGNAL PROCESSING (SUBMITTED)
29
101
100
10−1
r
o
r
r
e
n
o
i
t
a
m
i
t
s
e
e
v
i
t
a
l
e
R
10−2
0
Centralized
Distributed, ρ=3
Distributed, ρ=5
Distributed, ρ=10
True
Estimated
x 108
4
3
2
1
e
d
u
t
i
l
p
m
a
y
l
a
m
o
n
A
0
50
50
100
150
Iteration index
200
250
300
Flow index (f)
40
30
20
10
20
30
10
40
50
Time index (t)
Fig. 4.
Performance of DRPCA. (left) Relative estimation error for distributed and centralized algorithms under different ρ.
(right) Amplitude of true and estimated anomalies using Internet2-v1 network data when ρ = 5, PFA = 10−3 and PD = 0.98.
100
10−1
10−2
r
o
r
r
e
n
o
i
t
a
m
i
t
s
e
e
v
i
t
a
l
e
R
Distributed, σ=0.01, p=0.1
Centralized, σ=0.01, p=0.1
Distributed, σ=0.1, p=0.1
Centralized, σ=0.1, p=0.1
Distributed, σ=0.1, p=0.3
Centralized, σ=0.1, p=0.3
y
a
l
e
D
40
30
20
10
0
True
Estimated
10−3
0
100
200
300
400
600
500
Iteration index
700
800
900
1000
8
6
4
Node index
2
2
4
8
6
Node index
Fig. 5.
Performance of DMC. (left) Relative estimation error for distributed and centralized algorithms under various noise
strengths and percentage of missing entries. (right) Predicted and true ene-to-end delays of Internet2-v2 network for p = 0.2.
|
1301.3347 | 1 | 1301 | 2013-01-15T14:00:55 | Multi-agent learning using Fictitious Play and Extended Kalman Filter | [
"cs.MA",
"cs.LG",
"math.OC",
"stat.ML"
] | Decentralised optimisation tasks are important components of multi-agent systems. These tasks can be interpreted as n-player potential games: therefore game-theoretic learning algorithms can be used to solve decentralised optimisation tasks. Fictitious play is the canonical example of these algorithms. Nevertheless fictitious play implicitly assumes that players have stationary strategies. We present a novel variant of fictitious play where players predict their opponents' strategies using Extended Kalman filters and use their predictions to update their strategies.
We show that in 2 by 2 games with at least one pure Nash equilibrium and in potential games where players have two available actions, the proposed algorithm converges to the pure Nash equilibrium. The performance of the proposed algorithm was empirically tested, in two strategic form games and an ad-hoc sensor network surveillance problem. The proposed algorithm performs better than the classic fictitious play algorithm in these games and therefore improves the performance of game-theoretical learning in decentralised optimisation. | cs.MA | cs |
Multi-agent learning using Fictitious Play and
Extended Kalman Filter
Michalis Smyrnakis∗1
1Complex Systems and Statistical Physics Group, School of
Physics and Astronomy, University of Manchester, UK
August 21, 2018
Abstract
Decentralised optimisation tasks are important components of multi-
agent systems. These tasks can be interpreted as n-player potential games:
therefore game-theoretic learning algorithms can be used to solve decen-
tralised optimisation tasks. Fictitious play is the canonical example of
these algorithms. Nevertheless fictitious play implicitly assumes that play-
ers have stationary strategies. We present a novel variant of fictitious play
where players predict their opponents' strategies using Extended Kalman
filters and use their predictions to update their strategies.
We show that in 2 by 2 games with at least one pure Nash equilibrium
and in potential games where players have two available actions, the pro-
posed algorithm converges to the pure Nash equilibrium. The performance
of the proposed algorithm was empirically tested, in two strategic form
games and an ad-hoc sensor network surveillance problem. The proposed
algorithm performs better than the classic fictitious play algorithm in
these games and therefore improves the performance of game-theoretical
learning in decentralised optimisation.
Keywords: Multi-agent learning, game theory, fictitious play, decen-
tralised optimisation, learning in games, Extended Kalman filter.
1
Introduction
Recent advance in technology render decentralised optimisation a crucial com-
ponent of many applications of multi agent systems and decentralised control.
Sensor networks (Kho et al., 2009), traffic control (van Leeuwen et al., 2002)
and scheduling problems (Stranjak et al., 2008) are some of the tasks where
decentralised optimisation can be used. These tasks share common character-
istics such as large scale, high computational complexity and communication
constraints that make a centralised solution intractable. It is well known that
many decentralised optimisation tasks can be cast as potential games (Wolpert
and Turner, 1999.; Arslan et al., 2006), and the search of an optimal solution
∗[email protected]
1
can be seen as the task of finding Nash equilibria in a game. Thus it is feasi-
ble to use iterative learning algorithms from game-theoretic literature to solve
decentralised optimisation problems.
A game theoretic learning algorithm with proof of convergence in certain
kinds of games is fictitious play (Fudenberg and Levine, 1998; Monderer and
Shapley, 1996).
It is a learning process where players choose an action that
maximises their expected rewards according to the beliefs they maintain about
their opponents' strategies.The players update their beliefs about their oppo-
nents' strategies after observing their actions. Even though fictitious play con-
verges to Nash equilibrium, this convergence can be very slow. This is because
it implicitly assumes that other players use a fixed strategy in the whole game.
Smyrnakis and Leslie (2010) addressed this problem by representing the ficti-
tious play process as a state space model and by using particle filters to predict
opponents' strategies. The drawback of this approach is the computational cost
of the particle filters that render difficult the application of this method in real
time applications.
The alternative that we propose in this article is to use instead of particle
filters, extended Kalman filters (EKF) to predict opponents' strategies. There-
fore the proposed algorithm has smaller computational cost than the particle
filter variant of fictitious play algorithm that proposed by Smyrnakis and Leslie
(2010). We show that the EKF fictitious play algorithm converges to a pure
Nash equilibrium, in 2 by 2 games with at least one pure Nash equilibrium and
in potential games where players have two available actions. We also empirically
observe, in a range of games, that the proposed algorithm needs less iterations
than the classic fictitious play to converge to a solution. Moreover in our simu-
lations, the proposed algorithm converged to a solution with higher reward than
the classic fictitious play algorithm.
The remainder of this paper is organised as follows. We start with a brief
description of game theory, fictitious play and extended Kalman filters. Section
3 introduces the proposed algorithm that combines fictitious play and extended
Kalman filters. The convergence results we obtained are presented in Section
4. In Section 5 we propose some indicative values for the EKF algorithm pa-
rameters. Section 6 presents the simulation results of EKF fictitious play in a
2×2 coordination game, a three player climbing hill game and an ad-hoc sensor
network surveillance problem. In the final section we present our conclusions.
2 Background
In this section we introduce some definition from game theory that we will
use in the rest of this article and the relation between potential games and
decentralised optimisation. We also briefly present the classic fictitious play
algorithm and the extended Kalman filter algorithm.
2.1 Game theory definitions
We consider a game Γ with I players, where each player i,i = 1, 2, . . . , I, choose
his action, si, from a finite discrete set Si. We then can define the joint action
that is played in a game as the set product S = ×i=I
i=1Si. Each Player i receive a
reward, ui, after choosing an action . The reward is a map from the joint action
2
space to the real numbers, ui : S → R. We will often write s = (si, s−i), where
si is the action of Player i and s−i is the joint action of Player i's opponents.
When players select their actions using a probability distribution they use mixed
strategies. The mixed strategy of a player i, σi, is an element of the set ∆i,
where ∆i is the set of all the probability distributions over the action space Si.
The joint mixed strategy, σ, is then an element of ∆ = ×i=I
i=1∆i. Analogously
to the joint actions we will write σ = (σi, σ−i). In the special case where the
players choose an action with probabiity one we will say that players choose
their actions using pure strategies. The expected utility a player i will gain if
he chooses a strategy σi (resp. si), when his opponents choose the joint strategy
σ−i is ui(σi, σ−i) (resp. ui(si, σ−i)).
A common decision rule in game theory is best response (BR). The best
response is defined as the action that maximizes players' expected utility given
their opponents' strategies. Thus for a specific opponents' strategy σ−i we
evaluate the best response as:
BRi(σ−i) = argmax
si∈S
ui(si, σ−i)
(1)
Nash (1950) showed that every game has at least one equilibrium, which is
a fixed point of the best response correspondence, σi ∈ BR(σ−i). Thus when a
joint mixed strategy σ is a Nash equilibrium then:
ui(σi, σ−i) ≥ ui(si, σ−i)
for all si ∈ Si
(2)
Equation 2 implies that if a strategy σ is a Nash equilibrium then it is not
possible for a player to increase his utility by unilaterally changing his strategy.
When all the players in a game select their actions using pure strategies then
the equilibrium actions are referred as pure strategy Nash equilibria. A pure
equilibrium is strict if each player has a unique best response to his opponents
actions.
2.2 Decentralised optimisation tasks as potential games
A class of games that are of particular interest in multi agent systems and
decentralised optimisation tasks are potential games, because of their utility
structure. In particular in order to be able to solve an optimisation task de-
centrally the local functions should have similar characteristics with the global
function that we want to optimise. This suggests that an action which improves
or reduces the utility of an individual should respectively increase or reduce
the global utility. Potential games have this property, since the potential func-
tion (global function) depict the changes in the players' payoffs (local functions)
when they unilaterally change their actions. More formally we can write
ui(si, s−i) − ui( si, s−i) = φ(si, s−i) − φ( si, s−i)
where φ is a potential function and the above equality stands for every player
i, for every action s−i ∈ S−i, and for every pair of actions si, si ∈ Si, where Si
and S−i represent the set of all available actions for Player i and his opponents
respectively. Moreover potential games has at least one pure Nash equilibrium,
hence there is at least one joint action s where no player can increase their
reward, therefore the potential function, through a unilateral deviation.
3
It is feasible to choose an appropriate form of the agents' utility function
in order for the global utility to act as a potential of the system. Wonderful
life utility is a utility function that introduced by Wolpert and Turner (1999.)
and applied by Arslan et al. (2006) to formulate distributed optimisation tasks
as potential games. Player i's utility, when wonderful life utility is used, can
be defined as the difference between the global utility ug and the utility of the
system when a reference action is used as player's i action. More formally when
player i chooses an action si we write
ui(si) = ug(si, s−i) − ug(si
0, s−i)
where si
0 denotes the reference action of player i. Hence the decentralised op-
timisation problem can be cast as a potential game and any algorithm that is
proved to converge to a Nash equilibrium of a potential game, which is a local
or the global optimum of the optimisation problem, will converge to a joint
action from which no player can increase the global reward through unilateral
deviation.
2.3 Fictitious play
Fictitious play (Brown, 1951), is a widely used learning technique in game the-
ory.
In fictitious play each player chooses his action according to the best
response to his beliefs about his opponents' joint mixed strategy σ−i.
Initially each player has some prior beliefs about the strategy that each of
his opponents uses to choose an action based on a weight function κt. The play-
ers, after each iteration, update the weight function and therefore their beliefs
about their opponents' strategies and play again the best response according
to their beliefs. More formally in the beginning of a game Player i maintains
0, ∀j ∈ [1, I]\{i}, that are
some arbitrary non-negative initial weight functions κj
updated using the formula:
κj
t (sj) = κj
t−1(sj) + Isj
t =sj
(cid:26) 1
for each j, where Isj
The mixed strategy of opponent j is estimated from the following formula:
t =sj =
0
.
if sj
t = sj
otherwise.
σj
t (sj) =
.
(3)
(cid:80)
κj
t (sj)
s(cid:48)∈Sj κj
t (s(cid:48))
Player i based on his beliefs about his opponents' strategies, chooses the
action which maximises his expected payoffs. When player i uses equation (3)
to update the beliefs about his opponents' strategies he treats the environment of
the game as stationary and implicitly assumes that the actions of the players are
sampled from a fixed probability distribution. Therefore the recent observations
have the same weight as the initial ones. This approach leads to poor adaptation
when the other players choose to change their strategies.
2.4 Fictitious play as a state space model
We follow Smyrnakis and Leslie (2010) and we will represent fictitious play
process as a state-space model. According to this state space model each player
4
t(si) to play each of his available actions si ∈ Si, and then
has a propensity Qi
he forms his strategy based on these propensities. Finally he chooses his actions
based on his strategy and the best response decision rule. Because players have
no information about the evolution of their opponents' propensities, and under
the assumption that the changes in propensities are small from one iteration
of the game to another, we model propensities using a Gaussian autoregressive
prior on all propensities. We set Q0 ∼ N (0, I) and recursively update the value
of Qt according to the value of Qt−1 as follows:
Q(st) = Q(st−1) + ηt
where ηt ∼ N (0, χ2I). The action of a player then is related to his propensity
by the following sigmoid equation for every si ∈ Si
(cid:80)
si =
e(Qi(si)/τ )
s∈Si e(Qt(s)/τ )
.
Therefore players will assume that at every iteration t their opponents have a
different strategy σt.
2.5 Kalman filters and Extended Kalman filters
Our objective is to estimate player i's opponent propensity and thus to estimate
the marginal probability p(Qt, s1:t). This objective can be represented as a
Hidden Markov Model (HMM). HMMs are used to predict the value of an
unobserved variable xt, the hidden state, using the observations of another
variable z1:t. There are two main assumptions in the HMM representation. The
former one is that the probability of being at any state xt at time t depends
only at the state of time t − 1, xt−1. The latter one is that an observation
at time t depends only on the current state xt. One of the most common
methods to estimate p(x1:t, z1:t) is Kalman filters and its variations. Kalman
filter (Kalman et al., 1960) is based on two assumptions, the first is that the
state variable is Gaussian. The second is that the observations are the result of
a linear combination of the state variable. Hence Kalman filters can be used in
cases which are represented as the following state space model:
xt =Axt−1 + ξt−1 hidden layer
yt =Bxt + ζt observations
where ξt and ζt follow a zero mean normal distribution with covariance matrices
Ξ = qtI and Z = rtI respectively, and A, B are linear transformation matrices.
When the distribution of the state variable xt is Gaussian then p(xty1:t) is also
a Gaussian distribution, since yt is a linear combination of xt. Therefore it is
enough to estimate its mean and variance to fully characterise p(xty1:t).
Nevertheless in the state space model we want to implement, the relation
between Player i's opponent propensity and his actions is not linear. Thus we
should use a more general form of state space model such as:
xt = f (xt−1) + ξt
yt = h(xt) + ζt
5
(4)
where ξt and ζt are the hidden and observation state noise respectively, with
zero mean and covariance matrices Ξ = qtI and Z = rtI respectively. The
distribution of p(xty1:t) is not a Gaussian distribution because f (·) and h(·)
are non-linear functions. A simple method to overcome this shortcoming is to
use a first order Taylor expansion to approximate the distributions of the sate
space model in (4). In particular we let xt = mt−1 + , where mt denotes the
mean of xt and ∼ N (0, P ). We can rewrite (4) as:
xt = f (mt−1 + ) + wt−1 = f (mt−1) + Fx(mt−1) + ξt−1
yt = h(mt + ) + ζt = h(mt) + Hx(mt) + ζt
(5)
where Fx(mt−1) and Hx(mt) is the Jacobian matrix of f and h evaluated at
mt−1 and mt, respectively. If we use the transformations in (5) then p(xty1:t)
is a Gaussian distribution.
Since p(xty1:t) is a Gaussian distribution to fully characterise it we need to
evaluate its mean and its variance. The EKF process (Jazwinski, 1970; Grewal
and Andrews, 2011) estimates this mean and variance in two steps the prediction
and the update step. In the prediction step at any iteration t the distribution
of the state variable is estimated based on all the observations until time t − 1,
p(xty1:t−1). The distribution of p(xty1:t−1) is Gaussian and we will denote its
mean and variance as m−
respectively. During the update step the
estimation of the prediction step is corrected in the light of the new observation
at time t, so we estimate p(xty1:t). This is also a Gaussian distribution and we
will denote its mean and variance as mt and Pt respectively.
The prediction and the update steps of the EKF process (Jazwinski, 1970;
Grewal and Andrews, 2011) to estimate the mean and the variance of p(xty1:t−1)
and p(xty1:t) respectively are the following:
Prediction Step
t and P −
t
m−
t =f (mt−1)
P −
t =F (mt−1)Pt−1F (mt−1) + Ξt−1
where the j, j(cid:48) element of F (mt) is defined as
[F (m−
t )]j,j(cid:48) =
∂f (xj, r)
∂xj(cid:48)
x=m
−
t ,q=0
Update Step
vt = zt − h(m−
t )
t H T (m−
t )P −
St = H(m−
Kt = P −
t H T (m−
t )S−1
mt = m−
t + Ktvt
t − KtStK T
Pt = P −
t
t ) + Z
t
where zt is the observation vector (with 1 in the entry of the observed action
and 0 everywhere else) and the j, j(cid:48) element of H(mt) is defined as:
[H(m−
t )]j,j(cid:48) =
∂h(xj, r)
∂xj(cid:48)
x=m
−
t ,r=0
6
3 Fictitious play and EKF
For the rest of this paper we will only consider inference over a single opponent
mixed strategy in fictitious play. Separate estimates will be formed identically
and independently for each opponent. We therefore consider only one opponent,
and we drop all dependence on player i, and write st, σt and Qt for Player i's
opponent's action, strategy and propensity respectively. Moreover for any vector
x, x[j] will denote the jth element of the vector and for any matrix y, y[i, j] will
denote the (i, j)th element of the matrix.
We can use the following state space model to describe the fictitious play
process:
Qt = Qt−1 + ξt−1
st = h(Qt) + ζt
where ξt−1 ∼ N (0, Ξ), is the noise of the state process and ζt is is the error of
the observation state with zero mean and covariance matrix Z, which occurs
because we approximate a discrete process like best responses, equation (1),
using a continuous function h(·). Hence we can combine the EKF with fictitious
play as follows. At time t − 1 Player i has an estimation of his opponent's
propensity using a Gaussian distribution with mean mt−1 and variance Pt−1,
and has observed an action st−1. Then at time t he uses EKF prediction step
to estimate his opponent's propensity. The mean and variance of p(Qts1:t−1)
of the opponent's propensity approximation are:
m−
t = mt−1
P −
t = Pt−1 + Ξ
Player i then evaluates his opponents strategies using his estimations as:
(cid:80)
exp(m−
s∈S exp(m−
t [st]/τ )
t [s]/τ )
σt(st) =
.
(6)
where m−
t [st] is the mean of Player i's estimation about the propensity of his
opponent to play action st. Player i then uses the estimation of his opponent
strategy , equation (6), and best responses, equation (1), to choose an action.
After observing the opponent's action st, Player i correct his estimations about
his opponent's propensity using the update equations of EKF process. The
update equations are:
vt = zt − h(m−
t )
St = H(m−
t )P −
t H T (m−
Kt = P −
t )S−1
t H T (m−
mt = m−
t + Ktvt
t − KtStK T
Pt = P −
t
t
t ) + Z
(cid:80)
(cid:48)
where h = exp(Qt[s
s∈S exp(Qt[s]/τ ) , and τ is a temperature parameter. The Jacobian
matrix H(m−
t ) is defined as
]/τ )
7
[H(m−
t )]j,j(cid:48) =
(cid:80)
((cid:80)
j(cid:54)=j(cid:48) exp(m
((cid:80)
− exp(m
−
t [j(cid:48)])
j exp(m
−
t [j]) exp(m
−
t [j]))2
−
t [j(cid:48)])
−
t [j]) exp(m
j exp(m
−
t [j]))2
if j = j(cid:48)
if j (cid:54)= j'
.
Table 1 summarises the fictitious play algorithm when EKF is used to predict
opponents strategies.
At time t
1. Player i maintains some estimations about his opponents propensity up
to time t − 1, p(Qt−1s1 : t − 1). Thus he has an estimation of the mean
mt−1 and the covariance Pt−1 of this distribution.
2. Then Player i is updating his estimations about his opponents propensi-
ties p(Qts1 : t − 1) using equations, m−
(cid:80)
opponents strategies using σj
−
t (sj) = exp(m
t (j)/τ )
j(cid:48) exp(m
−
t (j)/τ )
.
t = mt−1, P −
t = Pt−1 + Wt−1.
3. Based on the weights of step 1 each player updates his beliefs about his
4. Choose an action based on the beliefs of step 3 according to best response.
5. Observe opponent's action st.
6. Update the propensities estimates using mt = m−
t + Ktvt and
Pt = P −
t − KtStK T
t .
7. set t=t+1
Table 1: EKF Fictitious Play algorithm
4 Theoretical Results
In this section we present the convergence results we obtained for games with at
least one pure Nash equilibrium and players who have 2 available actions, s =
(1, 2). We will denote as −s the action that a player does not choose, for example
if Player i's opponent chooses action 1, s = 1 and hence −s = 2. Also we will
denote as m[1] and m[2] the estimated means of opponent's propensity of action
1 and 2 respectively. Similarly P [1, 1] and P [2, 2] will represent the variance of
the propensity's estimation of action 1 and 2 respectively, and P [1, 2], P [2, 1]
their covariance.
The proposed algorithm has the following two properties:
Proposition 1. If at iteration t of the EKF fictitious play algorithm, action s is
played from Player i's opponent, then the estimation of his opponent propensity
to play action s increases, mt−1[s] < mt[s]. Also the estimation of his opponent
propensity to play action −s decreases, mt−1[−s] > mt[−s]
Proof. The proof of Proposition 1 is on Appendix A.
8
L
U 1,1
D 0,0
R
0,0
1,1
Table 2: Simple coordination game
Proposition 1 implies that players, when they use EKF fictitious play, learn
their opponent's strategy and eventually they will choose the action that will
maximise their reward base on their estimation. Nevertheless there are cases
where players may change their action simultaneously and trapped in a cycle
instead of converging in a pure Nash equilibrium. As an example we consider
the game that is depicted in Table 2.
This is a simple coordination game with two pure Nash equilibria the joint
actions (U, L) and (D, R). In the case were the two players start from joint action
(U, R) or (D, L) and they always change their action simultaneously then they
will never reach one of the two pure Nash equilibria of the game.
Proposition 2. In a 2 × 2 game where the players use EKF fictitious play
process to choose their actions, and the variance of the observation state is set
to Z = rI + I, with high probability the two players will not change their action
simultaneously infinitely often. We define as a random number from normal
distribution with zero mean and arbitrarily small covariance matrix, I is the
identity matrix.
Proof. The proof of Proposition 2 is on Appendix B.
We should mention here that the reason we set Z = rI + I is in order
to break any symmetries that occurred because the initialisation of the EKF
fictitious play algorithm. Based on Proposition 1 and 2 we can infer the following
propositions and theorems.
Proposition 3. (a) In a game where players have two available actions if s is a
Nash equilibrium, and s is played at date t in the process of EKF fictitious play,
s is played at all subsequent dates. That is, strict Nash equilibria are absorbing
for the process of EKF fictitious play. (b) Any pure strategy steady state of EKF
fictitious play must be a Nash equilibrium.
Proof. Consider the case where players beliefs σt, are such that their optimal
choices correspond to a strict Nash equilibrium s. In EKF fictitious play pro-
cess players' beliefs are formed identically and independently for each opponent
based on equation (6). By Proposition 1 we know that players' estimations
about their opponents' propensities and therefore their strategies, that each
player maintains for the other players, will increase for the actions that are
included in s and will be reduced otherwise. Thus the best response to their
beliefs σt+1 will be again s and since s is a Nash equilibrium they will not de-
viate from it. Conversely, if a player remains at a pure strategy profile, then
eventually the assessments will become concentrated at that profile, because of
Proposition 1, hence if the profile is not a Nash equilibrium, one of the players
would eventually want to deviate.
9
Proposition 4. Under EKF fictitious play, if the beliefs over each player's
choices converge, the strategy profile corresponding to the product of these dis-
tributions is a Nash equilibrium.
Proof. Suppose that the beliefs of the players at time t, σt, converges to some
profile σ. If σ were not a Nash equilibrium, some player would eventually want
to deviate and the beliefs would also deviate since based on Proposition 1 players
eventually learn their opponents actions.
Theorem 1. The EKF fictitious play process converges to the Nash equilibrium
in 2 × 2 games with at least one pure Nash equilibrium, when the covariance
matrix of the observation space error, Z, is defined as in Proposition 2, Z =
rI + I.
Proof. We can distinct two possible initial states in the game.
In the first
players' initial beliefs of the players actions are such that their initial joint action
s0 is a Nash equilibrium. From Proposition 3 and equation (6) we know that
they will play the joint action which is a Nash equilibrium for all the iterations
of the game.
The second case where the initial beliefs of the players are such that their
initial joint action s0 is not a Nash equilibrium is divided in 2 subcategories.
The first include 2× 2 games with only one pure Nash equilibrium. In this case,
one of the two players has a dominant action, thus for all the iterations of the
game he will choose the dominant action. This action maximises his expected
payoff regardless the other player's strategy and thus he will select this action
in every iteration of the game. Therefore because of Proposition 1 the other
player will learn his opponent's strategy and players will choose the joint action
which is the pure Nash equilibrium.
The second category includes 2 × 2 games with 2 pure Nash equilibria, like
the simple coordination game that is depicted in Table 2. In this case players
initial joint action s0 = (s1, s2) is not a Nash equilibrium. Then the players
will learn their opponent's strategy, Proposition 1 and Equation (6), and they
will change their action. We know from Proposition 2 that in a finite time with
high probability the players will not change their actions simultaneously, and
hence they will end up in a joint action that will be one of the two pure Nash
equilibria of the game.
We can extend the results of Theorem 1 in n × 2 games with a better reply
path. A game with a better reply path can be represented as a graph were its
edges are the join actions of the game s and there is a vertex that connects
s with s(cid:48) iff only one player i can increasing his payoff by changing his action
(Young, 2005). Potential games have a better reply path.
Theorem 2. The EKF fictitious play process converges to the Nash equilib-
rium in n × 2 games with a better reply path when the covariance matrix of the
observations space error, Z, is Z = r + I.
Proof. Similarly to the 2 × 2 games if the initial beliefs of the players are such
that their initial joint action s0 is a Nash equilibrium, from Proposition 3 and
equation (6), we know that they will play the joint action which is a Nash
equilibrium for the rest of the game.
10
Moreover in the case of the initial beliefs of the players are such that their
initial joint action s0 is not a Nash equilibrium based on Proposition 1 and
Proposition 2 after a finite number of iterations because the game has a better
reply path the only player that can improve his payoff by changing his actions
will choose a new action which will result in a new joint action s. If this action
is not the a Nash equilibrium then again after finite number of iterations the
player who can improve his payoff will change action and a new joint action s(cid:48)
will be played. Thus after the search of the vertices of a finite graph, and thus
after a finite number of iterations, players will choose a joint action which is a
Nash equilibrium.
5 Simulations to define algorithm parameters Ξ
and Z.
The covariance matrix of the state space error Ξ = qI and the measurement
error Z = rI are two parameters that we should define in the beginning of the
EKF fictitious play algorithm and they affect its performance. Our aim is to
find values, or range of values, of q and r that can efficiently track opponents'
strategy when it smoothly or abruptly change, instead of choosing q and r
heuristically for each opponent when we use the EKF algorithm. Nevertheless
it is possible that for some games the results of the EKF algorithm will be
improved for other combinations of q and r than the ones that we propose in
this section.
n +1
2
We examine the impact of EKF fictitious play algorithm parameters in its
performance in the following two tracking scenarios. In the first one a single
opponent chooses his actions using a mixed strategy which changes smoothly and
has a sinusoidal form over the iterations of the tracking scenario. In particular
= 1 − σt(2),
for t = 1, 2, . . . , 100 iterations of the game: σt(1) = cos 2πt
In the second toy example Player i's opponent change his
where n = 100.
strategy abruptly and chooses action 1 with probability σ2
t (1) = 1 during the
first 25 and the last 25 iterations of the game and for the rest iterations of
the game σ2
t (1) = 0. The probability of the second action is calculated as:
t (2) = 1 − σ2
σ2
We tested the performance of the proposed algorithm for the following range
of parameters 10−4 ≤ q ≤ 1 and 10−4 ≤ r ≤ 1. We repeated both examples
100 times for each of the combinations of q and r. Each time we measured the
absolute error of the estimated strategy against the real one. The combined
average absolute error when both examples are considered is depicted on Figure
1. The darkest areas of the contour plot represent the areas where the average
absolute error is minimised.
t (1).
The average absolute error is minimised for a range of values of q and r,
that form two distinct areas.
In the first area, the wide dark area of Figure
1, the range of q and r were 0.08 ≤ q ≤ 0.4 and 0.2 ≤ r ≤ 1 respectively. In
the second area, the narrow dark area of Figure 1, the range of q and r were
0.001 ≤ q ≤ 0.025 and 0.08 ≤ r ≤ 0.13 respectively. The minimum error which
we observed in our simulations was in the narrow area and in particular when
Ξ = 0.01I and Z = 0.1I, where I is the identical matrix.
11
Figure 1: Combined absolute error for both tracking scenarios. The range of
both parameters, q and r is between 10−4 and 1.
U
M
D
U M D
0
0
40
0
0
30
0
50
0
U
U
-300
-300
0
M D
80
70
0
60
0
0
M
U
100
M
-300
0
0
0
0
D
D
90
0
0
Table 3: Climbing hill game with three players. Player 1 selects rows, Player 2
selects columns, and Player 3 selects the matrix. The global reward depicted in
the matrices, is received by all players. The unique Nash equilibrium is in bold
6 Simulation results
This section is divided in two parts. The first part contains results of our simu-
lations in two strategic form games and the second part contains the results we
obtained in an ad-hoc sensor network surveillance problem. In all the simula-
tions of this section we set the covariance matrix of the hidden and the observa-
tions state to Ξ = 0.01I and Z = (0.1 + )I respectively, where ∼ N (0, 10−5)
and I is the identical matrix.
6.1 Simulations results in strategic form games
In this section we compare the results of our algorithm with those of fictitious
play in two coordination games. These games are depicted in Tables 2 and 3.
The game that is depicted in Table 2, as it was described in Section 4 , is a simple
coordination game with two pure Nash equilibria, its diagonal elements. Table
3 presents an extreme version of the climbing hill game (Claus and Boutilier,
1998) in which three players must climb up a utility function in order to reach
the Nash equilibrium where their reward is maximised.
We present the results of 50 replications of a learning episode of 50 iterations
for each game. As it is depicted in Figures 2 and 3 the proposed algorithm
performs better than fictitious play in both cases. In the simple coordination
game that is shown in Table 2, the EKF fictitious play algorithm converges to
12
Figure 2: Results of EKF and classic fictitious play in the simple coordination
game of Table 2
one of the pure equilibria after a few iterations. On the other hand fictitious
play is trapped in a limit cycle in all the replications where the initial joint
action was not one of the two pure Nash equilibria. For that reason the players'
payoff for all the iterations of the game was either 1 utility unit or 0 utility
units depending to the initial joint action. In the climbing hill game, Table 3
the proposed algorithm converges to the Nash equilibrium after 35 iterations
when fictitious play algorithm do not converge even after 50 iterations.
6.2 Ad-hoc sensor network surveillance problem.
We compared the results of our algorithm against those of fictitious play in a
coordination task of a power constrained sensor network, where sensors can be
either in a sense or sleep mode (Farinelli et al., 2008; Chapman et al., 2011).
When the sensors are in sense mode they can observe the events that occur in
their range. During their sleep mode the sensors harvest the energy they need
in order to be able function when they are in the sense mode. The sensors then
should coordinate and choose their sense/sleep schedule in order to maximise
the coverage of the events. This optimisation task can be cast as a potential
game. In particular we consider the case where I sensors are deployed in an area
where E events occur. If an event e, e ∈ E, is observed from the sensors then it
produce some utility Ve. Each of the sensors i = 1, . . . , I should choose an action
si = j, from one of the j = 1, . . . , J time intervals which they can be in sense
mode. Each sensor i when it is in sense mode can observe an event e, if it is
in its sense range, with probability pie = 1
, where die is the distance between
die
the sensor i and the event e. We assume that the probability each sensor has
to observe an event is independent from the other sensors. If we denote as iin
the sensors that are in sense mode when the event e occurs and e is in their
sensing range, then we can write the probability an event e to be observed from
13
Figure 3: Probability of playing the (U,U,D) equilibrium for the EKF fictitious
play (solid line) and fictitious play (dash line) for the three player climbing hill
game
the sensors, iin as
1 − (cid:89)
i∈iin
(1 − pie)
The expected utility that is produced from the event e is the product of its
utility Ve and the probability it has to be observed by the sensors, iin that are
in sense mode when the event e occurs and e is in their sensing range. More
formally we can express the utility that is produced from an event e as:
The global utility is then the sum of the utilities that all events, e ∈ E, produce
(1 − pie))
Ue(s) = Ve(1 − (cid:89)
(cid:88)
i∈iin
Uglobal(s) =
Ue(s).
e
Each sensor after each iteration of the game receives some utility which is
based on the sensors and the events that are inside his communication and sense
range respectively. For a sensor i we denote e the events that are in its sensing
range and s−i the joint action of the sensors that are inside his communication
range. The utility that sensor i will receive if his sense mode is j will be
(cid:88)
Ui(si = j, s−i) =
Ue(si = j, s−i).
e
We compared the performance of the two algorithms in 2 instances of the
above scenario one with 20 and one with 50 sensors that are deployed in a unit
square. In both instances sensors had to choose one time interval of the day that
they will be in sense mode and use the rest time intervals to harvest energy. We
consider cases where sensors had to choose their sense mode between 2, 3 and
14
4 available time intervals. Sensors are able to communicate with other sensors
that are at most 0.6 distance units away, and can only observe events that are
at most 0.3 distance units away. Moreover in both instances we assumed that
20 events took place in the unite square area. Those events were uniformly
distributed in space and time, so an event could evenly appear in any point of
the unit square area and it could occur at any time with the same probability.
The duration of each event was uniformly chosen between (0-6] hours and each
event had a value Ve ∈ (0 − 1]. Figures 4 and 5 depict the average results
of 50 replications of the game for the two algorithms. For each instance, both
algorithms run for 50 iterations. To be able to average across the 50 replications
we normalise the utility of a replication by the global utility that the sensors
will gain if they were only in sense mode during the whole day.
(a) Results when sensors have to
choose between two time intervals.
(b) Results when sensors have to
choose between three time intervals.
(c) Results when sensors have to
choose between four time intervals.
Figure 4: Results of the instance where 20 sensors should coordinate for both
algorithms. The results of EKF fictitious play are the solid lines and the results
of the classic fictitious play are the dash lines. The horizontal axis of the figures
depict the iteration of the game and the vertical axis the global utility as a
percentage of the global utility of the system in the case that sensors were
always in sense mode.
As we observe in Figures 4 and 5 EKF fictitious play converges to a stable
joint action faster than the fictitious play algorithm. In particular on average
the EKF fictitious play algorithm needed 10 "negotiation" steps between the
sensors in order to reach a stable joint action, when fictitious ply needed more
than 25. Moreover the classic fictitious play algorithm was always resulted in
joint actions with smaller reward than the proposed algorithm.
15
(a) Results when sensors have to
choose between two time intervals.
(b) Results when sensors have to
choose between three time intervals.
(c) Results when sensors have to
choose between four time intervals.
Figure 5: Results of the instance where 50 sensors should coordinate for both
algorithms. The results of EKF fictitious play are the solid lines and the results
of the classic fictitious play are the dash lines. The horizontal axis of the figures
depict the iteration of the game and the vertical axis the global utility as a
percentage of the global utility of the system in the case that sensors were
always in sense mode.
7 Conclusion
We have introduced a variation of fictitious play that uses Extended Kalman
filters to predict opponents' strategies. This variation of fictitious play addresses
the implicit assumption of the classic algorithm that opponents use the same
strategy in every iteration of the game.
We showed that, for 2 × 2 games with at least one pure Nash equilibrium,
EKF fictitious play converges in the pure Nash equilibrium of the game. More
over the proposed algorithm converges in games with a better reply path, like
potential games, and n players that have 2 available actions.
EKF fictitious play performed better than the classic algorithm algorithm
in the strategic form games and the ad-hoc sensor network surveillance problem
we simulated. Our empirical observations indicate that EKF fictitious play con-
verges to a solution that is better than the classic algorithm and needs only a
few iterations to reach that solution. Hence by slightly increasing the computa-
tional intensity of fictitious play less communication is required between agents
to quickly coordinate on a desired solution.
16
8 Acknowledgements
This work is supported by The Engineering and Physical Sciences Research
Council EPSRC (grant number EP/I005765/1).
A Proof of Proposition 1
We will base the proof of Proposition 1 on the properties of EKF when they
used to estimate opponent's strategy with two available actions. If player i's
opponent has two available actions 1 and 2, then we can assume that at time
t−1 Player i maintains beliefs about his opponent's propensity, with mean mt−1
and variance Pt−1. Moreover based on these estimations he chooses his strategy
σt−1. At the prediction step of this process he uses the following equations to
predict his opponent's propensity and choose an action using best response.
(cid:18) m−
(cid:19)
m−
(cid:18) P −
t−1[1]
m−
t =
t−1[2]
t−1[1, 1] P −
P −
t−1[2, 1] P −
t−1[1, 2]
t−1[2, 2]
(cid:19)
+ qI
P −
t =
without loss of generality we can assume that his opponent in iteration t chooses
action 2. Then the update step will be :
since Players i's opponent played action 2 and h =
write vt and Ht(m−
t ) as:
(cid:48)
exp(Qt[s
s∈S exp(Qt[s]/τ ) we can
]/τ )
(cid:80)
(cid:19)
vt = zt − h(m−
t )
vt =
=
1 − σt−1(1)
(cid:19)
(cid:18) 0
(cid:18) −σt−1(1)
−
1
σt−1(1)
(cid:18) σt−1(1)
(cid:19)
(cid:18) at −at
(cid:19)
(cid:18) b −b
−at
at
(cid:19)
−b
b
+ Z
Ht(m−
t ) =
where at is defined at = σt−1(1)σt−1(2). The estimation of St = H(m−
will be:
t )P −
t H T (m−
t ) + Z
t [1, 2]. The Kalaman gain, Kt = P −
t
t H T (m−
t )S−1
(cid:19)
b
(cid:19)(cid:18) b + r
b
b + r
where b = P −
can be written as
t [1, 1]+P −
St = a2
t [2, 2]−2P −
(cid:18) P −
Kt =
1
t [1, 1]
2rb + r2
k
k
P −
t [2, 2]
up to a multiplicative constant we can write
(cid:19)(cid:18) at −at
(cid:18) c −c
−at
(cid:19)
at
−d
d
K1 ∼
17
where c = P −
the mean and variance are:
t [1, 1]− P −
t [1, 2] and d = P −
t [2, 2]− P −
t [1, 2]. The updates then for
mt =m−
Pt =P −
t + Ktvt
t − KtStK T
t
The mean of the Gaussian distribution that is used to estimate opponent's
propensities is:
(cid:18) mt[1]
mt[2]
(cid:32)
(cid:19)
=
mt =
(cid:33)
t [1] − 2σ(1)
m−
m−
t [2] + 2σ(1)
a(b−k)
a(b−k)
4a2(b−k)+(r+)
4a2(b−k)+(r+)
(7)
Based on the above we observe that mt(1) < mt−1(1) and mt(2) > mt−1(2)
which completes the proof.
B Proof of Proposition 2
We consider 2 × 2 games with at least one pure Nash equilibrium. In the case
that only one Nash equilibrium exists, a dominant strategy exists and thus one
of the players will not deviate from this action. Hence we are interested in
in 2 × 2 games with two pure Nash equilibria. Without loss of generality we
consider a game with similar structure to the simple coordination game that is
depicted in Table 2. with two equilibria, the joint actions in the diagonal of the
payoff matrix, (U, L) and (D, R). We will present calculations for Player 1,but
the same results hold also for Player 2. We define λ as the necessary confidence
level that Player 1's estimation of σt(L) should reach in order to choose action
U . Hence we Player 1 will choose D if:
σt(1) > λ ⇔
> λ ⇔
exp(m−
t [1])
exp(m−
t [1]) + exp(m−
t [2])
m−
t [1] > ln(
mt−1[1] > ln(
λ
1 − λ
λ
1 − λ
) + m−
t [2] ⇔
) + mt−1[2]
In order to prove Proposition 2, we need to show that when a player changes
his action his opponent will change his action at the same iteration with proba-
bility less than 1. In the case where at time t − 1 the joint action of the players
is U, R then Player 1 believes that his opponent will play L, while he observing
him playing R. Assume that Player 2's beliefs about Player 1's strategies has
reached the necessary confident level about Players 1's strategy and at iteration
t he will change his action from R to L. Player 1 will also change his action at
the same time if
mt−1[2] > ln(
1 − λ
λ
) + mt−1[1]
We want to show that players will not change actions simultaneously with prob-
ability 1. Hence it is enough to show that
18
P rob(mt−1[1] > ln(
λ
1 − λ
) + mt−1[2]) > 0
(8)
We can replace mt−1[1] and mt−1[2] with their equivalent from (7) and write:
t [1] − 2σ(1)
m−
a(b − k)
4a2(b − k) + (r + )
> ln(
λ
1 − λ
) + m−
t [2] + 2σ(1)
a(b − k)
4a2(b − k) + (r + )
⇔
− 4σ(1)
4a2(b − k) + (r + )
> ln(
4a2(b − k) + (r + )
<
−4σ(1)
a(b − k)
a(b − k)
) + m−
λ
1 − λ
1−λ ) + m−
t [2] − m−
t [2] − m−
t [1]
ln( λ
t [1] ⇔
Solving this with respect to we have
a(b − k)σ(1)
>
ln( λ
1−λ ) + m−
t [2] − m−
t [1]
Thus we can write (8) as:
P rob( >
a(b − k)σ(1)
ln( λ
1−λ ) + m−
t [2] − m−
t [1]
− a2(b − k) − r
− a2(b − k) − r) > 0
(9)
Since is a Gaussian white noise (9) is always true.
We also consider the case where at time t − 1 the joint action of the players
is D, L then Player 1 believes that his opponent will play R, while he observing
him playing L. Assume that Player 2's beliefs about Player 1's strategies has
reached the necessary confident level and at t he will change his action from L
to R. Player 1 will also change his action at the same time if
mt−1[1] > ln(
λ
1 − λ
) + mt−1[2]
We want to show that Players will not change actions simultaneously with prob-
ability 1. Hence it is enough to show that
1 − λ
λ
P rob(mt−1[2] > ln(
) + mt−1[1]) > 0
(10)
We can rewrite (10) using the results we obtained for mt−1[1] and mt−1[2] in
(7) again as
P rob( >
a(b − k)σ(1)
ln( λ
1−λ ) + m−
t [2] − m−
t [1]
− a2(b − k) − r) > 0
(11)
Since is a Gaussian white noise (11) is always true.
If we define ξt the event that both players change their action at time t simul-
taneously, and assume that the two players have change their actions simultane-
ously at the following iterations t1, t2, . . . , tt, then the probability that they will
also change their action simultaneously at time tT +1, P (ξt1 , ξt2, . . . , ξtT , ξtT +1)
is almost zero for large but finite T .
19
References
Arslan, G., Marden, J., Shamma, J.. Autonomous vehicle-target assignment:
A game theoretical formulation. 2006.
Brown, G.W.. Iterative solutions of games by fictitious play. In: Koopmans,
T.C., editor. Activity Analysis of Production and Allocation. Wiley; 1951. p.
374 -- 376.
Chapman, A.C., Leslie, D.S., Rogers, A., Jennings, N.R.. Convergent learn-
ing algorithms for potential games with unknown noisy rewards. Working
Papers 05/2011; University of Sydney Business School, Discipline of Business
Analytics; 2011.
Claus, C., Boutilier, C.. The dynamics of reinforcement learning in cooperative
In: Proceedings of the fifteenth nationalon Artificial
multiagent systems.
intelligence. 1998. .
Farinelli, A., Rogers, A., Jennings, N.. Maximising sensor network efficiency
through agent-based coordination of sense/sleep schedules. In: Workshop on
Energy in Wireless Sensor Networks in conjuction with DCOSS 2008. 2008.
p. 43 -- 56.
Fudenberg, D., Levine, D.. The theory of Learning in Games. The MIT Press,
1998.
Grewal, M., Andrews, A.. Kalman filtering: theory and practice using MAT-
LAB. Wiley-IEEE press, 2011.
Jazwinski, A.. Stochastic processes and filtering theory. volume 63. Academic
press, 1970.
Kalman, R., et al. A new approach to linear filtering and prediction problems.
Journal of basic Engineering 1960;82(1):35 -- 45.
Kho, J., Rogers, A., Jennings, N.R.. Decentralized control of adaptive sam-
pling in wireless sensor networks. ACM Trans Sen Netw 2009;5(3):1 -- 35.
van Leeuwen, P., Hesselink, H., Rohlinga,
J.. Scheduling aircraft using
constraint satisfaction. Electronic Notes in Theoretical Computer Science
2002;76:252 -- 268.
Monderer, D., Shapley, L.. Potential games. Games and Economic Behavior
1996;14:124 -- 143.
Nash, J.. Equilibrium points in n-person games. In: Proceedings of the National
Academy of Science, USA. volume 36; 1950. p. 48 -- 49.
Smyrnakis, M., Leslie, D.S.. Dynamic Opponent Modelling in Fictitious Play.
The Computer Journal 2010;.
Stranjak, A., Dutta, P.S., Ebden, M., Rogers, A., Vytelingum, P.. A multi-
agent simulation system for prediction and scheduling of aero engine overhaul.
In: AAMAS '08: Proceedings of the 7th international joint conference on
Autonomous agents and multiagent systems. 2008. p. 81 -- 88.
20
Wolpert, D., Turner, K.. An overview of collective intelligence. Handbook of
Agent Technology 1999.;.
Young, H.P.. Strategic Learning and Its Limits. Oxford University Press, 2005.
21
|
1004.3809 | 1 | 1004 | 2010-04-21T21:33:32 | Artificial Immune Systems Metaphor for Agent Based Modeling of Crisis Response Operations | [
"cs.MA",
"cs.AI",
"cs.CY"
] | Crisis response requires information intensive efforts utilized for reducing uncertainty, calculating and comparing costs and benefits, and managing resources in a fashion beyond those regularly available to handle routine problems. This paper presents an Artificial Immune Systems (AIS) metaphor for agent based modeling of crisis response operations. The presented model proposes integration of hybrid set of aspects (multi-agent systems, built-in defensive model of AIS, situation management, and intensity-based learning) for crisis response operations. In addition, the proposed response model is applied on the spread of pandemic influenza in Egypt as a case study. | cs.MA | cs | Artificial Immune Systems Metaphor for Agent Based
Modeling of Crisis Response Operations
Khaled M. Khalil1, M. Abdel-Aziz1, Taymour T. Nazmy1,
Abdel-Badeeh M. Salem1
1Faculty of Computer and Information Science Ain shams University Cairo, Egypt
[email protected], [email protected], [email protected],
[email protected]
Abstract. Crisis response requires information intensive efforts utilized for
reducing uncertainty, calculating and comparing costs and benefits, and
managing resources in a fashion beyond those regularly available to handle
routine problems. This paper presents an Artificial Immune Systems (AIS)
metaphor for agent based modeling of crisis response operations. The presented
model proposes integration of hybrid set of aspects (multi-agent systems, built-
in defensive model of AIS, situation management, and intensity-based learning)
for crisis response operations. In addition, the proposed response model is
applied on the spread of pandemic influenza in Egypt as a case study.
Keywords: Crisis Response, Multi-agent Systems, Agent-Based Modeling,
Artificial Immune Systems, Process Model.
1 Introduction
The challenge of crisis response is reducing the influence crises cause to society, the
economy, and the lives of individuals and communities. This challenge is extreme in
several dimensions. The demand is highly diverse and largely unpredictable in terms
of location, time, and specific resources needed. Moreover, the urgency associated
with crisis has many implications, such as the need to rapidly identify information
about the developing situation, and to have the capability to make good decisions in
the face of an inevitable degree of uncertainty and incompleteness of information. An
efficient crisis response is of paramount importance, because if not responded to
promptly and managed properly, even a small mishap could lead to a very big
catastrophe with significantly severe consequences. Being equipped with a profusion
of resources does not ensure a successful response to the crisis situation. Thus, the key
to the successful response necessitates an effective and expedited allocation of the
requested resources to the emergency locations. Such complexity suggests the use of
intelligent agents for adaptive real-time modeling of the crisis response operations
[16]. Multi-agent Systems are computational systems where software agents
cooperate or compete with each other to achieve an individual or collective task [30].
In order to build multi-agent architecture, and increase the effectiveness of the
crisis response operations, a similar metaphor is required that mimics the crisis
response operations. AIS metaphor is selected in this study. AIS are a computational
systems inspired by the principles and processes of the biological immune system [5]
[9] . AIS represent an area of vast research over the last few years. For example,
developed AIS in a variety of domains, such as machine learning [14], anomaly
detection [4] [10], data mining [21], computer security [20] [6], adaptive control [24]
and fault detection [7]. The biological immune system is a robust, complex, adaptive
system that defends the body from foreign pathogens. It is able to categorize all cells
(or molecules) within the body as self-cells or non-self cells. It does this with the help
of a distributed task force that has the intelligence to take action from a local and also
a global perspective using its network of chemical messengers for communication [6].
A more detailed overview of the immune system can be found in many textbooks [23]
[26]. The immune system combines a priori knowledge with the adapting capabilities
of a biological immune system to provide a powerful alternative to currently available
techniques for pattern recognition, learning and optimization [22] . It uses several
computational models and algorithms such as Bone Marrow Model, Negative
Selection Algorithm, and Clonal Selection Algorithm [13].
In this paper we propose a multi-agent based model for crisis response. The
proposed model architecture and operations process are adopted from AIS. Then the
proposed response model is applied on controlling pandemic influenza in Egypt.
Section 2 provides the proposed response model, while section 3 presents design of
the proposed model for pandemic influenza in Egypt. Section 4 includes experiments.
Finally, section 5 includes conclusions.
2 The Proposed Response Model
The view of the biological immune system provides the basis for a representation of
AIS as systems of autonomous agents which exist within a distributed and
compartmentalized environment [29]. In what follows, we present the multi-agent
model based on the AIS metaphor for crisis response operations.
2.1 Proposed Hierarchical Architecture for Multi-Agent Response Model
The architecture of the AIS can be abstracted into hierarchy of three levels (cells,
tissue, and host) (see Fig. 1). Cells are able to interact with their environment and
communicate and coordinate their behavior with other cells by synthesizing and
responding to a range of molecules. Cells within the body aggregate to form tissue,
such as muscle or connective tissue. Tissues themselves combine to form hosts, such
as the heart, brain, or thymus. Hosts work together to form the immune system.
The proposed multi-agent architecture follows the same hierarchical architecture of
the biological immune systems with mapping of the functionalities of cells to agents
and adopting crisis response domain attributes and operations levels (operational,
tactical and strategic levels [2] [3]) (see Table 1). Pathogens represent source of
danger to the body entity, in which immune systems antibody cells tries to detect and
kill. Pathogens are mapped to danger sources or undesired situations in the crisis
domain, in which agents have to detect and overcome. Cells are represented by agents
working as first responders and voluntaries. Cells contain different type of receptors
which affect their capability to match pathogens and to kill them. Thus, receptors are
mapped to agents skills or resources required to overcome danger. Agents are working
in groups belonging to certain organization (tissue) which provide help by other
agents teams dedicated in the tactical level. Host represents the grouping of different
tissues working together. This can be mapped to emergency operations center (EOC)
of different working divisions and each division contains specialized teams.
Human Immune System
Hos t
Hos t
Hos t
Tissues
T issues
Tissues
.....
.....
Cells
Cells
Cells
.....
Fig. 1. Hierarchical architecture of human immune system
Table 1. Mapping biological immune system levels to crisis response operations and levels
Biological Immune
System Level
Pathogens
Cellular
Tissue
Host
System
Crisis
Response Level
-
Operational
Tactical
Strategic
-
Crisis Domain
Danger sources – harmful situations
First Responders and Voluntaries
Helper Agents staff
Emergency Operations Centers
Human Society and Important Properties
2.2 AIS Operational Architecture for Multi-Agent Model
However, the AIS layered structure which is adopted in previous section is not
complete from the conceptual framework perspective, which is required to allow
effective algorithms to be developed [28]. Brownlee [1] said that "The acquired
immune system provides a general pattern recognition and defense system that
modifies itself, adapting to improve its capability with experience".
Decision Making Process of AIS (Conceptual Model Formalization)
The effectiveness of the system is due to a set of internal strategies to cope with
pathogenic challenges. Such strategies remodel over time as the organism develops,
matures, and then ages. Towards determining the decision making process of AIS,
rational reconstructions approach is used. Rational reconstructions operate so as to
transform a given problematic philosophical scientific account-particularly of a
terminological, methodological or theoretical entity-into a similar, but more precise,
consistent interpretation [8] . Proposed rational reconstruction of AIS follows the same
steps of Grant et al. work [12] . Grant et al. steps include: definition of requirements of
the model, definition of the top-level use-cases, selection of notation, formalization of
the process model by walking through use-cases, implementation, and evaluation.
Definition of requirements of the proposed response model follows crisis response
systems design requirements [17] [18]. The following requirements were defined as:
• The model processes are concurrent.
• Support multiple instances of agents.
• Agents located in different layers should share required information only.
• Allow continuous monitoring of the situation and available resources.
• Permit relationships between agents to be collaborative.
• Allow separate behavior of each agent based on its role and objectives.
• Allow removal and adding of new agents and components at run time.
• Allow continuous planning/re-planning.
•
Integrate planning and learning processes.
• Each process defined can be done by one or more agents. Agents with
different roles differ by their allocated processes.
Definition of the top-level use cases is as follows (see Fig. 2):
• Use-case (0): no change in environment. This use-case applies when no
pathogens found in the environment.
• Use-case (1): antibody found a pathogen. This use-case applies when the
antibody finds a pathogen in the environment.
• Use-case (2): antibody receptors detect the required response for pathogens.
This use-case applies when the antibody receptors match the pathogen, and
can provide required response.
• Use-case (3): antibody receptors cannot detect the type of the pathogen thus
failed to provide required response. This use-case applies when the antibody
failed to match the pathogen and failed to provide response.
• Use-case (4): antibody cell asks for help for handling the unknown pathogen.
This use-case applies when antibody failed to response to the pathogen.
Antibody sends signals to activate the adaptive response.
• Use-case (5): adaptive cells mutate to match the pathogen and generate
required receptors. This use-case applies when the adaptive cells mutate to
match the pathogen and generate the required receptors.
• Use-case (6): required receptors are cloned to be applied to pathogen. This
use-case applies when the required receptors for response are cloned to
provide response to pathogen.
• Use-case (7): successful response is sustained as memory cells. This use-case
applies when a successful response is executed; the receptors are stored in
memory cells for later usage. Go to use-case (0).
• Use-case (8): failed response is ignored. This use-case applies when a failed
response is gained; the receptors are ignored and not stored in memory. Go
to use-case (5).
Selection of Notation includes selection of process notation for representing the
process model. Integrated DEFinition Methods Technique (IDEFS0) [15] represents a
common notation used in process modeling which is highly suited to specifying
systems in terms of functional processes [12] .
Fig. 2. Rationally reconstructed AIS model, formalized using SADT notation
The rationally reconstructed model (see Fig. 2) can be explained by starting at the
environment and considering the activity of a typical Cell. Antibody cell checks other
cells in the environment. Antibody cell determines if the examined cell is self or non-
self based on the available receptors in the system using specific discrimination
method. If cell is identified as non-self, cell immediately tries to match the receptors
of the cell with existing receptors for response using matching method. If response
receptors are found, clonal of the antibodies using clonal algorithm is applied to attack
the pathogen cell. In case of new pathogen receptors not currently recognized, cell
activates the adaptive response mechanism. Cells in the adaptive response mechanism
mutate to match the non-self receptors using available genes library and negative
selection algorithm. When reach an acceptable receptors form, the generated
antibodies are cloned to provide response to the pathogen cell. Then, the newly
generated receptors are added to the receptors library.
Table 2 shows comparison between the AIS rational reconstructed model and other
process models such as OODA, RPDM, and Rasmussen models. Comparison criteria
and OODA, RPDM, and Rasmussen models values are presented by Grant et al. [11].
Table 2. Comparing proposed AIS, OODA, RPDM, and Rasmussen models
Criteria/Process Model
Control Loop
Detailed
Tempo (fast decision making)
Planning
Learning
OODA
√
×
√
×
×
Rasmussen
√
×
×
√
×
RPDM
√
√
×
×
×
AIS Model
√
√
×
√
√
Formalizing the process model by walking through use-case. Use case (8)
represents failing to response to pathogen cells. To formalize the process model, steps
of the use case are presented as follows (see Fig. 3):
1. Antibody cell examines other cells looking for pathogens.
2. Antibody cell detects that cell is non-self.
3. Antibody cell tries to find proper receptors to kill the pathogen.
4. Antibody cell failed to response.
5. Antibody cell activates the adaptive system to generate proper receptors.
6. Adaptive cells mutate to match the pathogen and generate required receptors.
7. Required receptors are cloned and response is provided.
8. Failed response is reported as the pathogen is detected again which backs to
step 1.
Fig. 3. Walk through use case (8), failed response to pathogen
Mapping AIS Operational Model to Crisis Response Multi-Agent Model
After definition of the conceptual model of the AIS, mapping the proposed model to
multi-agent model for crisis response is a straight forward process (see Tables 3 and
4). An agent can examine environment searching for danger sources or undesired
feature. This can be mapped to operational agents. When agent finds an un-desired
situation in environment, agent tries to handle the problem using available procedures.
If the situation exceeds the available routine procedures, agents ask for help from
tactical agents. Tactical agents check available memory for similar situations and
check if the old experienced situation can be adopted for the current situation or not.
If an old situation matches the current state of the environment, apply the course of
actions coupled with the experienced situation. Otherwise, tactical agent asks for
decision making (strategic) agent help. Using nearest matched situation, decision
making agent mutates different course of actions to handle the current situation till
reach an acceptable course of actions (intensity-based learning [25]). Decision making
agent allocates course of actions to tactical agents to be deployed (cloned). During the
execution of actions, operational agents report status of tasks execution, and in case of
failed task re-planning is presented. Finally, generation and death of agents are related
to the application domain of the proposed model. In crisis response death of cells and
generation of new cells can be mapped to deployment of effective actions and
neglecting others, or removal and adding new responders to the response field.
Table 3. Mapping AIS operational model to crisis response multi-agent model
Biological Immune System Level
Cells
Helper Cells
Cells mutation
Memory Cells
Crisis Response Domain
Operational agent
Tactical agents
Decision making agents
Case Memory
Table 4. Mapping AIS Operational Model Processes to agents' roles
Biological Immune System
Process
Self or non-self discrimination
Matching existing receptors
Identify new receptors
Clone required receptors
Response
Cells generation and death
Crisis Domain
Agent Role
Reporting un-desired situation
Situation Recognition
Planning (Situation Assessment)
Allocation of plan tasks
Executing plan tasks
Deployment of effective actions
and neglecting non-effective
actions
Operational
Tactical
Decision making
Tactical
Operational
3 Crisis Response to Pandemic Influenza in Egypt
This section presents the design, and implementation of the proposed response model
for pandemic influenza in Egypt. Agent environment includes two parts: the pandemic
model which is implemented in the previous work [19], and the available resources
(control strategies or actions). Agent retrieves the current pandemic situation based on
the agents' health states. Each control strategy is represented by (resource type,
amount, cost, from/to date, and efficiency [19]). Agents' roles argue that each agent
has its own profile and roles which specify its responsibilities and skills. For example
decision making agent has the role of making decisions and adding new memory
cases, while tactical agent has the role of processing information and allocating tasks
to operational agents. Table 5 shows different agent roles, role responsibilities, and
number of agents allowed per emergency operations center. Added here a new role
titled tactical communication agent. Actually tactical communication agent is a
tactical agent which is specialized for managing communication among the EOC
parties. Tactical communication agent has to receive reports from operation agents,
send reports to other tactical agents, deliver reports to crisis decision maker, and
deliver plans back to the tactical agents.
Crisis decision making agent follows the BDI (Belief-Desire-Intention) cognitive
model [27]. While, tactical and operational agents are recognized as reactive helper
agents for the decision making agents, and do not have believes nor desires. Decision
making agents need to work with other agents in EOC through collaboration to deliver
resources information and to deploy actions. Decision making agents need to make
decisions to control the spread of pandemic influenza using available effective
resources. Decision making agents need to determine the course of actions to be
deployed to control the pandemic.
Agent Role
Decision
Making Agent
Tactical
Communication
Agent
Tactical Agent
Operational
Agent
Table 5. Agent roles and number of agents per EOC
Responsibilities
Providing course of actions
Adding new cases to memory
Receiving reports from operational agents
Receiving resources reports from tactical agents
Sending reports to decision making agent
Receiving plan from decision making agent
Assigning plan tasks to tactical agents
Receiving status of plan execution
Sending plans status to decision making agent
Handling reports from communication agent
Processing available resources
Reporting current environment state
Deployment of course of actions
# of Agents
1
1
1 or more
1 or more
Design of the AIS Planning Methodology in the Proposed Response Model
Design of AIS planning methodology follows Stepney et al. [28] structure of AIS
engineering. Pandemic situation is represented as a record of the total number of
agents based on agents' health states. For example, situation can be represented by:
(Susceptible: 50 agents, In-Contact: 10 agents, Infectious: 20 agents, Isolated
Infected: 10 agents, Recovered: 1 agent, and Dead: 3 agents). The required course of
actions to control given pandemic situation is represented as follows: (identifier,
successfulness of the course of actions, and current pandemic situation). In addition
the course of actions is coupled with deployed actions. The entry of course of actions
and its coupled actions constructs a memory case.
The city block distance is used here to find the similarity between situations. For
example: distance between situation 1 (Infectious: 2 agents, Isolated Infectious: 1
10
0
33
31
20
=−
+−+−
agent) and situation 2 (Immunized: 31 agents) is:
.
Immune algorithms present the mutation (dynamic) behavior of AIS. Bone marrow
algorithm, positive selection algorithm, clonal selection is used in the proposed model
implementation.
4 Experiments
The main goal of experiments scenarios is to validate the proposed response model.
Scenarios basically include simulation of pandemic influenza in a closed population
of 1000 agents and initially three infected agents located in Cairo. Cairo EOC
contains 3 operational agents, 2 tactical agents, and one decision making agent. The
duration of the simulation round is 50 days. All simulation rounds involve randomly
generated resources pool. Each action in the resource pool has efficacy of (0.75). The
basic simulation round with no control strategies gives pandemic peak on day 10 with
%60.8 infected agents [19].
Fig. 4 shows the flow of control during a simulation round, and the total cost of
deployed actions (total cost = 2821.4 and plan certainty = 0 due to there are no
previous plans in the system). While, Fig. 5 shows the stored case memory for the
finished simulation round (case id = 174, successfulness = 0.001198). It is found that
the pandemic peak is shifted to day 16 with %55 infected agents of the population.
Fig. 4. Round 1 - Crisis Response Log
Fig. 5. Round 1 - Case Memory
5 Conclusions
Currently, multi-agent architecture is the essence of response systems. The original
idea comes out from agent characteristics in MAS, such as autonomy, local view of
environment, capability of learning, planning, coordination and decentralized decision
making. The incorporation of multi-agent systems can be clarified by discussing
major disciplines involved during crisis response operations; situation management,
and decision making and planning.
The study of biological systems is of interest to scientists and engineers as they
turn out to be a source of rich theories. They are useful in constructing novel
computer algorithms to solve complex engineering problems. Immunology as a study
of the immune system inspired the evolution of artificial immune system, which is an
area of vast research over the last few years. Detecting non-self, matching receptors,
clone antibodies, response then detecting non-self cells again represents the control
loop in the AIS metaphor. This process loop (control loop) is the core of the decision
making process in AIS metaphor. The AIS model allows learning by mutating genes
to generate acceptable receptors and store them in memory cells. Selection of required
antibodies to be cloned and the cloning process represents the planning methodology
of AIS. Artificial immune systems represent an interesting metaphor for building
effective based and defensive multi-agent crisis response operations model. The
proposed architecture proposed by AIS, promises effective operations and system
architecture for crisis response.
According to experiments scenarios, AIS model shows slow learning process but it
is very fast in handling desired or un-desired situations. Number of cases represents
the growth of the system. While, scenarios results show that the effectiveness of
response operations and utilization of resources are improved within the growth of the
response model by ignoring low successfulness memory cases while deliberating new
course of actions. The simulation process is very slow and consumes a lot of
computation power. Each round takes at least 10 minutes to complete. It is
recommended to implement the model using high performance computing to enable
fast growth of case memory.
References
1. Brownlee, J.: A Hierarchical Framework of the Acquired Immune System. Technical
Report, Victoria, Australia: Complex Intelligent Systems Laboratory (CIS), Swinburne
University of Technology (2007).
2. Builder, C. H., Steven, C., Bankes, Nordin, R.: Command Concepts A Theory Derived
from the Practice of Command and Control. RAND Corporation's National Defense
Research Institute, ISBN/EAN: 0-8330-2450-7 (1999).
3. Chen, R., Sharman, R., Rao, H. R., and Upadhyaya, S.: Design Principles for Critical
Incident Response Systems. In Journal of Information Systems and E-Business
Management, Volume 5, Number 3, pp. 201-227 (2007).
4. Dasgupta, D., and Forrest, S.: Novelty detection in time series data using ideas from
immunology. In ISCA 5th International Conference on Intelligent Systems (1996).
5. Dasgupta, D., Attoh-Okine, N.: Immunity-Based Systems: A Survey. In the proceedings of
the IEEE International Conference on Systems, Man and Cybernetics, Orlando, Florida,
Volume 1, pp. 369-374 (1997).
6. Dasgupta, D.: Immune-based intrusion detection system: A general framework. In
proceedings of the National Information Systems Security Conference, pp.147–160
(1999).
7. Dasgupta, D., Krishna, K. K., Wong, D., and Berry, M.: Negative selection algorithm for
aircraft fault detection. In proceedings of the Third International Conference, ICARIS2004
on Artificial Immune Systems, pp. 1-13 (2004).
8. Davia, G.A.: Thoughts on a Possible Rational Reconstruction of the Method of "Rational
Reconstruction". In Twentieth World Congress of Philosophy Boston, Massachusetts
U.S.A., pp. 10-15 (1998).
9. Golzari, S., Doraisamy, S., Sulaiman, M. N. B., and Udzir, N. L.: A Review on Concepts,
Algorithms and Recognition-Based Applications of Artificial Immune System. Springer-
Verlag Berlin Heidelberg, CCIS 6, pp. 569–576 (2008).
10. Gonzalz, F., and Dasgupta, D.: Anomaly detection using real-valued negative
selection.Genetic Programming and Evolvable Machines, pp. 383-403 (2004).
11. Grant, T.J., and Kooter, B.M.: Comparing OODA and Other Models as Operational View
C2 Architecture. In Proceedings, 10th International Command and Control Research and
Technology Symposium (ICCRTS 2005), Washington DC, USA (2005).
12. Grant, T.: Unifying Planning and Control using an OODA-based Architecture. In
Proceedings of Annual Conference of South African Institute of Computer Scientists and
Information Technologists (SAICSIT), White River, Mpumalanga, South Africa, pp. 111 –
113 (2005).
13. Hofmeyr, S.A.: An interpretative introduction to the immune system. In Design Principles
for the Immune System and other Distributed Autonomous Systems, edited by Cohen I.,
and Segel L.A., Oxford University Press: New York (2000).
14. Hunt, J.E., and Cooke, D.E.: Learning using an artificial immune system. Journal of
Network Computing Applications, Volume19, pp. 189-212 (1996).
15. Integrated DEFinition
(IDEF) methods
IDEF0 – Function Modeling Method,
http://www.idef.com/IDEF0.htm.
16. Khalil, K.M., Abdel-Aziz, M. H., Nazmy, M. T., Salem, A. M.: The Role of Artificial
Intelligence Technologies in Crisis Response. Mendel Conference, 14th International
Conference on Soft Computing, pp. 293-298 (2008).
17. Khalil, K.M., Abdel-Aziz, M. H., Nazmy, M. T., Salem, A. M.: Multi-Agent Crisis
Response systems – Design Requirements and Analysis of Current Systems. Fourth
International Conference on Intelligence Computing and Information Systems, Cairo,
Egypt, pp. 920-925 (2009).
18. Khalil, K.M., Abdel-Aziz, M. H., Nazmy, M. T., Salem, A. M.: Bridging the Gap between
Crisis Response Operations and Systems. Annals of University of Craiova, Mathematics
and Computer Science Series, Volume 36, Number 2, ISSN: 1223-6934, pp. 141 -145
(2009).
19. Khalil, K.M., Abdel-Aziz, M. H., Nazmy, M. T., Salem, A. M.: An agent-based modeling
for pandemic influenza in Egypt. INFOS2010: 7th International Conference on Informatics
and Systems, Cairo, Egypt, AOM 24-30 (2010).
20. Kim, J., and Bentley, P.: Toward an artificial immune system for network intrusion
detection: An investigation of dynamic clonal selection. In Proceedings of the 2002
Congress on Evolutionary Computation, Honolulu, Hawaii, pp. 1244-1252 (2002).
21. Knight, T., and Timmis, J.: AINE: An immunological approach to data mining. IEEE
International Conference on Data Mining, pp. 297-304 (2001).
22. Krishnakumar, K.: Artificial Immune System Approaches for Aerospace Applications.
Technical Report, AIAA (2003).
23. Kubi, J.: Immunology. W H Freeman, Fifth Edition (2002).
24. Krishna, K.K., and Neidhoefer, J.: Immunized adaptive critic for an autonomous aircraft
control application. Springer-Verlag, Inc., Chapter 20 in the book entitled Artificial
Immune Systems and Their Applications, pp. 221-240 (1999).
25. Lewis, L.: Managing Computer Networks: A Case-Based Reasoning Approach. Artech
House (1995).
26. Perelson, A.S., and Weisbuch, G.: Immunology for physicists Reviews of Modern Physics.
Volume 69, pp. 1219-1267 (1997).
27. Shendarkar, A., Vasudevan, K., Lee, S., Son, Y.: Crowd simulation for emergency
response using BDI agent based on virtual reality. In WSC '06: Proceedings of the 38th
conference on Winter simulation, pp. 545-553 (2006).
28. Stepney, S., Smith, R., Timmis, Jon, and Tyrrell, A.: Towards a Conceptual Framework
for Artificial Immune Systems. In the proceedings of the CARIS 2004: Third International
Conference on Artificial Immune Systems, Catania, Sicily, Volume 3239 of Lecture Notes
in Computer Science, Springer, pp 53-64 (2004).
29. Twycross, J., Aickelin, U.: Information Fusion in the Immune System. Information
Fusion, Volume 11, Issue 1, pp. 35-44 (2010).
30. Weiss, G.: Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence.
MIT Press, pp. 1–23 (2000).
|
1503.07220 | 2 | 1503 | 2015-04-02T22:58:57 | Individual Planning in Agent Populations: Exploiting Anonymity and Frame-Action Hypergraphs | [
"cs.MA",
"cs.AI",
"cs.GT"
] | Interactive partially observable Markov decision processes (I-POMDP) provide a formal framework for planning for a self-interested agent in multiagent settings. An agent operating in a multiagent environment must deliberate about the actions that other agents may take and the effect these actions have on the environment and the rewards it receives. Traditional I-POMDPs model this dependence on the actions of other agents using joint action and model spaces. Therefore, the solution complexity grows exponentially with the number of agents thereby complicating scalability. In this paper, we model and extend anonymity and context-specific independence -- problem structures often present in agent populations -- for computational gain. We empirically demonstrate the efficiency from exploiting these problem structures by solving a new multiagent problem involving more than 1,000 agents. | cs.MA | cs |
Individual Planning in Agent Populations: Exploiting Anonymity and
Frame-Action Hypergraphs
Ekhlas Sonu, Yingke Chen and Prashant Doshi
THINC lab, Dept. of Computer Science
University of Georgia
Athens, Georgia 30602
{esonu,ykchen}@uga.edu, [email protected]
Abstract
Interactive partially observable Markov decision processes
(I-POMDP) provide a formal framework for planning for a
self-interested agent in multiagent settings. An agent operat-
ing in a multiagent environment must deliberate about the ac-
tions that other agents may take and the effect these actions
have on the environment and the rewards it receives. Tradi-
tional I-POMDPs model this dependence on the actions of
other agents using joint action and model spaces. Therefore,
the solution complexity grows exponentially with the num-
ber of agents thereby complicating scalability. In this paper,
we model and extend anonymity and context-specific indepen-
dence -- problem structures often present in agent populations
-- for computational gain. We empirically demonstrate the ef-
ficiency from exploiting these problem structures by solving
a new multiagent problem involving more than 1,000 agents.
Introduction
We focus on the decision-making problem of an in-
dividual
agent operating in the presence of other
self-interested agents whose actions may affect
the
the environment and the subject agent's re-
state of
In stochastic and partially observable environ-
wards.
ments,
this problem is formalized by the interactive
POMDP (I-POMDP)
(Gmytrasiewicz and Doshi 2005).
I-POMDPs cover an important portion of the multiagent
planning problem space
(Seuken and Zilberstein 2008;
Doshi 2012), and applications in diverse areas such as
security
Seymour and Peterson 2009),
Woodward and Wood 2012),
robotics
ad
and
human
(Doshi et al. 2010;
behavior
Wunder et al. 2011)
testify to its wide appeal while
critically motivating better scalability.
(Ng et al. 2010;
(Wang 2013;
teams
(Chandrasekaran et al. 2014)
modeling
hoc
value
iteration
Previous I-POMDP solution approximations such as in-
teractive particle filtering (Doshi and Gmytrasiewicz 2009),
(Doshi and Perez 2008)
point-based
and
policy
(I-
interactive
(Sonu and Doshi 2014) scale I-POMDP solutions
BPI)
to larger physical state, observation and model spaces.
introduced the specialized I-
Hoang and Low (2013)
POMDP Lite framework that promotes efficiency by
bounded
iteration
Copyright c(cid:13) 2015, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
modeling other agents as nested MDPs. However,
to
the best of our knowledge no effort specifically scales
I-POMDPs to many interacting agents -- say, a population
of more than a thousand -- sharing the environment.
For illustration, consider the decision-making problem of
the police when faced with a large protest. The degree of the
police response is often decided by how many protestors of
which type (disruptive or not) are participating. The individ-
ual identity of the protestor within each type seldom matters.
This key observation of frame-action anonymity motivates
us in how we model the agent population in the planning
process. Furthermore, the planned degree of response at a
protest site is influenced, in part, by how many disruptive
protestors are predicted to converge at the site and much less
by some other actions of protestors such as movement be-
tween other distant sites. Therefore, police actions depend
on just a few actions of note for each type of agent.
The example above illustrates two known and power-
ful types of problem structure in domains involving many
agents: action anonymity (Roughgarden and Tardos 2002)
and context-specific independence (Boutilier et al. 1996).
Action anonymity allows the exponentially large joint ac-
tion space to be substituted with a much more compact
space of action configurations where a configuration is a tu-
ple representing the number of agents performing each ac-
tion. Context-specific independence (wherein given a con-
text such as the state and agent's own action, not all actions
performed by other agents are relevant) permits the space
of configurations to be compressed by projecting counts
over a limited set of others' actions. We extend both action
anonymity and context-specific independence to allow con-
siderations of an agent's frame as well. 1 We list the specific
contributions of this paper below:
1. I-POMDPs are severely challenged by large numbers
of agents sharing the environment, which cause an ex-
ponential growth in the space of joint models and ac-
tions. Exploiting problem structure in the form of frame-
action anonymity and context-specific independence, we
present a new method for considerably scaling the so-
lution of I-POMDPs to an unprecedented number of
1I-POMDPs distinguish between an agent's frame and type
with the latter including beliefs as well. Frames are similar in se-
mantics to the colloquial use of types.
agents.
2. We present a systematic way of modeling the prob-
lem structure in transition, observation and reward func-
tions, and integrating it in a simple method for solving I-
POMDPs that models other agents using finite-state ma-
chines and builds reachability trees given an initial belief.
3. We prove that the Bellman equation modified to include
action configurations and frame-action independences
continues to remain optimal given the I-POMDP with
explicated problem structure.
4. Finally, we theoretically verify the improved savings in
computational time and memory, and empirically demon-
strate it on a new problem of policing protest with over a
thousand protestors.
The above problem structure allows us to emphatically
mitigate the curse of dimensionality whose acute impact on
I-POMDPs is well known. However, it does not lessen the
impact of the curse of history. In this context, an additional
step of sparse sampling of observations while generating the
reachability tree allows sophisticated planning with a popu-
lation of 1,000+ agents using about six hours.
Related Work
Building on graphical games (Kearns, Littman, and Singh 2001),
action graph games (AGG) (Jiang, Leyton-Brown, and Bhat 2011)
utilize problem structures such as action anonymity and
context-specific independence to concisely represent single
shot complete-information games involving multiple agents
and to scalably solve for Nash equilibrium. The indepen-
dence is modeled using a directed action graph whose nodes
are actions and an edge between two nodes indicates that
the reward of an agent performing an action indicated by
one node is affected by other agents performing action
of the other node. Lack of edges between nodes encodes
the context-specific independence where the context is the
action. Action anonymity is useful when the action sets of
agents overlap substantially. Subsequently, the vector of
counts over the set of distinct actions, called a configuration,
is much smaller than the space of action profiles.
We substantially build on AGGs in this paper by ex-
tending anonymity and context-specific independence to in-
clude agent frames, and generalizing their use to a par-
tially observable stochastic game solved using decision-
theoretic planning as formalized by I-POMDPs. Indeed,
Bayesian AGGs (Jiang and Leyton-Brown 2010) extend the
original formulation to include agent types. These result
in type-specific action sets with the benefit that the ac-
tion graph structure does not change although the number
of nodes grows with types: ΘA nodes for agents with
Θ types each having same A actions. If two actions
from different type-action sets share a node, then these ac-
tions are interchangeable. A key difference in our repre-
sentation is that we explicitly model frames in the graphs
due to which context-specific independence is modeled us-
ing frame-action hypergraphs. Benefits are that we natu-
rally maintain the distinction between two similar actions
but performed by agents of different frames, and we add
less additional nodes: Θ + A. However, a hypergraph
is a more complex data structure for operation. Tempo-
ral AGGs (Jiang, Leyton-Brown, and Pfeffer 2009) extend
AGGs to a repeated game setting and allow decisions to con-
dition on chance nodes. These nodes may represent the ac-
tion counts from previous step (similar to observing the ac-
tions in the previous game). Temporal AGGs come closest to
multiagent influence diagrams (Koller and Milch 2001) al-
though they can additionally model the anonymity and inde-
pendence structure. Overall, I-POMDPs with frame-action
anonymity and context-specific independence significantly
augment the combination of Bayesian and temporal AGGs
by utilizing the structures in a partially observable stochastic
game setting with agent types.
al.
Varakantham et
(2014) building on previous
work (Varakantham et al. 2012) recently introduced a de-
centralized MDP that models a simple form of anonymous
interactions: rewards and transition probabilities specific to
a state-action pair are affected by the number of other agents
regardless of their identities. The interaction influence is
not further detailed into which actions of other agents are
relevant (as in action anonymity) and thus configurations
and hypergraphs are not used. Furthermore, agent types
are not considered. Finally, the interaction hypergraphs in
networked-distributed POMDPs (Nair et al. 2005) model
complete reward independence between agents -- analogous
to graphical games -- which differs from the hypergraphs in
this paper (and action graphs) that model independence in
reward (and transition, observation probabilities) along a
different dimension: actions.
Background
Interactive POMDPs allow a self-interested agent to plan
individually in a partially observable stochastic environ-
ment in the presence of other agents of uncertain types. We
briefly review the I-POMDP framework and refer the reader
to (Gmytrasiewicz and Doshi 2005) for further details.
A finitely-nested interactive I-POMDP for an agent (say
agent 0) of strategy level l operating in a setting inhabited
by one of more other interacting agents is defined as the fol-
lowing tuple:
I-POMDP0,l = hIS0,l, A, T0, Ω0, O0, R0, OC0i
• IS0,l denotes the set of interactive states defined as,
IS0,l = S × QN
j=1 Mj,l−1, where Mj,l−1 = {Θj,l−1 ∪
SMj}, for l ≥ 1, and ISi,0 = S, where S is the set
of physical states. Θj,l−1 is the set of computable, inten-
tional models ascribed to agent j: θj,l−1 = hbj,l−1, θji,
where bj,l−1 is agent j's level l − 1 belief, bj,l−1 ∈
△
△(ISj,l−1), and θj
= hA, Tj, Ωj, Oj, Rj, OCj i, is j's
frame. Here, j is assumed to be Bayes-rational. At level 0,
bj,0 ∈ △(S) and a level-0 intentional model reduces to a
POMDP. SMj is the set of subintentional models of j, an
example is a finite state automaton;
• A = A0 × A1 × . . . × AN is the set of joint actions of all
agents;
• T0 : S × A0 × QN
function;
j=1 Aj × S → [0, 1] is the transition
• Ω0 is the set of agent 0's observations;
j=1 Aj × Ω0 → [0, 1] is the observation
function;
• O0 : S × A0 ×QN
• R0 : S × A0 ×QN
j=1 Aj → R is the reward function; and
• OC0 is the optimality criterion, which is identical to that
for POMDPs. In this paper, we consider a finite-horizon
optimality criteria.
Besides the physical state space, the I-POMDP's interac-
tive state space contains all possible models of other agents.
In its belief update, an agent has to update its belief about
the other agents' models based on an estimation about the
other agents' observations and how they update their models.
As the number of agents sharing the environment grows, the
size of the joint action and joint model spaces increases expo-
nentially. Therefore, the memory requirement for represent-
ing the transition, observation and reward functions grows
exponentially as well as the complexity of performing belief
update over the interactive states. In the context of N agents,
interactive bounded policy iteration (Sonu and Doshi 2014)
generates good quality solutions for an agent interacting
with 4 other agents (total of 5 agents) absent any problem
structure. To the best of our knowledge, this result illustrates
the best scalability so far to N > 2 agents.
Many-Agent I-POMDP
To facilitate understanding and experimentation, we intro-
duce a pragmatic running example that also forms our eval-
uation domain.
Figure 1: Protestors of different frames (colors) and police troops
at two of three sites in the policing protest domain. The state space
of police decision making is factored into the protest intensity lev-
els at the sites.
Example 1 (Policing Protest) Consider a policing sce-
nario where police (agent 0) must maintain order in 3 geo-
graphically distributed and designated protest sites (labeled
0, 1, and 2) as shown in Fig. 1. A population of N agents
are protesting at these sites. Police may dispatch one or two
riot-control troops to either the same or different locations.
Protests with differing intensities, low, medium and high (dis-
ruptive), occur at each of the three sites. The goal of the
police is to deescalate protests to the low intensity at each
site. Protest intensity at any site is influenced by the number
of protestors and the number of police troops at that loca-
tion. In the absence of adequate policing, we presume that
the protest intensity escalates. On the other hand, two police
troops at a location are adequate for de-escalating protests.
Factored Beliefs and Update
joint models of other agents, b0,l ∈ ∆(S × QN
As we mentioned previously, the subject agent in an I-
POMDP maintains a belief over the physical state and
j=1 Mj,l−1),
where ∆(·) is the space of probability distributions. For
settings such as Example 1 where N is large, the size of
the interactive state space is exponentially larger, IS0,l =
SMj,l−1N , and the belief representation unwieldy. How-
ever, the representation becomes manageable for large N if
the belief is factored:
b0,l(s, m1,l−1, m2,l−1, . . . , mN,l−1) = P r(s) P r(m1,l−1s)
(1)
× P r(m2,l−1s) × . . . × P r(mN,l−1s)
This factorization assumes conditional
independence of
models of different agents given the physical state. Conse-
quently, beliefs that correlate agents may not be directly rep-
resented although correlation could be alternately supported
by introducing models with a correlating device.
j ), where M ∗
The memory consumed in storing a factored belief is
j is the size of the largest
O(S + N SM ∗
model space among all other agents. This is linear in the
number of agents, which is much less than the exponentially
growing memory required to represent the belief as a joint
j N ).
distribution over the interactive state space, O(SM ∗
0 and the
, the updated belief at
Given agent 0's belief at time t, bt
0,l, its action at
subsequent observation it makes, ωt+1
time step t + 1, bt+1
0,l , may be obtained as:
0
P r(st+1, mt+1
bt
0,l, at
bt
0,l, at
0, ωt+1
0, ωt+1
N,l−1bt
0,l, at
1,l−1st+1, mt+1
1,l−1, . . . , mt+1
) P r(mt+1
) × . . . × P r(mt+1
0
0
2,l−1, . . . , mt+1
N,l−1st+1, bt
0,l, at
0, ωt+1
0
) = P r(st+1
N,l−1,
0, ωt+1
0
)
(2)
Each factor in the product of Eq. 2 may be obtained as
follows. The update over the physical state is:
P r(st+1bt
) ∝ P r(st+1, ωt+1
bt
0
1,l−1st) × . . . × bt
N,l−1st)
0,l, at
0)
0,l(mt
0, ωt+1
0
bt
0,l(mt
bt
0,l, at
0,l(st) Xmt
1mt
−0
P r(at
= Xst
×Xat
−0
1,l−1) × . . . × P r(at
N mt
N,l−1)
× Ot+1
0
(st+1, hat
0, at
−0i, ωt+1
0
) T0(st, hat
0, at
−0i, st+1)
(3)
and the update over the model of each other agent, j =
P r(mt+1
1 . . . N , conditioned on the state at t + 1 is:
N,l−1, bt
1,l−1st) × . . . × bt
j+1,l−1, . . . , mt+1
bt
0,l(mt
bt
j,l−1st+1, mt+1
0(st) Xmt
1mt
−j,l−1
P r(at
1,l−1) × . . . × P r(at
nmt
N,l−1)
0,l, at
0, ωt+1
0,l(mt
) =
0
N,l−1st)
Xst
Xat
Xωt+1
−j
j
Oj(st+1, haj , at
−ji, ωt+1
j
) P r(mt+1
j
mt
j, at
j, ωt+1
j
)
(4)
Derivations of Eqs. 3 and 4 are straightforward and not given
here due to lack of space. In particular, note that models of
agents other than j at t+1 do not impact j's model update in
the absence of correlated behavior. Thus, under the assump-
tion of a factored prior as in Eq. 1 and absence of agent cor-
relations, the I-POMDP belief update may be decomposed
into an update of the physical state and update of the models
of N agents conditioned on the state.
Frame-Action Anonymity
As noted by Jiang et al. (2011), many noncooperative and co-
operative problems exhibit the structure that rewards depend
on the number of agents acting in particular ways rather than
which agent is performing the act. This is particularly evi-
dent in Example 1 where the outcome of policing largely de-
pends on the number of protestors that are peaceful and the
number that are disruptive. Building on this, we additionally
observe that the transient state of the protests and observa-
tions of the police at a site are also largely influenced by the
number of peaceful and disruptive protestors moving from
one location to another. This is noted in the example below:
Example 2 (Frame-action anonymity of protestors) The
transient state of protests reflecting the intensity of protests
at each site depends on the previous intensity at a site and
the number of peaceful and disruptive protestors entering
the site. Police (noisily) observes the intensity of protest at
each site which is again largely determined by the number
of peaceful and disruptive protestors at a site. Finally, the
outcome of policing at a site is contingent on whether the
protest is largely peaceful or disruptive. Consequently, the
identity of the individual protestors beyond their frame and
action is disregarded.
Here, peaceful and disruptive are different frames of others
in agent 0's I-POMDP, and the above definition may be ex-
tended to any number of frames. Frame-action anonymity is
an important attribute of the above domain. We formally de-
fine it in the context of agent 0's transition, observation and
reward functions next:
Definition 1 (Frame-action anonymity) Let ap
−0 be a joint
−0 be a joint action of
action of all peaceful protestors and ad
all disruptive ones. Let ap
−0 be permutations of the
two joint action profiles, respectively. An I-POMDP models
frame-action anonymity iff for any a0, s, s′, ap
T0(s, a0, ap
−0, s′) = T0(s, a0, ap
−0 and ad
−0 and ad
−0, s′),
−0, ad
−0, ad
−0:
−0, ad
−0, ad
−0, ad
−0, ad
−0) ∀ ap
−0, ω0), and
−0.
−0, ω0) = O0(s′, a0, ap
−0) = R0(s, a0, ap
−0, ad
O0(s′, a0, ap
R0(s, a0, ap
Recall the definition of an action configuration, C, as the
vector of action counts of an agent population. A permu-
tation of joint actions of others, say ap
−0, assigns different
actions to individual agents. Despite this, the fact that the
transition and observation probabilities, and the reward re-
mains unchanged indicates that the identity of the agent per-
forming the action is irrelevant. Importantly, the configura-
tion of the joint action and its permutation stays the same:
C(ap
−0). This combined with Def. 1 allows re-
defining the transition, observation and reward functions
to be over configurations as: T0(s, a0, C(ap
−0), s′),
O0(s′, a0, C(ap
−0), C(ad
−0
)).
−0), o) and R0(s, a0, C(ap
−0) = C( ap
−0), C(ad
−0), C(ad
Let Ap
1, . . . , Ap
−0. Observe that multiple joint actions from Ap
n be overlapping sets of actions of n peace-
ful protestors, and Ap
−0 is the Cartesian product of these
sets. Let C(Ap
−0) be the set of all action configurations for
Ap
−0 may re-
sult in a single configuration; these joint actions are config-
uration equivalent. Consequently, the equivalence partitions
the joint action set Ap
−0) classes. Furthermore,
when other agents of same frame have overlapping sets of ac-
tions, the number of configurations could be much smaller
than the number of joint actions. Therefore, definitions of
the transition, observation and reward functions involving
configurations could be more compact.
−0 into C(Ap
Frame-Action Hypergraphs
In addition to frame-action anonymity, domains involving
agent populations often exhibit context-specific indepen-
dences. This is a broad category and includes the context-
specific independence found in conditional probability ta-
bles of Bayesian networks (Boutilier et al. 1996) and in
action-graph games. It offers significant additional structure
for computational tractability. We begin by illustrating this
in the context of Example 1.
Example 3 (Context-specific independence in policing)
At a low intensity protest site, reward for the police on
passive policing is independent of the movement of the
protestors to other sites. The transient intensity of
the
protest at a site given the level of policing at the site
(context) is independent of
the movement of protestors
between other sites.
The context-specific independence above builds on the
similar independence in action graphs in two ways: (i) We
model such partial independence in the transitions of fac-
tored states and in the observation function as well, in ad-
dition to the reward function. (ii) We allow the context-
specific independence to be mediated by the frames of other
agents in addition to their actions. For example, the rewards
received from policing a site is independent of the number of
protestors at another site, instead the rewards are influenced
by the number of peaceful and disruptive protestors present
at that site.
The latter difference generalizes the action graphs into
frame-action hypergraphs, and specifically 3-uniform hyper-
graph indicates that the reward for agent 0 on performing
action a0 at state s is affected (in part) by the agents of
frame θ−0 who perform action in A−0.
We illustrate a general frame-action hypergraph for context-
specific independence in a transition function and a reward
function as bipartite Levi graphs in Figs. 2(a) and (b), re-
spectively. We point out that the hypergraph for the reward
function comes closest in semantics to the graph in ac-
tion graph games (Jiang, Leyton-Brown, and Bhat 2011) al-
though the former adds the state to the context and frames.
Hypergraphs for the transition and observation functions dif-
fer substantially in semantics and form from action graphs.
To use these hypergraphs in our algorithms, we first define
the general frame-action neighborhood of a context node.
Definition 3 (Frame-action neighborhood) The
frame-
action neighborhood of a context node ψ ∈ Ψ, ν(ψ),
given a frame-action hypergraph G is defined as a subset
of A × Θ such that ν(ψ) = {(a−0, θ)a−0 ∈ A−0, θ ∈
Θ, (ψ, a−0, θ) ∈ E}.
As an example, the frame-action neighborhood of a state-
action pair, hs, a0i in a hypergraph for the reward function
is the set of all action and frame nodes incident on each hy-
peredge anchored by the node hs, a0i.
We move toward integrating frame-action anonymity in-
troduced in the previous subsection with the context-specific
independence as modeled above by introducing frame-
action configurations.
Definition 4 (Frame-action configuration) A configura-
tion over the frame-action neighborhood of a context node,
ψ, given a frame-action hypergraph is a vector,
C ν(ψ) △
= h C(A
θ1
−0), C(A
θ2
−0), . . . , C(A
θ Θ
−0 ), C(φ) i
θ
where each a included in A
−0 is an action in ν(ψ) with
θ
frame θ, and C(A
−0) is a configuration over actions by
agents other than 0 whose frame is θ. All agents with frames
other than those in the frame-action neighborhood are as-
sumed to perform a dummy action, φ.
Definition 4 allows further inroads into compacting
the transition, observation and rewards functions of the
I-POMDP using context-specific independence. Specifi-
cally, we may redefine these functions one more time
to limit
the configurations only over the frame-action
neighborhood of the context as, T0(s, a0, C ν(s,a0,s′), s′),
O0(s′, a0, C ν(s′,a0,ω0), ω0) and R0(s, a0, C ν(s,a0)). 2
Revised Framework
To benefit from structures of anonymity and context-specific
independence, we redefine I-POMDP for agent 0 as:
I-POMDP0,l = hIS0,l, A, Ω0, T0, O0, R0, OC0i
where:
2Context in our transition function is hs, a0, s′i compared with
the context of just hs, a0i in Varakantham et al's (2014) transitions.
(a)
(b)
Figure 2: Levi (incidence) graph representation of a generic frame-
action hypergraph for (a) the transition function, and (b) the re-
ward function. The shaded nodes represent edges in the hypergraph.
Each edge has the context, ψ, denoted in bold, agent's action, a,
and its frame, θ, incident on it. For example, the reward for a state
and agent 0's action, hs, a0i1 is affected by others' actions a1
j and
j performed by any other agent of frame θ1
a2
j only.
graphs where each edge is a set of 3 nodes. We formally
define it below:
Definition 2 (Frame-action hypergraph) A frame-action
hypergraph for agent 0 is a 3-uniform hypergraph G =
hΨ, A−0, Θ−0, Ei, where Ψ is a set of nodes that represent
the context, A−0 is a set of action nodes with each node rep-
resenting an action that any other agent may take; Θ−0 is a
set of frame nodes, each node representing a frame ascribed
to an agent, and E is a 3-uniform hyperedge containing one
node from each set Ψ, A−0, and Θ−0, respectively.
Both context and action nodes differ based on whether the
hypergraph applies to the transition, observation or reward
functions:
• For the transition function, the context is the set of all
pairs of states between which a transition may occur and
each action of agent 0, Ψ = S × A0 × S, and the ac-
tion nodes includes actions of all other agents, A−0 =
SN
j=1 Aj. Neighbors of a context node ψ = hs, a0, s′i are
all the frame-action pairs that affect the probability of the
transition. An edge (h s, a0, s′i, a−0, θ) indicates that the
probability of transitioning from s to s′ on performing a0
is affected (in part) by the other agents of frame θ perform-
ing the particular action in A−0.
• The context for agent 0's observation function is the state-
action-observation triplet, Ψ = S × A0 × Ω0, and the
action nodes are identical to those in the transition func-
tion. Neighbors of a context node, hs, a0, ω0i, are all those
frame-action pairs that affect the observation probability.
Specifically, an edge (hs, a0, ω0i, a−0, θ) indicates that
the probability of observing ω0 from state s on perform-
ing a0 is affected (in part) by the other agents performing
action, a−0, who possess frame θ.
• For agent 0's reward function, the context is the set of
pairs of state and action of agent 0, Ψ = S × A0, and the
action nodes the same as those in transition and observa-
tion functions. An edge (h s, a0i, a−0, θ−0) in this hyper-
• IS0,l, A, Ω0 and OC0 remain the same as before. The
k=1 Xk.
physical states are factored as, S = QK
• T0 is the transition function, T0(x, a0, C ν(x,a0,x′), x′)
where C ν(x,a0,x′) is the configuration over the frame-
action neighborhood of context hx, a0, x′i obtained from
a hypergraph that holds for the transition function. This
transition function is significantly more compact than the
original that occupies space O(X2A0A−0N ) com-
ν∗ )ν∗) of T0, where the frac-
pared to the O(X2A0( N
ν∗+1 (cid:1), ν∗ is the maxi-
ν∗ (cid:1) is obtained from
mum cardinality of the neighborhood of any context, and
( N
combinatorial compositions and represents the number of
ways ν∗ + 1 non-negative values can be weakly com-
posed such that their sum is N .
tion is the complexity of (cid:0)N +ν∗+1
ν∗ )ν∗ ≪ A−0N . The value(cid:0)N +ν∗
• The
is
redefined
observation
is
function
O0(x′, a0, C ν(x′,a0,ω0), ω0) where C ν(x′,a0,ω0)
the
configuration over the frame-action neighborhood of
context hx′, a0, ω0i obtained from a hypergraph that
holds for the observation function. Analogously to the
transition function,
the original observation function
consumes space O(XΩA0A−0N ), which is much
ν∗ )ν∗) occupied by
larger than space O(XΩA0( N
this redefinition.
• R0 is the reward function defined as R0(x, a0, C ν(x,a0))
where C ν(x,a0) is defined analogously to the configu-
rations in the previous parameters. The reward for a
state and actions may simply be the sum of rewards
for the state factors and actions (or a more general
function if needed). As with the transition and observa-
tion functions, this reward function is compact occupy-
ν∗ )ν∗) that is much less than
ing space O(XA0( N
O(XA0A−0N ) of the original.
Belief Update For this extended I-POMDP, we compute
the updated belief over a physical state as a product of its
factors using Eq. 5 and belief update over the models of each
other agent using Eq 6 as shown below:
P r(st+1bt
0,l, at
0, ωt+1
0
) ∝ (cid:26)Xst
bt
0,l(st)
P r(C ν(xt+1
k
,at
0,ωt+1
0
)bt
0,l(M1,l−1st), . . . , bt
O0(xt+1
k
, at
0, C ν(xt+1
k
,at
0,ωt+1
0
), ωt+1
0
K
)
,ω
,at
0
t+1
k
t+1
0
Cν(x
Yk=1 X
0,l(MN,l−1st))
Yk=1
bt
0,l(st)
K
)(cid:27) ×(cid:26)Xst
X
,at
0 ,x
Cν(xt
k
P r(C ν(xt
k ,at
0,xt+1
k
)bt
0,l(M1,l−1st), . . . ,
t+1
k
)
0,l(MN,l−1st))T0(xt
bt
k, C ν(xt
k,at
0,xt+1
k
), xt+1
k
)(cid:27)
(5)
the term, P r(C ν(xt+1
Here,
0,l(M1,l−1st), . . . ,
0,l(MN,l−1st)), is the probability of a frame-action con-
bt
figuration (see Def. 4) that is context specific to the triplet,
0,ωt+1
)bt
,at
k
0
hxt+1, a0, ωt+1i. It is computed from the factored beliefs
over the models of all others. We discuss this computation
in the next section. The second configuration term has an
analogous meaning and is computed similarly.
The factored belief update over the models of each other
agent, j = 1 . . . N , conditioned on the state at t+1 becomes:
0(mt
bt
j
0(st)Xmt
j
bt
0) = Xst
P r(C ν(xt+1,at
j ,ωj)
P r(mt+1
j,l−1st+1, mt+1
−j,l−1, bt
0,l, at
j mt
P r(at
st)Xat
bt
0,l(M1,l−1st), . . . , bt
C
j) X
ν(xt+1 ,at
j
j
,ωj )
Oj(xt+1,
0,l(MN,l−1st))Xot+1
j, at
) P r(mt+1
mt
j
j
j
j
)
(6)
j, ωt+1
j ,ωj ), ωt+1
j, C ν(xt+1,at
at
Proofs for obtaining Eqs. 5 and 6 are omitted due to space
restrictions. Notice that the distributions over configurations
are computed using distributions over other agents' models.
Therefore, we must maintain and update conditional beliefs
over other agents' models. Hence, the problem cannot be re-
duced to a POMDP by including configurations with physi-
cal states.
Value Function The finite-horizon value function of the
many-agent I-POMDP continues to be the sum of agent 0's
immediate reward and the discounted expected reward over
the future:
V h(mt
0,l) = max
0∈A0
at
ER0(bt
0,l, at
0)+
0
0
bt
(7)
0,l, at
P r(ωt+1
0)V h−1(mt+1
0,l )
γ Xωt+1
0) is the expected immediate reward of
where ER0(bt
0,l, at
agent 0 and γ is the discount factor. In the context of the re-
defined reward function of the many-agent I-POMDP frame-
work in this section, the expected immediate reward is ob-
tained as:
bt
0,l, at
ER0(bt
0) = Xst
bt
0,l(M1,l−1st), . . . , bt
0,l(st)(cid:18) K
Xk=1 XCν(xt
k
,at
0)
P r(C ν(xt
k,at
0)
0,l(MN,l−1st))R0(xt
k, at
0, C ν(xt
k,at
0 )(cid:19)
(8)
=
hxt
over
all
and
outermost
the
st
k,at
0)bt
the
Ki,
the
0,l(MN,l−1st))
sum is
1, . . . , xt
0,l(M1,l−1st), . . . , bt
state
where
term,
factors,
P r(C ν(xt
de-
notes the probability of a frame-action configuration that
is context-specific to the factor, xt
k. Importantly, Propo-
sition 1 establishes that the Bellman equation above is
exact. The proof is given in the extended version of this
paper (Sonu, Chen, and Doshi 2015).
Proposition 1 (Optimality) The dynamic programming in
Eq. 7 provides an exact computation of the value function
for the many-agent I-POMDP.
Algorithms
We present an algorithm that computes the distribution over
frame-action configurations and outline our simple method
for solving the many-agent I-POMDP defined previously.
for
Distribution Over Frame-Action Configurations
Algorithm 1 generalizes an algorithm by Jiang and
Lleyton-Brown (2011)
computing configurations
over actions given mixed strategies of other agents to
include frames and conditional beliefs over models of
other agents.
It computes the probability distribution
of configurations over
the frame-action neighborhood
of an action given the belief over the agents' models:
0,l(MN,l−1st)) and
P r(C ν(x,a0,ω0)bt
. . . , bt
P r(C ν(x,a0,x′)bt
in
0,l(MN,l−1st))
bt
Eq. 5, P r(C ν(x,ωj )bt
in
and
0,l(MN,l−1st)) in Eq. 8.
bt
0,l(M1,l−1st), . . . , bt
P r(Cν(x,a0)bt
0,l(M1,l−1st),
0,l(M1,l−1st), . . . ,
0,l(MN,l−1st))
0,l(M1,l−1st), . . . ,
Eq.
6,
cremented (lines 9-12). Similarly, we update the probability
of a configuration using the probability of aj and that of the
base configuration cj−1 (line 15). This algorithm is invoked
multiple times for different values of ν(·) as needed in the
belief update and value function computation.
We utilize a simple method for solving the many-agent I-
POMDP given an initial belief: each other agent is modeled
using a finite-state controller as part of the interactive state
space. A reachability tree of beliefs as nodes is projected for
as many steps as the horizon (using Eqs. 5 and 6) and value
iteration (Eq. 7) is performed on the tree. In order to mitigate
the curse of history due to the branching factor that equals
the number of agent 0's actions and observations, we utilize
the well-known technique of sampling observations from the
propagated belief and obtain a sampled tree on which value
iteration is run to get a policy. Action for any observation
that does not appear in the sample is that which maximizes
the immediate expected reward.
Computational Savings
Computing
P r(C ν(·)b0,l(M1,l−1s),
Algorithm 1
. . . , b0,l(MN,l−1s))
Input: ν(·), hb0,l(M1,l−1s), . . . , b0,l(MN,l−1s)i
Output: A trie Pn representing distribution over the frame-
action configurations over ν(·)
1: Initialize c0 ← (0, . . . , 0), one value for each frame-
action pair in ν(·) and for φ. Insert into empty trie P0
2: Initialize P0[c0] ← 1
3: for j ← 1 to N do
4:
5:
6:
7:
Initialize Pj to be an empty trie
for all cj−1 from Pj−1 do
for all mj,l−1 ∈ Mj,l−1 do
for all aj ∈ Aj such that P r(aj mj,l−1) > 0
do
8:
9:
10:
11:
12:
13:
14:
15:
cj ← cj−1
if haj, θji ∈ ν(·) then
cj[aj] ← cj[aj] + 1
else
cj[φ] ← cj[φ] + 1
if Pj[cj] does not exist then
Initialize Pj[cj] ← 0
Pj[cj] ← Pj[cj] + Pj−1[cj−1] ×
P r(aj mj,l−1) × b0,l(mj,l−1s)
16: return Pn
Algorithm 1 adds the actions of each agent one at a time.
A Trie data structure enables efficient insertion and access
of the configurations. We begin by initializing the configu-
ration space for 0 agents (P0) to contain one tuple of inte-
gers (c0) with ν + 1 0s and assign its probability to be 1
(lines 1-2). Using the configurations of the previous step, we
construct the configurations over the actions performed by j
agents by adding 1 to a relevant element depending on j's
action and frame (lines 3-15). If an action aj performed by
j with frame mj is in the frame-action neighborhood ν(·),
then we increment its corresponding count by 1. Otherwise,
it is considered as a dummy action and the count of φ is in-
The complexity of accessing an element in a ternary search
trie is Θ(ν). The maximum number of configurations
encountered at any iteration is upper bounded by total
ν∗ )ν∗).
number of configurations for N agents, i.e. O(( N
The complexity of Algorithm 1 is polynomial
in N ,
ν∗ )ν∗) where M ∗
j are largest
O(N M ∗
sets of models and actions for any agent.
j ν∗( N
j and A∗
j A∗
j N Ω∗
j N A∗
j N Ω∗
j N A∗
j N A∗
j S2M ∗
For the traditional I-POMDP belief update, the complex-
j N ) and that
ity of computing Eq. 3 is O(SM ∗
for computing Eq. 4 is O(SM ∗
j ) where ∗
denotes the maximum cardinality of a set for any agent.
For a factored representation, belief update operator in-
vokes Eq. 3 for each value of all state factors and it in-
vokes Eq. 4 for each model of each agent j and for all
values of updated states. Hence the total complexity of be-
lief update is O(N M ∗
j ). The com-
plexity of computing updated belief over state factor xt+1
ν∗ )ν∗) (recall
using Eq. 5 is O(SN KM ∗
the complexity of Algorithm 1). Similarly, the complex-
ity of computing updated model probability using Eq. 6 is
ν∗ )ν∗). These complex-
O((SN M ∗
ity terms are polynomial in N for small values of ν∗ as
opposed to exponential in N as in Eqs. 3 and 4. The overall
complexity of belief update is also polynomial in N .
immediate
ex-
pected reward in the absence of problem structure is
the com-
O(SKM ∗
plexity of computing expected reward using Eq. 8 is
ν∗ )ν∗), which is again polyno-
O(SKN M ∗
mial in N for low values of ν∗. These complexities are
discussed in greater detail in (Sonu, Chen, and Doshi 2015).
j N ). On the other hand,
j ν∗ + Ω∗
Complexity
j ν∗( N
j ν∗( N
computing
j N A∗
j )( N
j A∗
of
the
j A∗
j A∗
Experiments
We implemented a simple and systematic I-POMDP solv-
ing technique that computes reachable beliefs over the fi-
nite horizon and then calculates the optimal value at the
root node using the Bellman equation for the Many-Agents
I-POMDP framework. We evaluate its performance in the
aforementioned non-cooperative policing protest scenario
(S = 27, A0 = 9, Aj = 4, Oj = 8, Oi = 8). We
model the other agents as POMDPs and solve them using
bounded policy iteration (Poupart and Boutilier 2003), rep-
resenting the models as finite state controllers. This repre-
sentation enables us to have a compact model space. We set
the maximum planning horizon to 4 throughout the experi-
ments. The frame-action hypergraphs are encoded into the
transition, observation and reward functions of the Many-
Agent I-POMDP (Fig. 3). All computations are carried out
on a RHEL platform with 2.80 GHz processor and 4 GB
memory.
< x, (0,0), x'>
< x, (0,0)>
Peaceful
Disruptive
Peaceful
Disruptive
< x, (0,1), x'>
< x, (2,2), x'>
< x, (0,1)>
< x, (2,2)>
GoTo2
GoTo0
GoTo1
GoTo2
GoTo0
GoTo1
< x, (0,2), x'>
< x, (1,1), x'>
< x, (0,2)>
< x, (1,1)>
< x, (1,2), x'>
Disperse
< x, (1,2)>
Disperse
(a)
(b)
Figure 3: A compact Levi graph representation of policing protest
as a frame-action hypergraph for (a) the transition function, and (b)
the reward function at site 0. Variables x and x′ represent the start
and end intensities of the protest at site 0 and the action shows the
location of the two police troops. As two police troops are sufficient
to de-escalate any protest, the contexts in which both troops are
at site 0 are independent of the actions of other agents. All other
contexts depend on the agents choosing to protest at site 0 only.
To evaluate the computational gain obtained by exploiting
problem structures, we implemented a solution algorithm
similar to the one described earlier that does not exploit
any problem structure. A comparison of the Many-Agent
I-POMDP with the original I-POMDP yields two impor-
tant results: (i) When there are few other agents, the Many-
Agent I-POMDP provides exactly the same solution as the
original I-POMDP but with reduced running times by ex-
ploiting the problem structure. (ii) Many-Agent I-POMDP
scales to larger agent populations, from 100 to 1,000+, and
the new framework delivers promising results within reason-
able time.
Protestors H I-POMDP Many-Agent Exp. Value
2
3
4
5
2
3
2
3
2
3
2
3
1 s
19 s
3 s
38 s
39 s
223 s
603 s
2,480 s
0.55 s
17 s
0.56 s
17 s
0.57 s
17 s
0.60 s
18 s
77.42
222.42
77.34
222.32
76.96
221.87
76.88
221.77
Table 1: Comparison between traditional I-POMDP and Many-
Agent I-POMDP both following same solution approach of com-
puting a reachability tree and performing backup.
In the first setting, we consider up to 5 protestors with dif-
ferent frames. As shown in Table 1, both the traditional and
the Many-Agent I-POMDP produce policies with the same
expected value. However, as the Many-Agent I-POMDP
losslessly projects joint actions to configurations, it requires
much less running time.
For Horizon H = 3
With Observation Sampling
Without Observation Sampling
0
200
400
600
800 1000 1200 1400 1600
Number of Agents
(a)
With Observation Sampling
H = 3
H = 4
)
c
e
s
(
e
m
T
i
)
c
e
s
(
e
m
T
i
20000
15000
10000
5000
0
25000
20000
15000
10000
5000
0
0
200
400
600
800 1000 1200 1400 1600
Number of Agents
(b)
Figure 4: (a) The computational gain obtained by observation
sampling. (b) Performance of Many-Agent I-POMDP with obser-
vation sampling for horizon 3 and 4. The time required to solve a
problem is polynomial in the number of agents.
Our second setting considers a large number of protestors,
for which the traditional I-POMDP does not scale. Instead,
we first scale up the exact solution method using Many-
Agent I-POMDP to deal with a few hundreds of other agents.
Although the exploitation of the problem structures reduces
the curse of dimensionality that plagues I-POMDPs, the
curse of history is unaffected by such approaches. To miti-
gate the curse of history we use the well-known observation
sampling method (Doshi and Gmytrasiewicz 2009), which
allows us to scale to over 1,000 agents in a reasonable time
of 4.5 hours as we show in Fig. 4(a). This increases to about
7 hours if we extend the horizon to 4 as shown in Fig. 4(b).
Conclusion
The key contribution of the Many-Agent I-POMDP is its
scalability beyond 1,000 agents by exploiting problem struc-
tures. We formalize widely existing problem structures --
frame-action anonymity and context-specific independence --
and encode it as frame-action hypergraphs. Other real-world
examples exhibiting such problem structure are found in eco-
nomics where the value of an asset depends on the num-
ber of agents vying to acquire it and their financial standing
(frame), in real estate where the value of a property depends
on its demand, the valuations of neighboring properties as
well as the economic status of the neighbors because an up-
scale neighborhood is desirable. Compared to the previous
best approach (Sonu and Doshi 2014), which scales to an ex-
tension of the simple tiger problem involving 5 agents only,
the presented framework is far more scalable in terms of
number of agents. Our future work includes exploring other
types of problem structures and developing approximation
algorithms for this I-POMDP. An integration with existing
multiagent simulation platforms to illustrate the behavior of
agent populations may be interesting.
Acknowledgements
This research is supported in part by a NSF CAREER grant,
IIS-0845036, and a grant from ONR, N000141310870. We
thank Brenda Ng for valuable feedback that led to improve-
ments in the paper.
References
P.,
[Boutilier et al. 1996] Boutilier, C.; Friedman, N.; Gold-
szmidt, M.; and Koller, D. 1996. Context-specific indepen-
dence in bayesian networks.
In Twelfth international con-
ference on Uncertainty in artificial intelligence (UAI), 115 --
123.
[Chandrasekaran et al. 2014] Chandrasekaran, M.; Doshi, P.;
Zeng, Y.; and Chen, Y. 2014. Team behavior in interac-
tive dynamic influence diagrams with applications to ad hoc
teams (extended abstract). In Autonomous Agents and Multi-
Agent Systems Conference (AAMAS), 1559 -- 1560.
[Doshi and Gmytrasiewicz 2009] Doshi,
and Gmy-
trasiewicz, P. J. 2009. Monte Carlo sampling methods for
approximating interactive POMDPs. Journal of Artificial
Intelligence Research 34:297 -- 337.
[Doshi and Perez 2008] Doshi, P., and Perez, D.
2008.
Generalized point based value iteration for interactive
POMDPs.
In Twenty Third Conference on Artificial Intel-
ligence (AAAI), 63 -- 68.
[Doshi et al. 2010] Doshi, P.; Qu, X.; Goodie, A.; and Young,
D. 2010. Modeling recursive reasoning in humans using em-
pirically informed interactive POMDPs. In International Au-
tonomous Agents and Multiagent Systems Conference (AA-
MAS), 1223 -- 1230.
[Doshi 2012] Doshi, P. 2012. Decision making in complex
multiagent settings: A tale of two frameworks. AI Magazine
33(4):82 -- 95.
[Gmytrasiewicz and Doshi 2005] Gmytrasiewicz, P. J., and
Doshi, P. 2005. A framework for sequential planning in mul-
tiagent settings. Journal of Artificial Intelligence Research
24:49 -- 79.
[Hoang and Low 2013] Hoang, T. N., and Low, K. H. 2013.
Interactive POMDP lite: Towards practical planning to pre-
dict and exploit intentions for interacting with self-interested
agents. In International Joint Conference on Artificial Intel-
ligence (IJCAI), 2298 -- 2305.
[Jiang and Leyton-Brown 2010] Jiang, A. X., and Leyton-
Brown, K. 2010. Bayesian action-graph games. In Advances
in Neural Information Processing Systems (NIPS), 991 -- 999.
[Jiang, Leyton-Brown, and Bhat 2011] Jiang, A. X.; Leyton-
Brown, K.; and Bhat, N. A. 2011. Action-graph games.
Games and Economic Behavior 71(1):141 -- 173.
[Jiang, Leyton-Brown, and Pfeffer 2009] Jiang,
X.;
Leyton-Brown, K.; and Pfeffer, A. 2009. Temporal action-
graph games: A new representation for dynamic games.
In Twenty-Fifth Conference on Uncertainty in Artificial
Intelligence (UAI), 268 -- 276.
[Kearns, Littman, and Singh 2001] Kearns, M.; Littman, M.;
and Singh, S. 2001. Graphical models for game theory. In
Uncertainty in Artificial Intelligence (UAI), 253 -- 260.
[Koller and Milch 2001] Koller, D., and Milch, B.
2001.
Multi-agent influence diagrams for representing and solving
games. In IJCAI, 1027 -- 1034.
[Nair et al. 2005] Nair, R.; Varakantham, P.; Tambe, M.; and
Yokoo, M. 2005. Networked distributed POMDPs: A synthe-
sis of distributed constraint optimization and POMDPs. In
A.
Twentieth AAAI Conference on Artificial Intelligence, 133 --
139.
[Ng et al. 2010] Ng, B.; Meyers, C.; Boakye, K.; and Nitao, J.
2010. Towards applying interactive POMDPs to real-world
adversary modeling. In Innovative Applications in Artificial
Intelligence (IAAI), 1814 -- 1820.
[Poupart and Boutilier 2003] Poupart, P., and Boutilier, C.
2003. Bounded finite state controllers. In Neural Informa-
tion Processing Systems.
[Roughgarden and Tardos 2002] Roughgarden, T., and Tar-
dos, E. 2002. How bad is selfish routing? Journal of ACM
49(2):236 -- 259.
[Seuken and Zilberstein 2008] Seuken, S., and Zilberstein, S.
2008. Formal models and algorithms for decentralized de-
cision making under uncertainty. Journal of Autonomous
Agents and Multiagent Systems 17(2):190 -- 250.
[Seymour and Peterson 2009] Seymour, R., and Peterson,
G. L. 2009. A trust-based multiagent system.
In IEEE
International Conference on Computational Science and En-
gineering, 109 -- 116.
[Sonu and Doshi 2014] Sonu, E., and Doshi, P. 2014. Scal-
able solutions of interactive POMDPs using generalized and
bounded policy iteration. Journal of Autonomous Agents and
Multi-Agent Systems DOI: 10.1007/s10458 -- 014 -- 9261 -- 5, in
press.
[Sonu, Chen, and Doshi 2015] Sonu, E.; Chen, Y.; and
Doshi, P. 2015. Individual planning in agent populations:
Exploiting anonymity and frame-action hypergraphs. Tech-
nical Report http://arxiv.org/abs/1503.07220, arXiv.
[Varakantham, Adulyasak, and Jaillet 2014] Varakantham,
P.; Adulyasak, Y.; and Jaillet, P.
2014. Decentralized
stochastic planning with anonymity in interactions. In AAAI
Conference on Artificial Intelligence, 2505 -- 2511.
[Varakantham et al. 2012] Varakantham, P.; Cheng, S.; Gor-
don, G.; and Ahmed, A. 2012. Decision support for agent
populations in uncertain and congested environments. In Un-
certainty in artificial intelligence (UAI), 1471 -- 1477.
[Wang 2013] Wang, F. 2013. An I-POMDP based multi-
agent architecture for dialogue tutoring.
In International
Conference on Advanced ICT and Education (ICAICTE-13),
486 -- 489.
[Woodward and Wood 2012] Woodward, M. P., and Wood,
R. J. 2012. Learning from humans as an i-pomdp. CoRR
abs/1204.0274.
[Wunder et al. 2011] Wunder, M.; Kaisers, M.; Yaros, J.; and
Littman, M. 2011. Using iterated reasoning to predict oppo-
nent strategies. In International Conference on Autonomous
Agents and Multi-Agent Systems (AAMAS), 593 -- 600.
Appendix
Factored Belief Update:
, . . . , mt+1
n )
1
b0,l(st+1, mt+1
= P r(st+1, mt+1
1
= P r(st+1bt
0, at
× · · · × P r(mt+1
0, at
0, ωt+1
)
n bt
, . . . , mt+1
0, ωt+1
n st+1, mt+1
0
1
) × P r(mt+1
1
, . . . , mt+1
0
st+1, bt
n−1bt
)
0, at
0, at
0, ωt+1
0, ωt+1
)
0
0
Derivation of equation 5:
Starting with equation 3, we have:
P r(st+1bt
0,l, at
= P r(st+1bt
0, ωt+1
0
0)P r(ωt+1
0,l, at
) ∝ P r(st+1, ωt+1
bt
0,l, at
0)
0
0,l, at
0)
st+1, bt
0
The update term for physical states may be represented as
a product of its factors such that for any factor Xk:
P r(xt+1
k
bt
0,l, at
P r(at
−0mt
Xat
−0
0,l(mt
bt
−0st)×
bt
0) = Xst
−0) T0(xt
0,l(st) Xmt
k, hat
0, at
−0
−0i, xt+1
k
)
where bt
P r(at
0,l(mt
−0st) = bt
−0) = P r(at
0,l(mt
1mt
1st) × . . . × bt
1) × . . . × P r(at
0,l(mt
N mt
N st), and
N ).
−0mt
We introduce a projection function δν(ψ) that maps joint
actions to the corresponding frame-action configurations as
defined in definition 4. Formally δν(ψ) : a → Cν(ψ), where
Cν(ψ) is the set of all possible configurations such that for
all agents j with frame θ, C(a, θ) = {j : aj = a, θj =
θ, (aj, θ) ∈ ν(ψ)}.
Next we partition the set of joint action of all other agents
such that the
A−0 into smaller subsets A1
projection function δν(ψ) maps all joint actions belonging
to any given partition Ac
−0 to the same value configuration.
Hence, we may rewrite the above equation as:
. . . , ACν(ψ)
−0,
−0
P r(xt+1
k
bt
0,l, at
0) = Xst
0,l(st)
bt
Cν(xt
k
,at
0 ,x
t+1
k
)
Xc=1 Xmt
−0
bt
0,l(mt
−0st) Xat
k, hat
−0∈Ac
−0i, xt+1
0, at
−0
k
)
T0(xt
P r(at
−0mt
−0)
Under
independence,
T0(xt
0, at
k, hat
k,at
where C ν(xt
frame-action
for all
−0i, xt+1
k
0,xt+1
) = δν(xt
k
anonymity
and
joint actions at
−0
) = T0(xt
k, at
0,xt+1
k
0, C ν(xt,at
)(at
−0).
k,at
frame-action
∈ Ac
−0
0,xt+1), xt+1
),
k
P r(xt+1
k
bt
0,l, at
0) = Xst
bt
0,l(st)
Cν(xt
k
,at
0 ,x
t+1
k
)
Xc=1 Xmt
−0
P r(at
−0mt
−0)
bt
0,l(mt
−0st) Xat
0, C ν(xt
k, at
−0∈Ac
k,at
−0
0,xt+1
k
T0(xt
), xt+1
k
)
The cumulative probability of
ping
to
the
same
joint actions map-
−0st)
0,l(mt
bt
configuration, Pmt
−0
−0), is computed tractably using algo-
P r(at
−0mt
Pat
−0∈Ac
0,xt+1
rithm 1 as P r(C ν(xt
Hence, the equation becomes:
k ,at
−0
k
)b0,l(M t
1st), . . . b0,l(M t
nst)).
P r(xt+1
k
b0,l(M t
bt
bt
0,l, at
0) = Xst
1st), . . . b0,l(M t
0,l(st) X
0 ,x
nst)) × T0(xt, at
Cν(xt
,at
k
t+1
k
P r(C ν(xt
k,at
0,xt+1
k
)
)
0, C ν(xt
k,at
0,xt+1
k
), xt+1)
Similarly,
the observation probability may also be
) =
−0) may be substitued instead of the joint
obtained in a factored form and C ν(st+1,at
δν(st+1,at
action.
0,ωt+1
0,ωt+1
)(at
0
0
K
P r(ωt+1
0
st+1, bt
0,l, at
bt
0,l(st)
0) = Xst
)bt
0,l(M1,l−1st), . . . , bt
0,ωt+1
0, C ν(xt+1
,at
), ωt+1
0,ωt+1
,at
)
k
0
0
0
Yk=1 X
,at
t+1
k
t+1
0
)
0 ,ω
Cν(x
0,l(MN,l−1st))
k
P r(C ν(xt+1
O0(xt+1
, at
k
Therefore, we may rewrite equation 3 as follows:
P r(st+1bt
0,l, at
0, ωt+1
0
) ∝ (cid:26)Xst
bt
0,l(st)
P r(C ν(xt+1
k
,at
0,ωt+1
0
)bt
0,l(M1,l−1st), . . . , bt
K
Yk=1 X
0,l(MN,l−1st))
Cν(x
t+1
k
t+1
0
,at
0
,ω
)
O0(xt+1
k
, at
0, C ν(xt+1
k
,at
0,ωt+1
0
), ωt+1
0
)(cid:27) ×(cid:26)Xst
bt
0,l(st)
K
Yk=1
X
,at
0
,x
Cν(xt
k
P r(C ν(xt
k ,at
0,xt+1
k
)bt
0,l(M1,l−1st), . . . ,
t+1
k
)
bt
0,l(MN,l−1st))T0(xt
k, C ν(xt
k,at
0,xt+1
k
), xt+1
k
)(cid:27)
(9)
Derivation of equation 6:
The belief update over the models of agent j shown in equa-
N,l−1, bt
0,l, at
P r(at
0, ωt+1
1mt
1,l−1)
)
0
j A∗
j ν∗(cid:0)N +ν∗
O(S×{N M ∗
complexity terms are polynomial in N for small values of
ν∗ as opposed to exponential in N as in equations 3 and 4.
The overall complexity of belief update is also polynomial
in N .
ν∗ (cid:1)}). These
ν∗ (cid:1)+Ω∗
j (cid:0)N +ν∗
Proof of Proposition 1:
P r(at
j−1mt
j−1,l−1)
The expected reward of agent 0 is obtained as the sum of
reward factors.
bt
0,l(st)(cid:18) K
bt
0,l(mt
Xk=1 Xmt
1,l−1) × . . . × P r(at
−0
P r(at
1mt
1st) × . . . ×
N mt
N,l−1)
ER0(bt
bt
0,l(mt
0,l, at
0) = Xst
N st) ×Xat
−0i)(cid:19)
0, at
−0
R0(xt
k, hat
j )N (A∗
Complexity of computing expected reward using the
j )N ). Equation 8 is
above equation is O(SK(M ∗
derived similarly to the belief update by substituting dis-
tribution over frame-action configurations for distributions
over joint models and joint actions. This combined with the
proofs for Eqs. 5 and 6 allow us to obtain Eq.7 from the
Bellman equation of the original I-POMDP.
The complexity of computing expected reward using
equation 8 is
O(SK{N M ∗
nomial in N for low values of ν∗.
j ν∗+1}(cid:0)N +ν∗
j A∗
ν∗ (cid:1)) which is again poly-
tion 4 can be rewritten as follows:
j+1,l−1, . . . , mt+1
j,l−1st+1, mt+1
0(st) Xmt
1,l−1st)Xat
bt
bt
0,l(mt
bt
0,l(mt
1,l−1
1
j−1,l−1st) Xat
j mt
P r(at
j−1
j−1,l−1
bt
0,l(mt
j,l−1st)Xat
j
P r(mt+1
= Xst
. . . Xmt
Xmt
bt
0,l(mt
j,l−1
P r(at
j+1mt
j+1,l−1st) Xat
N,l−1st)Xat
j+1
N
bt
0,l(mt
P r(at
N mt
j,l−1) Xmt
j,l−1) . . . Xmt
j+1,l−1
N,l−1
N,l−1) Xωt+1
j, at
mt
j
j
Oj (st+1, haj, at
−ji, ωt+1
j
) P r(mt+1
j, ωt+1
j
)
Substituting frame-action configuration as in equation 9,
we get:
P r(mt+1
j+1,l−1, . . . , mt+1
j,l−1st+1, mt+1
0(st)Xmt
bt
bt
0(mt
P r(at
N,l−1, bt
j mt
= Xst
P r(C ν(xt+1,at
j ,ωj )at
bt
0,l(Mj+1st), . . . bt
j
j
jst)Xat
0, bt
0,l(M1st), . . . , bt
0,l(mN st)) × Xot+1
j, at
mt
j
j
j, ωt+1
j
)
0
0, ωt+1
0,l, at
j) X
0,l(Mj−1st),
Oj(xt+1, at
j,
ν(xt+1 ,at
j
C
,ωj )
C ν(xt+1,ωj ), ωt+1
j
) P r(mt+1
)
(10)
Where the probability over the configurations is computed
as in algorithm 1 using belief over models of all other agents
except j. In the end we add 1 to the count of action a0 in
every configuration.
Complexity of belief update:
j )N (A∗
j N A∗
For the traditional I-POMDP belief update, the complex-
j )N ) and
ity of computing equation 3 is O(S(M ∗
that for computing equation 4 is O(SM ∗
j N Ω∗
j )
where ∗ denotes the maximum cardianlity of a set for
any agent. For factored representation, belief update oper-
ator invokes equation 3 for each value of all state factors
and it invokes equation 4 for each model of each agent
j and for all values of updated states. Hence the total
complexity of belief update is O(KX ∗SM ∗
j N
+N M ∗
j N A∗
j S2(M ∗
j )N A∗
j N Ω∗
j ).
In equation 5, algorithm is called once for all values of st.
The two inner summations iterate over all possible config-
urations over transition and observation contexts. The num-
ber of configurations is upper bounded by (cid:0)N +ν∗
ν∗ is the maximum cardinality of the frame-action neigh-
borhood for any context. Hence the complexity of comput-
ing updated belief of state factor xt+1 using equation 5 is
O(SK × {N M ∗
the complexity of algorithm 1). Similarly, the complexity
of computing updated model probability using equation 6 is
ν∗ (cid:1) +(cid:0)N +ν∗
ν∗ (cid:1)}) (recall
j ν∗(cid:0)N +ν∗
ν∗ (cid:1) where
j A∗
|
0909.4349 | 1 | 0909 | 2009-09-24T02:05:06 | Leader-following Consensus Problems with a Time-varying Leader under Measurement Noises | [
"cs.MA",
"cs.IT",
"cs.IT"
] | In this paper, we consider a leader-following consensus problem for networks of continuous-time integrator agents with a time-varying leader under measurement noises. We propose a neighbor-based state-estimation protocol for every agent to track the leader, and time-varying consensus gains are introduced to attenuate the noises. By combining the tools of stochastic analysis and algebraic graph theory, we study mean square convergence of this multi-agent system under directed fixed as well as switching interconnection topologies. Sufficient conditions are given for mean square consensus in both cases. Finally, a numerical example is given to illustrate our theoretical results. | cs.MA | cs |
Leader-following Consensus Problems with a
Time-varying Leader under Measurement Noises
Yilun Shang1
Department of Mathematics, Shanghai Jiao Tong University, Shanghai 200240, P. R. China.
Abstract
In this paper, we consider a leader-following consensus problem for networks of
continuous-time integrator agents with a time-varying leader under measurement noises.
We propose a neighbor-based state-estimation protocol for every agent to track the
leader, and time-varying consensus gains are introduced to attenuate the noises. By
combining the tools of stochastic analysis and algebraic graph theory, we study mean
square convergence of this multi-agent system under directed fixed as well as switching
interconnection topologies. Sufficient conditions are given for mean square consensus
in both cases. Finally, a numerical example is given to illustrate our theoretical results.
Keywords: consensus problems; multi-agent system; leader-following; stochastic
system.
1. Introduction
In recent years, there has been an increasing research interest in the distributed co-
ordination for multi-agent systems. This is partly due to its broad applications in many
areas such as cooperative control of unmanned aerial vehicles, formation control [1, 26, 27]
and swarming behaviors of social living beings [11, 16, 30].
Consensus problems have a long history in computer science and formed the foundation
of the field of distributed computing [12]. In consensus control, it is critical to design a
decentralized network algorithm based on neighborhood information for agents to reach
an agreement on their states, asymptotically in some sense. For a variety of consensus
algorithms and convergence results we refer the reader to the comprehensive surveys [17,
24] and references therein. Most researches in the previous literature assume the exchange
of messages between agents is error-free. However, this is only an ideal approximation
for real communication processes. Recently, consensus of dynamic networks with random
measurement noises has attracted the attention of some researchers. In [8, 25], the authors
introduce time-varying consensus gains and design control schemes based on a Kalman
filter structure. The decreasing consensus gain a(k) (where k is the discrete time instant) in
the protocols is proposed in [5] to attenuate the measurement noises in a strongly connected
circulant network. The analysis in [5] is generalized to strongly connected digraphs in
[7] and digraphs containing a spanning tree in [6] by the same authors. The work in
[10] deals with discrete-time average consensus problems in switching balanced digraphs
under stochastic communication noises, while [9] investigates the continuous-time average
consensus control with fixed topology and Gaussian communication noises. The authors in
[13] treat a continuous-time leader-following consensus control under measurement noises
with a constant state leader.
In this paper, motivated by the above works, we consider a leader-following consensus
problem for networks of continuous-time integrator agents with a time-varying leader in
directed fixed and switching topologies. The control input of each agent is based on the
1email: [email protected]
1
measurement of its neighbors’ states and some estimated data of the leader which are both
corrupted by white noises. We design a leader-following consensus protocol such that the
leader has an underlying dynamics and some variables (e.g. velocity and acceleration) of
the leader cannot be measured and every follower can obtain the measured information
(e.g. position) of the leader only when they are connected with the leader directly. The
collective behavior of self-organized groups of agents with active (or dynamical) leaders
is one of the most interesting topics in distributed cooperative control. However, as [22]
suggests, the extension of consensus algorithms from a constant reference to a time-varying
one is non-trivial. Some related results can be found e.g. in [3, 20, 31], where the systems
considered are all deterministic and free of noise.
Inspired by [5, 9, 13], we introduce time-varying consensus gains in the followers control
protocol to attenuate the measurement noises, which lead to a time-varying stochastic dif-
ferential equation of the system. The state matrix of the equation is time-dependent and
no longer a Laplacian matrix, and is neither symmetric nor diagonalizable. To implement
the convergence study, we merge stochastic analysis and algebraic graph theory, by devel-
oping a Lyapunov-based approach and addressing the Ito integral by the stopping time
truncation method. Firstly, we derive a sufficient condition for the state of each follower
to converge to that of the leader in mean square under fixed and directed interconnection
topology. Then it is shown that the algorithm also render each follower track the leader
in mean square under switching topology when the subgraph induced by the followers is
balanced.
The rest of the paper is organized as follows. In Section 2, we provide some preliminar-
ies and present the leader-following consensus protocol. Section 3 contains the convergence
analyses under directed fixed and switching interaction topologies. A numerical example
is given in Section 4 and we conclude the paper in Section 5.
2. Problem formulation
Before we proceed, some basic concepts on graph theory (see e.g. [2]) are provided as
below.
Let G = (V, E, A) be a weighted digraph with the set of vertices V = {1, 2, · · · , n} and
the set of arcs E ⊆ V × V. The vertex i in G represents the ith agent, and a directed
edge (i, j) ∈ E means that agent j can directly receive information from agent i. The set
of neighbors of vertex i is denoted by Ni = {j ∈ V (j, i) ∈ E}. A = (aij) ∈ Rn×n is
called the weighted adjacency matrix of G with nonnegative elements and aij > 0 if and
j=1 aij
If din(i) = dout(i) for i = 1, 2, · · · , n, then the
digraph G is called balanced [18]. The Laplacian of G is defined as L = D − A, where
D = diag(din(1), din(2), · · · , din(n)). A digraph G is called strongly connected if there is a
directed path from i to j between any two distinct vertices i, j ∈ V. There exists a directed
path from vertex i to vertex j, then j is said to be reachable from i. For every vertex in
digraph G, if there is a path from vertex i to it, then we say i is globally reachable in G.
This is much weaker than strong connectedness.
only if j ∈ Ni. The in-degree and out-degree of vertex i are defined as din(i) =Pn
and dout(i) = Pn
j=1 aji, respectively.
Here, we consider a system consisting of n + 1 agents, in which an agent indexed by 0 is
assigned as the leader and the other agents indexed by 1, 2, · · · , n are referred as follower
agents. The information interaction topology among n followers are described by the
digraph G as defined above; and the whole system including n + 1 agents is conveniently
2
modeled by a weighted digraph G = (V, E, A) with V = {0, 1, · · · , n} and
A =
0
0
a10 a11
...
...
an0 an1
· · ·
· · ·
. . .
· · ·
0
a1n
...
ann
∈ R(n+1)×(n+1),
where the lower right block submatrix of order n can be viewed as A. We define a diagonal
matrix B = diag(b1, b2, · · · , bn) to be the leader adjacency matrix associated with G, where
bi = ai0 ≥ 0 and bi > 0 if and only if 0 ∈ Ni(G). Here, Ni(G) is the set of neighbors of
agent i in G.
The continuous-time dynamics of n followers is described as follows:
xi(t) = ui(t),
i = 1, 2, · · · , n,
(1)
where xi(t) ∈ R is the state of the ith agent, and ui(t) ∈ R is the control input. The leader
of this considered multi-agent system is described by a double integrator of the form:
x0(t) = g(t)v0(t)
v0(t) = a0(t)
y(t) = x0(t)
(2)
where g(t) : [0, ∞) → (0, ∞) is piecewise continuous, y(t) is the measured output and
a0(t) is the input. We assume g(t) and a0(t) are known, that is, the dynamical behavior of
the leader is precisely known (c.f. Remark 1). On the other hand, y(t) = x0(t) is the only
data that may be gotten by the followers when they are connected to the leader directly.
Since v0(t) cannot be measured, we have to estimate v0(t) in a distributed way during the
evolution. The estimate of v0(t) by agent i is denoted by vi(t), i = 1, 2, · · · , n.
In our model, the ith agent receives information from its neighbors with measurement
noises:
yji(t) = xj(t) + σjinji(t),
j ∈ Ni,
y0i(t) = x0(t) + σ0in0i(t),
(3)
(4)
where yji(t) (i ∈ V, j ∈ V) denotes the measurement of the jth agent’s state xj(t) by the
ith agent. The {nji(t) j ∈ V, i ∈ V} are independent standard white noises and σji ≥ 0
is the noise intensity.
A group of controls U = {ui i = 1, 2, · · · , n} is called a measurement-based distributed
protocol [9], if ui(t) ∈ σ(cid:0)xi(s),Sj∈N i
Herein σ(ξλ, λ ∈ Λ) denotes the σ-algebra generated by a family of random variables
{ξλ, λ ∈ Λ}. The so-called leader-following consensus problem is to design a measurement-
based distributed protocol such that each follower’s state will converge to the leader’s in
some sense as time goes on.
yji(s), 0 ≤ s ≤ t(cid:1), for t ≥ 0, i = 1, 2, · · · , n.
Consequently, we propose the distributed control protocol which consists of two parts:
• a neighbor-based feedback law:
ui(t) = h(t)(cid:16)Xj∈Ni
aij(yji(t) − xi(t)) + bi(y0i(t) − xi(t))(cid:17) + g(t)vi(t),
i = 1, 2, · · · , n (5)
where t ≥ 0 and h(t) : [0, ∞) → (0, ∞) is a piecewise continuous function, called a time-
varying consensus gain [9].
3
• a dynamic neighbor-based system to estimate v0(t):
vi(t) = a0(t) + γh(t)(cid:16)Xj∈Ni
aij(y′
ji(t) − xi(t)) + bi(y′
0i(t) − xi(t))(cid:17),
i = 1, 2, · · · , n (6)
where 0 < γ < 1 is some constant, and moreover y′
of yji(t) and y0i(t), respectively. In other words, we have
ji(t) and y′
0i(t) are independent copies
y′
ji(t) = xj(t) + σjin′
ji(t),
y′
0i(t) = x0(t) + σ0in′
0i(t),
j ∈ Ni,
(7)
(8)
where {n′
{nji(t) j ∈ V, i ∈ V}.
ji(t) j ∈ V, i ∈ V} are independent standard white noises and independent with
The set of neighbors Ni of agent i in (5) and (6) varies when the interconnection
topology is switching and we defer the corresponding protocol formulation to Section 3.2.
Remark 1. We take individual state xi as scalar for simplicity in (1) and it can be
extended to multi-dimensional scenarios as studied in [20, 31] without much effort. For
example, if xi ∈ R2, it can be thought as the position of agent i moving in a plane.
Therefore, gv0 and ga0 + gv0 are the velocity and acceleration of the leader respectively,
which are known since the exact dynamics of the leader is assumed.
Remark 2. We separate a factor g from the ‘velocity term’ of the leader in (2) in order
to tone the decreasing consensus gain h, which appears to be a notable feature distinct
from some kinds of uncertain environment (see e.g.
[28, 29]), where a random term is
directly appended behind the equation of the system. In such works, the consensus gains
are supposed to have positive lower bound.
Remark 3. From (5) and (6) it is clear that the designed protocol for the ith agent is
indeed a measurement-based distributed protocol since it relies only on the state of itself
and its neighbors.
Let x(t) = (x1(t), · · · , xn(t))T and v(t) = (v1(t), · · · , vn(t))T . Denote the ith row
of the matrix A by αi, and Σi := diag(σ1i, · · · , σni) for i = 1, 2, · · · , n. Hence Σ :=
diag(α1Σ1, · · · , αnΣn) is an n × n2 dimensional block diagonal matrix. Let n0(t) =
(n01(t), · · · , n0n(t))T and ni(t) = (n1i(t), · · · , nni(t))T for i = 1, 2, · · · , n.
In addition,
n′
0(t) and n′
i(t) can be defined in a similar way. The juxtaposed matrix Q := (B, Σ) is an
n × n(n + 1) dimensional block matrix. Combining (1) with (5) and (6), we may write the
protocol in a matrix form:
( dx(t)
dt = −h(t)(L + B)x(t) + h(t)B1x0(t) + g(t)v(t) + h(t)QZ(t)
dv(t)
dt = a0(t)1 − γh(t)(L + B)x(t) + γh(t)B1x0(t) + γh(t)QZ ′(t)
(9)
0 (t), nT
1 (t), · · · , nT
T (t))T are two
where Z(t) = (nT
n(n + 1) dimensional independent standard white noise sequences, and 1 = (1, · · · , 1)T ∈
Rn. The system (9) may be further written in the form of the Ito stochastic differential
equations:
n (t))T and Z ′(t) = (n′
0
T (t), · · · , n′
n
T (t), n′
1
(cid:26) dx(t) = −h(t)(L + B)x(t)dt + h(t)B1x0(t)dt + g(t)v(t)dt + h(t)GdW1(t)
dv(t) = a0(t)1dt − γh(t)(L + B)x(t)dt + γh(t)B1x0(t)dt + γh(t)GdW2(t)
(10)
where W1(t) = (W11(t), · · · , W1n(t))T and W2(t) = (W21(t), · · · , W2n(t))T are two n
dimensional standard Brownian motions which are independent with each other, and
G := diag(cid:16)qb2
1 +Pj∈N1
σ2
j1a2
1j , · · · ,qb2
n +Pj∈Nn
σ2
jna2
nj (cid:17).
4
3. Convergence analysis
In this section we will give the convergence analysis of the system (10) and show
that the state of every follower will track that of the leader in the sense of mean square
convergence, that is, Ekx(t) − x0(t)1k → 0, as t → ∞. Here k · k denotes Frobenius norm.
Two different cases, fixed topology and switching topology, are explored.
Remark 4. Mean square consensus protocols for stochastic systems are first introduced
in [5] and then further studied by several researchers (e.g. [6, 7, 9, 10, 13]). Mean square
convergence seems to be an important alternative for almost sure convergence in consensus
problems under noisy environments.
For a given symmetric matrix A, let λmax(A) and λmin(A) denote its maximum and
minimum eigenvalue, respectively. To get the main result, we need the following assump-
tions:
(A1) The vertex 0 is globally reachable in G.
(A2) There is a δ > 0, such that h(t)
Here, P is a positive definite matrix defined by Equation (13), see below.
(A3)
(A4)
Remark 5. Assumption (A1) is imposed on the network topology, which is much weaker
than strong connectedness. The technical Assumption (A2) roughly means that g is com-
parable with the consensus gain h. Assumptions (A3) and (A4) are called convergence
condition and robustness condition respectively in [9], and which are standard assump-
tions often used in the stochastic approximation [15].
0 h(s)ds = ∞.
0 h2(s)ds < ∞.
R ∞
R ∞
g(t) > λmax(P )
2γ(1−γ 2) + δ for t ≥ 0.
3.1. Fixed topology
Let x∗ = x − x01 and v∗ = v − v01. We then obtain an error dynamics of (10) as
follows:
dε(t) = F (t)ε(t)dt + G(t)dW (t),
t ≥ 0
(11)
where ε(t) = (x∗(t), v∗(t))T , W (t) = (W1(t), W2(t))T and
F (t) =(cid:18) −h(t)(L + B)
−γh(t)(L + B)
g(t)In
0
(cid:19) ,
Here In denotes the n × n dimensional identity matrix.
We will need a lemma for Laplacian matrix.
G(t) = h(t)(cid:18) G
γG (cid:19) := h(t)eG.
Lemma 1.[2, 23] The Laplacian matrix L of a digraph G = (V, E, A) has at least one zero
eigenvalue and all of the nonzero eigenvalues are in the open right half plane. Furthermore,
L has exactly one zero eigenvalue if and only if there is a globally reachable vertex in G.
The main result in this section is given as follows:
Theorem 1. For system (1) with the consensus protocols (5) and (6), if Assumptions
(A1)-(A4) hold, then
Ekε(t)k2 = 0.
lim
t→∞
(12)
Proof. By Assumption (A1) and Lemma 1, we know L + B is a positive stable matrix,
or in other words, −L − B is a stable matrix. From Lyapunov theorem, there is a unique
positive definite matrix P ∈ Rn×n such that
(L + B)T P + P (L + B) = In.
(13)
5
−γP
−γP
Let eP = (cid:18) P
P (cid:19) and define a Lyapunov function V (t) = εT (t)eP ε(t). Since
0 < γ < 1, eP is a positive definite matrix. In fact, we have λmin(eP ) = (1 − γ)λmin(P ) and
λmax(eP ) = (1 + γ)λmax(P ). Utilizing Ito formula and (11), we have
dV (t) = εT (t)(cid:0)eP F (t) + F T (t)eP(cid:1)ε(t)dt + tr(GT (t)eP G(t))dt + 2εT (t)eP G(t)dW (t).
Here tr(·) means the trace of a matrix. From the Lyapunov equation (13), we get
eP F (t) + F T (t)eP = −h(t) (1 − γ2)In −P g(t)
h(t) ! := −h(t)eQ(t).
h(t)
2γP g(t)
−P g(t)
h(t)
Invoking the Haynsworth inertia additivity formula [21], Assumption (A2) and the positive
definiteness of P , we know that eQ(t) is positive definite with ρ := mint≥0(cid:8)λmin(eQ(t))(cid:9) > 0.
Thereby, we have
dV (t) ≤ −h(t)
Next we want to prove
V (t)dt + h2(t)tr(eGTePeG)dt + 2h(t)εT (t)ePeGdW (t).
(14)
(15)
ρ
λmax(eP )
EZ t
t0
h(s)εT (s)ePeGdW (s) = 0,
∀ 0 ≤ t0 ≤ t.
For any given T ≥ t0 ≥ 0 and K ∈ N, let τ t0,T
is a stopping time. By (14) we have, for t0 ≤ t ≤ T ,
K = T ∧ inf{t ≥ t0 εT (t)eP ε(t) ≥ K}, which
ρ
[t≤τ
≤ −
K )1
E(cid:0)V (t ∧ τ t0,T
λmax(eP )Z t
≤ tr(eGTePeG)Z T
t0
t0
t0 ,T
K ](cid:1) − EV (t0)
h(s)E(cid:0)V (s ∧ τ t0,T
h2(s)ds.
K )1
[s≤τ
t0,T
K ](cid:1)ds + tr(eGTePeG)Z t
t0
h2(s)ds
This implies that there is a constant Ct0,T > 0 such that
E(cid:0)V (t ∧ τ t0,T
K )1
[t≤τ
t0 ,T
K ](cid:1) ≤ Ct0,T ,
∀ t0 ≤ t ≤ T.
Since limK→∞ t ∧ τ t0,T
K = t a.s., for t0 ≤ t ≤ T , by Fatou lemma, we derive
sup
t0≤t≤T
EV (t) ≤ Ct0,T .
Accordingly,
EZ t
t0
h2(s)V (s)ds ≤ sup
t0≤s≤t
EV (s) ·Z T
0
h2(s)ds < ∞,
∀ 0 ≤ t0 ≤ t.
Combining this with
EZ t
t0
h2(s)kεT (s)ePeGk2ds ≤ keP kkeGk2EZ t
t0
6
h2(s)V (s)ds
yields (15) (c.f. [19]).
Now employing (14) and (15), we obtain
0
0
ρ
λmax(eP )Z t
EV (t) ≤ EV (0) exp(cid:16) −
+ tr(eGTePeG)Z t
h(s)EV (s)ds + tr(eGTePeG)Z t
λmax(eP )Z t
h2(s) exp(cid:16) −
h(s)ds(cid:17)
λmax(eP )Z t
ρ
ρ
0
0
s
h2(s)ds,
∀ t ≥ 0.
h(u)du(cid:17)ds.
(16)
Clearly, by Assumption (A3) the first term on the right hand side of (16) tends to 0,
as t → ∞. For any η > 0, by Assumption (A4), there exists some s0 > 0 such that
EV (t) − EV (0) ≤ −
Thus, from the comparison principle [14],
s0
R ∞
Z t
0
h2(s)ds < η. Hence,
s
0
ρ
ρ
h2(s) exp(cid:16) −
=Z s0
≤ exp(cid:16) −
≤ exp(cid:16) −
λmax(eP )Z t
h2(s) exp(cid:16) −
λmax(eP )Z t
λmax(eP )Z t
h(u)du(cid:17)ds
λmax(eP )Z t
h(u)du(cid:17)Z s0
h(u)du(cid:17)Z ∞
ρ
ρ
s0
s0
0
s
h(u)du(cid:17)ds +Z t
h2(s)ds +Z t
s0
s0
0
h2(s) exp(cid:16) −
h2(s)ds
ρ
λmax(eP )Z t
s
h(u)du(cid:17)ds
h2(s)ds + η,
∀ t ≥ s0
By Assumptions (A3), (A4) and the arbitrariness of η, the last expression tends to zero,
as t → ∞. Therefore, (16) yields limt→∞ EV (t) = 0. Note that
kε(t)k2 ≤
V (t)
λmin(eP )
which concludes the proof. ✷
Remark 6. As is known, the solution to Lyapunov matrix equation may be obtained by
using Kronecker product. Thus the positive definite matrix P involved in Assumption (A2)
can be given explicitly.
Remark 7. Theorem 1 implies that in the fixed topology, under Assumptions (A1)-(A4),
the designed protocol guarantees the state of each follower tracks that of the leader in mean
square.
3.2. Switching topology
In this section we deal with the convergence of the protocol under switching topology.
Let σ(t) : [0, ∞) → SH = {1, 2, · · · , N } be a switching signal that determines the
communication topology. The set H is a set of digraphs with the common vertex set V
and can be denoted as H = {G1, G2, · · · , GN }, where N is the total number of digraphs
in H. Naturally, let Gσ(t) be the subgraph of Gσ(t) induced by V. Thereby, we rewrite the
7
consensus protocols (5) and (6) as:
ui(t) = h(t)(cid:16) Xj∈Ni(σ(t))
and
aij(σ(t))(yji(t) − xi(t)) + bi(σ(t))(y0i(t) − xi(t))(cid:17)
+ g(t)vi(t),
i = 1, 2, · · · , n (17)
vi(t) = a0(t) + γh(t)
·(cid:16) Xj∈Ni(σ(t))
aij(σ(t))(y′
ji(t) − xi(t)) + bi(σ(t))(y′
i = 1, 2, · · · , n (18)
0i(t) − xi(t))(cid:17),
where, Ni(σ(t)) is the set of neighbors of agent i in the digraph Gσ(t) formed by n fol-
lowers; aij(σ(t)) is the (i, j)-th element of the adjacency matrix of Gσ(t), and let Bσ(t) :=
diag(cid:0)b1(σ(t)), b2(σ(t)), · · · , bn(σ(t))(cid:1) represent the leader adjacency matrix associated with
Gσ(t) such that bi(σ(t)) > 0 if and only if 0 ∈ Ni(Gσ(t)).
In parallel with Section 2, substituting the protocols (17), (18) to the system (1), we
can describe the system in the form of the Ito differential equations:
(cid:26)dx(t) = −h(t)(Lσ(t) + Bσ(t))x(t)dt + h(t)Bσ(t)1x0(t)dt + g(t)v(t)dt + h(t)Gσ(t)dW1(t)
dv(t) = a0(t)1dt − γh(t)(Lσ(t) + Bσ(t))x(t)dt + γh(t)Bσ(t)1x0(t)dt + γh(t)Gσ(t)dW2(t)
(19)
where Lσ(t) is the Laplacian matrix of Gσ(t), and
Gσ(t) := diag(cid:16)sb2
1(σ(t)) + Xj∈N1(σ(t))
σ2
j1a2
1j(σ(t)) , · · · ,sb2
n(σ(t)) + Xj∈Nn(σ(t))
σ2
jna2
nj(σ(t))(cid:17).
Let x∗ = x − x01 and v∗ = v − v01 as in Section 3.1. We get an error dynamics of (19)
as follows:
dε(t) = Fσ(t)ε(t)dt + Gσ(t)dW (t),
where ε(t) = (x∗(t), v∗(t))T , W (t) = (W1(t), W2(t))T and
t ≥ 0
(20)
Fσ(t) =(cid:18) −h(t)(Lσ(t) + Bσ(t))
−γh(t)(Lσ(t) + Bσ(t))
g(t)In
0
(cid:19) ,
Gσ(t) = h(t)(cid:18) Gσ(t)
γGσ(t) (cid:19) := h(t)eGσ(t).
In the sequel, we show that under switching topology, the consensus protocols (17) and
(18) ensure that each follower tracks the leader in mean square. We will use the following
lemma.
Lemma 2.[4] Given t ≥ 0 and suppose the digraph Gσ(t) is balanced. Then Lσ(t) +Bσ(t) +
(Lσ(t) + Bσ(t))T is positive definite if and only if the vertex 0 is globally reachable in Gσ(t).
The matrix Lσ(t) + Bσ(t) + (Lσ(t) + Bσ(t))T plays a key role in the convergence analysis
below. Define µ := mint≥0(cid:8)λmin(Lσ(t) + Bσ(t) + (Lσ(t) + Bσ(t))T )(cid:9).
Prior to establishing the main result, we present a condition analogous with Assump-
tion (A2) in Section 3.1:
(A5) There is a δ > 0, such that h(t)
Remark 8.
It is easily shown that µ > 0 under the assumptions of Theorem 2 below by
exploiting Lemma 2 and the fact that H is a finite set. This validates the expression in
Assumption (A5).
2γ(1−γ 2)µ + δ for t ≥ 0.
g(t) >
1
8
Theorem 2. For system (1) with the consensus protocols (17) and (18), if for any t ≥ 0,
Gσ(t) is balanced, and vertex 0 is globally reachable in Gσ(t), then under Assumptions (A3)-
(A5), we have
Ekε(t)k2 = 0.
lim
t→∞
(21)
−γIn
−γIn
By Ito formula and (20), we have
Straightforward calculation yields
Proof. Let eI :=(cid:18) In
In (cid:19). Obviously, we have λmin(eI) = 1 − γ and λmax(eI) =
1 + γ. Hence eI is a positive definite matrix by recalling 0 < γ < 1. Define a Lyapunov
function V (t) = εT (t)eIε(t).
dV (t) = εT (t)(cid:0)eIFσ(t) + F T
σ (t)eIGσ(t))dt + 2εT (t)eIGσ(t)dW (t).
h(t) !
σ (t)eI = −h(t) (1 − γ2)(cid:0)Lσ(t) + Bσ(t) + (Lσ(t) + Bσ(t))T(cid:1) −In
eIFσ(t) + F T
eQσ(t) is positive definite with ν := mint≥0(cid:8)λmin(eQσ(t))(cid:9) > 0.
By using the Haynsworth inertia additivity formula [21], Assumption (A5), we get that
σ (t)eI(cid:1)ε(t)dt + tr(GT
:= −h(t)eQσ(t).
g(t)
h(t)
g(t)
2γIn
Therefore, we have
−In
g(t)
h(t)
dV (t) ≤ −h(t)
ν
1 + γ
V (t)dt + h2(t)tr(cid:0)eGT
σ(t)eIeGσ(t)(cid:1)dt + 2h(t)εT (t)eIeGσ(t)dW (t).
The remaining proofs are similar with those in Theorem 1 by noting that H is a finite
(22)
set, and hence omitted. ✷
Remark 9. From Theorem 2 we see that the designed protocol may guarantee the state
of each follower tracks that of the leader in mean square even under the switching topology.
4. Numerical example
In this section, we provide a numerical simulation to illustrate the theoretical results.
We consider a network consisting of four agents including one leader labeled by vertex 0,
as shown in Fig. 1. The digraph in this figure is assumed to have 0−1 weights. With simple
calculation, it is not hard to solve P from Equation (13) and obtain λmax(P ) = 0.9447. We
take σij = 0.1 for all i, j, h(t) = 1
6(t+2) and γ = 0.5. Therefore, Assumptions
(A1)-(A4) in Theorem 1 hold.
t+2 , g(t) = 1
The simulation results for the consensus errors x∗ and v∗ are shown in Fig. 2 and Fig.
3 respectively, with initial value ε(0) = (−2, 1.5, 3, 2, −1.5, −1)T . From Fig. 2 and Fig. 3,
we can see that three followers can eventually follow the leader.
5. Conclusion
This paper studies a leader-following coordination problem for multi-agent systems
with a time-varying leader under measurement noises. Although the state of the leader
keeps changing and the measured information by each follower is corrupted by white noises,
we propose a neighborhood-based protocol for each agent to follow the leader. We present
sufficient conditions for each follower to track the leader in mean square under directed
fixed topologies. Sufficient conditions are also provided when the interaction topology
9
is switching and the subgraph formed by the followers is balanced. Finally, numerical
simulations are presented to illustrate the theoretical results. Topics worth investigating
in the future include time-delay cases and the design of almost sure consensus protocols.
References
[1] J. A. Fax, R. M. Murray, Information flow and cooperative control of vehicle forma-
tions. IEEE Trans. Autom. Control, 49(2004), pp.1465–1476
[2] C. Godsil, G. Royle, Algebraic Graph Theory. Spring, New York, 2001
[3] Y. Hong, J. Hu, L. Gao, Tracking control for multi-agent consensus with an active
leader and variable topology. Automatica, 42(2006) pp.1177–1182
[4] J. Hu, Y. Hong, Leader-following coordination of multi-agent systems with coupling
time delays. Physica A, 374(2006) pp.853–864
[5] M. Huang, J. H. Manton, Coordination and consensus of networked agents with noisy
measurement: stochastic algorithms and asymptotic behavior. submitted to SIAM
Journal on Control and Optimization: Special issue on Control and Optimization in
Cooperative Networks, 2006
[6] M. Huang, J. H. Manton, Stochastic consensus seeking with measurement noise:
convergence and asymptotic normality. Proc. American Control Conference, 2008,
pp.1337–1342
[7] M. Huang, J. H. Manton, Stochastic approximation for consensus seeking: mean
square and almost sure convergence. Proc. 46th IEEE Conference on Decision and
Control, 2008, pp.306–311
[8] D. B. Kingston, W. Ren, R. W. Beard, Consensus algorithms are input-to-state stable.
Proc. American Control Conference, 2005, pp.1686–1690
[9] T. Li, J. Zhang, Mean square average consensus under measurement noises and fixed
topologies: necessary and sufficient conditions. Automatica, 2009, to appear.
[10] T. Li, J. Zhang, Consensus conditions of multi-agent systems with time-varying
topologies and stochastic communication noises. IEEE Trans. Autom. Control, 2009,
to appear.
[11] Y. Liu and F. Yang, Some theoretical results of synchronization of a linearly coupled
dynamical system with random perturbation on a network. J. Phys. A: Math. Theor.,
42(2009) 065101
[12] N. Lynch, Distributed Algorithms. Morgan Kaufmann, San Matero, 1996
[13] C. Ma, T. Li, J. Zhang, Leader-following consensus control for multi-agent systems
under measurement noises. Proc. 17th IFAC Congress, 2008, pp.1528–1533
[14] A. N. Michel, R. K. Miller, Qualitative Analysis of Large Scale Dynamical Systems.
Academic Press, New York, 1977
10
[15] M. B. Nevelson, R. A. Hasminskii, Stochastic Approximation and Recursive Estima-
tion. American Mathematical Society, Providence RI, 1976
[16] R. Olfati-Saber, Flocking for multi-agent dynamic systems: Algorithms and theory.
IEEE Trans. Autom. Control, 51(2006) pp.401–420
[17] R. Olfati-Saber, J. A. Fax, R. M. Murray, Consensus and cooperation in networked
multi-agent systems. Proceedings of the IEEE, 95(2007) pp.215–233
[18] R. Olfati-Saber, R. M. Murray, Consensus problems in networks of agents with switch-
ing topology and time-delays. IEEE Trans. Autom. Control, 49(2004) pp.1520–1533
[19] B. Øksendal, Stochastic Differential Equations: An Introduction with Applications.
Springer-Verlag, Berlin, 2003
[20] K. Peng, Y. Yang, Leader-following consensus problem with a varying-velocity leader
and time-varying delays. Physica A, 388(2009) pp.193–208
[21] S. Puntanen, G. P. H. Styan, Historical introduction: Issai Schur and the early de-
velopment of the Schur complement: in The Schur Complement and Its Applications,
F. Zhang, Ed. Springer, New York, 2004, pp.1–16
[22] W. Ren, Multi-vehicle consensus with a time-varying reference state. Systems & Con-
trol Letters, 56(2007) pp.474–483
[23] W. Ren, R. W. Beard, Consensus seeking in multi-agent systems under dynamically
changing interaction topologies. IEEE Trans. on Autom. Control, 50(2005) pp.655–
661
[24] W. Ren, R. W. Beard, E. M. Atkins, A survey of consensus problems in multi-agent
coordination, Proc. American Control Conference, 2005, pp.1859–1864
[25] W. Ren, R. W. Beard, D. B. Kingston, Multi-agent Kalman consensus with relative
uncertainty. Proc. American Control Conference, 2005, pp.1865–1870
[26] A. Sinha, D. Ghose, Generalization of linear cyclic pursuit with application to ren-
dezvous of multiple autonomous agents. IEEE Trans. Autom. Control, 51(2006),
pp.1819–1824
[27] H. Su, W. Zhang, Second-order consensus of multiple agents with coupling delay.
Commun. Theor. Phys., 51(2009) pp.101–109
[28] Y.-Z. Sun, J. Ruan, Leader-follower consensus problems of multi-agent systems with
noise perturbation and time delays. Chin. Phys. Lett., 25(2008) pp.3493–3495
[29] J. N. Tsitsiklis, D. P. Bertsekas, M. Athans, Distributed asynchronous determinis-
tic and stochastic gradient optimization algorithms. IEEE Trans. Autom. Control,
31(1986) pp.803–812
[30] T. Vicsek, A. Czir´ok, E. Ben-Jacob, I. Cohen, O. Shochet, Novel type of phase tran-
sition in a system of self-driven particles. Phys. Rev. Lett., 75(1995) pp.1226–1229
[31] J. Wang, J. Hu, D. Cheng, Consensus problem of multi-agent systems with an active
leader and time delay. Proc. Chinese Control Conference, 2006, pp.340–345
11
Figure captions
Fig. 1 Directed network G of four agents involving one leader. G has 0 − 1 weights.
Fig. 2 Consensus error x∗ for the agents.
Fig. 3 Consensus error v∗ for the agents.
0
✲
✒
2
✻
✠
❄
1
✲
3
3
2
1
agent 3
agent 2
*
x
0
agent 1
10
20
30
40
50
Time
60
70
80
90
100
agent 3
−1
−2
−3
0
4
3
2
agent 2
*
v
1
0
−1
−2
0
agent 1
10
20
30
40
60
70
80
90
100
50
Time
12
|
1905.08087 | 2 | 1905 | 2019-08-19T12:26:07 | A Regularized Opponent Model with Maximum Entropy Objective | [
"cs.MA",
"cs.AI",
"cs.LG"
] | In a single-agent setting, reinforcement learning (RL) tasks can be cast into an inference problem by introducing a binary random variable o, which stands for the "optimality". In this paper, we redefine the binary random variable o in multi-agent setting and formalize multi-agent reinforcement learning (MARL) as probabilistic inference. We derive a variational lower bound of the likelihood of achieving the optimality and name it as Regularized Opponent Model with Maximum Entropy Objective (ROMMEO). From ROMMEO, we present a novel perspective on opponent modeling and show how it can improve the performance of training agents theoretically and empirically in cooperative games. To optimize ROMMEO, we first introduce a tabular Q-iteration method ROMMEO-Q with proof of convergence. We extend the exact algorithm to complex environments by proposing an approximate version, ROMMEO-AC. We evaluate these two algorithms on the challenging iterated matrix game and differential game respectively and show that they can outperform strong MARL baselines. | cs.MA | cs | A Regularized Opponent Model with Maximum Entropy Objective
Zheng Tian1∗ , Ying Wen1∗ , Zhichen Gong1 , Faiz Punakkath1 , Shihao Zou2 and Jun Wang 1
1University College London
2University of Alberta
{zheng.tian, ying.wen, jun.wang}@cs.ucl.ac.uk
9
1
0
2
g
u
A
9
1
]
A
M
.
s
c
[
2
v
7
8
0
8
0
.
5
0
9
1
:
v
i
X
r
a
Abstract
less well studied.
In a single-agent setting, reinforcement learning
(RL) tasks can be cast into an inference problem
by introducing a binary random variable o, which
stands for the "optimality". In this paper, we rede-
fine the binary random variable o in multi-agent set-
ting and formalize multi-agent reinforcement learn-
ing (MARL) as probabilistic inference. We derive a
variational lower bound of the likelihood of achiev-
ing the optimality and name it as Regularized Op-
ponent Model with Maximum Entropy Objective
(ROMMEO). From ROMMEO, we present a novel
perspective on opponent modeling and show how
it can improve the performance of training agents
theoretically and empirically in cooperative games.
To optimize ROMMEO, we first introduce a tab-
ular Q-iteration method ROMMEO-Q with proof
of convergence. We extend the exact algorithm
to complex environments by proposing an approx-
imate version, ROMMEO-AC. We evaluate these
two algorithms on the challenging iterated matrix
game and differential game respectively and show
that they can outperform strong MARL baselines.
1 Introduction
Casting decision making and optimal control as an inference
problem have a long history, which dates back to [Kalman,
1960] where the Kalman smoothing is used to solve opti-
mal control in linear dynamics with quadratic cost. Bayesian
methods can capture the uncertainties regarding the transi-
tion probabilities, the rewards functions in the environment or
other agents' policies. This distributional information can be
used to formulate a more structured exploration/exploitation
strategy than those commonly used in classical RL, e.g. -
greedy. A common approach in many works [Toussaint and
Storkey, 2006; Rawlik et al., 2013; Levine and Koltun, 2013;
Abdolmaleki et al., 2018] for framing RL as an inference
problem is by introducing a binary random variable o which
represents "optimality". By this way, RL problems are able to
lend itself to powerful inference tools [Levine, 2018]. How-
ever, the Bayesian approach in a multi-agent environment is
∗The first two authors contributed equally.
In many single-agent works, maximizing entropy is part of
a training agent's objective for resolving ambiguities in in-
verse reinforcement learning [Ziebart et al., 2008], improv-
ing the diversity [Florensa et al., 2017], robustness [Fox et
al., 2015] and the compositionality [Haarnoja et al., 2018a]
of the learned policy.
In Bayesian RL, it often presents in
the evidence lower bound (ELBO) for the log likelihood of
optimality [Haarnoja et al., 2017; Schulman et al., 2017;
Haarnoja et al., 2018b], commonly known as maximum en-
tropy objective (MEO), which encourages the optimal policy
to maximize the expected return and long term entropy.
In MARL, there is more than one agent interacting with a
stationary environment. In contrast with the single agent en-
vironment, an agent's reward not only depends on the current
environment state and the agent's action but also on the ac-
tions of others. The existence of other agents increases the
uncertainty in the environment. Therefore, the capability of
reasoning about other agents' belief, private information, be-
havior, strategy, and other characteristics is crucial. A rea-
soning model can be used in many different ways, but the
most common case is where an agent utilize its reasoning
model to help its self decision making [Brown, 1951; Hein-
rich and Silver, 2016; He et al., 2016; Raileanu et al., 2018;
Wen et al., 2019; Tian et al., 2018].
In this work, we use
the word "opponent" when referring to another agent in the
environment irrespective of the environment's cooperative or
adversarial nature.
In our work, we reformulate the MARL problem into
Bayesian inference and derive a multi-agent version of MEO,
which we call the regularized opponent model with maxi-
mum entropy objective (ROMMEO). Optimizing this objec-
tive with respect to one agent's opponent model gives rise to
a new perceptive on opponent modeling. We present two off-
policy RL algorithms for optimizing ROMMEO in MARL.
ROMMEO-Q is applied in discrete action case with proof of
convergence. For the complex and continuous action environ-
ment, we propose ROMMEO Actor-Critic (ROMMEO-AC),
which approximates the former procedure and extend itself
to continuous problems. We evaluate these two approaches
on the matrix game and the differential game against strong
baselines and show that our methods can outperform all the
baselines in terms of the overall performance and speed of
convergence.
(cid:34) ∞(cid:88)
2 Method
2.1 Stochastic Games
For an n-agent stochastic game [Shapley, 1953], we define a
tuple (S,A1, . . . ,An, R1, . . . , Rn, p,T , γ), where S denotes
the state space, p is the distribution of the initial state, γ is a
discount factor, Ai and Ri = Ri(s, ai, a−i) are the action
space and the reward function for agent i ∈ {1, . . . , n} re-
spectively. States are transitioned according to T : S × A,
where A = {A1,··· ,An}. Agent i chooses its action
ai ∈ Ai according to the policy πi
θi(ais) parameterized by
θi conditioning on some given state s ∈ S. Let us define
the joint policy as the collection of all agents' policies πθ
with θ representing the joint parameter. It is convenient to
interpret the joint policy from the perspective of agent i such
θi(ais), π−i
θ−i(a−is)), where a−i = (aj)j(cid:54)=i,
that πθ = (πi
θ−i(a−is) is a compact representation
θ−i = (θj)j(cid:54)=i, and π−i
of the joint policy of all complementary agents of i. At each
stage of the game, actions are taken simultaneously. Each
agent is presumed to pursue the maximal cumulative reward,
expressed as
(cid:35)
t, a−i
t )
,
(1)
max ηi(πθ) = E
t, a−i
t=1
t ) sample from (πi
γtRi(st, ai
θi, π−i
θ−i). In fully cooperative
with (ai
games, we assume there exists at least one joint policy πθ
such that all agents can achieve the maximal cumulative re-
ward with this joint policy.
2.2 A Variational Lower Bound for Multi-Agent
Reinforcement Learning Problems
We transform the control problem into an inference problem
by introducing a binary random variable oi
t which serves as
the indicator for "optimality" for each agent i at each time
step t. Recall that in single agent problem, reward R(st, at) is
bounded, but the achievement of the maximum reward given
the action at is unknown. Therefore, in the single-agent case,
ot indicates the optimality of achieving the bounded maxi-
mum reward r∗
t . It thus can be regarded as a random vari-
able and we have P (ot = 1st, at) ∝ exp(R(st, at)). Intu-
itively, this formulation dictates that higher rewards reflect a
higher likelihood of achieving optimality, i.e., the case when
ot = 1. However, the definition of "optimality" in the multi-
agent case is subtlety different from the one in the single-
agent situation.
reinforcement
cooperative multi-agent
learning
(CMARL), to define agent i's optimality, we first introduce
the definition of optimum and optimal policy:
Definition 1.
In cooperative multi-agent reinforcement
learning, optimum is a strategy profile (π1∗, . . . , πn∗) such
that:
In
(cid:35)
where π = (πi, π−i) and Agent i's optimal policy is πi∗.
In CMARL, a single agent's "optimality" oi
t cannot imply
that it obtains the maximum reward because the reward de-
pends on the joint actions of all agents (ai, a−i). Therefore,
t = 1 only indicates that agent i's policy at time
we define oi
step t is optimal. The posterior probability of agent i's opti-
t is the probability that the action is
mality given its action ai
sampled from the optimal policy:
P (oi
tai
t) = P (ai
t ∼ πi∗ai
t) = πi∗(ai
t).
(3)
We also assume that given other players playing optimally
(o−i
t = 1), the higher the reward agent i receives the higher
the probability of agent i's current policy is optimal (oi
t = 1):
t, a−i
t ) ∝ exp(R(st, ai
t = 1o−i
t = 1, st, ai
t, a−i
t )).
P (oi
(4)
The conditional probabilities of "optimality" in both of
CMARL and single-agent case have similar forms. However,
it is worth mentioning that the "optimality" in CMARL has a
different interpretation to the one in single-agent case.
For cooperative games, if all agents play optimally, then
agents can receive the maximum rewards, which is the opti-
mum of the games. Therefore, given the fact that other agents
are playing their optimal policies o−i = 1, the probability that
agent i also plays its optimal policy P (oi = 1o−i = 1) is the
probability of obtaining the maximum reward from agent i's
perspective. Therefore, we define agent i's objective as:
max J ∆= log P (oi
1:T = 1o−i
1:T = 1)
(5)
1:T , a−i
1:T , a−i
1:T , s1:Toi
1:T , s1:Toi
As we assume no knowledge of the optimal policies and the
model of the environment, we treat them as latent variables.
To optimize the observed evidence defined in Eq. 5, there-
fore, we use variational inference (VI) with an auxiliary dis-
tribution over these latent variables q(ai
1:T =
1, o−i
1:T = 1). Without loss of generality, we here derive the
solution for agent i. We factorize q(ai
1:T =
1, o−i
1:T = 1) so as to capture agent i's conditional policy on
the current state and opponents actions, and beliefs regard-
ing opponents actions. This way, agent i will learn optimal
policy, while also possessing the capability to model oppo-
nents actions a−i. Using all modelling assumptions, we may
1:T , a−i
factorize q(ai
(cid:89)
1:T , s1:Toi
1:T , a−i
= P (s1)
(cid:89)
× q(a−i
= P (s1)
1:T = 1, o−i
P (st+1st, at)q(ai
t = o−i
P (st+1st, at)π(ai
1:T = 1)
ta−i
1:T = 1, o−i
1:T , s1:Toi
1:T = 1) as:
t = o−i
t )ρ(a−i
tst, a−i
st, oi
t = 1)
t = 1)
st),
, st, oi
q(ai
t
t
t
t
(cid:34) ∞(cid:88)
(cid:34) ∞(cid:88)
t=1
γtRi(st, ai∗
)
t
t , a−i∗
(cid:35)
t, a−i
t )
E
s∼ps,ai∗
t ∼πi∗,a
−i∗
t ∼π−i∗
s∼ps,ai
≥ E
∀π ∈ Π, i ∈ (1 . . . n),
t∼πi,a
−i
t ∼π−i
γtRi(st, ai
(2)
t
t=1
where we have assumed the same initial and states transitions
as in the original model. With this factorization, we derive a
lower bound on the likelihood of optimality of agent i:
log P (oi
≥ J (π, ρ) ∆=
+ H(π(ai
(cid:88)
1:T = 1o−i
1:T = 1)
t, a−i
E
t )
(st,ai
tst, a−i
t )) − DKL(ρ(a−i
st)P (a−i
st))]
(cid:124)
(cid:123)(cid:122)
t, a−i
t ∼ρ[Ri(st, ai
t ) + H(π(ai
−i
t∼π,a
ai
t )∼q[Ri(st, ai
−i
t,a
Est[E
t
t
t
−i
(cid:123)(cid:122)
t ∼ρ[DKL(ρ(a−i
t
Regularizer of ρ
MEO
st)P (a−i
st))]
].
t
(cid:125)
(cid:88)
(cid:124)
a
t
=
− E
tst, a−i
t ))
]
(6)
(cid:125)
(7)
t
t
t
t
t
st, o−i
st, o−i
tst, a−i
t = 1, o−i
1:T = 1o−i
t = 1) and P (a−i
Written out in full, ρ(a−i
t = 1) is agent i's op-
ponent model estimating optimal policies of its opponents,
t = 1, o−i
t = 1) is the agent i's conditional
π(ai
, oi
t = o−i
policy at optimum (oi
t = 1)
is the prior of optimal policy of opponents. In our work, we
st, o−i
set the prior P (a−i
t = 1) equal to the observed em-
pirical distribution of opponents' actions given states. As we
are only interested in the case where (oi
t = 1), we
st) here and thereafter. H(·) is
drop them in π, ρ and P (a−i
the entropy function. Eq. 6 is a variational lower bound of
1:T = 1) and the derivation is deferred to
log P (oi
Appendix B.1.
2.3 The Learning of Opponent Model
We can further expand Eq. 6 into Eq. 7 and we find that it
resembles the maximum entropy objective in single-agent re-
inforcement learning [Kappen, 2005; Todorov, 2007; Ziebart
et al., 2008; Haarnoja et al., 2017]. We denote agent i's ex-
pectation of reward R(st, ai
t ) plus entropy of the condi-
tional policy H(π(ais, a−i)) as agent i's maximum entropy
objective (MEO). In the multi-agent version, however, it is
worthy of noting that optimizing the MEO will lead to the
optimization of ρ. This can be counter-intuitive at first sight
as opponent behaviour models are normally trained with only
past state-action data (s, a−i) to predict opponents' actions.
t = 1) is modelling op-
ponents' optimal policies in our work. Given agent i's policy
πi being fixed, optimizing MEO with respect to ρ updates
agent i's opponent model in the direction of the higher shared
reward R(s, ai, a−i) and the more stochastic conditional pol-
icy πi(ais, a−i), making it closer to the real optimal policies
of the opponents. Without any regularization, at iteration d,
agent i can freely learn a new opponent model ρi
d+1 which
is the closest to the optimal opponent policies π−i∗ from its
d(ais, a−i). Next, agent i can optimize
perspective given πi
the lower bound with respect to πi
d+1.
Then we have an EM-like iterative training and can show
it monotonically increases the probability that the opponent
model ρ is optimal policies of the opponents. Then, by act-
ing optimally to the converged opponent model ρi∞, we can
recover agent i's optimal policy πi∗.
However, recall that ρ(a−i
d+1(ais, a−i) given ρi
st, o−i
t, a−i
However, it is unrealistic to learn such an opponent model.
As the opponents have no access to agent i's conditional pol-
d(ais, a−i), the learning of its policy can be different
icy πi
t
from the one of agent i's opponent model. Then the actual op-
ponent policies π−i
d+1 can be very different from agent i's con-
verged opponent model ρi∞ learned in the above way given
d(ais, a−i). Therefore, acting
agent i's conditional policy πi
optimally to an opponent model far from the real opponents'
policies can lead to poor performance.
The last
st)P (a−i
term in Eq. 7 can prevent agent i build-
ing an unrealistic opponent model. The Kullback-Leibler
(KL) divergence between opponent model and a prior
st)) can act as a regularizer of ρ. By
DKL(ρ(a−i
setting the prior to the empirical distribution of opponent past
behaviour, the KL divergence penalizes ρ heavily if it deviates
from the empirical distribution too much. As the objective in
Eq. 7 can be seen as a Maximum Entropy objective for one
agent's policy and opponent model with regularization on the
opponent model, we call this objective as Regularized Oppo-
nent Model with Maximum Entropy Objective (ROMMEO).
t
t
3 Multi-Agent Soft Actor Critic
To optimize the ROMMEO in Eq. 7 derived in the previous
section, we propose two off-policy algorithms. We first intro-
duce an exact tabular Q-iteration method with proof of con-
vergence. For practical implementation in a complex continu-
ous environment, we then propose the ROMMEO actor critic
ROMMEO-AC, which is an approximation to this procedure.
3.1 Regularized Opponent Model with Maximum
Entropy Objective Q-Iteration
In this section, we derive a multi-agent version of Soft Q-
iteration algorithm proposed in [Haarnoja et al., 2017] and we
name our algorithm as ROMMEO-Q. The derivation follows
from a similar logic to [Haarnoja et al., 2017], but the exten-
sion of Soft Q-learning to MARL is still nontrivial. From this
section, we slightly modify the objective in Eq. 7 by adding
a weighting factor α for the entropy term and the original ob-
jective can be recovered by setting α = 1.
We first define multi-agent soft Q-function and V-function
respectively. Then we can show that the conditional policy
and opponent model defined in Eq. 10 and 11 below are opti-
mal solutions with respect to the objective defined in Eq. 7:
Theorem 1. We define the soft state-action value function of
agent i as
a−i P (a−is)
sof t(s, ai, a−i))
(9)
Then the optimal conditional policy and opponent model for
Eq. 6 are
ai exp( 1
α Q∗
π∗(ais, a−i) =
(cid:80)
exp( 1
ai exp( 1
α Qπ∗,ρ∗
sof t (s, ai, a−i))
α Qπ∗,ρ∗
sof t (s, ai, a−i))
,
(10)
Qπ∗,ρ∗
sof t (st, ai
∗
+ αH(π
(ai
−i
t ) = rt + E
t+l, st+l)) − DKL(ρ
−i
(st+l,ai
t, a
t+la
t+l,a
∗
and soft state value function as
V ∗(s) = log(cid:80)
(cid:16)(cid:80)
∞(cid:88)
−i
t+l,...)∼q[
t+lst+l)P (a
t+lst+l))],
−i
−i
(a
γl(rt+l
l=1
(8)
(cid:17)α
,
and
P (a−is)
ρ∗(a−is) =
(cid:16)(cid:80)
sof t(s, ai, a−i))
ai exp( 1
α Q∗
exp(V ∗(s))
(cid:17)α
.
(11)
Proof. See Appendix C.2.
Following from Theorem 1, we can find the optimal solu-
tion of Eq. 7 by learning the soft multi-agent Q-function first
and recover the optimal policy π∗ and opponent model ρ∗ by
Equations 10 and 11. To learn the Q-function, we show that it
satisfies a Bellman-like equation, which we name it as multi-
agent soft Bellman equation:
Theorem 2. We define the soft multi-agent Bellman equation
sof t(s, ai, a−i) of
for the soft state-action value function Qπ,ρ
agent i as
sof t(s, ai, a−i) = rt + γE(st+1)[Vsof t(st+1)].
Qπ,ρ
(12)
Proof. See Appendix C.3.
sof t and V ∗
With this Bellman equation defined above, we can derive a
solution to Eq. 12 with a fixed point iteration, which we call
ROMMEO Q-iteration (ROMMEO-Q). Additionally, We can
show that it can converge to the optimal Q∗
sof t with
certain restrictions as stated in [Wen et al., 2019]:
Theorem 3. ROMMEO Q-iteration.
with only one global optimum,
Eπ
Qsof t(·,·,·) and Vsof t(·) be bounded and assume
Eπ∗(cid:2)Qi
t(s)(cid:3) ≥
t(s)(cid:3), where π∗ is the optimal strategy profile. Let
(cid:2)Qi
(cid:33)α
(cid:88)
sof t(s, ai, a−i))
Q∗
In a symmetric game
(cid:32)(cid:88)
P (a−is)
< ∞
exp(
i.e.
1
α
a−i
ai
and that Q∗
sof t < ∞ exists. Then the fixed-point iteration
where
(cid:16)(cid:80)
Qsof t(st, ai
log(cid:80)
t, a−i
Vsof t(st)
α Qsof t(st, ai
t ) ← rt + γE(st+1)[Vsof t(st+1)],
P (a−i
t, a−i
←
t, a−i
t ))
sof t respectively.
, ∀st, ai
(cid:17)α
−i
t
a
(13)
st) ×
t
, converges
t
ai
t
exp( 1
sof t and V ∗
to Q∗
Proof. See Appendix C.3.
3.2 Regularized Opponent Model with Maximum
Entropy Objective Actor Critic
The ROMMEO-Q assumes we have the model of the envi-
ronment and is impractical to implement in high-dimensional
continuous problems. To solve these problems, we pro-
pose the ROMMEO actor critic (ROMMEO-AC) which is
a model-free method. We use neural networks (NNs) as
function approximators for the conditional policy, opponent
model and Q-function and learn these functions by stochastic
gradient. We parameterize the Q-function, conditional policy
and opponent model by Qω(s, ai, a−i), πθ(ai
t ) and
ρφ(a−i
st) respectively.
tst, a−i
t
Without access to the environment model, we first replace
the Q-iteration with Q-learning. Therefore, we can train ω to
minimize:
(14)
JQ(ω) = E
(st,ai
t,a
− R(st, ai
with
¯V (st+1) = Q¯ω(st+1, ai
− α log πθ(ai
1
2
t, a−i
t )
(Qω(st, ai
t )∼D[
−i
t ) − γEst+1∼ps [ ¯V (st+1)])2],
t, a−i
t+1st+1)
t+1, a−i
t+1st+1, a−i
t+1) − log ρφ(a−i
t
t+1) + logP (a−i
t+1st+1),
(15)
where Q¯ω are target functions for providing relatively stable
target values. We use a−i
t denoting the action sampled from
st) and it should be distin-
agent i's opponent model ρ(a−i
guished from a−i
t which is the real action taken by agent i's
opponent. Eq. 15 can be derived from Eq. 10 and 11.
To recover the optimal conditional policy and oppo-
nent model and avoid intractable inference steps defined in
Eq. 10 and 11 in complex problems, we follow the method
in [Haarnoja et al., 2018b] where θ and φ are trained to min-
imize the KL-divergence:
−i
st∼D,a
t ∼ρ
πθ(·st, a−i
t )
Jπ(θ) = E
(16)
(cid:32)
ρ(·st)
(st,ai
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) exp( 1
α Qω(st,·, a−i
t ))
Zω(st, a−i
t )
(cid:17)α
(cid:16) exp( 1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) P (·st)
α Q(st,ai
tst,·)
πθ(ai
Zω(st)
t,·))
t)∼D
(cid:34)
DKL
,
(cid:33)(cid:35)
.
DKL
Jρ(φ) = E
(17)
By using the reparameterization trick: a−i
t = fθ(i
ai
; st) and
t ), we can rewrite the objectives above as:
st∼D,i
t = gφ(−i
Jπ(θ) = E
t; st, a−i
t ))
t; st, a−i
t
(19)
The gradient of Eq. 14, 18 and 19 with respect to the cor-
t
responding parameters are listed as below:
∇ωJQ(ω) = ∇ωQω(st, ai
t, a−i
tst, a−i
tst, a−i
− R(st, ai
∇θJπ(θ) = ∇θα log πθ(ai
α log πθ(ai
t, a−i
t )(Qω(st, ai
t ) − γ ¯V (st+1)),
t ) + ∇θfθ(i
t ) − ∇ai
(∇ai
t
Qω(st, ai
t
t, a−i
t )
(20)
t; st, a−i
t )
t, a−i
t )),
(21)
t
st) + (∇a
∇φJρ(φ) = ∇φ log ρφ(a−i
st) − ∇a
log P (a−i
α log πθ(aist, a−i
st)
t, a−i
t )
; st). (22)
We list the pseudo-code of ROMMEO-Q and ROMMEO-
log ρφi(a−i
t
Qωi(st, ai
t ))∇φgφ(−i
− ∇a
+ ∇a
−i
t
−i
t
−i
t
−i
t
t
t
AC in Appendix A.
t∼N,a
− Qω(st, fθ(i
(st,at)∼D,
Jρ(φ) = E
− logP (a−i
+ α log πθ(ai
t
t )],
t ), a−i
t ∼ρ[α log πθ(fθ(i
−i
t; st, a−i
t ∼N [log ρφ(gφ(−i
st) − Q(st, ai
t, gφ(−i
tst, gφ(−i
; st))].
−i
t
t
(18)
; st)st)
; st))
(a)
(b)
(c)
Figure 1: (a): Learning curves of ROMMEO and baselines on ICG over 100 episodes. (b): Probability of convergence to the global optimum
for ROMMEO and baselines on ICG over 100 episodes. The vertical axis is the joint probability of taking actions A for both agents. (c):
Probability of taking A estimated by agent i's opponent model ρi and observed empirical frequency P i in one trail of training, i ∈ {1, 2}.
4 Related Works
In early works, the maximum entropy principle has been
used in policy search in linear dynamics [Todorov, 2010;
Toussaint, 2009; Levine and Koltun, 2013] and path integral
control in general dynamics [Kappen, 2005; Theodorou et al.,
2010]. Recently, off-policy methods [Haarnoja et al., 2017;
Schulman et al., 2017; Nachum et al., 2017] have been pro-
posed to improve the sample efficiency in optimizing MEO.
To avoid complex sampling procedure, training a policy in su-
pervised fashion is employed in [Haarnoja et al., 2018b]. Our
work is closely related to this series of recent works because
ROMMEO is an extension of MEO to MARL.
A few related works to ours have been conducted in multi-
agent soft Q-learning [Wen et al., 2019; Wei et al., 2018;
Grau-Moya et al., 2018], where variants of soft Q-learning
are applied for solving different problems in MARL. How-
ever, unlike previous works, we do not take the soft Q-
learning as given and apply it to MARL problems with mod-
ifications. In our work, we first establish a novel objective
ROMMEO and ROMMEO-Q is only an off-policy method
we derive with complete convergence proof, which can opti-
mize the objective. There are other ways of optimizing ROM-
MEO, for example, the on-policy gradient-based methods,
but they are not included in the paper.
There has been substantial progress in combining RL with
probabilistic inference. However, most of the existing works
focus on the single-agent case. The literature of Bayesian
methods in MARL is limited. Among these are methods
performing on cooperative games with prior knowledge on
distributions of the game model and the possible strategies
of others [Chalkiadakis and Boutilier, 2003] or policy pa-
rameters and possible roles of other agents [Wilson et al.,
2010].
In our work, we assume very limited prior knowl-
edge of the environment model, optimal policy, opponents
or the observations during the play.
In addition, our al-
gorithms are fully decentralized at training and execution,
which is more challenging than problems from the cen-
tralized training [Foerster et al., 2017; Lowe et al., 2017;
Rashid et al., 2018].
In our work, we give a new definition of optimality in
CMARL and derive a novel objective ROMMEO. We provide
two off-policy RL algorithms for optimizing ROMMEO and
the exact version comes with convergence proof. In addition,
we provide a natural perspective on opponent modeling in co-
ordination problems: biasing one's opponent model towards
the optimum from its perspective but regularizing it with the
empirical distribution of opponent's real behavior.
Iterated Matrix Games
5 Experiments
5.1
We first present the proof-of-principle result of ROMMEO-
Q1 on iterated matrix games where players need to cooperate
to achieve the shared maximum reward. To this end, we study
the iterated climbing games (ICG) which is a classic purely
cooperative two-player stateless iterated matrix games.
(CG)
is
[Claus
a
fully
and Boutilier,
summarized
cooperative
1998]
follows:
Climbing
game
game
whose
R =
A
B
C
in
proposed
payoff matrix
(11, 11)
A
(−30,−30)
(0, 0)
is
B
(−30,−30)
(7, 7)
(0, 0)
as
.
C
(0, 0)
(6, 6)
(5, 3)
It
is
a challenging benchmark because of the difficulty of conver-
gence to its global optimum. There are two Nash equilibrium
(A, A) and (B, B) but one global optimal (A, A). The
punishment of miscoordination by choosing a certain action
increases in the order of C → B → A. The safest action is
C and the miscoordination punishment is the most severe for
A. Therefore it is very difficult for agents to converge to the
global optimum in ICG.
We compare our method to a series of strong baselines
including Joint Action Learner (JAL) [Claus
in MARL,
and Boutilier, 1998], WoLF Policy Hillclimbing (WoLF-
PHC) [Bowling and Veloso, 2001], Frequency Maximum Q
(FMQ) [Kapetanakis and Kudenko, 2002] and Probabilistic
Recursive Reasoning (PR2) [Wen et al., 2019]. ROMMEO-
Q-EMP is an ablation study to evaluate the effectiveness of
our proposed opponent model learning process, where we re-
place our opponent model with empirical frequency. Fig. 1a
shows the learning curves on ICG for different algorithms.
The difference of rewards between ROMMEO-Q and FMQ-
c10 may seem small because of the small reward margin
between the global optimum and the local one. However,
ROMMEO-Q actually outperforms all baselines significantly
in terms of converging to the global optimum, which is shown
1The
experiment
code
and appendix are
available
at
https://github.com/rommeoijcai2019/rommeo.
050100150200250300Episodes302010010RewardJALWoLF-PHCPR2FMQ-c10ROMMEO-QROMMEO-Q-EMP050100150200250300Episodes0.000.250.500.751.00P(A,A)JALWoLF-PHCPR2FMQ-c10ROMMEO-QROMMEO-Q-EMP050100150200250300Episodes0.00.20.40.60.81.0p(A)1212 (a)
(b)
(c)
Figure 2: Experiment on Max of Two Quadratic Game. (a) Reward surface and learning path of agents. Scattered points are actions taken at
each step; (b) Learning curve of ROMMEO and baselines. (c) Mean of agents' policies π and opponent models ρ.
in Fig. 1b. To further analyze the opponent modeling de-
scribed in Sec. 2.3, we visualize the probability of agent −i
taking the optimal action A estimated by agent i's opponent
model ρi and its true policy π−i in Fig. 1c. Agent i's op-
ponent model "thinks ahead" of agent −i and converges to
agent −i's optimal policy before agent −i itself converges to
the optimal policy. This helps agent i to respond to its op-
ponent model optimally by choosing action A, which in turn
leads to the improvement of agent −i's opponent model and
policy. Therefore, the game converges to the global optimum.
To note, the big drop of P (A) for both policies and opponent
models at the beginning of the training comes from the severe
punishment of miscoordination associated with action A.
)2 − ( a2−5
1
1
3
3
5.2 Differential Games
We adopt the differential Max of Two Quadratic Game [Wei
et al., 2018] for continuous case. The agents have continu-
ous action space of [−10, 10]. Each agent's reward depends
on the joint action following the equations: r1(cid:0)a1, a2(cid:1) =
r2(cid:0)a1, a2(cid:1) = max (f1, f2) , where f1 = 0.8 × [−( a1+5
)2], f2 = 1.0 × [−( a1−5
)2 −
( a2+5
)2] + 10. We
compare the algorithm with a series of baselines includ-
ing PR2 [Wen et al., 2019], MASQL [Wei et al., 2018;
Grau-Moya et al., 2018], MADDPG [Lowe et al., 2017]
and independent learner via DDPG [Lillicrap et al., 2015].
To compare against traditional opponent modeling methods,
similar to [Rabinowitz et al., 2018; He et al., 2016], we im-
plement an additional baseline of DDPG with an opponent
module that is trained online with supervision in order to
capture the latest opponent behaviors, called DDPG-OM. We
trained all agents for 200 episodes with 25 steps per episode.
This is a challenging task to most continuous gradient
based RL algorithms because gradient update tends to direct
the training agent to the sub-optimal point. The reward sur-
face is provided in Fig. 2a ; there is a local maximum 0 at
(−5,−5) and a global maximum 10 at (5, 5), with a deep
valley staying in the middle. If the agents' policies are initial-
ized to (0, 0) (the red starred point) that lies within the basin
of the left local maximum, the gradient-based methods would
tend to fail to find the global maximum equilibrium point due
to the valley blocking the upper right area.
A learning path of ROMMEO-AC is summarized in Fig.
2a and the solid bright circle on the right corner implies the
convergence to the global optimum. The learning curve is
presented in Fig. 2b, ROMMEO-AC shows the capability
of converging to the global optimum in a limited amount of
steps, while most of the baselines can only reach the sub-
optimal point. PR2-AC can also achieve the global optimum
but requires many more steps to explore and learn. Addition-
ally, fine tuning on the exploration noise or separate explo-
ration stage is required for deterministic RL methods (MAD-
DPG, DDPG, DDPG-OM, PR2-AC), and the learning out-
comes of energy-based RL method (MASQL) is extremely
sensitive to the annealing scheme for the temperature. In con-
trast, ROMMEO-AC employs a stochastic policy and con-
trols the exploration level by the weighting factor α. It does
not need a separate exploration stage at the beginning of the
training or a delicately designed annealing scheme for α.
Furthermore, we analyze the learning path of policy π and
modeled opponent policy ρ during the training, the results
are shown in Fig. 2c. The red and orange lines are mean of
modeled opponent policy ρ, which always learn to approach
the optimal ahead of the policy π (in dashed blue and green
lines). This helps the agents to establish the trust and con-
verge to the optimum quickly, which further justifies the ef-
fectiveness and benefits of conducting a regularized opponent
model proposed in Sec. 2.3.
6 Conclusion
In this paper, we use Bayesian inference to formulate MARL
problem and derive a novel objective ROMMEO which gives
rise to a new perspective on opponent modeling. We design
an off-policy algorithm ROMMEO-Q with complete conver-
gence proof for optimizing ROMMEO. For better general-
ity, we also propose ROMMEO-AC, an actor critic algorithm
powered by NNs to solve complex and continuous problems.
We give an insightful analysis of the effect of the new learning
process of the opponent modeling on agent's performance in
MARL. We evaluate our methods on the challenging matrix
game and differential game and show that they can outper-
form a series of strong base lines. It is worthy of noting that
Theorems 1,2 and 3 only guarantees the convergence to opti-
mal solutions with respect to ROMMEO objective but not the
optimum in the game. The achievement of the optimum in the
game also relies on the opponent learning algorithm. In our
work, we demonstrate that ROMMEO-Q/AC's convergence
to the optimum of the game in self-play setting. The conver-
gence to optimum of games in non-self-play settings will be
studied in our future work.
10505Action of Agent 110505Action of Agent 2-30.0-30.0-26.0-26.0-22.0-22.0-22.0-18.0-18.0-18.0-14.0-14.0-10.0-10.0-6.0-6.0-2.0-2.02.06.002004006008001000Step0255075100125150175200Episodes302010010ReturnROMMEO-ACPR2-ACDDPG-OMDDPGMADDPGMASQL01020304050Episodes10505Mean of Policy1212 References
[Abdolmaleki et al., 2018] Abbas Abdolmaleki, Jost Tobias Sprin-
genberg, Yuval Tassa, R´emi Munos, Nicolas Heess, and Mar-
tin A. Riedmiller. Maximum a posteriori policy optimisation.
CoRR, abs/1806.06920, 2018.
[Bowling and Veloso, 2001] Michael Bowling
and Manuela
Veloso. Rational and convergent learning in stochastic games. In
IJCAI, San Francisco, CA, USA, 2001.
[Brown, 1951] George W. Brown.
Iterative solution of games by
fictitious play. In AAPA. New York, 1951.
[Chalkiadakis and Boutilier, 2003] Georgios Chalkiadakis
and
Craig Boutilier.
Coordination in multiagent reinforcement
learning: A bayesian approach. In AAMAS, pages 709 -- 716, New
York, NY, USA, 2003. ACM.
[Claus and Boutilier, 1998] Caroline Claus and Craig Boutilier.
The dynamics of reinforcement learning in cooperative multia-
gent systems. AAAI '98/IAAI '98, Menlo Park, CA, USA, 1998.
American Association for Artificial Intelligence.
[Florensa et al., 2017] Carlos Florensa, Yan Duan, and Pieter
Abbeel. Stochastic Neural Networks for Hierarchical Reinforce-
ment Learning. arXiv e-prints, April 2017.
[Foerster et al., 2017] Jakob Foerster, Gregory Farquhar, Tri-
antafyllos Afouras, Nantas Nardelli, and Shimon Whiteson.
Counterfactual Multi-Agent Policy Gradients. arXiv e-prints,
page arXiv:1705.08926, May 2017.
[Fox et al., 2015] Roy Fox, Ari Pakman, and Naftali Tishby. Tam-
ing the Noise in Reinforcement Learning via Soft Updates. arXiv
e-prints, page arXiv:1512.08562, December 2015.
[Grau-Moya et al., 2018] Jordi Grau-Moya, Felix Leibfried, and
Haitham Bou-Ammar. Balancing Two-Player Stochastic Games
with Soft Q-Learning. IJCAI, 2018.
[Haarnoja et al., 2017] Tuomas Haarnoja, Haoran Tang, Pieter
Abbeel, and Sergey Levine. Reinforcement learning with deep
energy-based policies. CoRR, abs/1702.08165, 2017.
[Haarnoja et al., 2018a] Tuomas Haarnoja, Vitchyr Pong, Aurick
Zhou, Murtaza Dalal, Pieter Abbeel, and Sergey Levine. Com-
posable Deep Reinforcement Learning for Robotic Manipulation.
arXiv e-prints, page arXiv:1803.06773, March 2018.
[Haarnoja et al., 2018b] Tuomas Haarnoja, Aurick Zhou, Pieter
Abbeel, and Sergey Levine. Soft Actor-Critic: Off-Policy Max-
imum Entropy Deep Reinforcement Learning with a Stochastic
Actor. arXiv e-prints, page arXiv:1801.01290, January 2018.
[He et al., 2016] He He, Jordan Boyd-Graber, Kevin Kwok, and
Hal Daum´e III. Opponent modeling in deep reinforcement learn-
In International Conference on Machine Learning, pages
ing.
1804 -- 1813, 2016.
[Heinrich and Silver, 2016] Johannes Heinrich and David Silver.
learning from self-play in imperfect-
Deep reinforcement
information games. CoRR, abs/1603.01121, 2016.
[Kalman, 1960] Rudolf Kalman. A new approach to linear filtering
and prediction problems. Transactions of the ASME - Journal of
basic Engineering, 82:35 -- 45, 01 1960.
[Kapetanakis and Kudenko, 2002] Spiros Kapetanakis and Daniel
Kudenko. Reinforcement learning of coordination in coopera-
tive multi-agent systems. In Eighteenth National Conference on
Artificial Intelligence, Menlo Park, CA, USA, 2002.
[Kappen, 2005] Hilbert J. Kappen. Path integrals and symmetry
breaking for optimal control theory. Journal of Statistical Me-
chanics: Theory and Experiment, 2005(11), 2005.
[Levine and Koltun, 2013] Sergey Levine and Vladlen Koltun.
Variational policy search via trajectory optimization. In C. J. C.
Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q. Wein-
berger, editors, NIPS '13. 2013.
[Levine, 2018] Sergey Levine. Reinforcement Learning and Con-
trol as Probabilistic Inference: Tutorial and Review. arXiv e-
prints, page arXiv:1805.00909, May 2018.
[Lillicrap et al., 2015] Timothy P Lillicrap,
Jonathan J Hunt,
Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David
Silver, and Daan Wierstra. Continuous control with deep rein-
forcement learning. arXiv preprint arXiv:1509.02971, 2015.
[Lowe et al., 2017] Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb,
Pieter Abbeel, and Igor Mordatch. Multi-Agent Actor-Critic for
Mixed Cooperative-Competitive Environments. arXiv e-prints,
page arXiv:1706.02275, June 2017.
[Nachum et al., 2017] Ofir Nachum, Mohammad Norouzi, Kelvin
Xu, and Dale Schuurmans. Bridging the Gap Between Value
and Policy Based Reinforcement Learning. arXiv e-prints, page
arXiv:1702.08892, February 2017.
[Rabinowitz et al., 2018] Neil C Rabinowitz, Frank Perbet, H Fran-
cis Song, Chiyuan Zhang, SM Eslami, and Matthew Botvinick.
Machine theory of mind. arXiv preprint arXiv:1802.07740, 2018.
[Raileanu et al., 2018] Roberta Raileanu, Emily Denton, Arthur
Szlam, and Rob Fergus. Modeling others using oneself in multi-
agent reinforcement learning. CoRR, abs/1802.09640, 2018.
[Rashid et al., 2018] Tabish Rashid, Mikayel Samvelyan, Christian
Schroeder de Witt, Gregory Farquhar, Jakob Foerster, and Shi-
mon Whiteson. QMIX: Monotonic Value Function Factorisation
for Deep Multi-Agent Reinforcement Learning. arXiv e-prints,
page arXiv:1803.11485, March 2018.
[Rawlik et al., 2013] Konrad Rawlik, Marc Toussaint, and Sethu
Vijayakumar. On stochastic optimal control and reinforcement
learning by approximate inference. IJCAI '13, 2013.
[Schulman et al., 2017] John Schulman, Xi Chen, and Pieter
Abbeel. Equivalence Between Policy Gradients and Soft Q-
Learning. arXiv e-prints, page arXiv:1704.06440, April 2017.
[Shapley, 1953] Lloyd S Shapley. Stochastic games. Proceedings
of the national academy of sciences, 39(10):1095 -- 1100, 1953.
[Theodorou et al., 2010] Evangelos Theodorou, Jonas Buchli, and
Stefan Schaal. A generalized path integral control approach to re-
inforcement learning. J. Mach. Learn. Res., 11, December 2010.
[Tian et al., 2018] Zheng Tian, Shihao Zou, Tim Warr, Lisheng
Wu, and Jun Wang. Learning Multi-agent Implicit Communi-
cation Through Actions: A Case Study in Contract Bridge, a
Collaborative Imperfect-Information Game. arXiv e-prints, page
arXiv:1810.04444, October 2018.
[Todorov, 2007] Emanuel Todorov. Linearly-solvable markov de-
cision problems. In B. Scholkopf, J. C. Platt, and T. Hoffman,
editors, NIPS '19. 2007.
[Todorov, 2010] Emanuel Todorov. Policy gradients in linearly-
In J. D. Lafferty, C. K. I. Williams, J. Shawe-
solvable mdps.
Taylor, R. S. Zemel, and A. Culotta, editors, NIPS. 2010.
[Toussaint and Storkey, 2006] Marc Toussaint and Amos Storkey.
Probabilistic inference for solving discrete and continuous state
markov decision processes.
ICML '06, pages 945 -- 952, New
York, NY, USA, 2006. ACM.
[Toussaint, 2009] Marc Toussaint. Robot trajectory optimization
using approximate inference. ICML '09, pages 1049 -- 1056, New
York, NY, USA, 2009. ACM.
[Wei et al., 2018] Ermo Wei, Drew Wicke, David Freelan, and Sean
Luke. Multiagent soft q-learning. AAAI, 2018.
[Wen et al., 2019] Ying Wen, Yaodong Yang, Rui Luo, Jun Wang,
and Wei Pan. Probabilistic recursive reasoning for multi-agent
reinforcement learning. In ICLR, 2019.
[Wilson et al., 2010] Aaron Wilson, Alan Fern, and Prasad Tade-
palli. Bayesian policy search for multi-agent role discovery.
AAAI'10, pages 624 -- 629. AAAI Press, 2010.
[Ziebart et al., 2008] Brian D. Ziebart, Andrew Maas, J. Andrew
Bagnell, and Anind K. Dey. Maximum entropy inverse reinforce-
ment learning. In AAAI '08, 2008.
A Algorithms
Algorithm 1 Multi-agent Soft Q-learning
Result: policy πi, opponent model ρi
Initialization:
Initialize replay buffer M to capacity M.
Initialize Qωi(s, ai, a−i) with random parameters ωi, P (a−is) arbitrarily, set γ as the discount factor.
Initialize target Q¯ωi(s, ai, a−i) with random parameters ¯ωi, set C the target parameters update interval.
while not converge do
Collect experience
For the current state st compute the opponent model ρi(a−i
st) and conditional policy πi(ai
tst, a−i
t ) respectively from:
α
,
ρi(a−i
t
st) ∝ P (a−i
t
st)
exp(
1
α
Qωi(st, ai
t, a−i
t ))
πi(ai
tst, a−i
Qωi (st, ai
t, a−i
t )).
Compute the marginal policy πi(ai
tst) and sample an action from it:
t ∼ πi(ai
ai
tst) =
πi(ai
tst, a−i
t )ρ(a−i
t
st).
t
(cid:88)
ai
t
1
α
t ) ∝ exp(
(cid:88)
a−i
Observe next state st+1, opponent action a−i
t
and reward ri
t, save the new experience in the reply buffer:
Update the prior from the replay buffer:
P (a−i
t
st) =
M ← M ∪ {(st, ai
t, a−i
t
, st+1, ri
t)}.
(cid:80)M
m=1
(cid:80)M
m=1
I(s = st, a−i = a−i
t )
I(s = st)
∀st, a−i
t ∈ M.
Sample a mini-batch from the replay buffer:
{s(n)
t
, ai,(n)
t
−i,(n)
, a
t
, s(n)
t+1, r(n)
t }N
n=1 ∼ M.
t
t
t
, s(n)
−i,(n)
, a
t
Sample {a−i,(n,k)}K
Compute empirical ¯V i(s(n)
Update Qωi(s, ai, a−i):
for each tuple (s(n)
) do
, ai,(n)
t+1, r(n)
k=1 ∼ ρ, {ai,(n,k)}K
k=1 ∼ π.
(cid:16)
K(cid:88)
(cid:40)
1
t+1) = log
t+1) as:
¯V i(s(n)
P 1
Set
k=1
K
α (a−i,(n,k)s(n)
t+1) exp( 1
α Q¯ωi(s(n)
t+1, ai,(n,k), a−i,(n,k)))
π(ai,(n,k)s(n)
t+1, a−i,(n,k))ρ(a−i,(n,k)s(n)
t+1)
(cid:17)α
.
y(n) =
r(n)
t
r(n)
t + γ ¯V i(s(n)
t+1)
for terminal s(n)
t+1
for non-terminal s(n)
t+1
Perform gradient descent step on (y(n) − Qωi(s(n)
Every C gradient descent steps, reset target parameters:
t+1, ai,(n), a−i,(n)))2 with respect to parameters ωi
¯ωi ← ω
end for
end while
Compute converged πi and ρi
Algorithm 2 Multi-agent Variational Actor Critic
Result: policy πθi, opponent model ρφi
Initialization:
Initialize parameters θi, φi, ωi, ψi for each agent i and the random process N for action exploration.
Assign target parameters of joint action Q-function: ¯ωi ← ω.
Initialize learning rates λV , λQ, λπ, λφ, α, and set γ as the discount factor.
for Each episode d = (1, . . . , D) do
Initialize random process N for action exploration.
for each time step t do
For the current state st, sample an action and opponent's action using:
t ← gφ−i(−i; st), where −i
a−i
t ← fθi(i; st, a−i
t ), where i
ai
Observe next state st+1, opponent action a−i
t ∼ N ,
t ∼ N .
t
and reward ri
Di ← Di ∪ {(st, ai
t, save the new experience in the replay buffer:
t, a−i
, st+1, ri
, a−i
t)}.
t
t
Update the prior from the replay buffer:
ψi = arg max EDi[−P (a−is) log Pψi(a−is)]
Sample a mini-batch from the reply buffer:
{s(n)
t
, ai,(n)
t
−i,(n)
, a
t
−i,(n)
, a
t
, s(n)
t+1, r(n)
t }N
n=1 ∼ M.
t+1, sample an action and opponent's action using:
For the state s(n)
−i,(n)
t+1 ← gφ−i (−i; s(n)
a
t+1 ← f¯θi(i; s(n)
ai,(n)
t+1, a
¯V i(s(n)
t+1) = Q¯ω(s(n)
t+1, ai,(n)
Set
t+1), where −i
−i,(n)
), where i
t+1
−i,(n)
t+1 , a
t+1
t+1 ∼ N ,
t+1 ∼ N .
(cid:40)
)− α log πθi(ai,(n)
t+1 s(n)
t+1, a
−i,(n)
t+1
)− log ρφi(a
−i,(n)
t+1
s(n)
t+1) + log Pψi(a
−i,(n)
t+1
s(n)
t+1).
y(n) =
r(n)
t
r(n)
t + γ ¯V i(s(n)
t+1)
for terminal s(n)
t+1
for non-terminal s(n)
t+1
∇ωiJQ(ωi) = ∇ωiQωi(s(n)
t
, ai,(n)
t
−i,(n)
, a
t
)(Qωi(s(n)
t
, ai,(n)
t
−i,(n)
, a
t
) − y(n))
s(n)
∇θiJπ(θi) = ∇θiα log πθi(ai,(n)
−i,(n)
+ (∇ai,(n)
, a
t
α log πθi(ai,(n)
s(n)
−i,(n)
, a
t
) − ∇ai,(n)
)
t
t
t
t
t
t
Qω(s(n)
t
, ai,(n)
t
−i,(n)
, a
t
))∇θfθi(i
t; s(n)
t
−i,(n)
, a
t
)
−i,(n)
∇φiJρ(φi) = ∇φi log ρφi (a
t
−i,(n)
+ (∇a
s(n)
log ρφi(a
t
+ ∇a
α log πθi(ai,(n)s(n)
s(n)
) − ∇a
−i,(n)
, a
t
−i,(n)
t
−i,(n)
t
t
t
t
)
−i,(n)
t
−i,(n)
log P (a
t
))∇φigφi(−i
; s(n)
t
t
)
s(n)
t
) − ∇a
−i,(n)
t
Qωi (s(n)
t
, ai,(n)
t
−i,(n)
, a
t
)
Update parameters:
ωi = ωi − λQ∇ωiJQ(ωi)
θi = θi − λπ∇θiJπ(θi)
φi = φi − λφi∇φiJρ(φi)
end for
Every C gradient descent steps, reset target parameters:
ωi = βωi + (1 − β)ωi
.
end for
B Variational Lower Bounds in Multi-agent Reinforcement Learning
B.1 The Lower Bound of The Log Likelihood of Optimality
We can factorize P (ai
1:T , s1:To−i
1:T ) as :
(cid:89)
1:T , a−i
P (ai
1:T , s1:To−i
1:T , a−i
t ) is the conditional policy of agent i when other agents −i achieve optimality. As agent i has no
P (st+1st, at)P (ai
t )P (a−i
st, o−i
t ),
1:T ) = P (s1)
, st, o−i
ta−i
(23)
t
t
t
ta−i
t
, st, o−i
where P (ai
ta−i
knowledge about rewards of other agents, we set P (ai
(cid:89)
1:T , o−i
1:T ) as:
(cid:89)
Analogously, we factorize q(ai
1:T , s1:Toi
1:T , s1:Toi
1:T ) = P (s1)
1:T , a−i
1:T , o−i
1:T , a−i
q(ai
t
t
= P (s1)
, st, o−i
t ) ∝ 1.
P (st+1st, at)q(ai
P (st+1st, at)π(ai
ta−i
t
, st) is agent 1's conditional policy at optimum and ρ(a−i
t
t
where π(ai
policies.
With the above factorization, we have:
t
t
st, oi
t, o−i
t )
t )q(a−i
st),
t, o−i
, st, oi
t )ρ(a−i
ta−i
tst, a−i
st) is agent 1's model about opponents' optimal
(25)
(24)
t
log P (oi
(cid:88)
1:To−i
1:T )
= log
≥(cid:88)
ai
1:T ,a
−i
1:T ,s1:T
q(ai
1:T , a−i
P (oi
1:T , ai
1:T , a−i
1:T , s1:To−i
1:T )
1:T , o−i
1:T ) log
P (oi
q(ai
1:T , s1:Toi
T(cid:88)
= E
(ai
1:T ,a
1:T ,s1:T ∼q)[
−i
, st, ai
t
t=1
log P (oi
log P (st+1st, ai
to−i
T(cid:88)
XXXXXXXXXXXX
t, a−i
t )
t ) − T(cid:88)
ρ(a−i
tst, a−i
P (a−i
T(cid:88)
log
t=1
t=1
t
t
st)
st, o−i
t )
t, a−i
t ) − log π(ai
T(cid:88)
t=1
+
1:T ,s1:T ∼q)[
−i
t=1
Ri(st, ai
t, a−i
t )∼q[Ri(st, ai
−i
t,a
t ) + H(π(ai
log π(ai
XXXXXX
− log P (s1) −
− T(cid:88)
(cid:88)
= E
1:T ,a
t=1
(ai
E
(st,ai
=
t
1:T , ai
1:T , a−i
t, a−i
1:T , a−i
1:T , s1:Toi
1:T , s1:To−i
1:T )
T(cid:88)
1:T , o−i
1:T )
XXXXXXXXXXXX
t, a−i
t )
log P (st+1st, ai
XXXXX
log P (s1) +
t ) +
t=1
log P (ai
tst, a−i
t
, o−i
t )]
st)
st, o−i
t )
st)P (a−i
t
t
ρ(a−i
P (a−i
t
+ log 1]
st, o−i
t ))].
t ) − log
tst, a−i
t )) − DKL(ρ(a−i
t
tst, a−i
C Multi-Agent Soft-Q Learning
C.1 Soft Q-Function
We define the soft state-action value function Qπ,ρ
t, a−i
t )
Qπ,ρ
sof t(st, ai
= rt + E
(st+l,ai
t+l,a
t+l,...)∼q[
−i
∞(cid:88)
l=1
= E
= E
(st+1,ai
t+1,a
t+1)[rt + γ(αH(π(ai
−i
t+1)[rt + γ(αH(π(·st+1, a−i
−i
(st+1,a
sof t(s, a, a−i) of agent i in a stochastic game as:
γl(rt+l + αH(π(ai
t+la−i
t+l, st+l)) − DKL(ρ(a−i
t+lst+l)P (a−i
t+1)) − DKL(ρ(a−ist+1)P (a−ist+1)) + Qπ,ρ
t+1∼π[Qπ,ρ
t+1)) − DKL(ρ(a−ist+1)P (a−ist+1)) + E
t+1st+1, a−i
ai
t+lst+l))]
sof t(st+1, ai
sof t(st+1, ai
t+1, a−i
t+1, a−i
t+1))]
t+1)])]
= E(st+1)[rt + γ(E
+ E
−i
t+1∼ρ,ai
a
t+1∼π[Qπ,ρ
sof t(st+1, ai
−i
t+1∼ρ,ai
a
t+1∼π[αH(π(ai
t+1, a−i
t+1)])],
t+1st+1, a−i
t+1))] − DKL(ρ(·st+1)P (·st+1))]
(26)
(27)
(28)
(29)
(30)
(31)
(32)
(33)
(34)
(35)
(36)
Then we can easily see that the objective in Eq. 7 can be rewritten as:
J (π, φ) =
E
(st,ai
t,a
t )∼(ps,π,ρ)[Qπ,ρ
−i
sof t(st, ai
t, a−i
t ) + αH(π(ai
tst, a−i
t )) − DKL(ρ(a−i
t
st)P (a−i
t
st))],
(37)
(cid:88)
t
by setting α = 1.
C.2 Policy Improvement and Opponent Model Improvement
Theorem 4. (Policy improvement theorem) Given a conditional policy π and opponent model ρ, define a new conditional policy
π as
Assume that throughout our computation, Q is bounded and(cid:80)
π). Then Qπ,ρ
sof t(s, ai, a−i) ≥ Qπ,ρ
sof t(s, ai, a−i)∀s, a.
π(·s, a−i) ∝ exp(
sof t(s,·, a−i)),∀s, a−i.
Qπ,ρ
1
α
ai Q(s, ai, a−i) is bounded for any s and a−i (for both π and
(38)
(39)
(cid:88)
Theorem 5. (Opponent model improvement theorem) Given a conditional policy π and opponent model ρ, define a new oppo-
nent model ρ as
ρ(·s) ∝ exp(
sof t(s, ai,·)π(ai·, s) + αH(π(s)) + log P (·s)),∀s, ai.
Qπ,ρ
Assume that throughout our computation, Q is bounded and(cid:80)
ai
a−i exp((cid:80)
sof t(s, ai, a−i) ≥ Qπ,ρ
sof t(s, ai, a−i)∀s, a.
and ai (for both ρ and ρ). Then Qπ, ρ
ai Q(s, ai, a−i)π(ais, a−i)) is bounded for any s
The proof of Theorem 4 and 5 is based on two observations that:
αH(π(·s, a−i)) + Eai∼π[Qπ,ρ
sof t(s, ai, a−i)] ≤ αH(π(·s, a−i)) + Eai∼π[Qπ,ρ
sof t(s, ai, a−i)],
(40)
and
E
−i
t+1∼ρ,ai
a
≤ E
t+1∼π[αH(π(ai
t+1st+1, a−i
t+1))] − DKL(ρ(·st+1)P (·st+1))] + E
−i
t+1∼ ρ,ai
t+1∼π[αH(π(ai
a
t+1st+1, a−i
t+1))] − DKL(ρ(a−i
t+1∼π[Qπ,ρ
−i
t+1∼ρ,ai
t+1st+1)P (·st+1)) + E
−i
t+1∼ ρ,ai
a
sof t(st+1, ai
t+1∼π[Qπ,ρ
a
t+1, a−i
sof t(st+1, ai
t+1)]
t+1, a−i
(41)
t+1)].
(42)
First, we notice that
αH(π(·s, a−i)) + Eai∼π[Qπ,ρ
sof t(s, ai, a−i)] = −αDKL(π(·s, a−i)π(·s, a−i)) + α log
(cid:88)
ai
exp(
1
α
sof t(s, ai, a−i)). (43)
Qπ,ρ
Therefore, the LHS is only maximized if the KL-Divergence on the RHS is minimized. This KL-Divergence is minimized only
when π = π, which proves the Equation 40.
Similarly, we can have
Ea−i∼ρ,ai∼π[αH(π(ais, a−i))] − DKL(ρ(·s)P (·s))]) + Ea−i∼ρ,ai∼π[Qπ,ρ
= −DKL(ρ(·s)ρ(·s)) + log
exp(
Qπ,ρ(s, ai, a−i)π(ais, a−i) + αH(π(·s, ai)) + log P (a−is)),
sof t(s, ai, a−i)]
(44)
(cid:88)
a−i
(cid:88)
ai
which proves the Equation 42.
With the above observations, the proof of Theorem 4 and 5 is completed by as follows:
Qπ,ρ
sof t(st, ai
= E
t, a−i
t )
t+1)[rt + γ(αH(π(ai
−i
t+1)) − DKL(ρ(a−i
t+1st+1)P (a−i
t+1st+1, a−i
(st+1,ai
t+1,a
t+1st+1)) + Qπ,ρ
sof t(st+1, ai
t+1, a−i
t+1))]
= E
≤ E
(st+1,a
(st+1,a
−i
t+1)[rt + γ(αH(π(·st+1, a−i
t+1)[rt + γ(αH(π(·st+1, a−i
−i
t+1)) − DKL(ρ(a−i
t+1st+1)P (a−i
t+1st+1)) + E
t+1∼π[Qπ,ρ
sof t(st+1, ai
ai
t+1, a−i
t+1)])]
t+1)) − DKL(ρ(a−i
t+1st+1)P (a−i
t+1st+1)) + E
t+1∼π[Qπ,ρ
sof t(st+1, ai
ai
t+1, a−i
t+1)])]
t+1, a−i
t+1∼π[αH(π(ai
t+1)])]
t+1∼π[αH(π(ai
t+1)])]
t+1st+1, a−i
t+2)) − DKL(ρ(a−i
a
a
−i
t+1∼ ρ,ai
−i
t+1∼ρ,ai
= E(st+1)[rt + γ(E
+ E
≤ E(st+1)[rt + γ(E
+ E
= E
+ γ2E
≤ E
+ γ2E
...
(st+1,ai
(st+1,ai
(st+2,a
(st+2,a
t+1,a
t+1,a
−i
−i
t+1∼ρ,ai
a
t+1∼π[Qπ,ρ
sof t(st+1, ai
−i
t+1∼ ρ,ai
a
t+1∼π[Qπ,ρ
sof t(st+1, ai
−i
t+1, a−i
t+1)∼q[rt + γ(αH(π(ai
−i
t+2)[αH(π(·st+2, a−i
t+1)[rt + γ(αH(π(ai
−i
t+2)[αH(π(·st+2, a−i
∞(cid:88)
γl(rt+l + αH(π(ai
≤ rt + E
(st+l,ai
t+l,a
t+l,...)∼q[
−i
l=1
= Qπ, ρ
sof t(st, ai
t, a−i
t ).
t+1st+1, a−i
t+1))] − DKL(ρ(·st+1)P (·st+1))
t+1st+1, a−i
t+1))] − DKL(ρ(·st+1)P (·st+1))
t+1)) − DKL(ρ(a−ist+1)P (a−ist+1)) + rt+1)
t+2st+2)P (a−i
t+2st+2)) + E
t+2∼π[Qπ,ρ
sof t(st+2, ai
ai
t+2, a−i
t+2)]]]
t+1st+1, a−i
t+1)) − DKL(ρ(a−ist+1)P (a−ist+1)) + rt+1)
t+2)) − DKL(ρ(a−i
t+2st+2)P (a−i
t+2st+2)) + E
t+2∼π[Qπ,ρ
sof t(st+2, ai
ai
t+2, a−i
t+2)]]]
t+la−i
t+l, st+l)) − DKL(ρ(a−i
t+lst+l)P (a−i
t+lst+l))]
(45)
(46)
(47)
(48)
(49)
(50)
(51)
(52)
(53)
(54)
(55)
(56)
(57)
With Theorem 4 and 5 and the above inequalities, we can see that, if we start from an arbitrary conditional policy π0 and an
arbitrary opponent model ρ0 and we iterate between policy improvement as
πi+1(·s, a−i) ∝ exp(
sof t (s,·, a−i)),
Qπt,ρt
1
α
and opponent model improvement as
ρt+1(·s) ∝ exp(
Qπt+1,ρt
sof t
(s, ai,·)πt+1(ai·, s) + αH(πt+1(s)) + log P (·s)),
(cid:88)
ai
sof t (s, ai, a−i) can be shown to increase monotonically. Similar to [Haarnoja et al., 2017], we can show that with
then Qπt,ρt
certain regularity conditions satisfied, any non optimal policy and opponent model can be improved this way and Theorem 1 is
proved.
C.3 Soft Bellman Equation
As we show in Appendix C.2, when the training converges, we have:
π∗(ais, a−i) =
1
(cid:80)
α exp(Q∗(s, ai, a−i))
α Q∗(s, ai, a−i))
ai exp( 1
,
and
ρ∗(a−is) =
=
exp((cid:80)
(cid:80)
a−i exp((cid:80)
(cid:16)(cid:80)
P (a−is)
ai Q∗(s, ai, a−i)π∗(ais, a−i) + αH(π∗(ais, a−i)) + log P (a−is))
ai Q∗(s, ai, a−i)π∗(ais, a−i) + αH(π∗(ais, a−i)) + log P (a−is))
ai exp(Q∗
exp(V ∗(s))
sof t(s, ai, a−i))
(cid:17)α
,
where the equality in Eq. 57 comes from substituting π∗ with Eq. 56 and we define the soft sate value function V π,ρ
agent i as:
sof t(s) of
(cid:88)
−i
t
a
(cid:88)
ai
t
(cid:18) 1
α
(cid:19)α
V π,ρ
sof t(st) = log
P (a−i
t
st)
exp
Qπ,ρ
sof t(st, ai
t, a−i
t )
.
(58)
Then we can show that
Qπ∗,ρ∗
sof t (s, ai, a−i)
= rt + γEs(cid:48)∼ps [(E
−i
t+1∼ρ,ai
+ E
t+1∼π[Qπ,ρ
sof t(st+1, ai
= rt + γEs(cid:48)∼ps[V ∗(s(cid:48))].
−i
t+1∼ρ,ai
a
a
t+1∼π[αH(π(ai
t+1)])]
t+1, a−i
(cid:34)
log
t+1st+1, a−i
t+1))] − DKL(ρ(·st+1)P (·st+1))]
(59)
1
2
.
exp
ai(cid:48)
log
a−i(cid:48)
2019],
maxs,ai,a−i
P (a−i(cid:48)s(cid:48))
P (a−i(cid:48)s(cid:48))
(cid:18) 1
α
(cid:88)
a−i(cid:48)
Q(s(cid:48), ai(cid:48), a−i(cid:48))
(60)
(cid:13)(cid:13) ∆=
T Q(s, ai, a−i) = R(s, ai, a−i) + γEs(cid:48)∼ps
We define the soft value iteration operator T as:
In a symmetric fully cooperative game with only one global optimum, we can show as done in [Wen et al.,
(cid:0)s, ai, a−i(cid:1) − Qi
(cid:32)(cid:88)
(cid:18) 1
(cid:0)s, ai, a−i(cid:1)(cid:12)(cid:12). Let ε =(cid:13)(cid:13)Qi
(cid:19)(cid:33)α
(cid:32)(cid:88)
(cid:13)(cid:13), then we have:
(cid:32)(cid:88)
(cid:32)(cid:88)
(cid:19)(cid:33)α(cid:35)
the operator defined above is a contraction mapping. We define a norm on Q-values (cid:13)(cid:13)Qi
(cid:12)(cid:12)Qi
(cid:19)(cid:33)α
(cid:88)
(cid:33)α
(cid:19)(cid:33)α
(cid:19)(cid:33)α
α Q2(s(cid:48), ai(cid:48), a−i(cid:48))(cid:1)(cid:1)α
(cid:18) 1
(cid:18) 1
(cid:32)(cid:88)
(cid:18) 1
(cid:32)(cid:88)
(cid:18) 1
a−i(cid:48) P (a−i(cid:48)s(cid:48))(cid:0)(cid:80)
1 − Qi
(cid:88)
(cid:88)
(cid:88)
a−i(cid:48) P (a−i(cid:48)s(cid:48))(cid:0)(cid:80)
ai(cid:48) exp(cid:0) 1
(cid:13)(cid:13) ≤ γε = γ(cid:13)(cid:13)Qi
1 − T Qi
= αε + log
α Q1(s(cid:48), ai(cid:48), a−i(cid:48))(cid:1)(cid:1)α ≥ −αε+log(cid:80)
(cid:13)(cid:13), where α = 1.
1 − Qi
ai(cid:48) exp(cid:0) 1
Q2(s(cid:48), ai(cid:48), a−i(cid:48)) + ε
P (a−i(cid:48)s(cid:48)) exp(ε)α
Q2(s(cid:48), ai(cid:48), a−i(cid:48))
Q1(s(cid:48), ai(cid:48), a−i(cid:48))
Q2(s(cid:48), ai(cid:48), a−i(cid:48))
Q2(s(cid:48), ai(cid:48), a−i(cid:48))
(cid:88)
a−i(cid:48)
P (a−i(cid:48)s(cid:48))
P (a−i(cid:48)s(cid:48))
Similarly, log(cid:80)
Therefore(cid:13)(cid:13)T Qi
≤ log
P (a−i(cid:48)s(cid:48))
a−i(cid:48)
a−i(cid:48)
a−i(cid:48)
exp
α
ai(cid:48)
exp
α
ai(cid:48)
exp
ai(cid:48)
exp
exp
1 − Qi
2
ai(cid:48)
ai(cid:48)
= log
= log
2
2
2
α
. (61)
α
α
(cid:19)
exp(ε)
.
|
1702.02541 | 1 | 1702 | 2017-02-08T17:49:31 | Modelling community formation driven by the status of individual in a society | [
"cs.MA",
"cs.SI",
"physics.soc-ph"
] | In human societies, people's willingness to compete and strive for better social status as well as being envious of those perceived in some way superior lead to social structures that are intrinsically hierarchical. Here we propose an agent-based, network model to mimic the ranking behaviour of individuals and its possible repercussions in human society. The main ingredient of the model is the assumption that the relevant feature of social interactions is each individual's keenness to maximise his or her status relative to others. The social networks produced by the model are homophilous and assortative, as frequently observed in human communities and most of the network properties seem quite independent of its size. However, it is seen that for small number of agents the resulting network consists of disjoint weakly connected communities while being highly assortative and homophilic. On the other hand larger networks turn out to be more cohesive with larger communities but less homophilic. We find that the reason for these changes is that larger network size allows agents to use new strategies for maximizing their social status allowing for more diverse links between them. | cs.MA | cs |
Modelling community formation driven by the status of individual in a society
Jan E. Snellman1, Gerardo Iniguez2,1,3, Tzipe Govezensky4, R. A. Barrio5, and Kimmo K. Kaski1
1Department of Computer Science, Aalto University School of Science, FI-00076 AALTO, Finland,
2Instituto de Investigaciones en Matem´aticas Aplicadas y en Sistemas,
Universidad Nacional Aut´onoma de M´exico, 01000 M´exico D.F., Mexico
3Centro de Investigaci´on y Docencia Econ´omicas,
Consejo Nacional de Ciencia y Tecnolog´ıa, 01210 M´exico D.F., Mexico
4Instituto de Investigaciones Biom´edicas, Universidad Nacional Aut´onoma de M´exico, 04510 M´exico D.F., Mexico and
5Instituto de F´ısica, Universidad Nacional Aut´onoma de M´exico, 01000 M´exico D.F., Mexico
(Dated: November 14, 2018)
In human societies, people's willingness to compete and strive for better social status as well as
being envious of those perceived in some way superior lead to social structures that are intrinsically
hierarchical. Here we propose an agent-based, network model to mimic the ranking behaviour of
individuals and its possible repercussions in human society. The main ingredient of the model is the
assumption that the relevant feature of social interactions is each individual's keenness to maximise
his or her status relative to others. The social networks produced by the model are homophilous and
assortative, as frequently observed in human communities and most of the network properties seem
quite independent of its size. However, it is seen that for small number of agents the resulting network
consists of disjoint weakly connected communities while being highly assortative and homophilic.
On the other hand larger networks turn out to be more cohesive with larger communities but less
homophilic. We find that the reason for these changes is that larger network size allows agents to
use new strategies for maximizing their social status allowing for more diverse links between them.
community formation, opinion formation, social hierarchy
PACS numbers:
I.
INTRODUCTION
One of the most pervasive tendencies of humans is
putting things in ranking order. In human societies these
tendencies are reflected in their social interactions and
networks being hierarchical in many respects. Hierar-
chies and ranks emerge due to individuals' subjective per-
ceptions that some other individuals are in some respect
better. Then a relevant research question is whether or
not the formation and structure of hierarchies in human
societies can be understood by making the assumption
that the dominant driving force of people in social in-
teractions is to enhance their own "value" or "status"
relative to others. We call this assumption "better than-
hypothesis" (BTH) and note that it is closely related
to the thinking of the school of individual psychology
founded by Adler in the early 1900s [1], which, while
starting with the assumption that human individuals uni-
versally strive for "superiority" over others, emphasizes
inferiority avoidance as a motive for many human actions.
Further studies of this kind of individuals' status-
seeking behaviour, especially concerning consumer be-
haviour and economics, include the canonical references
by Veblen [2], Duesenberry [3] and Packard [4] (See also
Refs [5 -- 8]). In addition there is a closely related socio-
logical model called Social Dominance Theory [9], which
proposes that the construction and preservation of so-
cial hierarchies is one of the main motivations of humans
in their social interactions and networks. However, the
most relevant observational facts concerning BTH come
from the field of experimental economics, especially from
the results of experiments on the so-called "ultimatum
game" [10], where the human players have been shown
to reject too unequal distributions of money. The con-
cept of inequity aversion, that is the observed social phe-
nomenon of humans preferring equal treatment in their
societies, is often invoked to explain these observations.
Recently some models featuring inequity aversion have
been proposed in Refs. [11, 12].
All of these models, although from different fields of
study, have something to do with the relative standings
between different human individuals and groups, and so
they could all be considered to emerge from or be based
on a single principle such as BTH. It is this generality
which makes BTH an intriguing and interesting object
of study. There are even some studies on economic data,
such as [13], that suggest a link between relative social
standings and human well-being, and considerations of
social status have measurable effects on brain functions,
as shown in e.g. [14, 15]. These studies imply that BTH
could well be something fundamental to human nature.
The competition for a better hierarchical position
among humans can be intense and sometimes even vio-
lent. However, humans have other characteristics includ-
ing egalitarianism as well as striving for fairness. These
traits could be interpreted in the context of BTH by re-
marking that people need to live in societies and make
diverse social bonds, which in turn would contribute to
their social status. This means that the members of soci-
ety when they make decisions, need to take the feelings of
others into account. Hence the behavioral patterns of in-
dividuals in social networks should then be characterised
by sensitivity to the status of the other individuals in
the network. This sensitivity manifests itself as inequity
aversion and treating others fairly. To find out what in
this context are the plausible and relevant mechanisms of
human sociality driving societal level community forma-
tion we will focus on improving the BTH-based approach
by using the frame of agent-based models and studying
the emergence of social norms in such social systems, fol-
lowing the tradition presented in Refs. [16 -- 21].
In this study we use an agent-based network model
applying BTH-based approach to simulate social inter-
actions dependent on societal values and rank, to get
insight to their global effects on the structure of soci-
ety. We find that in such a model society with a given
constant ranking system the social network forms a de-
gree hierarchy on top of the ranking system under BTH,
such that the agents' degrees tend to increase, the further
away their rank is from the average. The structure of the
paper is as follows. In Section II we motivate the basics
of BTH using the simple and well-researched ultimatum
game as an example, and in Section III we show how the
findings from this can be utilised as a part of agent-based
models. In Section IV we present the numerical results
of the simulations from the model, and in Section V we
analyse them. The two final Sections discuss the possible
interpretations of the results and present the conclusions.
II. BETTER THAN-HYPOTHESIS AND
ULTIMATUM GAME
In this section we describe the theoretical basis for our
model. We start by analysing the ultimatum game first
proposed in [10], as it allows us to derive a basic form for
the social gain function in our model. The ultimatum
game is a game with two players, where one player has
the task to make a proposal to the other player about
how a given sum of money should be divided between
them. The second player then gets to choose if the pro-
posal is acceptable or not; if it is, the money is divided
as proposed. If not, neither player gets anything. Exper-
iments show that humans playing this game normally do
not accept deals that are perceived to be unfair, i.e.
in
situations in which the proposer gets too large a share
of the money (see, e.g. Refs.
[22 -- 27]). This is a classic
problem in the mainstream economics, where humans are
assumed to be rational and, therefore, accept something
rather than nothing.
We implement BTH in the ultimatum game by inter-
preting the money used in a deal as a way of compar-
ing the status between the one who accepts the proposal
(called from now on the accepter) and its proposer. We
denote the change of "status" of the accepter as ∆a,
which takes into account it's own monetary gain, and the
gain in relation to the proposer. Therefore, the simplest
expression for ∆a is,
∆a =Ra(t1) − Ra(t0) + [Ra(t1) − Rp(t1)]
− [Ra(t0) − Rp(t0)],
(1)
where Ra(t) and Rp(t) stand for the monetary reserves
2
(in the context of the game) of the accepter and proposer,
respectively, at time t, with t0 being the time before the
deal and t1 the time after the deal. In terms of economic
theory, ∆a would be called the accepter's change of util-
ity, which is ordinarily assumed to consist of the term
Ra(t1) − Ra(t0) or the absolute payoff of the accepter.
The additional terms [Ra(t1)−Rp(t1)]−[Ra(t0)−Rp(t0)]
that stem from the BTH measuring the change in rela-
tive standings of the accepter and proposer. The actual
BTH utility function for the accepter in the ultimatum
game takes the form
U BT H
a
= Ra + [Ra − Rp].
(2)
According to Eq. (1), the accepter will refuse the deal
for ∆a < 0, and will accept it for ∆a > 0, with ∆a = 0
being the borderline case. Should the deal be rejected,
Ra(t1) = Ra(t0) and Rp(t1) = Rp(t0), and, consequently,
∆a = 0.
If we denote by Rmax the total amount of
money to be shared and by Rshare the actual amount
money that the proposer has reserved for the accepter,
then in the case where the transaction does take place
we have Ra(t1) = Ra(t0) + Rshare and Rp(t1) = Rp(t0) +
(Rmax − Rshare).
If we further assume that Ra(t0) =
Rp(t0) (i.e.
the players start on equal footing, which
may very well be the case in the context of the game at
least), it follows from Eq.(1) that the smallest offer that
the accepter expects from the proposer is one third of the
maximum Rmax, i.e. the condition
Rshare >
Rmax
3
(3)
must hold for the proposal to be acceptable.
Previous literature shows that the minimum offers that
people are usually willing to accept are around 30% of
a given quantity [24], in close agreement with the cal-
culation above. Moreover, we note that if the term
Ra(t1)− Ra(t0) in Eq. (1) is neglected, then the accepter
will never settle for less than half of the total amount.
Next we use Eq. (1) to illustrate how the BTH-based
approach can be implemented in the context of agent
based social simulations. The trick is to generalise this
equation to the cases of many players, and with multiple
and different kinds of items being exchanged.
If there
are N players (denoted by i = 1, . . . , N ), the change of
status ∆i of the individual i may be written as follows
[Ri(t0) − Rj(t0)].
(4)
In the case that players have several ways to measure
their status, we may introduce normalisation factors to
compare the relative value of the exchanged items, and
∆i = Ri(t1) − Ri(t0)
+
[Ri(t1) − Rj(t1)]
(cid:88)
−(cid:88)
j(cid:54)=i
j(cid:54)=i
write the change of status as,
∆i =
i
(cid:26)
(cid:2)Rα
(cid:2)Rα
α
(cid:88)
(cid:88)
−(cid:88)
j(cid:54)=i
+
j(cid:54)=i
1
N α
i (t1) − Rα
Rα
i (t0)
j (t1)(cid:3)
j (t0)(cid:3)(cid:27)
i (t1) − Rα
i (t0) − Rα
(5)
where the index α runs over the various items determin-
ing the status, and N α
is the normalisation factor of each
item, which may vary from one player to another. The
utility function associated with Eq. (5) reads then
i
U BT H
i
=
1
N α
i
Rα
i +
i − Rα
j
.
(6)
(cid:26)
(cid:88)
α
(cid:2)Rα
(cid:88)
j(cid:54)=i
(cid:3)(cid:27)
III. THE BTH AND AGENT-BASED
NETWORK MODEL
In this section we present an agent-based model of a
ranked social system of N agents, in which the agents
exchange their views of the ranking system itself. To
each agent i we assign a parameter ai to describe the
rank of the agent, and a state variable xi to denote the
opinion of the agent i of the social value attached to pa-
rameter a. The social value is then a relative quantity
in the minds of the agents, and they value each other
in either ascending or descending order according to the
"ranking parameter", we call a, and xi determines which
order a given agent i prefers and how strongly. Generally
speaking, the sign of xi represents the chosen order, −
for descending and + for ascending order, while its mag-
nitude represents the strength of conviction: with xi= 0
the agent can be said to support equality of all the agents
irrespective of the ranking parameter, with xi= 1 the
agent thinks that a should directly define the hierarchy
of the society, and for the case xi< 1 or xi> 1 cor-
respondinly downplaying or emphasizing the significance
of a, respectively. Here, we adopt the maximum value
for xi to be 1000.
To put the relation of the social value and the opinion
parameter into more precise terms, we adopt the follow-
ing expression for the term we call "ranking pressure":
Pj =
1
max(a) − min(a)
(aj − ak),
(7)
where the summation is over the whole network. Now, we
define the social value of the agent i in the eyes of agent j
as xjPi, and assume that the agents take into account the
views of their neighbours in addition to their own when
evaluating their total social value Vi. Thus, we write
(cid:88)
k
(cid:88)
k∈m1(i)
Vi = xiPi +
xkPi,
(8)
3
(cid:88)
where m1(i) denotes agents that are one step away from
agent i. The first term on the right could be considered
as the agent's "self esteem" and the second the "social
value" given to it by its first neighbours. It should be
noted at this point that in defining Vi in terms of Pi we
have assumed that ai does not confer direct advantages
or disadvantages for the agents. Therefore, the most nat-
ural interpretation for ai is that it represents the owner-
ship of pure status symbols or Veblen goods, i.e. goods
that are only, or mostly, desirable due to their status-
enhancing properties, such as luxury items. If ai would
give some advantages or disadvantages for the agents, Eq.
(8) would have to be revised accordingly. The agents in
our simulations attempt to gain as much social value as
possible, both in absolute and relative terms, and do this
either by changing their opinion variables or adjusting
their relations to other agents. The system we use here
is purely reactive, with agents reacting to the changes in
their social environment in accordance with BTH. The
social gain function could then be written as
∆i(t1, t0) = Vi(t1) − Vi(t0) +
(Vi(t1) − Vk(t1))
−(cid:88)
k
(Vi(t0) − Vk(t0)),
(9)
k
where t0 and t1 are the two consecutive time steps.
The sign of this function determines the direction of the
changes in xi.
The decision making method employed by the agents
is thus a simple hill climbing algorithm: At a given time
step t, first an agent observes the quantity ∆i(t−1, t−2),
and then changes its variable xi(t − 1), which leads to a
recurrence relation of the form
(cid:26) xi(t − 1) + dx,
xi(t − 1) − dx,
xi(t) =
if Gi > 0,
otherwise
(10)
where Gi = sign[(∆i(t− 1, t− 2)× (xi(t− 2)− xi(t− 1))]
and dx is a small increment. In the spirit of simulated an-
nealing techniques, the magnitude of the change is larger
at the beginning of the dynamics and falls linearly with
time to a minimum value, the maximum and minimum
values being dx = 0.11 and dx = 0.01, respectively, and
the time period to reach the minimum is 1000 time steps.
In general, the links between the agents in the social
network may change in time for which purpose we use
the following rewiring scheme of Ref. [28]. The social
network of the agents is initially random, but will change
periodically, i.e. at every g time steps of the dynamical
Eq. (10). Given the definition of the total social value of
an agent in Eq. (8), the gain function Eq.(9) can be used
to calculate the loss or gain in total social status when
forming or breaking new social bonds. In this study, we
take any positive gain as sufficient to justify the rear-
rangement of social relations between the agents. When
agent i considers cutting an existing bond with agent j,
the gain function has the form
this study, homophily refers to agents with similar rank-
ing parameters forming connections with each other.
4
i,j = Vj − Vi − (ki + 2)xjPi,
∆c
(11)
where ki is the current number of neighbours of agent
i. Similarly, when agent i considers forming a new bond
with agent j, the social gain function reads
i,j = Vi − Vj + (ki + 2)xjPi − xiPj.
∆f
(12)
Since any positive change indicated by the functions
above leads to rewiring, it is the sign of these functions
that determines whether or not links between agents are
broken or created. For instance, if ∆c
i,j > 0, the link be-
tween agents i and j will be cut, and preserved if ∆c
i,j < 0.
In the same vein, a link between agents i and j will be cre-
ated if ∆f
i,j > 0, and not created otherwise. It should be
noted that when forming links the opinions of both agents
are taken into account: The relation formation only suc-
ceeds if ∆f
j,i are both positive. The agents will
form all the relationships they can in a rewiring cycle.
i,j and ∆f
The averaged numerical results extracted from the sim-
ulations consist then of the standard network properties,
i.e. degree (cid:104)k(cid:105) the shortest path (cid:104)L(cid:105), the average clus-
tering coefficient (cid:104)C(cid:105), the mean number of second neigh-
bours (cid:104)n(2)(cid:105), susceptibility (cid:104)s(cid:105) and average assortativity
coefficient (cid:104)ra(cid:105), and a homophily coefficient (cid:104)rh(cid:105). Sus-
ceptibility here refers to average cluster size, which is
calculated as the second moment of the number of s sized
clusters, ns:
(cid:80)
s nss2(cid:80)
s nss
(cid:104)s(cid:105) =
.
(13)
As customary in percolation theory, the largest connected
component of the network is not counted in calculating
s. For the assortativity coefficient we use the definition
given in [29], and the homophily coefficient is defined
using Pearson's product moment coefficient, which mea-
sures the goodness of a linear fit to a given data. For a
sample it can be defined as
(cid:80)M
(cid:113)(cid:80)M
i (vi − v)(wi − w)
i (vi − v)2
(cid:113)(cid:80)M
i (wi − w)2
IV. NUMERICAL RESULTS
rh =
,
(14)
The numerical simulations of the models of social sys-
tem described in the previous sections are performed as
follows. First, the initial state of the system is set at
random with the agents given a relatively small initial
opinion xi between −1 and 1, a ranking parameter ai
between 0 and maximum value of 100.0 and initial con-
nections to other agents, with initial average degree of
5. The opinion and ranking parameters are chosen using
a random number generator, which returns a flat dis-
tribution. The dynamics are then run for 200000 time
steps, which is, according to our test runs, sufficient for
the general structure of the network to settle. However,
the dynamics of the opinion variables do not have a set
stopping point, so they may experience fluctuations even
when such fluctuations do not have an effect on the net-
work structure anymore.
To obtain reliable statistics, the same simulations are
repeated 100 times with random initial values, and av-
erages are calculated from these repeated tests for the
quantities under study. The rewiring timescale g is fixed
to 100 in our simulations, since this value lies in the range
where communities are formed in the opinion formation
model of Ref. [28]. The main parameter whose effect is
studied here is the number of simulated agents, N .
The main objective of this research is to study the
structure of the social networks created under BTH as-
sumption in the case of a rigid ranking system, which we
perform using the model explained in Section III. The
most interesting properties of the system are then asso-
ciated with assortativity, or the tendency of agents with
high degrees connecting to other highly connected agents,
and homophily, or the inclination of similar agents form-
ing connections between each other.
In the context of
where v and w are vectors containing the value param-
eters of agents linked by link i, v and w are the mean
values of these vectors, respectively, and M is the total
number of links. More specifically, if agents α and β are
connected by link i, then vi = aα and wi = aβ. The links
are indexed as follows: the links involving the first agent
are given the first indices, then follow the links involving
the second agent but not the first, and so on, without re-
peating links that have already been indexed. It should
be noted, however, that rh only measures linear correla-
tion between the ranking parameters of linked agents, it
does not indicate how steep these trends are. To check
whether the system is truly homophilic, then, one needs
to make a linear fit to the data: the closer the obtained
linear coefficients are to 1, the greater the homophily.
The average network properties of the system, with
graphs illustrating the behaviour of the system are shown
in Fig. 1 as functions of the population size, which is
varied between 50 and 500. The main observations that
can be made about the graphs in Fig.1(a) are that at
lower population levels they show a tendency of breaking
apart into many subcomponents of different sizes, while
for larger population sizes they tend to consist of a sin-
gle large component and possibly some smaller separate
clusters. A noteworthy fact about these clusters is that
they consist of agents with similar values of the rank-
ing pressure Pi, which means that the network exhibits
homophily in this case. The largest clusters are found
at extreme values, and they become smaller when one
approaches 0, which also corresponds to average ranking
parameters.
In the high population case the picture becomes more
complicated due to the emergence of clusters that con-
5
signs then the "self esteem" part of Vi is negative, which
in itself does not mean that the agent cannot develop a
contrarian opinion, since the second sum in the equation
could be positive because it depends on the opinions of
the neighbours j. This allows the contrarian to be able
to make connections with agents of the same or opposite
ranking pressure. Additionally, by looking at all connec-
tions among contrarians we find that they mostly have
Pi of the same sign as xj, as illustrated in the example
of Fig. 2.
FIG. 1: (a) Graphs showing examples of the network structure
for three different sizes, N , where the colour represents the
value of the agent's normalized ranking pressure, Pi/Pmax.
Observe communities of tightly bound agents of the same
colour. The circular vertices represent agents whose opin-
ion variable xi and ranking pressure Pi have the same sign,
while the triangular vertices represent agents having oppo-
(b) Network properties, such as degree (cid:104)k(cid:105), the
site signs.
shortest path (cid:104)L(cid:105), the average clustering coefficient (cid:104)C(cid:105), the
mean number of second neighbours (cid:104)n(2)(cid:105), susceptibility (cid:104)s(cid:105),
assortativity coefficient (cid:104)ra(cid:105), and homophily coefficient (cid:104)rh(cid:105)
as averages over 100 realizations for the model as a function
of the population size, ranging from 50 to 500.
tain agents with opposing opinions as well. These new
clusters tend to be less connected than the previously de-
scribed homogeneous ones, and they tend to connect to
the large subgraphs, thus forming a single giant graph. A
closer look reveals that these agents generally have opin-
ion variables and ranking pressures with opposite signs
and are depicted as triangles in Fig. 1 and named 'con-
trarians" from now on.
A naive analysis would indicate that the agents with
positive ranking pressures should always support the as-
cending hierarchy, and the agents with negative rank-
ing pressures should always support the descending hier-
archy. However, the contrarian agents exhibit opposite
preferences. The reason why this behaviour is status-wise
profitable can be found by looking into the connections of
the contrarian agents, details of which are shown in Figs.
2 and 3. As it turns out, most of the connections they
form are to other similar agents but with opposite "po-
larity" to theirs, i.e. the contrarian agent with negative
ranking pressure forms connections mostly with contrar-
ian agents of positive ranking pressure, and vice versa.
An important quantity in the model is the social value
Vi, which is a product of the opinion and ranking pres-
sure, as it is seen in Eq. 8. If xi and Pi have opposite
FIG. 2: Scatterplot for ranking pressure Pi of agent i vs. the
opinion variable xj of agent j, taken from 10 realisations for
networks with N = 250 agents. Observe that in most cases
xj and Pi have the same sign (especially for normal-normal
and contrarian-contrarian links), while only in case of normal
- contrarian links they can have opposite signs.
This situation can be status-wise beneficial to all par-
ties involved, since the small penalty to an agent's self-
esteem is more than compensated by the respect that the
agent will gain in this case from other agents. The fact
that the agents could find this strategy using as primi-
tive an intelligence setup as hill climbing is astounding.
Another interesting thing about the contrarians is that
they appear mostly as connections between clusters that
are defined as communities of "normal" agents.
The various kinds of behaviour exhibited by the social
networks have a marked effect on the network properties
also shown in Fig. 1(b). The most obvious is the gradu-
ally rising normalised maximum cluster size ((cid:104)cmax(cid:105)/N ),
which is about 40% of total population size for N = 50,
and over 95% for N = 500. From the figure it seems
that the maximum cluster size reaches 50% of the popu-
lation size for approximately N = 180, after which point
we may assume that the contrarian behavioural patterns
start to become progressively more pronounced.
The susceptibility ((cid:104)s(cid:105)) at first rises pretty much lin-
early, which is not too surprising because of the tendency
of network to break into smaller subgraphs at low pop-
ulation sizes. However, once the population size reaches
about 200, the susceptibility starts to decay, most likely
due to the main component of the network becoming
aN=50N=250normal agentcontrariancontrarianN=5001.00.50.00.51.0Pi/Pmax0.00.20.40.60.81.0propertyb›cmaxfi/N›csfi/N›Cfi›rafi›rhfi›kfi/N›n(2)fi/N50100150200250300350400450500network size N01020304050property›sfi›Lfi›ncfi 150 100 5050100150ranking pressure Pi 1000 5005001000opinion xjN=250normal normalcontrarian contrariannormal contrarian more prominent, with a decay pattern that is almost
piecewise linear itself, apart from fairly large fluctuations.
While fairly high throughout, the homophily ((cid:104)rh(cid:105))
and clustering coefficients ((cid:104)C(cid:105)) gradually fall as func-
tions of population size, almost certainly due to the pro-
liferation of contrarians. There are no great changes in
other network properties, as in the average assortativity
coefficient ((cid:104)ra(cid:105)), although some faint systematic tenden-
cies can be discerned, a slight rising of the average path
length ((cid:104)L(cid:105)), as well as slightly decreasing average num-
ber of clusters ((cid:104)nc(cid:105)). The rest of the properties per agent
are nearly constant, a slight rise of the average number
number of second neighbours ((cid:104)n(2)(cid:105)/N ), and a just per-
ceptible decrease of the average cluster size ((cid:104)cs(cid:105)/N ) and
average degree ((cid:104)k(cid:105)/N ).
A way to illustrate the homophily of the system is to
make a scatterplot of the ranking parameters of linked
agents. As it is seen in Fig.3, the correlation turns out to
be very homophilous, as the ranking parameters of linked
agents correspond very closely to one other. The emer-
gence of the contrarians is also clearly seen: for N = 50,
the percentage of contrarians in 10 realisations is 5.4%,
while for N = 250 is 13.8%, for N = 500 is 24.9% and for
one realisation in in a network of 1000 agents is 38.5%.
There is also a clear decreasing linear trend due to the
contrarians.
The rising trend is without doubt caused by the normal
agents, who tend to associate with agents of similar rank,
and the decreasing trend is likewise due to the contrari-
ans. Both trends have a similar tendency to form square-
like patterns along the diagonals, with each "square" cor-
responding to some of the many visible communities of
the graphs.
6
As explained above, it is necessary to check whether
the correspondence of the rankings is truly homophilic.
In Table I we show the value of the homophily coefficient
of normal and contrarian agents for networks of various
sizes. The data were taken from 10 different numerical
realisations in each case. Observe that normal agents
have values very near one, and contrarians are around
one half. Also in the table we show the slope of the re-
gression of ai vs. aj. Normal agents are very close to one,
indicating high degree of homophily, and the contrarians
are negative and around 0.6, indicating that they mostly
form connection with agents that have opposite signs of
the ranking pressure, and are much less homophilic.
Normals
Contrarians
N=50 N=250 N=500 N=50 N=250 N=500
-0.606 -0.607 -0.707
-0.588 -0.636 -0.731
Slope 0.955 0.968
rh
0.959 0.963
0.963
0.959
TABLE I: Homophily measurement from 10 numerical reali-
sations of networks with different sizes.
In Fig. 4 we show the degree and the final state of the
opinion variable as functions of the ranking parameter of
the agents. One can see that the agents with relatively
large or low values for the ranking parameter seem to
have more neighbours than the agents with the ranking
parameters close to the average rank. The communities
are also visible in this figure in the form of plateaus at
progressively more extreme values of the ranking param-
eter. The appearance of contrarians is also very clearly
visibly in this figure.
FIG. 3: A scatterplot of the ranking parameters ai vs. aj of
linked agents i and j for one realisation of networks of differ-
ent sizes. The colour code: dark blue for connections between
normal agents with positive ranking pressure and opinion vari-
ables, light blue for connections between normal agents with
negative ranking pressure and opinion variables, dark green
for connections between contrarian agents of different "polar-
ity", light green for connections between contrarian agents of
same "polarity" (these are almost non-existent), light red for
connections between normal and contrarian agents with differ-
ent signs of the opinion variable, and dark red for connections
between normal and contrarian agents with same signs of the
opinion variable.
FIG. 4: The degree ki and the final state of the opinion
variable xi/xmax as a function of the ranking parameter ai
of agent i . The colour code is as follows:
light blue for
normal agents with positive ranking pressure, light red for
normal agents with negative ranking pressure, dark blue for
contrarian agents with positive ranking pressure, and dark
red for contrarian agents with negative ranking pressure.
The lower row of panels in Fig.4 shows the correla-
tion of agents' ranking parameters and opinion variables.
As expected, most of the agents below the average rank-
ing parameters have negative opinion variables, and sim-
050100020406080100ranking parameter ajN=50050100ranking parameter aiN=250normal agents, P,x>0normal agents, P,x<0contrarians, different polaritycontrarians, same polaritynormal/contrarian, sgnxi=sgnxjnormal/contrarian, sgnxisgnxj050100N=500051015degree kiN=50normal agent, Pi>0normal agent, Pi<0contrarian, Pi>0contrarian, Pi<00204060N=250050100N=5000501001.00.50.00.51.0opinion xi/xmax050100ranking parameter ai050100 ilarly, the agents with the above average ranking param-
eters have positive opinion variables, at least for small
N . There are some exceptions to this rule. The contrar-
ian agents are seen clearly in the figure as agents whose
opinion variables have opposing signs to agents of similar
ranking. On the other hand, when they are present, there
are often such normal agents that have more egalitarian
views, i.e. less extreme opinions.
There is a clear tendency for the agents to attain the
most extreme values for opinion variables in either case,
although this tendency is smaller as the size of the net-
work increases. A clear visualisation of this phenomenon
is depicted in Fig. 5 where one can see that the cumula-
tive relative frequency of values of x is almost vertical at
the extremes for normal agents. The proportion of nor-
mal agents with x > 980 is 51.6, 47.4, 39.5 and 34.3%,
for N = 50, 250, 500 and 1000, respectively. The picture
is very different for contrarians, their distribution tends
to be uniform independent of the network size, except at
the extreme values where the percentage of agents with
x > 980 is 4.8, 10.4, 11.6 and 6.4%, for N = 50, 250,
500 and 1000, respectively.
FIG. 5: Empirical distributions of opinions for 10 realisations
on networks of sizes 50, 250, and 500 and for single realisation
for the system size of 1000 agents, where blue and orange
symbols stand for normal and contrarian agents, respectively.
As simulations with over 500 agents are very time-
consuming, we have not tried to obtain results with large
statistics for these cases. We did, however, run some sin-
gular simulations with very high agent numbers, to see
whether the patterns observed above would hold even
there. Fig. 6 shows the resultant graph, rank-rank, and
rank-degree correlation scatterplots for a simulation with
1000 agents. As we can see, the contrarians have become
more numerous and have ever more extreme ranking pa-
rameters, as could be expected from the earlier results.
Another interesting feature that emerges from the anal-
ysis of 10 numerical realisations is that the degree distri-
bution is bimodal for normal agents, that is, there is a
large number of agents with low or high degree and very
few around the mean degree value. The range of degrees
7
FIG. 6: The rank-rank (a) and rank-degree correlations (b)
for a single simulation with N = 1000 agents, along with the
final configuration of the network (c); the colour codes and
symbols are the same as in Figs 1 (c), 3 (a), and 4 (b).
and the median of the distribution increase with the size
of the network.
For contrarians the picture is different. their degree
distribution is unimodal (although not normal) and their
median is smaller. Furthermore, their degree range is
about half the one for normal agents. This is clearly seen
in Fig. 7, were we plot the relative frequency as a function
of the degree (k). Observe that the proportion of isolated
agents (k = 0) is larger for the contrarians although this
tendency diminishes as the networks become larger.
FIG. 7: Histogram showing the degree distribution of agents
for 10 realisations on networks of sizes 50, 250, and 500 and
for single realisation for the system size of 1000 agents.
0.00.20.40.60.81.0cumulative relative frequencyN=50normal agentcontrarianN=250100050005001000opinion xi0.00.20.40.60.81.0cumulative relative frequencyN=500100050005001000opinion xiN=1000050100ranking parameter ai020406080100ranking parameter ajN=1000anormal agents, P,x>0normal agents, P,x<0contrarians, different polaritynormal/contrarian, sgnxi=sgnxjnormal/contrarian, sgnxisgnxj050100ranking parameter ai050100150200degree kibnormal agent, Pi>0normal agent, Pi<0contrarian, Pi>0contrarian, Pi<0cnormal agentcontrariancontrarian1.00.50.00.51.0Pi/Pmax0510150.000.100.200.300.40relative frequencyN=500204060N=250normal agentcontrarian050100degree ki0.000.050.100.150.20relative frequencyN=5000100200degree kiN=1000 V. APPROXIMATIVE ANALYSIS OF THE
NUMERICAL RESULTS
In order to better understand the behaviour of the
model we need to perform a thorough analysis of its me-
chanics. A convenient starting point for this effort is
studying the tendency of the agents with extreme rank-
ing parameters to form large fully connected communi-
ties, while the agents of more average rank form small
fully connected communities. Just from considering the
nature of the BTH one might formulate the hypothesis
that this may be because of agents with average rankings
not forming links with agents with extreme rankings due
to it not being status-wise beneficial.
If an agent sees
that the other agent would gain more in having a link
between them, that link will probably not be created by
the agent. By inspecting the rewiring rules Eqs.(11) and
(12) this hypothesis can be verified. Without loss of gen-
erality, we can limit our investigation to the case of two
agents with positive ranking pressures and opinion vari-
ables considering forming a link, as the case with agents
with negative ranking pressures and opinion variables fol-
lows from a similar line of reasoning. From Eq. (12) one
finds that if agent i considers forming a link with agent
j, the ranking pressures and the opinion variables of the
agents and their neighbours must satisfy the condition
xj +(cid:80)
(ki + 2)xj + xi +(cid:80)
k∈m1(j) xk + xi
k∈m1(i) xk
Pi
Pj
>
(15)
for the potential link to be acceptable to agent i. If we
assume, for simplicity, that the opinion variables of the
agents and their neighbours have about the same value,
then condition (15) simplifies to
Pi
Pj
>
1
2
kj + 1
ki + 3/2
,
which in turn simplifies to
Pi
Pj
>
1
2
,
(16)
(17)
if we further assume that ki, kj (cid:29) 1 and ki ≈ kj, which
is reasonable considering that most of the agents at least
have large number of neighbours and tend to belong to
almost fully connected communities. Here the implicit
assumption is that the agents i and j would belong to
the same block, if they formed a link. Written in terms
of ranking parameters (17) becomes
where (cid:104)a(cid:105) = N−1(cid:80)
ai >
(aj + (cid:104)a(cid:105)) ,
1
2
(18)
k ak is the average of the rank-
ing parameters. From (18) it directly follows that if
agent j has the maximum allowed ranking parameter,
aj = amax, the link between the agents i and j will
only be formed if ai > 0.75amax, since (cid:104)a(cid:105) ≈ 0.5amax.
8
Similarly,
if aj = 0.75amax, the two agents will only
bond if ai > 0.625amax, and if aj = 0.625amax, only
if ai > 0.5625amax, and with every iteration the range of
possible ranking parameters of agent i shrinks. This pat-
tern is remarkably apparent in the rank-rank correlation
Figs. 3 and 6, in which the emergence of the contrarians
becomes increasingly clear as the population numbers are
increased. For aj = 0.5amax we find that ai must also
be 0.5amax for a link to be formed, which explains the
tendency of the agents with average ranking parameters
to have so few neighbours, as is seen in Figs.3 and 6.
To understand the emergence of the contrarians one
must analyze the dynamics of the opinion variables, en-
capsulated in Eqs.(9) and (10). Naively thinking, one
would expect agents to always choose the orientation of
their opinion parameters according to their ranking pres-
sures. This means that agents with positive ranking pres-
sures would prefer positive opinion variables, and simi-
larly agents with negative ranking pressures would pre-
fer negative opinion variables. The first hypothesis as to
how some agents would choose to go against these logi-
cal positions is related to the competitive nature of the
model's social interactions. It may be that some agents
cannot compete in this setting and, therefore, choose to
use contrarian strategies instead.
Let us approach this question the same way as above,
focusing on an agent with positive ranking pressure and
initially positive opinion variable, connected to other sim-
ilar agents in the way revealed in Section IV. Let us as-
sume that, initially, all the agents with positive ranking
pressure increase their opinion variables and, conversely,
that all the agents with negative ranking pressure de-
crease their opinion variables in a simulation step.
In
this case the social gain function of agent i is
∆i(t1, t0) = dx(N + 1)(ki + 1)Pi − dx
(cid:88)
(kj + 1)Pj
j∈P +
+dx
(kj + 1)Pj,
(19)
(cid:88)
j∈P −
where P + is the set of agents with positive ranking pres-
sure and P − the set of agents with negative ranking pres-
sure. Now,∆i(t1, t0) needs to be positive if agent i is to
continue raising its opinion variable. This yields a con-
dition of the form
(cid:88)
j∈P +
(kj + 1)Pj − (cid:88)
(kj + 1)Pj
j∈P −
.
(ki+1)Pi >
1
N + 1
(20)
If we denote the number of agents in P + and P − with
N + and N−, respectively, we can write this formula in
terms of averages over P + and P − as follows:
(cid:0)N +(cid:104)(k + 1)P(cid:105)+ − N−(cid:104)(k + 1)P(cid:105)−(cid:1) ,
(ki+1)Pi >
1
N + 1
(21)
where (cid:104)(cid:105)+ denotes average over P +, and, likewise, (cid:104)(cid:105)−
denotes average over P −. If we assume N (cid:29) 1 and that
the situation is symmetric, i.e.
N +(cid:104)(k + 1)P(cid:105)+ ≈ −N−(cid:104)(k + 1)P(cid:105)−,
which further simplifies to
(ki + 1)Pi > (cid:104)(k + 1)P(cid:105)+,
since N = N + + N− and N + ≈ N−.
(22)
(23)
Condition (23) basically states that revising opinion
parameter upwards, which is the "natural" direction for
an agent in P +, is only status-wise beneficial for the agent
if the product of the agent's degree (plus one) and the
ranking pressure are above the average of the same prod-
uct over the whole group P +. We can see from this that
the agents with low ranking pressures and the amounts
of connections fare badly under the condition (23). As
observed earlier, in the simulations the agents with av-
erage ranking parameters, who correspondingly have low
ranking pressures, are also the ones with least connec-
tions, which accordingly means that they are most likely
to become contrarians, at least with regard to (23).
But what about the observed phenomenon of more and
more of agents with higher and higher ranking parame-
ters (and thus ranking pressures) to become contrarians
as the total numbers of simulated agents increases?. The
answer to this question can be found by substituting the
numbers obtained for the sizes of the different communi-
ties to the right hand side of inequality (23), along with
the average ranking parameters for these communities,
as shown in Table II.
(cid:104)a(cid:105)n
n Mn
0.875amax
1 0.25N
2 0.125N
0.6875amax
3 0.0625N 0.59375amax
4 0.03125N 0.546875amax
TABLE II: The approximate number of members and the
average ranking parameters of the 4 largest communities in
P +.
From condition(18) it follows that the largest commu-
nity in P + comprises of those agents with ranking param-
eters over 0.75amax, which means that the community
will have approximately 0.25N members, when one takes
into account the fact that the ranking parameters are
uniformly distributed. The second largest community,
likewise, consists of those agents with ranking param-
eters between 0.625amax and 0.75amax, and has about
0.125N members. The nth (n > 1) largest group will
i=2 2−i)
i=2 2−n), and have 2−(n+1)N members.
While ranking parameters naturally vary from agent to
agent within the communities, the average value of the
ranking parameters of each group (cid:104)a(cid:105)n falls approxi-
mately to the middle point of each ranking range due
to the uniform distribution of the parameters:
have ranking parameters between amax(1 −(cid:80)n+1
and amax(1 −(cid:80)i+2
(cid:104)a(cid:105)n = amax
1 − (
2−i + 2−(n+1))
.
(24)
(cid:33)
(cid:32)
n+1(cid:88)
i=2
9
From Figs. 3 and 6 we see that only a maximum of
four to five of these communities exist in practice at any
one time, so we limit our approximation to these groups.
By assuming the groups to be fully connected, as they
seem to be in the graphs, and approximating sums of the
products of the degrees and ranking parameters with the
products of their average values we get
(cid:104)(k + 1)P(cid:105)+ ≈ 2
M 2
n
,
(25)
4(cid:88)
n=1
(cid:16)(cid:104)a(cid:105)n − amax
(cid:17)
2
where (cid:104)(cid:105)n denotes the average ranking parameter and
Mn the number of members of the nth largest group.
Substituting the values given in Table II we get
(cid:104)(k + 1)P(cid:105)+ ≈ 0.05N 2amax,
which in turn can be inserted into (23):
(ki + 1)Pi > 0.05N 2amax.
(26)
(27)
Finally, using the definition of Pi the condition (27) can
be written in the form
αi >
1
2
+ 0.05
N
ki + 1
,
(28)
where αi = ai/amax.
From condition (28) we can see that the probability
of agent i following the conventional wisdom diminishes
with rising N and decreasing ki, which is what we saw
happening in our simulations judging from the results
shown in the previous Section. To take an example, for
the N = 500 case the largest community of agents with
positive ranking pressures comprises of about 125 agents.
This means, according to the inequality (28), that the
agents with αi > 0.7 could definitively be expected to
always choose to have positive opinion variables.
From Fig. 3 we can tell that the real threshold is closer
to αi > 0.8, which is, however, in remarkably good agree-
ment with the approximate value of 0.7 when one takes
into the consideration the fact that the appearance of
the contrarians themselves was not taken into account in
the derivation of (28), and that for N = 500 they are al-
ready very prominent. If one were to derive the condition
equivalent to (28) with contrarian strategies taken into
account, one would need to consider the effect that the
contrarians have on their neighbours' total social value
V . Thus, condition (28) will most likely not hold for
networks with larger N .
The last question we need to address as regards to
the contrarians is the fact that they are often embed-
ded in the groups of normal agents. So why would it be
status-wise beneficial for an agent using normal strategy
to retain, let alone form, a link with a contrarian agent?
Let us consider a situation where an agent pursues con-
trarian strategies in a group of normal agents. Returning
to inequality (15), we see that it is acceptable for a nor-
mal agent i with Pi > 0 and maximal xi = xmax to form
10
a link with a contrarian agent j with Pj > 0 and minimal
xj = −xmax if
VI. DISCUSSION
> −kj
Pi
Pj
(29)
and we assume that all the neighbours of agents i and j
also have the maximal opinion parameters. The striking
fact about this relation is that it is always fulfilled in this
case, meaning that i would always find formation of links
with contrarians acceptable. From the point of view of
agent j, however, linking to i is only acceptable if
Pj
Pi
>
1
2
ki
kj + 1/2
,
(30)
i.e.
which leads to the very same result as before.
j
will not form a link with i unless Pj > Pi/2, if i and j
are to belong to the same group. Thus the contrarian
agents would behave and be treated as normal agents
when forming relations, which is surprising considering
that their contribution to the total social value of other
agents is negative. The latter fact is demonstrated in the
simulations with some of the most counterintuitive be-
haviours of the model, namely, relations between agents
being first broken and immediately reinstated. Let us
use Eq. (11) to determine, whether the agent i from the
previous calculations would benefit from cutting the link
with agent j, even when expression (29) says that i would
also form a link with j in the event that such a link did
not exist. With the previously stated assumptions, we
arrive to the following condition
Pj
Pi
> − 1
kj − 1
(31)
for the link to be cut. From (31) we see that unless
kj = 0, agent i will cut its ties with the contrarians.
Since kj (cid:29) 1 for the largest groups, the inequality (31)
is likely to be true most of the time, leading to links be-
tween agents i and j being cut and immediately reformed
repeatedly, since inequality (29) also holds. As suggested
above, we have observed this behavioural pattern in our
simulations, and to some extent it can be observed in
Fig. 2, in which it is seen that only the connections be-
tween contrarian and normal agents i and j can the signs
of xi and Pj being of opposite signs. It should again be
stressed, however, that the calculations above do not take
into account the existence of more than one contrarian.
Having more contrarians in the system allows them to
form links between each other, which has a sizeable ef-
fect on the overall structure of the network.
In summary, it could be said that while conditions (17)
and (30) provide surprisingly well fitting approximations
as to how a given agent chooses to link with other agents,
the conditions (28), (29) and (31) (though pointing to the
right direction) only give vague qualitative explanations
for the behaviour of contrarians and can not be expected
to yield precise numerical predictions.
The interpretation of the ranking parameter ai serves
as the key to find possible parallels between our model
and the real world. As it describes a single property of
an agent, the links between the agents only correspond to
exchanges of opinion on whether the agents with larger
ai are "better" than the agents with smaller ai, or vice
versa.
Agents could be considered as being embedded a larger
social context, and in this context they could, in prin-
ciple, have other social connections.
In this case, the
results presented in the previous Section are best in-
terpreted in terms of echo chambers, which means that
agents prefer such a social hierarchy in which they have
better relative rank, and seek to communicate their opin-
ion to others. The agents whose ranking parameters are
further away from the average, are more vocal in broad-
casting their views and gather supporters, since they rank
highly in their chosen hierarchy and, therefore, would
benefit from their hierarchy becoming more widely ac-
cepted.
On the other, the agents with average ranking param-
eters are much more reluctant to take part in the conver-
sation at all, since they do not rank highly in either of the
hierarchies. Then the end result for small system sizes is
that agents divide themselves according to their ranking
pressure into two or more distinct communities support-
ing opposing hierarchies, in which the agents with similar
rankings lump together and refuse to communicate with
those that disagree. It is the shutting out of the opposing
point of view that makes this system's behaviour remi-
niscent of echo chambers found in reality. However, with
increasing system size the agents develop more nuanced
positions on their preferred hierarchies due to mounting
social competition, as is seen in the emergence of the
contrarians.
There is, however, an alternative way to interpret the
ranking parameter. It could be taken to represent an ag-
gregate of all the social properties of an agent, thereby
representing its total standing in the societal status mea-
sures. In this case the connections could represent the
totality of the agents' social interactions, and the opin-
ion variables the agents' attitude to the (current) state
society at large. With this interpretation the rupture be-
tween the different communities observed for smaller sys-
tem sizes would actually represent a real disintegration
of the society. This might have implications concerning
early human migrations, as they could easily have been
influenced by social pressures as well as material needs.
If the environmental pressures define a minimum group
size necessary for a comfortable life for a tribe, and this
minimum is smaller than the limit at which the tribe is
forced to be adopting more advanced strategies to en-
hance social stability, as exemplified by the contrarians
in our simulations, the tribe may well split, with splinter
groups migrating elsewhere.
Other than the different economic games, BTH can
also shed some light into the well known paradox of value,
also known as diamond-water paradox, which refers to
the fact that diamonds are far more valued in monetary
terms than water, even though water is necessary for life
and diamonds not. From the BTH view point the so-
lution to this paradox is obvious: Water, being neces-
sary requirement for life, has to be available in sufficient
quantities to all living humans, which means that owning
water or its source does not set an individual apart from
others, that is, an individual cannot really compare fa-
vorably to others on grounds of having water. Diamonds,
on the other hand, are relatively rare, and thus cannot
be owned by everyone. Therefore, an individual possess-
ing diamonds is compared favourably to others, and so
diamonds acquire a relatively high value in comparison
to water in the minds of humans, in a very similar man-
ner with which the Veblen goods become valuable. Then
BTH, in a sense, contains in itself a natural definition of
value, although further work is needed to determine how
exactly this status-value relates to other forms of value,
such as value derived from usefulness or necessity.
In Section II we only analysed the behaviour of the
accepter, since this is straightforward in comparison to
predicting the behaviour of the proposer. The experi-
ments on the Ultimatum Game often find that the pro-
posers tend to offer fair shares to accepters, which is eas-
ily explained in the context of BTH by the desire of the
proposer to have the proposal accepted: the proposers
only offer shares that they would accept themselves, and
in this way Eq.(3) also restricts the proposers offers, al-
though it cannot tell the exact amount of money offered.
To be able to give a better estimate for the offers one
would need to study the learning processes that shape
the proposers experience on how uneven treatment peo-
ple are usually willing to tolerate. This is, however, out-
side the scope of this paper.
The behaviour of dictators in the dictator game [30]
is somewhat more difficult to analyse using BTH. The
dictator game is similar to the Ultimatum game, the only
difference being that the other player does not even get
to make a choice, and only receives what the first player,
or dictator, endows. It has been observed that [24] in this
game the dictators tend to be rather generous, which is
difficult but not impossible to explain in the context of
BTH, if one takes into account the effect of reputation
and other "social goods". The nature of such influence
on the behaviour of the dictator will be studied in a later
work.
However, there are some indications that BTH could
very well be applied to the dictator game when all the
social effects are taken into account. It has been reported
[31] that when the rules of the dictator game are modified
so that instead of giving money to the other players, the
dictator gets to take some or all of the money given to
the other players (thus turning the game into a "taking
game"), the dictator's behaviour changes from egalitar-
ian to self serving, i.e. taking often the majority or even
all of the available money. From the BTH point of view,
11
the dictator's observed behaviour change can potentially
be explained in terms of social norms. In the ordinary
dictator game the dictator may still feel bound by the
usual norms of the society, while in the "taking game" it
is encouraged to go against these norms. This sets the
"taker" apart from the other player in particular, and
other members of the society in general. Hence the dic-
tator feels "better" than the others when breaking the
norms with impunity, and act on this feeling by taking
money from the other players. The fact that BTH can
possibly lead to formation of norms as well as rebellion
against these norms is well worth of further studies.
VII. CONCLUSIONS
Relating to the known results of the Ultimatum game,
we have formulated a hypothesis explaining the observed
behaviour of humans in terms of superiority maximiza-
tion, or "better than"-hypothesis, and presented a sim-
ple agent-based model to implement this hypothesis. The
model describes agents with constant ranking parameters
and raises the question whether the agents with larger
ranks are "better" than agents with smaller ranks or the
other way around.
We have found that the social system produced by our
model, features homophily, meaning that agents form-
ing social ties with other agents with similar ranking pa-
rameters, and assortativity, describing the tendency of
highly/lowly connected agents forming links with other
highly/lowly connected agents. In addition we find com-
munity formation, both in terms of there being commu-
nities with opposing opinions and in terms of the commu-
nities with the same opinion fracturing into smaller ones
according to their ranking parameters. Furthermore, we
have observed the formation of a hierarchy, in the sense of
a connectivity hierarchy being formed on top of the one
defined by the ranking parameters, with the agents with
extreme ranking parameters presenting higher connectiv-
ity than the agents with average ranking parameters.
Moreover, we have found that the resulting social net-
works tend to be disconnected for small system sizes, but
mostly connected for larger system sizes. This fact may
have some relevance for research of early human migra-
tions, hinting of the effects of social pressure in shaping
the social network.
Acknowledgments
J.E.S. acknowledges financial support from Niilo He-
lander's foundation, G.I. acknowledges a Visiting Fellow-
ship from the Aalto Science Institute, and K.K. acknowl-
edges financial support by the Academy of Finland Re-
search project (COSDYN) No. 276439 and EU HORI-
ZON 2020 FET Open RIA project (IBSEN) No. 662725.
R.A.B. wants to thank Aalto University for kind hospital-
ity during the development of this work. RAB acknowl-
edges financial support from Conacyt through project
799616. We acknowledge the computational resources
provided by the Aalto Science-IT project.
12
[1] Adler, A. (1924) The Practice and Theory of Individual
Psychology. Routledge, Trench and Trubner & Co, Ltd.
[2] Veblen, T. (1899) The Theory of the Leisure Class.
Macmillan.
[3] Duesenberry, J. S. (1949) Income, savings, and the
theory of consumer behaviour. Harvard University Press:
Cambridge, Massachusetts.
[4] Packard, V. (1959) The Status Seekers. LONGMANS,
GREEN AND CO LTD6 & 7 CLIFFORD STREET,
LONDON WI.
[5] Eastman, J. K., Goldsmith, R. E. & Flynn, L. R.
(1999) Status Consumption in Consumer Behaviour:
Scale Development and Validation. Journal of Market-
ing Theory and Practice, 7(3), 41 -- 51.
[6] Weiss, Y. & Fershtman, C. (1998) Social status and
economic performance: A survey. European Economic
Review, 42, 801 -- 820.
[7] Rege, M. (2008) Why do people care about social sta-
tus?. Journal of Economic Behavior & Organization, 66,
233 -- 242.
[8] Gaspart, F. & Seki, E. (2008) Cooperation, status
seeking and competitive behaviour: theory and evidence.
Journal of Economic Behavior & Organization, 51, 51 --
77.
[9] Sidanius, J. & Pratto, F. (1999) Social Dominance:
An Intergroup Theory of Social Hierarchy and Oppres-
sion. Cambridge University Press.
[10] Guth, W., Schmittberger, R. & Schwarze, B.
(1982) An Experimental Analysis of Ultimatum Bargain-
ing. Journal of Economic Behavior & Organization, 3(4),
367 -- 388.
[11] Fehr, E. & Schmidt, K. (1999) A theory of fairness,
competition and cooperation. Quarterly Journal of Eco-
nomics, 114(3), 817 -- 868.
[12] Bolton, G. & Ockenfels, A. (2000) Erc: A theory of
equity, reciprocity and competition. The American Eco-
nomic Review, 90, 166 -- 193.
[13] Luttmer, E. F. P. (2005) Neighbors as Negatives: Rel-
ative Earnings and Well Being. Quarterly Journal of Eco-
nomics, pp. 963 -- 1002.
[14] Izuma, K., Saito, D. N. & Sadato, N. (2008) Pro-
cessing of Social and Monetary Rewards in the Human
Striatum. Neuron, 58(2), 284 -- 294.
[15] Zink, C. F., Tong, Y., Chen, Q., Bassett, D. S.,
Stein, J. L. & Meyer-Lindenberg, A. (2008) Know
Your Place: Neural Processing of Social Hierarchy in Hu-
mans. Neuron, 58(2), 273 -- 283.
[16] Axelrod, R. & Hamilton, W. D. (1981) The Evolu-
tion of Cooperation. Science, 211, 1390 -- 1396.
[17] Axelrod, R. (1981) An Evolutionary Approach to
Norms. American Political Science Review, 80(4), 1095 --
1111.
[18] Young, H. P. (1993) The evolution of conventions.
Econometrica, 61, 57 -- 84.
[19]
(1995) The economics of convention. The Jour-
nal of Economic Perspectives, 10, 105 -- 122.
[20] Centola, D., Willer, R. & Macy, M. (2005) The
Emperor's Dilemma: A Computational Model of Self-
Enforcing Norms. American Journal of Sociology, 10(4),
1009 -- 1040.
[21] Fagundes, M. S., Ossowski, S., Cerquides, J. &
Noriega, P. (2016) Design and evaluation of norm-
aware agents based on Normative Markov Decision Pro-
cesses. International Journal of Approximate Reasoning,
78, 33 -- 61.
[22] Thaler, R. H. (1988) Anomalies: The Ultimatum
Game. The Journal of Economic Perspectives, 2(4), 195 --
206.
[23] Guth, W. & Tietze, R. (1990) Ultimatum bargaining
behavior: A survey and comparison of experimental re-
sults. Jornal of Econonomic Psychology, 11(3), 417 -- 449.
[24] J., H., Boyd, R., Bowles, S., Camerer, C., Fehr,
E. & Gintis, H. (2004) Foundations of Human Sociality:
Economic Experiments and Ethnographic Evidence from
Fifteen Small-Scale Societies. Oxford University Press.
[25] Oosterbeek, H., Sloof, R. & van de Kuilen, G.
(2004) Cultural Differences in Ultimatum Game Exper-
iments: Evidence from a Meta-Analysis. Experimental
Economics, 7(2), 171 -- 188.
[26] Forsythe, R., Horowitz, J., Savin, N. E. & Sefton,
M. (1994) Fairness in Simple Bargaining Experiments.
Games and Economic Behavior, 6, 347 -- 369.
[27] Bolton, G. E. & Zwick, R. (1995) Anonymity versus
punishments in ultimatum bargaining. Games and Eco-
nomic Behavior, 10(1), 95 -- 121.
[28] Iniguez, G., Kert´esz, J., Kaski, K. K. & & Bar-
rio, R. A. (2009) Opinion and community formation in
coevolving networks. Physical Review E, 80, 066119.
[29] Piraveenan, M., Prokopenko, M. & Zomaya, A. Y.
(2008) Local assortativeness in scale-free networks. Eu-
rophysics Letters, 84(2), 28002.
[30] Kahneman, D., Knetsch, J. L. & Thaler, R. H.
(1986) Fairness and the Assumptions of Economics. The
Journal of Business, 59(4), 285 -- 300.
[31] List, J. A. (2007) On the Interpretation of Giving in
Dictator Games. Journal of Political Economy, 115(3),
482 -- 493.
|
1905.08085 | 5 | 1905 | 2019-11-28T03:39:46 | Arena: A General Evaluation Platform and Building Toolkit for Multi-Agent Intelligence | [
"cs.MA",
"cs.AI"
] | Learning agents that are not only capable of taking tests, but also innovating is becoming a hot topic in AI. One of the most promising paths towards this vision is multi-agent learning, where agents act as the environment for each other, and improving each agent means proposing new problems for others. However, existing evaluation platforms are either not compatible with multi-agent settings, or limited to a specific game. That is, there is not yet a general evaluation platform for research on multi-agent intelligence. To this end, we introduce Arena, a general evaluation platform for multi-agent intelligence with 35 games of diverse logics and representations. Furthermore, multi-agent intelligence is still at the stage where many problems remain unexplored. Therefore, we provide a building toolkit for researchers to easily invent and build novel multi-agent problems from the provided game set based on a GUI-configurable social tree and five basic multi-agent reward schemes. Finally, we provide Python implementations of five state-of-the-art deep multi-agent reinforcement learning baselines. Along with the baseline implementations, we release a set of 100 best agents/teams that we can train with different training schemes for each game, as the base for evaluating agents with population performance. As such, the research community can perform comparisons under a stable and uniform standard. All the implementations and accompanied tutorials have been open-sourced for the community at https://sites.google.com/view/arena-unity/. | cs.MA | cs |
Arena: A General Evaluation Platform and Building Toolkit for Multi-Agent
Intelligence
Yuhang Song,1 Andrzej Wojcicki,3 Thomas Lukasiewicz,1 Jianyi Wang,4 Abi Aryan,5
Zhenghua Xu,1,2∗ Mai Xu,4 Zihan Ding,6 Lianlong Wu1
1Department of Computer Science, University of Oxford, United Kingdom
2State Key Laboratory of Reliability and Intelligence of Electrical Equipment, Hebei University of Technology, China
3Lighthouse, 4School of Electronic and Information Engineering, Beihang University, China
5University of California, Los Angeles, United States, 6Imperial College London, United Kingdom
{yuhang.song,thomas.lukasiewicz}@cs.ox.ac.uk, [email protected], [email protected]
Abstract
Learning agents that are not only capable of taking tests, but
also innovating is becoming a hot topic in AI. One of the most
promising paths towards this vision is multi-agent learning,
where agents act as the environment for each other, and im-
proving each agent means proposing new problems for others.
However, existing evaluation platforms are either not compat-
ible with multi-agent settings, or limited to a specific game.
That is, there is not yet a general evaluation platform for re-
search on multi-agent intelligence. To this end, we introduce
Arena, a general evaluation platform for multi-agent intel-
ligence with 35 games of diverse logics and representations.
Furthermore, multi-agent intelligence is still at the stage where
many problems remain unexplored. Therefore, we provide a
building toolkit for researchers to easily invent and build novel
multi-agent problems from the provided game set based on
a GUI-configurable social tree and five basic multi-agent re-
ward schemes. Finally, we provide Python implementations
of five state-of-the-art deep multi-agent reinforcement learn-
ing baselines. Along with the baseline implementations, we
release a set of 100 best agents/teams that we can train with
different training schemes for each game, as the base for eval-
uating agents with population performance. As such, the re-
search community can perform comparisons under a stable
and uniform standard. All the implementations and accompa-
nied tutorials have been open-sourced for the community at
https://sites.google.com/view/arena-unity/.
Introduction
Modern learning algorithms are more of outstanding test-
takers, but less of innovators, i.e., the ceiling of an agent's
intelligence may be limited by the complexity of its environ-
ment (Leibo et al. 2019). Thus, the emergence of innovation
is becoming a hot topic for AI. One of the most promising
paths towards such a vision is learning via social interaction,
i.e., multi-agent learning. In multi-agent learning, how the
agents should beat the opponents or collaborate with each
other is not defined or limited by the creator of the environ-
ment, e.g., the inventor of the ancient Go never defines what
strategies are good. However, enormous and sophisticated
∗Corresponding author: Zhenghua Xu.
Copyright c(cid:13) 2020, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
strategies are invented while a population of human play-
ers/artificial agents evolves by improving themselves over
the others, i.e., each agent is acting as an environment for the
others and improving itself means proposing new problems
for the others.
To study a new class of intelligence, general evaluation
platforms with diverse games are milestones that push for-
ward the research to the next levels. For example, ALE
(Bellemare et al. 2013), Mujoco (Todorov, Erez, and Tassa
2012), and DM-Suite (Tassa et al. 2018) are the most
spread general evaluation platforms that greatly acceler-
ate the research in general reinforcement learning. How-
ever, there is no such general evaluation platform for multi-
agent intelligence. Although some platforms support multi-
agent settings (Wydmuch, Kempka, and Ja´skowski 2018;
Vinyals et al. 2017), they are not general evaluation plat-
forms, i.e., built for specific games. Thus, in this paper, we
propose the first general evaluation platform for multi-agent
intelligence, called Arena, containing 35 multi-agent games
in total, with diverse logics and representations; see Fig. 1.
Apart from training and evaluation, multi-agent intelli-
gence research is still at a stage where many problems remain
undiscovered or unexplored. Thus, the second contribution
of Arena is a building toolkit for multi-agent intelligence,
enabling the easy creation of different multi-agent scenarios.
For example, in the sample game in Fig. 2 (a), after defining
the basic behavior of the agent (i.e., moving and turning) and
the "alive" state of the agent (i.e., it stays on the playground),
it can be extended to different multi-agent scenarios with
minimal effort. For example, (1) five players fight each other
until only one agent is left alive (see Fig. 2 (b)), or (2) 5 2
players form 2 teams, and each agent fights for its own team
until all players in a team are dead (see Fig. 2 (c)), or (3)
multiple players form multiple teams in hierarchies, where
the collaboration and competition relationships between the
teams are customized (see Fig. 3).
Thus, Arena is not just a research platform for the evalua-
tion with a fixed set of games, but also a building toolkit for
researchers to invent and build novel multi-agent problems.
To achieve the above vision of building a toolkit for multi-
agent intelligence, (1) we provide a GUI-configurable tree
that defines the social structure of agents, called social tree;
Figure 1: Game set of Arena.
and (2) based on the social tree, we propose 5 basic multi-
agent reward schemes (BMaRSs), which define different
social paradigms at each node in the social tree. Specifically,
each BMaRS is a restriction applied to the reward function, so
it corresponds to a batch of reward functions that can lead to
a specific social paradigm. For each BMaRS, Arena provides
multiple ready-to-use reward functions, simplifying the con-
struction of games with complex social relationships. Further-
more, if the agent is controlling each joint of a robot, it has
long been a burden for researchers that low-level intelligence
(such as the basic skill of moving) must first be built, before
they can study high-level multi-agent intelligence (Heess et
al. 2017). Thus, Arena provides many ready-to-use dense
reward functions in each BMaRS that handle such low-level
intelligence. Additionally, Arena also offers a verification
option for customized reward functions, so the researchers
can make sure that the programmed reward functions lie in
one of the BMaRSs that produces a specific social paradigm.
Thus, with the above efforts towards a building toolkit for
multi-agent intelligence and the provided set of 35 games for
a general evaluation platform, one can easily customize a set
of games of a new social paradigm to study a yet unexplored
problem.
Finally, we provide Python implementations of several
state-of-the-art deep multi-agent reinforcement learning base-
lines, which can be used as starting points for the develop-
ment of novel multi-agent algorithms, as well as the valida-
tion of new environments. Along with the baseline imple-
mentations, we also release a set of 100 best agents/teams
that we can train with different training schemes for each
game, as the base for evaluating agents with population per-
formance (Balduzzi et al. 2019; Balduzzi et al. 2018). So, the
research community can perform comparisons under a stable
and uniform standard.
To summarize, this paper's contributions are as follows:
(1) a general evaluation platform for multi-agent intelligence
with a set of diverse games, most of which are new to the
community or still stand as a challenge for state-of-the-art
algorithms, (2) a building toolkit for multi-agent games, en-
abling the easy creation of new social paradigms based on
GUI-configurable social trees and BMaRSs, (3) the baseline
implementations of 5 state-of-the-art multi-agent algorithms
for both competitive and collaborative settings, and (4) sets
of benchmark agents/teams for the community to conduct
stable and uniform population evaluation (Balduzzi et al.
2018). Code for games, building toolkit, and baselines, as
well as all corresponding tutorials have been released online
at https://sites.google.com/view/arena-unity/.
The Platform
State-of-the-Art Engine. The engine behind Arena is the
world-leading game engine Unity (Juliani et al. 2018), which
provides Arena with several desirable features on render-
ing, physics, customizability, and community. There are
also other choices of popular engines. Some platforms con-
tain a wide set of diverse games (Bellemare et al. 2013;
Nichol et al. 2018; Perez-Liebana et al. 2016; OpenAI 2016).
However, they are designed mostly for single-agent scenarios
and are extremely hard to customize (adding multiple players
or creating new games), since the games are provided as com-
piled binary ROMs. Other downsides of these choices include
deterministic environments, unrealistic rendering, and unreal-
istic physics. Other platforms (Todorov, Erez, and Tassa 2012;
Tassa et al. 2018) are, in nature, more physics engines
than game engines, which lack a visual editor for eas-
ily creating customized games, and cannot handle more
"game-like" features, such as instantiating and destroying
objects in real-time during the simulation. The rest of
the platforms are limited in the sense that they are built
for specific tasks, such as for first-person shooting (Wyd-
AirHockeyBilliardsBlowBlowBoomerCrawlerPushCrossroadsOffTheGroundFighterKickBoxingMiniPUBG-VillageMiniPUBG-WindridgeMiniPUBG-ForestMiniPUBG-FloodedGroundMiniPUBG-SunTemplePingPongRealRace-LakesRealRace-NightBattleCity-FPRealRace-DriftFightForFoodSoccersBattleCity-TPTennisPushBoxDestroyPyramidTransportBricks RunToGoalYouShallNotPassSumoPushBallTowerDefenseLiftBallReacherGoBangCaptureFlagRunner Figure 2: Game examples of the extensible multi-agent platform.
much, Kempka, and Ja´skowski 2018), Real-Time Strategy
(RTS) (Tian et al. 2017), vision understanding (Qiu et al.
2017), in-door scene understanding (Handa et al. 2016;
Brodeur et al. 2017; Savva et al. 2017; Chang et al. 2017;
Puig et al. 2018; Gao et al. 2019), surviving (Suarez et al.
2019), and interaction (Wu et al. 2018; Savva et al. 2019;
Kolve et al. 2017), or specific games, such as Starcraft
(Vinyals et al. 2017) and Dota2 (OpenAI 2018). Thus, creat-
ing a general evaluation platform on these engines is not a
reasonable choice. DeepMind Lab (Beattie et al. 2016), Psy-
chlab (Leibo et al. 2018), and Malmo (Johnson et al. 2016)
are more appropriate choices when building a customizable
general evaluation platform. However, the main drawbacks
of the above engines are tied to their dated nature. The render-
ing system of these engines are either low-polygon pixelated
(Malmo, based on Minecraft) or outdated (DeepMind Lab
and Psychlab, based on Quake III). The physics systems of
these engines are either rudimentary (Malmo), or have a gap
(Juliani et al. 2018) to the physical world (DeepMind Lab
and Psychlab). Besides, they are all incompatible with a vi-
sual editor, making it quite cumbersome to build customized
scenarios.
To summarize, built on Unity, Arena has the following
advantages over other platforms: (1) realistic rendering, so
that features, such as complex lighting, textures, and shaders,
are fully handled by the background engine and easily pro-
duced in a customized game, (2) realistic physics, so that
enough and realistic stochasticity is introduced in the game
and transferring a policy learned within a simulator to the
real world is easier, (3) user-friendly visual editor, so that
building new multi-agent scenarios in Arena is easy, and (4)
a large and active development community, so that creating
new games is easy with millions of off-the-shelf assets.
Game Sets Towards General Intelligence. The first con-
tribution of Arena is to provide a set of multi-agent games
with diverse game logics and representations, so that it may
push forward the research of general multi-agent intelligence.
Specifically, Arena provides: (1) 27 new games that are not
yet studied in the community, (2) 8 games, of which the ba-
sic logics are inspired by other research, but equipped with
realistic rendering effects, physics engine, and all features
described in the following two paragraphs, such as extensibil-
ity to other social paradigms, and (3) interface to the popular
stand-alone domain StarCraft. The game set is shown in Fig.
1. For more detailed information, see Tables 2-7 in the ex-
tended paper (Song et al. 2019).
Building Toolkit for Multi-Agent Environments. As the
second contribution, we provide a building toolkit for multi-
agent environments: we provide (1) a GUI-configurable so-
cial tree that defines how agents are grouped together with
each other, and (2) 5 basic multi-agent reward schemes
(BMaRSs) applied on each node in the social tree, so that
different social relationships can be easily built and verified,
and low-level intelligence (like motor skills) can be handled.
Other Features. Learning to communicate is an important
research area in multi-agent intelligence (Das et al. 2017;
Mordatch and Abbeel 2018). Thus, Arena provides a broad-
cast board at each node of the social tree (accessible for any
agent as a child of the node), which enables the study of learn-
ing communication at each level. Also, global states may be
used in research for different purposes (Lowe et al. 2017;
Gupta, Egorov, and Kochenderfer 2017; Foerster et al. 2017;
Foerster et al. 2018). Thus, Arena provides the option to
broadcast it to all agents. Besides, a top-down view of the
global game is often appreciated for visualizing population
behavior (Johnson et al. 2016; Wydmuch, Kempka, and
Ja´skowski 2018; Jaderberg et al. 2018; Liu et al. 2019). Thus,
Arena by default enables this option. Finally, there is a neces-
sity for competitive agents to evaluate against human players,
and also a research trend for collaborative agents to team up
with human players. Thus, Arena provides a gaming interface
for humans, so that a human player can take the place of any
agent in the game.
Basic Multi-Agent Reward Schemes and
Social Trees
Preliminaries. We consider a Markov game as defined in
(Littman 1994), consisting of multiple agents x P X , a fi-
nite global state space S, a finite action space Ax for each
Ś(cid:32)
(
agent x, and a bounded-step reward space rx,t P R for each
`
agent x. The environment consists of a transition function
Ś(cid:32)
(
Ax : x P X
Ñ S, which is a stochastic function
g : S
`
st`1 „ g
st,pax,tqxPX
, a reward function for each agent
fx : S
Ax : x P X
Ñ R, which is a deterministic
st,pax,tqxPX
function rx,t`1 " fx
, a joint reward function
f " pfxqxPX , and episode reward Rf
x "
t"1 rx,t for each
agent x under the joint reward function f. For the agent, we
consider that it observes sx,t P Sx, where Sx consists of a
part of the information from the global state space S. Thus,
we have a policy πx : Sx Ñ Ax, which is a stochastic func-
tion ax,t „ πxpsx,tq. Besides, we consider that the agent x
can take a policy πx from a set of policies Πx and assume
that the random seed of all sampling operations is k, which
is sampled from the whole seed space K.
We investigate the effect of tx : x P Xu and tπx : πx P
ř
T
(c)5 VS 5(a)(b)Action SpacePlayground1 VS 1 VS 1 VS 1 VS 1Single agent x : x P Xu. By applying different restrictions on
Πxu on tRf
the effect, we have different BMaRSs, each one of which is
a set of joint reward functions F " tf : u that produce a
similar effect on the population X . The term reward scheme
first appears in (Tampuu et al. 2017) as a tabular, which is
applied to a special case of Pong. While we define it in a
general form and show that many examples are special cases
within this general form.
In a non-sequential setting (normal-form game), the re-
ward scheme serves a similar purpose as the payoff matrix
(Myerson 2013), which is also represented as a tabular. See
Lemma 2 in the extended paper (Song et al. 2019) for how the
payoff matrix is aligned with BMaRSs. In the following, we
define 5 different BMaRSs. Along defining these BMaRSs,
we also describe the ready-to-use reward functions f within
these BMaRSs, which is provided by Arena as a dropdown
list.
Non-learnable BMaRSs (F N L) are a set of joint reward
functions f as follows:
(cid:32)
(
f : @k P K,@x P X ,@πx P Πx,
x {Bπx " 0
,
BRf
F N L "
(1)
,
(2)
F IS "
BRf
xBπx1
@x1 P Xztxu,@πx P Πx,@πx1 P Πx1 ,
where 0 is a zero matrix of the same size and shape as the
parameter space that defines πx. Intuitively, F N L means that
x for any agent x P X cannot be optimized by improving
Rf
its policy πx.
Isolated BMaRSs (F IS) are a set of joint reward functions
f as follows:
(cid:32)
(
f : f R F N L and @k P K,@x P X ,
" 0
Intuitively, F IS means that the episode reward Rf
x received
by any agent x P X is not related to any policy πx1 taken by
any other agent x1 P Xztxu.
Reward functions fx in f of F IS are often called internal
reward functions in other multi-agent approaches (Hendtlass
2004; Jaderberg et al. 2018; Bansal et al. 2018), meaning
that apart from the reward functions applied at a population
level (such as win/loss), which are too sparse to learn, there
are also reward functions directing the learning process to-
wards receiving the population-level rewards, but are more
frequently available, i.e., more dense (Heess et al. 2017).
F IS is especially practical if the agent is a robot requiring
continuous control of applying force on each of its joints,
which means basic motor skills (such as moving) need to be
learned before generating population-level intelligence. Thus,
we provide f in F IS of energy cost, punishment of applying
a big force, encouragement of keeping a steady velocity, and
moving distance towards target.
Competitive BMaRSs (F CP ),
Daskalakis 2011), are defined as
F CP "
(cid:32)
(
f : f R F N L Y F IS and @k P K,@x P X ,
" 0
@πx P Πx,@πx1 P Πx1 ,
inspired by (Cai and
x1dx1
ş
(3)
B
,
x1PX Rf
Bπx
which intuitively means that for any agent x P X , taking any
possible policy πx P Πx, the sum of the episode reward of all
agents will not change. If the episode length is 1, it expresses
a classic multi-player zero-sum game (Cai and Daskalakis
2011). Useful examples of f within F CP are: (1) agents fight
for a limited amount of resources that are always exhausted
at the end of the episode, and the agent is rewarded for the
amount of resources that it gained, and (2) fight till death,
and the reward is given based on the order of death (the
reward can also be based on the reversed order, so that the
one departing the game first receives the highest reward,
such as in some poker games, the one who first discards all
cards wins). Rock, Paper, and Scissors in normal-form game
(Myerson 2013) and Cyclic Game in (Balduzzi et al. 2019)
are both special cases of F CP ; see Lemmas 2 and 3 in the
extended paper (Song et al. 2019).
Collaborative BMaRSs (F CL),
Daskalakis 2011), are defined as
F CL "
(cid:32)
(
f : f R F N L Y F IS and @k P K,@x P X ,
ě 0
@x1 P Xztxu,@πx P Πx,@πx1 P Πx1,
inspired by (Cai and
(4)
,
BRf
x1
BRf
x
x1 {BRf
x1 {BRf
which, intuitively, means that there is no conflict of interest
x ă 0) for any pair of agents px1, xq. Besides,
(BRf
since f R F N L Y F IS, there is at least one pair of agents
x ą 0. This indicates that this
px, x1q that makes BRf
pair of agents shares a common interest, so that improving
x1 for agent x1. The most
x for agent x means improving Rf
Rf
common example of f within F CL is that fx for all x P X is
identical, such as the moving distance of an object that can be
pushed forward by the joint effort of multiple agents, or the
alive duration of the population (as long as there is at least
one agent alive in the population, the population is alive).
Thus, we provide f in F CL: living time of the team (both
positive and negative, since some games require the team to
survive as long as possible, while other games require the
team to depart as early as possible, such as poker).
Competitive and Collaborative Mixed BMaRSs (F CC)
ş
are defined as a catch-all for any other than the above four
ones. First, the term B
x1dx1 {Bπx " 0 in (3) can be
xdx1 " 0 (see Lemma 1 in ex-
x1PX BRf
written as
tended paper (Song et al. 2019), which makes an alternative
(3). Considering F CP in this alternative (3) and F CL in (4),
an intuitive explanation of F CC is that there exist circum-
ş
stances when BRf
x ă 0, meaning that the agents are
x1 {BRf
competitive at this point. But the derivative of total interest
x1 {BRf
x1PX BRf
xdx1 is not always 0; thus, the total interest
can be maximized with specific policies, meaning that the
agents are collaborative at this point.
ş
x1 {BRf
x1PX Rf
Apart from providing several practical f in each BMaRS,
we also provide a verification option for each BMaRS, mean-
ing that one can customize an f and use this verification
option to make sure that the programmed f lies in a specific
BMaRS. The implementation of verification option can be
found in Section 1 in the extended paper (Song et al. 2019).
The Social Tree. The BMaRSs defined above apply to an
Figure 3: An example of a social tree and BMaRSs applied on it.
agent group of all sizes. To define more complex and struc-
tured social paradigms, we use a tree structure (social tree)
to organize the agents and apply BMaRSs on each node of
the tree. We illustrate this by an example. The GUI interface
in Fig. 3 (a) defines a tree structure in Fig. 3 (b), representing
a population of 4 agents.
The tree structure can be easily reconfigured by dragging,
duplicating, or deleting nodes in the GUI interface in Fig. 3
(a). In this example, each agent has an agent-level BMaRS.
The agent is a robot ant, so that the agent-level BMaRSs are
F IS, specifically, the option of ant-motion that directs the
learning towards basic motion skills such as moving forward,
as shown in Fig. 3 (c). Each two agents form a team (which
is a set of agents or teams), the two agents have team-level
BMaRSs. In this example, the two robot ants collaborate with
each other to push a box forward, as shown in Fig. 3 (d). Thus,
the team-level BMaRSs are F CL, specifically, the moving
distance of the box. On the two teams, we have global-level
BMaRSs. In this example, the two teams are set to have a
match regarding which team pushes its box to the target point
first, as shown in Fig. 3 (e). Thus, the global-level BMaRSs
are F CP , specifically, the ranking of the box reaching the
target. The final reward function applied to each agent is a
weighted sum of the above three BMaRSs at three levels.
One can imagine defining a social tree of more than three
levels, where small teams form bigger teams, and BMaRSs
are defined at each node to give more complex and structured
social problems. After defining the social tree and applying
BMaRSs on each node, the environment is ready for use
with an abstraction layer handling everything else, such as
assigning viewports to each agent in the window, applying
the team color, displaying the agent ID, and generating a
top-down view.
The Learning Agents
The Baselines. We provide Python implementations of sev-
eral state-of-the-art baselines that can be used as starting
points for the development of novel multi-agent algorithms,
as well as for the validation of new environments. Specifically,
we first implement a fully decentralized system, where each
agent is a self-contained PPO (Schulman et al. 2017), with in-
dependent critic, actor, and optimizer. We also implement two
state-of-the-art methods based on self-play in (OpenAI 2018)
(SP) and population-based training in (Jaderberg et al. 2018;
DeepMind 2019) (PB). For collaborative agents, we imple-
ment two state of the arts: centralized critic (Lowe et al. 2017)
(CC) and centralized critic with a counterfactual baseline (Fo-
erster et al. 2018) (CF).
The Evaluation Metric. It is recently raising attention that
evaluating an agent against a single-agent or hand-coded bot
is unstable and misleading (Balduzzi et al. 2018). Thus, the
population performance is introduced to evaluate an agent's
(or an agent group's) performance among a base population.
To enable population evaluation, we release 100 best agents,
which we can train with different training schemes for each
game as the base population. One can call the provided func-
tion to get the ranking of an agent among the base population,
or get the averaged ranking of a population among the base
population. Moreover, we provide a human ranking among
the base population, which provides an indication of human-
level intelligence in the game. We will accept the submission
of agents from the community as well as keep implementing
algorithms introduced in the future, so that the base popula-
tion will be upgraded, as the level of research in multi-agent
intelligence advances.
Experiments
Experiments are conducted from three aspects. First, we eval-
uate our game set from the perspective of stochasticity, realis-
tic rendering, and simulation speed, all of which are presented
in the extended paper (Song et al. 2019) due to page limit.
Other advantages from the Unity engine have been verified
by (Juliani et al. 2018). Second, we evaluate our design of
the extensible multi-agent building toolkit with a case study,
showing that by applying different social trees and BMaRSs,
different population-level strategies can be learned. Third,
we report the experimental results of 5 baselines that we im-
plemented and show that by using the provided population
performance evaluation metric, the training progress can be
visualized in a less noisy and more analyzable way.
Case Study of Social Tree and BMaRSs. We use the game
Crossroads from Arena to study the effectiveness of the pro-
posed social tree and BMaRSs via designing different social
paradigms. Specifically, in the game Crossroads shown in Fig.
5 (a), the agent can move and turn, the final goal of the agent
is to reach the target on the other side of the crossroad. By
defining different social trees and applying different BMaRSs,
as shown in Fig. 5 (b) to (d), the agents learn different strate-
gies. In Fig. 5 (b), isolated BMaRSs (F IS) are applied to
all agents, i.e., each agent minimizes the time that it takes
to reach the target. The result shows that the learned agents
simply rush forward, and they easily crash with each other at
the center of the crossroad, producing a traffic jam. In Fig. 5
(c), collaborative BMaRSs (F CL) are applied to the parent
Team-level(b)(c)(d)(e)Agent-levelGlobal-level(a) Figure 4: Visualizing training progress over episode reward (a,b) and population performance (c,d) of different baselines: D-PPO (Decentralized
Proximal Policy Optimization), SP (Self-Play), PB (Population-Based training), CC (Centralized Critic), and CF (Counterfactual Baseline).
with clear performance gaps.
Related Work
Surveys of multi-agent intelligence research can be found in
(Hernandez-Leal, Kartal, and Taylor 2018). Different ideas
have been explored on competitive and collaborative multi-
agent settings.
Collaborative Settings. The simplest way to deploy multi-
agent collaborative systems is to make each agent have a
completely independent learning process (fully decentral-
ized) (Matignon, Laurent, and Le Fort-Piat 2012). However,
collaborative behaviors are hardly observed under such fully
decentralized setting; thus, a fully centralized system is uti-
lized in (Peng et al. 2017), where the policy has access to the
global state and is shared by all agents. However, it is imprac-
tical, since the global state is mostly unavailable in practice,
and the system does not support extending the number of
agents. Thus, centralized training and decentralized execu-
tion are gaining attention (Kraemer and Banerjee 2016). For
multi-agent systems, this idea is mostly explored under actor-
critic algorithms (Foerster et al. 2018). Other ideas include
using a joint action-value function, (Lauer and Riedmiller
2004) addressing the variance problem by a large batch size
(Bansal et al. 2018), and learning grounded cooperative com-
munication protocols between agents (Foerster et al. 2016).
Competitive Settings. Competitive multi-agent intelligence
originally comes from computational game theory (Bowl-
ing et al. 2015). Later on, deep multi-agent reinforcement
learning (D-MARL) is preferred, due to its scalability,
and as it achieves notable advances on two-player games,
such as Poker and Go (Moravcík et al. 2017; Silver et al.
2017). Later, D-MARL was applied to more diverse prob-
lems, such as high-dimensional video games (OpenAI 2018;
DeepMind 2019) and those involving physics control (Bansal
et al. 2018). When solving more practical problems, many
issues have been raised, such as ensuring diversity amongst
agents (Marivate 2015), avoiding overfitting to the policy of
the opponents (Lanctot et al. 2017). Many ideas address such
issues (Kleiman-Weiner et al. 2016). Following on D-MARL,
a very promising recent direction is self-play (Tesauro 1995).
Fictitious self-play (Heinrich and Silver 2016) first shows
promising performance on the competitive game Leduc Poker.
However, as the stability and parallelizability are improving
with the invention of new reinforcement learning algorithms,
state-of-the-art approaches adopt a simpler form of self-play
(OpenAI 2018), which produces a superior-human intelli-
gence on large video games, like Dota2. Another promising
Figure 5: Case Study of Social Tree and BMaRSs.
node of all agents, i.e., all agents are rewarded with the time
that the last one of them takes to reach the target. The result
shows that the agents learn to wait for each other to go across
the crossroad, so that they can all get across as efficiently
as possible. In Fig. 5 (d), collaborative BMaRSs (F CL) are
applied on the parent node of every 4 agents (which form a
team), and competitive BMaRSs (F CP ) are applied on the
parent node of the two teams. Specifically speaking, each two
agents in the same team are rewarded with the same reward,
and the reward is 1 for the team that gets all of its agents to
the target first, 0 for the other team. The results show that
each team learns to block the road of the other team with
one agent, so that the other agents in the team can get across
undisturbed. Then, the agent that blocks the road leaves for
the target, after all its teammates have reached the target.
Baselines and Evaluation Metric. We compare 5 baselines
on two games: (1) Crossroads in Fig. 5 (a) with the BMaRS
settings of Fig. 5 (d) and (2) PushBox in Fig. 3 (e) with the
BMaRS settings of Fig. 3 (b). The BMaRS settings of both
games contain competitive as well as collaborative social
relationships, i.e., multiple agents form collaborative teams,
and teams compete with each other. Thus, we investigate SP
and PB baselines at the level of teams competing with each
other, as well as investigate CC and CF baselines at the level
of agents collaborating with each other in a team. As can be
seen, the curve of episode reward shown in Fig. 4 (a) and
(b) is extremely noisy, as the environment is non-stationary
with the strategy of other collaborators and/or competitors
evolving during the training. However, in Fig. 4 (c) and (d),
which is the curve of ranking in the released base population,
i.e., population performance, all methods are comparable
(a) Crossroads(b) PushBox(c) Crossroads(d) PushBoxEpisode RewardEpisode RewardPopulation PerformancePopulation Performance(b)(a)(c)(d) recent idea is population-based training, as adopted in Star-
Craft (DeepMind 2019).
Summary and Outlook
This paper has introduced the first general evaluation plat-
form for multi-agent intelligence research. Besides, with the
efforts on a building toolkit of multi-agent environments,
the platform also allows for easily building new multi-agent
problems. Additionally, with the released implementations of
several state-of-the-art baselines, researchers can start their
adventure instantly. Finally, by releasing a base population,
the community can conduct comparisons under a stable and
uniform evaluation metric.
Acknowledgments. This work was supported by the China
Scholarship Council under the State Scholarship Fund, by
the Graduate Travel and Special Project Grants from the
Somerville College of the University of Oxford, by the Alan
Turing Institute under the UK EPSRC grant EP/N510129/1,
by the AXA Reseach Fund, by the National Natural Science
Foundation of China under the grants 61906063, 61876013,
and 61922009, by the Natural Science Foundation of Tianjin
under the grant 19JCQNJC00400, by the "100 Talents Plan"
of Hebei Province, and by the Yuanguang Scholar Fund of
Hebei University of Technology.
Experiments: Realistic Rendering, Simulation Speed and Stochasticity
Realistic Rendering. Realistic rendering in games is gaining more consideration, as the research community is moving
towards transferring the algorithms to real-world scenarios. Some of such platforms are (Brodeur et al. 2017; Qiu et al. 2017;
Yan et al. 2018; Kolve et al. 2017; Wu et al. 2018). In Fig. 6, we report an objective comparison of the most realistic scenes
provided in these works against those in our platform. The results show that our platform provides a realistic rendering effect at
the same level as the best of them.
Simulation Speed. Simulation speed and parallelism of an environment are two important factors for carrying out research.
Thus, we compare our game Boomer in Fig. 7 (a) with MsPacman in Fig. 7 (b) from the most widely used general evaluation
platform ALE (Bellemare et al. 2013), both of which run on our parallel implementation of the PPO (D-PPO) baseline on a
server with 32 CPU threads. Then, we compare these two games, as they are of similar complexity. The result in Fig. 7 (c) shows
that Arena allows for parallel implementations while maintaining a similar simulation speed as ALE (Bellemare et al. 2013) on
games of similar complexity when the number of concurrent threads is below the number of CPU threads of the machine, i.e.,
smaller than 32.
Stochasticity. According to (Jaderberg et al. 2018), having enough stochasticity is essential for researchers to verify that their
algorithms are learning general knowledge instead of memorizing action sequences. Thus, we conduct a stochasticity study on
the existing general evaluation platforms ALE (Bellemare et al. 2013), Retro (Nichol et al. 2018), GVG-AI (Perez-Liebana
et al. 2016), Mujoco (Todorov, Erez, and Tassa 2012), and DM-Suite (Tassa et al. 2018) by running a fixed sequence of 1000
actions repeatedly 1000 times and investigating how many branches are produced (averaged over all games in the corresponding
platform). Table 1 shows that our platform generates most stochasticity among them, measured by the number of branches that
the environment produces when applying the same action sequence on it. This is accomplished by introducing stochasticity from
the initial setup (e.g., a randomly generated map), the rendering effect (e.g., randomized light conditions and particle systems),
and the physics system (e.g., randomized physics properties).
How Verification for BMaRSs is Implemented
Since the definitions of BMaRSs contain the statement of @, which requires enumerating over the entire space of some variables,
we first set a hyper-parameter of N, denoting the number of samples that we will apply the verification operation to. If a statement
Y of sample i P I holds true for all N samples from I, we approximately make the judgement that p@i P I, Yq is true. If a
statement Y of sample i P I holds false for any one of N samples from I, we make the judgement that p@i P I, Yq is false. Thus,
in the following, we describe how to make the judgment of whether a statement Y of sample i holds true or not.
Non-learnable BMaRSs (F N L) are defined as:
(cid:32)
f : @k P K,@x P X ,@πx P Πx,
(
" 0
,
(5)
F N L "
x and denote it Rf
x
x. We denote it Rf
x
x1 1´Rf
x1 2
x´π2
π1
Isolated BMaRSs (F IS) are defined as:
BRf
xBπx
where 0 is a zero matrix of the same size and shape as the parameter space that defines πx. For each sample i, we first sample
k P K, x P X ,tπx P ΠxuxPX randomly. Then, we first run one episode to get Rf
1, since it is from the first
2, since it is from the second episode. In the second episode,
episode. We run the second episode to get Rf
we keep the following variables the same as the first episode: k, tπxu for all x P Xztxu. In the second episode, we sample a
different πx from that in the first episode. The πx in the first and second episode are denoted π1
x respectively. Then, we
can compute one sample of BRf
xBπx
(cid:32)
f : @k P K,@x P X ,@x1 P Xztxu,@πx P Πx,@πx1 P Πx1,
(6)
For each sample i, we first sample k P K, x P X , x1 P Xztxu,tπx P ΠxuxPX randomly. Then, we first run one episode to get
2, since it is
x. We denote it Rf
Rf
from the second episode. In the second episode, we keep the following variables the same as the first episode: k, tπxu for all
x
x P Xztx1u. In the second episode, we sample a different πx1 from that in the first episode. The πx1 in the first and second
episode are denoted π1
x1 and make the judgement
whether BRf
1, since it is from the first episode. We run the second episode to get Rf
and make the judgement whether BRf
xBπx
" 0 for sample i holds true.
x and denote it Rf
x
(
" 0
BRf
xBπx1
F IS "
x1 and π2
" Rf
x
x and π2
.
xBπx1 " 0 for sample i holds true.
Competitive BMaRSs (F CP ) are defined as:
xBπx1 " Rf
x1 respectively. Then, we can compute one sample of BRf
(cid:32)
(
f : f R F N L Y F IS and @k P K,@x P X ,@πx P Πx,@πx1 P Πx1,
" 0
1´Rf
x1´π2
x1 dx1
ş
x
π1
B
.
x
2
x1PX Rf
Bπx
F CP "
(7)
Figure 6: Comparison of realistic rendering effect.
(a) Boomer (Arena)
(b) MsPacman (Atari) (c) Comparison of train Frames Per Second (FPS) under different
numbers of concurrent threads.
Figure 7: Comparison of simulation speed.
Table 1: Comparison of the stochasticity of platforms.
Platform ALE Retro GVG-AI Mujoco DM-Suite Arena
922.4
Branches
431.2
50.2
0.0
23.1
12.2
HoMEAI2-THORCHALETUnrealCVHouse3DArenaFPSNumber of Concurrent ThreadsNumber of CPU Threads ş
ş
BRf
x1
BRf
x
ş
1
x1PX Rf
x1
For each sample i, we first sample k P K, x P X ,tπx P ΠxuxPX randomly. Then, we first run one episode to get
x1dx1.
ş
x1dx1 and denote it
We denote it
dx1, since it is from the second episode. In the second episode, we keep the following variables the same as the first
x1PX Rf
x1
episode: k, tπxu for all x P Xztxu and we sample a different πx from that in the first episode. The πx in the first and second
dx1
episode are denoted π1
dx1, since it is from the first episode. We run the second episode to get
x respectively. Then, we can compute one sample of
x1PX Rf
ş
x1PX Rf
x and π2
x1PX Rf
x1PX Rf
x1 dx1
"
x1 1
x1 2
ş
B
2
ş
dx1´
x´π2
π1
x
x1PX Rf
Bπx
ş
B
and make the judgement of whether
Collaborative BMaRSs (F CL) are inspired by (Cai and Daskalakis 2011) and defined as,
x1 dx1
x1PX Rf
Bπx
" 0 for sample i holds true.
(cid:32)
(
f : f R F N L Y F IS and @k P K,@x P X ,@x1 P Xztxu,@πx P Πx,@πx1 P Πx1,
ě 0
.
F CL "
x
For each sample i, we first sample k P K, x P X , x1 P Xztxu,tπx P ΠxuxPX randomly. Then, we first run one episode to get
1 and Rf
x and Rf
x1. We denote it Rf
x and
Rf
x1
x
2, respectively, since it is from the second episode. In the second episode, we keep the following
2 and Rf
Rf
x1 and denote it Rf
x1
variables the same as the first episode: k, tπxu for all x P Xztxu and we sample a different πx from that in the first episode.
Then, we can compute one sample of
ě 0 for sample i holds true.
Competitive and Collaborative Mixed BMaRSs (F CC) are defined to be any situations other than the above four ones:
1, respectively, since it is from the first episode. We run the second episode to get Rf
BRf
x1
BRf
BRf
x1
BRf
x1 1´Rf
x1 2
2 and make the judgement whether
1´Rf
Rf
x
" Rf
(cid:32)
f : f R F N L Y F IS Y F CP Y F CL
(9)
which means if a reward function f is verified to be not among above F N L, F IS, F CP or F CL, it is verified to be among F CC.
Lemma 1 An equivalent expression of
F CC "
(
,
x
x
x
Proof of Lemma 1: Following the defination of F CP , since f R F N L, we have:
‰ 0.
Dk P K,Dx P X ,Dπx P Πx,
(8)
(10)
(11)
(12)
(13)
(14)
F CP "
F CP "
is:
Thus, F CP equals to:
F CP "
"
ş
(cid:32)
(
f : f R F N L Y F IS and @k P K,@x P X ,@πx P Πx,@πx1 P Πx1,
" 0
(cid:32)
(
f : f R F N L Y F IS and @k P K,@x P X ,@πx P Πx,@πx1 P Πx1,
dx1 " 0
x1PX Rf
Bπx
x1dx1
ż
B
.
BRf
x1
BRf
x
x1PX
BRf
xBπx
ż
(cid:32)
(
f : f R F N L Y F IS and @k P K,@x P X ,@πx P Πx,@πx1 P Πx1,
" 0
(cid:32)
(
f : f R F N L Y F IS and @k P K,@x P X ,@πx P Πx,@πx1 P Πx1,
dx1 " 0
dx1 BπxBRf
BRf
x1
Bπx
x1PX
ż
x
,
BRf
x1
BRf
x
x1PX
completing the proof.
Lemma 2 Rock, Paper, and Scissors in normal-form game (Myerson 2013) is a special case of F CP , which is defined in (11).
Player 1
Strategy 1: Rock
Strategy 2: Paper
Strategy 3: Scissors
Strategy 1: Rock
p0, 0q
p1,´1q
p´1, 1q
Strategy 2: Paper
Player 2
p´1, 1q
p0, 0q
p1,´1q
Strategy 3: Scissors
p1,´1q
p´1, 1q
p0, 0q
Table 2: Payoff matrix of Rock, Paper, and Scissors.
BRf
x1
BRf
x
BRf
x1
BRf
x
" 1´p´1q
, we can verify that the payoff matrix satisfies the definition of F CP in (11).
Proof of Lemma 2: The payoff matrix in a normal-form game (Myerson 2013) is a matrix aligning the tuple of tplayer, strategyu
with the payoff received by each player. The reward scheme in a multi-agent learning context is a map aligning the tuple of
tagent x, policy πxu to the episode reward Rf
x received by the agent x. Thus, player is equivalent to agent, strategy is a special
case of policy when the episode length is one, payoff is equivalent to episode reward Rf
x when the episode length is one. Payoff
matrix of Rock, Paper, and Scissors is shown in Table 2. By selecting any pair of cells in the payoff matrix, say p1,´1q and
p´1, 1q, we can compute a
p´1q´1. Thus, by enumerating over all possible pairs of cells in the payoff matrix to compute
all
Lemma 3 Symmetric zero-sum games and cyclic games in (Balduzzi et al. 2019) are a special case of F CP , which is defined in
(11).
Proof of Lemma 3: In (Balduzzi et al. 2019), W denotes a set of agents parameterized by the weights of a neural net. For
instance, w P W and v P W are two agents of specific policies from the full policy set W . φpv, wq evaluates a pair of agents
pw, vq
(15)
The higher φpv, wq, the better for agent v. They refer to φ ą 0, φ ă 0, and φ " 0 as wins, losses, and ties, respectively, for v. A
game is defined as symmetric zero-sum game if:
φ : W W Ñ R.
@w P W,@v P W, φpv, wq " ´φpv, wq.
A cyclic game is a symmetric zero-sum game. As can be seen, when the game is symmetric
@x P X , Πx is the same and Πx " ΠX ,
(16)
(17)
(19)
(20)
W is equivalent to ΠX . When the game has only two players x1 and x2
tx1u " Xztx2u,
(18)
φ is equivalent to Rx1. Thus, v and w are equivalent to πx1 and πx2, respectively and the equivalent expression of symmetric
zero-sum game in (16) is:
The equivalent expression of (11) when there are only two players x1 x2, i.e., satisfying (18), is:
F CP "
@πx1 P ΠX ,@πx2 P ΠX , Rx1 " ´Rx2 .
(cid:32)
(
f : f R F N L Y F IS and @x1 P X ,tx2u " Xztx1u,@πx1 P ΠX ,@πx2 P ΠX ,
" ´1
.
BRx1
BRx2
This shows that symmetric zero-sum games are a special case of F CP . Since cyclic games are symmetric zero-sum games, cyclic
games are also a special case of F CP .
Game
Name
Observation
Action Space
Description
Crossroads-2T1P-
v1.
Visual;
Continuous
Discrete;
/
ArenaCrawlerMove-
2T1P-v1.
Visual
RAM;
/
Continuous;
Get to the target at the other side of the crossroad. The game
uses dense reward function of IS_Target at agent level (Reward-
SchemeScale at this level is 1), and use reward function of
CP_Ranking at global level (RewardSchemeScale at this level
is 100).
Get to the target in the middle. The game uses dense reward func-
tion of IS_Target at agent level (RewardSchemeScale at this level
is 1), specifically, it uses IsRewardMovingTowardsTarget and Is-
PenaltyHeadDown, and use reward function of CP_Ranking at
global level (RewardSchemeScale at this level is 100).
AirHockey.
Visual;
Continuous
Discrete;
/
See (Wiki Page).
Billiards.
Visual;
Discrete;
See (Wiki Page).
BlowBlow.
Visual;
Continuous
Discrete;
/
Stay on the playground, push the opponent off the playground
using your body, or the particles blown from you.
Boomer.
Visual;
Discrete;
Inspired by (Wiki Page).
PushBox.
Visual;
Continuous
(forces at every
joint
of
the
robot ant);
Push the box forward as a team, and try to push the box to the
target point first.
OffTheGround.
Visual;
Continuous
Discrete;
/
Stay on the playground, push the opponent using your body and
try to take advantages of the terrain.
KickBoxing.
Visual;
Discrete;
MiniPUBG-
Village.
Visual;
Discrete;
MiniPUBG-
FloodedGround.
Visual;
Discrete;
MiniPUBG-
Windridge.
MiniPUBG-
SunTemple.
Visual;
Discrete;
Visual;
Discrete;
Learn to use 18 different movements, such as Jab, Cross, Hook,
Uppercut and etc. See (Wiki Page) for more information about
the movements. Different movements cost different amount of
energy. If the energy runs out, no movements can be made. Hitting
different parts of the opponent's body results in different levels of
harm. Thus, the agent is expected to learn to hit the soft part of the
opponent's body with the hard part of its own body, meanwhile,
try to keep a budget of the energy cost.
(Wiki Page) for basic rules of PlayerUnknown's Battlegrounds
(PUBG). In this map of Village, the agent is expected to learn to
take various advantages of the buildings, grass and trees.
See (Wiki Page) for basic rules of PlayerUnknown's Battlegrounds
(PUBG). In this map of FloodedGround, the playground is rela-
tively large, so this game is designed for agents to learn in a larger
population and learn long term planning.
See (Wiki Page) for basic rules of PlayerUnknown's Battlegrounds
(PUBG). In this map of FloodedGround, there are both city areas
and grass land areas. Thus, the agent is expected to learn the
strategies of surviving between city and grassland areas.
See (Wiki Page) for basic rules of PlayerUnknown's Battlegrounds
(PUBG). In this map of SunTemple, there are mainly various build-
ings. Thus, the agent is expected to learn to hide, fight and take
the advantages of the buildings.
PingPong.
Visual;
Continuous
Discrete;
/
Adopt with full rules of standard table tennis, see (Wiki Page).
RealRace-Lakes.
Visual;
Discrete;
This track is general.
RealRace-Sprint.
Visual;
Discrete;
This track contains many long straight roads, where driving too
fast will result in the car easily losing control. Thus, the agent is
expected to learn to maintain the speed within a range that is high
enough but safe and controllable at the same time.
RealRace-Drift.
Visual;
Discrete;
This track is designed for learning drafting.
RealRace-Night.
Visual;
Discrete;
Snake.
Visual;
Soccers.
Visual;
BattleCity-TP.
Visual;
Tennis.
Visual;
Continuous
Discrete;
Continuous
Discrete;
Continuous
Discrete;
Continuous
Discrete;
/
/
/
/
This scene is at night, so the agent is dealing with more complex
night rendering conditions. Besides, this track contains lots of
crossroads, where the agent needs to learn to recognize the road
signs.
Based on a classic video game, see (Wiki Page). If two players
collide with each other, the one of smaller length will be killed.
Adopted from (Liu et al. 2019).
See (Wiki Page). This version of BettleCity is of third-person view.
Adopted from Unity ML-Agents (Juliani et al. 2018).
DestroyPyRAMid.
TransportBricks.
RunToGoal.
YouShallNotPass.
Sumo.
Reacher.
LiftBall.
Visual
RAM;
Visual
RAM;
Visual
RAM;
Visual
RAM;
Visual
RAM;
Visual
RAM;
Visual
RAM;
/
/
/
/
/
/
/
Continuous;
The goal is to destroy the pyramid.
Continuous;
The task is to transport all bricks to the target point.
Continuous;
Adopted from (Bansal et al. 2018).
Continuous;
Adopted from (Bansal et al. 2018).
Continuous;
Adopted from (Bansal et al. 2018).
Continuous;
Reach the target first.
Continuous;
Collaborate to lift the ball.
TowerDefense.
Visual;
Discrete;
Adopted from (Tian et al. 2017).
PushBall.
GoBang.
CaptureFlag.
Visual
RAM;
Visual
RAM;
Visual
RAM;
/
/
/
Continuous;
Control the robot arm to push the ball to the base of the opponent.
Discrete;
See (Wiki Page).
Discrete;
Adopted from (Jaderberg et al. 2018).
BattleCity-FP.
Visual;
Fighter.
Visual;
Continuous
Discrete;
Continuous
Discrete;
/
/
See (Wiki Page). This version of BettleCity is of first-person view.
Take advantages of various bunkers and destroy the opponent.
MiniPUBG-Forest. Visual;
Discrete;
See (Wiki Page) for basic rules of PlayerUnknown's Battlegrounds
(PUBG). In this map of FloodedGround, there are mainly grass-
lands. Thus, the agent is expected to learn the strategies of hiding
and attacking.
FightForFood.
Visual;
Continuous
Discrete;
/
A simple game where agents fight for limited food. When the food
is exhausted, the episode ends.
Runner.
Visual;
Discrete;
A simple game where agents fight for higher score. When there is
a score reach the specified threshold, the episode ends.
References
[Balduzzi et al. 2018] Balduzzi, D.; Tuyls, K.; Perolat, J.; and
Graepel, T. 2018. Re-evaluating evaluation. In NIPS.
[Balduzzi et al. 2019] Balduzzi, D.; Garnelo, M.; Bachrach,
Y.; Czarnecki, W. M.; Perolat, J.; Jaderberg, M.; and Graepel,
T. 2019. Open-ended learning in symmetric zero-sum games.
arXiv:1901.08106.
[Bansal et al. 2018] Bansal, T.; Pachocki, J.; Sidor, S.;
Sutskever, I.; and Mordatch, I. 2018. Emergent complexity
via multi-agent competition. In ICLR.
[Beattie et al. 2016] Beattie, C.; Leibo, J. Z.; Teplyashin, D.;
Ward, T.; Wainwright, M.; Küttler, H.; Lefrancq, A.; Green,
S.; Valdés, V.; Sadik, A.; et al.
2016. Deepmind lab.
arXiv:1612.03801.
[Bellemare et al. 2013] Bellemare, M. G.; Naddaf, Y.; Veness,
J.; and Bowling, M. 2013. The Arcade learning environment:
An evaluation platform for general agents. JAIR.
[Bowling et al. 2015] Bowling, M.; Burch, N.; Johanson, M.;
and Tammelin, O. 2015. Heads-up limit hold'em poker is
solved. Science.
[Brodeur et al. 2017] Brodeur, S.; Perez, E.; Anand, A.;
Golemo, F.; Celotti, L.; Strub, F.; Rouat, J.; Larochelle, H.;
and Courville, A. 2017. Home: A household multimodal
environment. arXiv:1711.11017.
[Cai and Daskalakis 2011] Cai, Y., and Daskalakis, C. 2011.
On minmax theorems for multiplayer games. In SIAM.
[Chang et al. 2017] Chang, A.; Dai, A.; Funkhouser, T.; Hal-
ber, M.; Niessner, M.; Savva, M.; Song, S.; Zeng, A.; and
Zhang, Y. 2017. Matterport3D: Learning from RGB-D data
in indoor environments. arXiv:1709.06158.
[Das et al. 2017] Das, A.; Kottur, S.; Moura, J. M.; Lee, S.;
and Batra, D. 2017. Learning cooperative visual dialog
agents with deep reinforcement learning. In ICCV.
[DeepMind 2019] DeepMind. 2019. AlphaStar: Mastering
the real-time strategy game StarCraft II. In DeepMind.
[Foerster et al. 2016] Foerster, J.; Assael, I. A.; de Freitas, N.;
and Whiteson, S. 2016. Learning to communicate with deep
multi-agent reinforcement learning. In NIPS.
[Foerster et al. 2017] Foerster, J.; Nardelli, N.; Farquhar, G.;
Afouras, T.; Torr, P. H.; Kohli, P.; and Whiteson, S. 2017. Sta-
bilising experience replay for deep multi-agent reinforcement
learning. In ICML.
[Foerster et al. 2018] Foerster, J. N.; Farquhar, G.; Afouras,
T.; Nardelli, N.; and Whiteson, S. 2018. Counterfactual
multi-agent policy gradients. In AAAI.
[Gao et al. 2019] Gao, X.; Gong, R.; Shu, T.; Xie, X.; Wang,
S.; and Zhu, S.-C. 2019. VRKitchen: An interactive 3D
environment for learning real life cooking tasks. In ICML
Workshop RL4RealLife.
[Gupta, Egorov, and Kochenderfer 2017] Gupta,
K.;
Egorov, M.; and Kochenderfer, M. 2017. Cooperative
multi-agent control using deep reinforcement learning. In
AAMAS.
[Handa et al. 2016] Handa, A.; Patraucean, V.; Stent, S.; and
J.
Cipolla, R. 2016. SceneNet: An annotated model generator
for indoor scene understanding. In ICRA.
[Heess et al. 2017] Heess, N.; Sriram, S.; Lemmon, J.; Merel,
J.; Wayne, G.; Tassa, Y.; Erez, T.; Wang, Z.; Eslami, S.; Ried-
miller, M.; et al. 2017. Emergence of locomotion behaviours
in rich environments. arXiv:1707.02286.
[Heinrich and Silver 2016] Heinrich, J., and Silver, D. 2016.
Deep reinforcement learning from self-play in imperfect-
information games. arXiv:1603.01121.
[Hendtlass 2004] Hendtlass, T. 2004. An introduction to
collective intelligence. In SAP.
[Hernandez-Leal, Kartal, and Taylor 2018] Hernandez-Leal,
P.; Kartal, B.; and Taylor, M. E. 2018. Is multiagent deep
reinforcement learning the answer or the question? A brief
survey. arXiv preprint arXiv:1810.05587.
[Jaderberg et al. 2018] Jaderberg, M.; Czarnecki, W. M.; Dun-
ning, I.; Marris, L.; Lever, G.; Castaneda, A. G.; Beattie, C.;
Rabinowitz, N. C.; Morcos, A. S.; Ruderman, A.; et al. 2018.
Human-level performance in first-person multiplayer games
with population-based deep reinforcement learning. In CoRR.
[Johnson et al. 2016] Johnson, M.; Hofmann, K.; Hutton, T.;
and Bignell, D. 2016. The Malmo platform for artificial
intelligence experimentation. In IJCAI.
[Juliani et al. 2018] Juliani, A.; Berges, V.-P.; Vckay, E.; Gao,
Y.; Henry, H.; Mattar, M.; and Lange, D. 2018. Unity: A
general platform for intelligent agents. arXiv:1809.02627.
[Kleiman-Weiner et al. 2016] Kleiman-Weiner, M.; Ho,
M. K.; Austerweil, J. L.; Littman, M. L.; and Tenenbaum,
J. B. 2016. Coordinate to cooperate or compete: abstract
goals and joint intentions in social interaction. In CogSci.
[Kolve et al. 2017] Kolve, E.; Mottaghi, R.; Gordon, D.; Zhu,
Y.; Gupta, A.; and Farhadi, A. 2017. AI2-THOR: An interac-
tive 3D environment for visual AI. arXiv:1712.05474.
[Kraemer and Banerjee 2016] Kraemer, L., and Banerjee, B.
2016. Multi-agent reinforcement learning as a rehearsal for
decentralized planning. Neurocomputing.
[Lanctot et al. 2017] Lanctot, M.; Zambaldi, V.; Gruslys, A.;
Lazaridou, A.; Tuyls, K.; Pérolat, J.; Silver, D.; and Graepel,
T. 2017. A unified game-theoretic approach to multiagent
reinforcement learning. In NIPS.
[Lauer and Riedmiller 2004] Lauer, M., and Riedmiller, M.
2004. Reinforcement learning for stochastic cooperative
multi-agent systems. In AAMAS.
[Leibo et al. 2018] Leibo, J. Z.; d'Autume, C. d. M.; Zoran,
D.; Amos, D.; Beattie, C.; Anderson, K.; Castañeda, A. G.;
Sanchez, M.; Green, S.; Gruslys, A.; et al. 2018. Psychlab:
A psychology laboratory for deep reinforcement learning
agents. arXiv:1801.08116.
[Leibo et al. 2019] Leibo, J. Z.; Hughes, E.; Lanctot, M.; and
Graepel, T. 2019. Autocurricula and the emergence of inno-
vation from social interaction: A manifesto for multi-agent
intelligence research. arXiv:1903.00742.
[Littman 1994] Littman, M. L. 1994. Markov games as a
framework for multi-agent reinforcement learning. In ICML.
[Schulman et al. 2017] Schulman, J.; Wolski, F.; Dhariwal,
P.; Radford, A.; and Klimov, O. 2017. Proximal policy
optimization algorithms. arXiv:1707.06347.
[Silver et al. 2017] Silver, D.; Schrittwieser, J.; Simonyan, K.;
Antonoglou, I.; Huang, A.; Guez, A.; Hubert, T.; Baker, L.;
Lai, M.; Bolton, A.; et al. 2017. Mastering the game of Go
without human knowledge. Nature.
[Song et al. 2019] Song, Y.; Wang, J.; Lukasiewicz, T.; Xu,
Z.; Xu, M.; Ding, Z.; and Wu, L. 2019. Arena: A gen-
eral evaluation platform and building toolkit for multi-agent
intelligence. arXiv preprint arXiv:1905.08085.
[Suarez et al. 2019] Suarez, J.; Du, Y.; Isola, P.; and Mor-
datch, I. 2019. Neural MMO: A massively multiagent game
environment for training and evaluating intelligent agents.
arXiv preprint arXiv:1903.00784.
[Tampuu et al. 2017] Tampuu, A.; Matiisen, T.; Kodelja, D.;
Kuzovkin, I.; Korjus, K.; Aru, J.; Aru, J.; and Vicente, R.
2017. Multiagent cooperation and competition with deep
reinforcement learning. PloS one.
[Tassa et al. 2018] Tassa, Y.; Doron, Y.; Muldal, A.; Erez,
T.; Li, Y.; Casas, D. d. L.; Budden, D.; Abdolmaleki, A.;
Merel, J.; Lefrancq, A.; et al. 2018. Deepmind control suite.
arXiv:1801.00690.
[Tesauro 1995] Tesauro, G. 1995. Temporal difference learn-
ing and TD-Gammon. Communications of ACM.
[Tian et al. 2017] Tian, Y.; Gong, Q.; Shang, W.; Wu, Y.; and
Zitnick, C. L. 2017. Elf: An extensive, lightweight and
flexible research platform for real-time strategy games. In
NIPS.
[Todorov, Erez, and Tassa 2012] Todorov, E.; Erez, T.; and
Tassa, Y. 2012. MuJoCo: A physics engine for model-based
control. In IROS.
[Vinyals et al. 2017] Vinyals, O.; Ewalds, T.; Bartunov, S.;
Georgiev, P.; Vezhnevets, A. S.; Yeo, M.; Makhzani, A.;
Küttler, H.; Agapiou, J.; Schrittwieser, J.; et al.
2017.
StarCraft II: A new challenge for reinforcement learning.
arXiv:1708.04782.
[Wu et al. 2018] Wu, Y.; Wu, Y.; Gkioxari, G.; and Tian, Y.
2018. Building generalizable agents with a realistic and rich
3D environment. arXiv:1801.02209.
[Wydmuch, Kempka, and Ja´skowski 2018] Wydmuch, M.;
Kempka, M.; and Ja´skowski, W. 2018. ViZDoom compe-
titions: Playing doom from pixels. IEEE Transactions on
Games.
[Yan et al. 2018] Yan, C.; Misra, D.; Bennnett, A.; Walsman,
A.; Bisk, Y.; and Artzi, Y. 2018. Chalet: Cornell house agent
learning environment. arXiv:1801.07357.
2015.
[Liu et al. 2019] Liu, S.; Lever, G.; Merel, J.; Tunyasuvu-
nakool, S.; Heess, N.; and Graepel, T. 2019. Emergent
coordination through competition. In ICLR.
[Lowe et al. 2017] Lowe, R.; Wu, Y.; Tamar, A.; Harb, J.;
Abbeel, O. P.; and Mordatch, I. 2017. Multi-agent actor-
critic for mixed cooperative-competitive environments. In
NIPS.
Improved Em-
[Marivate 2015] Marivate, V. N.
pirical Methods in Reinforcement-Learning Evaluation.
Ph.D. Dissertation, Rutgers University-Graduate School-New
Brunswick.
[Matignon, Laurent, and Le Fort-Piat 2012] Matignon, L.;
Laurent, G. J.; and Le Fort-Piat, N. 2012.
Independent
reinforcement learners in cooperative Markov games: A
survey regarding coordination problems. The Knowledge
Engineering Review.
[Moravcík et al. 2017] Moravcík, M.; Schmid, M.; Burch, N.;
Lis`y, V.; Morrill, D.; Bard, N.; Davis, T.; Waugh, K.; Johan-
son, M.; and Bowling, M. 2017. DeepStack: Expert-level
artificial intelligence in heads-up no-limit poker. Science.
[Mordatch and Abbeel 2018] Mordatch, I., and Abbeel, P.
2018. Emergence of grounded compositional language in
multi-agent populations. In AAAI.
[Myerson 2013] Myerson, R. B. 2013. Game theory. Harvard
University Press.
[Nichol et al. 2018] Nichol, A.; Pfau, V.; Hesse, C.; Klimov,
O.; and Schulman, J. 2018. Gotta learn fast: A new bench-
mark for generalization in RL. arXiv:1804.03720.
[OpenAI 2016] OpenAI.
https://github.com/openai/
universe.
[OpenAI 2018] OpenAI.
2018.
https://blog.openai.com/openai-five/.
[Peng et al. 2017] Peng, P.; Yuan, Q.; Wen, Y.; Yang, Y.; Tang,
Z.; Long, H.; and Wang, J. 2017. Multiagent bidirectionally-
coordinated nets for learning to play StarCraft combat games.
arXiv:1703.10069.
[Perez-Liebana et al. 2016] Perez-Liebana, D.; Samothrakis,
S.; Togelius, J.; Schaul, T.; and Lucas, S. M. 2016. General
video game AI: Competition, challenges and opportunities.
In AAAI.
[Puig et al. 2018] Puig, X.; Ra, K.; Boben, M.; Li, J.; Wang,
T.; Fidler, S.; and Torralba, A. 2018. VirtualHome: Simulat-
ing household activities via programs. In CVPR.
[Qiu et al. 2017] Qiu, W.; Zhong, F.; Zhang, Y.; Qiao, S.;
Xiao, Z.; Kim, T. S.; and Wang, Y. 2017. UnrealCV: Virtual
worlds for computer vision. In International Conference on
Multimedia.
[Savva et al. 2017] Savva, M.; Chang, A. X.; Dosovitskiy, A.;
Funkhouser, T.; and Koltun, V. 2017. MINOS: Multimodal
indoor simulator for navigation in complex environments.
arXiv:1712.03931.
[Savva et al. 2019] Savva, M.; Kadian, A.; Maksymets, O.;
Zhao, Y.; Wijmans, E.; Jain, B.; Straub, J.; Liu, J.; Koltun, V.;
Malik, J.; et al. 2019. Habitat: A platform for embodied AI
research. arXiv:1904.01201.
2016. Openai universe.
In
Openai five.
|
1709.08505 | 1 | 1709 | 2017-09-25T14:17:50 | Key Management and Learning based Two Level Data Security for Metering Infrastructure of Smart Grid | [
"cs.MA",
"cs.CR"
] | In the smart grid, smart meters, and numerous control and monitoring applications employ bidirectional wireless communication, where security is a critical issue. In key management based encryption method for the smart grid, the Trusted Third Party (TTP), and links between the smart meter and the third party are assumed to be fully trusted and reliable. However, in wired/wireless medium, a man-in-middle may want to interfere, monitor and control the network, thus exposing its vulnerability. Acknowledging this, in this paper, we propose a novel two level encryption method based on two partially trusted simple servers (constitutes the TTP) which implement this method without increasing packet overhead. One server is responsible for data encryption between the meter and control center/central database, and the other server manages the random sequence of data transmission. Numerical calculation shows that the number of iterations required to decode a message is large which is quite impractical. Furthermore, we introduce One-class support vector machine (machine learning) algorithm for node-to-node authentication utilizing the location information and the data transmission history (node identity, packet size and frequency of transmission). This secures data communication privacy without increasing the complexity of the conventional key management scheme. | cs.MA | cs | A Novel Key Management and Data Encryption
Method for Metering Infrastructure of Smart Grid
Imtiaz Parvez, Student Member, IEEE, Arif I. Sarwat, Member, IEEE, My T. Thai, Senior Member, IEEE,
and Anurag K. Srivastava, Senior Member, IEEE
1
Furthermore, from the fine grained energy consumption data,
the home appliance companies receive information about the
life style patterns of consumers and the energy utilization of
their home appliances. Thus competing companies can use this
valuable information in their businesses. The consumers might
want to alter the consumption data to reduce their electricity
bill. The most crucial thing is that the opponent/hacker might
jam or take over the AMI network by sending false signal
to meters on an unsecured system, which may cause power
outage in a wide area as well as an imbalance in the demand
generation model.
7
1
0
2
p
e
S
5
2
]
A
M
.
s
c
[
1
v
5
0
5
8
0
.
9
0
7
1
:
v
i
X
r
a
Abstract-In the smart grid, smart meters, and numerous
control and monitoring applications employ bidirectional wire-
less communication, where security is a critical issue. In key
management based encryption method for the smart grid, the
Trusted Third Party (TTP), and links between the smart meter
and the third party are assumed to be fully trusted and reliable.
However, in wired/wireless medium, a man-in-middle may want
to interfere, monitor and control the network, thus exposing its
vulnerability. Acknowledging this, in this paper, we propose a
novel two level encryption method based on two partially trusted
simple servers (constitutes the TTP) which implement this method
without increasing packet overhead. One server is responsible
for data encryption between the meter and control center/central
database, and the other server manages the random sequence of
data transmission. Numerical calculation shows that the number
of iterations required to decode a message is large which is
quite impractical. Furthermore, we introduce One-class support
vector machine (machine learning) algorithm for node-to-node
authentication utilizing the location information and the data
transmission history (node identity, packet size and frequency of
transmission). This secures data communication privacy without
increasing the complexity of the conventional key management
scheme.
Keywords-AMI, Key Management Scheme (KMS), localization,
machine learning, One-class Support Vector Machine, RSS, smart
grid, smart meter.
I.
INTRODUCTION
Smart grid is the modern power system infrastructure whose
distribution system is upgraded via bidirectional communica-
tion, and pervasive and intelligent computing capabilities in
order to provide improved control, efficiency, reliability and
safety. In the arena of smart grid, the Advanced Metering
Infrastructure (AMI) is the distribution level building block. It
consists of millions of meters interlinked by a hierarchical or
mesh-connected or hybrid network. The meters communicate
with the control center in a wireless fashion. The typical
communication standards for the AMI are ZigBee, WiFi and
LTE [1]. In the AMI, smart meters collect and store the
instantaneous energy consumption and report it periodically
to the back office (the control center) as against reporting of
monthly total energy consumption recorded in the conventional
meters. It also allows the consumers to actively engage in
electricity trade by selling home- generated, unused electricity
(from solar panel, windmill etc.) to the grid. AMI also outfits
the service provider with control and monitoring equipment,
including outage management, demand response, and disaster
prevention and recovery information.
Since AMI has to use wireless communication, security
issues become a critical problem. Observing the usage patterns,
an adversary/thief can predict the presence of the targeted con-
sumers at home which can be a threat to civil lives and privacy.
This research was supported in part through U.S. National Science Foun-
dation under the grant RIPS-1441223.
Like all other systems, AMI needs to comply with the
requirements of security- confidentiality, integrity, availability
and accountability (non-repudiation) [2], [3]. Confidentially
implies the accessibility of data by any authorized entity, and
any intentional or unintentional disclosures of data must be
denied. Since the main objective of the AMI is to collect
energy consumption information that conveys consumer life
patterns, habits and energy usage, this data must be concealed.
Integrity requires reflecting authentic data correctly without
any modifications, additions, or deletions. If any unauthorized
subject performs (or attempts to perform) such actions, this
must be detected. Since not only the hacker, but also the
consumer, might want to alter the consumption data by trap-
ping and resending, the integrity of data is very important.
Availability requires the accessibility of data by an authorized
user on demand. If the required data is not found at the time
of need, the system violates the availability aspect of the
security requirement of the system. Any natural or intentional
incidents (i.e. hacking) must not hamper the system from
operating correctly. For example, if the hacker wants to jam the
network, the system must comply with the availability aspect.
Accountability (non-repudiation) means non-deniability of an
action, i.e., the entities used in receiving or transmitting data
must not deny it. If an entity does not receive data, it will not,
subsequently, state that it has received it. In the AMI network,
accountability ensures a timely response to the command and
control, and integrity of billing profile, etc.
The main drawback of implementing security scheme in
AMI is the limited memory and low computational ability
of the smart meters. Furthermore, AMI is a huge network,
consists of millions of meters. So we need a lightweight but
robust security scheme. In the literature, the key management-
based encryption method has been proposed as a prominent
security scheme for smart grid, which includes a Trusted Third
Party (TTP) [4]–[6]. In all TTP management systems, it is
assumed that the TTP is fully trusted. However, the TTP itself,
and the meters and communication links among the TTP and
the meters could also be compromised.
Based on the semi-trusted servers of the TTP and un-
trustworthy/unreliable communication links, in this paper, we
propose a key management scheme which includes a node
(meter) to the Service Provider (SP) data encryption as well
as randomized packet transmission. In our early work [7], the
scheme consists of two independent servers. The master server
manages the public-private key before data is sent to the back
office of the SP. Additionally, the auxiliary server receives the
random sequence of data from the smart meter, in response
to the public key sent by the master server. Both the private
key associated with public key and the random sequence are
used to retrieve the data at the back office. In this paper, we
extend [7] by using Received Signal Strength (RSS) and One
class Support Vector Machine (OCSVM) techniques for node
to node authentication. RSS is used for localization of meters
using the received signal strength from neighbor meters. On
the other hand, OCSVM is used to detect new and outlier
data/packet, using current and previous data transmission his-
tory, which has not been considered in the literature to our
best knowledge. The introduction of two separate servers for
key management and random sequenced packet transmission
increases robustness in security in untrustworthy communica-
tion medium and servers. Additionally, OCSVM based node
authentication increases sturdiness in key management system
without increasing any overhead and it is easy to implement
in devices with limited memory and computational ability.
The rest of the paper is organized as follows. Section II
describes the literature review of the security schemes pro-
posed for the AMI. In Section III, details of the proposed
structure of the AMI are introduced. In Section IV, the the-
oretical background of RSS based localization, OCSVM and
entropy of data packet are presented. Section V presents the
communication and traffic flow among smart meters, servers
and SPs. Furthermore, in Section VI, simulation results and
theoretical security strength of a data packet are analyzed.
Finally, concluding remarks are provided in Section VII.
II. LITERATURE REVIEW
In the recent years, security issues in AMI attracted signif-
icant attention of different communities (electrical engineers,
computer science graduates, IT experts etc.) due to extensive
use of wireless communication. In [8], a 128-bit Asymmetric
Encryption Scheme Galois Counter Mode (AES GCM) cryp-
tography based security IC has been proposed and performance
comparison between the hardware-based and software-based
crypto-engines is presented. In [9], [10], a game theory based
security system has been introduced where two parties (an
attacker and a defender) interact pro-actively and reactively
for different attack severity levels. In [11], several nodes are
selected randomly for intermediate node-to-node authentica-
tion. Even though this reduces the packet overhead, it still
has vulnerability in the middle of communication. In [12],
randomization of the AMI configuration has been proposed
to make its behavior unpredictable to the hacker, whereas
the behavior is predictable to the control center. In [13],
Costas et al. introduced anonymization of data by randomizing
node identity using a TTP. But communication overhead is
increased due to the need for the TTP to communicate with all
nodes simultaneously. Moreover, retrieving information from
a large network becomes complicated. In [14], homomorphic
encryption has been introduced, which also has the problem
of retrieving data from a large network being complicated.
In [15], an Identity Based Signcryption (IBS) for zero
configuration encryption and authentication has been proposed
for end-to-end communication solution. In [16], a node-to-
node encryption with its own secret key has been proposed. But
for a large network, the packet overhead increases for both IBS
and node-to-node authentication. To distribute the keys and
2
Fig. 1: AMI architecture consisting of a cluster of
mesh/hybrid connected meters, data concentrator, KMS and
control center.
manage the network, a wireless sensor network based Public
Key management Infrastructure (PKI) has been proposed in
[17] which has problem of generation large number of unique
keys. In [18], a key management system has been introduced
based on DLMS/COSEM standard providing two main infor-
mation security features: data access security and data transport
security. Since DLMS/COSEM is an open standard and allows
a number of variations in the protocol
it
increases the complexity in the client side. In [19], a two layer
security scheme for meter to Data Concentrator (DC) and DC
to control center of SP has been provided in Taiwan. For meter
to DC, IEC 62056 based encryption method, and for DC to the
back office, public key management system has been proposed.
But here in each time step, encryption and decryption need to
be performed twice.
implementation,
Different than these earlier work in the literature, in our
proposed scheme, we provide improvements in the key man-
agement schemes for AMI systems. In particular,
in our
approach, we introduce randomization of data transmission
as well as node-to-node authentication method considering
unreliable communication scenario and using machine learning
technique. The introduction of randomization and the machine
learning based node authentication increases the robustness of
the key management system without increasing any overhead.
III. ARCHITECTURE OF AMI
The AMI is a web like network with millions of meters
as shown in Fig. 1. In here we provide the insights into
AMI architecture, which encompasses components including
home appliances, control center of utility service provider and
security infrastructure.
•
Smart meter: It is a solid state device responsible for
collecting, storing and sending data to the back office
using wireless communication and at interval less than
1 hour. The home appliances are connected to a smart
meter by a network, forming Home Area Network
(HAN).
• Neighborhood Area Network (NAN): The meters are
connected among themselves through a mesh or hierar-
chical or hybrid network termed as NAN. The network
may be wired (e.g. PLC) or wireless (e.g. WiFi, ZigBee,
GPRS). The head end of the NAN is the data concen-
trator or gateway which is connected to back office by
a dedicated wired or wireless connection (optical fiber,
cellular network, etc.).
• Control center/Hardware and software control sys-
tem/Utility back office/SP: The control center receives
MDMS (Metering data management service)Back OfficeMaster serverAuxiliary server jSMnSMiSMkSMWireless communication linkDedicated network (wired/wireless)Data collector/concentrator the consumption data from the AMI network and uses
them for making bills. This fine grained data can be used
to optimize power generation and distribution. Besides
that, electricity distribution, control, and monitoring is
performed from the control center.
• Master server: In our architecture, the master server is
semi- trusted. When a meter requests the key, the master
server generates a public key and private key pair. The
master server uni-casts the public key to the specific
meter for data encryption and the corresponding private
key is sent to the auxiliary server and control center for
data decryption.
• Auxiliary server: Before the encryption of data with the
public key sent by the master server, the smart meter
generates a random sequence. This random sequence
is sent as encrypted by public key to auxiliary server.
The auxiliary server receives random sequence and au-
thenticates it by the private key of smart meter before
forwarding it to the control center for final decryption.
IV. OC-SVM AND RSS ALGORITHM, AND ENTROPY OF
A DATA PACKET
In our scheme, we use RSS for the localization of meter, and
OC-SVM for node-to-node authentication. In the rest of this
section, we describe the two algorithms, and security strength
calculation of a data packet by entropy in further detail.
A. Received Signal Strength (RSS) based localization
RSS based localization can be used in localizing a new
meter in collaboration with meters whose positions are already
known. The location information is used to authenticate the
source node using OCSVM.
Let us consider an unknown positioned meter at a location
(x, y) accompanied by partially dispersed known position
meters at locations (xl, yl), where 1 ≤ l ≤ n. The received
signal strength at location (xl, yl) can be denoted by ψl
ψl = c − 10γ log(dl) + wl,
(1)
where c is an unknown constant that depends on transmitted
power, frequency etc, and γ is the path loss constant. For a
lossy environment, the typical value of γ is 4-6. In our model,
γ = 2.93 has been used considering residential area. The
parameter dl is the euclidean distance between the known and
unknown position meter defined as follows:
dl =(cid:112)(x − xl)2 + (y − yl)2,
(2)
and wl is the zero mean random Gaussian noise with standard
deviation σl. The value of σl ranges from 6 to 12 dB.
Let us define,
the θ and ψ as θ = [x, y, z]T and
ψ = [ψ1, ψ2, ......., ψn]T .
The likelihood function of θ for a given RSS measurement ψ,
f (θψ) is given by
(cid:41)
{ψl − c + 10γ log(dl)}2
2σ2
l
,
(3)
(cid:40)
− n(cid:88)
f (θψ) = c1 exp
where c1 is a constant.
l=1
The Maximum Likelihood (ML) estimate of θ, denoted by θ,
can be found from the following equation
3
(cid:41)
θ = arg max f (θψ)
= arg min
(cid:40)
− n(cid:88)
l=1
{ψl − c + 10γ log(dl)}2
2
2σl
(4)
,
The above equation is an optimization problem. Various op-
timization techniques such as differential evolution, dynamic
relaxation and particle swarp optimization (PSO) can be used
to solved (4). In our problem, we used PSO to solve the non-
linear optimization problem. Finally, the ML estimator yields
the location (x, y) and reference power z of the unknown
positioned meter
(x, y, z) = {θ(1), θ(2), θ(3)}.
(5)
B. OCSVM algorithm
The Support Vector Machine (SVM) is a machine learning
algorithm based on modern statistical learning theory [20],
[21]. It separates two classes by constructing a hyper surface
in the input space. In this input space, the input is mapped
to a higher dimensional feature space by non-linear mapping.
In this section, we describe OCSVM for segregating outliers
from a cluster of data.
Let us consider a data space Ψ = (xi, yi) ∈ Rd
i = {1, 2, 3, ...., n} where xi ∈ Rd is the input data and
yi ∈ {−1, +1} is the corresponding output pattern in the
dedicating class membership. SVM first projects the input
vector x to a higher dimensional space H by a non-linear
operator Φ(·) : Rn ←→ H where the data projection is linearly
separable. The non-linear SVM classification is defined as
Ω(x) = wT Φ(x) + b, w ∈ H, b ∈ R,
(6)
which is linear in terms of projected data Φ(x) and non-linear
in terms of original data x.
Fig. 2: Separation of two class in hyperplane.
SVM tries to maximize the separation margin between two
classes of data in hyperplane as shown in Fig. 2. To prevent
SVM from over fitting with noisy data (or create soft margin),
slack variables (ξi) are introduced to allow some data to lie
within the margin. Then the objective function which includes
the minimization of (cid:107)w(cid:107) can be written as
(cid:107)w(cid:107)
2
min
w,b,ξi
+ c
n(cid:88)
i=1
ξi,
such that yi(wT Φ(xi) + b) ≥ 1 − ξi,
and ξi ≥ 0.
(7)
(8)
Class 1 Class 2Optimal Hyperplane where c > 0 is the regularization parameter that determines the
trade-off of maximizing the margin and the number of training
data within the margin (thus reducing the training errors). To
minimize the objective function of (7) by using the Lagrange
multipliers technique, the necessary condition for w is
w =
γiyiΦi(x)
(9)
n(cid:88)
i=1
where γi > 0, i = 0, 1, 2.....n are Lagrange multipliers
corresponding to the constraints in equation (8).
The Lagrange multipliers (γi) can be solved from equation
(7) and written as
max w(γi) =
n
nγiγjyiyjk(xi, xj) (10)
(cid:88)
j=1
(cid:88)
i=1
nγi − 1
2
(cid:88)
n(cid:88)
i=1
such that 0 ≤ γi ≤ c,
and
γiyi = 0.
(11)
i=1
In (10), k(x, y) = Φ(x)T Φ(y) is known as the kernel
function. It determines the mapping of the input vector to a
high dimensional feature space. In this paper, we consider
two kernel functions. The Guassian RBP kernel is given by:
Gaussian RBF kernel: k(x, y) = exp(
(cid:107)x−y(cid:107)2
σ
)
while the Polynomial kernel is given by:
Polynomial kernel: k(x, y) = ((xy) + 1)p
where p ∈ N is the degree of the polynomial function and
σ ∈ R is the width of the RBF function.
OCSVM detects abnormal data within a class [21]. OCSVM
maps the input vector to feature dimension according to the
kernel function and separates it from the origin with maximum
margin. It penalizes the outliers by employing slack variables
ξ in the objective function and controls carefully the trade off
between empirical risk and regularize the penalty.
(cid:88)
i=1
Fig. 3: Feature dimensional space for OCSVM
The quadratic programming minimization function
min
w,ξi,ρ
1
2
(cid:107)w(cid:107)2 +
1
vn
nξi − ρ
such that (w · Φ(xi) ≥ ρ − ξi,
and ξi ≥ 0,
i = 1, 2, 3, ...., n.
(12)
(13)
where Φ is the kernel function for mapping, ξi is the slack
variables, v ∈ (0, 1] is a prior fixed constant, and ρ is the
decision value that determines whether a given point falls
within the estimated high density region. Then the resultant
decision function f m
w,p(x) takes the form
4
f (x) = sgn(w∗T Φ(x) − ρ∗)
(14)
where ρ∗ and w∗ are the values of w and ρ solving from the
equation (12).
In OCSVM, v characterizes the solution instead of c
•
•
It determines an upper bound on the fraction of outliers.
It is the lower bound on the number of training examples
used as support vector.
(smoothness operation)-
Due to the high significance of v, OCSVM is also termed as
v − SVM.
Since RSS algorithm can pinpoint the location of neigh-
boring meters based on received electromagnetic signals, and
OCSVM can detect anomaly in input data set, the combination
of these two algorithms makes it possible to identify false
nodes. In our security scheme, OCSVM utilizes packet size,
frequency of data transmission and node identity along with
the node position for meter authentication. The details have
been provided in the next section.
C. Entropy of a data packet
Entropy is used to measure the uncertainty of a random
variable or data packet. The more certain about a value, the
smaller is the entropy value.
The entropy for a sequence S
S P (S = x) log2 P (S = x)
where P (S = x) is the probability of taking S a value over
H(s) =(cid:80)
If the size of a random variable or packet generated by the
meter is n bit, then the entropy and security strength of the
data packet are n and 2n respectively.
x.
V. TRAFFIC FLOW PROCESS
In this section, the privacy scheme implementation process
and the data flow are described in detail.
ability.
meters are not fully reliable.
In our architecture, we assume the following assumption:
• The Master and Auxiliary servers are independent and
semi-trusted. However, the servers might physically be
one but virtually divided into two servers.
• The wireless communication links between servers and
• The meters have limited memory and computational
• The control center has adequate computational ability.
• The meters keep the records of position of neighboring
meters, frequency of transmission, packet size and node
identity. Frequency of transmission, node identity and
packet size are extracted from the packet header. The
node position is derived from electromagnetic signals
using RSS based localization as explained in previous
section IV(A).
• Every meter transmits data at a constant transmission
• The packet size is constant for every meter.
• When a new meter is installed,
it starts to record
the position of the neighbor meters, frequency of data
transmission, node identity and packet size.
power.
wwpw 5
consumption data is encrypted and processed by the following
method:
Encryption of data: pk1 ⊕ M1 −→ C1
Segmentation of the encrypted data into n packets: C1 −→
sequence:
Ordering
(m1, m2, m3......mn)
random
(m1, m2, m3......mn) S−→ (h1, h2, h3........hn)
packets
by
STEP 3: Data transmission
The packets are transmitted by the transmitting algorithm as
explained in Algorithm 1:
time instant t
Go to Initialization
Proceed to next step
Algorithm 1 Transmitting algorithm
1: Initialization:
2: Generate Random Sequence, St = (s1, s2, ........, sn) at
3: if St−1 == St then
4:
5: else
6:
7: end if
8: Segment C −→ (m1, m2, m3.....mn)
9: Set Pi = 1
Si
, p3 = 1
s3
P = (p1, p2, p3, ......., pn) where p1 =
10: Transmission:
11: Transmit packet mi chosen from (m1, m2, m3, ....., mn)
with greater probability probability distribution
(p1, p2, p3, ...., pn)
..........., pn = 1
sn
, p2 = 1
s2
1
s1
12: End
STEP 4: Hop to hop data aggregation and forwarding
Based on previous data receiving records containing node
identity, sender node's position, frequency of data received and
packet size- the node SM2 verifies source node SM1 and
forwards data to the next node SM3. OCSVM algorithm is
used to authenticate the source node. The data aggregation and
forwarding algorithm pseudocode is tabulated in Algorithm 2:
STEP 5: Data retrieval
The control center receives the randomized and encrypted
packets and decodes them by the secret key pk and random
sequence S.
S−→
Reordering
(h1, h2, h3, ...., hn)
(m1, m2, m3......., mn)
Message unification: (m1, m2, m3......., mn) −→ C1
pk−→ M1
Decryption: C1
data:
the
VI. SIMULATION RESULT AND SECURITY ANALYSIS
A. Simulation result
To get insights into the localization of the meters, we used
a suburban building block topology (Manhattan grid) [22] in
which we used rectangle and hexagon shape as area of interest
(AOI). In an AOI, each edge is a known positioned meter and
the center is the emitter (unknown positioned meter). In Matlab
simulation, we used γ = 2.93 and σl = 12 dB. Furthermore,
for optimization PSO was used whereas residential pathloss
model was considered for pathloss calculation.
With an increase in the number of nodes (meters), the Mean
Square Error (MSE) from the exact position of the meter
decreases as illustrated in Fig. 5. Furthermore, as the variance
TABLE I: Symbol notation.
Fig. 4: Data flow among various components of AMI.
Notation
SMi
AEi
ki
Ei
Di
pk
sk
Mi
Ci
t
(m1, m2, ......., mn) ∈ C
(s1, s2, ...., sn) ∈ S
(p1, p2, ......, pn) ∈ P
γ
δl
PSO
One-class SVM
for
encryption scheme
Description
Smart meter i
Asymmetric
node/ meter i
Randomized key generation algorithm
for node/ meter i
Encryption algorithm for node i
Deterministic decryption algorithm for
node i
Public key
Secret key
Message/data of meter/node i
Encrypted data/ Message of node/ meter
i
Time instance
Segmented packets of message C
Random sequences
Probability of ith packet transmission
at each time instant t
Path loss component
Variance of random noise
Particle swarm optimization
One class Support Vector Machine
A. Basic Notifications and definitions
The notifications and definitions used in our algorithm are
stated in Table I.
B. Data flow
The data flow (as shown in Fig. 4) among the meters, servers
and control center has be explained here in details.
STEP 1: Initialization
SM1 sends a request for a public key before sending data.
The Master server generates a public key and private key
pair. The public key is unicasted to SM1 for data encryption.
Additionally,
to the
Auxiliary server and the control center.
the corresponding private key is sent
Key generation by asymmetric algorithm:
AE1 = (k1,E1, D1)
k1 −→ (pk1, sk1)
STEP 2: Encryption
The SM1 generates a random sequence (S). S is encrypted
by the received public key and is sent to the Auxiliary server.
The Auxiliary server receives the cypher text and decodes
it by the secret key. After that the Auxiliary server sends
the sequence S to the control center. At the same time, the
SM_1MSASSM_2Control CenterRequest key Generates keys (p_k,s_k)Send public key(p_k) Random sequence (S) generationSend S encrypted by public keyDecrypt S by node secret key (s_k)Send random sequence (S)Data encryption by public key (p_k)Data send according to sequence (S)Node (SM_1) authentication by OC-SVMSM_iData/packet forwardNode (SM_2) authentication by OC-SVMPacket forwardSend secret key (p_k)Send secret key (p_k) Algorithm 2 Data aggregation and forwarding
1: Initialization:
2: if source meter SMi is new meter then
3:
Record data transmission history (frequency of trans-
mission, location of neighbor meter, packet size, node
identity ) for time h
Send attach request to control center
Receive approval of new meter
4:
5:
6: else
7:
signal
8:
9:
10:
11:
12:
13:
then
else
ter
14: end if
15: end if
Derive SMi's location from received electromagnetic
Get frequency (time difference between two packet
sending), node identity, packet size from header informa-
tion of packet
Run OCSVM algorithm
if the new data and previous data belong to same group
Forward data to the next meter
Cease data transmission and report to control cen-
Fig. 5: MSE from the exact position of the meter vs. Number
of nodes
of noise increases, the MSE also increases as depicted in Fig.
6.
In the second part of simulation using Python, we use
OCSVM for novelty/anomaly detection in a given data set. It
takes node identity, node's position, frequency of transmission
and packet size as new data set. The previous authenticate data
set or fraction of new data set can be used as training data. In
the simulation environment, we generate random 100 training
observations, 20 normal regular observations and 20 abnormal
observations. The red line (as shown in Fig. 7 and Fig. 8) is
the learned decision function that separates abnormal data from
the normal regular data. As shown in Fig.7, OCSVM draws a
learned decision boundary based on 100 training clean data, of
which 12 data falls outside the decision boundary erroneously.
Following this, among the 20 regular novel data 3 data falls
outside the boundary and all 20 abnormal data falls outside
the decision boundary. So for 100 training data, OCSVM can
detect anomaly with 100% accuracy (0 error in novel abnormal
detection). If we reduce the number of training samples to 20,
the performance of OCSVM decreases (1 error in 20 novel
abnormal detection) which is depicted in Fig. 8.
Now we use the polluted data for training purposes, which
is a fraction of the input data. There is no separate clean
training sample in this case. We also compare OCSVM with
6
Fig. 6: MSE from the exact position of the meter vs.
Variance (σ2) of Gaussian noise.
robust covariance estimator for different kinds of input data
distribution. In this case, we generate 100 samples, 10% of
which are outliers/abnormal. For a well-centered and elliptic
input data set, OCSVM and robust covariance estimator show
the same performance as illustrated in Fig. 9. For a bi-modal
distribution of data, OCSVM shows better performance than
the robust covariance estimator. This is depicted in Fig. 10.
For the non- Gaussian distributed input data, the performance
of both the methods degrades, wherein OCSVM shows better
performance than the robust covariance estimator, illustrated
in Fig. 11.
B. Security strength analysis
Let us assume, a meter sends a consumption unit packet
of size 256 bit encrypted by 256 bit asymmetric key to the
central database of control center. If the packet is subdivided
into 32 blocks with each block having a size of 8, and a block
is transmitted each time according to a random sequence, then
the entropy is 256. The security strength of the data packet is
2256.
Furthermore, the security strength of a 256 bit asymmetric
So, for 32 random sequenced packets and 256 bit asymmet-
key is 2256/2.
ric key,
Total security strength of the data packet = 2256 + 2256/2
Hence, a hacker needs a maximum (2256 + 2256/2) number
of iterations (tries) to decrypt a message, which is impractical.
VII. CONCLUSION
In our security scheme, a two-level security method has
been proposed- data encryption and node authentication. In
the data encryption level, encryption by asymmetric keys and
randomization of data packets have been proposed. In the
conventional key management system, only data encryption
is used. On the other hand, in our scheme, randomization
of packets along with data encryption ensures enhanced data
security. Another contribution of our scheme is the introduction
of node-to-node authentication by OCSVM, which utilizes four
variables- meter ID, frequency of data reception from a specific
meter, packet size and meter position. The information of meter
ID, data frequency and packet size is easily extractable from
the packet header. On the otherhand, for meter's position RSS
is used and it is almost constant due to stationary position of
the meters.
For TTP-to-smart meter communication, we need a bi-
directional communication similar to meter data communi-
cation. Since the communication between meters and servers
7
Fig. 7: Anomaly detection in 40 data with training data size 100. (a) Contour diagram of decision function. (b) Statistics of
error.
(a)
(b)
Fig. 8: Anomaly detection in 40 data with training data size 20. (a) Contour diagram of decision function. (b) Statistics of error.
(a)
(b)
happens once per every session of sending the meter data to
the control center, it doesn't hamper the normal traffic flow
between the meters and control center. Further, since a random
sequence along with the asymmetric key is used to retrieve data
in the control center, it also helps to verify data flow from a
specific smart meter.
REFERENCES
[1] V. Gungor, D. Sahin, T. Kocak, S. Ergut, C. Buccella, C. Cecati, and
G. Hancke, "Smart Grid Technologies: Communication Technologies
and Standards," Industrial Informatics, IEEE Transactions on, vol. 7,
no. 4, pp. 529–539, Nov 2011.
[2] F. Cleveland, "Cyber security issues for Advanced Metering Infrast-
tructure (AMI)," in Proc. IEEE conf. on Power and Energy Society
General Meeting - Conversion and Delivery of Electrical Energy in the
21st Century, July 2008, pp. 1–5.
[3] Z. Lu, X. Lu, W. Wang, and C. Wang, "Review and evaluation of
security threats on the communication networks in the smart grid,"
in Proc. Conf. on MILITARY COMMUNICATIONS CONFERENCE
(MILCOM), Oct 2010, pp. 1830–1835.
[4] Y. Yan, Y. Qian, and H. Sharif, "A secure and reliable in-network col-
laborative communication scheme for advanced metering infrastructure
in smart grid," in Proc. IEEE Int. conf. on Wireless Communications
and Networking Conference (WCNC), March 2011, pp. 909–914.
[5] S. Das, Y. Ohba, M. Kanda, D. Famolari, and S. Das, "A key
management framework for AMI networks in smart grid," IEEE Com-
munications Magazine, vol. 50, no. 8, pp. 30–37, August 2012.
[6] Z. Wan, G. Wang, Y. Yang, and S. Shi, "SKM: Scalable Key Man-
agement for Advanced Metering Infrastructure in Smart Grids," IEEE
Trans. on Industrial Electronics, vol. 61, no. 12, pp. 7055–7066, Dec
2014.
I. Parvez, A. Islam, and F. Kaleem, "A key management-based two-level
encryption method for AMI," in PES General Meeting - Conference
Exposition, 2014 IEEE, July 2014, pp. 1–5.
[7]
[8] W. Somkaew, S. Thepphaeng, and C. Pirak, "Data security implemen-
tation over ZigBee networks for AMI systems," in Proc. 11th Int. Conf.
on Electrical Engineering/Electronics, Computer, Telecommunications
and Information Technology (ECTI-CON, May 2014, pp. 1–5.
[9] Z. Ismail, J. Leneutre, D. Bateman, and L. Chen, "A Game Theoretical
Analysis of Data Confidentiality Attacks on Smart-Grid AMI," IEEE J.
on Selected Areas in Communications,, vol. 32, no. 7, pp. 1486–1499,
July 2014.
[10] S. Amin, G. Schwartz, A. Cardenas, and S. Sastry, "Game-Theoretic
Models of Electricity Theft Detection in Smart Utility Networks:
Providing New Capabilities with Advanced Metering Infrastructure,"
IEE J. on Control Systems, vol. 35, no. 1, pp. 66–81, Feb 2015.
[11] Y. Yan, R. Hu, S. Das, H. Sharif, and Y. Qian, "An efficient security
protocol for advanced metering infrastructure in smart grid," IEEE J.
on Network, vol. 27, no. 4, pp. 64–71, July 2013.
[12] M. Ali, E. Al-Shaer, and Q. Duan, "Randomizing AMI configuration
for proactive defense in smart grid," in IEEE Int. Conf. on Smart Grid
Communications (SmartGridComm), Oct 2013, pp. 618–623.
[13] C. Efthymiou and G. Kalogridis, "Smart Grid Privacy via Anonymiza-
tion of Smart Metering Data," in Proc. First IEEE Int. Conf. on Smart
Grid Communications (SmartGridComm), Oct 2010, pp. 238–243.
[14] N. Yukun, T. Xiaobin, C. Shi, W. haifeng, Y. Kai, and B. Zhiyong, "A
security privacy protection scheme for data collection of smart meters
based on homomorphic encryption," in Proc. IEEE conf. on EUROCON,
July 2013, pp. 1401–1405.
[15] H. So, S. Kwok, E. Lam, and K.-S. Lui, "Zero-Configuration Identity-
Based Signcryption Scheme for Smart Grid," in Proc. First IEEE Int.
Conf. on Smart Grid Communications (SmartGridComm), Oct 2010,
pp. 321–326.
[16] Y. Yan, Y. Qian, and H. Sharif, "A secure and reliable in-network col-
laborative communication scheme for advanced metering infrastructure
in smart grid," in Proc. IEEE Int. Conf. on Wireless Communications
and Networking Conference (WCNC), March 2011, pp. 909–914.
[17] R. Bhatia and V. Bodade, "Defining the framework for wireless-
AMI security in smart grid," in Proc. Int. conf. on Green Computing
Communication and Electrical Engineering (ICGCCEE), March 2014,
pp. 1–5.
8
Fig. 9: Comparison of OCSVM and robust covariance estimator for outliers (abnormal data) detection when inlier mode
well-centered and elliptic.
Fig. 10: Comparison of OCSVM and robust covariance estimator for outliers (abnormal data) detection when inlier distribution
is bimodal.
Fig. 11: Comparison of OCSVM and robust covariance estimator for outliers (abnormal data) detection when inlier distribution
is strongly non Gaussian.
[18] M. Thomas, I. Ali, and N. Gupta, "A secure way of exchanging the
secret keys in advanced metering infrastructure," in Proc. IEEE Int.
Conf. on Power System Technology (POWERCON), Oct 2012, pp. 1–7.
[19] P.-H. Hsu, W. Tang, C. Tsai, and B.-C. Cheng, "Two-Layer Security
Scheme for AMI System in Taiwan," in Proc. Ninth IEEE Int. Symp.
on Parallel and Distributed Processing with Applications Workshops
(ISPAW), May 2011, pp. 105–110.
[20] V. N. Vapnik, Statistical learning theory, 1st ed. Wiley, Sep. 1998.
[21] R. Zhang, S. Zhang, S. Muthuraman, and J. Jiang, "One Class Support
Vector Machine for Anomaly Detection in the Communication Network
Performance Data," in Proceedings of the 5th Conference on Applied
Electromagnetics, Wireless and Optical Communications, ser. ELEC-
TROSCIENCE'07, 2007, pp. 31–37.
J. Markkula and J. Haapola, "LTE and hybrid sensor-LTE network
performances in smart grid demand response scenarios," in Proc. IEEE
Conf. on Smart Grid Communications (SmartGridComm), Oct 2013,
pp. 187–192.
[22]
|
1812.05296 | 1 | 1812 | 2018-12-13T07:31:54 | Aerial Robot Model based design and verification of the single and multi-agent inspection application development | [
"cs.MA"
] | In recent decade, potential application of Unmanned Aerial Vehicles (UAV) has enabled replacement of various operations in hard-to-access areas, such as, inspection, surveillance or search and rescue applications in challenging and complex environments. Furthermore, aerial robotics application with multi-agent systems are anticipated to further extend its potential. However, one of the major difficulties in aerial robotics applications is the testing of the elaborated system within safety concerns, especially when multiple agents are simultaneously applied. Thus, virtual prototyping and simulation-based development can serve in development, assessment and improvement of the aerial robot applications. In this research, two examples of the specific applications are highlighted, harbor structure and facilities inspection with UAV, and development of autonomous positioning of multi-UAVs communication relaying system. In this research, virtual prototype was designed and further simulated in multi-body simulation (MBS) feigning the sensing and actuating equipment behaviors. Simultaneous simulation of the control and application system running with software in the loop (SITL) method is utilized to assess the designed hardware behavior with modular application nodes running in Robot Operating System. Furthermore, prepared simulation environment is assessed with multi-agent system, proposed in previous research with autonomous position control of communication relaying system. Application of the virtual prototype's simulation environment enables further examination of the proposed system within comparison degree with postfield tests. The research aims to contribute through case assessment of the design process to safer, time and cost-efficient development and application design in the field of aerial robotics. | cs.MA | cs | Aerial Robot Model based design and verification
of the single and multi-agent inspection application development
Seiko P. Yamaguchi1a,2, Masaru Sakuma1a, Takaki Ueno1a, Filip Karolonek 2
Tadeusz Uhl 2, Ankit A. Ravankar1b, Takanori Emaru1b, Yukinori Kobayashi1b*
1 Laboratory of Robotics & Dynamics, Div. of Human Mechanical Systems and Design
a Graduate School / b Faculty of Engineering, Hokkaido University, Sapporo, Japan
2 Department of Robotics and Mechatronics, Faculty of Mechanical Engineering and Robotics,
AGH University of Science and Technology, Cracow, Poland
[email protected], *[email protected]
Abstract
In recent decade, potential application of Unmanned Aerial Vehicles (UAV) has enabled replacement of various
operations in hard-to-access areas, such as: inspection, surveillance or search and rescue applications in
challenging and complex environments. Furthermore, aerial robotics application with multi-agent systems are
anticipated to further extend its potential. However, one of the major difficulties in aerial robotics applications
is the testing of the elaborated system within safety concerns, especially when multiple agents are simultaneously
applied. Thus, virtual prototyping and simulation-based development can serve in development, assessment and
improvement of the aerial robot applications. In this research, two examples of the specific applications are
highlighted: harbor structure and facilities inspection with UAV, and development of autonomous positioning of
multi-UAVs communication relaying system. In this research, virtual prototype was designed and further
simulated in multi-body simulation (MBS) feigning the sensing and actuating equipment behaviors.
Simultaneous simulation of the control and application system running with software in the loop (SITL) method
is utilized to assess the designed hardware behavior with modular application nodes running in Robot Operating
System. Furthermore, prepared simulation environment is assessed with multi-agent system, proposed in
previous research with autonomous position control of communication relaying system. Application of the
virtual prototype's simulation environment enables further examination of the proposed system within
comparison degree with postfield tests. In future work, the proposed simulation environment can be further
applied for motion and vibration control of the UAV applications. The research aims to contribute through case
assessment of the design process to safer, time and cost-efficient development and application design in the field
of aerial robotics.
Keywords : Aerial Robotics, UAV, Mechatronic Design, Multi-UAV System, System Integration, FANET, MBS ,ROS
1. Introduction
Evolution and development of UAVs have enabled their utilization in variety of robotic operations [1] [2]. Although,
enormous potential of aerial robots exists, high concern to safety and regulatory restrictions is a major bottleneck during
system design and development phase. Hence, the article presents the design procedure with mechatronic approach,
verified and confirmed to be effective and time-efficient for innovative product and system design [3].Virtual prototyping
and verification with simulation tools decreases the risk of damage while failures. In this research, two cases of aerial
robots anticipated applications are concerned, presenting the system and software design process with mechatronic
model-based approach of: hard-to-access area structure inspections, and multi-UAV communication relaying [4].
2. Problem Formulation
As one of the challenges considered, operation range of aerial platforms are limited due to its wireless communication.
Increasing operation range of UAV missions can be achieved by multi-agent communication relaying, anticipated to serve
in better endurance of the system applications [5] (Fig.1). However, multi-UAV system design where several agents are
simultaneously applied in air faces a high risk of damage during failures. Hence, virtual-prototype and simulation-based
development should decrease design time and improve its reliability before the hardware laboratory model is tested.
Preprint Submitted to: The 14th International Conference on Motion and Vibration Control (MoViC 2018)
1
Application requirements considering monitoring of port facilities
including pier structure and breakwater monitoring had been revised (Fig.2).
Challenges of UAV utilization in this area arise due to several reasons; Wide
area coverage and long-range flight required for a single mission, hard-to-
access environment as well as possibility of GPS-denied missions. The large
aspect of the wide requirements needs reflection in application development.
3. Virtual Prototyping and Simulation based Design
Aerial robotic system includes UAV platform, sensing and actuating
components, flight controller and computational responsibility. Fig.3(i)
presents the set-up of the aerial robot. Virtual prototype of the UAV's rigid
body model was developed with CAD and converted to simulation supported
format to be applied for multi-body simulations. Thus, integrating the
simulation model with flight simulator and software operating system, one
can imitate the behavior of the aerial robot within SITL simulation, as in
Fig.3(ii). In combination, HIL setup is enabled prior to field tests.
4. Implementation and Verification
Model based designing process was applied to following application
design and development with proposed methods to highlight its strength.
4.1 Multi-UAV Communication Relaying
System structure as well as autonomous positioning system for formation
control [4] was developed and verified. Fig.4(i) presents dynamic MBS
simulation interfacing with system developed, Fig.4(ii) formation control
monitoring with SITL simulation. It is remarkable that during the software
design of the fail-safe collision avoidance between UAVs, proposed system
enabled to verify algorithm developed with no hardware damaged due to
SITL tests proceeded within software development stage.
4.2 Harbor Inspection Scenarios
In order to test capabilities in challenging environments, application
development for data acquisition, localization and navigation methods were
considered. Modular experimental possibilities of
the simulation
environment accelerated the choice of components and verification for
application building. For example, 3D point-cloud data accusation and
processing presented in Fig.5, or SLAM [6]and navigation techniques [7] for
GPS-denied environment [8] [9].
5. Conclusion and Future Work
Figure 1: Operation principle of the Multi-UAV
communication relaying and its system design through
model based simulations, followed by hardware tests.
Figure 2: Multi-copter requirements on port facilities
inspection. System design process should concern wide
requirements characteristics within its application.
Figure 3: System diagram comparison of aerial robot
components and model-based simulation environment
Figure 4: Multi-UAV Communication relaying system
design and verification with simulation environment,
The applied method has presented its effectiveness in acceleration of
design process and prototyping, as well as increased reliability in safety.
Although, further verification of the simulation accuracy is required,
communication relaying had been developed with this method. Thus,
practical challenges highlighted during actual hardware tests are brought
over to improvements with model-based simulations, including necessary
improvement in simulations. Harbor inspection application is still in the
development phase. Advantages of simulation-based development was
highlighted due to the access difficulties to aimed structures. The research
plans to further verify the solution developed with testbed implementations.
6. References
[1] N. Eduardo, A. A. Ravankar, A. Ravankar, Y. Kobayashi and T. Emaru, "Vision based autonomous docking of
VTOL UAV using a mobile robot manipulator," in IEEE/SICE International Symposium on System Integration
Figure 5: Simulation of the UAV-based breakwater 3D
point-cloud acquisition with 2D lidar and flight data.
Preprint Submitted to: The 14th International Conference on Motion and Vibration Control (MoViC 2018)
2
(SII), Taipei, 2017.
[2] E. Feron and E. N. Johnson, "Aerial robotics," in Springer Handbook of Robotics, Berlin Heidelberg, Springer,
2008, pp. 1009-1029.
[3] A.Martowicz, M.Ciszewski, T.Buratowski, A.Gallina, M.Rosiek, K.Seweryn, W.Teper, A.J.Zwierzyński and T.Uhl,
"Mechatronic approach in application to solution of research and design problems," Mechatronics, vol. 36, pp. 1-
17, 2016.
[4] S. P. Yamaguchi, F. Karolonek, T. Emaru, Y. Kobayashi and T. Uhl, "Autonomous position control of multi-
unmanned aerial vehicle network designed for long range wireless data transmission," in IEEE/ SICE, SII 2017,
Taipei, 2017.
[5] W. Akihisa, Y. Toshiaki, M. Masaaki, A. Takanari, A. Hideo and T. Hirokazu, "A Surveillance System Using Small
Unmanned Aerial Vehicle (UAV) Related Technologies," NEC Technical Journal, Vols. 8-1, no. Special Issue on
Solving Social Issues, pp. 68-72, 2016.
[6] A. Ravankar, A. A. Ravankar, Y. Hoshino, E. Takanori and Y. Kobayashi, "On a hopping-points svd and hough
transform-based line detection algorithm for robot localization and mapping," International Journal of Advanced
Robotic Systems, vol. 13, no. 3, p. 98, 2016.
[7] A. Ravankar, A. A. Ravankar, Y. Kobayashi, Y. Hoshino and C.-C. Peng, "Path smoothing techniques in robot
navigation: State-of-the-art, current and future challenges," Sensors, vol. 18, no. 9, p. 3170, 2018.
[8] T. Yamaguchi, T. Emaru, Y. Kobayashi and A. A. Ravankar, "3D map-building from RGB-D data considering noise
characteristics of Kinect," in IEEE/SICE International Symposium on System Integration (SII), Sapporo, 2016.
[9] M. Sakuma, Y. Kobayashi, T. Emaru and A. A. Ravankar, "Mapping of pier substructure using UAV," in IEEE/SICE
International Symposium on System Integration (SII), Sapporo, Japan, 2016.
[10] M. Sakuma, "Development of Port Facilities Inspection Robot Using Multi-Copter," Graduate School of
Engineering, Hokkaido University, Sapporo, Japan, 2018.
[11] OSRF, "Robot Operating System," Open Source Robotics Foundation, [Online]. Available: http://www.ros.org/.
[Accessed July 2018]
Preprint Submitted to: The 14th International Conference on Motion and Vibration Control (MoViC 2018)
3
|
cs/0407024 | 1 | 0407 | 2004-07-10T11:06:57 | An agent-based intelligent environmental monitoring system | [
"cs.MA",
"cs.CE"
] | Fairly rapid environmental changes call for continuous surveillance and on-line decision making. There are two main areas where IT technologies can be valuable. In this paper we present a multi-agent system for monitoring and assessing air-quality attributes, which uses data coming from a meteorological station. A community of software agents is assigned to monitor and validate measurements coming from several sensors, to assess air-quality, and, finally, to fire alarms to appropriate recipients, when needed. Data mining techniques have been used for adding data-driven, customized intelligence into agents. The architecture of the developed system, its domain ontology, and typical agent interactions are presented. Finally, the deployment of a real-world test case is demonstrated. | cs.MA | cs | In Management of Environmental Quality, An International Journal
vol.15, no.3, pp.238-249, 2004. Emerald Publishers.
An agent -based intelligent environmental monitoring system
Ioannis N. Athanasiadis and Pericles A. Mitkas
ABSTRACT
Fairly rapid environmental changes call for continuous surveillance and on- line
decision- making. Two areas where IT technologies can be valuable. In this paper we
present a multi-agent system for monitoring and assessing air-quality attributes, which
uses data coming from a meteorological station. A community of software agents is
assigned to monitor and validate measurements coming from several sensors, to
assess air -quality, and, finally, to fire alarms to appropriate recipients, when needed.
Data mining techniques have been used for adding data-driven, customized
intelligence into agents. The architecture of the developed system, its domain
ontology, and typical agent interactions are presented. Finally, the deployment of a
real-world test case is demonstrated.
Keywords : Multi-Agent Systems, Intelligent Applications, Data Mining, Inductive
Agents, Air-Quality Monitoring
Introduction
Enviro nmental Information Systems (EIS) is a generic term that describes the class of
systems that perform one or more of the following tasks: environmental monitoring,
data storage and access, disaster description and response, environmental reporting,
planning and simulation, modeling and decision- making. As the requirements for
accurate and timely information in these systems are increasing, the need for
incorporating advanced, intelligent features in EIS is revealed. In this context
advances in Information Te chnology (IT) sector are promising to satisfy these
requirements. Enviromatics (an abbreviation of the term (cid:147)environmental informatics(cid:148))
is the research initiative examining the application of Information Technology in
environmental research, monitoring, assessment, management, and policy (IFIP
1999).
The deployment of modern, intelligent systems for monitoring and evaluating
meteorological data series and assessing air quality is only one of several
enviromatics applications. Our focus in this work is to take advantage of powerful
tools for software development and demonstrate their application for monitoring air-
quality attributes.
Air Quality Operational Centers (AQOC) have been established worldwide in areas
with potential air pollution problems. Their responsibility is to monitor critical
atmospheric variables and publish regularly their analysis results. Currently, real- time
decisions are made by human experts, whereas mathematical models are used for off-
line study and understanding of the atmospher ic phenomena involved. For example, in
the London Air Quality Network (http://www.erg.kcl.ac.uk/london/) flexible data
analysis is supported through statistical tools and in the Texas Natural Resource
Conservation Commission (http://www.tnrcc.state.tx.us), meteorologists set
criteria for making predictions.
the
Several Environmental Monitoring Information Systems (EMIS) have been developed
worldwide for assisting environmentalists in their tasks. These systems vary from
platforms for integrating heterogeneous data sources, to real- time monitoring systems.
For example, the New Zealand Distributed Information System (NZDIS) is an agent -
based architecture for integrating environmental information systems (Purvis et al.
2003). Similarly, the InfoSleuth project used agents for querying distributed
environmental data-clusters in a transparent way (Nodine et al. 2000) and Java Agents
for Monitoring and Management use a collection of software agents that can perform
various administrative tasks in a monitoring network (Tierney et al. 2000). Other
approaches towards EMIS include the Knowledge -based Environmental Data
Analysis Assistant project , which provides an interface between the user and off- the-
shelf analysis packages for off- line study (Fine 1998) and DNEMO, a multi-agent
system for managing air quality in real- time (Kalapanidas & Avouris 2002). These
systems incorporate domain knowledge and their reasoning capabilities are deployed
in the form of expert systems, case-based reasoning, or knowledge bases.
The system described in this paper, adapts a different approach since it employs data
mining techniques for adding customized intelligence into an EMIS. The system,
2
called O3RTAA, is developed as a multi- agent system. Several software agents co(cid:150)
operate in a distributed agent society, in order to monitor both meteorological and air (cid:150)
pollutants attributes in an effort to evaluate air quality and, ultimately, to trigger
alarms.
O3RTAA relies on the Agent Paradigm for building intelligent software applications,
while taking advantage of Machine Learning algorithms and Data Mining
methodologies for extracting knowledge. The implemented system is equipped with
powerful, advanced features, such as measurement validation, missing measurement
estimation and custom alarm identification, while it also performs more rudimentary
tasks, such as data monitoring, storage, and access.
The approach
The system developed, improves existing environmental monitoring systems by
adding customized intelligence into their modules. Our approach is to couple the
software agent paradigm with the knowledge discovery roadmap, in order to provide
Intelligent Environmental Software Applications. This approach has been enabled by
the use of Agent Academy, an integrated platform for creating intell igent agents, with
training and retraining capabilities. (Agent Academy Consortium 2000 )
According to the software agent paradigm (Jennings et al. 1998), agents in our system
are autonomous problem- solvers that co-operate to achieve the overall goals of the
system. Thus, certain system goals are assigned to specific agents. Advanced tasks,
such as incoming measurement validation or custom alarm identification, are
performed with the use of decision models discovered using data mining algorithms.
In general, the knowledge discovery roadmap is used for identifying useful patterns in
vast volumes of data. In the case of environmental systems, there is certainly an
abundance of data. In our approach, interesting patterns hidden in environmental data
sets are discovered and subsequently embedded into our system, essentially adapting
the problem- solving method to local conditions. As a consequence, decision- making
is performed more accurately since the characteristics of the problem at hand may
differ from general trends or (cid:145)rules of thumb(cid:146).
3
In the following sections the approach for developing an agent-based environmental
monitoring system is described. The system(cid:146)s functionality and architecture are
detailed and deployment steps are outlined.
Functional Description
The overall goal of O3RTAA is to monitor air -quality data, including pollutants(cid:146)
distributions, measured by a single meteorological station. Currently, the station(cid:146)s
field sensors measure these attributes and human experts are responsible for assess ing
these measurements and identifying possible alarms. Our system intervenes between
the sensors and the experts and undertakes several tasks in order to assist humans in
their evaluation. Besides the usual housekeeping tasks, such as the updating of the
database with incoming measurements, O3RTAA is empowered with advanced
features including measurement validation, estimation of missing values, and custom
alarm identification. These advanced features are enabled through the exploitation of
data mining tec hniques.
- The system operates between the field sensors and human experts in order to
achieve specific goals, organized in three layers, as shown in Figure 1. The system
goals form a pyramid, as the success of higher- level goals depends on the
fulfillment o f lower- level ones.
4
System Goals
Agent Roles
Distribution Layer
Alarm distribution
User profiling
Management Layer
Data analysis & storing
Alarm identification
D
M
Distribution Agents
Alarm distribution to appropriate
users based on their profiles
Alarm Agents
Formal alarm triggering
Custom alarm identification
Database Agents
Database updates
Contribution Layer
Measurement caption
Data conditioning-verification
C Diagnosis Agents
Measurement validation
Estimation of missing or
erroneous values
Take in Figure 1
Caption: System goals and corresponding agent roles.
At the first layer, the Contribution Layer, system goals related to preprocessing
activities are gathered. Preprocessing includes efficient measurement caption,
incoming data- series verification, possible malfunctions identification and restoration,
and, finally, delivery of preprocessed measurement values. These tasks are assigned to
the diagnosis agent role.
The second layer of goals, the Management Layer, is responsible for the manipulation
of the preprocessed measurement values arriving from the contribution layer. Its
major objective is the identification of alarms. There are two types of alarms: a) the
(cid:147)formal alarms(cid:148), regulated by law- imposed thresholds, and b) the (cid:147)custom alarms(cid:148),
defined by user requirements and taking under account local phenomena and trends.
Alarm agent and database agent roles deliver these tasks in close collaboration.
At the top level, the Distribution Layer, goals are related to the distribution of the
identified alarms to the appropriate recipients, based on previously described profiles
5
of the interested parties. The alarms triggered at the management layer are resolved
and finally delivered to the appropriate recipients at the manner they have specified
(i.e., email message, SMS, etc).
System Architecture
The implemented multi-agent system is shown in Figure 2. Agents are organized in
three layers, contribution, management and distribution, following the classification
of system goals discussed in the previous section. Thick arrows below the agent layers
indicate the critical tasks realized through each one of them.
Contribution
Management
Distribution
Sensor 1
Diagnosis Agent 1
Sensor 2
Diagnosis Agent 2
Sensor n
Diagnosis Agent n
Incoming
Measurement
Validation
Missing Values
Estimation
Alarm Agent
Database Agent
Measurements
Database
Custom Alarm
Identification
Distribution Agent
User A
User B
Use r
Profiles
Formal Alarm
Triggering
Alarm
Distribution
Database
Updates
Contribution
Management
Distribution
Take in Figure 2.
Caption: Agent System Architecture and related critical tasks.
6
Information flows from field sensors (on the left in Figure 2) to the users (on the
right) through the three agent layers. Air -quality measurements arrive into the system
from the sensors. Diagnosis Agents capture them, and after the validation process,
deliver them to the Management layer. The Alarm Agent receives the validated
measurements and determines whether a formal or custom alarm must be issued. The
validated measurements are stored into the database for future use by the Database
Agent. Possible alarms are forwarded to the Distribut ion Agent, which delivers them
in the appropriate format to the corresponding end - users.
Diagnosis Agent Type
Diagnosis Agents are in charge of monitoring various air quality attributes including
pollutants emissions and meteorological attributes. Each one of the Diagnosis Agent
instances is assigned to monitor a certain field sensor. For the developed system there
are eleven attributes monitored: SO2 (Sulfur dioxide), O3 (Ozone), NO (Nitrogen
oxide), NO2 (Nitrogen dioxide), NOX (Nitrogen oxides), VEL (Wind velocity), DIR
(Wind direction) , TEM (Temperature), HR (Relative humidity), RAD (Radiation),
and PRE (Pressure). Diagnosis agents are also responsible for ensuring the efficient
operation of sensors. In case of a sensor breakdown, diagnosis agents are in charge of
estimating the missing values.
7
Process
message
Reply with
current value
Yes
Measurement
request
No
Reply: (cid:147)Not
understood(cid:148)
No
Incoming
measurement
Yes
Incoming
Measurement
Validation
Valid
Measurement
Yes
Sensor
breakdown
check
Semantic
Representation
Missing Value
Estimation
Send
measurement
to Alarm Agent
Take in Figure 3
Caption: Diagnosis Agent type internal structure. (Reasoning Engines are in bold)
The internal structure of the Diagnosis Agent type is shown in Figure 3. When a
message is received by a diagnosis agent it is checked. If it is a measurement request
by another agent, the diagnosis agent will respond by sending the value of its current
8
measurement. If a new measurement has arrived, it is checked for its validity. In any
other case, the incoming messa ge is not understood and the agent responds with an
appropriate message. The incoming measurement validation (IMV) procedure
involves the application of the corresponding inference engine. The IMV engine runs
a decision model extracted with the use of data mining on historical data. The
outcome of this decision model is the classification of the incoming measurement
either as (cid:145)valid(cid:146) or (cid:145)invalid(cid:146).
In the first case, the valid measurement is preprocessed in a semantic way. Qualitative
interpretations of the real value are calculated, such as value range, trend, or
persistence. The real measurement along with the qualitative attributes is sent to the
A larm Agent.
When the measurement is classified as invalid, the Missing Value Estimation (MVE)
rule engine is activated. This rule engine incorporates an inductive decision strategy
discovered also by performing data mining on historical data. The MVE engine
estimates the missing value level in a qualitative form (i.e. (cid:145)low value(cid:146)). The
qualitative indication of the missing measurement is forwarded to the Alarm Agent.
When a Diagnosis Agent identifies that the sensor(cid:146)s measurements are persistently
invalid, triggers a sensor malfunction message to the alarm agent.
The Diagnosis Agent behavior implements JADE(cid:146)s cyclic behavior. This is indicated
in Figure 3 with the dotted arrow connecting the end state to the starting node.
Alarm Agent Type
Alarm Agents are responsible for triggering formal alarms and custom alarms. Formal
alarms are the ones imposed by law, indicating dangerous situations in the atmosphere
exceeding legal thresholds. Custom alarms are alerts for the system users about
situations of their concern. In the present implementation custom alarms warn
environmentalists for events based on their scientific interest related to ozone
pollution. An expanded version of the system may include other type of custom
alarms for patients, public administration, or industry, among others.
9
Process
message
Reply: (cid:147)Not
understood(cid:148)
No
Measurement
from sensor
Yes
No
Add to tuple
Tuple
completed
Yes
Send tuple to
Database Agent
Formal Alarm
Triggering
Custom Alarm
Identification
Alarm exists
Yes
Send Alarm(s) to
Distribution Agents
Take in Figure 4
Caption: Alarm Agent type internal structure. (Reasoning Engines are in bold)
No
10
The Alarm Agent workflow is shown in Figure 4. Alarm Agent gathers the
measurements coming from field sensors through the Contribution Layer.As the
Alarm Agent receives the validated measurements from the Diagnosis Agent, a tuple
(record) of all concurrent measurements is created. When the tuple is completed, the
Alarm Agent performs three parallel activities:
- It sends the measurements to the Database Agent in order for them to be stored to
the Database.
- It may trigger (cid:145)formal alarms (cid:146). This activity involves the application of an
inference engine for Formal Alarm Triggering (FAT Engine).
- It may identify and subsequently issue a custom alarm through the Identification
of Custom Alarms (ICA) engine. The ICA rule engine implements a dec ision model
extracted using data mining techniques.
Whenever Formal or Custom alarms are identified, appropriate messages are
forwarded to the Distribution Agent. Note that the behavior of the Alarm Agent is also
cyclic.
Database Agent Type
The Database Agent is responsible for updating environmental databases with data
from field sensors. This task, although trivial, is vital for the system performance, as it
relieves humans from manipulating all this amount of information.
The Database Agent receives a message from the Alarm Agent, containing the
measurement tuple. It establishes a connection to the related databases and stores all
information in the appropriate format to the corresponding table.
Distribution Agent Type
The Distribution Agent pushes the alarms raised by the alarm agent to the appropriate
users. As an alarm message is received, the Distribution Agent queries user profiles
for selecting target users interested in the alarm and selects the appropriate medium of
notification (i.e. email or SMS). As the set of recipients and mediums has been
prespecified, the alarm is transformed properly into an alert and finally transmitted to
the users.
11
System Deployment
The aforementioned agent types have been developed using the Agent Academy
platform following a certain procedure discussed in Mitkas et al. 2003. This procedure
involves the definition of agent type functions, domain ontology, along with agent
interactions, and communication language.
Agent Academy
Agent Academy is a framework for developing agents empowered with training
capabilities exploiting data mining techniques and is distributed under the less-GPL
license. (Agent Academy Consortium 2000, Mitkas et al. 2002). Agent Academy
facilitates the whole procedure for developing a multi-agent community. Agent
Academy supports the creation of agents with limited initial referenc ing capabilities,
along with a certain agent training process targeting to augment agent intelligence
efficiently, through the embedding of data(cid:150)driven decision strategie s into them
(Athanasiadis et al. 2003b).
Technologies adopted
The development of O3RTAA involved a variety of technologies. The JADE platform
was used for agent creation (Bellifemine et al. 2000) and JESS engine for rule
execution (Friedman-Hill 2003). Ont ology design and specification was done with
ProtØgØ 2000 Ontology Editor (Grosso et al. 1999). Data mining experiments were
performed using the core of the WEKA tool for knowledge analysis (Witten & Frank
1999) and PMML format was selected for knowledge model representation (Data
Mining Group 2001). FIPA standards were adopted for agent communication (FIPA
2000).
Information flow among agents is structured in agent messages, while rule engines
support data-driven decision- making.
Agent Ontology and Messaging
In our system, the content of agent messages is structured using a common ontology,
created with the ProtØgØ 2000 ontology editor. Part of the developed ontology is
shown in Figure 5, containing the main concepts of the system. The slots of the
12
various concepts have been configured in order to contain the appropriate information
communicated by the agents. For example, concept (cid:145)Pollutants(cid:146) has the following
slots: value, level, and variability for describing a measurement in both
quantitative and qualitative form.
Meteorological
Station
Alarms
Measurements
Station
Info
Formal
Custom
Pollutants
T ime
Meteorological
CO
NO2
SO2
NO
O3
NOX
Temperature
Pressure
Wind
Relative
Humidity
Speed
Direction
Take in Figure 5
Caption: Main concept structure as a part of the domain ontology.
Information shifts between agent layers via messages communicated among agent
instances. The ontology is exported in RDFS format, which is compatible with JADE
agents. The agent communication language is FIPA-ACL and a sample message
between a Diagnosis Agent and the Alarm agent is shown in Figure 6. Using the
implemented ontology, the ozone Diagnosis Agent reports its measurement to the
alarm agent, as needed. Agent predicate concept (cid:145)sendMeasurement(cid:146) is used for
informing the Alarm Agent about a measurement and has two slots: one is instance of
the (cid:145)TimeStamp(cid:146) concept and the other an instance of the (cid:145)O3(cid:146) concept. Note that
O3 concept inherits its slots from t he Pollutants concept.
In a similar way, all information is exchanged between agents, conforming to FIPA
specifications and realizing the developed domain ontology.
13
Contribution
Diagnosis Agent
Content:
((sendMeasurement
(TimeStamp :date 01/06/2001
:time 08:30)
(O3 :value 53.00
:level low
:variability increasing)
))
Management
Alarm Agent
Take in Figure 6
Caption: Ozone Diagnosis Agent reports
measurement.
Agent Decision Making Models
to
the Alarm Agent
the current
Agent decision- making is the most critical part of the system as it is related with all
advanced features of the system. O3RTAA contains four different reasoning engines.
IVM, MVE and ICA Engines realize inductive decision models and FAT Engine
realizes a deductive decision model. Inductive decision models are those created using
data mining techniques and incorporating data-driven knowledge into the system.
In general, the goal of the data mining procedure is to reveal interesting patterns from
data volumes and use them for future decision making. A predictive model is built
based on these patterns . Predictive models are comprised of two parts: the predictor
and the response. The predictor consists of a set of indepe ndent attributes used to
make the prediction, while the response is the target value or class. Predictive models
could be in the form of decision trees, neural networks or association rules. A
predictive model can be transformed into logical rules having the form: (cid:145)If
assumption then consequence(cid:146) and thus is easy to implement and execute through a
Rule Engine.
To develop our system, we have used data mining techniques for adding customized
intelligence in the form of data-driven decision strategies. For the knowledge
extraction process, Agent Academy(cid:146)s Data Miner component was used, which
extends the Waikato Environment for Knowledge Analysis (WEKA) tool. Predictive
models extracted with the Agent Academy Data Miner are formed using Predictive
Model Markup Language (PMML). PMML documents are transformed into JESS
rules and finally incorporated into software agents.
14
The test case deployment
Historical data coming from the Onda meteorological station, in the Community of
Valencia, Spain, were used for creating data-driven rules. More specifically, several
meteorological attributes and air -pollutant values, along with validation tags, were
recorded on a quarter- hourly basis during years 2000 and 2001. There are about
70,000 records in the volume. Raw measurements have been preprocessed properly in
order to create training and validation datasets compliant with agent goals and suitable
for data mining.
As an example, let us focus on the Incoming Measurement Validation process. The
dataset was preprocessed in order to co ntain the corresponding validation tag as the
response attribute and the current value of a specific pollutant along with a set of
previous values and measures as predictor attributes. These measures are shown in
Table 1. The preprocessed dataset was split into two sets. Data recorded in year 2000
were used as the training set and data recorded in 2001 were used as the test set.
Quinlan(cid:146)s C4.5 algorithm for decision tree induction (Quinlan 1993) was applied on
the data. The resulted decision tree contains 15 leaves, i.e. it can be implemented as a
set of fifteen rules. The predictive accuracy of the decision model on the validation set
is 99.71%, which is satisfactory. Our data mining experiments are discussed in detail
in Athanasiadis et al (2003a,b).
O3
O3_30
O3_90
The current ozone value
The ozone value 30 min ago
The ozone value 90 min ago
MinMax60 The difference between the maximum and the minimum ozone value in the last 60
min
MinMax150 The difference between the maximum and the minimum ozone value in the last 150
O3va l
min
The corresponding validation tag (valid/erroneous)
Take in Table 1
Caption: Attributes used for the validation decision model
Summary and future work
In this paper we have presented an intelligent environmental monitoring system,
developed with software agents and using data mining techniques for adding
customized intelligence into them. System goals have been assigned successfully to
15
agents, which act as media tors and deliver validated information to the appropriate
stakeholders. Agent communication and semantic representation of information is
performed using state of the art tools.
System customization has been based on the application of data mining techniques on
historical environmental data for adding data-driven intellige nce. The use of C4.5
algorithm yielded trustworthy decision models for validating incoming measurements,
estimating the erroneous ones and identifying custom alarms in the described
application.
The system developed will be installed as a pilot case at the Mediterra nean Centre for
Environmental Studies Found a tion (CEAM), Valencia, Spain, in collaboration with
IDI-EIKON, Valencia, Spain. Having the system validated for a single meteorological
station, future work will be focused on expanding the architecture for covering the
whole network.
Acknowledgements
Authors would like to express their gratitude to the Agent Academy Consortium for
their valuable help. Special thanks go to the IDI-EIKON team for their efforts within
Agent Academy project to deploy the O3RTAA system and to CEAM for the
provision of the ONDA dataset. The Agent Academy project is partially funded by the
European Commission under the IST programme (IST-2000-31050).
References
Agent Academy Consortium (2000), Agent Academy: A Data Mining Framework for
Developing Intelligent Agents. (Available at: http://AgentAcademy.iti.gr).
Athanasiadis, I. N., Kaburlasos, V. G., Mitkas, P. A. and Petridis, V. (2003),
Applying machine learning techniques on air quality data for real- time decision
support, in (cid:147)Information Technologies in Environmental Engineering(cid:148), ICSC-
NAISO, Canada.
Athanasiadis, I. N., Mitkas, P. A., Laleci, G. B.and Kabak, Y. (2003), Embedding
data-driven decision strategies on software agents: The case of a multi-agent
system for monitoring air-quality indexes, in Jardim- Goncalves, R., Cha J. and
Steiger-Garcao A. (eds), (cid:147)Concurrent Engineering: The Vision for the Future
Generation in Research and Applications(cid:148), Balkema Publishers, Madeira,
Portugal, pp.23-30.
Bellifemine, F., Poggi, A. and Rimassa, G. (2000), Developing multi-agent systems
with JADE,
in (cid:147)Seventh International Workshop on Agent Theories,
Architectures and Languages (ATAL-2000)(cid:148), Boston, MA. (Available at:
http://jade.cselt.it).
Data Mining Group (2002), Predictive Model Mark up Language (PMML),
Specifications, (Available at: http://www.dmg.org).
16
Fine, S. S., Smith, W. T., Thorpe, S. R., Wheeler, N. J. M. and Eyth, A. M. (1998),
An automated assistant for environmental data analysis, in (cid:147)Fourteenth
International Conference on Interactive Information and Processing Systems for
Meteorology, Oceanography and Hydrology(cid:148), American Meteorological
Society, Phoenix, AZ, pp. 38-40.
FIPA (1999), Foundation for Intelligent Physical Agents, Specifications, (Available
at: http://www.fipa.org).
Friedman-Hill, E. J. (2003), Jess, the expert system shell for the java platform, version
6.1,
Sandia
National
Laboratories,
CA.
(Available
at:
http://herzberg.ca.sandia.gov/jess/).
Grosso, W. E. et al. (1999), Knowledge modeling at the millennium, The design and
evolution of Prot_ eg_e-2000. (Available at:http://protege.stanford.edu).
IFIP (1999), Computers and the environment, Working Group 5.11, International
Federation
for
Information
Processing.
(Available
at:
http://www.environmatics.org).
Jennings, N. R., Sycara, K. and Wooldridge, M. J. (1998), A roadmap of agent
research and development, in (cid:147)Autonomous Agents and Multi-Agent Systems
1(cid:148), Kluwer Academic Publishers, Boston, pp. 7-38.
Kalapanidas, E. and Avouris, N. (2002), (cid:147)Air quality management using a multi-agent
system(cid:148), International Journal of Computer Aided Civil and Infrastructure
Engineering 17(2), 119-130.
Mitkas, P., Symeonidis, A., Kehagias, D., Athanasiadis, I. et al. (2002), An agent
framework for dynamic agent retraining: Agent academy, in B. Stanford-Smith,
E. Chiozza and M. Edin, eds, (cid:147)Challenges and Achievements in e-business and
e-work(cid:148), Prague, Czech Republic, pp. 757-764.
Mitkas, P. A., Symeonidis, A. L., Kehagias, D. and Athanasiadis, I. N. (2003), A
framework for constructing multi- agent applications and training intelligent
agents,
in (cid:147)Fourth International Workshop on Agent-oriented Software
Engineering (AOSE-2003). Held at Autonomous Agents and Multi- Agent
Systems (AAMAS 2003)(cid:148), Melbourne, Australia.
Nodine, M., Fowler, J., Ksiezyk, T., Perry, B., Taylor, M. and Unruh, A. (2000),
(cid:147)Active
information gathering
in InfoSleuth(cid:148), International Journal of
Cooperative Information Systems 9(1-2), 3-28.
Purvis, M., Cranefield , S., Ward, R., Nowostawski, M., Carter, D. and Bush, G.
(2003), (cid:147)A multi-agent system for the integration of distributed environmental
information(cid:148), Environmental Modelling and Software 18(6), 565-572.
Tierney, B., Crowley, B. et al. (2000), A monitoring sensor management system for
grid environments, in (cid:147)High Performance Distributed Computing (HPDC -9)(cid:148),
American Meteorological Society, Pittsburg, Pennsylvania.
Witten, I. H. and Frank, E. (1999), Data Mining: Practical Machine Learning Tools
and Techniques with Java Implementations, Morgan Kaufmann Publishers,
USA.
17
|
1909.06537 | 1 | 1909 | 2019-09-14T05:38:50 | Speeding Up Distributed Pseudo-tree Optimization Procedure with Cross Edge Consistency to Solve DCOPs | [
"cs.MA",
"cs.AI"
] | Distributed Pseudo-tree Optimization Procedure (DPOP) is a well-known message passing algorithm that has been used to provide optimal solutions of Distributed Constraint Optimization Problems (DCOPs) -- a framework that is designed to optimize constraints in cooperative multi-agent systems. The traditional DCOP formulation does not consider those constraints that must be satisfied (also known as hard constraints), rather it concentrates only on soft constraints. However, the presence of both types of constraints are observed in a number of applications, such as Distributed Radio Link Frequency Assignment and Distributed Event Scheduling, etc. Although the combination of these types of constraints is recently incorporated in DPOP to solve DCOPs, scalability remains an issue for them as finding an optimal solution is NP-hard. Additionally, in DPOP, the agents are arranged as a DFS pseudo-tree. Recently it has been observed that the constructed pseudo-trees in this way often come to be chain-like and greatly impair the algorithm's performance. To address these issues, we develop an algorithm that speeds up the DPOP algorithm by reducing the size of the messages exchanged and increasing parallelism in the pseudo tree. Our empirical evidence suggests that our approach outperforms the state-of-the-art algorithms by a significant margin. | cs.MA | cs | SPEEDING UP DISTRIBUTED PSEUDO-TREE OPTIMIZATION
PROCEDURE WITH CROSS EDGE CONSISTENCY
TO SOLVE DCOPS
9
1
0
2
p
e
S
4
1
]
A
M
.
s
c
[
1
v
7
3
5
6
0
.
9
0
9
1
:
v
i
X
r
a
Department of Computer Science and Engineering
Department of Computer Science and Engineering
Department of Computer Science and Engineering
Department of Computer Science and Engineering
Mashrur Rashik
University of Dhaka
[email protected]
Md. Mamun-or-Rashid
University of Dhaka
[email protected]
Md. Musfiqur Rahman
University of Dhaka
[email protected]
Md. Mosaddek Khan
University of Dhaka
[email protected]
ABSTRACT
Distributed Pseudo-tree Optimization Procedure (DPOP) is a well-known message passing algorithm
that has been used to provide optimal solutions of Distributed Constraint Optimization Problems
(DCOPs) -- a framework that is designed to optimize constraints in cooperative multi-agent systems.
The traditional DCOP formulation does not consider those constraints that must be satisfied (also
known as hard constraints), rather it concentrates only on soft constraints. However, the presence of
both types of constraints are observed in a number of applications, such as Distributed Radio Link
Frequency Assignment and Distributed Event Scheduling, etc. Although the combination of these
types of constraints is recently incorporated in DPOP to solve DCOPs, scalability remains an issue
for them as finding an optimal solution is NP-hard. Additionally, in DPOP, the agents are arranged
as a DFS pseudo-tree. Recently it has been observed that the constructed pseudo-trees in this way
often come to be chain-like and greatly impair the algorithm's performance. To address these issues,
we develop an algorithm that speeds up the DPOP algorithm by reducing the size of the messages
exchanged and increasing parallelism in the pseudo tree. Our empirical evidence suggests that our
approach outperforms the state-of-the-art algorithms by a significant margin.
1
Introduction
Distributed Constraint Optimization Problems (DCOP) are a framework involving multiple agents that are used to
interact with one another to achieve a common goal [1]. A number of real world problems, such as distributed event
scheduling [2], scheduling smart home devices [3] and allocating tasks in mobile sensor networks [4], can be modelled
with this framework. Specifically, a DCOP consists of a number of distributed cost functions which collectively form
a global objective function (i.e. the common goal). Each of these cost functions represents a constraint relationship
among a set of variables that are controlled by the agents contributing to that constraint. In more detail, each agent is
responsible for setting value(s) of its own variable(s) from a finite domain(s). However, they can communicate with their
neighbouring agents, and thus can influence value assignment of each other. The goal of a DCOP solution approach is
to set every variable to a value from its domain in order to minimize the number of constraint violations or maximize
the global objective function.
Over the last couple of decades, a number of algorithms have been proposed to solve DCOPs, and they have been
primarily classified into two types: incomplete and complete algorithms. The former experiences better computation
and communication cost at the expense of solution quality. Among the incomplete DCOP algorithms DBA [5], DSA
[6] and Max-Sum [7] are the most notable ones. Although it is obvious that this class of algorithms perform well in
terms of computation and communication cost, a good number of applications, such as Wi-Fi Channel Assignment
[8], Reactive Network Resilience [9] and many other besides, cannot afford sacrificing the quality of solution. In
x5
0
0
1
1
x6 Cost
0
12
3
1
7
0
1
3
(a) A constraint graph representation of
a DCOP. Here, the edges having rela-
tional operators are the hard constraints.
(b) A sample cost table for soft constraint involving variable x5
and x6.
Figure 1: Example of a DCOP Instance containing both soft and hard constraints
effect, a number of complete DCOP algorithms have been proposed in the literature, and a lot of efforts can be seen
to improve those algorithms. This class of algorithms can be further classified as search-based and inference-based
algorithms. The former use a search technique to find the optimal solution from a set of possible assignments. Some
of the notable search-based complete algorithms are SyncBB [10], ConcFB [11], ADOPT [12]. On the other hand,
the latter, such as DPOP [13], Action-GDL [14], BrC-DPOP [15], are based on dynamic programming techniques.
Among them, Distributed Pseudo-tree Optimization Procedure (DPOP) has gained particular attention from the DCOP
community. This is due to the fact that DPOP requires a linear number of messages compared to the search-based
complete algorithms.
To date, several DPOP variants have been proposed. Specifically, O-DPOP [16] and MB-DPOP [17] have made
improvements in terms of memory requirements of the original algorithm. Then an extension of DPOP, named SS-
DPOP [18], improves participating agents' privacy. Whereas, a partially centralized version of DPOP (i.e. PC-DPOP)
achieves shorter runtime but sacrifices some privacy [19]. A notable issue with all of the above variants is that they
are not able to handle such constraints that must be satisfied (i.e. hard constraints). In contrast, a soft constraint poses
a profit/loss for each possible value assignment to its corresponding variables. Nonetheless, the hard constraints are
frequently seen in a number of well-known DCOPs, such as distributed Radio Link Frequency Assignment (RLFA)[20]
and distributed event scheduling problem[2]. In the wake of this shortcoming, two notable extensions of DPOP, H-DPOP
[21] and BrC-DPOP [15], have been proposed.
In more detail, H-DPOP reduces the computation cost of DPOP by ruling out infeasible combination of the variables,
and thus generating smaller messages. This is done by a Constraint Decision Diagram (CDD), which graphically
represents a solution set for n-ary constraints [22]. To do so, H-DPOP performs join and projection operations on CDDs
that are computationally expensive. At the same time, it is not possible to fully exploit hard constraints to prune the
domain of a variable using this approach. This particular issue has been addressed by BrC-DPOP through the use of
Value Reachability Matrix (VRM) which is a representation of a constraint between two variables in the form of a
matrix. It is worth noting that similar to the aforementioned DPOP extensions, BrC-DPOP uses depth-first search pseudo
tree to graphically represent a DCOP. Notably, it is shown in [23] that a depth first search pseudo tree often results in
a chain-like structure thus impairing the performance of the algorithm due to the lack of parallelism. Nevertheless,
the algorithm proposed in the paper, the so-called BFS-DPOP, has shown the significance of an alternative graphical
representation− breadth-first search pseudo tree. To be exact, BFS-DPOP enhances parallelism, and thus reduces the
runtime of the algorithm. However, BFS-DPOP cannot handle hard constraints, and thus is not directly applicable to
BrC-DPOP.
Against this background, we propose a new variant of the DPOP algorithm, that we call CeC-DPOP. CeC-DPOP
takes the advantage of parallelism through the use of BFS pseudo tree as the communication structure. It can also
deal with hard constraints. However, unlike BrC-DPOP that enforces branch consistency, CeC-DPOP uses a new form
of consistency, namely Cross-edge Consistency (CeC). This particular phenomenon enables CeC-DPOP to produce
smaller message size and improve DPOP's runtime of by pruning the domain of the corresponding variables. To be
precise, We empirically evaluate the performance of our approach, and observe a significant reduction of runtime,
average of 60% by using this technique.
2
X1X2X3X4X5X6X7Soft>==<SoftSoft 2 Background and Problem Formulation
A DCOP model can be formally expressed as a tuple (cid:104)A, X, D, F, α(cid:105) where:
• A = {a1, a2, ...., ak} is a set of agents.
• X = {x1, x2, ...., xn} is a set of variables, where n≥k.
• D = {d1, d2, ...., dn} is a set of domains for the variables in X, where di ∈ D is the available domain for the
corresponding variable xi ∈ X.
• F = {f1, f2, ...., fm} is a set of constraint functions (also known as utility or cost functions). Here, each
function fi(xi) depends on a subset of variables xi ⊆ X that can be mentioned as the scope of that function.
In order to represent the relationship among the variables in xi, the function fi(xi) denotes the utility value
for each possible assignment of those variables. Each constraint fi ∈ F can be hard in which the value
combinations that must be avoided are denoted as the cost 0 and the combinations that are allowed have the
cost 1. The remaining type is the soft constraint indicating that each value combination results in a finite
utility/cost value and need not to be avoided. The dependencies among the variables can be used to construct a
constraint graph that has been used to represent DCOPs graphically. In this representation, each variable is
associated with a node and connected to each other through an edge.
• α : X → A is an onto mapping function that assigns the variables X to the set of agents A.
(cid:88)
fi∈F
X∗ = arg max
x
fi(xi)
(1)
Within this model, the main objective of a DCOP algorithm can be expressed as each agent assigning the values to its
associated variable(s) from the corresponding domain(s) that can be expressed as X∗, in the pursuit of maximization
or minimization of the sum of the utility functions (i.e. the global objective function). In this paper, we are going to
consider the maximization problem only (Equation 1). For example, in Figure 1a ,a DCOP instance is graphically
represented as a constraint graph. Here, we consider the set of variables X = {x1, x2, ..., x7}, each having domain
di = {0, 1}. The cost matrix of the soft constraint involving variables x5 and x6 is showed in figure 1b. The remaining
constraints in the graph that are defined by relational operators are hard constraints.
As aforementioned, Distributed Pseudo-tree Optimization Procedure is a complete, synchronous message passing
algorithm for solving DCOPs. Specially it uses the dynamic programming technique on a DFS pseudo-tree in a
distributed manner. DPOP is executed through three phases. In the first phase, a distributed DFS traversal is started
from the root(held by an agent) of the constraint graph using the distributed DFS algorithm like in [24]. As a result,
a DFS pseudo-tree structure is built where each agent labels its neighbours as parents, pseudo-parents, children or
pseudo-children and edges are identified as tree or back edges. For example, after this phase, the constraint graph in
Figure 1a will result in a DFS pseudotree like in Figure 2a. The resulting DFS pseudo-tree serves as a communication
structure for the subsequent phases of DPOP. The next phase is the Util propagation phase in which each agent,
starting from the leaves of the constraint graph, sends UTIL message to its parent. The UTIL message is generated by
aggregating the constraint utilities between the current node and the variables in its separator that is the ancestors of
the current node that are connected directly to this node or its descendants and also the utilities in the UTIL message
received from its children and finally projecting out its own variables by optimizing over them. At last, the value
propagation phase is initiated from the root agent. Each agent selects its optimal assignment using the cost function
computed in the UTIL propagation phase and the VALUE message received from its parent. Afterwards, each agent
broadcasts its assignment to its children. When every agent has chosen its optimal assignment, the algorithm terminates.
DPOP can be executed on different branches independently using DFS pseudo-tree as communication structure. Though
DPOP produces linear number of messages as mentioned before, message size in this algorithm is exponential. Another
notable limitation of the DPOP algorithm is that it does not exploit hard constraint along with soft constraints which
has been found useful in many real life DCOP problems. These two situation can be resolved by another algorithm
BrC-DPOP proposed by [15].
BrC-DPOP exploits hard constraints by enforcing arc consistency and introducing a weaker form of the path consistency
which can be applied along the path of a pseudo-tree in pursuit of reducing message size.The algorithm starts with
generating a pseudo-tree structure followed by a path construction Phase which is later used to get the knowledge
of the direct paths from each agent to its parent and pseudo-parents. In the next phase, arc consistency is enforced
in a distributed environment.Then the most important phase is executed where branch consistency is exploited in a
distributed way. The aim of this phase is to ensure mutual reachability of every pair of values of an agent and its
pseudo-parents considering every pseudo-tree path between them. Finally, the UTIL and VALUE propagation phase
are executed considering the updates of the pseudo-tree. The advantages of BrC-DPOP includes smaller message
size due to BrC propagation enforcement as well as faster runtime since it prunes the values of the variables. Though
3
(a) DFS Pseudo-tree
Figure 2: DCOP Pseudo-trees
(b) BFS Pseudo-tree
Algorithm 1 Path-Construction(Gbf s, P, C)
Input: Pseudo tree Gbf s, set of parents P, set of child C.
Output: A list containing path information from current variable to enforce cross-edge consistency.
1: for each cross edge (xi, xj) of Gbf s do
xl ← LCA(xi, xj)
2:
send N EXT _U P DAT E(xl, xi) to P i
3:
4: if xi is not an end point of a cross edge then
5:
6: while cnt_nexti < Ci do
7:
N EXT i ← N EXT i ∪ (xl, xc)
8:
9:
cnt_nexti ← cnt_nexti + 1
10:
11: if N EXT i not equals NULL then
12:
13:
14: send complete(xi) to P i
send N EXT _U P DAT E(N U LL, xi) to P i
if receive N EXT (xl, xc) from xc ∈ Ci then
if receive complete(xc) from xc ∈ Ci then
for each xl such that (xl, xc) ∈ N EXT i do
send N EXT _U P DAT E(xl, xi) to P i
BrC-DPOP improves the DPOP algorithm to a great scale, the communication structure is DFS pseudo-tree, and
as previously mentioned often becomes chain-like in many experiments for example in Figure 2a. This condition
greatly reduces performances of solving DCOPs by Brc-DPOP or other variants of DPOP that use DFS pseudo-tree
as the communication structure. To deal with this drawback, [23] propose a variant of DPOP algorithm (the so-called
BFS-DPOP) which uses the Breadth First Search (BFS) pseudo-tree as the communication structure.
In more detail, BFS-DPOP operates on Breadth First Search (BFS) pseudo-tree that is used as the communication
structure intending to increase parallelism. This is because it produces more branches than that of the DFS counterpart.
Here, Figure 2b depicts the transformed BFS pseudo-tree of the corresponding constraint graph of Figure 1a. In
BFS-DPOP, following the construction of BFS Pseudo-tree, the cluster removal phase occurs wherein the allocation of
cross-edges are decided so that it can reduce the maximal message size as much as possible by the disposal of cross-edge
constraints. Finally, the UTIL and VALUE propagation phase is executed on the BFS Pseudo-tree considering the
changes occurred in the previous phases. Even though BFS-DPOP experiences shorter communication paths, and hence
less communication time, through the use of BFS pseudo-tree, the algorithm produces messages with exponential
size as the system grows, as in traditional DPOP. Moreover, as aforementioned, this algorithm can not deal with hard
constraints which is utilized in BrC-DPOP by enforcing branch consistency. On the other hand, BFS-DPOP is not
suitable for exploiting branch consistency. In light of the above background, we address these issues in the section that
follows.
3 Cross-Edge Consistent DPOP (CeC-DPOP)
CeC-DPOP improves DPOP by enforcing cross-edge consistency that reduces the domain size of the variables of a
DCOP. In effect, it reduces the message size and runtime of the DPOP algorithm. Moreover, unlike the traditional
DPOP algorithms, CeC-DPOP uses BFS pseudo tree instead of a DFS pseudo tree in order to take its inherent benefits
of increased parallelism and shorter tree depth. Specifically, this algorithm comprises of four phases, BFS pseudo tree
construction, consistency enforcement, UTIL propagation and VALUE propagation phase.
4
X1X2X3X4X5X6X7>==<SoftSoftSoftX1X2X3X4X5X6X7Soft>==<SoftSoft (a) A part of pseudo-tree depicted in
Figure 2b, used as an example to enforce
cross-edge consistency
after
pairs
arc-
(b) Consistent
consistency enforcement. Here, an
arrow indicates an allowable pair
Figure 4: Simulation of the algorithm
(c) Consistent pairs along cross-edge af-
ter applying CeC-DPOP which shows
an overall 76% domain size reduction
Initially, a pseudo tree is constructed from the constraint graph. In order to generate the corresponding BFS pseudo tree,
we use the same method as prescribed in the BFS-DPOP algorithm. For example, Figure 2b illustrates a sample BFS
pseudo tree of the constraint graph depicted in Figure 1a. For simplicity, we use a part (Figure 4a) of the pseudo tree of
Figure 2b as the worked example of our algorithm. Having a BFS pseudo tree Gbf s constructed, CeC-DPOP enforces
arc-consistency. This phase uses the distributed Arc-Consistency (AC) algorithm that is introduced in BrC-DPOP
algorithm. This algorithm results in a reduced domain for all the variables having hard constraints, as shown in Figure 4b,
where domain of each variable is {0, 1, 2, 3, 4}.
After arc-consistency is enforced, CeC-DPOP enforces a new form of consistency (i.e. the so-called cross edge
consistency) on the BFS pseudo tree. To do so, we need the lowest common ancestor LCA(xi, xj) for every pair of
variables xi and xj in Gbf s. To find the LCA of every pair of variables, we followed the distributed algorithm shown in
[25]. In order to represent hard constraints we use consistency matrices, where a matrix Mij represents a hard constraint
between variables, xi and xj. The consistency matrix between x1 and x2, which represents the constraint x1 < x2 as
shown in Figure 3.
M12 =
0
1
2
3
4
0
0
0
0
0
0
1
1
0
0
0
0
2
1
1
0
0
0
3
1
1
1
0
0
4
1
1
1
1
0
Figure 3: An example consistency matrix (M12) between x1 and x2 where a 1 along a particular row (ri) and column
(ci) indicates that x1 = ri and x2 = ci is allowable for x1 < x2
Now the algorithm enforces cross edge consistency on the pseudo tree Gbf s. For this, we need to construct a path for
each cross-edge in Gbf s (Algorithm 1). The BFS pseudo tree Gbf s, parent set P and set of child C are the inputs of the
algorithm. Throughout the algorithm, we construct a N EXTi list containing information in the form of (xl, xc). It
informs the current agent xi about the next agent xc to enforce cross-edge consistency for an edge whose endpoints
have LCA at xl. The for loop in line 1 selects a cross edge having one end point xi from Gbf s, and sends a message
N EXT _U P DAT E(xl, xi) to its parent Pi. This message contains information about the LCA xl of two variables xi
and xj and the current variable xi. To do this, line 2 computes a LCA, xl of xi with another variable xj with whom it
holds cross edge. Then in line 3 xi sends a N EXT _U P DAT E(xl, xi) to its parent Pi. In line 4, CeC-DPOP checks
whether xi is a member of a cross edge, and if this is not the case, it sends a N EXT _U P DAT E(N U LL, xi) to
parent Pi. Here, NULL indicates that xi is not an end point of any cross edge. In our exemplary pseudo tree of Figure
4a, x5 computes a LCA (i.e. x1) for its cross edge connecting x5 and x6 and then sends N EXT _U P DAT E(x1, x5)
to its parent x2.
Afterwards, the while loop in line 6 compares a counter variable, cnt_nexti with the child count of current variable
(i.e. Ci) to check whether the current variable received a N EXT _U P DAT E message from each child in Ci.
Within this loop, if a N EXT _U P DAT E(xl, xc) is received from a child, appends (xl, xc) to the list, N EXTi (line
7-8). Then line 9 checks for any complete(xc) received from a child. This message informs the current variable xi
5
X1X2X3X5X6Soft>==<X1 < X20011223344X2 = X50011223344X3 > X60011223344X1 = X30011223344X5 - X60011223344Possible assignments52=25After enforcing cross-edge consistency 6% reduction76% // CeC message propagation initiated by root
Algorithm 2 Cross-Edge-Consistency-Propagation(Gbf s, M, C, N EXT )
Input: Pseudo tree Gbf s, set of consistency matrix M, set of child C, next list N EXT .
Output: A cross edge consistent pseudo tree.
1: if xi is root then
2:
3:
4: if received CeC(xp, M pl) from xp such that xp is parent of xi then // propagating the updated consistency matrix
5:
6:
7:
8:
9:
10:
11:
12: for each xc ∈ Ci such that xc received no CeCi message do
13:
14: for each ceij ∈ CEi such that xl equals LCAij do
15:
16:
send CeC(xi, NULL) to xc
M lj ← M jl
M ij ← M il × M lj
// computing reduced domain along cross-edge
// matrix transpose
if xc not equals NULL then
send CeC(xi, M il) to xc
if xl equals xi then
M il ← M ii
M il ← M ip × M pl
else
// initializing to unary constraint
for each xc in Ci do
send CeC(xi, M ii) to xc
for each (xl, xc) ∈ N EXT i do
T
that path construction for the sub tree rooted at xc is complete. For each received complete(xc) message, line 10
increments cnt_nexti by 1. The while loop terminates when each child xc in Ci sends a complete(xc) message. In
our example, after receiving N EXT _U P DAT E(x1, x5) from x5, x2 appends (x1, x5) to N EXT2. Now, in line 11,
the algorithm checks whether N EXTi is not empty. If this is true, the for loop in line 12 selects each (xl, xc) pair
from the N EXTi list and line 13 sends a N EXT _U P DAT E(xl, xi) message to Pi. Next, the algorithm terminates
after sending a complete(xi) message to Pi after line 14. In this context, in our example after x5 sends a next message
to x2, x5 generates a complete(x5) message and sends it to x2. Upon receiving a complete message from its only
child x2, increments cnt_nexti by 1 and the while loop of line 6 terminates. In our example of Figure 4a, x2 sends
N EXT _U P DAT E(x1, x2) to x1. After receiving this, x1 obtains information about its next child x2 to enforce cross
edge consistency. At the same time, x2 obtains information about x5. Thus, a path from x1 to x5 through the variable
x2 is established. Similarly, another path from x1 to x6 through the variable x3 is established.
Finally, we enforce cross edge consistency on the path that is established on the pseudo tree (Algorithm 2). The BFS
pseudo tree Gbf s, set of consistency matrices M, set of child C and the N EXT list are the inputs of the algorithm. The
algorithm works as follows. Line 1 of the algorithm checks whether the current variable xi is root. If this true, it initiates
CeC message propagation by iterating every child using the for loop in line 2. Line 3 then sends a CeC(xi, Mii) to
every xc in Ci, where xi is the variable which sent the message along with its consistency matrix Mii. Line 4 of the
algorithm checks whether any CeC message has been received from its parent. If this is the case, line 5 iterates over
each pair (xl, xc) of the N EXTi list to propagate CeC message. For this purpose, line 6-7 checks whether the current
variable equals to LCA xl of a cross edge in the subtree. If this is the case, it initializes Mil with its unary constraint Mii
which represents the domain of the current variable xi. Otherwise, line 8-9 computes Mil, which is a multiplication of
Mip and Mpl. Next, line 10 checks whether xc is not null. If this is true, line 11 sends a CeC(xi, Mil) message to xc. In
our example of Figure 4a, x1 initializes CeC message propagation and sends a CeC(x1, M11) to both x2 and x3. When
x2 receives CeC(x1, M11) from x1, it computes M21 by multiplying M21 with M11 and sends this new consistency
matrix to its child x5 through a CeC(x2, M21) message. Now, line 12-13 of the algorithm checks whether any child
exists that did not receive any CeC message. If this is true, xi sends a CeC(xi, N U LL) to that child. Line 14-16 finally
computes the consistency matrices along each cross edge by iterating over every cross edge and multiplying the matrices
obtained for each end point of the cross edge. In our example, after x5 receives a CeC(x2, M21) from x2, it computes
M51. In a similar way x6 computes M61. To obtain the reduced set of assignable pairs, CeC-DPOP multiplies M51
with M16. The output for our simulation is shown in Figure 4c. Here a total number of allowable pairs is 52 = 25. After
cross-edge consistency is enforced, the number of allowable pairs is reduced to 6. Hence, the reduction is 25−6
25 = 0.76,
which is a 76% reduction in the total number of assignable pairs along x5 and x6. After cross edge consistency is
enforced, we obtain a set of variables with reduced domain size. Now, we execute the UTIL and VALUE propagation
phase. These two steps correspond to the UTIL and VALUE propagation phase of BFS-DPOP algorithm.
4 Complexity Analysis
For enforcing cross-edge consistency we have constructed a path and next enforced arc-consistency. The path con-
struction phase needs to send a message containing information about its subtree to its parent for each endpoint of the
6
(a) number of variables vs runtime (random graph)
(b) domain size vs runtime (random graph)
(c) graph density vs runtime (random graph)
(d) number of variables vs runtime (distributed RLFA prob-
lem)
Figure 5: Experimental results for random graphs and the distributed RLFA problem
cross-edge. Therefore, the complexity of this phase is O(CE log(X)), where CE is the number of cross-edge and
X is the number of variables. The path construction phase requires the lowest common ancestor for each pair of nodes,
which is found in a preprocessing phase having a complexity of O(log(X)). The next phase is the arc-consistency
enforcement phase. In this phase, each hard constraints are evaluated to check whether the domain of both variables
connecting the endpoints is consistent with each other. Given the number of hard constraints is CH and the average
domain of each variable is d, the complexity of this phase is O(CH d3). Here, in order to check whether each value
in the domain of an endpoint is consistent with every value of the other endpoint, it requires three nested loops thus
resulting in d3 computations. The final phase is enforcing cross-edge consistency. In this phase, each agent waits for
its parent agent to send a CeC message which it uses to find the final cross-edge consistent matrix which requires a
complexity of O(d3) which is required for the multiplication of two matrices. Apart from regular matrix multiplication,
entry-wise matrix multiplication has a complexity of O(d2). The process continues for each variable, and as such the
total complexity of cross-edge consistency enforcement phase is O(X(d3 + d2)).
The arc-consistency enforcement phase requires O(dX) messages, where the size of each message is O(d). In each
step of arc-consistency enforcement, the domain information of a variable is only required to be propagated. Therefore,
the cross-edge consistency enforcement phase requires CH messages and the size of each message is O(d2). In this
phase, we only propagate CeC messages which contain the consistency matrices and the size of a message depends on
the size of these matrices.
5 Experimental Results
We now empirically evaluate how much performance improvement can be attained using CeC-DPOP in comparison
to the original DPOP algorithm and two important variants of DPOP named BFS-DPOP and BrC-DPOP. Unlike
CeC-DPOP, the original DPOP uses DFS pseudo-tree as the communication structure and do not actively exploit hard
constraints. Therefore, it is reasonable to observe the attributes of CeC-DPOP (i.e. inclusion of soft constraints along
7
with hard constraints and use of BFS pseudo-tree as the communication structure) with respect to the original DPOP.
Additionally, We consider with BFS-DPOP algorithm as a benchmark because it also uses BFS pseudo-tree as the
communication structure. Finally, we compare CeC-DPOP with BrC-DPOP as both the algorithms can deal with DCOPs
having both type of constraints. Note that another DPOP variants H-DPOP has not been considered as a benchmark
because it already shown in the work of BrC-DPOP that it is outperformed for its high runtime. To benchmark the
runtime of our algorithm CeC-DPOP as well as the benchmarks, we run our experiments on two different types of
DCOP settings: random constraint graph and the distributed RLFA problems. These choices are made following the
experimental settings of BrC-DPOP.
In case of random DCOPs, the runtime of the algorithms have been reported varying three parameters: number of
variables, their domain size and graph density (i.e. the ratio of the constraint number and n ∗ (n − 1)/2 where n is
the number of varaibles). For the first parameter, we vary the number of nodes from 5 to 30 in Figure 5a setting the
domain size, D = 10 and edges are created by taking pairs of variables randomly and connecting them considering fix
graph density, ρ = 0.5. For the second parameter, we execute the algorithms by changing domain size from 5 to 30
in Figure 5b where we consider the parameters number of variables X at 20 and graph density ρ at 0.5. Then for the
third parameter, we increase the graph density from 0.1 to 0.9 in Figure 5c setting the number of variables and domain
size as above mentioned. In case of Distributed RLFA Problem, we observe runtime by varying number of nodes from
5 to 30 in Figure 5d and setting other parameters with the same as the previous setting. In both type of settings, we
generated 30 instances and calculated the average runtime that we found by running each of the algorithms. All of the
experiments were performed on a simulator implemented in an Intel i7 Octacore 3.4GHz machine with 16GB of RAM.
Our experimental results for solving random DCOPs are depicted in Figures 5a -- 5c. In so doing, we generate three
synthesized graphs. Specifically, we use hard constraints that are either "less than", "greater than" or "equal" alongside
soft constraints for which we randomly generated utility values from the range [0, 100]. In Figure 5a, we observe that
runtime of Cec-DPOP increases in a steady way with respect to other algorithms as we increase the number of nodes.
The result can be understood by observing the fact that the larger the constraint graph, the greater the advantages in
parallelism and communication efficiency are found by CeC-DPOP. Another notable advantage of CeC is that it can
avoid performing operations on the values pruned during the consistency enforcement phase. As a result, our algorithm
is slightly faster than both BFS-DPOP and BrC-DPOP. More precisely, we observer that CeC-DPOP experiences
17 − 81% reduction in runtime compared to DPOP, 25 − 62% compared to Bfs-DPOP and 36 − 66% compared to
Brc-DPOP.
Figure 5b illustrates the results based on the next setting that is varying the domain size while setting the number of
nodes and graph density, we observe that runtime of Cec-DPOP increases at smaller rate than that of other algorithms.
To be precise, we find 49 − 75% smaller runtime than DPOP, 42 − 65% than Bfs-DPOP and 22 − 48% than Brc-DPOP.
The reason behind this performance is that when domain size increases, more values in each domain is pruned by
CeC-DPOP through consistency enforcement producing UTIL message of smaller dimension. As a result, the required
time to compute messages decreases at a significant rate. Though Brc-DPOP have relatively smaller runtime than
DPOP and BFS-DPOP for enforcing branch consistency, Cec-DPOP is always the winner through enforcing cross edge
consistency.
In the third experimental setting, we vary the the graph density and set the other two parameters (Figure 5c). we
observe a notable performance gain of CeC-DPOP in terms of runtime compared to other algorithms. To be exact,
we detect 47 − 85%, 35 − 84% and 10 − 50% reduction of runtime with comparison to DPOP, Bfs-DPOP and Brc-
DPOP, respectively. This behavior can be explained when we notice that CeC-DPOP uses BFS pseudo-tree as the
communication structure which is generated from dense constraint graph resulting more branches. As a result, more
parallelism is experienced. Another reason is that the number of edges is relatively higher in the dense constraint graphs
creating the opportunity of cross edge consistency enforcement at a significant level. Thus, more domain values are
pruned and more shorter messages are produces resulting in a smaller computation time. Overall, a great reduction
in runtime is observed. Another important behavior that should be mentioned is phase transition occurs for DPOP,
BFS-DPOP and BrC-DPOP in Figure 5a when number of variables goes across 20. DPOP and BFS-DPOP experiences
phase transition in Figure 5b when domain size increase towards 20. In both cases, transition for CeC-DPOP is relatively
smooth than other algorithms.
As aforementioned, Distributed RLFA Problem is considered as the second type of problem to evaluate CeC-DPOP
against the benchmarking algorithms. The distributed RLFA problem [20] consists of a set of channels, each having
a transmitter and receiver at both ends. The aim is to assign a frequency from a given set F by minimizing the total
interference at the receivers below an acceptable level and at the same time using as few and also as low frequencies
as possible. For our experiment, we mapped a transmitter as a variable and for simplicity, we assigned a single agent
to a variable. The domain of a variable consists of frequencies (chosen from available spectral resources) that can
be assigned to a variable. The interference between transmitter is modeled as a constraint of the form xi − xj > s
where xi, xj are variables and s is a random frequency separation. We varied the number of variables in Figure 5d and
observed the runtime with DPOP, BFS-DPOP, and BrC-DPOP. We set the domain size, D = 10, s ∈ {3, 4} and graph
8
density, ρ = 0.5. Significantly, the results are similar as observed to that in case of random DCOPs. More specifically,
CeC-DPOP outperforms other algorithms contributing 27− 75%, 12− 66% and 16− 35% reduction in runtime relative
to DPOP, Bfs-DPOP and Brc-DPOP. It is worth mentioning that in the distributed RLFA problem and also in random
DCOPs that was previously described, CeC-DPOP reduces message size around 5% than Brc-DPOP.
6 Conclusion
We present a new algorithm, CeC-DPOP, that significantly reduces the runtime of the DPOP algorithm that can be used
to solve DCOPs having both soft and hard constraints. We empirically observe that our algorithm performs around
10-85% faster than the state of the art algorithms. This is mainly possible because of the introduction of cross edge
consistency. In addition, the use of the BFS pseudo tree as a communication structure enables CeC-DPOP to perform
even faster. As a result, CeC-DPOP extends the use of DPOP in solving real-life problems that include both hard and
soft constraints. In future work, we intend to investigate whether our approach can be applied to other DPOP extensions,
as well as how much speedup can be achieved for them.
References
[1] M. Yokoo, E. H. Durfee, T. Ishida, and K. Kuwabara, "The distributed constraint satisfaction problem: Formal-
ization and algorithms," IEEE Transactions on knowledge and data engineering, vol. 10, no. 5, pp. 673 -- 685,
1998.
[2] R. T. Maheswaran, M. Tambe, E. Bowring, J. P. Pearce, and P. Varakantham, "Taking dcop to the real world:
Efficient complete solutions for distributed multi-event scheduling," in Proceedings of the Third International
Joint Conference on Autonomous Agents and Multiagent Systems-Volume 1, pp. 310 -- 317, IEEE Computer Society,
2004.
[3] F. Fioretto, W. Yeoh, and E. Pontelli, "A multiagent system approach to scheduling devices in smart homes," in
Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems, pp. 981 -- 989, International
Foundation for Autonomous Agents and Multiagent Systems, 2017.
[4] M. Jain, M. Taylor, M. Tambe, and M. Yokoo, "Dcops meet the real world: Exploring unknown reward matri-
ces with applications to mobile sensor networks," in Twenty-First International Joint Conference on Artificial
Intelligence, 2009.
[5] K. Hirayama and M. Yokoo, "The distributed breakout algorithms," Artificial Intelligence, vol. 161, no. 1-2,
pp. 89 -- 115, 2005.
[6] W. Zhang, G. Wang, Z. Xing, and L. Wittenburg, "Distributed stochastic search and distributed breakout: properties,
comparison and applications to constraint optimization problems in sensor networks," Artificial Intelligence,
vol. 161, no. 1-2, pp. 55 -- 87, 2005.
[7] A. Farinelli, A. Rogers, A. Petcu, and N. R. Jennings, "Decentralised coordination of low-power embedded devices
using the max-sum algorithm," in Proceedings of the 7th international joint conference on Autonomous agents
and multiagent systems-Volume 2, pp. 639 -- 646, International Foundation for Autonomous Agents and Multiagent
Systems, 2008.
[8] D. Orden, J. Gimenez-Guzman, I. Marsa-Maestre, and E. de la Hoz, "Spectrum graph coloring and applications to
wi-fi channel assignment," Symmetry, vol. 10, no. 3, p. 65, 2018.
[9] E. de la Hoz, J. M. Gimenez-Guzman, I. Marsa-Maestre, L. Cruz-Piris, and D. Orden, "A distributed, multi-agent
approach to reactive network resilience," in Proceedings of the 16th Conference on Autonomous Agents and
MultiAgent Systems, pp. 1044 -- 1053, International Foundation for Autonomous Agents and Multiagent Systems,
2017.
[10] K. Hirayama and M. Yokoo, "Distributed partial constraint satisfaction problem," in International Conference on
Principles and Practice of Constraint Programming, pp. 222 -- 236, Springer, 1997.
[11] A. Netzer, A. Grubshtein, and A. Meisels, "Concurrent forward bounding for distributed constraint optimization
problems," Artificial Intelligence, vol. 193, pp. 186 -- 216, 2012.
[12] P. J. Modi, W.-M. Shen, M. Tambe, and M. Yokoo, "Adopt: Asynchronous distributed constraint optimization with
quality guarantees," Artificial Intelligence, vol. 161, no. 1-2, pp. 149 -- 180, 2005.
[13] A. Petcu and B. Faltings, "A scalable method for multiagent constraint optimization," tech. rep., 2005.
[14] M. Vinyals, J. A. Rodriguez-Aguilar, and J. Cerquides, "Generalizing dpop: Action-gdl, a new complete algorithm
for dcops," in Proceedings of The 8th International Conference on Autonomous Agents and Multiagent Systems-
Volume 2, pp. 1239 -- 1240, International Foundation for Autonomous Agents and Multiagent Systems, 2009.
9
[15] F. Fioretto, T. Le, W. Yeoh, E. Pontelli, and T. C. Son, "Improving dpop with branch consistency for solving
distributed constraint optimization problems," in International Conference on Principles and Practice of Constraint
Programming, pp. 307 -- 323, Springer, 2014.
[16] A. Petcu and B. Faltings, "Odpop: An algorithm for open/distributed constraint optimization," in AAAI, vol. 6,
pp. 703 -- 708, 2006.
[17] A. Petcu and B. Faltings, "Mb-dpop: A new memory-bounded algorithm for distributed optimization.," in IJCAI,
pp. 1452 -- 1457, 2007.
[18] R. Greenstadt, B. Grosz, and M. D. Smith, "Ssdpop: improving the privacy of dcop with secret sharing," in
Proceedings of the 6th international joint conference on Autonomous agents and multiagent systems, p. 171, ACM,
2007.
[19] A. Petcu, B. Faltings, and R. Mailler, "Pc-dpop: A new partial centralization algorithm for distributed optimization.,"
in IJCAI, vol. 7, pp. 167 -- 172, 2007.
[20] B. Cabon, S. De Givry, L. Lobjois, T. Schiex, and J. P. Warners, "Radio link frequency assignment," Constraints,
vol. 4, no. 1, pp. 79 -- 89, 1999.
[21] A. Kumar, A. Petcu, and B. Faltings, "H-dpop: Using hard constraints for search space pruning in dcop.," in AAAI,
pp. 325 -- 330, 2008.
[22] K. C. Cheng and R. H. Yap, "Constrained decision diagrams," in Proceedings of the national conference on
artificial intelligence, vol. 20, p. 366, Menlo Park, CA; Cambridge, MA; London; AAAI Press; MIT Press; 1999,
2005.
[23] Z. Chen, Z. He, and C. He, "An improved dpop algorithm based on breadth first search pseudo-tree for distributed
constraint optimization," Applied Intelligence, vol. 47, no. 3, pp. 607 -- 623, 2017.
[24] A. Petcu, B. Faltings, and D. C. Parkes, "M-dpop: Faithful distributed implementation of efficient social choice
problems," Journal of Artificial Intelligence Research, vol. 32, pp. 705 -- 755, 2008.
[25] B. Schieber and U. Vishkin, "On finding lowest common ancestors: Simplification and parallelization," SIAM
Journal on Computing, vol. 17, no. 6, pp. 1253 -- 1262, 1988.
10
|
1908.08637 | 1 | 1908 | 2019-08-23T01:57:30 | Immediate Observation in Mediated Population Protocols | [
"cs.MA",
"cs.DC"
] | In this paper we analyze the computational power of variants of population protocols (PP), a formalism for distributed systems with anonymous agents having very limited capabilities. The capabilities of agents are enhanced in mediated population protocols (MPP) by recording the states in the edges of the interaction graph. Restricting the interactions to the communication model of immediate observation (IO) reduces the computational power of the resulting formalism. We show that this enhancement and restriction, when combined, yield a model (IOMPP) at least as powerful as the basic PP. The proof requires a novel notion of configurations in the MPP model allowing differentiation of agents and uses techniques similar to methods of analyzing encoding criteria, namely operational correspondence. The constructional part of the proof is generic in a way that all protocols can be translated into the new model without losing the desirable properties they might have besides a stable output. Furthermore, we illustrate how this approach could be utilized to prove our conjecture of IOMPP model being even as expressive as the MPP model. If our conjecture holds, this would result in a sharp characterization of the computational power and reveal the nonnecessity of two-way communication in the context of mediated population protocols. | cs.MA | cs | Immediate Observation in Mediated Population Protocols
Tobias Prehn
Modelle und Theorie Verteilter Systeme
TU Berlin, Germany
Myron Rotter
TU Berlin, Germany
[email protected]
[email protected]
In this paper we analyze the computational power of variants of population protocols (PP), a formal-
ism for distributed systems with anonymous agents having very limited capabilities. The capabilities
of agents are enhanced in mediated population protocols (MPP) by recording the states in the edges
of the interaction graph. Restricting the interactions to the communication model of immediate
observation (IO) reduces the computational power of the resulting formalism. We show that this en-
hancement and restriction, when combined, yield a model (IOMPP) at least as powerful as the basic
PP. The proof requires a novel notion of configurations in the MPP model allowing differentiation
of agents and uses techniques similar to methods of analyzing encoding criteria, namely operational
correspondence. The constructional part of the proof is generic in a way that all protocols can be
translated into the new model without losing the desirable properties they might have besides a sta-
ble output. Furthermore, we illustrate how this approach could be utilized to prove our conjecture
of IOMPP model being even as expressive as the MPP model. If our conjecture holds, this would
result in a sharp characterization of the computational power and reveal the nonnecessity of two-way
communication in the context of mediated population protocols.
1
Introduction
Population protocols have been introduced in 2004 as a computational model for passively mobile fi-
nite state sensors by Angluin et al. [2, 3]. They feature a finite state space, making them suitable for
computation units with very limited capabilities and full anonymity, resulting directly from this restric-
tion. Since the number of possible states that each agent could be in may not grow with the number of
participating agents, there is no space for memorizing the ids of already met communication partners or
similar constructs. Therefore, the outcome of any binary communication does not depend on whether
the participants have communicated before. Another feature is the fully distributed approach of the base
version for population protocols that does not need a base station, leader, or scheduler of any kind. The
impact of such extensions has been studied [5, 7, 1].
It is well known that predicates computable by population protocols are exactly the semilinear pred-
icates. The first study on the computational power of this model was in 2007 by Angluin et al. [4]. In
this context, also several different communication patterns have been modeled in population protocols
and their computational power have been studied as well. One of those mechanisms has been the imme-
diate observation model, which is a special kind of one-way communication as opposed to the two-way
communication that comes with the base model. The idea is that an agent may observe another agent
without it noticing being observed. Clearly the observed agent cannot change its state in such an in-
teraction whereas the observer can use the information given by its own and the observed agent's state.
In contrast to stronger mechanisms no synchronization between the communication partners is needed.
Consequently the communication in such a model is asynchronous and applicable to a broader variety
of systems. With the fully distributed setting in mind the immediate observation communication seems
to be a desirable feature. But these qualities come with a price. Protocols with this limitation to the
J.A. P´erez and J. Rot (Eds.): Combined Workshop on Expressiveness in
Concurrency and Structural Operational Semantics (EXPRESS/SOS 2019).
EPTCS 300, 2019, pp. 102 -- 113, doi:10.4204/EPTCS.300.7
c(cid:13) T. Prehn & M. Rotter
This work is licensed under the
Creative Commons Attribution License.
T. Prehn & M. Rotter
103
communication can only compute predicates in COUNT∗, i.e. predicates that count multiplicities of input
values and compare them to previously given thresholds.
Another approach to altering population protocols has been the work of Michail et al. [12]. In their
model agents are allowed to store distinct information for different communication partners. To achieve
this they extended the base formalism by states for each pair of agents residing in the edges of the
interaction graph. Some previous work on mediated population protocols modelled directed interaction
graphs with one state per edge [12] and others used undirected graphs where each edge has a state
for each of its two endpoints [8]. This extension is a reasonable compromise between maintaining the
anonymity of each agent and being able to memorize the already met communication partners. An
agent is capable of telling an agent, that it has not yet communicated with, apart from one it has already
met. But two other agents being in the same state and with the same communication history are still
indistinguishable. Aside from this the edge states can be used for storing several other information. This
mediated population protocols are able to compute all symmetric predicates in NSPACE(n2).
We now present a model in this paper that combines the extended storage possibilities of mediated
population protocols and the limited communication model of immediate observation protocols. The
computational power of our resulting formalism has to be studied as it is unclear how this extension and
restriction interact.
In section 2 the basic formalisms and existing models are defined. We are using a representation of
population that allows the distinction of agents from a global point of view. This does not interfere with
the anonymity of the agents and is for analysis purposes only. Based on this we define our model of
immediate observation mediated population protocols (IOMPP) in section 3. Subsequently we study the
computational power of our model in section 4. We take the approach simulating population protocols
in immediate observation mediated population protocols in 3.1. Additionally, we give a translation of
configurations from one model to the other and define criteria such a translation has to meet for it to
express desirable attributes in 4.1. Our work is inspired by and makes use of the encodability criteria
stated by Gorla in [11]. To the best of our knowledge this technique is novel to population protocols in
the way we utilize it in 4.2 to prove that immediate observation mediated population protocols are at least
as expressive as the base model of population protocols. In 4.3 we conjecture that our approach could
also be used to show that immediate observation does not restrict the computational power of mediated
population protocols. We conclude our paper by a discussion on the given results and possible application
of the used techniques in section 5. We also give an outlook on future work and open questions.
2 Technical preliminaries
First we introduce populations, which form the base of all population protocols. They are often modeled
as multisets to emphasize the indistinguishability of the participating agents. We will use vectors where
each entry represents the state of a specific agent, because we want to efficiently compare populations
in different models from a global viewpoint. Note that this will not give agents a distinct id they could
make use from their local point of view. A different kind of vector representation can be found in [10]
and is not to be confused with ours. They use vectors where each entry describes for an agent state the
multiplicities of agents in that state. Their vector representation efficiently stores sets of agents with their
states, making it easy to identify equivalent protocol states. In contrast to this, our representation can
be used to compare sets of agents in population protocols with sets of agents in extended variants like
mediated population protocols.
104
Immediate Observation in Mediated Population Protocols
Definition 1 (Populations). Let A be a nonempty finite set and n ∈ N. Then An describes the set of
n-tuples over A also referred to as vectors of length n. A population over A, denoted by POP (A), is the
set of all vectors of arbitrary but finite length over A. If v ∈ An we use v = n to denote the length of v
and (v)i to reference the ith element of v (with i ∈ N+
≤n).
Next we define how calculations in population protocols are modeled from a local point of view. Each
agent has the same set of states and rules how to change states according to a communication partner's
state. Additionally, functions to map input to an initial state and a state to some output are defined to be
identical for each agent.
δ) where
Definition 2 (Population Protocols [6]). A population protocol P is a 5-tuple P = (Q,
Q is a finite set of agent states, Σ a finite input alphabet, I : Σ → Q describes the input function, O :
Q → {0, 1} is the output function, and δ : Q2 → Q2 is referred to as transition function and describes all
possible pairwise interactions. We also making use of a set representation of δ ⊆ Q4 whenever we write
t = (p, q) →δ (p′
, q′) referring to a specific transition t ∈ δ.
, I , O,
Σ
To analyze protocols we need a global perspective. Here states are configurations holding state
information for each agent. This kind of global view has been used to study the computational power of
population protocols [4].
,
,
,
I
Σ
Σ
O
, I , O,
δ) be a population protocol. The global protocol
Definition 3 (Global Protocols [4]). Let P = (Q,
to P is a 5-tuple GP = (C
, −→) where C = POP (Q) is the set of configurations, i.e., vectors of
agent states Q, I : POP (Σ) → C maps input vectors to initial configurations, O : C → {0, 1, ⊥} maps
configurations to outputs, and −→: C → C is the global transition function with −→∗ being its reflexive
and transitive closure. For C,C′ ∈ C it holds that C −→ C′ iff there is a transition t ∈ δ and i, j ∈ N+
with i 6= j such that t = ((C)i , (C) j) →δ ((C′)i , (C′) j) and (C)k = (C′)k for every k ∈ N+ \ {i, j}. We also
ti, j−→ C′ and call agent i the initiator of t and j the responder or (if the state of i is not changed by
write C
t) observer. The global input function takes use of I to get an agent state for each single value in its input
and O aggregates the outputs of the agents according to O. It holds that O(C) = x ∈ {0, 1} iff O((C)i) = x
for each i ∈ N+
≤C and O(C) = ⊥ in every other case. When the underlying protocol P is clear from the
context we often omit the index of GP and simply state that G is the global protocol to P.
Based on a global protocol we can describe what it means for a protocol to compute some predicate.
For this we need to define executions and fairness.
,
,
,
I
Σ
O
Definition 4 (Computation). Let G = (C
, −→) be a global protocol. A configuration C ∈ C is
output stable with output x ∈ {0, 1} iff O(C′) = x for each C′ ∈ C with C −→∗ C′. We call a sequence of
configurations C0 ,C1 ,C2 , · · · ∈ C with Ci −→ Ci+1 for each i ∈ N an execution. An execution is fair iff for
each C ∈ C with Ci = C for infinitely many i ∈ N it holds that if there is a transition C −→ C′ then also
C j = C′ for infinitely many j ∈ N. A population protocol P is well-specified if for each input Inp, it holds
that all fair executions of P starting in I (Inp) reach a configuration that is output stable. P computes a
predicate if this reached configuration is output stable with output 1 if Inp satisfies the predicate and with
output 0 otherwise.
In the context of population protocols several communication mechanisms have been studied [4].
Immediate observation is one of those mechanisms. It reduces the class of computable predicates to
predicates counting multiplicities of input values COUNT∗. To model this kind of communication, restric-
tions to the allowed form of transitions are made.
δ) be a population protocol. P is an immedi-
Definition 5 (Immediate Observation). Let P = (Q,
ate observation protocol, if there is no transition that changes the state of the initiator. In other words, all
transitions t ∈ δ have to be of the form t = (p, q) →δ (p, q′).
, I , O,
Σ
T. Prehn & M. Rotter
105
Another extension to population protocols is the mediated variant. The idea is to introduce states in
all edges of the communication graph. Since the most general graph is the complete graph, each pair of
agents is given such a state. In the context of an immediate observation communication mechanism, it is
not reasonable to assume a storage that both agents can write to. Therefore, we introduce a pair of edge
states for each pair of agents. Edge states are always initialized with the same value.
Σ
, S, s0 , I , O,
Definition 6 (Mediated Population Protocols [12]). A mediated population Protocol P is a 7-tuple P =
(Q,
, I and O are analogous to population protocols. The set of edge states S
includes the initial edge state s0 ∈ S and the transition function δ : (Q × S)2 → (Q × S)2 incorporates the
edge states for each pair of agents.
δ) where Q,
Σ
Configurations in mediated population protocols cannot be represented by simple vectors. We need
to introduce matrices as configurations containing the agent states on the diagonal and the states of the
edge between agents a and b in fields Ca,b (side of agent a) and Cb,a (side of agent b).
Definition 7 (Mediated Populations). An×n describes the set of square matrices over A of size n × n. A
mediated population over A denoted by POPM (A) is the set of all matrices of arbitrary but finite length
over A. If m ∈ An×n we use (m)i, j to reference the element of m at column i and row j with i, j ∈ N+
≤n
and m = n to denote the length as well as the height of a square matrix m.
We can now proceed with lifting our global protocol definitions to represent mediated population
protocols as well.
δ) be a
Definition 8 (Global Protocols for Mediated Population Protocols). Let P = (Q,
Σ
mediated population protocol. The global protocol to P is again a 5-tuple G = (C
In
contrast to global protocols for simple population protocols C = POPM (Q) is the set of configurations
and I : POP (Σ) → C maps input vectors to initial configurations, initializing the diagonal fields with
the corresponding agent states and every other field with s0. The output function O : C → {0, 1, ⊥}
ignores all fields not on the diagonal and −→: C → C now also changes the respective edge states.
For C,C′ ∈ C it holds that C −→ C′ iff there is a transition t ∈ δ and i, j ∈ N+ with i 6= j such that
t = ((C)i,i , (C)i, j , (C) j, j , (C) j,i) →δ ((C′)i,i , (C′)i, j , (C′) j, j , (C′) j,i) and (C′)k,l = (C′)k,l for every k, l ∈
N+ \ {i, j}.
, S, s0 , I , O,
, −→).
O
Σ
I
,
,
,
3 Modelling immediate observation in mediated population protocols
From the technical preliminaries in section 2 we can easily combine the models for mediated population
protocols and immediate observation conform communication. We get our model of population protocols
with two edge states in every edge, one per communication partner, and transitions that keeps the states
of the initiator unaltered and changes the states of the observer.
Definition 9 (Immediate Observation Mediated Population Protocols). An immediate observation me-
δ) where Q is a finite set of agent states,
diated population protocol P is a 7-tuple P = (Q,
Σ a finite input alphabet, I : Σ → Q describes the input function, O : Q → {0, 1} is the output func-
tion, and δ : (Q × S)2 → (Q × S)2 is referred to as transition function and describes all possible pair-
wise interactions. We also making use of a set representation of δ ⊆ (Q × S)4 whenever we write
, r′) referring to a specific transition t ∈ δ. Since our model uses the im-
t = (p, s, q, r) →δ (p′
mediate observation communication mechanism, all transitions t = (p, s, q, r) →δ (p′
, r′) have to
satisfy p = p′ and s = s′.
, S, s0 , I , O,
, q′
, q′
, s′
, s′
Σ
106
Immediate Observation in Mediated Population Protocols
3.1 Simulating population protocols by immediate observation mediated population pro-
tocols
We can now simulate protocols in the basic population protocol model by immediate observation me-
diated population protocols. The main idea is to split every two-way communication with an initiator
and a responder into 4 steps. Two steps are required to signal the request and the acknowledgement of a
communication and another two steps are needed to finish the communication resolving all pending state
changes. Additionally, a reset transition is given for the case of an unsuccessful communication.
Simulation 10. Let P = (Q,
mediated population protocol P′ simulates the protocol P and is given by the tuple (Q′
where
δ) be a population protocol. The following immediate observation
δ′)
, I , O,
, O′
, s′
0 , I′
Σ
Σ′
, S′
,
,
Σ′
Q′
S′
s′
0
I′(σ)
O′(l , q)
:= Σ
,
:= {L,U } × Q,
:= {sinit , sponr} ∪ Q,
:= sinit ,
:= (U , I(σ)) for all σ ∈ Σ
:= O(q) for all (l , q) ∈ Q′
,
.
The states of the agents are of the form (l , q) where l ∈ {L,U } indicates whether the agent is locked or
unlocked and q ∈ Q is the computation state according to the original population protocol P. For easier
referencing we call the first component of an agent state (l , q) the locking state l of this agent and the
second component its computation state q. We use the formulation of an agent being locked whenever its
locking state is L and say this agent is unlocked otherwise. W.l.o.g. we assume that {sinit , sponr} ∩ Q = /0.
The input function I maps each input symbol σ ∈ Σ to the state (U , I(σ)). The output function O′
maps each state (l , q) ∈ Q to O(q) independent of the locking indicator l. We specify for each transition
t = (p, q) →δ (p′
, q′ ∈ Q the following transitions for δ′.
, q′) of δ with p, q, p′
t(1) = ((U , p), sinit , (U , q), sinit ) →δ′
t(2) = ((L, q′), q, (U , p), sinit ) →δ′
t(3) = ((L, p′), sponr , (L, q′), q) →δ′
t(4) = ((x, y), sinit , (L, p′), sponr) →δ′
t(5) = ((x, y), z, (L, q′), q) →δ′
((U , p), sinit , (L, q′), q)
((L, q′), q, (L, p′), sponr)
((L, p′), sponr , (U , q′), sinit )
((x, y), sinit , (U , p′), sinit )
for every (x, y) ∈ Q′
((x, y), z, (U , q), sinit )
for every (x, y) ∈ Q′
and z ∈ S′ \ {sponr}
(1)
(2)
(3)
(4)
(5)
The locking state of each agent prohibits simultaneous participation in several different communi-
cations. Whenever an agent took part in a two-way communication t in the original protocol, it could
be the observer of a transition of type (1) in the simulation. A t(1) transition locks the observing agent
and puts its old computation state in the edge state this agent controls on the edge with the observed
agent. This has two reasons: First it signals the interest in a communication with the other agent and
second it backups the old state for a potential future reset. If the other agent observes the change in the
edge state, it may signal the acknowledgement of requested communication by locking itself, changing
T. Prehn & M. Rotter
107
its computation state according to the transition t and putting sponr in its edge state. This is achieved by
transition t(2). Now the two agents have to reset their edge states to sinit and unlock themselves. The
agent mimicking the responder of the original transition t starts by taking t(3), followed by the simulator
of the original initiator taking t(4). If a communication was not successful, either because the initiators
surrogate has taken an other transition with another agent in the meantime or because responder sim-
ulating agent observes its partner before it could acknowledge the communication, t(5) is taken. This
transition assures that in the described cases an agent can give up on a communication attempt and reset
its state, readying itself for another attempt, potentially with a different partner.
Note that if a transition t = (p, q) →δ (p, q′) is already immediate observation compliant we do not
need to add the whole set of transitions. We could instead add a slightly altered version of the original
transition as follows.
t(6) = ((U , p), sinit , (U , q), sinit ) →δ′
((U , p), sinit , (U , q′), sinit )
(6)
Clearly the result would be the same. This kind of transition is only needed if the simulation needs to
be more efficient in the sense of steps needed to get to an output stable configuration. We will therefore
omit this type of transitions in our analyses.
Observation 11 (Output Changing Transitions). By definition of O′ the output of an agent only depends
on its computation state. As transitions t(3), t(4) do not change the computation states, only transitions
t(1), t(2), and t(5) can have an impact on the output of an agent. Since Simulation 10 is an immediate
observation protocol, this agent has to be the observer of such transitions.
Observation 12 (Number of Started Conversations). Every agent has at most one started and not yet
concluded conversation at any point in time. Starting a conversation by taking transition t(1) as respon-
der brings an agent to a locked state. Therefore, no other conversation can be started or acknowledged by
this agent until the conversation is concluded with transition t(3) or aborted with transition t(5). Acknowl-
edging a conversation by taking transition t(2) as responder also brings an agent to a locked state. Again
no other conversation can be started or acknowledged by this agent until the conversation is concluded
with transition t(4) or aborted with transition t(5).
(2)
Observation 13 (Point of no Return). Every occurrence of transition t
i, j with acting agents i and j is
(2)
i, j agent i is locked with q ∈ Q in
eventually followed by transitions t
its edge state to j and agent j is locked with sponr in its edge state to i. From Observation 12 we know
that neither i nor j can be observer of any transition with some agent different from i and j. From
(4)
i, j . After execution of t
(3)
j,i and t
is enabled and will be taken at some point because of the fairness
the transitions with i and j only t
assumptions in population protocols. After that agent j is still locked with sponr in its edge state to i.
Therefore, j can only be observer of transitions t(3)
,t(4), or t(5). Again from Observation 12 we know
that only t
is possible.
(3)
j,i
(4)
i, j
4 Computational power
We now show that our model can compute all predicates computable in population protocols by giving a
translation, that relates configurations from a protocol to configurations from its simulation representing
the same state of computation. Additionally, we identify requirements imposed on such a translation to
be helpful in proving the equality of computed predicates. We will ultimately show how this proof is
executed.
108
Immediate Observation in Mediated Population Protocols
4.1 Translation and criteria
We provide a translation, that constructs configurations in mediated population protocols from configu-
rations of population protocols by giving all agents an unlocked state and setting all edges to the neutral
sinit state.
, s′
0 , I′
,
Definition 14 (Translation of Configurations). Let P = (Q,
(Q′
δ) be a population protocol and P′ =
δ′) a immediate observation mediated population protocol constructed from P using
, I , O,
, O′
, S′
Σ
Σ′
,
Simulation 10. ByJ·K : POP (Q) → POPM (Q′) we denote the translation of configurations C ∈ POP (Q)
in the population protocol into configurations D ∈ POPM (Q′) from the mediated population protocol.
This translation is defined as follows:
(U , q1)
sinit
sinit
...
sinit
(U , q2)
. . .
. . .
. . .
. . .
. . .
sinit
sinit
...
sinit
(U , qn)
J(q0 , q1 , . . . , qn)K =
From the criteria for good encodings defined by Gorla [11] we adopt the notion of operational corre-
spondence.
corresponding if it is
Definition 15 (Operational Correspondence). A translation J·K : POP (Q) → POPM (Q′) is operationally
(1) (operationally) complete, i.e., for all C,C′ ∈ POP(Q) with C −→∗ C′ it holds that JCK −→∗ JC′K,
(2) (operationally) sound, i.e., for all C ∈ POP (Q) and D ∈ POPM (Q′) withJCK −→∗ D there exists a
and
C′ ∈ POP (Q) with D −→∗ JC′K and C −→∗ C′.
If our translation in Def. 14 instantiated with concrete population protocol P and mediated popula-
tion protocol P′ is operationally corresponding, we get that every configuration reachable in P is also
reachable in P′ and vice versa.
Definition 16 (Input/Output Correspondence). Let P be a population protocol, P′ be a mediated popula-
POPM (Q′) is I/O corresponding if it is
tion protocol and G, G′ be the global protocols to P and P′ respectively. A translation J·K : POP(Q) →
(1) input corresponding, i.e., for all V ∈ POP (Σ) it holds that JI (V )K = I′ (V ), and
(2) output corresponding, i.e., for all C ∈ POP (Q) it holds that O (C) = O′ (JCK).
Input/Output Correspondence gives us the assurance that input and output functions of the protocols
related by a translation behave in a similar way. If it holds, translating an input configuration of the
original protocol or directly using the input function of the corresponding mediated population protocol
yields the same result. Additionally, configurations are always translated into configurations with the
same output.
Definition 17 (Output Stability Preservation). Let P be a population protocol, P′ be a mediated popula-
tion protocol and G, G′ be the global protocols to P and P′ respectively. A translation J·K : POP(Q) →
POPM (Q′) is output stability preserving if for each C ∈ POP (Q) it holds that JCK is output stable iff C
is output stable.
From the output stability preservation we get that each output stable configuration is translated into
a configuration also being output stable.
T. Prehn & M. Rotter
Lemma 18. Let P = (Q,
Σ
, I , O,
δ) be a population protocol and P′ = (Q′
,
Σ′
, S′
109
, O′
,
δ′) a mediated
, s′
0 , I′
population protocol. If there is a translation J·K : POP(Q) → POPM (Q′) that is operationally corre-
sponding, input/output corresponding, and output stability preserving, then P and P′ compute the same
predicate.
Proof. Since the translation is input corresponding every initial configuration in P has a correspond-
ing initial configuration in P′. From the operational correspondence we know that a configuration is
reachable from an initial configuration in P′ iff it has a corresponding configuration reachable from the
corresponding initial configuration in P. With output correspondence both configurations clearly have
the same output and, since the translation is output stability preserving, the output is either stable in both
configurations or in none. Therefore, the protocol P calculates the same semilinear predicate as P′.
4.2 All semilinear predicates can be computed by immediate observation mediated pop-
ulation protocols
Lemma 19. The translation J·K given in Def. 14 is operationally corresponding for any population pro-
δ) and the mediated population protocol P′ = (Q′
δ′) constructed
0 , I′
, O′
, S′
, s′
Σ′
,
,
tocol P = (Q,
, I , O,
using Simulation 10.
Σ
Proof. To prove operational completeness assume that C,C′ ∈ POP (Q) with C −→∗ C′. Since −→∗
is defined as reflexive-transitive closure of −→ we get that C1 ,C2 , . . . ,Cn ∈ POP (Q) exist with C −→
C1 −→ C2 −→ . . . −→ Cn −→ C′. We can always simulate a step Ci −→ Ci+1 in P by making 4 steps in P′.
Assume that the step is due to transition t ∈ δ and agents at a and b are acting as initiator and responder
respectively in Ci
ta,b−→ Ci+1. This can be simulated as JCiK
JCiK −→4 JCi+1K holds. Thus, JCK −→4 JC1K −→4
b,a−→ JCi+1K and thus
. . . −→4 JCnK −→4 JC′K exemplifies JCK −→∗ JC′K.
To prove operational soundness assume that C ∈ POP (Q) and D ∈ POPM (Q′) withJCK −→∗ D. We
construct C′ as follows. For this assume (D)i,i = (li , di).
t
b,a−→ C3
i
t
a,b−→ C4
i
t
a,b−→ C2
i
t
(1)
(2)
(3)
(4)
(C′)i =(g
di
, if there is exactly one j ∈ N+
≤D,6=i such that (D)i, j = g and (D) j,i 6= sponr
, otherwise
We can show that D −→∗ JC′K by taking the appropriate transitions for all i, j ∈ N+
follows.
≤D with i 6= j as
if (D)i, j = q
if (D)i, j = sinit
if (D)i, j = q
and
and
and
(D) j,i = sponr
(D) j,i = sponr
(D) j,i 6= sponr
take transitions t
take transition t
take transition t
(4)
i, j
(3)
j,i ,t
(4)
i, j
(5)
j,i
otherwise
do nothing
(2)
is a step Dx
t
i, j−→ Dy in this path, take transition Cv
SinceJCK −→∗ D by assumption and D −→∗JC′K we get thatJCK −→∗JC′K. We can construct C −→
C1 −→ C2 −→ . . . −→ C′ from the pathJCK −→ D1 −→ D2 −→ . . . −→JC′K as follows. Whenever there
ti, j−→ Cw in the path of C −→∗ C′. By Observation 12 we
get that each such step of type (2) is eventually followed by transitions of type (3) and (4) and since JC′K
has only edge states sinit , this has to happen beforeJC′K is reached. Therefore, there exists a C′ ∈ POP (Q)
with D −→∗ JC′K and C −→∗ C′.
110
Immediate Observation in Mediated Population Protocols
Lemma 20. The translation J·K given in Def. 14 is I/O corresponding for any population protocol
δ) and the mediated population protocol P′ = (Q′
δ′) constructed using
0 , I′
, O′
, S′
, s′
Σ′
,
,
Σ
P = (Q,
, I , O,
Simulation 10.
Proof. The input function I′ of P′ makes use of the input function I of P by putting an agent in the
unlocked state (U , I(σ)) iff the same agent would be in state I(σ) in P. Also every edge state is initialized
agents state q to a tuple with q being the second component i.e. its computation state. For each agent a
The output function O′ of P′ makes use of the output function O of P by ignoring the locking state of
with sinit . This matches J·K that translates each state q to (U , q) and sets every edge state to sinit . Thus,
JI (V )K = I′ (V ) and J·K is input corresponding.
an agent and giving back the output of O for the computation state. As seen above, J·K translates every
it holds that O ((C)a) = O′(cid:16)(D)a,a(cid:17). As O aggregates the outputs of all agents, which are the same in P
and P′, O (C) = O′ (JCK) and consequently J·K is output corresponding.
Lemma 21. The translationJ·K given in Def. 14 is output stability preserving for any population protocol
δ) and the mediated population protocol P′ = (Q′
δ′) constructed using
0 , I′
, O′
, S′
, s′
Σ′
,
,
Σ
P = (Q,
, I , O,
Simulation 10.
output stable in P′. From the definition of output stability follows that a configuration D exists with
Proof. Let C be a configuration that is output stable in P. Assume towards contradiction JCK is not
O′ (JCK) 6= O′ (D) and D is reachable from JCK, i.e. there is a path JCK −→ D1 −→ D2 −→ . . . −→ D.
W.l.o.g assume that D is the first such configuration in this path, i.e. O′ (JCK) = O′ (Di) for each such Di.
Since O′ aggregates the values of O′ for each agent and because P′ is an immediate observation protocol,
there has to be a single agent a that has changed its output because of the transition leading to D. From
Observation 11 we know that this transition has to be of type (1), (2), or (5) and a has to be its observer.
If it is (1) or (2) we can construct a configuration in the same way as C′ was constructed in the proof
O (C) 6= O (C′), a contradiction to C being output stable.
If the transition agent a took was of type (5), there has to be a configuration along the path from
changes the output of a, the corresponding type (1) transition must also have changed it, contradicting
(1), (2), or (5) and therefore does not change its output. Because our translation maintains outputs for
of operational soundness for Lemma 19. This C′ is reachable from C in P and JC′K is reachable from
D in P′. Note that on the path from D to JC′K agent a is never an observer of any transition with type
each agent O ((C)a) = O′(cid:16)(JCK)a,a(cid:17) 6= O′(cid:16)(D)a,a(cid:17) = O′(cid:16)(JC′K)a,a(cid:17) = O ((C′)a) holds. This results in
JCK to D where agent a took a transition of type (1). Observation 12 states that there has to be such a
transition in advance and by the definition of J·K this has to be after JCK. But if the transition of type (5)
our assumption of D being the first configuration with an output different from O′ (JCK). This is due
For the other direction assume towards contradiction JCK is output stable in P′ and C is not output
stable in P. Then there exists a configuration C that is reachable from C with O (C) 6= O(cid:0)C(cid:1). BecauseJ·K
is operationally complete by Lemma 19 it holds that qCy is reachable from JCK. From the output corre-
spondence ofJ·K in Lemma 20 follows that O′ (JCK) = O (C) 6= O(cid:0)C(cid:1) = O′(cid:0)qCy(cid:1). This is a contradiction
to the output stability of JCK in P′.
to a type (5) transition resetting the computation state of an agent back to the state it had before the
corresponding type (1) transition.
Theorem 22. For any population protocol P = (Q,
ulation protocol P′ constructed from P using Simulation 10 calculates the same semilinear predicate.
δ), the immediate observation mediated pop-
, I , O,
Σ
T. Prehn & M. Rotter
111
Proof. We have shown thatJ·K from Definition 14 is operationally corresponding, I/O corresponding and
output stability preserving in Lemmas 19, 20, and 21. By Lemma 18 the statement directly follows.
Corollary 23. Immediate observation mediated population protocols can compute every semilinear
predicate and are therefore at least as expressive as population protocols.
4.3 Immediate observation does not restrict the computational power of mediated pop-
ulation protocols
The approach in the previous sections can be used to prove that two-way communication in mediated
population protocols does not add to the computational power of the model. To achieve this we define a
simulation of mediated population protocols into the variant with immediate observation communication.
Simulation 24. Let P = (Q,
δ) be a mediated population protocol. The following imme-
diate observation mediated population protocol P′ simulates the protocol P and is given by the tuple
(Q′
δ′) where
, S, s0 , I , O,
, O′
, S′
Σ′
Σ
,
,
, s′
0 , I′
Σ′
Q′
S′
s′
0
I′(σ)
O′(l , q)
:= Σ
,
:= {L,U } × Q,
:= ({sinit , sponr} ∪ (Q × S)) × S,
:= (sinit , s0) ,
:= (U , I(σ)) for all σ ∈ Σ
:= O(q) for all (l , q) ∈ Q′
.
,
W.l.o.g. we assume that {sinit , sponr} ∩ (Q ∪ S) = /0. In contrast to Simulation 10 the edge state has two
components. The first component again signals the current state of the simulated communication and
serves as a backup for the condition prior to the communication. Here we need to save both, computation
state and edge state. The second component represents the actual edge state present in the original
protocol. We specify for each transition t = (p, r, q, s) →δ (p′
, q′ ∈ Q and
r, s, r′
, s′ ∈ S the following transitions for δ′.
, s′) of δ with p, q, p′
, q′
, r′
t(1) = ((U , p), (sinit , r), (U , q), (sinit , s)) →δ′
t(2) = ((L, q′), ((q, s), s′), (U , p), (sinit , r)) →δ′
t(3) = ((L, p′), (sponr , r′), (L, q′), ((q, s), s′)) →δ′
t(4) = ((x, y), (sinit , s′), (L, p′), (sponr , r′)) →δ′
t(5) = ((x, y), (v, w), (L, q′), ((q, s), s′)) →δ′
((U , p), (sinit , r), (L, q′), ((q, s), s′))
((L, q′), ((q, s), s′), (L, p′), (sponr , r′))
((L, p′), (sponr , r′), (U , q′), (sinit , s′))
((x, y), (sinit , s′), (U , p′), (sinit , r′))
for every (x, y) ∈ Q′
((x, y), (v, w), (U , q), (sinit , s))
for every (x, y) ∈ Q′
and (v, w) ∈ S′ \ {(sponr , r′)}
This simulation follows the same ideas as the Simulation 10. The only difference is the edge state of
the original protocol that needs to be taken into account by transitions of the simulation and that needs to
be backed up for possible future resets. We can now give a translation similar to Definition 14 required
for our line of argumentation.
112
Immediate Observation in Mediated Population Protocols
Definition 25 (Translation of Mediated Configurations). Let P = (Q,
ulation protocol and P′ = (Q′
δ) be a mediated pop-
δ′) a immediate observation mediated population protocol
, S, s0 , I , O,
, O′
, S′
, s′
Σ′
Σ
,
,
constructed from P using Simulation 24. By J·K : POPM (Q) → POPM (Q′) we denote the translation of
configurations C ∈ POPM (Q) in the population protocol into configurations D ∈ POPM (Q′) from the
mediated population protocol. This translation is defined as follows:
0 , I′
(sinit , sn,1)
...
(sinit , sn,n−1)
(sinit , sn−1,n)
(U , qn)
u
wwwwv
q1
s2,1
s1,2
...
s1,n
q2
. . .
. . .
. . .
. . .
. . .
sn−1,n
sn,1
...
sn,n−1
qn
}
(cid:127)(cid:127)(cid:127)(cid:127)~
=
(U , q1)
(sinit , s2,1)
(sinit , s1,2)
...
(sinit , s1,n)
(U , q2)
. . .
. . .
. . .
. . .
. . .
With this simulation and translation we conjecture that each mediated population protocol can be
simulated by a immediate observation mediated population protocol that shares several attributes, espe-
cially computing the same predicates.
Conjecture 26. For any mediated population protocol P = (Q,
tion mediated population protocol P′ = (Q′
calculates the same semilinear predicate.
0 , I′
, O′
, S′
, s′
Σ′
,
,
Σ
, S, s0 , I , O,
δ), the immediate observa-
δ′) constructed from P using Simulation 24
5 Conclusion and future work
We have given a proof for the model of immediate observation mediated population protocols to compute
all semilinear predicates. Thus they are as least as powerful in computation as population protocols.
Additionally, we have given arguments why we believe this model is even equivalent to the model of
mediated population protocols with two-way communication. Consequently allowing the initiator of a
transition to change its agent and edge states does not contribute to the computational power. The proof
of our Conjecture 26 can hopefully be done in our future research.
Our approach asks for a simulation and a translation which might seem overly complicated for a
proof of equal computational power. But additionally several other attributes, besides the computation
of the same predicate, carry over from the one protocol to the other if our Simulation 10 and translation
from Definition 14 are used. Consider for example livelock freedom, i.e. no configuration is reached that
has no successor besides itself. The simulation can reach a livelock iff the original protocol can reach
such a configuration. This can easily be derived from the operational correspondence in Definition 15.
In the context of protocols computing some predicate, a livelock is only possible if an output is reached.
Otherwise the requirements for a well-specified protocol are not met. A livelock can be a desirable state
as the computation can clearly be stopped in such a configuration. If the protocol does something else
than computing a predicate, livelocks can be even more important to be reached or avoided, depending
on the situation.
Another example is the analysis of a required communication structure. Whereas some protocols
need a full interaction graph to carry out a computation, a path structure would suffice for others to get a
correct result. The interaction graphs supporting a protocol do also support its simulation.
As a last example consider failure resistance [9].
If a protocol is designed to tolerate a certain
number and type of faults, the simulation of this protocol could be capable of a comparable behaviour.
This however depends on the type of failure and the chosen strategy to handle it. Crash failures, where
an agent may leave the population at any time, should be manageable in the simulation with the same
mechanisms as the original protocol did. Message losses could lead to new problems in the simulation
T. Prehn & M. Rotter
113
like deadlocked communication partners. Some error handling and error masking strategies could lead
to the number of failures tolerable by the simulation being reduced in contrast to the original protocol.
A study on desirable attributes and how they carry over from one protocol to another by our simula-
tion is something we wish to address in the future.
References
[1] Dan Alistarh, Rati Gelashvili & Milan Vojnovi´c (2015): Fast and Exact Majority in Population Protocols.
In: Proceedings of the 2015 ACM Symposium on Principles of Distributed Computing, PODC '15, ACM,
New York, NY, USA, pp. 47 -- 56, doi:10.1145/2767386.2767429.
[2] Dana Angluin, James Aspnes, Zoe Diamadi, Michael J. Fischer & Ren´e Peralta (2004): Computation in net-
works of passively mobile finite-state sensors. In: Proceedings of the twenty-third annualACM symposium
on Principlesof distributedcomputing, ACM, pp. 290 -- 299, doi:10.1145/1011767.1011810.
[3] Dana Angluin, James Aspnes, Zoe Diamadi, Michael J. Fischer & Ren´e Peralta (2006): Computation in
networks of passively mobile finite-state sensors. Distributed Computing 18(4), pp. 235 -- 253, doi:10.1007/
s00446-005-0138-3.
[4] Dana Angluin, James Aspnes, David Eisenstat & Eric Ruppert (2007): The computational power of popula-
tion protocols. DistributedComputing 20(4), pp. 279 -- 304, doi:10.1007/s00446-007-0040-2.
[5] Dana Angluin, James Aspnes, Michael J. Fischer & Hong Jiang (2008): Self-stabilizing Population Protocols.
ACM Trans.Auton.Adapt. Syst. 3(4), pp. 13:1 -- 13:28, doi:10.1145/1452001.1452003.
[6] James Aspnes & Eric Ruppert (2009): An Introduction to Population Protocols, pp. 97 -- 120. Springer Berlin
Heidelberg, Berlin, Heidelberg, doi:10.1007/978-3-540-89707-1_5.
[7] Shantanu Das, Giuseppe Antonio Di Luna, Paola Flocchini, Nicola Santoro & Giovanni Viglietta (2017):
Mediated Population Protocols: Leader Election and Applications. In T V Gopal, Jager Gerhard & Silvia
Steila, editors: TheoryandApplicationsofModelsofComputation, Springer International Publishing, Cham,
pp. 172 -- 186, doi:10.1007/978-3-319-55911-7_13.
[8] Shantanu Das, Giuseppe Antonio Di Luna, Paola Flocchini, Nicola Santoro & Giovanni Viglietta (2017):
Mediated Population Protocols: Leader Election and Applications. In: Theory and Applications of Mod-
els of Computation, Lecture Notes in Computer Science, Springer, Cham, pp. 172 -- 186, doi:10.1007/
978-3-319-55911-7_13.
[9] Carole Delporte-Gallet, Hugues Fauconnier, Rachid Guerraoui & Eric Ruppert (2006): When Birds Die:
Making Population Protocols Fault-tolerant. In: Proceedings of the Second IEEE InternationalConference
on Distributed Computing in Sensor Systems, DCOSS'06, Springer-Verlag, Berlin, Heidelberg, pp. 51 -- 66,
doi:10.1007/11776178_4.
[10] Javier Esparza, Pierre Ganty, J´erome Leroux & Rupak Majumdar (2017): Verification of population proto-
cols. Acta Informatica 54(2), pp. 191 -- 215, doi:10.1007/s00236-016-0272-3.
[11] Daniele Gorla (2010): Towards a unified approach to encodability and separation results for process calculi.
Informationand Computation 208(9), pp. 1031 -- 1053, doi:10.1016/j.ic.2010.05.002.
[12] Othon Michail, Ioannis Chatzigiannakis & Paul G. Spirakis (2011): Mediated population protocols. Theo-
reticalComputerScience 412(22), pp. 2434 -- 2450, doi:10.1016/j.tcs.2011.02.003.
|
1909.09087 | 1 | 1909 | 2019-09-19T16:42:21 | Real-Time Verification for Distributed Cyber-Physical Systems | [
"cs.MA"
] | Safety-critical distributed cyber-physical systems (CPSs) have been found in a wide range of applications. Notably, they have displayed a great deal of utility in intelligent transportation, where autonomous vehicles communicate and cooperate with each other via a high-speed communication network. Such systems require an ability to identify maneuvers in real-time that cause dangerous circumstances and ensure the implementation always meets safety-critical requirements. In this paper, we propose a real-time decentralized reachability approach for safety verification of a distributed multi-agent CPS with the underlying assumption that all agents are time-synchronized with a low degree of error. In the proposed approach, each agent periodically computes its local reachable set and exchanges this reachable set with the other agents with the goal of verifying the system safety. Our method, implemented in Java, takes advantages of the timing information and the reachable set information that are available in the exchanged messages to reason about the safety of the whole system in a decentralized manner. Any particular agent can also perform local safety verification tasks based on their local clocks by analyzing the messages it receives. We applied the proposed method to verify, in real-time, the safety properties of a group of quadcopters performing a distributed search mission. | cs.MA | cs | Real-Time Verification for Distributed
Cyber-Physical Systems
Hoang-Dung Tran1, Luan Viet Nguyen2, Patrick Musau1, Weiming Xiang3,
and Taylor T. Johnson1
1 Institute for Software Integrated Systems, Vanderbilt University, TN, USA
2 Department of Computer Science and Engineering, University of Notre Dame, USA
3 School of Computer and Cyber Sciences, Augusta University, USA
Abstract. Safety-critical distributed cyber-physical systems (CPSs) have
been found in a wide range of applications. Notably, they have displayed
a great deal of utility in intelligent transportation, where autonomous
vehicles communicate and cooperate with each other via a high-speed
communication network. Such systems require an ability to identify ma-
neuvers in real-time that cause dangerous circumstances and ensure the
implementation always meets safety-critical requirements. In this paper,
we propose a real-time decentralized reachability approach for safety veri-
fication of a distributed multi-agent CPS with the underlying assumption
that all agents are time-synchronized with a low degree of error. In the
proposed approach, each agent periodically computes its local reachable
set and exchanges this reachable set with the other agents with the goal
of verifying the system safety. Our method, implemented in Java, takes
advantages of the timing information and the reachable set information
that are available in the exchanged messages to reason about the safety
of the whole system in a decentralized manner. Any particular agent can
also perform local safety verification tasks based on their local clocks
by analyzing the messages it receives. We applied the proposed method
to verify, in real-time, the safety properties of a group of quadcopters
performing a distributed search mission.
1
Introduction
The emergence of 5G technology has inspired a massive wave of the research
and development in science and technology in the era of IoT where the commu-
nication between computing devices has become significantly faster with lower
latency and power consumption. The power of this modern communication tech-
nology influences and benefits all aspects of Cyber-Physical Systems (CPSs) such
as smart grids, smart homes, intelligent transportation and smart cities. In par-
ticular, the study of autonomous vehicles has become an increasingly popular
research field in both academic and industrial transportation applications. Au-
tomotive crashes pose significant financial and life-threatening risks, and there
is an urgent need for advanced and scalable methods that can efficiently verify
a distributed system of autonomous vehicles.
9
1
0
2
p
e
S
9
1
]
A
M
.
s
c
[
1
v
7
8
0
9
0
.
9
0
9
1
:
v
i
X
r
a
Over the last two decades, although many methods have been developed to
conduct reachability analysis and safety verification of CPS, such as the ap-
proaches proposed in [1, 3, 4, 6, 12, 13, 16, 18, 21, 27, 29], applying these techniques
to real-time distributed CPS remains a big challenge. This is due to the fact
that, 1) all existing techniques have intensive computation costs and are usu-
ally too slow to be used in a real-time manner and, 2) these techniques target
the safety verification of a single CPS, and therefore they naturally cannot be
applied efficiently to a distributed CPS where clock mismatches and communi-
cation between agents (i.e., individual systems) are essential concerns. Since the
future autonomous vehicles systems will work distributively involving effective
communication between each agent, there is an urgent need for an approach
that can provide formal guarantees of the safety of distributed CPS in real-time.
More importantly, the safety information should be defined based on the agents
local clocks to allow these agents to perform "intelligent actions" to escape from
the upcoming dangerous circumstances. For example, if an agent A knows based
on its local clock that it will collide with an agent B in the next 5 seconds, it
should perform an action such as stopping or quickly finding a safe path to avoid
the collision.
In this paper4, we propose a decentralized real-time reachability approach for
safety verification of a distributed CPS with multiple agents. We are particularly
interested in two types of safety properties. The first one is a local safety property
which specifies the local constraints of the agent operation. For example, each
agent is only allowed to move within a specific region, does not hit any obstacles,
and its velocity needs to be limited to specific range. This type of property does
not require the information of other agents and can be verified locally at run-
time. The second safety property is a global property defined on the states of
multiple agents. Particularly, we consider a peer-to-peer collision free property
and a generalized property where we want to verify if all agents satisfy a set of
linear constraints (on the states of all agents) defining the property, e.g., two
agents do not go into the same region at the same time.
Our decentralized real-time reachability approach works as follows. Each
agent locally and periodically computes the local reachable set from the current
local time to the next T seconds, and then encodes and broadcasts its reachable
set information to the others via a communication network. When the agent
receives a reachable set message, it immediately decodes the message to read the
reachable set information of the sender, and then performs peer-to-peer collision
checking based on its current state and the reachable set of the sender. Verifying
a generalized global property involving the states of N agents is done at the
time an agent receives all needed reachable sets from other agents. Additionally,
the local safety property of the agent is verified simultaneously with the reach-
able set computation process at run-time. The proposed verification approach is
based on an underlying assumption that is, all agents are time-synchronized to
some level of accuracy. This assumption is reasonable as it can be achieved by
using existing time synchronization protocols such as the Network Time Proto-
4 This paper is an extension of [28]
col (NTP). Our approach has successfully verified in real-time the local safety
properties and collision occurrences for a group of quadcopters conducting a
search mission.
The rest of the paper is organized as follows. Section 2 presents briefly the
distributed CPS modeling and its verification problems. Section 3 gives the de-
tail of real-time reachability for single agent and how to use it for real-time
local safety verification. Section 4 addresses the utilization reachable set mes-
sages for checking peer-to-peer collision. Section 5 investigates the global safety
verification problem. Section 6 presents the implementation and evaluation of
our approach via a distributed search application using quadcopters.
2 Problem Formulation
In this paper, we consider a distributed CPS with N agents that can communi-
cate with each other via an asynchronous communication channel.
Communication Model The communication between agents is implemented by
the actions of sending and receiving messages over an asynchronous communi-
cation channel. We formally model this communication model as a single au-
tomaton, Channel, which stores the set of in-flight messages that have been sent,
but are yet to be delivered. When an agent sends a message m, it invokes a
send(m) action. This action adds m to the in-flight set. At any arbitrary time,
the Channel chooses a message in the in-flight set to either delivers it to its re-
cipient or removes it from the set. All messages are assumed to be unique and
each message contains its sender and recipient identities. Let M be the set of
all possible messages used in communication between agents. The sending and
receiving messages by agent i are denoted by Mi,∗ and M∗,i, respectively.
Agent Model The ith agent is modeled as a hybrid automaton [15,25] defined by
the tuple (cid:104)Ai = Vi, Ai,Di,Ti(cid:105), where:
a) Vi is a set of variables consisting of the following: i) a set of continuous vari-
ables Xi including a special variable clki which records the agent's local time,
and ii) a set of discrete variables Yi including the special variable msghisti
that records all sent and received messages. A valuation vi is a function that
associates each vi ∈ Vi to a value in its type. We write val(Vi) for the set of
all possible valuations of Vi. We abuse the notion of vi to denote a state of
Ai, which is a valuation of all variables in Vi. The set Qi
∆= val(Vi) is called
the set of states.
b) Ai is a set of actions consisting of the following subsets: i) a set {sendi(m) m ∈
Mi,∗} of send actions (i.e., output actions), ii) a set {receivei(m) m ∈ M∗,i}
of receive actions (i.e., input actions), and iii) a set Hi of other, ordinary
actions.
c) Di ⊆ val(Vi) × Ai × val(Vi) is called the set of transitions. For a transition
i in short. i) If ai = sendi(m) or receivei(m),
i) ∈ Di, we write vi
(vi, ai, v(cid:48)
ai→ v(cid:48)
then all the components of vi and v(cid:48)
i are identical except that m is added to
msghist in v(cid:48)
i. That is, the agent's other states remain the same on message
sends and receives. Furthermore, for every state vi and every receive action
ai, there must exist a v(cid:48)
i, i.e., the automaton must have
well-defined behavior for receiving any message in any state. ii) If ai ∈ Hi,
then vi.msghist = v(cid:48)
d) Ti is a collection of trajectories for Xi. Each trajectory of Xi is a function
mapping an interval of time [0, t], t ≥ 0 to val(Vi), following a flow rate
that specifies how a real variable xi ∈ Xi evolving over time. We denote the
duration of a trajectory as τdur, which is the right end-point of the interval
t.
i such that vi
i.msghist.
ai→ v(cid:48)
Agent Semantics The behavior of each agent can be defined based on the concept
of an execution which is a particular run of the agent. Given an initial state v0
i ,
an execution αi of an agent Ai is a sequence of states starting from v0
i , defined
as αi = v0
i , . . ., and for each index j in the sequence, the state update from
vj
i to vj+1
is either a transition or trajectory. A state vj
i is reachable if there
exists an executing that ends in vj
i . We denote Reach(Ai) as the reachable set
of agent Ai.
i , v1
i
System Model The formal model of the complete system, denoted as System, is
a network of hybrid automata that is obtained by parallel composing the agent's
models and the communication channel. Formally, we can write, System ∆=
A1(cid:107) . . .AN(cid:107)Channel. Informally, the agent Ai and the communication channel
Channel are synchronized through sending and receiving actions. When the agent
Ai sends a message m ∈ Mi,j to the agent Aj, it triggers the sendi(m) action.
At the same time, this action is synchronized in the Channel automaton by
putting the message m in the in-flight set. After that, the Channel will trigger
(non-deterministically) the receivej(m) action. This action is synchronized in the
agent Aj by putting the message m into the msghistj.
In this paper, we investigate three real-time safety verification problems for
distributed cyber-physical systems as defined in the following.
Problem 1 (Local safety verification in real-time). The real-time local safety ver-
ification problem is to compute online the reachable set Reach(Ai) of the agent
and verify if it violates the local safety property, i.e., checking Reach(Ai) ∩ Ui =
∅?, where Ui (cid:44) Cixi ≤ di, xi ∈ Xi is the unsafe set of the agent.
Problem 2 (Decentralized real-time collision verification). The decentralized real-
time collision verification problem is to reason in real-time whether an agent Ai
will collide with other agents from its current local time ti
c to the computable,
safe time instance in the future Tsaf e based on i) the clock mismatches, and
ii) the exchanging reachable set messages between agents. Formally, we require
c ≤ t ≤ Tsaf e, dij(t) ≥ l, where dij(t) is the distance between agents
that ∀ ti
Ai and Aj at the time t of the agent Ai local clock, and l is the allowable safe
distance between agents.
Problem 3 (Decentralized real-time global safety verification). The decentralized
real-time global safety verification problem is to construct online (at each agent)
the reachable set of all agents globalReach and verify if it violates the global
safety property, i.e., checking globalReach ∩ U = ∅, where U (cid:44) Cx ≤ d, x =
1 , . . . , xT
[xT
N ]T , xi ∈ Xi, is the unsafe set of the whole system.
3 Real-Time Local Safety Verification
The first important step in our approach is, each agent Ai computes forwardly its
reachable set of states from the current local time ti to the next (ti + T ) seconds
which is defined by Ri[ti, ti +T ]. Since there are many variables used in the agent
modeling that are irrelevant in safety verification, we only need to compute the
reachable set of state that is related to the agent's physical dynamics (so called
as motion dynamics) which is defined by a nonlinear ODE xi = f (xi, ui), where
xi ∈ Rn is state vector and ui ∈ Rm is the control input vector. The agent can
switch from one mode to the another mode via discrete transitions, and in each
mode, the control law may be different. When the agent computes its reachable
set, the only information it needs are its current set of states xi(ti) and the
current control input ui(ti). It should be clarified that although the control law
may be different among modes, the control signal ui is updated with the same
control period T i
c . Consequently, ui is a constant vector in each control period.
j = j × Tc, using its local sensors
and GPS, we have the current state of the agent xi. Note that the local sensors
and the provided GPS can only provide the information of interest to some
accuracy, therefore the actual state of the agent is in a set xi ∈ Ii. The control
signal ui is computed based on the state xi and a reference signal, e.g., a set
point denoting where the agent needs to go to, and then computed control signal
is applied to the actuator to control the motion of the agent. From the current
set of states Ii and the control signal ui, we can compute the forward reachable
set of the agent for the next ti
j +T seconds. This reachable set computation needs
runtime ≥ T i
to be completed after an amount of time T i
c ,
a new ui will be updated. The control period T i
c is chosen based on the agent's
motion dynamics, and thus to control an agent with fast dynamics, the control
period T i
c needs to be sufficiently small. This is the source of the requirement
that the allowable run-time for reachable set computation be small.
Assuming that the agent's current time is ti
runtime < T i
c because if T i
To compute the reachable set of an agent in real-time, we use the well-known
face-lifting method [5, 8] and a hyper-rectangle to represent the reachable set.
This method is useful for short-time reachability analysis of real-time systems.
It allows users to define an allowable run-time T i
runtime, and has no dynamic
data structures, recursion, and does not depend on complex external libraries as
in other reachability analysis methods. More importantly, the accuracy of the
reachable set computation can be iteratively improved based on the remaining
allowable run-time.
Algorithm 3.1 describes the real-time reachability analysis for one agent. The
Algorithm works as follows. The time period [ti, ti + T ] is divided by M steps.
Algorithm 3.1 Real-time reachability analysis for agent Ai.
Input: Ii, ui, ti, T , hi, T i
Output: Ri[ti, ti + T ], saf e = true or saf e = uncertain
runtime, Ui
% Current reachable set
% Remaining reach time
% Reach time step
runtime % Remaining run-time
1 > 0) do
while (T i
CR = Ii
saf e = true
T i
while T i
2 = T
2 > 0 do
step = hi
T i
1 = T i
1: procedure Initialization
2:
3:
4: procedure Reachability Analysis
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
% Update remaining runtime
T i
1 = T i
if T i
else
return Ri[ti, ti + T ] = CR, saf e
1 ≤ 0 then break
% Do Single Face Lifting
R, T (cid:48) = SF L(CR, step, T i
2, ui)
CR = R % Update reach set
2 = T (cid:48) % Update remaining reach time
T i
if (CR ∩ Ui (cid:54)= ∅) then: saf e = uncertain
Ri[ti, ti + T ] = CR
1 − (Ai.currentT ime() − ti)
step = hi/2 % Reduce reach time step
The reach time step is defined by hi = T /M . Using the reach time step and
the current set Ii, the face-lifting method performs a single-face-lifting opera-
tion. The results of this step are a new reachable set and a remaining reach
time T i
remainReachT ime < T . This step is iteratively called until the reachable
set for the whole time period of interest [ti, ti + T ] is constructed completely,
i.e., the remaining reach time is equal to zero. Interestingly, with the reach time
step size hi defined above, the face-lifting algorithm may be finished quickly
after an amount of time which is smaller than the allowable run-time T i
runtime
specified by user, i.e., there is still an amount of time called remaining run
time T i
runtime that is available for us to recall the face-lifting
algorithm with a smaller reach time step size, for example, we can recall the
face-lifting algorithm with a new reach time step hi/2. By doing this, the con-
servativeness of the reachable set can be iteratively improved. The core step of
face-lifting method is the single-face-lifting operation. We refer the readers to [5]
for further detail. As mentioned earlier, the local safety property of each agent
can be verified at run-time simultaneously with the reachable set computation
process. Precisely, let Ui (cid:44) Cixi ≤ di be the unsafe region of the ith agent, the
agent is said to be safe from ti to ti + t ≤ ti + T if Ri[ti, ti + t] ∩ Ui = ∅. Since
remainRunT ime < T i
Fig. 1: Timeline for reachable set computing, encoding, transferring, decoding
and collision checking.
the reachable set Ri[ti, ti + t] is given by the face-lifting method at run-time,
the local safety verification problem for each agent can be solved at run-time.
Since the Algorithm 3.1 computes an over-approximation of the reachable set of
each agent in a short time interval, it guarantees the soundness of the result as
described in the following lemma.
[5, 8] The real-time reachability analysis algorithm is sound, i.e.,
Lemma 1.
the computed reachable set contains all possible trajectories of agent Ai from ti
to ti + T .
4 Decentralized Real-Time Collision Verification
Our collision verification scheme is performed based on the exchanged reachable
set messages between agents. For every control period Tc, each agent executes
the real-time reachability analysis algorithm to check if it is locally safe and to
obtain its current reachable set with respect to its current control input. When
the current reachable set is available, the agent encodes the reachable set in a
message and then broadcasts this message to its cooperative agents and listens to
the upcoming messages sent from these agents. When a reachable set message
arrives, the agent immediately decodes the message to construct the current
reachable set of the sender and then performs peer-to-peer collision detection.
The process of computing, encoding, transferring, decoding of the reachable set
along with collision checking is illustrated in Figure 1 based on the agent's local
clock.
Let ti
rs, ti
e, ti
c respectively be the instants that we compute,
encode, transfer, decode the reachable set and do collision checking on the agent
d, and ti
tf , ti
Agent Ai 's (local) time lineGlobal timeAgent Aj's (local) time line≤δjCompute reach setFinish computationEncode reach set≤δiFinish encodingBroadcast reach set Receive reach setDecode reach set Finish decodingCheck collisionReach set computation timeτrsiReach set computation timeτrsjEncoding timeτeiEncoding timeτejReach set transferring timeτtfjReach set transferring timeτtfiDecoding timeτdjDecoding timeτdi+1trsitrsteittfitditcitettftdtctrsjtejttfjtdjtcjElapsed time between computing reachable set and checking collision of agent AjτelapsedjτelapsediElapsed time between computing reach set and checking collision of agent Ai Fig. 2: Useful reachable set.
Ai. Note that these time instants are based on the agent Ai's local clock. The
actual run-times are defined as follows.
τ i
rs = ti
τ i
e = ti
tf ≈ tj
τ i
τ i
d = ti
e − ti
tf − ti
d − ti
c − ti
rs, % reachablet set computation time,
e, % encoding time,
tf , % transferring time,
d, % decoding time.
Note that we do not know the exact transfer time τ i
tf since it depends on
two different local time clocks. The above transfer time formula describes its
approximate value when neglecting the mismatch between the two local clocks.
The actual reachable set computation time is close to the allowable run-time
chosen by user, i.e., τ i
runtime. We will see later that the encoding time
and decoding time are fairly small in comparison with the transferring time, i.e.,
e ≈ τ i
τ i
tf . All of these run-times provide useful information for selecting an
appropriate control period Tc for an agent. However, for collision checking pur-
pose, we only need to consider the time instants that an agent starts computing
reachable set ti
rs and checking collision ti
c.
rs ≈ T i
d (cid:28) τ i
A reachable set message contains three pieces of information: the reachable
set which is a list of intervals, the time period (based on the local clock) in
which this reachable set is valid, i.e., the start time ti
rs + T
and the time instant that this message is sent. Based on the timing information
of the reachable set and the time-synchronization errors, an agent can examine
whether or not a received reachable set contains information about the future
behavior of the sent agent which is useful for checking collision. The usefulness
of the reachable sets used in collision checking is defined as follows.
rs and the end time ti
Definition 1 (Useful reachable sets). Let δi and δj respectively be the time-
synchronization errors of agent Ai and Aj in comparison with the virtual global
time t, i.e, t − δi ≤ ti ≤ t + δi and t − δj ≤ tj ≤ t + δj, where ti and tj are
current local times of Ai and Aj respectively. The reachable sets Ri[ti
rs + T ]
and Rj[tj
rs + T ] of the agent Aj that are available at the agent Ai at time ti
c
rs, tj
rs, ti
Agent i 's (local) time lineGlobal timeAgent j's (local) time linetrsitrstcitctrsjtcj Useful reach set R itrsi+TUseless reach set L itrsi+T>tcitrsi+T≤tci Useful reach set Rjtrsj+TUseless reach set L jtrsj+T≤tcjtrsj+T>tcjFuture timePast time Future time Past time are useful for checking collision between Ai and Aj if:
ti
c < tj
ti
c < ti
rs + T − δi − δj,
rs + T.
(1)
rs, tj
Assume that we are at a time instant where the agent Ai checks if a collision
occurs. This means that the current local time is ti
c. Note that agent Ai and
Aj are synchronized to the global time with errors δi and δj respectively. The
reachable set Rj[tj
rs + T ] is useful if it contains information about the future
behavior of agent Aj under the view of the agent Ai based on its local clock. This
can be guaranteed if we have: tj
c + δi. Additionally, the
current reachablet set of agent Ai contains information about its future behavior
if ti
rs + T + δi + δj,
then the reachable set of Aj contains a past information, and thus it is useless
for checking collision. One interesting case is when tj
c <
tj
rs + T + δi + δj. In this case, we do not know whether the received reachable
set is useful or not.
rs + T as depicted in Figure 2. We can see that if ti
rs + T − δi − δj < ti
rs − δj + T > ti
rs + T ≥ ti
c < ti
c > tj
Remark 1. We note that the proposed approach does not rely on the concept
of Lamport happens-before relation [20] to compute the local reachable set of
each agent. If the agent could not receive reachable messages from others until
a requested time-stamp expires, it still calculates the local reachable set based
on its current state and the state information of other agents in the messages it
received previously. In other words, our method does not require the reachable
set of each agent to be computed corresponding to the ordering of the events
(sending or receiving a message) in the system, but only relies on the local clock
period and the time-synchronization errors between agents. Such implementation
ensures that the computation process can be accomplished in real-time, and is
not affected by the message transmission delay.
The peer-to-peer collision checking procedure depicted in Algorithm 4.2 works
as follows: when a new reachable set message arrives, the receiving agent decodes
the message and checks the usefulness of the received reachable set and its cur-
rent reachable set. Then, the agent combines its current reachable set and the
received reachable set to compute the minimum possible distance between two
agents. If the distance is larger than an allowable threshold l, there is no collision
between two agents in some known time interval in the future, i.e., Tsaf e.
Lemma 2. The decentralized real-time collision verification algorithm is sound.
Proof. From Lemma 1, we know that the received reachable set Rj[tj
rs +
T ] contains all possible trajectories of the agent Aj from tj
rs + T . Also,
the current reachable set of the agent Ai, Ri[ti
rs + T ], contains all possible
trajectories of the agent from ti
rs + T . If those reachable sets are useful,
then they contains all possible trajectories of two agents from ti
c to sometime
rs + T − δi − δj, ti
Tsaf e = min(tj
rs + T ) in the future based on the agent Ai clock.
Therefore, the minimum distance dmin between two agents computed from two
rs to tj
rs to ti
rs, tj
rs, ti
Algorithm 4.2 Decentralized Real-Time Collision Verification at Agent Ai.
Input: l, % safe distance between agents
Output: collision, Tsaf e % collision flag and safe time interval in the future
rs, tj
rs + T ] arrive then
if new message Rj[tj
rs % current reachable set start time
rs + T − δi − δj and ti
decode message
ti
c = Ai.current time() % current time
rs = Ri.ti
ti
c < tj
if ti
compute possible minimum distance dmin between two agents
if dmin > l then
1: procedure Peer-to-Peer Collision Detection
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
Collision = false
Tsaf e = min(tj
Collision = uncertain, Tsaf e = [ ]
rs + T − δi − δj, ti
store the message
rs + T )
c < ti
rs + T then % check usefulness
else
reachable sets is the smallest distance among all possible distances in the time
interval [ti
c, Tsaf e]. Consequently, the collision free guarantee is sound in the time
interval [ti
c, Tsaf e].
We have studied how to use exchanged reachable sets to do peer-to-peer
collision detection. Next, we consider how to verify online the global behavior of
a distributed CPS in decentralized manner.
5 Decentralized Real-Time Global Safety Verification
Definition 2 (Globally useful reachable set.). Consider a distributed CPS
with N agents with time synchronization errors δi, i = 1, 2, . . . , N , a globally
useful reachable set of the whole system under the view of agent Ai based on its
current local time clock ti
c is defined below:
N(cid:94)
globalReach =
Ri[ti
rs, ti
c ≤ t ≤ T + min{ti
i=1
rs + T ] ∧ T ,
rs − δi − δj}, j (cid:54)= i, 1 ≤ j ≤ N ).
T ∆= (ti
(2)
For any time t such that ti
we have Ri(t) ⊆ Ri[ti
possible trajectories of all agents from the current local time ti
the future time defined by T + min{ti
rs−δi−δj} for ∀ 1 ≤ j ≤ N, i (cid:54)= j,
rs + T ],∀i. In other words, globalReach contains all
c of agent Ai to
rs − δi − δj}, j (cid:54)= i, 1 ≤ j ≤ N .
c ≤ t ≤ T +min{ti
rs, ti
It should be noted that to construct a global reachable set, an agent needs
to wait for all messages arrive and then decodes all these messages. This process
may have an expensive computation cost especially when the number of agents
increases. Since this global reachable set is only valid in an interval of time, the
amount of time that is available for verify the global property may be small
and not enough for the agent to perform the global safety verification. Having
additional hardware for handling in parallel the processes of receiving/decoding
messages is a good solution to overcome this challenge.
Using the globally useful reachable set, the global safety verification problem
is equivalent to checking whether the globally useful reachable set intersects with
the global unsafe region defined by U ∆= Cx ≤ d, where x = [xT
N ]T
and xi is the state vector of agent Ai. The procedure for global safety verification
is summarized in Algorithm 5.3.
2 ,··· , xT
1 , xT
Algorithm 5.3 Decentralized Real-Time Global Safety Verification at Agent
Ai.
Input: U, % global unsafe constraints
Output: global saf e, Tglobal saf e % global safe flag and safe time interval in the
future
global saf e = true
% global safety flag
1: procedure Initialization
2:
3: procedure Global Safety Verification
4:
if all useful messages are available then
5:
6:
7:
8:
9:
10:
11:
12:
13:
ti
c = Ai.current time()
recheck if all messages are still useful
construct globally useful reach set globalReach
if (globalReach ∩ U (cid:54)= ∅) then
global saf e = uncertain
Tglobal saf e = [ ]
global saf e = true
Tglobal saf e = T + min{ti
else
rs − δi − δj}, j (cid:54)= i, 1 ≤ j ≤ N
Lemma 3. The decentralized real-time global safety verification algorithm is
sound.
Proof. Similar to Lemma 2, the soundness of the verification algorithm is guar-
antee because of the soundness of the globally useful reachable set containing
all possible trajectories of all agents at any time t ∈ T , where T ∆= (ti
c ≤ t ≤
T + min{ti
rs − δi − δj}, j (cid:54)= i, 1 ≤ j ≤ N ).
6 Case study
The decentralized real-time safety verification for distributed CPS proposed in
this paper is implemented in Java as a package called drreach. This package is
Fig. 3: Distributed Search Application Using Quadcopters.
currently integrated as a library in StarL, which is a novel platform-independent
framework for programming reliable distributed robotics applications on An-
droid [22]. StarL is specifically suitable for controlling a distributed network
of robots over WiFi since it provides many useful functions and sophisticated
algorithms for distributed applications. In our approach, we use the reliable com-
munication network of StarL which is assumed to be asynchronous and peer-to-
peer. There may be message dropouts and transmission delays; however, every
message that an agent tries to send is eventually delivered with some time guar-
antees. All experimental results of our approach are reproducible and available
online at: http://www.verivital.com/rtreach/.
6.1 Experiment setup
We evaluate the proposed approach via a distributed search application using
quadcopters5 in which each quadcopter executes its search mission provided by
users as a list of way-points depicted in Figure 3. These quadcopters follow
the way-points to search for some specific objects. For safety reasons, they are
required to work only in a specific region defined by users. In this case study, the
quadcopters are controlled to operate at the same constant altitude. It has been
shown from the experiments that the proposed approach is promisingly scalable
as it works well for a different number of quadcopters. We choose to present in
this section the experimental results for the distributed search application with
eight quadcopters.
The first step in our approach is locally computing the reachable set of each
quadcopter using face-lifting method. The quadcopter has nonlinear motion dy-
namics given in Equation 3 in which θ, φ, and ψ are the pitch, roll, and yaw
angles, f = Σ4
i=1Ti is the sum of the propeller forces, m is the mass of the
5 A video recording is available at: https://youtu.be/YC_7BChsIf0
𝓐𝟏𝓐𝟐𝓐𝟔𝓐𝟕𝓐𝟖𝓐𝟑𝓐𝟓𝓐𝟒𝓐𝟏𝓐𝟐𝓐𝟑𝓐𝟒𝓐𝟓𝓐𝟔𝓐𝟖𝓐𝟕 quadcopter and g = 9.81m/s2 is the gravitational acceleration constant. As the
quadcopter is set to operate on a constant altitude, we have z = 0 which yields
cos(θ)cos(φ) . Let vx and vy be the velocities of a
the following constraint: f =
quadcopter along with x- and y- axes. Using the constraint on the total force,
the motion dynamics of the quadcopter can be rewritten as a 4-dimensional
nonlinear ODE as depicted in Equation 4.
mg
x =
y =
f
m
f
m
f
m
(sin(ψ)sin(φ) + cos(ψ)sin(θ)cos(φ)),
(sin(ψ)sin(θ)cos(φ) − sin(φ)cos(ψ)),
cos(θ)cos(φ) − g,
x = vx,
vx = gtan(θ),
y = vy,
(4)
(3)
tan(φ)
cos(θ)
.
vy = g
z =
A PID controller is designed to control the quadcopter to move from its cur-
rent position to desired way-points. Details about the controller parameters can
be found in the available source code. The PID controller has a control period of
Tc = 200 milliseconds. In every control period, the control inputs pitch (θ) and
roll (φ) are computed based on the current positions of the quadcopter and the
current target position (i.e., the current way-point it needs to go). Using the con-
trol inputs, the current positions and velocities given from GPS and the motion
dynamics of the quadcopter, the real-time reachable set computation algorithm
(Algorithm 3.1) is executed inside the controller. This algorithm computes the
reachable set of a quadcopter from its current local time to the next T = 2 sec-
onds. The allowable run-time for this algorithm is Truntime = 10 milliseconds.
The local safety property is verified by the real-time reachable set computation
algorithm at run-time. The computed reachable set is then encoded and sent
to another quadcopter. When a reachable set message arrives, the quadcopter
decodes the message to reconstruct the current reachable set of the sender. The
GPS error is assumed to be 2%. The time-synchronization error between the
quadcopters is δ = 3 milliseconds. We want to verify in real-time: 1) local
safety property for each quadcopter; 2) collision occurrence; and 3) geospatial
free property. The local safety property is defined by vx ≤ 500, i.e., the maximum
allowable velocities along the x-axis of two arbitrary quadcopters are not larger
than 500m/s. The collision is checked using the minimum allowable distance
between two arbitrary quadcopters dmin = 100. The geospatial free property
requires that the some quadcopters never go into a specific region at the same
time.
6.2 Verifying local safety property and collision occurrence
Figure 4 presents a sample of a sequence of events happening in the distributed
search application. One can see that each quadcopter can determine based on its
local clocks if there is no collision to some known time in the future. In addition,
the local safety property can also be verified at run-time. For example, in the
figure, the quadcopter 1 receives a reachable set message from the quadcopter
0 which is valid from 17 : 29 : 49.075 to 17 : 29 : 51.074 of the quadcopter 0's
Fig. 4: A sample of events for verifying local safety property and collision occur-
rence.
Fig. 5: One sample of the reachable sets of eight quadcopters in [0, 2s] time
interval and their interval hulls.
clock. After decoding this message, taking into account the time-synchronization
error δ, quadcopter 1 realizes that the received reachable set message is useful for
checking collision for the next 1.645 seconds of its clock. After checking collision,
quadcopter 1 knows that it will not collide with the quadcopter 0 in the next
1.645 seconds (based on its clock).
It should be noted that we can intuitively verify the collision occurrences by
observing the intermediate reachable sets of all quadcopters and their interval
hulls. The intermediate reachable sets of the quadcopters in every [0, 2s] time
interval computed by the real-time reachable set computation algorithm (i.e.,
Algorithm 3.1) is described in Figure 5. The zoom plot within the figure presents
a very short-time interval reachable set of the quadcopters. We note that the
intermediate reachable set of a quadcopter is represented as a list of hyper-
rectangles and is used for verifying the local safety property at run-time. The
reachable set that is sent to another quadcopter is the interval hull of these
hyper-rectangles. The intermediate reachable set cannot be transferred via a
network since it is very large (i.e., hundreds of hyper-rectangles). The interval
hull of all hyper-rectangles contained in the intermediate reachable set covers all
possible trajectories of a quadcopter in the time interval of [0, 2s]. Therefore, it
can be used for safety verification. One may question why we use the interval
hull instead of using the convex hull of the reachable set since the former one
results in a more conservative result. The reason is that we want to perform the
safety verification online, convex hull of hundreds of hyper-rectangles is a time-
consuming operation. Therefore, in the real-time setting, interval hull operation
is a suitable solution. From the figure, we can see that the interval hulls of the
reachable set of all quadcopters do not intersect with each other. Therefore, there
is no collision occurrence (in the next 2 seconds of global time).
Time
Quad. 1 Quad. 2 Quad. 3 Quad. 4 Quad. 5 Quad. 6 Quad. 7 Quad. 8
Ecoding Time τe (ms)
0.058
0.055
0.0553
0.0525
0.0557
0.0583
0.0584
0.0597
Decoding Time τd (ms)
0.0169
0.0193
0.0197
0.019
0.0210
0.0181
0.0177
0.022
Transferring Time τtf (ms)
Collision Checking Time τc (ms)
2.64
0.04
2.48
0.05
1.42
0.07
1.11
0.05
1.12
0.03
1.08
0.07
1.05
0.07
1.13
0.14
Total Verification Time V T (ms) 28.9363
19.1037
Table 1: The average encoding time τe, decoding time τd, transferring time τtf ,
collision checking time τc and total verification time V T of the quadcopters.
18.2527
18.0223
18.235
27.9
20.6232
18.3055
Since we implement the decentralized real-time safety verification algorithm
inside the quadcopter's controller, it is important to analyze whether or not
the verification procedure affects the control performance of the controller. To
reason about this, we measure the average encoding, decoding, transferring and
collision checking times for all quadcopters using 100 samples which are pre-
sented in Table 1. We note that the transferring time τtf is the average time for
one message transferred from other quadcopters to the ith quadcopter. It can
be seen that the encoding, decoding and collision checking times at each quad-
copter constitute a tiny amount of time. The total verification time is the sum
of the reachable set computation, encoding, transferring, decoding and collision
e + (N − 1) × (τ i
checking times. Note that the allowable runtime for reachable set computation
algorithm is specified by users as Truntime = 10 milliseconds. Therefore, the
(average) total time for the safety verification procedure on each quadcopter is
V Ti = Truntime + τ i
c), where i = 1, 2, . . . , N , and N
is the number of quadcopters. As shown in the Table, the (average) total veri-
fication time for each quadcopter is small (< 30 milliseconds), compared to the
control period Tc = 200 milliseconds. Besides, from the experiment, we observe
that the computation time for the control signal of the PID controller τ i
control
(not presented in the table) is also small, i.e., from 5 to 10 milliseconds. Since
V Ti + τ i
control < Tc/4 = 50 milliseconds, we can conclude that the verification
procedure does not affect the control performance of the controller.
tf + τ i
d + τ i
Interestingly, from the verification time formula, we can estimate the range
of the number of agents that the decentralized real-time verification procedure
can deal with. The idea is that, in each control period Tc, after computing
the control signal, the remaining time bandwidth Tc − τcontrol can be used for
verification. Let ¯τe(τ e), ¯τtf (τ tf ), ¯τd(τ d), ¯τc(τ c) be the maximum (minimum)
encoding, transferring, decoding and collision checking times on a quadcopter,
¯τcontrol(τ control) be the maximum (minimum) control signal computation time
for each control period Tc, then the number of agents that the decentralized
real-time safety verification procedure can deal with (with assumption that the
communication network works well) satisfies the following constraint:
Tc − ¯τcontrol − Truntime − ¯τe
+ 1 ≤ N ≤ Tc − τ control − Truntime − τ e
+ 1. (5)
¯τtf + ¯τd + ¯τc
τ tf + τ d + τ c
Let consider our case study, from the Table, we assume that ¯τe = 0.0597,
τ e = 0.0525, ¯τtf = 2.64, τ tf = 1.05, ¯τd = 0.022, τ d = 0.0169, ¯τc = 0.14,
τ c = 0.03 milliseconds. Also, we assume that ¯τcontrol = 10 and τ control = 5
milliseconds. We can estimate theoretically the number of quadcopters that our
verification approach can deal with is 64 ≤ N ≤ 168.
6.3 Verifying geospatial free property
To illustrate how our approach verifies the global behavior of a distributed CPS,
we consider the geospatial free property which requires that the some (or all)
quadcopters never go into a specific region at the same time. For simplification,
we reconsider the distributed search application with two quadcopters (quad 1
and quad 2) whose forbidden region is defined by 900 < x0 < 1200∧ 900 < x1 <
1200. Figure 6 describes a sample of events describing that the quadcopter 2 can
verify based on its local clock that it will not collide with the quadcopter 1 and
the global geospatial free property is guarantee in the next 1.838 seconds.
7 Discussion
The current implementation of our approach deploys the safety verifier of each
agent inside the controller, and a single thread is used to execute the control
Fig. 6: A sample of events for verifying geospatial free property.
Fig. 7: Software architecture for deploying decentralized real-time safety verifi-
cation approach on a real platform.
and verification tasks. The main drawback of this implementation is that it may
decrease the overall performance of the controller and even cause the controller to
crash. To prevent this happens, in practice, the controller and verifier should be
implemented in two separate software components. In this case, the computation
burden for safety checks in the verifier does not affect the performance of the
quadcopter0 computes it reach set from 2019-09-16 17:26:20.957 to 2019-09-16 17:26:22.956quadcopter0 encodes its reach set to send out in 0.017749 millisecondsquadcopter0 broadcasts its reach set to othersquadcopter0 may violates its local safety specification at time 2019-09-16 17:26:22.956quadcopter1 receives reach set (hull) from quadcopter0Time for transferring this reach set over network is around (not considering clock mismatch) 66 millisecondsDecoding message from quadcopter0 takes 0.029057 milliseconds Reach set (hull) of quadcopter0 that is valid from 2019-09-16 17:26:20.957 to 2019-09-16 17:26:22.956 of its local time is:dim = 0 -> [89.18, 280.21] dim = 1 -> [39.31, 58.43] dim = 2 -> [804.58, 1240.06] dim = 3 -> [91.37, 126.26] Current reach set (hull) of quadcopter1 that is valid from 2019-09-16 17:26:20.119 to 2019-09-16 17:26:22.118 of its local time is:dim = 0 -> [1946.28, 2060.66] dim = 1 -> [25.17, 33.29] dim = 2 -> [787.92, 1566.04] dim = 3 -> [184.97, 203.86] Current local time of quadcopter1 is 2019-09-16 17:26:21.112Useful time for checking collision and global safety property is 1838 millisecondsThe received reachable set from quadcopter0 is useful quadcopter1 will not collide with quadcopter0 in the next 1.838 secondsThe geospatial free property is guarantee in the next 1.838 secondsComplex ControllerVerifierPlanerPlantSensingSafe ControllerMapDecision MakerActuator CommandsEncoderDecoderSafety CheckerDecoder 1...Decoder 2Decoder NBroadcast ReachSet messagesComing ReachSet MessagesChecker 1Checker 2Checker N...Global Checker Reachable Set CalculatorLocal Checker Global Safety SpecificationsLocal Safety SpecificationsCollision Checkers controller. The control task and the verification task can be executed efficiently
in parallel as depicted in Figure 7. More importantly, this software architecture
adopts the architecture of a fault-tolerant system [14] to prevent the propagation
of failure from one component to others. It also benefits the use of simplex-
architecture for safety control in the case of dangerous circumstances.
As shown in Figure 7, the verifier component consists of four sub-components
including reachable set calculator, encoder, decoder, and safety checker. These
sub-components should also be implemented conveniently for parallel execution.
The local safety property is verified inside the reachable set calculator at run-
time. As the number of reachable set messages needs to be decoded increases
with the number of participating agents, it is necessary to have multiple decoders
working in parallel. These decoders listen to upcoming reachable set messages
on different ports assigned to them by the verifier and immediately decode any
arrived message. This parallel decoding helps to reduce the decoding time signif-
icantly. The decoded reachable sets are then sent to the safety checker containing
multiple checkers run in parallel in which each checker is responsible for checking
collision between the agent with another. The ith checker and the ith decoder
is a pair worker, i.e., the checker only waits for the decoded reachable set of
its corresponding co-worker. Therefore, the pair to pair collision detection task
can be done very quickly. The safety checker also has a global checker which is
responsible for checking global properties. The global checker is only triggered
when the decoder component finishes decoding all arrived reachable set mes-
sages. For this reason, having parallel working decoders is essential to speed up
the overall verification time which is required to be very small to work in the
real-time setting.
To analyze how fast our verification technique can achieved with the proposed
software architecture, let ¯τrs, ¯τe, ¯τtf and ¯τd respectively be the worst case times
of reachable set computation, encoding, transferring and decoding, ¯τcc and ¯τgc
be the worst case times of peer-to-peer collision detection and global safety
verification. For a system with N agents, the total worst-case verification time
is ¯τtotal = ¯τrs + ¯τe + ¯τtf + ¯τd + ¯τcc + ¯τgc. If we do the verification in sequential
way, i.e., using only one port for reachable set communication and one checker
for all peer-to-peer collision detection and global safety verification, the total
worst-case verification is: ¯τ∗
total = ¯τrs + ¯τe + ¯τtf + N ¯τd + N ¯τcc + ¯τgc >> ¯τtotal.
Scalability. From the above discussion, one can see that the software ar-
chitecture plays an important role when we implement our approach in a real
platform. In practice, if each participating agent has the powerful hardware for
communication and computation, and the software for our approach is imple-
mented in a parallel manner as proposed above, then the worst-case verification
time does not depend on the number of agents in the system. Therefore, our
decentralized real-time safety verification approach is scalable for systems with
a large number of agents. Also, the proposed software architecture is especially
useful in the case that there are losses of reachable set messages. In this hazardous
situation, the agent still has some partial information to check if a collision oc-
curs based on the available, reachable set messages. Therefore, the planner still
can re-perform path planning algorithm based on the current information and
past information it has to find the safest path for the agent for this incomplete
information situation.
8 Related Work
Our work is inspired by the static and dynamic analysis of timed distributed
traces [10] and the real-time reachability analysis for verified simplex design [5].
The former one proposes a sound method of constructing a global reachable set
for a distributed CPS based on the recorded traces and time synchronization
errors of participating agents. Then the global reachable set is used to verify a
global property using Z3 [9]. This method can be considered to be a centralized
analysis where the reachable set of the whole system is constructed and verified
by one analyzer. Such a verification approach is offline which is fundamentally
different from our approach as we deal with online verification in a decentralized
manner. Our real-time verification method borrows the face-lifting technique
developed in [5] and applies it to a distributed CPS.
Another interesting aspect of real-time monitoring for linear systems was
recently published in [7]. In this work, the authors proposed an approach that
combines offline and online computation to decide if a given plant model has
entered an uncontrollable state which is a state that no control strategy can be
applied to prevent the plant go to the unsafe region. This method is useful for a
single real-time CPS, but not a distributed CPS with multiple agents.
Additionally, there has been other significant works for verifying distributed
CPS. Authors of [11, 26, 30] presented a real-time software for distributed CPS
but did not perform a safety verification of individual components and a whole
system. The works presented in [2,17, 19] can be used to verify distributed CPS,
but they do not consider a real-time aspect. An interesting work proposed in [24]
can formally model and verify a distributed car control system against several
safety objectives such as collision avoidance for an arbitrary number of cars.
However, it does not address the verification problem of distributed CPS in a
real-time manner. The novelty of our approach is that it can over-approximate
of the reachable set of each agent whose dynamics are non-linear with a high
precision degree in real-time.
The most related work to our scheme was recently introduced in [23]. The au-
thors proposed an online verification using reachability analysis that can guaran-
tee safe motion of mobile robots with respective to walking pedestrians modeled
as hybrid systems. This work utilizes CORA toolbox [1] to perform reachability
analysis while our work uses a face-lifting technique. However, this work does not
consider the time-elapse for encoding, transferring and decoding the reachable
set messages between each agent, which play an important role in distributed
systems.
9 Conclusion and Future Work
We have proposed a decentralized real-time safety verification method for dis-
tributed cyber-physical systems. By utilizing the timing information and the
reachable set information from exchanged reachable set messages, a sound guar-
antee about the safety of the whole system is obtained for each participant
based on its local time. Our method has been successfully applied for a dis-
tributed search application using quadcopters built upon StarL framework. The
main benefit of our approach is that it allows participants to take advantages of
formal guarantees available locally in real-time to perform intelligent actions in
dangerous situations. This work is a fundamental step in dealing with real-time
safe motion/path planing for distributed robots. For future work, we seek to
deploy this method on a real-platform and extend it to distributed CPS with
heterogeneous agents where the agents can have different motion dynamics and
thus they have different control periods. In addition, the scalability of the pro-
posed method can be improved by exploiting the benefit of parallel processing,
i.e., each agent handles multiple reachable set messages and checks for collision
in parallel.
Acknowledgments
The material presented in this paper is based upon work supported by the Air
Force Office of Scientific Research (AFOSR) through contract number FA9550-
18-1-0122 and the Defense Advanced Research Projects Agency (DARPA) through
contract number FA8750-18-C-0089. The U.S. Government is authorized to re-
produce and distribute reprints for Government purposes notwithstanding any
copyright notation thereon. The views and conclusions contained herein are those
of the authors and should not be interpreted as necessarily representing the offi-
cial policies or endorsements, either expressed or implied, of AFOSR or DARPA.
References
1. Althoff, M.: An introduction to cora 2015. In: Proc. of the Workshop on Applied
Verification for Continuous and Hybrid Systems (2015)
2. Bae, K., Krisiloff, J., Meseguer, J., Olveczky, P.C.: Designing and verifying dis-
tributed cyber-physical systems using multirate pals: An airplane turning control
system case study. Science of Computer Programming 103, 13 -- 50 (2015)
3. Bak, S., Duggirala, P.S.: Hylaa: A tool for computing simulation-equivalent reach-
ability for linear systems. In: Proceedings of the 20th International Conference on
Hybrid Systems: Computation and Control. pp. 173 -- 178. ACM (2017)
4. Bak, S., Duggirala, P.S.: Simulation-equivalent reachability of large linear systems
with inputs. In: International Conference on Computer Aided Verification. pp.
401 -- 420. Springer (2017)
5. Bak, S., Johnson, T.T., Caccamo, M., Sha, L.: Real-time reachability for verified
simplex design. In: Real-Time Systems Symposium (RTSS), 2014 IEEE. pp. 138 --
148. IEEE (2014)
6. Chen, X., ´Abrah´am, E., Sankaranarayanan, S.: Flow*: An analyzer for non-linear
hybrid systems. In: International Conference on Computer Aided Verification. pp.
258 -- 263. Springer (2013)
7. Chen, X., Sankaranarayanan, S.: Model predictive real-time monitoring of linear
systems. In: Real-Time Systems Symposium (RTSS), 2017 IEEE. pp. 297 -- 306.
IEEE (2017)
8. Dang, T., Maler, O.: Reachability analysis via face lifting. In: Hybrid Systems:
Computation and Control (HSCC '98). pp. 96 -- 109. Springer (1998), lNCS 1386
9. De Moura, L., Bjørner, N.: Z3: An efficient smt solver. In: International conference
on Tools and Algorithms for the Construction and Analysis of Systems. pp. 337 --
340. Springer (2008)
10. Duggirala, P.S., Johnson, T.T., Zimmerman, A., Mitra, S.: Static and dynamic
analysis of timed distributed traces. In: Real-Time Systems Symposium (RTSS),
2012 IEEE 33rd. pp. 173 -- 182. IEEE (2012)
11. Eidson, J.C., Lee, E.A., Matic, S., Seshia, S.A., Zou, J.: Distributed real-time
software for cyber -- physical systems. Proceedings of the IEEE 100(1), 45 -- 59 (2012)
12. Frehse, G., Le Guernic, C., Donz´e, A., Cotton, S., Ray, R., Lebeltel, O., Ripado, R.,
Girard, A., Dang, T., Maler, O.: Spaceex: Scalable verification of hybrid systems.
In: Computer Aided Verification. pp. 379 -- 395. Springer (2011)
13. Girard, A., Le Guernic, C., Maler, O.: Efficient computation of reachable sets of
linear time-invariant systems with inputs. In: Hybrid Systems: Computation and
Control, pp. 257 -- 271. Springer (2006)
14. Goodloe, A.E., Pike, L.: Monitoring distributed real-time systems: A survey and
future directions (2010)
15. Henzinger, T.A.: The theory of hybrid automata. In: IEEE Symposium on Logic
in Computer Science (LICS). p. 278. IEEE Computer Society, Washington, DC,
USA (1996)
16. Henzinger, T.A., Ho, P.H., Wong-Toi, H.: Hytech: A model checker for hybrid
systems. In: Computer aided verification. pp. 460 -- 463. Springer (1997)
17. Johnson, T.T., Mitra, S.: Parametrized verification of distributed cyber-physical
systems: An aircraft landing protocol case study. In: Cyber-Physical Systems (IC-
CPS), 2012 IEEE/ACM Third International Conference on. pp. 161 -- 170. IEEE
(2012)
18. Kong, S., Gao, S., Chen, W., Clarke, E.: dreach: δ-reachability analysis for hybrid
systems pp. 200 -- 205 (2015)
19. Kumar, P., Goswami, D., Chakraborty, S., Annaswamy, A., Lampka, K., Thiele,
L.: A hybrid approach to cyber-physical systems verification. In: Proceedings of
the 49th Annual Design Automation Conference. pp. 688 -- 696. ACM (2012)
20. Lamport, L.: Time, clocks, and the ordering of events in a distributed system.
Communications of the ACM 21(7), 558 -- 565 (1978)
21. Le Guernic, C., Girard, A.: Reachability analysis of hybrid systems using support
functions. In: Computer Aided Verification. pp. 540 -- 554. Springer (2009)
22. Lin, Y., Mitra, S.: Starl: Towards a unified framework for programming, simulating
and verifying distributed robotic systems. CoRR abs/1502.06286 (2015), http:
//arxiv.org/abs/1502.06286
23. Liu, S.B., Roehm, H., Heinzemann, C., Lutkebohle, I., Oehlerking, J., Althoff,
M.: Provably safe motion of mobile robots in human environments. In: Intelligent
Robots and Systems (IROS), 2017 IEEE/RSJ International Conference on. pp.
1351 -- 1357. IEEE (2017)
24. Loos, S.M., Platzer, A., Nistor, L.: Adaptive cruise control: Hybrid, distributed,
and now formally verified. In: International Symposium on Formal Methods. pp.
42 -- 56. Springer (2011)
25. Lynch, N., Segala, R., Vaandrager, F., Weinberg, H.B.: Hybrid i/o automata.
Springer (1996)
26. Tang, Q., Gupta, S.K., Varsamopoulos, G.: A unified methodology for scheduling in
distributed cyber-physical systems. ACM Transactions on Embedded Computing
Systems (TECS) 11(S2), 57 (2012)
27. Tran, H.D., Nguyen, L.V., Hamilton, N., Xiang, W., Johnson, T.T.: Reachability
analysis for high-index linear differential algebraic equations (daes). In: 17th Inter-
national Conference on Formal Modeling and Analysis of Timed Systems (FOR-
MATS'19). Springer International Publishing (August 2019)
28. Tran, H.D., Nguyen, L.V., Musau, P., Xiang, W., Johnson, T.T.: Decentralized
real-time safety verification for distributed cyber-physical systems. In: P´erez, J.A.,
Yoshida, N. (eds.) Formal Techniques for Distributed Objects, Components, and
Systems (FORTE'19). pp. 261 -- 277. Springer International Publishing, Cham (June
2019)
29. Tran, H.D., Nguyen, L.V., Xiang, W., Johnson, T.T.: Order-reduction abstractions
for safety verification of high-dimensional linear systems. Discrete Event Dynamic
Systems 27(2), 443 -- 461 (2017)
30. Zhang, Y., Gill, C., Lu, C.: Reconfigurable real-time middleware for distributed
cyber-physical systems with aperiodic events. In: Distributed Computing Systems,
2008. ICDCS'08. The 28th International Conference on. pp. 581 -- 588. IEEE (2008)
|
cs/0612012 | 1 | 0612 | 2006-12-04T03:20:25 | Geographic Gossip on Geometric Random Graphs via Affine Combinations | [
"cs.MA",
"cs.IT",
"cs.IT"
] | In recent times, a considerable amount of work has been devoted to the development and analysis of gossip algorithms in Geometric Random Graphs. In a recently introduced model termed "Geographic Gossip," each node is aware of its position but possesses no further information. Traditionally, gossip protocols have always used convex linear combinations to achieve averaging. We develop a new protocol for Geographic Gossip, in which counter-intuitively, we use {\it non-convex affine combinations} as updates in addition to convex combinations to accelerate the averaging process. The dependence of the number of transmissions used by our algorithm on the number of sensors $n$ is $n \exp(O(\log \log n)^2) = n^{1 + o(1)}$. For the previous algorithm, this dependence was $\tilde{O}(n^{1.5})$. The exponent 1+ o(1) of our algorithm is asymptotically optimal. Our algorithm involves a hierarchical structure of $\log \log n$ depth and is not completely decentralized. However, the extent of control exercised by a sensor on another is restricted to switching the other on or off. | cs.MA | cs |
Geographic Gossip on Geometric Random Graphs via
Affine Combinations
Hariharan Narayanan
Department of Computer Science, University of Chicago
[email protected]
February 2, 2018
Abstract
In recent times, a considerable amount of work has been devoted to the develop-
ment and analysis of gossip algorithms in Geometric Random Graphs. In a recently
introduced model termed "Geographic Gossip," each node is aware of its position but
possesses no further information. Traditionally, gossip protocols have always used
convex linear combinations to achieve averaging. We develop a new protocol for Geo-
graphic Gossip, in which counter-intuitively, we use non-convex affine combinations as
updates in addition to convex combinations to accelerate the averaging process. The
dependence of the number of transmissions used by our algorithm on the number of
sensors n is n exp(O(log log n)2) = n1+o(1). For the previous algorithm, this depen-
dence was O(n1.5). The exponent 1+ o(1) of our algorithm is asymptotically optimal.
Our algorithm involves a hierarchical structure of log log n depth and is not completely
decentralized. However, the extent of control exercised by a sensor on another is re-
stricted to switching the other on or off.
1 Introduction
Geometric Random Graphs have become an accepted model for wireless ad hoc and sensor
networks. Due to applications in distributed sensing, a significant amount of effort has
been directed towards developing energy efficient algorithms for information exchange on
these graphs. The problem of distributed averaging has been studied intensively because
it appears in several applications such as estimation on ad hoc networks, and encapsulates
many of the difficulties faced in asynchronous distributed computation. Let v1, . . . , vn be n
points independently chosen uniformly at random from a unit square in R2. A Geometric
Random Graph G(n, r) is obtained from these points by connecting any two points within
1
Euclidean distance r. A Gossip Algorithm is an averaging algorithm that, after a certain
number of information exchanges and updates, leaves each node with a value close to the
average of all the originally held values.
1.1 Related Work
There is an extensive body of work surrounding the subject of gossip algorithms in various
contexts. Here, we only survey the results relevant in a narrow sense to the question under
consideration.
Gupta and Kumar [4] gave conditions under which G(n, r) is connected with high probability
n ) in order that G(n, r) be connected with
(w.h.p.). It is sufficient that r scales as Ω(q log n
probability greater than 1 − n−Θ(1).
A distributed Gossip Algorithm for arbitrary graphs was presented by Boyd et al [1]. In
this algorithm, when the clock of a sensor s ticks, s sends its value xs to a sensor v chosen
uniformly at random from its neighbors, and receives the value xv of v. Thereafter s and v
set their values to xs+xv
. The dependence of the number of transmissions required by this
algorithm on n is O(n2). The performance was related to the mixing time of the natural
random walk on that graph. In fact they showed that if the connectivity graph is G, the
number of transmissions made in the course of the algorithm is Θ(nTmix(G)), where Tmix(G)
is the mixing time of G.
2
In the standard framework for modeling sensor networks, n sensors are placed at random
n ). One does not assume
that a sensor possesses any information about its own location. In this model, the number
of transmissions that the best known algorithm uses is O(n2) as described above.1
on a unit square (cid:3) and have a radius of connectivity r = Θ(q log n
A more powerful model was proposed by Dimakis et al [5], wherein each sensor is aware of
its own location with reference to (cid:3) , but possess no further information. It is mentioned
in [5] that this is reasonable in typical scenarios. With this model, by exploiting geographic
information, they were able to provide an algorithm that requires O(n1.5) transmissions. In
their algorithm, each node exchanges its value with the node nearest to a position chosen
randomly on (cid:3), and both nodes replace their values by the average as in the algorithm of
Boyd et al [1]. Rejection sampling is used to make the distribution roughly uniform on nodes.
The routing takes O(√n) hops w.h.p, but since the mixing time on the complete graph is
O(1), one obtains an algorithm using O(n1.5) transmissions, which is an improvement over
[1] by a factor of O(√n).
A natural approach to obtaining more efficient algorithms would be to engage in long-range
1In using O, we ignore polylogarithmic factors and depending on context, the dependence on parameters
other than n.
2
information exchanges less frequently than short-range ones. However, it appears that the
benefit derived from an improved mixing time with long-range transmissions more than
compensates for the additional cost in terms of hops for a long-range routing. Due to this
fact, simply altering the probability distribution with which a node picks targets seems to
be counterproductive.
1.2 Our Contribution
An affine combination of two vectors a and b has the form αa + (1− α)b. Unlike the case of
convex combinations, α need not belong to [0, 1]. We introduce counter-intuitive update rules
which are affine combinations rather than convex combinations (with coefficients possibly as
large as Ω(√n)) to achieve faster averaging. The total number of transmissions used by the
proposed algorithm in order that the ℓ2-distance of the output from the average diminish by
a multiplicative factor of ǫ w.h.p, is n exp(O((log log n) log log n
log log n )
the number of transmissions is n1+o(1). The exponent 1+o(1) is asymptotically optimal, since
every node must make at least one transmission for an averaging algorithm to work. Like
previous algorithms, ours makes packet exchanges with random nodes. Due to the instability
introduced into the system by the use of non-convex combinations, for the present analysis
to hold, a certain amount of control needs to be exercised and our algorithm is not truly
decentralized. However, the extent of control exerted by any sensor on another is restricted
to switching the other on or off.
ǫ )). When ǫ = exp(n
o(1)
2 Preliminaries
The standard model for a sensor network is as follows. We assume that each node or sensor
has a clock that is a Poisson process with rate 1, and that these processes are independent.
This model is equivalent to having a single clock that is Poisson of rate n, and assigning
clock ticks to nodes uniformly at random. We assume that the time units are adjusted so
communication time between any two adjacent nodes is insignificant in comparison with the
length of an average time slot n−1. Our algorithm involves packet forwarding when two
non-adjacent nodes communicate. We shall assume that the time taken to forward a packet
is also insignificant in comparison with n−1, and that a single packet exists in the network
in each time slot w.h.p.. We assume some limited computational power, which amounts to
memory of logarithmic size, and the ability to do floating point computations.
For our purposes, a Geometric Random Graph is defined in the following way. Let v1, . . . , vn
be n points independently chosen uniformly at random from a unit square in R2. A Geometric
Random Graph G(n, r) is obtained from these points by connecting any two points within
Euclidean distance r.
3
2.1 Problem Statement
Let node vi for i = 1, . . . , n hold a value xi(t) at the tth global clock tick, the initial values
being xi(0). Without loss of generality, we assume x(0) = 0. Given ǫ, δ > 0, the task
is to design an algorithm such that kx(t)k < ǫkx(0)k for all possible choices of x(0) with
probability > 1− δ. The cost of the algorithm is the expected number of transmissions made
until t.
nectivity r(n) = Θ(q log n
In the rest of the paper, we shall make the standard assumption that the radius of con-
n ) (eg [5].) Under this assumption, the probability of the graph
G(n, r) being disconnected is Ω(n−O(1)), for an appropriate constant a. As a consequence,
it is not possible to drive δ below n−O(1). For this reason, in the analysis, we shall assume
that δ = n−O(1). On the other hand ǫ can be made arbitrarily small by running the aver-
aging algorithm for a sufficiently long interval of time. In this paper, we shall assume that
log 1
log log n . This does not allow ǫ to be exponentially small but permits it to be the
reciprocal of a quasipolynomial. A sufficiently large constant a will appear in the parameters
of our algorithm described later.When we use the term high probability, we shall mean with
probability 1 − n−Θ(1).
ǫ = n
o(1)
3 Overview of Algorithm
Let (cid:3) be the unit square in which the n sensors are randomly placed. Let the initial values
carried by sensors be xi(0), for i = 1 to n. We consider a partition of (cid:3) into ∼ n1/2 smaller
squares (cid:3)i. Let (cid:3)i contain #((cid:3)i) sensors. Let time(n) represent the expected number of
transmissions until kx(t)k ≤ ǫkx(0)k w.h.p., where ǫ is some function of n that we shall not
investigate at the moment. Suppose that we had a "nearly perfect" averaging protocol A on
the smaller squares (cid:3)i, i. e. when A is run on each square, after t = timeA(√n) transmissions,
within (cid:3)i the values are for practical purposes equal to the the average of the original values.
That is,
xs(0)
(∀i)(∀s ∈ (cid:3)i)xs(t) ⋍ Ps∈(cid:3)i
.
#((cid:3)i)
Definition 1 For each square (cid:3)i, let s((cid:3)i) be the sensor closest to the center of (cid:3)i.
This can be determined by each square, using a constant number of transmissions w.h.p.
The s((cid:3)i) exchange values among themselves by Greedy Geographic Routing (see [5]).
Consider the following protocol. Suppose that A has been run on each subsquare of the form
(cid:3)i independently, and the values carried by the nodes within (cid:3)i are all equal. When s((cid:3)i)
4
becomes active, the following round takes place.
1. si := s((cid:3)i) picks a square (cid:3)j uniformly at random. si geographically routes a packet
with its value to sj := s((cid:3)j).
2. sj routes its own value to si by greedy geographic routing.
3. xsi ← xsi + 2√n
4. xsj ← xsj + 2√n
5. A is independently run on (cid:3)i (the process being activated by si by switching certain
5 (xsj − xsi).
5 (xsi − xsj ).
nodes on) and on (cid:3)j (initiated by sj similarly).
6. A is ended on square (cid:3)i by si (by turning certain nodes off), and A is ended on (cid:3)j by
sj (by switching certain nodes off.)
xs(t). Without loss of generality, we assume that Pi xi = 0, since this
only adds a constant offset and does not affect the rate of convergence. An application of
10 w.h.p . If we examine the evolution
Now, let zi(t) := Ps∈(cid:3)i
the Chernoff Bound tells us that (∀i)(cid:12)(cid:12)(cid:12)
of z, we see that after a round of the kind described above
#((cid:3)i)
√n − 1(cid:12)(cid:12)(cid:12)
< 1
• zi(t) = (1 − αi)zi(t − 1) + αjzj(t − 1)
• zj(t) = (1 − αj)zj(t − 1) + αizi(t − 1)
where ∀i, αi ∈ ( 1
2 , 1
3). From Lemma 1, it follows that
2√n )tkz(0)k2. Roughly speaking after O(√n log( n
E[kz(t)k2] < (1− 1
a distribution x(t′) such that kx(t′)k < ǫkx(0)k.
Each geographical routing mentioned above takes O(√n) transmissions w.h.p (see [5]). Also,
each process of initiating or ending A on a square (cid:3)i takes O(√n) transmissions.
ǫ )) of these steps, we have
So, the total number of transmissions with n nodes time(n) satisfies a recurrence of the form:
time(n) ⋍ O(cid:16)√n log(
n
ǫ
)(timeA(√n) + O(√n))(cid:17) .
Ignoring the dependence on ǫ, it would allows us to recursively define the algorithm A on
(cid:3), for which timeA(n) = n exp(O(log log n)2).
5
4 Description of the Algorithm
4.1 Notation
The square (cid:3) is partitioned into n1 subsquares (cid:3)i, where n1 is the nearest integer to √n
that is the square of an even number. For a square (cid:3)i1...ir , let E#(cid:3)i1...ir denote the expected
number of sensors within (cid:3)i1...ir . Then, while E#(cid:3)i1...ir > (log n)8,
the square (cid:3)i1...ir is partitioned into nr+1 subsquares (cid:3)i1...ir+1, where nr+1 is the nearest
integer to pE#(cid:3)i1...ir that is the square of an even number. Let
r,
ℓ := 1 + sup
(cid:3)i1...ir
i. e. the number of levels in this recursion. Given a square (cid:3)<i>, let s((cid:3)<i>) denote the
sensor nearest to its center. By our construction, these centers are well separated, and any
sensor has this property with respect to at most one square w.h.p.. We shall denote this
by (cid:3)(s). We assign a Level to each node by the following rule: If s = s((cid:3)i1...ir ), s has
level ℓ − r. These nodes are have Levels 1, . . . , ℓ. There is a single root node at Level ℓ,
namely s((cid:3)). The nodes at Level 0 are the nodes not of the form s((cid:3)i1...ir ). In the informal
discussion earlier, we did not concern ourselves with the error in the averaging carried out
on subsquares (cid:3)i. However, these errors propagate up the hierarchy rapidly, and hence it
is necessary to obtain results with greater accuracy in smaller squares. Thus we define the
desired accuracy recursively. Let ǫr be the accuracy for the averaging process in a square
ǫr−1
poly(n) for a polynomial of
(cid:3)i1...ir−1. Lemma 2 tells us that it is sufficient to take ǫr, to be
sufficiently large degree.
Let ǫ0 = ǫ, δ0 = δ. We recursively define ǫr+1 := ǫr
25n
2 +a and δr+1 = δr
n2a
r
7
.
We define time(n, ℓ−1, ǫr, δr) to be(cid:16)(log n
1, ǫr−1, δr−1) := time(n, r, ǫr, δr)na(cid:16)log( nr
Let s ∈ (cid:3)i1...iℓ−1.
ǫr
ǫℓ−1
) log(δ−1
) log(δ−1
ℓ−1)(cid:17)16
r )(cid:17)16
.
. Thereafter, we define time(n, r−
4.2 The Protocol
Every node s has two states, a local.state and a global.state, both of which are initially
= of f , but can also take the value on. Each node s possesses a private counter counter(s).
During initialization, the global.state of s((cid:3)) is set to on but every other global.state is 0.
The local.state of all nodes is set to of f at this juncture.
Let us suppose that the clock of s ticks. We describe the protocol followed by it below. We
consider two cases. If s is at Level 0, it obeys the following protocol: {
6
1. If local.state(s) = on
N ear(s);
}
N ear(s){
1. s picks an adjacent node v contained in (cid:3)i1...iℓ−1 uniformly at random.
2. s sets xs(t + 1) = xs(t)+xv (t)
v sets xv(t + 1) = xs(t)+xv (t)
2
2
;
;
}
We next describe the protocol if s is at a Level greater than 0. The subroutine N ear is the
same as above. Let (cid:3)(s) =: (cid:3)i1...ir .
{
1. If global.state(s) = on
(a) If counter(s) = 0 Activate.square(s);
(b) With probability n−atime(n, r, ǫr, δr)−1
• F ar(s);
• counter(s) ← 0;
2. If local.state(s) = on
N ear(s);
3. If counter(s) ≥ time(r, n, ǫr, δr) Deactivate.square(s);
Else counter(s) ← counter(s) + 1;
}
F ar(s){
1. s picks a square (cid:3)i′
its value to s′ geographically.
1...i′
r 6∋ s uniformly at random. Let s′ := s((cid:3)i′
r ) . Node s routes
1...i′
2. xs(t + 1) = xs(t) + 2
5(E#(cid:3)i1...ir xs′(t) − E#(cid:3)i1...ir xs(t)).
3. s′ sends back to a packet with its value xs′(t) to s by greedy geographic routing.
4. Node s computes xs(t + 1) = xs(t) + 2
5(E#(cid:3)i1...ir xs′(t) − E#(cid:3)i1...ir xs(t)).
7
5. counter(v) ← 0.
}
Activate.square(s){
1. If s ∈ Level 1, send packets to each node s′ in (cid:3)(s) setting local.state(s′) ← on by
flooding.
2. If s ∈ Level i > 1, send packets to each Level i−1 node s′ in (cid:3)(s) by greedy geographic
routing, setting global.state(s′) ← on.
}
Deactivate.square(s){
1. If s ∈ Level 1, send packets to each node s′ in square(s) setting local.state(s′) ← of f
by flooding.
2. If s ∈ Level i > 1, send packets to each Level i−1 node s′ in (cid:3)(s) by greedy geographic
routing, setting global.state(s′) ← of f .
}
5 Analyzing the number of Transmissions
Let H(n, r, ǫr, δr) denote the number of transmissions used in our protocol in one round of
(cid:3)i1...ir , in order to diminish the variance (of the values carried by sensors in (cid:3)i1...ir ) by a
factor ǫr, with probability 1 − δr.
Observation 1 In one round, i. e. the duration between s activating (cid:3)(s) := (cid:3)i1...ir and
deactivating (cid:3)(s), the number of long-range packet exchanges between sensors of the kind
s((cid:3)i1...irir+1) is Θ(cid:16)n log( n
ǫr
)(cid:17) w.h.p, where
E#[(cid:3)i1...ir ]
E#[(cid:3)i1...irir+1]
n =
.
Each of these involves O(pE#[(cid:3)(s)])n hops w.h.p (see [5]). Therefore the total number of
transmissions here is O(cid:16)n2 log( n
)(cid:17) w.h.p.
ǫr
8
Each of these long-range packet exchanges is followed by a period of averaging within the
involved subsquares, and this takes H(n, r + 1, ǫr+1, δr) = Ω(n) transmissions. Thus we have
the recurrence
H(n, r, ǫr, δr) = O(cid:18)(H(n, r + 1, ǫr+1, δr+1) + n)n log(
)(cid:19) .
= O(cid:18)H(n, r + 1, ǫr+1, δr+1)n log(
n
ǫr
n
ǫr
)(cid:19)
. For these parameters, δr = Ω(
As mentioned in subsection 4.1, we let ǫ0 = ǫ, δ0 = δ and recursively define ǫr+1 := ǫr
25n7/2 and
δr+1 = δr
poly(n)) and the n telescope.
n2
r
ǫr = ǫ0Ωn−O(log log n) since ℓ ∼ log log n. Now, the smallest squares that we create have
O(polylogn) sensors each w.h.p. Since the ordinary averaging that we do there (described
by the procedure "Near(s)") has an averaging time that is quadratic [1, 2], H(n, ℓ, ǫℓ, δℓ) =
Ω(polylog( n
)). And so using the recurrence for H and telescoping, we see that the total
ǫℓ
number of transmissions is
poly(n) ), since δ0 = Ω(
1
1
H(n, 0, ǫ0, δ0) = (H(n, ℓ, ǫr+1, δr+1))Yr (cid:26) E#[(cid:3)i1...ir ]
E#[(cid:3)i1...irir+1]
log
n
ǫr(cid:27)
= n(log
)O(log log n).
n
ǫ
This is n1+o(1) if ǫ = exp(−n
o(1)
log log n ), and δ = n−O(1).
6 Notes on Correctness
In the algorithm proposed in this paper, each square (cid:3)(s) has a certain latency, which is
the averaging time restricted to that square. In order for our algorithm to be correct, we
require that (cid:3)(s) be undisturbed by the long-range exchanges that s is involved in, during
this period. This is not a condition that can be imposed without the long-range exchanges
of s losing their i.i.d property, which is crucial in our analysis of convergence. In order to
retain this, and have an algorithm that is successful w.h.p we have set the rates at which
long-range exchanges of s occur to be lower than the inverse of the latency by a factor
na. As a consequence, w.h.p, in the course of the entire algorithm, there are no long-range
transmissions made by any node s while (cid:3)(s) is active. The only issue that we have not
dealt with in detail is of showing that our choice of errors ǫr achieves the desired end. This
follows from Lemma 2 interpreted as follows: The nodes i represent subsquares (cid:3)i1...irir+1 of
(cid:3)i1...ir and the yj(t) for different j represent the sum of the values held by the nodes in a
subsquare (cid:3)i1...irj after t long distance transmissions between subsquares since the activation
of (cid:3)i1...ir . We set ǫ := ǫr+1kx(0)k. The perturbations n(t) represent the errors generated
from imperfect averaging within these subsquares.
9
7 Concluding Remarks
We introduced non-convex affine combinations, in our averaging protocol in order to accel-
erate Geographic Gossip in Geometric random graphs. The number of transmissions used
in the course of our protocol is n1+o(1). This exponent is asymptotically optimal. Our algo-
rithm, unlike the previous one in [5] is not completely decentralized. However as far as we
can see, this is not a necessary feature associated with the use of affine combinations.
8 Future Directions
It would be interesting to study whether affine combinations can be used to develop a
completely decentralized algorithm for Geographic Gossip that is also energy efficient.
References
[1] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah. Gossip algorithms : Design, analysis
and applications. In Proceedings of the 24th Conference of the IEEE Communications
Society (INFOCOM 2005), 2005.
[2] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah. Mixing Times for Random Walks on
Geometric Random Graphs. SIAM ANALCO 2005.
[3] S. Carruthers, V. King. Connectivity of Wireless Sensor Networks with Constant
Density.ADHOC-NOW, 2004, 149-157
[4] P. Gupta and P. Kumar. The capacity of wireless networks. IEEE Transactions on
Information Theory, 46(2):388–404, March 2000.
[5] A. Dimakis, A. Sarwate, M. Wainwright. Geographic gossip: efficient aggregation
for sensor networks. In Proceedings of the fifth international conference on information
processing in sensor networks (IPSN), 2006.
[6] R. Karp, C. Schindelhauer, S. Shenker, and B. Vöcking. Randomized rumor spreading.
In Proc. IEEE Conference of Foundations of Computer Science, (FOCS), 2000.
[7] D. Kempe, J. Kleinberg, A. Demers. Spatial gossip and resource location protocols. in
Proc. 33rd ACM Symposium on Theory of Computing, 2001.
[8] D. Kempe, J. Kleinberg. Protocols and Impossibility Results for Gossip-Based Com-
munication Mechanisms. In Proc. 43rd IEEE Symposium on Foundations of Computer
Science, 2002.
10
[9] D. Mosk-Aoyama and D. Shah. Information dissemination via gossip: Applications to
averaging and coding. http://arxiv.org/cs.NI/0504029, April 2005.
[10] R. Motwani and P. Raghavan. Randomized Algorithms. Cambridge University Press,
Cambridge, 1995.
[11] M. Penrose.Random Geometric Graphs.Oxford studies in probability. Oxford University
Press, Oxford, 2003.
[12] L. Xiao, S. Boyd, and S. Lall. A scheme for asynchronous distributed sensor fusion
based on average consensus. In 2005 Fourth International Symposium on Information
Processing in Sensor Networks (IPSN), 2005.
A Appendix
Let Kn be the complete graph on n vertices {1, . . . , n}. ∀i, let αi ∈ ( 1
2). At time t ≥ 0, for
i = 1, . . . , n, let node i hold the value xi(t). Consider the following update rule. If the tth
clock tick belongs to node i, then, i chooses a node j uniformly at random, and the following
update occurs:
3 , 1
• xi(t) = (1 − αi)xi(t − 1) + αjxj(t − 1).
• xj(t) = (1 − αj)xj(t − 1) + αixi(t − 1).
2n )tx(0)T x(0).
Lemma 1 E[x(t)T x(t)] < (1 − 1
Proof:Let the update rule for x(t) be given by A(t − 1), i. e. x(t) = A(t − 1)x(t − 1). Note
j ), if the ith vector of the standard basis is denoted
that A(t − 1) = I − (αiei − αjej)(eT
by ei.
i − eT
E[x(t)T x(t)x(t − 1)] = E[x(t − 1)T A(t − 1)T A(t − 1)x(t − 1)x(t − 1)]
= x(t − 1)T E[A(t − 1)T A(t − 1)]x(t − 1).
Let αiei−αjej = αij and ei−ej = eij. Then, E[A(t−1)T A(t−1)] = E[(I−eij αT
Let Eij denote the n × n matrix whose ijth entry is 1 and every other entry is 0.
Then, by expanding, one finds that
ij)T (I−eijαT
ij)].
E[A(t − 1)T A(t − 1)] = I +Xi
= I(1 −
(1 − 2αi)2 − 1
n
Eii +Xi6=j
1
n − 1
) +
11T
n(n − 1) −
(1 − 2α)(1 − 2α)T
n(n − 1)
n(n − 1)
+Xi
(1 − (1 − 2αi)(1 − 2αj))Eij
(1 − 2αi)2Eii
n − 1
.
11
An application of the formula for E[x(t)T x(t)x(t − 1)], now gives us the following:
E[x(t)T x(t)x(t − 1)] = E[x(t − 1)T A(t − 1)T A(t − 1)x(t − 1)x(t − 1)]
= x(t − 1)T E[A(t − 1)T A(t − 1)]x(t − 1)
We know that ∀i, 1 − 2αi ∈ (0, 1
3).
Let us upper bound x(t − 1)T E[A(t − 1)T A(t − 1)]x(t − 1) using the the expression for
E[A(t − 1)T A(t − 1)] derived earlier.
1
(1)
(2)
x(t − 1)T I(1 −
)x(t − 1) = (1 −
n − 1
x(t − 1)T 11T x(t − 1)
1
n − 1
= 0,
)kx(t − 1)k2,
−
and,
x(t − 1)T (1 − 2α)(1T − 2αT )x(t − 1)
≤ 0
n − 1
n(n − 1)
x(t − 1)T Xi
(1 − 2αi)2Eii
n − 1
! x(t − 1) ≤ kx(t − 1)k2
9(n − 1)
.
Adding up the above inequalities,
E[x(t)T x(t)x(t − 1)] ≤(cid:18)1 −
8
9(n − 1)(cid:19) x(t − 1)T x(t − 1).
As a consequence,
E[kx(t)k2 x(t − 1)] <(cid:18)1 −
1
2n(cid:19)kx(t − 1)k2.
Successively conditioning on x(t − 2), . . . , x(0), we see that
E[kx(t)k2] <(cid:18)1 −
1
2n(cid:19)t
kx(0)k2.
This proves the lemma.
(cid:3)
An application of Markov's inequality gives us the following corollary.
Corollary 1
P (kx(t)k > ǫkx(0)k) ≤ ǫ−2(cid:18)1 −
1
2n(cid:19)t
.
12
Proof:
> ǫ2(cid:19)
P (kx(t)k > ǫkx(0)k) = P(cid:18)kx(t)k2
kx(0)k2
≤ ǫ−2E(cid:18)kx(t)k2
kx(0)k2(cid:19) (Markov's inequality)
2n(cid:19)t
≤ ǫ−2(cid:18)1 −
1
(cid:3)
An application of Markov's inequality gives us the following corollary.
Corollary 2
P (kx(t)k > ǫkx(0)k) ≤ ǫ−2(cid:18)1 −
1
2n(cid:19)t
.
We now consider a modified update rule, and prove a lemma similar to Lemma 1.
Let Kn be the complete graph on n vertices {1, . . . , n}. ∀i, let αi ∈ ( 1
2 ). At time t ≥ 0,
for i = 1, . . . , n, let node i hold the value xi(t). Let n(0), n(1), . . . be a sequence of real
numbers. Consider the following update rule. If the tth clock tick belongs to node i, then, i
chooses a node j uniformly at random, and the following update occurs:
3 , 1
• yi(t) = (1 − αi)yi(t − 1) + αjyj(t − 1) + n(t − 1).
• yj(t) = (1 − αj)yj(t − 1) + αiyi(t − 1) − n(t − 1).
Lemma 2 Suppose that for each t, n(t) < ǫ, and that a > 0. Then,
)t/2ky(0)k + 8√2n3/2ǫ(cid:19)(cid:21) ≤
P(cid:20)ky(t)k > n
2 (cid:18)(1 −
a
1
2n
5
na
.
Proof:y(t) = A(t − 1)y(t − 1) + n(t − 1), where A(t) = I − (αiei − αjej)(eT
j ), and
n(t − 1) = n(t − 1)(ei − ej). Let x(0) = y(0), and let the x(t) satisfy x(t + 1) = A(t)x(t) as
in Lemma 1. We observe that
i − eT
y(1) = x(1) + n(0)
and more generally,
y(t + 1) = x(t + 1) + n(t) +
A(t)A(t − 1) . . . A(i + 1)n(i).
t−1
Xi=0
13
An application of the triangle inequality now gives us
ky(t + 1)k ≤ kx(t + 1)k + kn(t)k +
t−1
Xi=0
kA(t)A(t − 1) . . . A(i + 1)n(i)k.
Our approach to proving this Lemma is to upper bound each term in the right hand side.
Observation 2
P(cid:20)kx(t)k > (1 −
1
2n
)t/2na/2kx(0)k(cid:21) ≤ (cid:18)(1 −
≤ (cid:18)(1 −
=
1
na
.
1
2n
1
2n
)t/2na/2(cid:19)−2
)t/2na/2(cid:19)−2
E(cid:18)kx(t)k2
kx(0)k2(cid:19)
(1 −
1
2n
)t
The above inequalities follow from Lemma 1 and Corollary 2. We shall now upper bound
the other terms as well with high probability. Using Corollary 2
P(cid:20)kA(t − 1) . . . A(i)n(i − 1)k
kn(i − 1)k
> (1 −
1
2n
t−i
4 n
)
a+1
2 (cid:21) ≤ ((1 −
1
2n
t−i
4 n
)
However,
and so,
= n−(a+1)(1 −
(1 −
1
2n
t−1
Xi=1
t−i
2 n−(a+1) <
)
4
na
a+1
2 )−2(cid:18)1 −
1
2n(cid:19)t−i
1
2n
t−i
2 .
)
P(cid:20)∃i(cid:26)kA(t − 1) . . . A(i)n(i − 1)k
kn(i − 1)k
> (1 −
1
2n
t−i
4 n
)
a+1
2 (cid:27)(cid:21) ≤
4
na
.
We next observe that
As a consequence we have
Observation 3
(1 −
1
2n
Xi≤t
t−i
4 n
)
a+1
2 < 8n
a+3
2 .
P"Xi
kA(t − 1) . . . A(i)n(i − 1)k
kn(i − 1)k
> 8n
a+3
2 # ≤
4
na
.
14
Once we put the above two observations together and note that (∀i)√2ǫ ≥ kn(i)k, an
application of the union bound gives
P(cid:20)ky(t)k > n
a
2 (cid:18)(1 −
1
2n
)t/2ky(0)k + 8√2n3/2ǫ(cid:19)(cid:21) ≤
5
na
.
(cid:3)
15
|
1802.07280 | 1 | 1802 | 2018-02-19T01:58:28 | Simulating the Ridesharing Economy: The Individual Agent Metro-Washington Area Ridesharing Model | [
"cs.MA",
"nlin.AO"
] | The ridesharing economy is experiencing rapid growth and innovation. Companies such as Uber and Lyft are continuing to grow at a considerable pace while providing their platform as an organizing medium for ridesharing services, increasing consumer utility as well as employing thousands in part-time positions. However, many challenges remain in the modeling of ridesharing services, many of which are not currently under wide consideration. In this paper, an agent-based model is developed to simulate a ridesharing service in the Washington D.C. metropolitan region. The model is used to examine levels of utility gained for both riders (customers) and drivers (service providers) of a generic ridesharing service. A description of the Individual Agent Metro-Washington Area Ridesharing Model (IAMWARM) is provided, as well as a description of a typical simulation run. We investigate the financial gains of drivers for a 24-hour period under two scenarios and two spatial movement behaviors. The two spatial behaviors were random movement and Voronoi movement, which we describe. Both movement behaviors were tested under a stationary run conditions scenario and a variable run conditions scenario. We find that Voronoi movement increased drivers' utility gained but that emergence of this system property was only viable under variable scenario conditions. This result provides two important insights: The first is that driver movement decisions prior to passenger pickup can impact financial gain for the service and drivers, and consequently, rate of successful pickup for riders. The second is that this phenomenon is only evident under experimentation conditions where variability in passenger and driver arrival rates are administered. | cs.MA | cs | ORCID #: 0000-0003-2234-8756
Simulating the Ridesharing Economy: The Individual Agent
Metro-Washington Area Ridesharing Model (IAMWARM)
Joseph A. E. Shaheen
[email protected]
Computational and Data Science Department, George Mason University
4400 University Drive, MS 6A2
Fairfax, VA 22030-4444
Abstract
The ridesharing economy is experiencing rapid growth and innovation. Com-
panies such as Uber and Lyft are continuing to grow at a considerable pace while
providing their platform as an organizing medium for ridesharing services, in-
creasing consumer utility as well as employing thousands in part-time positions.
However, many challenges remain in the modeling of ridesharing services, many
of which are not currently under wide consideration. In this paper, an agent-based
model is developed to simulate a ridesharing service in the Washington D.C. met-
ropolitan region. The model is used to examine levels of utility gained for both
riders (customers) and drivers (service providers) of a generic ridesharing service.
A description of the Individual Agent Metro-Washington Area Ridesharing
Model (IAMWARM) is provided, as well as a description of a typical simulation
run. We investigate the financial gains of drivers for a 24-hour period under two
scenarios and two spatial movement behaviors. The two spatial behaviors were
random movement and Voronoi movement, which we describe. Both movement
behaviors were tested under a stationary run conditions scenario and a variable
run conditions scenario. We find that Voronoi movement increased drivers' util-
ity gained but that emergence of this system property was only viable under var-
iable scenario conditions. This result provides two important insights: The first is
that driver movement decisions prior to passenger pickup can impact financial
gain for the service and drivers, and consequently, rate of successful pickup for
riders. The second is that this phenomenon is only evident under experimentation
conditions where variability in passenger and driver arrival rates are adminis-
tered.
ORCID #: 0000-0003-2234-8756
2
1. Introduction
Ridesharing services such as Uber and Lyft have been experiencing explosive
growth driven by rider demand and a number of other factors in industry [1]. One
challenge for both ride-sharing service providers as well as for their driver con-
tractors is how to maximize driver acceptance of new customers while ensuring
drivers gain maximum utility from rides given. In other words, how to maximize
both financial gains for the service and its drivers while ensuring maximum ser-
vice-level quality for its customers.
Perhaps as a reflection of the growth of the ridesharing industry and of the
aforementioned challenges, empirical research in this area is also experiencing a
surge, exemplified by a growing number of journal publications [1-18] that ex-
plore the multidimensional challenges and opportunities produced by the wide-
spread adoption of ridesharing services.
This paper aims to investigate spatial behavioral conditions under which driv-
ers can gain increased financial returns (utility) on their invested time, while
simultaneously ensuring that a maximum number of potential passengers reach
their destination. The model produced utilizes an agent-based modeling (ABM)
framework that has the potential to be extended, expanded, and tested under many
variable conditions. And therefore, while it would be immediately salient that
much can and should be tested with the model, we reserve future extensions and
testing for future papers.
It is worthwhile to note that the agent simulation perspective is highly suitable
for testing spatial behaviors; when utilizing agent simulations it is considered
trivial to create many autonomous, heterogeneous agents following one or more
behavioral rule-sets and to simulate behaviors for various initial conditions and
parameters without the constraints of rigid assumptions. Additionally, as will be
shown by the results of our experimentation, other modeling techniques may not
fully capture the true temporal dynamics of a ridesharing service because of het-
erogeneity in agent decision-making, the spatial significance to end results, and
the variable scenario conditions under which emergent properties could arise. In
a subsequent section, we will show the relevance of the later.
The model described in this paper focuses on simulating drivers and riders in
the Washington, D.C. metro region and attempts to simulate the movement of
drivers under two spatial movement conditions.
Ultimately, the aim of this model is to gain insight into whether drivers, riders
and ridesharing services benefit more or less from optimized decision-making
during the drive-pickup-drop-off lifecycle familiar to ridesharing customers,
ORCID #: 0000-0003-2234-8756
3
while the aim of this paper is to highlight testing of some specific conditions.1 In
later iterations of the model, an increased variety of behaviors will be investi-
gated.
1.1 Background
There have been a number of studies in the last few years taking aim at under-
standing ride-sharing services and carpooling schemes – each of which takes a
different investigative position on the challenges faced by resource pooling ser-
vices as a whole [1, 22], while some consider some facet of modeling behaviors
[3] using agent-based approaches. The majority of papers reviewed were of mod-
eling carpooling decisions as an optimization problem [7, 11, 12] and finally,
some approaches intended to make early-stage predictions about carpooling and
ridesharing trends [6, 9] were pre-existent in the literature.
What becomes very clear during a topical literature review over the last few
years is that no real attempt had been made to provide for a comprehensive rides-
haring agent-based simulation that captures prevalent dynamics; though much of
the research attempts to understand the effects of ridesharing in general.
For example, Cho et. al. [3] provided a full description of a hypothetical agent-
based model for a carpooling application without offering an actual build of the
model hypothesized. The authors focused on the systemic theoretical structure of
the proposed model, the mathematics and optimization techniques that would be
used and the general form of social network types that could be used between the
driver agents of said model. The same group [4] later proposed another agent-
based model – this time only based on social network interactions without
implementation.
Significant advances in the area of heuristics and algorithm development that
propose better route optimization techniques have also been made over the last
few years and this is an area where high-value and productive work has been put
forward. For example, Pelzer et. al. [12] developed a method which aims to best
utilize "ridesharing potential while keeping detours below a specific limit" using
a spatial partitioning method.
IAMWARM aims to build a foundational baseline to test a small number of
interesting spatial problems for which answers have not been provided as of yet
and to use the model created as the basis for future improvements, extensions,
expansions, and experiments. We begin that endeavor by discussing our primary
and most central question: Can ridesharing utility for both riders, drivers, and
service be increased through varying the information-shared among agents,
1 The author of this paper registered with one ridesharing service in order to
gain insight into the natural behaviors of drivers and riders of the service. A total
of 30 trips were carried out.
ORCID #: 0000-0003-2234-8756
4
ultimately affecting spatial movement behaviors? And if so, under what condi-
tions could one note a difference in system-level properties?
1.2 Information Asymmetry vs Information Symmetry
Our path is to find the simplest method of testing ridesharing utility schemes
spatially, so we begin by discussing the problem of increased information sharing
briefly implied in the previous section.
For our specific context, we define information-asymmetry as a lack of infor-
mation regarding the location of other drivers by other drivers. That is-drivers,
in an information-asymmetric service, would not be given the locations of other
drivers, or riders2, with exception of a single potential rider within their vision's
radius who has just requested a pick-up, and thus without that information and
without clear route planning driven by spatial demographics, drivers would
simply move about randomly hoping to 'luck out' and be near a potential cus-
tomer when they request a pickup. This is currently the method by which all rides-
haring services manage their respective platforms. Drivers of those services are
not given location information of other drivers, and must move about based on
randomness, their own past experiences and information gleaned from their social
networks; and so ultimately, must make their spatial movement and positioning
decisions based on either luck or experience gained from learning. We will omit
learning behavior from this iteration of the model.
Symmetry represents a condition such that driver agents have all the available
information about other driver agents and rider agents. For this model however
we bound true information symmetry to a localized version that limits driver
agents' knowledge to the nearest driver agent and only to the nearest rider agent.
The comparison between information asymmetry and symmetry which will be
established by the comparison between random movement and Voronoi
movement will be applied such that agents have no vision for information
asymmetry scenario runs (no knowledge of the position of any other driver) and
have only local vision in the information symmetry variation of the model
(knowledge of the nearest driver agent's position). This modification is a direct
result of the platform chosen for the development of the simulation and its ability
to perform, and due to a lack of a clear theoretical or even observed cognitive
standard to base spatial behaviors upon in this case.
This model will investigate how drivers would benefit from having local in-
formation about peer drivers available in real time using a hypothesized spatial
2 We will later explain our terminology in detail, but for now we define a driver
as an agent who is picking up a rider from one location on our model's spatial
grid to another. Once a rider is "picked up" we will refer to him as a passenger.
In our model a passenger is no longer an agent but is a data point in the driver
agent's attributes list.
ORCID #: 0000-0003-2234-8756
5
behavior and whether that additional information would maximize driver utility.
And, though there are a number of differing spatial behaviors that can be consid-
ered, we will only test one behavior which we theorize would result from access
to that information. We will extend this model in future iterations with more be-
haviors.
Fig. 1. Shows an example of a Euclidean space with generator points and their corresponding
Voronoi polygons. The polygons are "emerged" from the collective positions of the generator
points such that all points in each area corresponding to a polygon are closer to the
corresponding generator point than to any other point [20].
We call the assumed behavior resulting from information symmetry Voronoi
behavior or Voronoi movement. We propose this behavior using the spatial con-
cept of a Voronoi polygon [19] as a base. A Voronoi polygon or diagram-as it
is commonly known-is a partitioning of a spatial plane such that "all locations
in the Voronoi polygon are closer to the generator point of that polygon than any
other generator point…in Euclidian plane" [18]. In other words, it is the space
such that maximum territory is created for each generator point without
overlapping the area belonging to any other generator point. Figure 1. shows an
illustration of generator points and their respective Voronoi polygons.
Voronoi movement essentially amounts to driver agents receiving location in-
formation about the nearest driver agent and moving away from them so as to
increase the potential of picking up a new customer and reducing local competi-
tion-a diverging topological behavior where each agent maximizes the distance
and the territory between self and all other agents. This viewpoint is a corollary
to the Voronoi polygon-from the view of the generator point (agent) not the
adjacent spatial points in the polygon, hence we call this behavior Voronoi be-
havior.
We compare Voronoi movement behavior with a random movement pattern
where driver agents move randomly across our spatial grid until they are close
enough to a rider agent to execute a pickup. The random movement behavior is a
ORCID #: 0000-0003-2234-8756
6
reflection of driver agents having no knowledge of where other drivers and where
customers might be. In this iteration of the model, we assume agents do not learn.
In a high-fidelity model, it would likely be the case that drivers would learn
about rider behaviors and adjust their own behaviors accordingly. However, as
you will see in the results section, even with this simplification, our baseline
model offers interesting conclusions nonetheless. Finally, rider (potential passen-
ger) agents do not move in this iteration of the model but enter the simulation at
a spatially random location on the Washington, D.C. geographic lattice. Figure 2
and 3 show the graphical representation of the model which was developed in
NetLogo [20] and subsequent sections will discuss model design particulars.
Fig. 2. This is the graphical representation of the model. Yellow agent types (person icons) are
rider agents. Red (car icons) are driver agents. When a driver agent performs a pickup, their
color turns from red to white to display that they are no longer available.
ORCID #: 0000-0003-2234-8756
7
Fig. 3. This model was implemented in NetLogo 5.3 and utilized open access data from the
Washington, D.C. government website. In this figure, we show the graphical user interface of
the model. Inputs, such as the number of drivers and the number of potential riders active at any
moment are complemented by outputs on the far right such as profitability, average cash on
hand, and passenger pick-ups.
ORCID #: 0000-0003-2234-8756
8
2. Model & Methods
IAMWARM was implemented in NetLogo 5.3 and utilized the GIS extension
native to the platform to import map and GIS data into the model. The model's
spatial configuration was based on a road network imported from the Washing-
ton, D.C. government Open Data Project website3, which included highly accu-
rate, editable shapefiles. The data included feature labels for roads and intersec-
tions. These features are used in the instantiation and location initialization of
agents, and for certain critical agent behaviors to be discussed in later sections of
this paper.
At the current iteration of the model we chose not to include additional layers
of geographic information for simplicity (only the road network was included),
but in future iterations utilizing the Open Data project more broadly can be ad-
vantageous in increasing the efficacy of our model specifically by adding more
spatial configuration data. Figure 4. provides a complete graphical summary of
the model's logic.
The road network was imported to NetLogo and an internal spatially-equiva-
lent configuration was assigned (labeling). All roads were labeled internally by a
variable to help identify spatial cells that contained a road, versus spatial cells
that did not. This would later be an important step when designing the movement
choices of agents in the simulation since all movement and agent entry will occur
on road cells as one might expect.
The model contains two agent types: drivers and riders. Drivers can move
across the model space, but only on cells that contain a road, while riders do not
move, but can only be initialized on roads, specifically intersections.
The movement of the driver agents was designed to be based on a direct line
of sight-that is-although the drivers must always remain on roads, we assume
that following actual traffic routes would not provide a negligible difference in
destination arrival times. This is mainly due to the size of our spatial lattice which
numbers in the several thousand. Moreover, for our research goals, it suffices that
driver agents move in a direct path to their destinations once rider agents are
picked up. In future iterations of the model, traffic and road direction movement
could be taken into account to create a greater sense of realism. For now, driver
agents move on roads in a direct fashion to their destinations.
2.1 Agent Behaviors
Before we discuss the specifics of agents used in our model we define the ter-
minology used in the model. We define driver agents as those agents who are
intending to pick up a rider. A rider agent is an agent who has been instantiated
and can be picked up by a driver. Once a driver agent picks up a rider agent, the
3 www.dcogc.org
ORCID #: 0000-0003-2234-8756
9
rider agent becomes a passenger. Passengers are not agents and do not interact
with their environment. In other words, riders who become passengers simply
become an attribute of the driver agents, releasing with them certain data points
which are then used by the driver agents post pick-up. This terminology will be
used throughout the model description.
Fig. 4. A complete breakdown of the model's process is presented in this figure. The area
labeled inside the yellow box is where the majority of driver agent behavior takes place. The
small orange box at the top of the diagram represents the model moving to the next time
unit/time cycle. Note that the majority of rider agent interactions are included as part of the
driver agent behaviors since much of the rider agent behaviors in the model is restricted to
entering or leaving the simulation.
Load Spatial Data, and Build Envrionment from GIS Faile Data.Start of Simulation SetupLoad original setting parameters (number of drivers and riders) at random shapefile locationsStart SimulationMovement parameter setting: Voronoi or RandomStep/TurnVoronoiAre there drivers nearby (within Vision setting)Check and record all other driver agents within vision settingPick Random driver agent from collected list, record relative position.Find target patch as set by distance setting at an angle of 180deg from alternative driver agent. (Opposite direction)Does target cell contain a road?YesMove to Target CellNoStayAny Rider Agents within Vision parameter setting?YesPerform Rider Agent PickupRandomChoose a random angle between 0 and 360.YesAccording to Distance parmater setting, assign a target cellNoRider agent becomes Passenger; Shares Destination with driver agent.Driver agent identifies destination coordinates; faces direction of coordinates.Driver agent chooses random angle between +45 and -45 of destination; Currently carrying passenger?Yes24-hour simulation?YesStep/turn = 24 hours eqivilant?YesStop SimulationNoVariable or Stationary run selected?NoVariableSet agent parameter settings according to programmed schedule (arrival rates).StationaryNoUpdate measures, counts, statistics, locations for driver, rider agents, and environment properties.Is entry & exit module checked?YesAllow for random creation of new rider and driver agents according to parameter settings.Allow for the removal of rider and driver agents randomly within limits of parameter settings.If driver agent has no more energy, remove from simulation. If rider agent has exceeded maximum wait time remove from simulation.NoDriver Agent ProcessSimulation Setup ORCID #: 0000-0003-2234-8756
10
2.2 Driver Agents
Driver agents are instantiated in the initial setup of the model according to a user
input parameter and a random distribution set to uniform properties. Spatially, the
driver agents are initialized at random road intersections of the geospatial model,
and according to a randomization test score that allows for a greater randomized
spatial distribution of agent objects. Driver agents are also instantiated throughout
the model run according to the aforementioned user set parameter. Those agents
are also instantiated at road intersections and their rates of arrival are also set
according to a user set parameter.
Driver agents move from one spatial cell to another by first checking whether
a road exists in the cell ahead. If it does, then an agent may move to that cell. If
no road exists in that cell then for the random movement behavior without an
active passenger procedure, agents are instructed to rotate a random number of
degrees between 0 and 360 and repeat the process.
For driver agents who are carrying a passenger, the process is the same. How-
ever, the randomized degree value is set to be between -45 and +45 degrees if a
road is not found directly in the heading of the driver agent. By doing so, we
ensure that drivers are constantly moving in the direction of their destination, but
are still able to overcome the majority of obstacles in their way, such as the lack
of available roads to travel on. This method does have grounds in reality in that
drivers who may not necessarily know precisely how to optimize their routes,
may in general, pick a random route that they know to be in the general direction
of their destination. In future iterations of the model a more advanced pathfinding
algorithm could be adopted such as the A* pathfinding algorithm, but for our
purposes, we assume that the difference is negligible and it is trivial to show so
through a model run.4
Driver agents are assigned a number of attributes at instantiation and some are
assigned as the model is run situationally. Attributes include energy level, cash-
on-hand, time driven, current driver destination (if carrying a passenger), riders
who are nearby, current passenger id, time the current passenger has been on a
trip, how many riders the driver has picked up, and how many passengers the
driver has dropped off, as well as a Boolean passenger variable indicating if the
driver agent is currently carrying a passenger. The attributes are more critical to
the progress of the model at varying times through simulation runs, depending on
the active phase of the drive-pickup-drop-off-drive cycle of the driver agents.
Therefore, a deeper explanation of the attributes and their relevance is appropriate
at this time.
We assign a level of energy to every driver agent set to be a random number
following a normal distribution between a 4 and 8-hour range translated into
model time units (which is set to be at 1 minute per time unit) by estimation. The
a
video
of
4
For
a
typical model
run,
please
visit
https://www.youtube.com/watch?v=apJEvDl4aqc
ORCID #: 0000-0003-2234-8756
11
underlying assumption is that the majority of drivers will only be able to drive,
regardless of their level of success, for a period determined by time availability
and physical and/or mental fatigue. This is an appropriate assumption and the
Gaussian shape of this distribution is not far-fetched. Time-driven is a variable
that counts the amount of time driven by each driver and is used to display and
calculate the model's summary statistics.
Cash-on-hand is the variable attribute that stores the accumulation of fares
each driver agent has gained, as well as the variable in which cash is deducted
(vehicle and transportation costs) for drivers who are not carrying active passen-
gers. In other words, it is the driver agents' total utility and financial gain at any
given time-period.
Pick-up count and drop-off count are variables that store the total number of
riders successfully picked up from their initialized location and successfully
dropped off at their destination, respectively. The Boolean passenger variable
shows whether the current driver is currently seeking a rider or already has a pas-
senger (as discussed earlier, riders that receive a pickup by a driver agent become
'passengers'-simply an attribute of the driver agent) and is used in a number of
important model mechanics. Passenger-id is a variable that stores the id of the
rider currently within the driver agent's vehicle when she becomes a passenger.
It is equivalent to the driver asking for the rider's name prior to pick-up and is
used in the verification process of the model to ensure that drivers are successfully
picking up intended riders.
Variables for nearby riders and nearby drivers were also implemented as at-
tributes of the driver agents. The first being the number, and id of any nearby
riders waiting for pickup. This attribute is used to evaluate whether there are any
riders nearby available for pickup. The second is whether there are any nearby
driver agents, and is used in the Voronoi movement mechanism. The details of
the Voronoi movement mechanism will be discussed in a subsequent section.
2.3 Rider Agents
Rider agents are instantiated at initialization of the model at random locations
(intersections). Rider agents do not move but await their intended pick up in the
same location. This is in line with expected behaviors of ridesharing service cus-
tomers. Rider agents are instantiated utilizing a user-set input and a random var-
iable to allow for some stochasticity in the model runs. Spatially, and in a similar
fashion to driver agents, they are placed randomly across the available intersec-
tions of the model's geographic configuration, and only on road cells.
Rider agents are instantiated throughout the model's runs but at a rate per time
unit user fixed parameter, unlike driver agents who typically remain in the
model's space until they decide to leave (randomly) or because they have ex-
hausted their energy variable and are replaced stochastically up to a maximum
user set parameter. In other words, while the number of driver agents is set by a
ORCID #: 0000-0003-2234-8756
12
maximum capacity global (exogenous) variable, the number of rider agents is set
to be a rate of arrival following a probability distribution. For simplicity, we chose
the arrival rate probability distribution to be normal, though in subsequent itera-
tions of the model testing of other probability distribution types, such an expo-
nential arrival function, would be necessary.
Rider agents possess two attributes for this iteration of the model. The first is
the rider destination, which is a randomly assigned destination converted to the
spatial coordinate equivalent. This destination is assigned at the moment of in-
stantiation of the rider agents. The second is the wait time variable which is a
count of how long a rider has been waiting for pickup and which is used to com-
pare to a user-set input to determine whether the rider agent should look for 'al-
ternative' transportation methods (like metro or bus service). A user designated
input allows for varying the waiting time of rider agents. For typical model runs
we assigned this variable to 20 minutes. Once a rider agent reaches their maxi-
mum waiting time assigned they leave the simulation. This is a proxy behavior
for the rider agent attempting to find alternative modes of transportation to their
destination.
2.4 Model Mechanics
The model relies on user inputs for the number of drivers (capacity), the num-
ber of riders arriving per given time unit, the maximum waiting time for rider
agents, whether or not to use random movement while attempting to find a rider
or whether to maximize distance from any other drivers in the area of assigned
vision (Voronoi movement). Based on these inputs and the parameters of the
model a typical model run behaves as follows.
Rider and driver agents are instantiated on a highly accurate road map of
Washington, D.C. at random intersections in accordance with the input parame-
ters assigned by the model user. Once the model is run, rider agents are spawned
while others, according to the user input, will leave the simulation. The same is
applied to driver agents.
Driver agents move according to one of the predetermined movement methods
(random movement or Voronoi movement based on which version of the simula-
tion is run) until they are within a user-set proximity setting of a rider agent (for
our model the vision was set to 3 cells). During this time, driver agents lose en-
ergy at a rate of 0.75 per time unit and lose cash at a rate of $0.1 per time unit.5
Once a driver agent is within a designated (by modeler) vicinity of the rider
agent, a pick-up occurs, and the rider is converted to a passenger. A transfer of
5 We assigned these cash variables based on a rough estimate of distance trav-
elled versus fare/ride gained from observations and experiences with a rideshar-
ing service.
ORCID #: 0000-0003-2234-8756
13
the destination of the rider occurs from rider agent to the driver agent at this time.
Once a driver agent has successfully executed a pickup, the agent can no longer
make any additional pickups, and it is at this point that the driver agent begins to
accumulate 'cash', set at a rate of $2.00 + 0.60/time unit. Calibration of the model
was conducted to reach a dollar amount that could be probable through qualitative
observations of distance and time of real-world trips versus the model's spatial
geometry. The key was to set the fare rate to include a fixed amount and a variable
amount so as to reflect actual ridesharing services.
Driver agents then proceed in a direct path to the coordinates of the transferred
destination while earning 'cash' and losing 'energy'. Once they reach their prox-
imate destination, a drop-off is executed, their passenger-carry variable is reduced
from 1 to 0, and a successful trip is recorded as being now completed. All relevant
attribute and model-level variables are updated with this new information. The
driver then continues to move searching for new riders and repeats the drive-
pickup-drop-off process.
Rider agents who are not picked up within their waiting time-period limit find
alternative transportation and leave the simulation, while driver agents who are
not carrying a passenger could "give up" and leave the simulation. The latter
could also run out of energy and leave the simulation due to fatigue. Typical runs
are for a 24-hour period, but a model user can run the model indefinitely if they
desire.
2.5 Model Inputs and Parameters
We implemented 7 inputs in our model that can be assigned and varied by the
user. Table 1. summarizes those inputs and contains their descriptions. The most
important of which are the maximum capacity for driver agents (drivers-count)
and the rate of entry of new rider agents into the model (riders-per-time-unit).
Other inputs are also critical but were not tested in a significant way-though
adjusted for calibration and realism. Those are the Voronoi vision setting which
controls how far driver agents can see other driver agents, the local-regional scale
which amounts to an adjustment for the speed of movement of the driver agents,
and a binary-switch which turns on or off the possibility for both driver and rider
agents leaving the simulation randomly.
In Table 2, we describe the parameters and distributions used in various parts
of the model. As mentioned in an earlier section we used a normal distribution of
varying means and standard deviations as the basis for a number of statistical tests
and parameter values so as not to add any additional unverified assumptions or
complexity to the model. Hence, we rely on the Central Limit Theorem heavily.
However, in future iterations of this model specific testing and data collection of
the distributions' parameters must be undertaken and compared to real data from
a ridesharing service.
ORCID #: 0000-0003-2234-8756
14
Table 1. Inputs of the model with typical value assignments.
Type
Input
Description
Typical
Value(s)
Driver
Agents
Driver Agent Number Assigns a maximum number of drivers to be ac-
50-150
tive at any given time unit
Normal Random Move-
ment (Choice)
Sets the movement behavior of drivers to be of
a random nature while they await a rider pickup
Voronoi Movement
(Choice)
Sets the movement behavior of driver agents to
follow a Voronoi-distance maximizing method
Voronoi Vision
If Voronoi Movement is chosen, sets the Voro-
noi movement vision distance
Local-Regional Scale Sets the vision and movement range for driver
agents. This amounts to a speed setting and is
used to calibrate the model.
N/A
N/A
3
0.5
Rider Agents Riders Active Per Time
Unit
Sets the rate by which new riders enter the sim-
ulation and await pickup
20-75
Environment
Scenario (Choice)
Sets the model into a run type where an expected
rate of arrival for riders and an expected maxi-
mum capacity for drivers is set at different hours
of the day.
Saturday
Table 2. Parameters used in model mechanics.
Parameter
Type
Value
Description
Driver Agent Placement Test X > 0.5
Rider Agent Placement Test X > 0.5
Normal (1,1) Tests whether a random number from a nor-
mal distribution with a mean of 1 and standard
deviation of 1 is greater than 0.5. If so, place-
ment of a driver agent succeeds at a given in-
tersection.
Normal (1,1) Tests whether a random number from a nor-
mal distribution with a mean of 1 and standard
deviation of 1 is greater than 0.5. If so, place-
ment of a rider agent succeeds at a given inter-
section.
Driver Agent Energy
Attribute Normal(360,120) Sets the energy of a driver agent at instantia-
tion as a number drawn from a random distri-
bution with a mean of 360 and a standard de-
viation of 120 (minutes)
Kill Count
Variable
(Normal (0,1) Sets the number of driver and rider agents who
will leave the simulation, randomly without
depleting their energy (driver agents) or reach-
ing maximum wait time (rider agents) to be the
absolute value of a random number drawn
from a normal distribution with mean 0 and
standard deviation of 1.
ORCID #: 0000-0003-2234-8756
15
2.6 Model Outputs
A number of outputs were included in the model to assist in the verification
process, to understand model mechanics and to derive results from model runs.
Table 3. lists those outputs and their descriptions. Our focus was to understand
driver agent utility given some set of inputs, parameters, and pre-conditions.
There are many forms of driver agent utility to consider, each of which would
require a focus on a different set of output measures. For this iteration of the
model, we chose to focus our attention on total driver agent utility in the form of
total profit from each model run. We include no outputs to measure ridesharing
service utility or rider/passenger utility in our final analysis and conclusions,
however, a number of outputs aimed towards the measurement of rider agents,
passengers, and ridesharing service utility are designed into our model and are
displayed to the user. We hope to expand on our analysis of system utility by
considering rider, passenger and service utility in future iterations of the model.
Table 3. Model outputs and measures
Output
Description
Number of driver agents active
The number of driver agents active in the model
Number of rider agents active
The number of rider agents active in the model
Total riders giving up
The total riders giving up based on randomly set pa-
rameters
Average number of riders picked
up per time unit
The average number of riders picked up per time is cal-
culated for each time unit and displayed
Total number of rider agents
picked up
This is the total number of rider agents converted to
passengers
Total number of successful drop-
offs
Total number of successful drop-offs, which tends to
be lower than the number of pickups as some driver
agents don't reach their destinations
Number of idle driver agents
Number of driver agents without an active passenger
Number of working driver agents Number of driver agents with an active passenger
Average cash on hand
The total amount of cash held by all driver agents (ac-
tive)
Number of agents who left (ran-
domly)
Number of agents who left the simulation due to ran-
dom tests
Number of passengers in active
trips
Number of passengers carried by driver agents
Average wait time
Rider agent average wait time
Average energy level
Driver agent average energy level
Total cash with active driver agents Total cash for all active driver agents at any given time
unit
Average fare per ride
Average fare per ride at any given time unit
Total profit generated
Total profit generated by all activity of the model
ORCID #: 0000-0003-2234-8756
16
2.7 Scenarios
Much of the model can be run in a stationary mode-that is-it can be run in a
form of equilibrium where driver agents and rider agents arrive at predetermined
rates resulting in a constantly changing but variably fixed dynamic. This is inter-
esting for general runs, verification of model mechanics and quality, as well as to
gain a general understanding of the ridesharing process. It is trivial to hypothesize
that in any transportation system the rates by which riders and drivers arrive, in-
teract, and exit are variable but also subservient to the city (the spatial lattice) in
which the ridesharing service operates. This would include seasonal variables
such as the time of year, month, day, and time of day. Additionally, rider and
driver rates and activity are also affected by current events, traffic, roadworks,
weather patterns and other exogenous factors. Therefore, though running the
model in situ yields important insights, it is important to run experimentation in
some variable scenario for comparative reasons and for a closer approximation of
real-world dynamics simply because the variance itself could yield insight.
Therefore, based on anecdotal evidence gained from the author's registering
with a ridesharing service and gaining first-hand experience in typical driver de-
cisions made, we develop a scenario which is not entirely hypothetical in order to
test the model's effectiveness under varying conditions. In future iterations of this
model, we intend to develop scenarios grounded in real data collected and to de-
velop a number of them to test different scenarios under different conditions with-
out such heavy reliance on qualitative observations.
For this model iteration, we conducted a test of one scenario-the "Saturday"
scenario which varies only arrival rates of both driver and rider agents according
to what might be expected on a typical weekend day-Saturday. Table 4. provides
a summary of the scenario and reasoning, where appropriate, for selection of sce-
nario inputs and parameters. As you will see from the table, we varied the arrival
rates and expected capacities of rider agents and driver agents respectively. For
example, on 'Saturday' we would expect high customer demand for the hour be-
fore 'lunch' as many city dwellers may be engaging in social activities in the
subsequent hour and so they intend to arrive before 'lunch-hour'. In this time-
period (11 AM) and in this specific scenario, we expect that in anticipation of
higher customer demand the number of driver agents may also increase, and so
we increase the maximum capacity of the driver agents.
ORCID #: 0000-0003-2234-8756
17
Table 4. This is a variable model scenario used in testing the model under realistic condi-
tions drawn from anecdotal observations.
Hour Driver Capacity Rider Rate of En-
Reasoning/Explanation
5AM
8AM
11AM
12PM
1.30PM
2.30PM
3.30PM
5.30PM
6.30PM
7.30PM
8.30PM
10
20
50
5
45
50
25
40
45
60
80
try
5
10
20
10
25
15
10
5
5
30
Early Morning - Airport Traffic - Mostly Quiet
Drivers starting their day for the Saturday
Brunch/Lunch
Lunch Time - low activity
Post-Lunch Rush
Post-Lunch Rush
Stationary Activity
Evening Drivers Beginning Their Shifts
Evening Drivers Beginning Their Shifts
Night Activity Period - Riders are going out to so-
cial events
40
Night Activity Period - Riders are going out to so-
cial events
9.30PM
100
40
Night Activity Period - Riders are going out to so-
cial events
10.30PM
90
10
Low Rider Activity - Riders are at their destina-
tions. Drivers still on the road expecting a rush of
new riders.
80
75
65
35
11.30PM
12.30AM
2AM
4AM
10
30
Some drivers give up, exit the simulation
More drivers give up. Riders beginning to end their
work shifts.
30
More drivers give up. Riders beginning to end their
work shifts.
10
End of night traffic. End of 24 our cycle.
2.8 Testing, Verification, and Validation
To test whether spatial movement behaviors can affect driver financial gain, we
aimed to compare Voronoi movement prior to agent-pickup with random move-
ment prior to agent-pickup. In other words, having drivers access information
ORCID #: 0000-0003-2234-8756
18
about other local drivers and having chosen to 'spread out' maximizing their per-
sonal territory and likelihood of rider pickup, when compared with random move-
ment, regardless of where other driver agents may be. This collection of tests
translates to:
1. Comparing Voronoi movement with random movement under stationary con-
ditions (constant capacity and arrival times through entire run)
2. Comparing Voronoi movement with random movement under variable condi-
tions, namely a 'Saturday' scenario (varying arrival rates and capacity for rider
agents' entry and driver agent entry).
Therefore, to test our model we conducted 4 standard runs: A scenario-based set
of runs with a comparison of random movement and Voronoi movement, and a
stationary standard run with Voronoi movement and random movement. We also
conducted a number of verification and validation tests to ensure that the model
is run correctly as well as that it is running as intended. We summarize those
efforts in Table 5.
Table 5. Verification and Validation Methods
Goal
Method
Result
Verify that road network imported cor-
rectly
Display and check spatial cell attributes
Success
Verify that agents instantiate on roads,
specifically intersections.
Compared spatial coordinates of agents with cell
coordinates of intended intersection
Success
Verify that agents instantiate with the
correct attribute values
Verify that the movement of agents is
as intended
Verify that the model mechanics for
rider entry functions as intended
Verify that the rider capacity functions
as intended
Validate Pickup Mechanics
Validate Drop-off Mechanics
Validate agent attribute changes
through model runs
Displayed agent attributes at random
Success
Visual observation and numerous model runs
Success
Raised and lowered the rider entry value and
monitored expected increases or decreases in out-
puts
Success
Raised and lowered the rider entry value and
monitored expected increases or decreases in out-
puts
Success
Verified through the transfer of rider id and desti-
nation
Success
Observed attribute changes for increases in drop-
off values
Success
Through many runs and observations
Success
ORCID #: 0000-0003-2234-8756
19
3. Results
3.1 A Typical Run
We present typical model results for the 4 possible variations of our test runs.
Figure 5-7 and table 6-7 show plots and summary results of our model run. Figure
5-7 contain the total and accumulated financial gain of all agents with total profit
on the y-axis and time units on the x-axis. This measure includes the financial
gain made by agents who have left the simulation due to fatigue or for any other
reason.
We see no real and substantial difference in terms of total profit (financial
gain) between either of our stationary conditions model runs for random move-
ment and Voronoi movement (yellow and grey). That is-whether drivers chose
to 'spread out' or move about randomly in the hopes of picking up more custom-
ers did not affect, on average, their financial gain. In fact, the difference was com-
fortably within 2 standard deviations for both Voronoi and random movement
runs.
Remarkably, for the variable scenario runs a stark difference emerged between
the two spatial movement types, unexpectedly. Divergence in the profitability
between random movement choices and Voronoi movement choices for driver
agents was clear, and exhibited in both the total profit made by driver agents in a
24-hour run (Figure. 6-7) and in the summary statistics of the model run as a
whole (Table. 7).
Specifically, we can comfortably note that Voronoi movement for driver
agents provides greater utility (financial gains) for drivers when varying rates are
executed on the agents' arrival rates i.e. when a scenario is utilized. Where for
stationary model runs neither movement method prior to rider pickup provided
any visible change in driver agent utility. Consequently, our observations and
analysis of the model run took a focus on the variable scenario runs, and more
precisely on the moments of divergence of the variable scenario random move-
ment run when compared to the Voronoi movement variable scenario model run.
Table 6. Summary statistics for stationary run for both random and Voronoi movement
Stationary Random
Stationary Voronoi
Mean
$ 6,284.02 Mean
$ 6,483.38
Standard Error
$ 94.35
Standard
Error
$ 88.71
Median
$ 6,161.00 Median
$ 6,400.90
Standard Deviation
$ 3,580.49
Standard
Deviation
$ 3,366.40
Maximum
$ 12,426.55 Maximum
$12,059.45
ORCID #: 0000-0003-2234-8756
20
Fig. 5. This figure shows the total profit made by all driver agents over a 24-hour period-a
single run of the model for both random driver movement (gray) and Voronoi driver movement
(yellow). The figure shows that while there are stochastic gains made at different times in the
model run by both random and Voronoi movement behavior, there are no clear advantages in
utilizing either behavior type when the arrival conditions of drivers and riders are stationary.
Table 7. Summary statistics for variable properties run for random and Voronoi move-
ment
Variable (Saturday) Random
Variable (Saturday) Voronoi
Mean
$ 2,845.05
Mean
$ 3,931.65
Standard Error
$ 68.96
Standard Error
$ 87.31
Median
$ 1,828.90
Median
$ 3,332.50
Standard Deviation $ 2,617.00
Standard Devia-
tion
$ 3,313.05
Maximum
$ 8,195.65
Maximum
$ 10,565.45
ORCID #: 0000-0003-2234-8756
21
More importantly than the observation that Voronoi movement outperformed
random movement only in variable run conditions, is whether we can deduce pre-
cisely where the divergence between the two behaviors began to take shape under
variable entry conditions.
The first instance of divergence in a typical run occurs at around 1 PM (marked
on Figure 7) into the model run which we hypothesize as being representative of
"lunchtime" activity and model as being a constant driver agent capacity of 5
drivers and an arrival rate of new riders of 10. Both rates are a reduction from the
11 AM hour which had a maximum capacity of 50 driver agents and 20 respec-
tively. At 1.30 PM the capacity for new driver entry increased to 45 and the rate
of arrival of rider agents also increases to 25 (Table. 8). Figure 7 has both the
random movement and Voronoi movement drawn with separate y-axis on the
same time-scale (x-axis) so as to allow us a better visual comparison of both run-
types, and we can see that at this lunch-time hour a slight divergence of perfor-
mance begins to emerge, allowing drivers who are using Voronoi movement to
make placement decisions that outperform those that move randomly.
Table 8. This table shows the relevant arrival rates for drivers (left) and riders(right) for the
first point of divergence in the variable conditions model run.
11AM
12PM
1.30PM
50
5
45
20
10
25
Drivers starting their day for the Saturday
Brunch/Lunch
Lunch Time - low activity
Post-Lunch Rush
The most salient divergence between the performance of the two behaviors we
tested occurred at around the 3.30 PM time-period. Figure 7 shows the stark dif-
ference in performance and thus in utility-gain between the two behaviors. Our
scenario at this time-period calls for the decrease of both driver capacity and rider
entry from the 2.30 PM period (from 50, 15 to 25, 10, for driver and rider agents
respectively). Table 9 summarizes the relevant part of the scenario run.
Table 9. This table shows the relevant arrival rates for drivers and riders (left, right) for the
second point of divergence in performance of the variable conditions model run.
50
25
40
15
10
5
Post-Lunch Rush
Stationary Activity
Evening Drivers Beginning Their Shifts
2.30PM
3.30PM
5.30PM
ORCID #: 0000-0003-2234-8756
22
Thus, we saw a divergence in performance for a case when both driver and
rider entry were increasing simultaneously, and for a case where they were de-
creasing simultaneously as well, which dispels any notion that divergence in per-
formance would solely be due to a decreasing rate of one arrival rate while an-
other was increasing. We will propose candidate theories in the discussion section
of this paper.
The objective of our "lunchtime" change in both driver and rider agent demand
and supply was to create drastic changes similar to that what would be expected
in a major metropolitan area during this time period. What is critical to note is
that this performance difference-this emergent pattern-is only seen under var-
iable run conditions, and not stationary run conditions. Consequently, during a
statistical analysis for our model's stationary run, we find that the mean, median
and maximum financial gain (by all agents) during a 24-hour period was not sig-
nificantly different between driver agents employing a Voronoi movement versus
random movement behavior. Wherein the variable ("Saturday") scenario run, the
median, mean, and maximum were contrasted, with Voronoi movement outper-
forming random movement decisions on the aggregate; though the majority of
the performance improvements came from the time-periods where Voronoi
movement allowed a greater rate of customer pick-ups (1 PM and 3.30 PM). It
was not immediately apparent that Voronoi movement outperformed random
movement for every time-period of the variable run scenario.
ORCID #: 0000-0003-2234-8756
23
Fig. 6. This figure shows the total profit made by all agents over time. Not the divergence be-
tween Voronoi movement (Orange), and the random movement (Blue) under variable condi-
tions. Voronoi movement outperforms random movement.
Fig. 7. This figure utilizes 2 axes, one for the variable conditions scenario run (Saturday) Voro-
noi movement (Left axis) and another for the random movement (right axis). This makes for
easier visual comparison and yields insight into precisely which point in time the divergence
between the two behaviors would typically occur. In this particular run it is clear that Voronoi
movement begins to outperform random movement at 630 minutes.
ORCID #: 0000-0003-2234-8756
24
4. Discussion
Clearly, the model shows that driver agent positioning prior to rider agent
pickup influences the financial utility gained by driver agents (consequently this
also means that riders receive more consistent pickups with shorter wait-times).
But this emergent phenomenon is not recognizable unless a realistic variable ar-
rival rate scenario is utilized. I argue that this is true because of the spreading
nature of Voronoi movement. In highly volatile rider-supply and driver-demand
areas of our variable scenario runs, spreading-out behavior ensures that drivers
are more evenly distributed, and by being so they are better positioned to "catch"
riders in a moment of higher demand. Where, if driver agents choose to continue
moving and placing themselves randomly in moments of drastic change to supply
and demand, their catching behavior is set to be limited and thus are unable to
maximize their financial gain and adapt to their surroundings. This is the case for
when both rates of entry for drivers and riders are increasing and decreasing sim-
ultaneously, thus it should be noted that opposite signs for the first derivative of
the profit variable are not a requirement for this phenomenon to occur.
The details of when this occurred are also important. In the "Saturday" scenar-
ios, Voronoi movement did not outperform random movement in all variance
combinations. For example, Table 10 shows a portion of the variable scenario at
the 10.30 PM time-period. Note the decrease in driver capacity throughout the
listed time-periods from 90 to 75, while rider arrival rates remained constant and
then increased to 30 per time unit. Figure 7 shows that in this time-period there
was no divergence in performance between Voronoi and random movement, even
though elements of both the rise and fall seen in the scenario portions (where
visible drastic change was present) was also embedded in this particular sequence
of agent arrivals.
Table 10. A portion of the "Saturday" scenario that contains elements of the scenario portions
where performance divergence was seen, yet, no divergence emerged for this sequence.
10.30PM
90
10
Low Rider Activity - Riders are at their destina-
tions. Drivers still on the road expecting a rush of
new riders.
11.30PM
12.30AM
80
75
10
30
Some drivers give up, exit the simulation
More drivers give up. Riders beginning to end their
work shifts.
Thus, we must then conclude that sudden increase and decrease in rider de-
mand in conjunction with a steady or a slightly decreasing capacity (supply) pro-
vides a sudden spatial vacuum in the model's geographic configuration which is
best enclosed by agents who are actively trying to move away from each other –
ORCID #: 0000-0003-2234-8756
25
a Voronoi movement pattern. But, that this pattern-that allows drivers to cover
more space and 'catch' more riders-occurs only where there is enough drivers
on the spatial geography such that there are actual additional riders who will be
picked up by this movement. In other words, if decreasing or increasing variations
in both agents' entry are occurring, Voronoi movement will provide drivers with
an advantage over random movement, given that there are enough free drivers
(without a passenger) and enough riders (without a driver) ready for pickup; an
opportunity for maximizing utility must exist. After all, an increase in the spatial
spread of driver agents allows for an increased probability of executing a pickup
of a rider agent, but only when there are riders to be picked up.
The key to reproducing this pattern is that it must be part of a sudden and/or
variable change scenario for arrival rates. We hypothesize that a stationary run
will not emerge this phenomenon because, with unchanging arrival rates, incre-
mental improvements will not allow for a critical mass of spatial imbalance in the
location of riders and drivers.
There is a connected phenomenon observed in supply chain management the-
ory that can be associated with this system property-what is known as the bull-
whip effect. The phenomenon is widely understood as that of being a powerful
reaction at the far end of a long supply chain which is often created from a small
change in the point of origin of the chain. If the change is more sudden the effect
is more compounded. In this case, the effect can be seen in the time-delayed spa-
tial response of one agent group to another, not in a supply chain.
This behavior can be described as emergent. The pattern of maximizing utility
through the prior, strategic positioning of driver agents is somewhat unexpected
since all entry and exit of agents and their locations on the geography of the model
are random. One might surmise (incorrectly, as we have shown) that if the random
placement of rider agents and random placement of driver agents forms the core
of the topological interactions of agents in this model, that through intuition alone
there would be no clear gain in Voronoi movement behavior over random move-
ment behavior. But as we have shown that while this is true for constant arrival
rates, there is a difference in variable run conditions.
This emergent behavior does not seem to occur in the stationary runs of the
model because at a constantly random and stable rate of entry for both agent types
there is never a sudden vacuum to be capitalized upon.
This result provides researchers in this area with several important lessons: If
one seeks to test movement behaviors of a ridesharing or an autonomous vehicle
system, it should be tested under highly variable conditions in order to observe
true emergent behavior. Stationary testing of spatial models would seem to be
misleading and ineffective in this regard. Moreover, I propose that the testing of
agent-agent interaction on any topological space for which there exists an entry
or arrival dynamic would be subservient to the conclusions presented herein;
though additional testing remains to conclude so irrefutably.
It's also paramount to realize that typical system dynamics modeling (differ-
ential equations) would most likely fail in producing the phenomenon as we have
ORCID #: 0000-0003-2234-8756
26
observed it, assuming a closed form equation for the mechanics of the model can
be found, to begin with. Thus, we consider the experiment an added piece of im-
portant evidence for the utilization of agent-based techniques in autonomous ve-
hicle and ridesharing service modeling and simulation (theoretically the differ-
ence between autonomous vehicle and ridesharing service modeling is negligible
from the modeler's perspective.)
5. Summary & Future Work
In this paper, we described the development of an agent-based model for rides-
haring services in the Washington, D.C. area. The model simulates riders and
drivers through simple interactions on an accurate data-driven geospatial config-
uration. This model forms the basis for a number of experiments and model ex-
tensions that could yield greater insights into the ridesharing economy as it de-
velops, expands, and evolves.
Our conclusions showed the importance of running experiments utilizing agent-
based modeling runs not only in the form of stationary runs but in the form of
variable scenario runs designed to create unpredictable effects that can-and in
our case did-yield greater insights which otherwise would not have been ob-
served.
Specifically, we found the emergence of a pattern where prior positioning of
driver agents had a significant effect on pickup rates, and thus on the financial
gain (utility) of drivers. We also found that this pattern emerged from a simple
spreading-out behavior, which we called Voronoi movement and that this move-
ment pattern outperformed random movement patterns even with randomly dis-
tributed arrival rates for both agent types. However, this emergent phenomenon
was not observable unless a variable scenario was utilized in the experimentation
process.
Consequently, we showed that driver to driver agent interactions, which form
a symmetrical information environment can provide increased utility for drivers,
and consequently for the ridesharing service and riders as well in some cases.
Current operating procedures of the leading ridesharing services do not allow
drivers to gain access to location information of other drivers, and thus do not
allow for movement behaviors that are dependent on that additional information.
The symmetrizing of information can yield greater utility for all sides of this
equation, including service, rider, and driver. Perhaps ridesharing services believe
that giving less information to drivers would allow them more centralized control
which they can use to better optimize the ridesharing experience, but evidence
that this is true is not without question if we consider the natural fluctuations in
drivers' and customers' supply and demand. The question posed by us here is
whether more information may allow drivers to create adaptive and cooperative
ORCID #: 0000-0003-2234-8756
27
strategies to maximize their financial gain and by consequence, all other parties.
We showed that this may be the case for one simple behavior and we demon-
strated the conditions under which future spatial behavioral testing should be im-
plemented if we are to be confident in the outputs of our simulations.
There are many pathways that this model can take-going forward. Primarily,
the most interesting extension would be to add more spatially complex behaviors
in agent-to-agent interactions and then to observe the results. It is not trivial that
we test expected utility for an information symmetry scenario with only one
movement-type behavior. More spatial movement patterns grounded in expected
behaviors should be tested to quantify the difference in a service that allows more
drivers to have more information and one that does not. There are improvements
to be made in the spatial configuration of the model itself as well. For example,
the inclusion of spatial demographics to enrich probability distribution calcula-
tions, adding road direction and traffic patterns, as well as utilizing more detailed
geographic datasets would all make significant improvements to the model's ef-
ficacy and predictive power.
ORCID #: 0000-0003-2234-8756
28
References
1. Anderson, D. N. (2014). Not just a taxi: For-profit ridesharing, driver strategies, and VMT.
Transportation, 41(5), 1099–1117. doi:10.1007/s11116-014-9531-8
2. Bash, E. (2015). Dynamic Ridesharing: An Exploration of the Potential for Reduction in
Vehicle Miles Traveled. Ph.D. Proposal.
3. Cho, S., Yasar, A. U. H., Knapen, L., Bellemans, T., Janssens, D., & Wets, G. (2012). A
Conceptual Design of an Agent-based Interaction Model for the Carpooling Application.
Procedia Computer Science, 10, 801–807. doi:10.1016/j.procs.2012.06.013
4. Cho, S., Yasar, A.-U.-H., Knapen, L., Patil, B., Bellemans, T., Janssens, D., & Wets, G.
(2013). Social networks in agent-based models for carpooling. In Transportation Research
Board 92nd Annual Meeting (p. no. 13–2055).
5. Furuhata, M., Dessouky, M., F., Brunet, M. E., Wang, X., & Koenig, S. (2013). Rideshar-
ing: The state-of-the-art and future directions. Transportation Research Part B: Methodo-
logical, 57, 28–46. doi:10.1016/j.trb.2013.08.012
6. Hussain, I., Knapen, L., Galland, S., Yasar, A.-U.-H., Bellemans, T., Janssens, D., & Wets,
G. (2015). Agent-based Simulation Model for Long-term Carpooling: Effect of Activity
Planning Constraints. Procedia Computer Science, 52(Ant), 412–419.
doi:10.1016/j.procs.2015.05.006
7. Knapen, L., Keren, D., Yasar, A. U. H., Cho, S., Bellemans, T., Janssens, D., & Wets, G.
(2012). Analysis of the co-routing problem in agent-based carpooling simulation. Procedia
Computer Science, 10(270833), 821–826. doi:10.1016/j.procs.2012.06.106
8. Knapen, L., Keren, D., Yasar, A.-U.-H., Cho, S., Bellemans, T., Janssens, D., & Wets, G.
(2013). Estimating Scalability Issues While Finding an Optimal Assignment for Carpool-
ing. Procedia Computer Science, 19(Ant), 372–379. doi:10.1016/j.procs.2013.06.051
9. Martin, C. J. (2016). The sharing economy: A pathway to sustainability or a nightmarish
form of neoliberal capitalism? Ecological Economics, 121, 149–159.
doi:10.1016/j.ecolecon.2015.11.027
10. Nielsen, J. R., Hovmøller, H., Blyth, P. L., & Sovacool, B. K. (2015). Of "white crows"
and "cash savers:" A qualitative study of travel behavior and perceptions of ridesharing in
Denmark. Transportation Research Part A: Policy and Practice, 78, 113–123.
doi:10.1016/j.tra.2015.04.033
11. Nourinejad, M., & Roorda, M. J. (2015). Agent-based model for dynamic ridesharing.
Transportation Research Part C: Emerging Technologies, 1, 117–132.
doi:10.1016/j.trc.2015.07.016
12. Pelzer, D., Xiao, J., Zehe, D., Lees, M. H., Knoll, A. C., & Aydt, H. (2015). A Partition-
Based Match Making Algorithm for Dynamic Ridesharing. IEEE Transactions on Intelli-
gent Transportation Systems, 16(5), 2587–2598. doi:10.1109/TITS.2015.2413453
13. Rayle, L., Shaheen, S., Chan, N., Dai, D., & Cervero, R. (2014). App-Based, On-Demand
Ride Services: Comparing Taxi and Ridesourcing Trips and User Characteristics in San
Francisco. University of California Transportation Center, 94720(August), 1–20.
doi:10.1007/s13398-014-0173-7.2
14. Rayle, L., Dai, D., Chan, N., Cervero, R., & Shaheen, S. (2016). Just a better taxi? A sur-
vey-based comparison of taxis, transit, and ridesourcing services in San Francisco.
Transport Policy, 45, 168–178. doi:10.1016/j.tranpol.2015.10.004
16. Shaheen, S. A. (2016). Shared mobility innovations and the sharing economy. Transport
Policy, 1–2. doi:10.1016/j.tranpol.2016.01.008
17. Shaheen, S. A., Chan, N. D., & Gaynor, T. (2016). Casual carpooling in the San Francisco
Bay Area: Understanding user characteristics, behaviors, and motivations. Transport Pol-
icy, 1–9. doi:10.1016/j.tranpol.2016.01.003
ORCID #: 0000-0003-2234-8756
29
18. Tran, Q. T.; Tainar, D.; Safar, M. (2009). Transactions on Large-Scale Data- and
Knowledge-Centered Systems. p. 357. ISBN 9783642037214.
19. Voronoi, G. (1908). Nouvelles applications des paramètres continus à la théorie des
formes quadratiques. Premier mémoire. Sur quelques propriétés des formes quadratiques
positives parfaites. Journal Für Die Reine Und Angewandte Mathematik (Crelle's Jour-
nal), 1908(133). doi:10.1515/crll.1908.133.97
20. Wilensky, U. (1999). NetLogo. http://ccl.northwestern.edu/netlogo/. Center for Connected
Learning and Computer-Based Modeling, Northwestern University, Evanston, IL
21. Zargayouna, M., Zeddini, B., Scemama, G., & Othman, A. (2013). Agent-based simulator
for travelers multimodal mobility. Frontiers in Artificial Intelligence and Applications,
252, 81–90. doi:10.3233/978-1-61499-254-7-81
22. Zhang, W., Guhathakurta, S., Fang, J., & Zhang, G. (2015). Exploring the impact of
shared autonomous vehicles on urban parking demand: An agent-based simulation ap-
proach. Sustainable Cities and Society, 19, 34–45. doi:10.1016/j.scs.2015.07.006
23. Zolnik, E. J. (2015). The effect of gasoline prices on ridesharing. Journal of Transport Ge-
ography, 47, 47–58. doi:10.1016/j.jtrangeo.2015.07.009
|
1812.04848 | 1 | 1812 | 2018-12-12T08:43:12 | Optimal Prizes for All-Pay Contests in Heterogeneous Crowdsourcing | [
"cs.MA",
"cs.GT"
] | Incentives are key to the success of crowdsourcing which heavily depends on the level of user participation. This paper designs an incentive mechanism to motivate a heterogeneous crowd of users to actively participate in crowdsourcing campaigns. We cast the problem in a new, asymmetric all-pay contest model with incomplete information, where an arbitrary n of users exert irrevocable effort to compete for a prize tuple. The prize tuple is an array of prize functions as opposed to a single constant prize typically used by conventional contests. We design an optimal contest that (a) induces the maximum profit---total user effort minus the prize payout---for the crowdsourcer, and (b) ensures users to strictly have the incentive to participate. In stark contrast to intuition and prior related work, our mechanism induces an equilibrium in which heterogeneous users behave independently of one another as if they were in a homogeneous setting. This newly discovered property, which we coin as strategy autonomy (SA), is of practical significance: it (a) reduces computational and storage complexity by n-fold for each user, (b) increases the crowdsourcer's revenue by counteracting an effort reservation effect existing in asymmetric contests, and (c) neutralizes the (almost universal) law of diminishing marginal returns (DMR). Through an extensive numerical case study, we demonstrate and scrutinize the superior profitability of our mechanism, as well as draw insights into the SA property. | cs.MA | cs | T. Luo, S. S. Kanhere, S. K. Das and H-P. Tan, "Optimal prizes for all-pay contests in heterogeneous crowdsourcing", Proc. IEEE MASS, 2014, pp. 136-144.
Optimal Prizes for All-Pay Contests in
Heterogeneous Crowdsourcing
Tie Luo∗, Salil S. Kanhere†, Sajal K. Das‡, Hwee-Pink Tan∗
∗Institute for Infocomm Research, A*STAR, Singapore
†School of Computer Science and Engineering, The University of New South Wales, Australia
‡Department of Computer Science, Missouri University of Science and Technology, USA
E-mail: [email protected], [email protected], [email protected], [email protected]
8
1
0
2
c
e
D
2
1
]
A
M
.
s
c
[
1
v
8
4
8
4
0
.
2
1
8
1
:
v
i
X
r
a
Abstract -- Incentive is key to the success of crowdsourcing which
heavily depends on the level of user participation. This paper
designs an incentive mechanism to motivate a heterogeneous crowd
of users to actively participate in crowdsourcing campaigns. We
cast the problem in a new, asymmetric all-pay contest model
with incomplete information, where an arbitrary n of users exert
irrevocable effort to compete for a prize tuple. The prize tuple is
an array of prize functions as opposed to a single constant prize
typically used by conventional contests. We design an optimal
contest that (a) induces the maximum profit -- total user effort
minus the prize payout -- for the crowdsourcer, and (b) ensures
users to strictly have incentive to participate. In stark contrast
to intuition and prior related work, our mechanism induces an
equilibrium in which heterogeneous users behave independently of
one another as if they were in a homogeneous setting. This newly
discovered property, which we coin as strategy autonomy (SA), is
of practical significance: it (a) reduces computational and storage
complexity by n-fold for each user, (b) increases the crowdsourcer's
revenue by counteracting an effort reservation effect existing in
asymmetric contests, and (c) neutralizes the (almost universal)
law of diminishing marginal returns (DMR). Through an extensive
numerical case study, we demonstrate and scrutinize the superior
profitability of our mechanism, as well as draw insights into the
SA property.
Index terms -- Incentive mechanism, all-pay auction,
asymmetric contest, strategy autonomy, participatory sens-
ing, network economics.
I. INTRODUCTION
Crowdsourcing offers a cost-effective approach to distributed
problem solving and data collection by soliciting contributions
(solutions, ideas, data, etc.) from a large group of people. Com-
pared to conventional means of hiring employees, crowdsourc-
ing can be potentially more cost-efficient. It has thus catalyzed
new computing and sensing paradigms such as participatory
sensing [1].
Key to the viability of crowdsourcing is providing incentives
to attain a sufficient level of user participation. While incentives
can be classified into intrinsic motivation (e.g., self-fulfillment,
enjoyment, and esteem) and extrinsic incentives (e.g., peer pres-
sure, financial rewards) [2], we focus on monetary incentives
which fall in the second category and have wider applications
in practice. Specifically, in this paper, we employ the theory
of mechanism design, in particular auctions [3], to design an
incentive mechanism for crowdsourcing.
We choose auction theory to be the tool primarily because,
to reward users for their contributions, a unilaterally stipulated
pricing scheme -- say by some authority -- for such informational
products as ideas, data, solutions, etc., is hard to satisfy both
the crowdsourcer and users. On the other hand, auctions offer
a perfect mechanism for a principal (crowdsourcer) and agents
(users) to implicitly negotiate and mutually agree on a deal
(e.g., solving a problem, contributing data). Indeed, auctions as
a bilateral pricing scheme have been widely adopted by a sizable
body of literature [4] -- [7].
In this paper, we design an incentive mechanism by modeling
crowdsourcing as an all-pay contest [8]. All-pay contests are
isomorphic to all-pay auctions [3] -- given an equilibrium in
one model, one can construct one and only one equilibrium
in the other model1 -- yet contests are semantically closer to
our mechanism which uses prizes as auctioned goods. Auctions
have many flavors in which the most common and well-known
ones, such as English (or second-price) and Dutch (or first-
price) auctions, belong to the intuitive "winner-pay" genre --
only the winner (i.e., the highest bidder) pays for a bid. On
the contrary, "all-pay" auctions require every bidder to pay for
his bid regardless of who wins the auction. This seemingly
peculiar form, however, precisely describes most crowdsourc-
ing platforms (e.g., Amazon mechanical turk (AMT), task.cn,
TaskRabbit.com) in which there is only one winner but all
participants will have to exert their irrevocable effort before
the winner is announced. This is why we model crowdsourcing
as an all-pay auction or contest.
The vast majority of prior work using auctions (either all-
pay or winner-pay) [5] -- [7], [9] -- [11] is based on symmetric
auctions where bidders are ex post or ex ante identical, i.e.,
either they have exactly the same type2 (ex post), or their
types follow the same, single probabilistic distribution (ex ante).
As a result, they all model an environment of homogeneous
agents, which is amicable to analysis. However, in many real
crowdsourcing scenarios (e.g., AMT, task.cn, TaskRabbit.com,
TopCoder.com), users come from dissimilar backgrounds and
possess different skill levels, constituting a heterogeneous en-
vironment. Therefore, a more realistic model would be asym-
metric auctions. Unfortunately, asymmetric auctions are much
less studied and understood because their associated analysis
is much more challenging, due to the fact
the most
celebrated revenue equivalence theorem3 [12], [13] breaks under
asymmetry. Thus, as a compromise for analytical tractability,
most research on asymmetric auctions is limited to two-player
cases [14] -- [16] or complete-information settings [8], [17] in
which all types are commonly known. A significant progress
that
1Drawing an analogy may shed light on the isomorphism: contestants exerting
effort in a contest is like bidders tendering bids in an auction, and the highest-
effort contestant winning the prize is like the highest bidder securing the
auctioned good.
2In Bayesian games and mechanism design, type is a term that refers to an
agent's private information or signal (e.g., ability, skill level, valuation of the
auctioned good), which essentially characterizes an agent.
3This theorem states that, under a set of standard assumptions including
symmetric agents, any auction generates the same revenue for the principal.
was made by [18] toward understanding asymmetric first-price
auctions with multiple players and incomplete information, but
only an approximate solution was obtained and it only applies
to weakly asymmetric agents. Other studies in this context resort
to numerical methods [19].
Unlike prior work, our model accommodates asymmetric --
regardless of weak or strong -- all-pay contests with an arbitrary
n of players and incomplete information (i.e., agent types are
uncertain). This is a much closer characterization of most
real crowdsourcing scenarios. Furthermore, we obtain precise,
analytical solutions rather than approximate or numerical ones.
Thus, this work presents the first attempt to offer a rigorous
understanding toward more realistic crowdsourcing campaigns,
with results of a much wider applicability. This constitutes the
first contribution of this paper.
The second contribution is that we explore another degree of
freedom in contest design, by furnishing the contest with a prize
tuple that consists of an array of n prize functions to cater for
the heterogeneous agents. This is distinct from all the standard
contests or auctions where a single fixed prize or good is used.
The rationale for this exploration is to understand whether and
how such a prize tuple can elicit higher profit -- revenue (agent
effort) minus cost (prize payout) -- for the principal. In this
paper, we derive the optimal prize tuple under asymmetry, and
demonstrate that it can induce significantly higher profit than
both symmetric and asymmetric fixed-prize contests.
this paper is the first
The third contribution is that
to
incorporate the principal's valuation of the prize into profit
formulation, thereby enabling our model and results to apply
to different crowdsourcers. It also means that, even if a contest
adopts a fixed prize, the cost (valuation of the prize) can still
be variable. Therefore, regardless of fixed or variable prizes,
this work provides an example of how to cover different
crowdsourcers or the same crowdsourcer's varying valuations.
Our fourth contribution, which we stress, is the discov-
ery and investigation of a new and counter-intuitive property
pertaining to asymmetric auctions and contests, called strategy
autonomy (SA). It captures the phenomenon that agents in an
asymmetric equilibrium behave independently of one another
as if they were in a symmetric one. This is in stark contrast to
all prior work on asymmetric auctions, and has three practical
significances: it (a) reduces computational and storage complex-
ity from O(n) to O(1) for each agent, (b) increases principal's
revenue by counteracting an effort reservation effect engendered
by asymmetric belief, and (c) dramatically enhances the system
scalability by neutralizing the (almost universal) law of dimin-
ishing marginal returns (DMR). Moreover, in addition to SA,
our mechanism also strictly satisfies individual rationality (IR),
which means that all agents strictly have incentive to participate
in our mechanism.
The rest of this paper proceeds as follows. Section II reviews
the literature. Section III presents our model and Section IV
provides the analysis. A detailed numerical case study is then
given in Section V which demonstrates key results, offers
intuitions, as well as draw some insights. Section VI concludes
the paper.
II. RELATED WORK
A. Incentive mechanisms and Symmetric auctions
While auctions are widely recognized as an effective incentive
mechanism for various activities such as online sales, govern-
2
ment procurement, crowdsourcing and participatory sensing, the
vast majority of prior work on auction-based incentive presumes
symmetric auctions. For example, in [5] and [6], users bid
their costs and the crowdsourcer or service provider determines
their payments such that the users bid their costs truthfully.
Along a different spirit, [7] investigates an all-pay auction
based incentive mechanism that is tailored to realistic settings,
including uncertain population sizes, unknown (yet symmetric)
user types, and risk-averse (subsuming risk-neutral) users.
Other studies on optimal design of auctions or contests
consider different approaches and objectives. For example, [11],
[20] investigate whether a single or multiple prizes is optimal in
terms of maximizing the highest k bids [11] or total bids [20].
Under a similar setting, [9] shows that the highest bid is at least
half of the total bids, and [21] finds that there is no advantage
to have multiple prizes under certain conditions. Similarly
in a symmetric model, [22] introduces variable rewards and
examines a paradoxical behavior where a reduction in reward
or an increase in cost may increase the total or the highest bid
in expectation.
Indeed, almost all auction theory, including the above, con-
cerns symmetric auctions, as concurred by [23]. On the contrary,
this paper tackles the challenge of asymmetry in order to
accommodate heterogeneous and more realistic crowdsourcing
environments.
B. Asymmetric auctions
This domain is relatively much less understood due to its
analytical complexity. Most work in this domain is devoted
to two-player cases or complete-information settings for the
sake of tractability. Amann and Leininger's seminal work [15]
offers an analysis of the equilibrium strategies for a two-
player asymmetric case. It was then extended by Maskin and
Riley [24], [25] who proved the monotonicity and uniqueness
of the equilibrium. Unfortunately, till now, there still lacks a
closed-form solution to general n-player cases with incomplete
information. Fibich and Gavious [18] proposed a perturbation
approach to obtain an analytical, but approximate, solution to
equilibria on the premise of weak asymmetry. Improvement in
terms of some other mathematical properties was made later
using a dynamic-system approach [23]. Another work [26] on
asymmetric contests focuses on risk aversion and gives some
exploratory yet inconclusive results.
Other studies assume complete information. Siegel [8] was
probably the first who coined the term "all-pay contests",
where he analyzed closed-form player payoffs in equilibrium
with complete information. Under a similar model, Xiao [17]
studied the problem of allocating more than one prizes to a
number of winners and proposed an algorithm to construct the
equilibrium. Franke et al. [27] aimed to maximize the revenue
(which is part of the profit we study in this paper) through
discriminating players by associating differentiated weights to
players, assuming complete information is available.
Moreover, all the above studies presume fixed prizes.
In this paper, we allow for more generality by not assuming
the availability of complete information and by accommodat-
ing multiple asymmetric players. Secondly, we empower the
crowdsourcer to provision a tuple of individually different prize
functions, in order to induce the highest possible profit. All these
are clearly different from prior art.
III. THE MODEL
The overall problem setting is that a principal aims to
crowdsource from n heterogeneous agents for the maximal
profit -- total effort (revenue) minus prize (cost). In this paper,
"effort" is a general term that can be interpreted according to
different contexts; for example, it can refer to the quantity of
tasks completed, the quality of solutions submitted [28], or a
quality-modulated quantity of sensing data [29].
We design this crowdsourcing campaign under an all-pay
contest framework, in which all participating agents exert effort
(which is irrevocable and will be sunk, i.e., "all-pay", regardless
of the outcome) in order to win some prize (i.e., as if in a "con-
test"). An agent i = 1, 2, ..., n is characterized by his type vi
which is private information (e.g., skillfulness or competency):
agent i knows his own type vi but does not know any other
vj,∀j (cid:54)= i. On the other hand, it is common knowledge that
all the agent types vin
i=1 are independently drawn from [v, ¯v]
according to Fi(v)n
i=1, respectively, where Fi(v) are the c.d.f.
of vi and [v, ¯v] is a continuous and nonnegative support. This
setting corresponds to an incomplete-information setting and,
essentially, constitutes an asymmetric Bayesian game where the
common prior consists of n generally different distributions.
Without loss of much generality, we assume that each Fi(v) is
differentiable and the corresponding p.d.f. fi(v) is continuous
and positive over (v, ¯v).
To provide incentive, the principal will reward the agent who
exerts the highest effort, i.e., the "winner", a prize. In this
paper, we allow the prize to depend on the winner's effort
rather than fixing the prize ex ante. Furthermore, in view of
the asymmetric agents, we provision the prize as a prize tuple
Z := (cid:104)Z1(b1), Z2(b2), ..., Zn(bn)(cid:105), where Zi(bi) is a prize
function in agent i's effort bi and takes effect when i is the
winner. This prize tuple is known to all the agents before the
contest.
The value of a prize to an agent is characterized by a value
function V (v, Z): an agent of type v values a prize Z (the face
value) to be of real worth V (v, Z) to him. That is, agent i values
the prize he competes for to be of real value V (vi, Zi(bi)). An
example of the value function is V (v, Z) = vZ, meaning that an
agent with higher ability can gain more benefit out of a prize; in
fact, V (·) = vZ is a generalization of standard all-pay auctions
where the fixed auctioned good is normalized and V (·) = v.
With out loss of much generality, we assume that V (v, Z) is
differentiable with respect to v.
Exerting effort incurs cost; an agent i who exerts effort bi has
to pay for his cost as per a payment function p(bi, vi). We do
not assume a specific form for p(·) or V (·), thereby covering
a broad class of auctions and contests.4 Now we can formulate
the expected utility of agent i, as
ui := qiV (vi, Zi(bi)) − p(bi, vi)
(1)
where qi is the probability that agent i wins the contest.
The principal aims to maximize his expected profit π, defined
as the total crowdsourced effort minus his prize payout, or
formally,
π := E(cid:2)(cid:88)
bi − V (λ, Zw(bw))(cid:3)
(2)
i
4For example, letting p(b, v) = b and V (v, Z(b)) = v yields standard all-
pay auctions; letting p(b, v) = 0 and V (v, Z(b)) = v − b yields standard
first-price auctions.
3
where w ∈ [1...n] is the winner's index which is a random
variable, and λ > 0 is the principal's type (valuation of prize)
which is common knowledge. In this profit formulation, the
revenue is defined in terms of total effort, which covers data-
gathering kinds of applications; it can also be defined in terms
of the highest effort, which covers solution-eliciting kinds of
applications, but the corresponding analysis completely parallels
this paper.
Throughout, we follow the notation convention of g(cid:48)
x := ∂g
∂x,
g(cid:48)(cid:48)
xy := ∂2g
∂x2∂y , for any differentiable function
g(x, y). We assume that the payment function p(b, v) is twice
continuously differentiable, p(0, v) = 0, p(cid:48)
b(b, v) > 0 which
v(b, v) ≤ 0 which
means higher effort, higher payment (cost); p(cid:48)
means higher type (ability), lower payment; p(cid:48)(cid:48)
bb(b, v) > 0 which
means striving from higher effort levels is more costly than from
lower effort levels, or conversely, the marginal output by adding
b2v(b, v) ≤ 0 which means lower
effort is decreasing; finally, p(cid:48)(cid:48)(cid:48)
types are more vulnerable to the decreasing marginal output.
We note that these assumptions are not restrictive.
∂x∂y , and g(cid:48)(cid:48)(cid:48)
x2y := ∂3g
In practice, our mechanism modeled in this section can
be realized as follows. The agent type distributions Fi can
be constructed by the crowdsourcer from user contribution
history, or from social acquaintance in the case of a small
local community of users, and then published on a website or
via a mobile app. The prize tuple can be either published or
downloaded, where each user has software (e.g., a mobile app)
to act as his agent. The user effort can be measured (in terms of
time, data volume, sampling rate, etc.) (a) by each agent itself,
or (b) by the principal (e.g., in a cloud) and continually fed
back to each corresponding agent, so that the agent can decide
how much effort to contribute.
IV. ANALYSIS
We first analyze the asymmetric equilibrium strategy for each
agent (Lemma 2), which is a function of any given prize tuple.
Then we determine the optimal prize tuple that induces the
maximum profit for the principal (Theorem 1). Following that
is an exposition of three important properties (Section IV-C).
A. Equilibrium strategy
Definition 1 (Bayesian Nash equilibrium). A pure-strategy
∗
Bayesian Nash equilibrium is a strategy profile b
:=
(b∗
1, b∗
n) that satisfies
2, ..., b∗
ui(b∗
i , b∗
−i) ≥ ui(bi, b∗
−i),∀bi,∀i.
In words, each agent in a Bayesian Nash equilibrium plays a
strategy that maximizes his expected payoff given his belief
about other agents' types and that other agents play their
respective equilibrium strategies.
Lemma 1 (Existence, monotonicity, uniqueness, and common
support). Our asymmetric all-pay contest with incomplete in-
formation admits a unique, asymmetric, pure-strategy Bayesian
Nash equilibrium that is strictly monotonic, i.e., the equilibrium
bids are strictly increasing in type. Furthermore, given that the
agent types have a common and nonnegative support, i.e., [v, ¯v],
the equilibrium bids also have a common support [0, ¯b], where
¯b is unknown.
We defer the proof to [30] due to space constraint.
Notation convention: Henceforth, we will exclusively deal
with the equilibrium state. Hence for brevity, we slightly deviate
i (·).
from the general national convention by dropping the superscript
∗ on equilibrium variables. For example, we write bi instead of
i and vi(·) instead of v∗
b∗
Lemma 1 tells that an agent's equilibrium strategy bi is a
strictly monotone (increasing) function of vi, which we denote
by βi(·), i.e., bi = βi(vi). Thus, its inverse function exists and is
also increasing, which we denote by vi(·) := β−1
(·). Moreover,
because of the strict monotonicity, the event bi = bj is of zero
probability and tie-breaking is trivial. Thus,
i
It is equivalent to
arg max
bi
Zi(bi)
(cid:89)
j(cid:54)=i
Fj(vj(bi)) − p(bi, vi)
4
(6)
where p(b, v) = p(b, v)/h(v) for v > 0.
Theorem 1. The optimal prize tuple that maximizes the princi-
pal's profit is given by Z = (cid:104)Z1(b1), Z2(b2), ..., Zn(bn)(cid:105) where
p(bi, vi(bi)) −(cid:82) bi
(cid:81)
(bi, vi(bi)) dvi(bi)
0 p(cid:48)
j(cid:54)=i Fj(vj(bi))
vi
(cid:81)
Pr(bi > bj) = Pr(β−1
j
(bi) > vj) = Fj(vj(bi)).
Zi(bi) =
Furthermore, because agent i's winning probability qi =
j(cid:54)=i Pr(bi > bj), (1) can be rewritten as
ui = V (vi, Zi(bi))
Fj(vj(bi)) − p(bi, vi).
(3)
in which the optimal effort bi(vi) is implicitly given by
p(cid:48)
bi
(bi, vi) =
1
h(λ)
+ p(cid:48)(cid:48)
bi,vi
(bi, vi)
1 − Fi
fi
Lemma 2. Given a prize function Zi(·), an agent i's equilib-
rium strategy bi(vi) is determined by
V (vi, Zi(bi))
Fj(vj(bi)) − p(bi, vi)
(cid:89)
j(cid:54)=i
[V (cid:48)
vi
(vi, Zi(bi))
Fj(vj(bi)) − p(cid:48)
vi
(bi, vi)] dvi.
(4)
(cid:90) vi
=
v
Proof: As the equilibrium bid bi is also the solution to the
optimization problem maxbi{ui} (3), we invoke the envelope
theorem [31] on (3) with respect to vi and obtain
Fj(vj(bi)) − p(cid:48)
(vi, Zi(bi))
= V (cid:48)
(bi, vi)
vi
vi
where vi(bi) is the inverse function of bi(vi). The maximum
profit achieved is given by
π =
bi(vi) − h(λ)p(bi, vi) + h(λ)p(cid:48)
(bi(vi), vi)
vi
dFi.
(cid:90) ¯v
(cid:104)
(cid:88)
v
i
(9)
Proof: First, we expand the principal's expected profit (2)
by calculating the expected cost, i.e., prize. Noticing that an
j(cid:54)=i Fj(vj(bi)), we use
agent i's winning probability is qi =(cid:81)
(cid:89)
the law of total expectation to have
(cid:88)
(cid:90) ¯v
Zi(bi(vi))
Fj(vj(bi)) dFi(vi).
Ew[Zw(bw)] =
(7)
(8)
(cid:105)
1 − Fi
fi
(cid:89)
j(cid:54)=i
(cid:89)
j(cid:54)=i
(cid:89)
(cid:89)
j(cid:54)=i
j(cid:54)=i
⇒ ui(vi) = ui(v)
∂ui
∂vi
(cid:90) vi
(cid:104)
+
v
V (cid:48)
vi
(vi, Zi(bi))
Fj(vj(bi)) − p(cid:48)
(bi, vi)
dvi.
vi
(5)
Since an agent with the lowest possible type never wins the
auction, he will bid zero (i.e., exert no effort) in an all-pay
auction (rather than bidding bi = v as in first or second-price
auctions). As a result, he reaps zero utility, i.e., ui(v) = 0.
Then, equating the r.h.s of (5) to that of (3) yields the result.
Recall in (4) that the derivatives in the integrand denote partial
derivatives, and hence no further reduction is allowed.
Remark: Asymmetric auctions, regardless of winner-pay or
all-pay, do not have closed-form expressions for equilibrium
strategies in general (while an approximate solution to the first-
price flavor can be found in [18]). However, we can solve for
the optimal prize tuple using Lemma 2 without obtaining the
closed form of equilibrium strategies, as shown next.
B. Optimal prize tuple
Solving for the optimal, i.e., profit-maximizing, prize tuple Z
requires an explicit form of the value function V (·), for which
we consider V (v, Z) := h(v)Z where h(·) satisfies h(0) = 0
and h(cid:48)(v) > 0. This form further generalizes the form V = vZ
which, as mentioned in Section III, is already a generalization
of the standard all-pay auctions.
First note that the utility maximization problem of each agent,
{ui} (3), can be reformulated as arg maxbi{ui} without
max
bi
change in principle, i.e.,
(cid:89)
j(cid:54)=i
(cid:90) ¯v
(cid:104)(cid:88)
(cid:90) ¯v
(cid:104)
(cid:88)
v
i
v
i
π =
=
(cid:105)
j(cid:54)=i
Then, by expanding the revenue portion,
v
i
(cid:105) − h(λ)Ew[Zw(bw)]
bi(vi) dFi(vi)
bi(vi) − h(λ)Zi(bi(vi))
Fj(vj(bi))
(10)
With Lemma 2, we substitute Zi(bi) for V (vi, Zi(bi)) in (4)
(cid:105)
dFi(vi).
(cid:89)
j(cid:54)=i
(cid:90) vi
v
p(cid:48)
vi
(bi(vi), vi) dvi,
(11)
where note that Zi
Substituting (11) into (10) yields
(cid:48)
vi
(bi) = 0 due to the envelope theorem.
π =
bi(vi) − h(λ)p(bi, vi) + h(λ)
p(cid:48)
vi
(bi(vi), vi) dvi
dFi.
(cid:105)
(cid:90) vi
v
and p(bi, vi) for p(bi, vi), and obtain
Fj(vj(bi)) − p(bi, vi) = −
Zi(bi)
(cid:89)
j(cid:54)=i
i
v
(cid:90) ¯v
(cid:104)
(cid:88)
(cid:90) ¯v
(cid:90) vi
(cid:90) ¯v
(cid:90) ¯v
=
v
v
v
=
v
By integrating the last term by parts,
p(cid:48)
vi
(bi(vi), vi) dvi dFi
(cid:90) ¯v
v
dFi.
v(bi(vi), vi) dvi −
p(cid:48)
1 − Fi
(bi(vi), vi)
p(cid:48)
vi
fi
Fi(vi)p(cid:48)
vi
(bi(vi), vi) dvi
arg max
bi
h(vi)Zi(bi)
Fj(vj(bi)) − p(bi, vi).
Substituting this back proves the principal's profit (9).
The principal originally faces the problem of maxZ{π}, yet
it is equivalent to maxb{π} because, essentially, the prize tuple
is used to induce the optimal effort vector b. Furthermore, in
(9) we have decoupled each agent i from other agents j (cid:54)= i.
Therefore, maximizing π can be achieved by maximizing each
individual integrand Ii over bi, where
Ii := bi(vi) − h(λ)p(bi, vi) + h(λ)p(cid:48)
vi
(bi, vi)
1 − Fi
.
fi
Simply applying the first order condition to Ii with respect to
bi proves the optimal effort bi (8) for each agent i.
To verify that the above effort bi is the unique maximizer,
we examine the second derivative
(cid:48)(cid:48)
b2
i
Ii
= −h(λ)p(cid:48)(cid:48)
b2
i
(bi, vi) + h(λ)p(cid:48)(cid:48)(cid:48)
b2
i vi
(bi, vi)
1 − Fi
fi
.
Since p = p/h(v), and v > 0 is treated as a fixed value due
to the envelope theorem, our assumptions on p(·) also hold for
≤ 0. Since h(λ) > 0 for λ > 0,
p(·), i.e., p(cid:48)(cid:48)
b2
therefore I(cid:48)(cid:48)
i
i < 0. Thus Ii is strictly concave, which validates
the existence and uniqueness of bi given by (8).
> 0 and p(cid:48)(cid:48)(cid:48)
b2
i vi
The optimal prize function (7) is then obtained, by rearrang-
ing (11) and changing variables from vi to bi. Note that the
lower limit of integral, 0, is due to bi(v) = 0 as the lowest-type
agent will bid zero (cf. proof of Lemma 2).
C. Qualitative properties
This section states three properties pertaining to our mech-
anism, denoted by OPT since it is equipped with an optimal
prize tuple.
1) Strategy Autonomy (SA): This is perhaps the most salient
property of OPT, particularly in the presence of asymmetry.
SA is of practical significance and none of prior work on
asymmetric mechanisms possesses this property.
Definition 2 (Strategy Autonomy). A mechanism satisfies strat-
egy autonomy if, given that the common prior Fi(vi)n
i=1 are
individually different, the equilibrium strategy bi(viF−i) =
bi(vi),∀i.
In words, in an asymmetric incomplete-information setting
where agents are ex ante heterogeneous, each agent adopts in
equilibrium a strategy that is independent of (the prior about) the
other agents. In other words, despite an asymmetric environment
(belief), agents behave autonomously as if they were in a
symmetric one.
Proposition 1. The OPT mechanism satisfies strategy autonomy.
Proof: Immediately follows from Theorem 1, where the
equilibrium strategy bi (8) is independent of any j (cid:54)= i.
This is a rather counter-intuitive, and somewhat surprising
result. This is because Lemma 2 shows that the equilibrium
strategy bi does depend on Fjj(cid:54)=i, or the strategy is not
autonomous, which also conforms to our intuition as the en-
vironment is asymmetric. Indeed, SA is in direct opposition to
all prior work on asymmetric auctions, regardless of winner-pay
or all-pay, with complete or incomplete information; see, e.g.,
[15], [17], [18] and a comprehensive survey [32]. So the key
question is: why do agents behave autonomously in the OPT
mechanism?
The fundamental reason is that the asymmetric belief about
agent types is endogenized (i.e., "absorbed") by the optimal
prize functions (7), or in other words, any bidder i's reasoning
about other bidders' bids is implicitly captured by the function
5
Zi(·). The rationale of this,
isolating asymmetry from
agents, is to counteract an effort reservation effect arising from
asymmetric belief, which is explained below.
i.e.,
SA has three important practical implications:
• Reduces complexity and saves energy: SA remarkably
reduces the computational complexity and storage requirement,
from O(n) to O(1), for each agent. The O(n) can be understood
from (4) where each agent's strategy involves reasoning about
all the j (cid:54)= i, which is also the case in, e.g., [18], [23], [27]. This
advantage enables each agent, which is embodied in practice
typically by software that resides on distribute portable devices
(e.g., mobile phones), to shed substantial computational and
storage burden and, as a result, save considerable energy.
• Counteracts effort reservation: SA overcomes an effort
reservation effect that exists in standard (fixed-prize) asymmet-
ric auctions [33]: when the prize is fixed, any agent only needs to
win the other agents by an infinitesimal winning margin; there-
fore, using a two-player scenario to illustrate, if the stronger
agent believes that the other agent is statistically weaker, he
has the incentive to reserve effort in order to reduce his winning
margin since a larger margin does not make the winner better
off at all. This effect outweighs the strategy adjustment of the
weaker agent and results in a reduced total revenue compared
to symmetric auctions [33], which will also be demonstrated
in Section V. However, SA insulates such negative inter-agent
influence, allowing agent not to be concerned with other agents
and to concentrate on exerting higher effort to increase the
winning margin which is now qualified for the (variable) prize.
• Neutralizes the law of DMR: The prevailing law of di-
minishing marginal returns (DMR) governs many phenomena
in (network) economics. It states that, as the number of new
employees increases,
the marginal product of an additional
employee will at some point be less than the marginal product
of the previous employee [34]. Mathematically, DMR leads to
a concave growth of profit or revenue as employees are being
added, which is also demonstrated by our evaluation of a stan-
dard (fixed-prize) auction in Section V (Fig. 4). However, the
independence connoted by SA neutralizes this submodularity-
resembling law of DMR, and indeed, we will show in Section V
that the principal's profit increases linearly as the number of
agents increases. This translates to a dramatically enhanced
system scalability.
2) Individual Rationality (IR):
Definition 3 (Individual Rationality). A mechanism satisfies
individual rationality if, in equilibrium, all participating agents
expect (weakly) higher surplus than from not participating. That
is, ui(bi, b−i) ≥ ui(0, b−i) for all i in equilibrium.
In other words, a mechanism satisfying IR ensures that any
agent has incentive to participate.
Proposition 2. The OPT mechanism satisfies individual ratio-
nality. In particular, an agent reaps strictly positive utility if he
exerts positive effort.
Proof: Combining (11) and (6) yields
(cid:90) vi
ui
= −
h(vi)
⇒ ui = −h(vi)
v
p(cid:48)
vi
(cid:90) vi
(bi(vi), vi) dvi
p(cid:48)
vi
(bi, vi)h(vi) − p(bi, vi)h(cid:48)(vi)
dvi.
v
h2(vi)
Based on the monotonicity of equilibrium (Lemma 1),
the
assumptions on the payment function p (cf. Section III) imply
that p(b, v) > 0 for any b > 0 = b(v). Furthermore, as
v(b, v) ≤ 0 and h(v) ≥ 0, we conclude that ui ≥ 0,
h(cid:48)(v) > 0, p(cid:48)
which proves IR, and that the equality holds iff vi = v (which
subsumes the case of vi = 0 since v ≥ 0). However, an agent
of vi = v will choose not to participate (bi = 0) as explained in
the proof of Lemma 2. Therefore, any agent who exerts nonzero
effort reaps a strictly positive payoff.
3) Incentive Compatibility (IC): We say a (direct revelation)
mechanism satisfies IC if agents will report their types truthfully.
In our mechanism, prize allocation is based on agents' observ-
able efforts and the common prior about all the agents' types,
instead of on reported (if any), unobservable types. Therefore,
the issue of truthful type-reporting is technically irrelevant to
our mechanism.
V. CASE STUDY
In order to derive an intuitive understanding and draw further
insights, we provide a numerical case study that involves six
mechanisms: OPT, FIX, SYM-1, SYM-2, OPT-n, and FIX-n.
OPT is instantiated with two agents of types v1, v2 ∈ [0, 1]
which are independently drawn from F1(v) = v (uniform
2 , respectively. Hence, f2(v) =
distribution) and F2(v) = v+1
2 where δ(·) is the Dirac delta function.5 That is,
2 δ(v) + 1
1
agent 2 is of type zero with probability 0.5, and draws his
type uniformly from (0,1] with probability of the other half; so,
agent 1 is statistically stronger than agent 2. The value function
V (v, Z) = vZ and the payment function is p(b, v) = b2. Hence,
h(v) = v and p(b, v) = b2/v.
We compare OPT with all counterpart mechanisms which are
three canonical auctions:
• FIX: Fixed-prize asymmetric all-pay auctions.
• SYM: Fixed-prize symmetric all-pay auctions, including
-- SYM-1: both types follow F1(v);
-- SYM-2: both types follow F2(v).
In order to investigate how SA neutralizes the law of DMR
and enhances scalability, which requires a larger-scale simula-
tion, we compare OPT-n and FIX-n, which are OPT and FIX
both with n symmetric agents (choosing agent 1 for illustration).
A. Theoretical underpinnings
To carry out the comparison, we need the following analytical
results for FIX and SYM. The proofs are available in [30].
Proposition 3 (Equilibrium strategy in FIX). In a two-player
asymmetric all-pay contest with incomplete information, if the
common prior is F1(v), F2(v), v ∈ [v, ¯v], and the payment
function p(b) satisfies p(0) = 0 and p(cid:48)(b) > 0, there exists a
unique Bayesian Nash equilibrium b = (b1, b2) which is given
by
b1(v1) = p−1(cid:16)(cid:90) v1
k−1(v)
b2(v2) = b1(k−1(v2)),
(cid:17)
k(v)F (cid:48)
1(v) dv
,
(12)
(13)
5Neither our model nor analysis assumes continuity of the p.d.f. at the
boundary of the support, and hence our results still apply. We also chose a
power-law distribution F2(v) = vα, α > 0, in the comparison and obtained
similar results; however, the actual expressions are too long (due to the inverse
effort function v2(b2)) to suit a neat presentation and hence omitted.
6
where b1(v) = 0 iff v1 = k−1(v), and k(v) is determined by
k(cid:48)(v) =
1(v)
2(k(v))
k(v)F (cid:48)
vF (cid:48)
with boundary condition k(¯v) = ¯v.
Proposition 4 (Equilibrium strategy in SYM). In a n-player
symmetric all-pay auction with incomplete information, if the
common prior is F (v), v ∈ [v, ¯v], and the payment function
p(b) satisfies p(0) = 0 and p(cid:48)(b) > 0, there exists a unique
Bayesian Nash (symmetric) equilibrium which is given by
(cid:90) v
(cid:17)
vF n−1(v) −
F n−1(t) dt
.
(14)
b(v) = p−1(cid:16)
B. Agent strategy, Contest prize, and Principal's profit
v
1) Agent strategy:
(cid:112)
(cid:112)
OPT: Using Theorem 1, we apply (8) with p(bi, vi) = b2
and F1 = v1 to obtain
2b1
v1
(1 − v1),
− 2b1
v2
1
1
λ
=
which gives the optimal equilibrium strategy for agent 1:
i /vi
b1(v1) =
v2
1
2λ
,
v1(b1) =
2λb1.
(15)
Similarly, applying (8) with F2 = v2+1
yields for agent 2:
2
b2(v2) =
v2
2
2λ
,
v2(b2) =
2λb2
(16)
f (v)
which is the same (i.e., symmetric) as agent 1. This is because
the two type distributions happen to have identical hazard rate
[12],
1−F (v), which is used (inversely) in (8). Clearly, this
should not be generalized to all c.d.f.'s; indeed, we shall see
later that the optimal prizes for the two agents (21)(22) as well
as their individual contributions to the principal's profit (23)(24)
are different.
FIX: Instantiating Proposition 3 with F1(v) = v and F2(v) =
2 yields
v+1
k(cid:48)(v) =
2k(v)
v
⇒ k(v) = v2,
(cid:16)(cid:90) v1
v2 dv
(cid:17) 1
2
k−1(v) =
=
v3/2
1√
3
,
Therefore
bf ix
1
(v1) =
bf ix
2
(v2) =
0
v3/4
2√
3
.
√
v.
(17)
(18)
SYM: For SYM-1, applying Proposition 4 with F (v) = v
gives
bsym
1
(v) =
v√
2
.
(19)
For SYM-2, applying Proposition 4 with F (v) = v+1
2 gives
Theorem 1 gives the optimal prize for agent 1 via (7):
(v) =
2) Optimal prize tuple in OPT:
bsym
2
4λ2 +(cid:82) v1
v3
1
0
v1+1
2
(cid:12)(cid:12)(cid:12)v1=
v2
4λ2 dv1
1
Z1(b1) =
v
2
.
(20)
√
√
2(
3λ2(
2λb1)3
2λb1 + 1)
.
√
=
2λb1
(21)
The optimal prize for agent 2 is similarly obtained to be
Z2(b2) =
2
3λ
b2.
(22)
3) Principal's profit:
OPT: The principal's maximized profit can again be calculated
using Theorem 1. Particularly in (9), we calculate each agent's
contribution corresponding to each summation term of π = π1+
π2, as follows:
(cid:90) 1
(cid:90) 1
0
(cid:2) v2
(cid:2) v2
2λ
π1 =
π2 =
∴ π = π1 + π2 =
2λ
0+
− v3
4λ
− v3
4λ
1
8λ
− v2
4λ
− v2
4λ
.
(1 − v)(cid:3) dv =
(1 − v)(cid:3) dv
=
2
1
12λ
1
24λ
,
.
(23)
(24)
(25)
In the above when calculating π2, although there is a probability
atom of 0.5 at v = 0, the effort and payment are both zero, and
hence it does not contribute to the profit and we can take the
integral from 0+ onward.
2 − λ where
FIX: The profit in this case is πf ix = πf ix
1 + πf ix
πf ix
1 =
πf ix
2 =
∴ πf ix =
bf ix
1
(v1) dF1(v1) =
bf ix
2
0+
√
24
(v2) dF2(v2) =
− λ.
35
3
√
2
5
3
√
2
7
3
,
.
(26)
(cid:90) 1
(cid:90) 1
0
Like in OPT, bf ix
(0) = 0 nullifies the atom at f2(0).
SYM: The profits of SYM-1 and SYM-2 are, respectively,
bsym
1
(v1) dF1(v1) − λ =
bsym
2
(v2) dF2(v2) − λ =
− λ,
1√
2
− λ.
1
4
(27)
(28)
2
(cid:90) 1
(cid:90) 1
0
0
πsym
1 = 2
πsym
2 = 2
C. Results
√
In line with the organizer's ultimate objective, we first com-
pare the profit of the above four mechanisms in Fig. 1, based
on formulae (25) -- (28). The plot clearly shows that OPT garners
the highest profit compared to all the other mechanisms over
all possible λ. In particular, even though SYM-1 is privileged
to benefit from two strong agents, it is still outperformed by
OPT, apart from being tangent
to OPT at only one point
2/4). Specifically, eight profit values are also marked in
(λ =
Fig. 1 at λ = 0.1 and 0.3, where we see that OPT significantly
outperforms the other three mechanisms, by about 105%, 315%
and 730%, respectively (at λ = 0.1). If λ is sufficiently high,
FIX, SYM-1, and SYM-2 even run into deficit (negative profit),
at λ > 0.396, λ > 0.707, and λ > 0.25, respectively. On the
other hand, as λ becomes smaller (i.e., the principal values the
prize less), OPT reaps exponential profit growth whereas the
other mechanisms only have linear profit increase.
1) Profit ranking and rationale: According to Fig. 1, the
profit of the four mechanisms can be ranked as SYM-2 ≺ FIX ≺
SYM-1 ≺ OPT, where ≺ denotes "is inferior to". To understand
the rationale behind this ranking result, we examine the agent
strategies, by plotting formulae (15) -- (20) in Fig. 2.
In general, the ranking SYM-2 ≺ FIX ≺ SYM-1 can be
understood by the composition of the three contests: SYM-1
and SYM-2 are composed of two strong and two weak agents,
respectively, and FIX is a mixture. However, it is worth noting
that FIX is much lower than the average of SYM-1 and SYM-2;
in fact, it is even lower than half of SYM-1 alone. This is due to
7
Figure 1: Profit comparison of different mechanisms.
the effort reservation effect existing in asymmetric auctions, as
mentioned in Section IV-C, where a stronger agent shades his
bid when facing a weaker agent. Indeed, Fig. 2 shows that agent
1 in FIX bids significantly lower than in SYM-1, although agent
2 exerts higher effort than in SYM-2.6 The reason why the effort
reduction of the stronger agent outweighs the effort increase
of the weaker agent is that, mathematically, the p.d.f. of the
stronger type concentrates on the higher region of the common
support [v, ¯v] and thus has a larger impact on the revenue
(calculated by an integral). Intuitively, this tells that "stronger
agents matter more", and offers us the following insight: it is
more productive for a mechanism to focus on incentivizing
stronger agents who constitute the main contributors to the
revenue. This hints toward a discriminatory contest design, and
indeed, this discriminatory design principle is used by both this
work (the agent-specific prize functions) and some prior work
such as [35].
To understand why SYM-1 ≺ OPT, first we see in Fig. 2
that OPT incentivizes agents to exert significantly higher effort
than all the other mechanisms, particularly at higher types,
which concurs the productivity of incentivizing stronger agents
mentioned above.
To draw deeper insights into how this is achieved, we examine
the optimal prize tuple of OPT, by plotting formulae (21)(22)
in Fig. 3. We see that OPT gives slightly higher reward to
agent 2 if he exerts the same amount of effort as agent 1. The
rationale is to motivate the weaker agent insofar as he becomes
a competitive rival to the stronger agent, thereby "threatening"
the stronger agent not to reserve effort. In principle, the prize
tuple endogenizes agent asymmetry and enables the contest
to recuperate from the fierceness of competition existing in
symmetric contests. Moreover, the increasing monotonicity of
6The reason why agent 2 works harder in FIX than in SYM-2 is because he
can deduce that the stronger agent will reserve effort and hence he (agent 2)
sees a better chance to win by striving above his (usual) effort level as in the
symmetric case (SYM-2).
00.10.20.30.40.500.20.40.60.811.21.41.61.82λ: Principal's Valuation of Prizeπ: Profit OPTFIXSYM−1SYM−20.6250.510.200.050.150.300.611.25 8
(a) λ = 0.1.
(b) λ = 0.3.
(c) λ = 0.5.
Figure 2: Equilibrium strategy (agent effort). The same line-spec is used for both agents in OPT as they adopt the same strategy.
(a) λ = 0.1; the maximum effort is 5.
(b) λ = 0.3; the maximum effort is 5
3 .
(c) λ = 0.5; the maximum effort is 1.
Figure 3: Optimal prizes as functions of winner efforts in OPT. The range of X axis is determined by the maximum effort.
the prize functions also motivates agents to work harder as the
reward grows with their effort. These incentives render OPT
superior to SYM-1 and the other mechanisms.
Lastly, as a side note, Fig. 2 also indicates that the agent
strategy in all the mechanisms is monotone increasing in agent
type, which conforms to Lemma 1.
2) Neutralizing DMR: In this subsection, we investigate how
SA neutralizes the law of DMR, by comparing OPT-n and FIX-
n. In OPT-n, since the n agents are now homogeneous (same
as agent 1, for illustration), the prize tuple collapses into a
single prize function. Using Theorem 1, this single, optimal
prize function can be calculated as
Z opt-n(b) =
(2λb)2− n
2
3λ2
.
In addition, the equilibrium agent strategy becomes
and the resultant profit is
,
b =
v2
2λ
πopt-n =
n
12λ
.
(29)
(30)
For FIX-n, the equilibrium agent strategy is calculated using
Proposition 4, as
(cid:114) n − 1
n
bf ix-n(v) =
n
2 ,
v
(31)
and the resultant profit is
πf ix-n = n
bf ix-n(v) dF1(v) − λ =
(cid:90) 1
0
2(cid:112)n(n − 1)
n + 2
− λ (32)
We plot πopt-n and πf ix-n in Fig. 4 with respect to n for
different λ values. As we can see, as a standard auction, FIX-n
is indeed governed by the law of DMR, and exhibits concave
profit growth as n increases. Eventually, it saturates at the upper
bound of limn→∞ πf ix-n = 2 − λ, which is indicated in Fig. 4
as well.
n(cid:82) 1
λ(cid:82) 1
On the contrary, OPT-n is not confined by the law of DMR
and its profit grows linearly as n increases. To understand why,
we note that in the symmetric case, SA is reinterpreted as that
the agent strategy is independent of the number of agents. This
is also evidenced by (29), whereas the strategy in FIX-n (31)
does depend on n. Therefore the revenue -- the sum of all the
agents' bids -- is a linear function of n (in detail, revenue is
6λ). The cost is also linear in n: λE[Z] =
12λ. Therefore, the profit is a linear function
6λ − n
12λ which coincides with (30).
Moreover, when n is not too small, as in most real scenarios,
Fig. 4 shows that OPT-n garners much larger profit which
is even superior to the upper bound of FIX-n by far. These
observations manifest a very healthy scalability for OPT-based
crowdsourcing systems.
v2
2λ dF (v) = n
v4−n
3λ2 dvn = n
of n, verified by n
12λ = n
0
0
00.20.40.60.8100.511.522.533.544.55Agent TypeEquilibrium Strategyλ=0.1 OPT: Agent 1OPT: Agent 2FIX: Agent 1FIX: Agent 2SYM−1SYM−200.20.40.60.8100.20.40.60.811.21.41.61.8Agent TypeEquilibrium Strategyλ=0.3 OPT: Agent 1OPT: Agent 2FIX: Agent 1FIX: Agent 2SYM−1SYM−200.20.40.60.8100.10.20.30.40.50.60.70.80.91Agent TypeEquilibrium Strategyλ=0.5 OPT: Agent 1OPT: Agent 2FIX: Agent 1FIX: Agent 2SYM−1SYM−201234505101520253035Winner EffortOptimal Prizeλ=0.1 Agent 1Agent 200.511.5200.511.522.533.54Winner EffortOptimal Prizeλ=0.3 Agent 1Agent 200.20.40.60.8100.20.40.60.811.21.4Winner EffortOptimal Prizeλ=0.5 Agent 1Agent 2 9
(a) λ = 0.1.
(b) λ = 0.3.
(c) λ = 0.5.
Figure 4: Strategy autonomy neutralizes the law of diminishing marginal returns.
VI. CONCLUSION
This paper frames the problem of incentive mechanism design
for crowdsourcing into an all-pay contest model. To the best
of our knowledge, our model represents the first contest or
auction model that (a) accommodates multiple heterogeneous
users with incomplete information, and (b) is instrumented with
a prize function tuple as opposed to the conventional, single,
fixed prize. Not only does it more closely characterizes realistic
crowdsourcing campaigns, but it also induces the highest pos-
sible total effort from self-interested agents, which is of utmost
importance to most crowdsourcing campaigns.
The strategy autonomy (SA) property, which is discovered
during the course of investigating this new model, captures a
counter-intuitive and surprising phenomenon: agents in a het-
erogeneous environment behave independently of one another
as if they were in a homogeneous one. SA bears practical
significances pertaining to system complexity, profitability, and
scalability. It could also be an enrichment to the theory of
mechanism design.
REFERENCES
[1] D. Estrin, "Participatory sensing: Applications and architecture," in ACM
MobiSys, 2010, pp. 3 -- 4.
[2] D. M. Kreps, "Intrinsic motivation and extrinsic incentives," American
Economic Review, vol. 87, no. 2, pp. 359 -- 364, 1997.
[3] V. Krishna, Auction theory, 2nd ed. Academic Press, 2009.
[4] J.-S. Lee and B. Hoh, "Sell your experiences: A market mechanism based
incentive for participatory sensing," in IEEE PerCom, 2010.
[5] D. Yang, G. Xue, X. Fang, and J. Tang, "Crowdsourcing to smartphones:
Incentive mechanism design for mobile phone sensing," in ACM Mobi-
Com, 2012.
[6] I. Koutsopoulos, "Optimal incentive-driven design of participatory sensing
systems," in IEEE INFOCOM, 2013.
[7] T. Luo, H.-P. Tan, and L. Xia, "Profit-maximizing incentive for participa-
tory sensing," in IEEE INFOCOM, 2014, pp. 127 -- 135.
[8] R. Siegel, "All-pay contests," Econometrica, vol. 77, no. 1, pp. 71 -- 92,
2009.
[9] S. Chawla, J. D. Hartline, and B. Sivan, "Optimal crowdsourcing contests,"
in ACM-SIAM Symposium on Discrete Algorithms, 2012.
[10] D. DiPalantino and M. Vojnovic, "Crowdsourcing and all-pay auctions,"
in ACM EC, 2009, pp. 119 -- 128.
[11] N. Archak and A. Sundararajan, "Optimal design of crowdsourcing
contests," in 30th International Conference on Information Systems, 2009.
[12] R. Myerson, "Optimal auction design," Mathematics of Operations Re-
search, vol. 6, no. 1, pp. 58 -- 73, 1981.
[13] J. Riley and W. Samuelson, "Optimal auctions," American Economic
Review, vol. 71, no. 3, pp. 381 -- 392, 1981.
[14] E. S. Maskin and J. G. Riley, "Asymmetric auctions," Review of Economic
Studies, vol. 67, no. 3, pp. 413 -- 438, 2000.
[15] E. Amann and W. Leininger, "Asymmetric all-pay auctions with incom-
plete information: The two-player case," Games and Economic Behavior,
vol. 14, no. 1, pp. 1 -- 18, May 1996.
[16] N. Szech, "Asymmetric all-pay auctions with two types," Working Paper,
University of Bonn, January 2011.
[17] J. Xiao, "Asymmetric all-pay contests with heterogeneous prizes," Work-
ing Paper, University of Melbourne, September 2013.
[18] G. Fibich and A. Gavious, "Asymmetric first-price auctions - a perturba-
tion approach," Math. Oper. Res., vol. 28, no. 4, pp. 836 -- 852, 2003.
[19] G. Fibich and N. Gavish, "Numerical simulations of asymmetric first-price
auctions," Games and Economic Behavior, vol. 73, pp. 479 -- 495, 2011.
[20] B. Moldovanu and A. Sela, "The optimal allocation of prizes in contests,"
American Economic Review, vol. 91, no. 3, pp. 542 -- 558, 2001.
[21] C. Cohen, T. Kaplan, and A. Sela, "Optimal rewards in contests," The
RAND Journal of Economics, vol. 39, no. 2, pp. 434 -- 451, 2008.
[22] T. Kaplan, I. Luski, A. Sela, and D. Wettstein, "All-pay auctions with
variable rewards," Journal of Industrial Economics, vol. 50, no. 4, pp.
417 -- 430, 2002.
[23] G. Fibich and N. Gavish, "Asymmetric first-price auctions - a dynamical-
systems approach," Mathematics of Operations Research, vol. 37, no. 2,
pp. 219 -- 243, 2012.
[24] E. S. Maskin and J. G. Riley, "Equilibrium in sealed high bid auctions,"
Review of Economic Studies, vol. 67, no. 3, pp. 439 -- 454, 2000.
[25] -- -- , "Uniqueness of equilibrium in sealed high-bid auctions," Games
and Economic Behavior, vol. 45, no. 2, pp. 395 -- 409, 2003.
[26] S. O. Parreiras and A. Rubinchik, "Contests with three or more hetero-
geneous agents," Games and Economic Behavior, vol. 68, pp. 703 -- 715,
2010.
[27] J. Franke, C. Kanzow, W. Leininger, and A. Schwartz, "Effort maxi-
mization in asymmetric contest games with heterogeneous contestants,"
Economic Theory, vol. 52, no. 2, pp. 589 -- 630, 2013.
[28] T. Luo, S. S. Kanhere, and H.-P. Tan, "SEW-ing a simple endorsement
web to incentivize trustworthy participatory sensing," in IEEE SECON,
2014, pp. 636-644.
[29] C.-K. Tham and T. Luo, "Quality of contributed service and market
equilibrium for participatory sensing," in IEEE DCOSS, 2013, pp. 133 --
140.
[30] Appendix. [Online]. Available: https://tonylt.github.io/pub/mass14app.pdf
[31] P. Milgrom and I. Segal, "Envelope theorems for arbitrary choice sets,"
Econometrica, vol. 70, no. 2, pp. 583 -- 601, 2002.
[32] K. A. Konrad, Strategy and Dynamics in Contests. Oxford University
Press, 2009.
[33] E. Cantillon, "The effect of bidders asymmetries on expected revenue in
auctions," Games and Economic Behavior, vol. 62, pp. 1 -- 25, 2008.
[34] P. A. Samuelson and W. D. Nordhaus, Microeconomics, 17th ed.
McGraw-Hill, 2001.
[35] F. Naegelen and M. Mougeot, "Discriminatory public procurement policy
and cost reduction incentives," Journal of Public Economics, vol. 67, no. 3,
pp. 349 -- 367, 1998.
24681012141618051015Number of AgentsProfitλ=0.1 OPT−nFIX−nUpper bound: 1.92468101214161800.511.522.533.544.55Number of AgentsProfitλ=0.3 OPT−nFIX−nUpper bound: 1.72468101214161800.511.522.53Number of AgentsProfitλ=0.5 OPT−nFIX−nUpper bound: 1.5 |
1910.08942 | 1 | 1910 | 2019-10-20T10:10:21 | Autonomous Industrial Management via Reinforcement Learning: Self-Learning Agents for Decision-Making -- A Review | [
"cs.MA",
"cs.AI"
] | Industry has always been in the pursuit of becoming more economically efficient and the current focus has been to reduce human labour using modern technologies. Even with cutting edge technologies, which range from packaging robots to AI for fault detection, there is still some ambiguity on the aims of some new systems, namely, whether they are automated or autonomous. In this paper we indicate the distinctions between automated and autonomous system as well as review the current literature and identify the core challenges for creating learning mechanisms of autonomous agents. We discuss using different types of extended realities, such as digital twins, to train reinforcement learning agents to learn specific tasks through generalization. Once generalization is achieved, we discuss how these can be used to develop self-learning agents. We then introduce self-play scenarios and how they can be used to teach self-learning agents through a supportive environment which focuses on how the agents can adapt to different real-world environments. | cs.MA | cs |
AUTONOMOUS INDUSTRIAL MANAGEMENT VIA
REINFORCEMENT LEARNING: SELF-LEARNING AGENTS FOR
DECISION-MAKING -- A REVIEW
A PREPRINT
Leonardo A. Espinosa Leal∗
Magnus Westerlund
Department of Business Management and Analytics
Department of Business Management and Analytics
Arcada University of Applied Sciences
Helsinki, Finland
Arcada University of Applied Sciences
Helsinki, Finland
Anthony Chapman
Computing Department
University of Aberdeen
Aberdeen, UK.
October 22, 2019
ABSTRACT
Industry has always been in the pursuit of becoming more economically efficient and the current focus
has been to reduce human labour using modern technologies. Even with cutting edge technologies,
which range from packaging robots to AI for fault detection, there is still some ambiguity on the aims
of some new systems, namely, whether they are automated or autonomous. In this paper we indicate
the distinctions between automated and autonomous system as well as review the current literature
and identify the core challenges for creating learning mechanisms of autonomous agents. We discuss
using different types of extended realities, such as digital twins, to train reinforcement learning agents
to learn specific tasks through generalization. Once generalization is achieved, we discuss how these
can be used to develop self-learning agents. We then introduce self-play scenarios and how they can
be used to teach self-learning agents through a supportive environment which focuses on how the
agents can adapt to different real-world environments.
Keywords Autonomous systems · reinforcement learning · self-play · digital twin · industry4.0
1
Introduction
Recent developments in industry, such as Industry4.0, have emphasized the need for a comprehensive automation of
operational processes [1]. Similar developments in both the automobile [2] and maritime industry [3] has highlighted
the need for vehicle/vessel driving automation. In recent media, we often see the term autonomous used as a synonym
for automated. Hence, automated driving often become autonomous driving, without a deeper reflection on what an
autonomous system would entail. Within information systems research we also have a long history of researching
autonomous entities, sometimes as an abstract futuristic entity that is often ill-defined in terms of true autonomy,
although some efforts have been made [4]. We consider an autonomous entity as an evolutionary leap compared to an
automated entity. Meanwhile, the automation component is far from solved for many physical environments. On the
other hand, for digital environments such as stock exchanges and advertisement auctions, the process has, to a large
extent, already been automated, but autonomous behaviour is rarely needed or desired in such environments. To address
this gap we delimit our focus in this paper, to consider one part of an autonomous system, the learning mechanisms
∗Corresponding author: [email protected]
A PREPRINT - OCTOBER 22, 2019
of an autonomous entity. To achieve this, we first expand on the subject of what an autonomous system entails and
relevant areas that may assist in achieving autonomy.
The need for automation has led to a realization that we also need to digitize our industrial environments and processes.
These processes include manufacturing, warehousing, building management, or hospital management. Digitization
may happen by introducing IoT sensors, communication networks, and cameras into an environment and then digitally
capture the physical reality. Advanced types of automation will often form a model using past events and test the
model using more recent events in a controlled environment before testing the model in real scenarios. A common
approach has been to use fuzzy logic to describe expert behaviour for an entity or an agent that operates in or oversees
the environment [5]. Prediction is sometimes required in automation and neural networks are a popular way to compare
on the outcomes of the predictions [6].
In order to verify a number of algoritmically derived scenarios in the physical factory may often be expensive, dangerous,
and disruptive. Instead a digital representation of the process/environment, a so called extended reality, can be used for
verification and optimization, before ideas/changes are moved into the physical realm. An example of an extended
reality is the digital twin technology [7], which can be described in many formats, e.g. process graphs, a 2D flat model,
a 3D space, or a 4D space-time environment (can be seen as stacked 3D models with temporal dependencies). Although,
the more realistic the environment and simulation is, the better the final result can be expected. Complexity tend to
increase for each added dimension and different strategies must be devised to conquer the different environments.
A state-of-the-art form of digital twin technology is physics-based (follows physical laws) and allows the agent to
interact with the environment, i.e. bi-directional data-exchange [8]. Furthermore, the environment may be also be
modified within certain parameters to allow for an improved process. Hence, both the agent and the extended reality
environment are part of the modelling process and may be enhanced in order to improve an outcome. In this article,
we focus on a method and experimental study for self-playing agents which can use deep reinforcement learning, in
what can best be generalized as an extended reality environment [9]. Our aim is to address the topic of designing
self-play scenarios for self-learning agents that combine various sources for discovering and generating training data.
By reviewing the literature we noticed a gap which could improve existing architectures used for creating self-learning
agents that can be used for enabling autonomous industrial management. [8] also highlight the need for more concrete
digital twin case-studies. Combining AI research and digital twin technology offer an interesting avenue for further
research.
The structure of the paper is the following. Section 2 reflect on three core terms used throughout the paper. The first
subsection provides a discussion on autonomous vs. automated systems, and notes that although media and information
systems literature provide a considerable history of autonomous systems, real-world examples of such systems do
not yet exist in a strict language or legal sense. Following this, eXtended Reality (XR) is introduced as a self-play
environment for teaching models through interactivity. The last sub-section discusses the need for data augmentation to
bridge the gap between training in extended reality and inference in the physical reality. The third section introduces
reinforcement learning and recent advancements. The fourth section presents how algorithms can learn by self-play, i.e.
learning from the interaction in XR, and a conceptual architecture. The fifth section contributes a design and proposes
applications in some relevant areas. The final section concludes our findings and presents future directions for research.
2 Research methods and concepts
In the first part of this paper, a literature review of the relevant concepts is presented. Through the literature review
we provide a foundational understanding of root definitions for the second part, where we provide an analysis of our
findings. In this latter part, we follow a soft systems methodology (SSM) by framing a problem formulation (model
learning) and an action plan (conceptual model for learning) aimed at future research [10]. [11] state that SSM consist
of an analysis of the current status of the system. This includes inherent problems and activities, and a definition of the
system that derives the actual goal of the targeted system ("root definition") in order to propose a conceptual system
model.
2.1 Requirements for achieving autonomous systems
Autonomous systems have long been a promise of both academia and industry alike. One of the early academic papers
discussing autonomous systems was presented in the 1950s, where the author discusses an autonomous nonlinear
oscillation system [12]. Much later, in the 1990's with the introduction of the personal computer and precursors to the
Internet, autonomous agents became a topic of interest. The agent implementations tended to cover relatively narrow
and software localized areas. [13] defines the agent as a persistent software entity dedicated to a specific purpose. [4]
reviews the early definitions of autonomous agents, and then summarizes that the autonomous property can be defined
2
A PREPRINT - OCTOBER 22, 2019
Figure 1: Agent properties adapted from [4].
as something which exercises control over its own actions. Given the progress since, we would argue these definitions
do not qualify an entity for being autonomous, but mere automated.
The definition of the word autonomous, in a language sense, can be describe as an autonomous entity is independent
and has the freedom to govern itself (adapted from the Cambridge Online Dictionary). Using a strict interpretation, it
suggests that truly autonomous systems have yet to come into existence, and that traditional methods have not been
able to deliver beyond the automation component. Automation can here be defined as an entity made to operate by
the use of machines or computers in order to reduce the work done by humans (adapted from the Cambridge Online
Dictionary). This latter definition of automated systems corresponds better to the original definition of autonomous
agents, as presented above. An autonomous system should have the freedom to govern itself, and governance implies
some form of understanding of real-world repercussions of ones actions, for example the limits the legal framework
provide us with. Additionally, an autonomous entity should show independence, suggests that a mere exploitation of
the historical world is not sufficient, but rather that the entity must learn to explore unknown worlds through self-play.
Hence, an entity that employs an intelligence derived from a rule-set or a neural network trained on historical data, will
likely fail the test of true autonomy.
A similar standpoint has been taken by organizations regulating and standardizing the automotive industry [14]. The
Society of Automotive Engineers (SAE) through the On-Road Automated Vehicle Standards Committee is responsible
for defining levels of driving automation for cars and provides a classification of levels from 0-5.
In the SAE
classification, the highest level, "Level 5 (steering wheel optional)" is defined as a fully automated sysem, and not as an
autonomous system [15].
There have been several methods proposed for creating agents, e.g. fuzzy logic or neural networks. In [5] the authors
use the concept of intelligent agents for describing an important property of the entity. Intelligence or rather reasoning,
allows an agent to react to different defined circumstances or even unforeseen circumstances. Some early considerations
realized the challenge of defining a globally validated view of a contained entity, with so called artificial intelligence,
and instead highlighted that at the essence of an intelligent entity is an ability to perform subjective probability analysis
over both logical and natural (physical) conditions, [16] and [5] approach this through an expert system governed by
fuzzy logic and highlight certain abilities an agent must achieve to be considered intelligent. They highlight certain
abilities an agent must achieve to be considered intelligent. For example, an agent must be able to plan a set of actions
and needs to adapt its plan to changing conditions. [4] provide a classification of properties an agent may have (see
Fig. 1).
On a more philosophical note, for an autonomous entity to achieve independence and freedom to govern itself, one can
argue only an AI with consciousness can achieve such a feat. However, [17] highlights that consciousness, although not
properly understood, may indeed be a representation rather than an agency. Hence, the consciousness created in our
minds are a result of the brain attempting to make sense of sensory inputs, and consciousness is but an illusion, a model
of the contents of our attention. Therefore, the state of dreaming and the state of awakeness is a change in sensory
details and not a termination/commencement of conscious experience. Assuming that the cortical conductor theory
interpretation is correct [18], we can reduce an autonomous system into four main concepts that feeds an attention layer,
3
Agentautonomouslearninggoal-orientedtemporally continuouscommuni-cativereactivemobileflexiblecharacter A PREPRINT - OCTOBER 22, 2019
learning, reasoning, control, and selection. In this paper we delimit our focus to the learning mechanism that occur
from an interaction and describe some industrial environments were learning can occur. Our position is that without an
interactive learning mechanism and an ability to self-improve, such that reinforcement learning provide us with, the
system cannot be considered autonomous. Still, as discussed above there are other properties that must also be fulfilled
for a system to be considered truly autonomous.
2.2 Extended reality
The term eXtended Reality (XR) has become known as a common term for fields were digitally enhanced environments
and human-interaction are combined for various purposes. This includes Virtual Reality (VR), Mixed Reality (MR), and
Augmented Reality (AR). Compared to, for example, autonomous driving, these environments offer AI researchers an
important and relatively low-cost setting for implementing human-interacting algorithms, such algorithms may contain
reasoning that can be considered "true" AI. Perhaps most importantly, these digital environments allow us to study an
agent's decision making, and thereby, offering a feed-back loop between the agent-environment-user.
Still, the need for training new abilities often require that agents are presented with big data. The traditional approach
was often to gather the data from users, e.g. playing a game, and then train on this data. Today, the deep networks'
need for massive data and relatively complex training scenarios for reinforcement learning presents researchers with
a problem that is often better solved by augmenting additional data for training networks, than using real data. Data
augmentation methods depend on the problem at hand, an agent may for example have to learn how to deal with object
recognition, spatial actions to take in relation to detected objects, or temporal differences in scenarios, to name a few.
The following sub-section reviews some of the relevant literature regarding the performance of data augmentation for
training agents.
2.3 Data augmentation
Data augmentation has gained prominence as a studied method for extending available datasets [19]. The ability to
train deep networks often depends on the availability of big data. The fields of both image recognition and voice
recognition has been strongly influenced by deep learning methods [20, 21], and this has motivated the emergence of
data augmentation. For RL many of the same concepts can be utilized, but there is also a need for methods that work
particularly in the temporal dimension.
Traditional, naïve approaches tend to manipulate the investigated environment or dataset in various ways. For visual
tasks, these have included scaling of objects, translating i.e. moving objects spatially to various positions, rotation
of objects at various angles, flipping objects as to remove bias from any direction, adding noise, changing lightning
conditions, and transforming perspective of a known object by changing the angle of view [19, 22].
For audio tasks, data augmentation often includes deformations into a temporal dimension. Approaches include time
stretching by changing audio speed, pitch shifting by raising or lowering the tone frequency by various degrees, dynamic
range compression, and introducing background noise using gaussian or natural noise methods [23].
The naïve data augmentation approaches tend to produce limited alternative data for RL agents to learn from in an
extended reality setting. For an RL agent to learn new abilities, data augmentation must support the agent's scenario
learning process. As suggested by [24], we shift from learning to generalize on spatial data to reacting to continuous-time
dynamical systems without a-priori discretization of time, state, and action.
Several approaches exist for the creation of these scenarios. An important method is adversarial learning, as it can
produce new and complex augmented datasets by pitting a generative model against an adversary [25]. A generative
model in combination with XR, can also address the exploration problem, as exploring some states in the physical
reality could be very costly and dangerous. This combination also allows the system developer to understand which
state spaces in the virtual environment has been visited and trained upon, and the model's ability to generalize in the
extended reality environment. These generative techniques for extending the learning environment are further explored
in the following section.
3 Reinforcement learning
Although XR is slowly receiving more recognition, AI and machine learning's use of XR to enhance learning is lagging
behind. XR could help improve an AI's behaviour by providing information from either pure virtual or semi-real
environments. Reinforcement learning is one machine learning technique which, when combined with XR, could
produce interesting and beneficial results for many applications, such as driverless cars, autonomous factories, smart
cities, gaming and more.
4
A PREPRINT - OCTOBER 22, 2019
RL's primary purpose is to calculate the best action an agent should take when an environment is provided. With RL,
we could be able to calculate what best action to take by maximizing the cumulative reward from previous actions, thus
learning a policy. RL has long roots from areas such as dynamical programming [26], and for a historical review see
[27]. However, recently it has received a lot of attention for its potential to advance and improve applications in gaming,
robotics, natural language processing (including dialogue systems, machine translation, and text generation), computer
vision, neural architecture design, business management, finance, healthcare, intelligent transportation systems, and
other computer systems [28].
Attention and memory are two parts from RL which, if done impetuously, could negatively affect performance. Attention
is the mechanism which focuses on the salient parts. Whereas, memory provides long term data storage, and attention is
an approach for memory addressing [28]. Using XR and self-play, agents may be able to learn desired behaviour before
an action an agent makes become crucial to their performance. As an example, autonomous helicopter software could
learn fundamental mechanisms for flight using virtual data in simulations in order to achieve high level of attention
using the memory required, without the risks posed by real world applications. Once the attention has reached a desired
level, it can be applied to real agents in the physical world.
General value functions can be used to represent knowledge. RL, arguably, mimics knowledge in the sense that it
(generally) learns from the results of actions taken. Thus, one may be able to represent knowledge with general value
functions using policies, termination functions, reward functions, and terminal reward functions as parameters [29].
Doing so, an agent may be able to predict the values of sensors, and policies to maximize those sensor values, and
answer predictive or goal-oriented questions.
Generative Adversarial Networks (GANs) estimate generative models via an adversarial process by training two models
simultaneously [25], a generative model G to capture the data distribution, and a discriminative model D to estimate
the probability that a sample comes from the training data but not the generative model G. Such an approach could be
extended to XR by training a generative model G on virtual / simulated test data and then a discriminative model D to
estimate the probability that a sample comes from the real world. This could help tackle some of the issues with RL
within virtual environments and extended to the real world. RL and XR could be used before the agent is applied to a
real environment, this could save on resources and make autonomous systems a more viable option for general use.
GANs together with transfer learning could advance self-play using virtual environments for real world agents [30]. By
combining virtual data generative models and transferring the learning model to a discriminative model, we may be
able to accurately express what was learned from the virtual learning environment to the real agent. Again, unforeseen
problem will inevitably arise due to the nature of modelling. By using RL both in the virtual learning phase and
embedded into the real agent, we may drastically improve a real agent's learning time.
[31] proposed a strategic attentive writer (STRAW), a deep recurrent neural network architecture, for learning high-level
temporally abstracted macro-actions in an end-to-end manner based on observations from the environment. Macro-
actions are sequences of actions commonly occurring. STRAW builds a multi-step action plan, updated periodically
based on observing rewards, and learns for how long to commit to the plan by following it without re-planning. Similar
to GANs, STRAW could be used after the simulation learning stage so the agent copes with any discrepancies between
the simulation and the real world.
Adaptive learning is a core characteristic to achieving strong AI [32]. Several adaptive learning methods have been
proposed which utilize prior knowledge [33, 34, 28]. [33] proposed to represent a particular optimization algorithm as a
policy, and convergence rate as reward. [34, 28] proposed to learn a flexible recurrent neural network (RNN) model to
handle a family of RL tasks, to improve sample efficiency, learn new tasks in a few samples, and benefit from prior
knowledge.
The notion of self-play is one of the biggest advancements of modern AI. AlphaGo AI is Deepmind's newest Go playing
AI [35], that learns, tabula rasa, superhuman proficiency in challenging domains. Starting with the basic rules, they
used self-play for the AI to learn strategies by playing against itself and storing efficient / rewarding moves. Fictitious
Self-Play is a machine learning framework that implements fictitious play in a sample-based fashion [36]. The three
strategies that are compared are: Learning by self-play, learning from playing against a fixed opponent, and learning
from playing against a fixed opponent while learning from the opponent's moves as well [37].
4 Self-play Scenarios and Architectures
Generalizing from training into real scenarios is not easy for self learning agents. This problem is known as the reality
gap [38]. In the initial stages of AI research, the training of self-learning agents included rules or limited scenarios
where it can learn and improve upon competition against other introduced players. Interestingly, video games have
emerged as one of the main source of benchmark environments for the training and testing of such agents, mostly due
5
A PREPRINT - OCTOBER 22, 2019
Figure 2: Proposed general architecture for a self-learning agent interacting with its environment.
to its realistic, yet controlled approach to the real world, and the easy access to large amounts of data. For example, an
AI agent is trained by playing with a perfect copy of itself without any supervision [39]. In this scenario, a set of basic
rules of the game have been introduced at the beginning and the agent improves much faster using a vector of rewards
instead of the classical scalar quantity [40].
Initially, self-play agents were trained to play board games (such as chess and go, among others) [35] but it has now
been successfully extended from the classic and simpler Atari 2600 video games [41] to more complex first-person
shooters: Doom [42], Battlefield [43], Quake [44]; Role Playing games2: Warcraft [45]; Real-Time Strategic Games:
Starcraft [46, 47, 48] and more recently Multiplayer Online Battle Arena (Dota 2, [49]). For a more comprehensive
review see the work by [50].
Motivation for studying self-play scenarios has increased in recent years mainly due to the advances in neural network
architectures suitable to the reinforcement learning paradigm: DQN [41], AC3 [51], DSR [52], Dueling networks[53]
among others as well as the development of powerful and accessible graphics processing unit (GPU) computing
resources. This area of research is broadly known as Deep Reinforcement Learning [28, 54, 55].
The challenge of training self-playing agents in order to develop more complex policies inside realistic and highly
specific or general environments remain as an open problem. Most of the recent developments tend to focus on very
particular properties of the learning agent or the way that they interact with their surroundings. To address this issue, we
identify two general mechanism that can be improved in order to design a better self-learning agent: self-play scenarios
and self-learning architectures.
4.1
Improving self-play scenarios and self-learning agents: closing the reality gap
Constructing realistic self-play scenarios plays a fundamental role in training self-learning agents. Once an agent is
immersed in a specific environment, we expect (independently of the self-learning architecture) that it will learn a set of
policies accordingly to the received experiences3. A problem which is widely understood, is that when agents learn
from strict simulated scenarios, they may not be prepared for unexpected situations when the environment changes,
such as a pigeon flying towards the sensor of a driverless car. Here, we propose a general scheme that uses the versatility
of the video games or simulators as a source of synthetic data and the wide array of capabilities of modern extended
reality technologies, to enrich the properties of the real environment during the training of self-learning agents. An
agent may learn independently, but the environment can be controlled to persuade the agent to learn a set of additional
policies for unexpected scenarios. In addition to the enriched data, the self-learning agent may be trained using purely
synthetic data. But the limitations of this method rely in the accuracy of the representation of the real scenarios.
For the self-learning mechanism, we have identified three key steps which could improve the design of architectures
for self-learning agents, which may improve policies both in terms of effectiveness and robustness. The three areas
2Including the Massively Multiplayer Online type of games.
3A learning agent can learn a set of policies, or will optimize the parameters of a given set of policy. New policies can emerge
even without previous knowledge. The sum of the whole policies is called general policy.
6
A PREPRINT - OCTOBER 22, 2019
Figure 3: Illustration of how different components from Figure 2 can be used on complete self-learning systems. a.
Adapted from [56], b. Adapted from [57], c. Adapted from [58], d. Adapted from [59], e. Adapted from [60] and f.
Adapted from [61].
are: Data collection, task generalization and emergent behaviour (see Fig. 2). For a given agent, in the first stage the
agent will need to interact with the environment, possibly by accessing a data collection, then the agent should be
able to generalize a set of given tasks and, simultaneously, new skills should emerge (independently or due to the task
generalization). In the final stage, the emergent and the generalization skills interact with the environment to create a
continuous self-learning agent.
To illustrate the steps, we present a set of representative developments in the area of (deep) reinforcement learning.
Note that this list is not comprehensive but the examples are presented to highlight their own specific properties
then to introduce them in our general model (see Figure 3). Each can be used as a building block inside a complete
self-play scenario, for example, Figure 3a shows a low-fidelity rendered images with random camera positions, lighting
conditions, object positions, and non-realistic textures use to train a self-learning agent [56]; Figure 3b shows an agent
which uses a compressed representation of a real scenario to learn a set of policies which are successfully use back to
the real environment [57]. Figure 3c shows an image of a robot used as a one-demonstration example during the training
stage [58]. Figure 3d shows another image of a robot used during a training stage to teach a robot to place an object in
a target position [59], Figure 3e depicts an agent stacking two blocks, behaviour learn from sparse rewards [60] and
Figure 3f illustrates one competitive environment where one the agents develops a new set skills [61].
1. Data collection:
Domain randomization (DR): DeepMind have recently shown how an agent can be trained on artificially
generated scenarios. In their paper, the authors successfully transferred the knowledge from a neural network
purely trained on low resolution rendered RGB images: domain randomization [56]. This method can be
extensively used for training agents in the case that the amount of data available is low or when the separation
between the real and the train environment is immense.
World models (WM): Self-learning agents can be trained in a compressed spatial and temporal representation
of real environments by [57]. This method is highly powerful because an agent can learn in a more compact or
hallucinated universe and then go back to the original environment exporting the set of learned abilities. One
of the main advantages of this method is the possibility to perform a much faster and accurate in situ training
of the agents by using less demanding computational resources.
2. Task generalization:
One-shot imitation learning (OSIL): the authors present an improved method that uses a meta-learning
framework built upon the soft attention model [62] named one-shot imitation learning [58]. Here, the agents
are able to learn a new policy and solve a task after being exposed to a few demonstrations during the training
stage.
7
A PREPRINT - OCTOBER 22, 2019
One-shot visual imitation learning (OSVIL): Meta-imitation learning algorithm that teaches an agent to
learn how to learn efficiently [59]. In this work, a robot reuses past experiences and then upon a single
demonstrations, it develops new skills.
3. Emergent behaviour:
Scheduled auxiliary control (SAC): another research team from DeepMind introduced a new framework that
allows agents to learn new and complex behaviours in presence of a stream of sparse rewards [60].
Multi-agent competition (MAC): a paper by one research team from OpenAI, the authors showed that multi-
agents self-play on complex environments can produce behaviours which can be more complex than the
environment itself [61]. The emergent skills can improve the capabilities of the agents upon unexpected
changes in the real environment.
To summarize, Figure 2 illustrates how an agent can retrieves data from an environment and then generalizes to a
specific task and simultaneously develops new abilities. The new skills can emerge independently or due to the task
generalization process. In the final stage, the environment gets modified by the agent itself. A combination of such
methods could be used to create more effective architectures for teaching self-learning agents. The novelty of this idea
is that it builds upon the concept of modularity to create more complex architectures which develop new untrained
behaviours. This notion has been explored before to improve learning agents, for instance by synergistically combining
two modelling methods such as type-based reasoning and policy reconstruction [63]. Interestingly, it has been stated
recently that this is still a promising area under research [64]. In addition, the proposed architecture establishes a
general framework for the design of intelligent agents or rational agents searching to maximize their reward upon
interaction with an external environment [65] via specific goals (task generalization and emergent behaviour). Once an
architecture is defined, targeting the most adequate general policies are dependent on the external information gathered
by the agent. In the data collection module, the main goal is to optimize the data used by the agent for learning. In
the general case however, a proper design should include a non-static complete representation of the environment for
exploring extensively all possible scenarios.
In the next section, we present a complete general design for a self-learning agent including an extra module for
extending the description of the world experienced by the agent and then, upon an ad-hoc division and by using the
concept of multi-agent [66]. We discuss the application in the design of a central autonomous agent able provide via
rational decision-making, strategies to tackle the possible consequences derived from changes on any stage of a specific
industrial infrastructure.
5 Proposed Design
As already discussed, our goal of designing general self-play scenarios for teaching self-learning agents can be tackled
by separating the data retrieved from the environment and the agent's self-learning architecture. In the spirit of the
SMM (see Section 2) we present a general scheme in which we divide the general architecture into two modules, in
Module 1, the agent retrieves the data from its surroundings as a combination of information from the real world and
synthetic data (or pure synthetic data), and in Module 2 (equivalent to the structure of the Fig 3), the agent creates its
final policies. The general scheme is depicted in the Fig 4.
For a self-learning agent inside a specific self-learning scenario, there is no difference between synthetic or real
data. Here we call real data the information extracted from physical world without any previous or further digital
modifications. The agent uses exclusively the information, in terms of raw bytes, independently of the sensors that
connect it with the environment. The use of synthetic data arises as a need to expose the self-learning agent to
unexpected situations or conditions that allow it to create a set of optimal related policies. The representation of the
real world can be done also via pure synthetic data, however the best case scenario is such where the synthetic data
is used to extend the real world. The agent also can modify its own environment during the learning process, and if
the whole process is fully done in a simulation environment, the use of engines that mimic the physical rules of the
real world become necessary. The proposed general design can be encapsulated and used as a basic element in a more
complex multi-agent based learning architecture [66]. The communication among its moieties can be performed via the
local information modules that represents their individual real worlds. Posteriorly, a central agent retrieves and updates
its state providing a final decision. Several mathematical strategies can be applied to interchange the information, for
instance, by defining scalar, vectorial or probabilistic representations.
In the next part we aim to present an example or how we can build a minimal supra-modular architecture for the specific
case of industrial management. We discuss the advantages and the characteristics for the modular parts of the design, as
well as review (not exhaustively), the state-of-the-art developments for learning agents in decision making.
8
A PREPRINT - OCTOBER 22, 2019
Figure 4: Two module design of a general architecture for a self-learner agent interacting with its enriched or altered
environment.
Figure 5: Proposed architecture for a reinforcement learning agent in Autonomous Industrial Management (AIM) for
adaptive industrial infrastructure.
5.1 Application in Autonomous Industrial Management
The aforementioned improvements can be employed in specific applications using specific designs. In particular, by
using the new and open simulation frameworks such as OpenAI Gym4 or Dopamine5 among others. Here we discuss a
particular application for industrial environments solving a managerial problem, however the developed ideas can be
used for other purposes upon a clear definition of their parts.
In the industrial regime, decision-making is one of the key elements in the adaptive business intelligence discipline [67].
Despite the advances in the computational tools, still many relevant decisions are taken by real persons. Here, we
propose a general model where self-learning agents can be trained to make decisions on industrial scale upon self-
4https://github.com/openai/gym
5https://github.com/google/dopamine
9
A PREPRINT - OCTOBER 22, 2019
playing using extended reality training scenarios. We devised an strategy where an industrial infrastructure is divided
ad-hoc in three independent sections managed by three independent self-learning agents (see Figure 5). Each agent
replicates the physical process but it has the property to explore (and improve) in its own learning space based in the
perturbation via self-learning mechanisms. Upon external perturbations in their environments, the final decision is
done by a central agent which learns thanks to the information gathered for the independent agents. This Multi-Agent
Reinforcement Learning (MARL) structure will bring a modularity structure that can be exploited in the industrial realm
to improve and optimize the decision-making processes. Our example focuses mainly in the type of manufacturing
industry, but the methodology can be extended easily to other industrial fields. Similar conceptual frameworks have been
presented and described in the literature [68, 69] and in some cases including full computational frameworks or software
applied on specific industrial environments [70, 71] were described. The niche of industrial management is a fruitful
source of academic research with the application of the cutting-edge methods for the design of intelligent systems. Our
division in three different entities: Supply Chain Management (SCM), Warehouse or Inventory Management and Digital
Twin responds to the fact that a lot of work on modelling and creation of autonomous agents has been done in the last
years independently in each area, therefore a low level description of an autonomous agent for industrial management is
possible by following our proposed architecture (see Fig. 4). In the next subsection we present a non exhaustive review
of methods by highlighting some of the relevant strategies employed in the creation of autonomous systems.
5.1.1 Supply Chain Management (SCM)
The use of RL for optimization in the SCM has been discussed largely in the last decades [72, 73], in particular, in the
seminal works by [74] where the authors propose a multi-agent systems for the optimization of tasks in the SCM, [75]
that solves the multi-agent problem as a semi-Markov decision problem, [76] that reduces the complexity by dividing
the agents into three components (in a similar concept presented in this work), among others. Here, the self-learning
agent must be able to decide about the different suppliers (external and internal) in coordination with the warehouse
manager agent.
5.1.2 Warehouse or Inventory Management
Models for warehouse management have been developed in parallel to the state-of-the-art technological developments
to achieve short and optimized responses in delivering goods [77] meanwhile an optimal stock level is maintained. In
particular, it is worth to highlight two types of warehousing systems: automatic and automated warehousing systems.
Several methods for self-learning agents in warehouses have shown to be effective in solving real problems, for instance,
Stochastic Learning [78], temporal difference Actor-Critic algorithm [79] or more recently Deep Reinforcement
Learning [80]. In general, the information can be controlled and retrieved via an optimized wireless network of sensors
inside the warehouse [81]. In this case, the self-learning agent must be able not only to gather the internal information
about the location of the goods inside the warehouse but also optimize the architecture of the network via self-play
mechanisms.
5.1.3 Digital Twin
The use of digital tools in the industry has outgrown the level in which only individual applications were able to
be modelled accurately for very specific topics, in the sixties, towards a complete simulation of the systems during
the whole entire life cycle in modern times [82]. The concept that includes a detailed model and description of a
product, system or component in all its possible phases (product design, production system engineering, production
planning, production execution, production intelligence and closed-loop optimization) [83] is known as digital twin6
In general terms, the digital twin is an extension of the model-based systems engineering (MBSE) concept [85]. The
three important aspects of the digital twin have been discussed previously by other authors [82, 83, 8]: modularity,
connectivity and autonomy. These aspects can be linked with our proposed design in the Fig 4. Our design can be
extended into the MARL context.
6 Conclusion & Future Work
The design of self-learning and self-play scenarios is still an area of fruitful development and research. Many critics
have pointed out that AI research is limited by its own ideas [86] as well as it's usefulness for industrial purposes.
The creation and discussion of general architectures can open the door to new proposals, in particular, for the final
emergence of the expected autonomous systems (see Section 2.1). Despite a boom in the field during the last years, there
are many open questions about how to enable self-learning agents to achieve specific tasks without any supervision. We
6A term coined by NASA during the Apollo missions, but brought recently to the general public [84].
10
A PREPRINT - OCTOBER 22, 2019
propose to tackle this question by using our general architecture, which can be, in principle, limited by the modular
developments and the availability of specific data sets or sensors.
In this work we grounded our position in autonomous systems by providing a definition of requirements and presented a
general review of the relevant literature. We provided a proposal for a design of a general architecture for self-learning
agents. Our design included two separate modules, one for the creation of the data and the second for the independent
self-learning of an agent. We conclude by stating that the second module is, in general, divided into three stages,
where each stage is in charge of accomplishing an independent task: data collection, task generalization and emergent
behaviour. In very particular designs, generalization can influence emergent behaviours, but only in one direction.
We expanded our findings by presenting a concrete example of our ideas in the industrial manufacturing sector. Here
we suggest a division in three modular elements that feed a central agent for decision-making. Each module is able
to explore and learn from all possible scenarios and available data, taking into account changes or disruptions in its
own infrastructure and the information provided by the other modules. In the near future, we plan to present the first
implementation of our ideas and publish the conclusions about the feasibility of our design.
References
[1] Klaus Schwab. The Fourth Industrial Revolution. Crown Publishing Group, New York, NY, USA, 2017.
[2] Tesla. Future of driving. https://www.tesla.com/autopilot, 2019. Accessed: 2019-07-04.
[3] Rolls-Royce. Rolls-royce and finferries demonstrate world's first fully autonomous ferry. https://www.rolls-
royce.com/media/press-releases/2018/03-12-2018-rr-and-finferries-demonstrate-worlds-first-fully-autonomous-
ferry.aspx, 2018. Accessed: 2019-07-04.
[4] Stan Franklin and Art Graesser. Is it an agent, or just a program?: A taxonomy for autonomous agents. In
International Workshop on Agent Theories, Architectures, and Languages, pages 21 -- 35. Springer, 1996.
[5] Christer Carlsson, Mario Fedrizzi, and Robert Fullér. Fuzzy logic in management, volume 66. Springer Science &
Business Media, 2012.
[6] Tong Wang, Huijun Gao, and Jianbin Qiu. A combined adaptive neural network and nonlinear model predictive
control for multirate networked industrial process control. IEEE Transactions on Neural Networks and Learning
Systems, 27(2):416 -- 425, 2016.
[7] Stefan Boschert, R Rosen, and C Heinrich. Next generation digital twin. In Proceedings of the 12th International
Symposium on Tools and Methods of Competitive Engineering TMCE, pages 209 -- 218, 2018.
[8] Werner Kritzinger, Matthias Karner, Georg Traar, Jan Henjes, and Wilfried Sihn. Digital twin in manufacturing:
A categorical literature review and classification. IFAC-PapersOnLine, 51(11):1016 -- 1022, 2018.
[9] Leonardo Espinosa Leal, Anthony Chapman, and Magnus Westerlund. Reinforcement learning for extended
reality: designing self-play scenarios. In Proceedings of the 52nd Hawaii International Conference on System
Sciences, pages 156 -- 163, 2019.
[10] Peter Checkland and John Poulter. Soft systems methodology. In Martin Reynolds and Sue Holwell, editors,
Systems Approaches to Managing Change: A Practical Guide, pages 1430 -- 1436. Springer, 2010.
[11] CG Sørensen, S Fountas, E Nash, Liisa Pesonen, Dionysis Bochtis, Søren Marcus Pedersen, B Basso, and
SB Blackmore. Conceptual model of a future farm management information system. Computers and electronics
in agriculture, 72(1):37 -- 47, 2010.
[12] Lawrence Lee Rauch. Oscillation of a third order nonlinear autonomous system.
In S. Lefschetz, editor,
Contributions to the theory of nonlinear oscillations, volume 1, page 39. Princeton University Press, 1950.
[13] David Canfield Smith, Allen Cypher, and Jim Spohrer. Kidsim: programming agents without a programming
language. Communications of the ACM, 37(7):54 -- 67, 1994.
[14] Miltos Kyriakidis, Riender Happee, and Joost CF de Winter. Public opinion on automated driving: Results of
an international questionnaire among 5000 respondents. Transportation research part F: traffic psychology and
behaviour, 32:127 -- 140, 2015.
[15] SAE On-Road Automated Vehicle Standards Committee. Taxonomy and definitions for terms related to on-road
motor vehicle automated driving systems. SAE International, 2014.
[16] Francis J Anscombe, Robert J Aumann, et al. A definition of subjective probability. Annals of mathematical
statistics, 34(1):199 -- 205, 1963.
[17] Joscha Bach. Phenomenal experience and the perceptual binding state. In Papers of the 2019 Towards Conscious
AI Systems Symposium. CEUR-WS, 2019.
11
A PREPRINT - OCTOBER 22, 2019
[18] Joscha Bach. The cortical conductor theory: Towards addressing consciousness in ai models. In Biologically
Inspired Cognitive Architectures Meeting, pages 16 -- 26. Springer, 2018.
[19] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural
networks. In Advances in neural information processing systems, pages 1097 -- 1105, 2012.
[20] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436, 2015.
[21] Leonardo Espinosa Leal, Kaj-Mikael Björk, Amaury Lendasse, and Anton Akusok. A web page classifier library
based on random image content analysis using deep learning. In Proceedings of the 11th PErvasive Technologies
Related to Assistive Environments Conference, PETRA '18, pages 13 -- 16, New York, NY, USA, 2018. ACM.
[22] Prasad Pai. Data augmentation techniques in CNN using Tensorflow. https://bit.ly/2KLm8K6, 2017. Accessed:
2018-06-12.
[23] Justin Salamon and Juan Pablo Bello. Deep convolutional neural networks and data augmentation for environmental
sound classification. IEEE Signal Processing Letters, 24(3):279 -- 283, 2017.
[24] Kenji Doya. Reinforcement learning in continuous time and space. Neural computation, 12(1):219 -- 245, 2000.
[25] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville,
and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pages
2672 -- 2680, 2014.
[26] R Bellman. Dynamic Programming. Princeton University Press, 1957.
[27] Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. Reinforcement learning: A survey. Journal of
artificial intelligence research, 4:237 -- 285, 1996.
[28] Yuxi Li. Deep reinforcement learning: An overview. arXiv preprint arXiv:1701.07274, 2017.
[29] Richard S Sutton, Joseph Modayil, Michael Delp, Thomas Degris, Patrick M Pilarski, Adam White, and Doina
Precup. Horde: A scalable real-time architecture for learning knowledge from unsupervised sensorimotor
interaction. In The 10th International Conference on Autonomous Agents and Multiagent Systems-Volume 2, pages
761 -- 768. International Foundation for Autonomous Agents and Multiagent Systems, 2011.
[30] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data
engineering, 22(10):1345 -- 1359, 2010.
[31] Alexander Vezhnevets, Volodymyr Mnih, Simon Osindero, Alex Graves, Oriol Vinyals, John Agapiou, et al.
Strategic attentive writer for learning macro-actions. In Advances in neural information processing systems, pages
3486 -- 3494, 2016.
[32] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn
and think like people. Behavioral and Brain Sciences, 40, 2017.
[33] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies.
The Journal of Machine Learning Research, 17(1):1334 -- 1373, 2016.
[34] Yutian Chen, Matthew W Hoffman, Sergio Gómez Colmenarejo, Misha Denil, Timothy P Lillicrap, Matt
Botvinick, and Nando de Freitas. Learning to learn without gradient descent by gradient descent. arXiv preprint
arXiv:1611.03824, 2016.
[35] David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot,
Laurent Sifre, Dharshan Kumaran, Thore Graepel, et al. Mastering chess and shogi by self-play with a general
reinforcement learning algorithm. arXiv preprint arXiv:1712.01815, 2017.
[36] Johannes Heinrich, Marc Lanctot, and David Silver. Fictitious self-play in extensive-form games. In International
Conference on Machine Learning, pages 805 -- 813, 2015.
[37] Michiel Van Der Ree and Marco Wiering. Reinforcement learning in the game of othello: learning against a fixed
opponent and learning from self-play. In Adaptive Dynamic Programming And Reinforcement Learning (ADPRL),
2013 IEEE Symposium on, pages 108 -- 115. IEEE, 2013.
[38] Nick Jakobi, Phil Husbands, and Inman Harvey. Noise and the reality gap: The use of simulation in evolutionary
robotics. In European Conference on Artificial Life, pages 704 -- 720. Springer, 1995.
[39] Alexey Dosovitskiy and Vladlen Koltun. Learning to act by predicting the future. arXiv preprint arXiv:1611.01779,
2016.
[40] Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. Cambridge, MA: MIT Press,
2011.
12
A PREPRINT - OCTOBER 22, 2019
[41] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex
Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep
reinforcement learning. Nature, 518(7540):529, 2015.
[42] Michał Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Ja´skowski. Vizdoom: A
doom-based ai research platform for visual reinforcement learning. In Computational Intelligence and Games
(CIG), 2016 IEEE Conference on, pages 1 -- 8. IEEE, 2016.
[43] Jack Harmer, Linus Gisslén, Henrik Holst, Joakim Bergdahl, Tom Olsson, Kristoffer Sjöö, and Magnus Nordin.
Imitation learning with concurrent actions in 3d games. arXiv preprint arXiv:1803.05402, 2018.
[44] Charles Beattie, Joel Z Leibo, Denis Teplyashin, Tom Ward, Marcus Wainwright, Heinrich Küttler, Andrew
Lefrancq, Simon Green, Víctor Valdés, Amir Sadik, et al. Deepmind lab. arXiv preprint arXiv:1612.03801, 2016.
[45] Per-Arne Andersen, Morten Goodwin, and Ole-Christoffer Granmo. Towards a deep reinforcement learning
approach for tower line wars. In International Conference on Innovative Techniques and Applications of Artificial
Intelligence, pages 101 -- 114. Springer, 2017.
[46] Oriol Vinyals, Timo Ewalds, Sergey Bartunov, Petko Georgiev, Alexander Sasha Vezhnevets, Michelle Yeo,
Alireza Makhzani, Heinrich Küttler, John Agapiou, Julian Schrittwieser, et al. Starcraft ii: a new challenge for
reinforcement learning. arXiv preprint arXiv:1708.04782, 2017.
[47] Gabriel Synnaeve, Nantas Nardelli, Alex Auvolat, Soumith Chintala, Timothée Lacroix, Zeming Lin, Florian
Richoux, and Nicolas Usunier. Torchcraft: a library for machine learning research on real-time strategy games.
arXiv preprint arXiv:1611.00625, 2016.
[48] Yuandong Tian, Qucheng Gong, Wenling Shang, Yuxin Wu, and C Lawrence Zitnick. Elf: An extensive,
lightweight and flexible research platform for real-time strategy games. In Advances in Neural Information
Processing Systems, pages 2656 -- 2666, 2017.
[49] OpenAI. Dota 2. https://blog.openai.com/dota-2/, 2018. Accessed: 2018-06-12.
[50] Niels Justesen, Philip Bontrager, Julian Togelius, and Sebastian Risi. Deep learning for video game playing. arXiv
preprint arXiv:1708.07902, 2017.
[51] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David
Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International
Conference on Machine Learning, pages 1928 -- 1937, 2016.
[52] Tejas D Kulkarni, Ardavan Saeedi, Simanta Gautam, and Samuel J Gershman. Deep successor reinforcement
learning. arXiv preprint arXiv:1606.02396, 2016.
[53] Ziyu Wang, Tom Schaul, Matteo Hessel, Hado Van Hasselt, Marc Lanctot, and Nando De Freitas. Dueling network
architectures for deep reinforcement learning. arXiv preprint arXiv:1511.06581, 2015.
[54] Seyed Sajad Mousavi, Michael Schukat, and Enda Howley. Deep reinforcement learning: an overview. In
Proceedings of SAI Intelligent Systems Conference, pages 426 -- 440. Springer, 2016.
[55] Kai Arulkumaran, Marc Peter Deisenroth, Miles Brundage, and Anil Anthony Bharath. A brief survey of deep
reinforcement learning. arXiv preprint arXiv:1708.05866, 2017.
[56] Joshua Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain ran-
domization for transferring deep neural networks from simulation to the real world. CoRR, abs/1703.06907,
2017.
[57] David Ha and Jürgen Schmidhuber. World models. CoRR, abs/1803.10122, 2018.
[58] Yan Duan, Marcin Andrychowicz, Bradly C. Stadie, Jonathan Ho, Jonas Schneider, Ilya Sutskever, Pieter Abbeel,
and Wojciech Zaremba. One-shot imitation learning. CoRR, abs/1703.07326, 2017.
[59] Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning
via meta-learning. CoRR, abs/1709.04905, 2017.
[60] Martin A. Riedmiller, Roland Hafner, Thomas Lampe, Michael Neunert, Jonas Degrave, Tom Van de Wiele,
Volodymyr Mnih, Nicolas Heess, and Jost Tobias Springenberg. Learning by playing - solving sparse reward tasks
from scratch. CoRR, abs/1802.10567, 2018.
[61] Trapit Bansal, Jakub Pachocki, Szymon Sidor, Ilya Sutskever, and Igor Mordatch. Emergent complexity via
multi-agent competition. CoRR, abs/1710.03748, 2017.
[62] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align
and translate. arXiv preprint arXiv:1409.0473, 2014.
13
A PREPRINT - OCTOBER 22, 2019
[63] Stefano V Albrecht and Subramanian Ramamoorthy. A game-theoretic model and best-response learning
method for ad hoc coordination in multiagent systems. In Proceedings of the 2013 international conference on
Autonomous agents and multi-agent systems, pages 1155 -- 1156. International Foundation for Autonomous Agents
and Multiagent Systems, 2013.
[64] Stefano V Albrecht and Peter Stone. Autonomous agents modelling other agents: A comprehensive survey and
open problems. Artificial Intelligence, 258:66 -- 95, 2018.
[65] Stuart J Russell and Peter Norvig. Artificial intelligence: a modern approach. Malaysia; Pearson Education
Limited„ 2016.
[66] Michael Wooldridge. An introduction to multiagent systems. John Wiley & Sons, 2009.
[67] Zbigniew Michalewicz, Martin Schmidt, Matthew Michalewicz, and Constantin Chiriac. Adaptive business
intelligence. Springer, 2006.
[68] Michael Wooldridge, Nicholas R. Jennings, and David Kinny. The gaia methodology for agent-oriented analysis
and design. Autonomous Agents and Multi-Agent Systems, 3(3):285 -- 312, Sep 2000.
[69] Jacques Ferber, Olivier Gutknecht, and Fabien Michel. From agents to organizations: An organizational view
of multi-agent systems. In Paolo Giorgini, Jörg P. Müller, and James Odell, editors, Agent-Oriented Software
Engineering IV, pages 214 -- 230, Berlin, Heidelberg, 2004. Springer Berlin Heidelberg.
[70] Jaelson Castro, Manuel Kolp, and John Mylopoulos. Towards requirements-driven information systems engineer-
ing: the tropos project. Information systems, 27(6):365 -- 389, 2002.
[71] Rafael H Bordini, Jomi Fred Hübner, and Michael Wooldridge. Programming multi-agent systems in AgentSpeak
using Jason, volume 8. John Wiley & Sons, 2007.
[72] Annapurna Valluri, Michael J North, and Charles M Macal. Reinforcement learning in supply chains. International
journal of neural systems, 19(05):331 -- 344, 2009.
[73] Xianyu Zhang, Xinguo Ming, Zhiwen Liu, Dao Yin, Zhihua Chen, and Yuan Chang. A reference framework and
overall planning of industrial artificial intelligence (i-ai) for new application scenarios. The International Journal
of Advanced Manufacturing Technology, pages 1 -- 23, 2018.
[74] Mihai Barbuceanu and Mark S Fox. Coordinating multiple agents in the supply chain.
In Proceedings of
WET ICE'96. IEEE 5th Workshop on Enabling Technologies; Infrastucture for Collaborative Enterprises, pages
134 -- 141. IEEE, 1996.
[75] Pierpaolo Pontrandolfo, Abhijit Gosavi, O Geoffrey Okogbaa, and Tapas K Das. Global supply chain management:
a reinforcement learning approach. International Journal of Production Research, 40(6):1299 -- 1317, 2002.
[76] Tim Stockheim, Michael Schwind, and Wolfgang Koenig. A reinforcement learning approach for supply chain
management. In 1st European Workshop on Multi-Agent Systems, Oxford, UK, 2003.
[77] Jeroen P Van den Berg and Willem HM Zijm. Models for warehouse management: Classification and examples.
International journal of production economics, 59(1-3):519 -- 528, 1999.
[78] Reza Moazzez Estanjini, Yingwei Lin, Keyong Li, Dong Guo, and Ioannis Ch Paschalidis. Optimizing warehouse
forklift dispatching using a sensor network and stochastic learning. IEEE Transactions on Industrial Informatics,
7(3):476 -- 486, 2011.
[79] Reza Moazzez Estanjini, Keyong Li, and Ioannis Ch Paschalidis. A least squares temporal difference actor -- critic
algorithm with applications to warehouse management. Naval Research Logistics (NRL), 59(3-4):197 -- 211, 2012.
[80] Joren Gijsbrechts, Robert N Boute, Jan A Van Mieghem, and Dennis Zhang. Can deep reinforcement learning
improve inventory management? performance and implementation of dual sourcing-mode problems. Performance
and Implementation of Dual Sourcing-Mode Problems (December 17, 2018), 2018.
[81] Lukasz Golab and Theodore Johnson. Data stream warehousing. In 2014 IEEE 30th International Conference on
Data Engineering, pages 1290 -- 1293. IEEE, 2014.
[82] Stefan Boschert and Roland Rosen. Digital twin -- the simulation aspect. In Mechatronic Futures, pages 59 -- 74.
Springer, 2016.
[83] Roland Rosen, Georg Von Wichert, George Lo, and Kurt D Bettenhausen. About the importance of autonomy and
digital twins for the future of manufacturing. IFAC-PapersOnLine, 48(3):567 -- 572, 2015.
[84] Mike Shafto, Mike Conroy, Rich Doyle, Ed Glaessgen, Chris Kemp, Jacqueline LeMoigne, and Lui Wang. Model-
ing, simulation, information technology & processing roadmap. National Aeronautics and Space Administration,
2012.
14
[85] A Wayne Wymore. Model-based systems engineering. CRC press, 1993.
[86] Judea Pearl and Dana Mackenzie. The Book of Why: The New Science of Cause and Effect. Basic Books, 2018.
A PREPRINT - OCTOBER 22, 2019
15
|
1904.08310 | 1 | 1904 | 2019-04-13T18:36:39 | LegenDary 2012 Soccer 2D Simulation Team Description Paper | [
"cs.MA",
"cs.RO"
] | In LegenDary project, we started a new research based on Agent2D in RoboCup 2D soccer simulation. In this paper, we mainly present the team algorithms and structures which we used to develop our team in separated section. We have focused on passing, dribbling and blocking skills. We improved them and made the team ready for this competition. Through pass is the most important part of our team that we work a lot on it. | cs.MA | cs | LegenDary 2012 Soccer 2D Simulation
Team Description Paper
Pourya Saljoughi1, Reza Ma'anijou1, Ehsan Fouladi1, Narges Majidi1,
Saber Yaghoobi1, Houman Fallah1 and Saeideh Zahedi1
1 Islamic Azad University of Lahijan
[email protected]
Abstract. In LegenDary project, we started a new research based on Agent2D
in RoboCup 2D soccer simulation. In this paper, we mainly present the team
algorithms and structures which we used to develop our team in separated
section. We have focused on passing, dribbling and blocking skills. We
improved them and made the team ready for this competition. Through pass is
the most important part of our team that we work a lot on it.
Keywords: 2D, Block, Through Pass, Agent, Dribble
1 Introduction
The main goals of LegenDary project are both increasing our knowledge level in
artificial intelligence and helping to improve this field of study in the world. For
achieving these purposes we tried so hard in last few years and improved so well as it
is described in continue.
LegenDary team is competing for the first time in RoboCup 2012. The first
implementation of team was based on UVA base source code [1] after a few
competitions we were involved we decided to use agent 2d base source code [2] as
code to work on due to the good implementation of low level layer and also by
experience; it has been proven that it performs much better than UVA base. We've
added our own AI methods to the code and made many changes in original code to
have our own team strategies and make it more flexible and powerful than it was.
The project's research focus is also concerned about developing cooperation
models for soccer playing. Although most approaches to cooperation in multi-robot
systems consider the control of individual robot behavior separately from the
cooperative group behavior [3], we tried to have a good cooperation in our team
between different agents. Any Multi-Agent cooperation is based on how the single
agent adapts to the Multi-Agent System. If every element in the system can
accommodate with the system, the system is steady [ 4]. Agent's action is done based
on current world model state and zone and every decision is based on main goal of
team in that state and agent's goal don't conflict each other. Therefore corporation
will be possible.
Our team's code consists of many different parts that are discussed in following
sections. We've focused on passing, dribbling and defensive skills. Through pass and
blocking system are powerful parts of our team strategy. In defensive part of team, we
used both blocking system and positioning system together to have a strong defensive
system and for advantage of using these abilities together we used an experiment
against three different teams by using and not using blocking system.
2 Pass
Passing is the most important skill in every team. A good passing system can make a
good scoring position and also can clear the ball from the danger area. When a player
can't make a better situation for himself, and the others have a better one, the pass
function will be called. For this skill we inspire from MarliK[5]. We generally can
classify the passing in some categories which are mentioned below.
2.1 Direct Pass
A direct pass is a simple pass that its target is one of our teammate's exact positions.
This skill will be used when our teammates are close enough. So when the target is
far, the direct pass won't have enough accuracy and it won't arrive to the target.
First of all the appropriate target will be chosen. In choosing a good target
different factors will be considered and we will rate our targets. So the summation of
them will be the final result of that pass rate. A pass rate will be calculated with the
formula 1. Cn is the factors coefficient and En is the value of that factor in formula 1.
Score = C1 * E1 + C2 * E2 + C3 * E3 + … .
(1)
After sorting targets by their rate, the best target will be chosen and the player will
kick the ball with an appropriate angle and speed to that target.
2.2 Through Pass
If there is no possibility to dribble or direct pass, through pass is used as a strategy to
break the opponent's defense line. This skill is useful in offensive situations when
passing and dribbling are not effective enough to make an attacking situation.
To see if there is a possibility to pass, we use prediction system for different points
around each player to find possible points for passing. The prediction system that we
used is similar to Ball Interception Model in Bahia2D team [6].The algorithm for this
process is shown in table 1.
Table 1. A simplified algorithm for through pass
While r < 5
Initialize canReachBall to false
Initialize distancePoint with angle of t and radius of r
Initialize targetPoint array and its members according to:
targetPoint[i] ← teammate player position + distancePoint
SORT targetPoint array from longest distance to shortest distance
Repeat for every member of targetPoint array:
Initialize the environment's current situations.
Repeat for every teammate in offensive situation:
Initialize r to 0
Initialize t to 35
While t>3
Until next member of targetPoint is not checked.
Until next seen teammate is not checked and targetPoint is not reliable.
RETURN
kick the ball to this reliablePoint
RETURN
Call predict to see if teammate is first player who reach the ball and SET
canReachBall.
If canReachBall THEN
SET r TO r+1
SET t TO t-1
For each point, prediction algorithm considers the possibility of getting the ball to
the teammate. Different factors are used. We consider ball direction, ball speed and
other parameters. These points are usually closest points to the goal with different
angels and different distances from each player.
To send more through pass and make better situations, it's better for teammates to
be near the offside line. Therefore they can reach faster than other player to the ball.
This method is implemented separately for each player who is in offensive situation
and each player tries to keep a minimum distance to offside line. So more through
passes happen. An example of through pass is shown bellow.
Fig. 1. Player 7 makes a good through Pass to player 9 which is positioned near the offside
line in order to reach the ball faster.
3 Dribble
Dribbling is one of the most useful abilities of a team. We get the idea for this skill
from ESKILAS [10]. Moving ball forward and keeping it away from the opponents is
the main part of this skill. In the algorithm which we used, our main purpose is to
reach the opponent's goal in predetermined ways as fast as possible. As a matter of
fact paths which lead us closer to opponent's goal are in priority. When agent has the
ball, this algorithm specifies which direction can place him in safer and better
situation.
This process will be done with aid of player's individual vision and information
about teammates' and opponents' position in the last cycles and his position towards
the opponent's goal. With this information the predetermined ways will be rated and
the best way will be chosen.
In our dribble skill, the dribble path is based on the predetermined paths (figure 2).
So the player would rather to choose the path which is near by the main dribble path.
And if he (or his chosen path) got blocked by the opponent(s), according to the
position of the ball and the player itself, he decides to change the path or pass the ball
to the safest teammate. If the player decides to change the path, the new path will be
the nearest one to the main path and the furthest to the opponents, to keep the ball
safe.
Fig. 2. Predetermined Paths for Dribbling
4 Block and Defensive system
Our Defensive system is divided into two parts. The first part is positioning. We
realized that with an optimized formation system and a simple intercept action for the
ball we can make base defense more powerful. Player's position is based on ball
position in each cycle. We created our formation by fedit software that is available for
agent2d base [7].The second part is blocking system.
Block system is one of the most important and efficient way for defending. We
inspire this skill from LEAKIN'DROPS [8].
For capturing the opponent's ball, the player can choose between two different
ways. The first and simple way is forcing straight ahead to the ball's position and
intercepts it. The other way is to predict the target of opponent and get to the position
to block the player and intercept the ball. It seems to be more effective and faster, also
this method saves the stamina.
The second way is called "blocking system". As a matter of fact the first method is
useless, because the opponent's position may change every moment and the player is
not fixed. And the better way is forecasting the opponent's target and move through
that direction to face the opponent. This method is better than the first one, because
the player is not going to get dribbled and he saves time for other teammates to arrive
to defense position.
Our blocking algorithm calculates a point (which is the best) between the
opponent's direction and our goal according to the information he receives from the
servers, like position, velocity and direction of the ball. According to the calculations,
if the chosen point wasn't too far to approach and the agent was not so weak on
stamina, the blocking commands will be executed.
Our blocking system is somehow dynamic and intelligent. Its decision changes
according to the zone and opponent goal. We defined two ways to intercept the ball
after standing in the proper block position. The player decides whether standing right
there and waits for the opponent (with ball) to get close, or forcing the opponent and
intercepts the ball.
We made an experiment to test our blocking system. In this experiment, we use 3
teams in 3 level of attacking power, Oxsy [9], MarliK, Helios Base. We selected these
teams according to average scored in RoboCup 2011. We run 10 games against each
teams (5 with blocking system and 5 without blocking system). The results of these
matches are mentioned in table 2.
Table 2. Experiment's results
Team
Name
Blocking
System
Condition
M1
M2
M3
M4
M5
Total
score
Average
Scored
Oxsy
MarliK
Helios
Base
on
off
on
off
on
off
4:2
1-2 0-0 1-0 1-0 1-0
20:0
3-0 2-0 6-0 6-0 3-0
4:1
0-0 2-0 0-0 1-0 1-1
15:0
3-0 3-0 4-0 2-0 3-0
1-6 0-6 1-4 0-3 1-4
3:23
2-3 6-3 6-3 3-4 4-6 21:19
0.8
4
0.8
3
0.6
4.2
opponent
average score
in RoboCup
2011
8.5
3.92
-
As we see in the table, we can decrease the average against scored from 3.73 to
0.73 per match with blocking system. And this result shows that our blocking system
is useful.
5 Future Work
In this paper we described some parts of our team. We spent a lot of time to improve
our team's way of playing in different situation. But we have to work more and more
to achieve our goal which we mentioned in the introduction section and there are
several problems which must be corrected and some bugs that need to be fixed. In the
next phase of LegenDary project we are going to improve our team and make it more
stable and realistic by using AI algorithms.
References
1. UvA Trilearn 2003. http://www.science.uva.nl/jellekok/robocup/2003/
2. agent2d source: http://sourceforge.jp/projects/rctools/
3. Jung, D., Zelinsky, A.: An Architecture for Distribute Cooperative-Planning in A Behavior-
Based Multi-Robot System n: Journal of Robotics & Autonomous Systems (RA&S), special
issue on Field & Service Robotics.
4. Rongya Chen, Yiming Li, Hao Wang, Baofu Fang: HfutEngine2011 Simulation 2D
Team Description Paper. In: RoboCup 2011 Symposium and Competitions, Turkey (2011)
5. Tavafi, A., Nozari, N., Vatani, R., Rad Yousefi, M., Rahmatinia, S., Pirdeyr, P.: MarliK
2011 Team Description Paper. In: RoboCup 2011 Symposium and Competitions, Turkey
(2011)
6. Silva, B., Frias, d., Eloi, I., Grimaldo, J,. Simões, M.:Bahia2D 2010 Team Description
Paper. In: RoboCup 2010 Symposium and Competitions, Singapore(2010)
7. Hidehisa Akiyama, Hiroki Shimora, Tomoharu Nakashima, Yosuke Narimoto,
and Tomohiko Okayama. HELIOS2011 Team Description Paper. In: RoboCup 2011
Symposium and Competitions, Turkey (2011)
8. Tavafi, A., Nozari, N., Vatani, R., Rahmatinia, S.:LEAKIN'DROPS 2D 2010 Team
Description Paper. In: IranOpen 2010 Symposium and Competitions, Iran(2010)
9. Marian, S., Luca, D., Sarac, B., Cotarlea O.: Oxsy 2011 Team Description Paper. In:
RoboCup 2011 Symposium and Competitions, Turkey (2011)
10. Bakhtiari, M., Montazeri, M., Saharkhiz, S., Kaviani, P.: ESKILAS 2011 Team Description
Paper. In: RoboCup 2011 Symposium and Competitions, Turkey (2011)
|
1001.5275 | 1 | 1001 | 2010-01-28T21:38:04 | An Agent-Based Modeling for Pandemic Influenza in Egypt | [
"cs.MA",
"cs.CY"
] | Pandemic influenza has great potential to cause large and rapid increases in deaths and serious illness. The objective of this paper is to develop an agent-based model to simulate the spread of pandemic influenza (novel H1N1) in Egypt. The proposed multi-agent model is based on the modeling of individuals' interactions in a space time context. The proposed model involves different types of parameters such as: social agent attributes, distribution of Egypt population, and patterns of agents' interactions. Analysis of modeling results leads to understanding the characteristics of the modeled pandemic, transmission patterns, and the conditions under which an outbreak might occur. In addition, the proposed model is used to measure the effectiveness of different control strategies to intervene the pandemic spread. | cs.MA | cs | An Agent-Based Modeling for Pandemic Influenza in Egypt
Khaled M. Khalil, M. Abdel-Aziz, Taymour T. Nazmy, Abdel-Badeeh M. Salem
Faculty of Computer and Information Science Ain shams University Cairo, Egypt
[email protected], [email protected], [email protected],
[email protected]
Abstract
Pandemic influenza has great potential to cause large
and rapid increases in deaths and serious illness. The
objective of this paper is to develop an agent-based model
to simulate the spread of pandemic influenza (novel
H1N1) in Egypt. The proposed multi-agent model is based
on the modeling of individuals' interactions in a space-
time context. The proposed model involves different types
of parameters such as: social agent attributes, distribution
of Egypt population, and patterns of agents' interactions.
Analysis of modeling results leads to understanding the
characteristics of the modeled pandemic, transmission
patterns, and the conditions under which an outbreak
might occur. In addition, the proposed model is used to
measure the effectiveness of different control strategies to
intervene the pandemic spread.
Keywords
Pandemic Influenza, Epidemiology, Agent-Based Model,
Biological Surveillance, Health Informatics.
1. Introduction
The first major pandemic influenza H1N1 is recorded
in 1918-1919, which killed 20-40 million people and is
thought to be one of the most deadly pandemics in human
history. In 1957, a H2N2 virus originated in China,
quickly spread throughout the world and caused 1-4
million deaths world wide. In 1968, an H3N2 virus
emerged in Hong Kong for which the fatalities were 1-4
million [16]. In recent years, novel H1N1 influenza has
appeared. Novel H1N1 influenza is a swine-origin flue
and is often called swine flue by the public media. The
novel H1N1 outbreak began in Mexico, with evidence that
there had been an ongoing epidemic for months before it
was officially recognized as such. It is not known when
the epidemic will occur or how sever it will be. Such an
outbreak would cause a large number of people to fall ill
and possibly die.
In the absence of reliable pandemic detection systems,
computer models and systems have become an important
information tools for both policy-makers and the general
public [15]. Computer models can help in providing a
global insight of the infectious disease outbreaks' behavior
by analyzing the spread of infectious diseases in a given
population, with varied geographic and demographic
features [12]. Computer models promise an improvement
in representing and understanding the complex social
structure as well as the heterogeneous patterns in the
contact networks of real-world populations determining
the transmission dynamics [4]. One of the most recent
approaches of such sophisticated modeling is agent-based
modeling [3]. Agent-based modeling of pandemics
recreates
the entire populations and their dynamics
through incorporating social structures, heterogeneous
connectivity patterns, and meta-population grouping at the
scale of the single individual [3].
In this paper we propose a stochastic multi-agent
model to mimic the daily person-to-person contact of
people in a large scale community affected by a pandemic
influenza in Egypt. The proposed model is used to: (i)
assess the understanding of transmission dynamics of
pandemic influenza, (ii) assess the potential range of
consequences of pandemic influenza in Egypt, and (iii)
assess the effectiveness of different pandemic control
strategies on the spread of the pandemic. We adopt
disease parameters and
recommended control
the
strategies from WHO [16]. While, we use Egypt census
data of 2006 [7] to create the population structure, and the
distribution of social agent attributes. Section 2 reviews
approaches:
epidemiological modeling
different
mathematical modeling,
cellular
automata
based
modeling, and agent based modeling. While, section 3
reviews related multi-agent models in literature. Section 4
discusses the proposed model, and section 5 validates the
proposed model. Section 6 discusses the pandemic control
strategies and their effect on the spread of the pandemic.
Section 7 presents the modeling experiments and analysis
of results, and then we conclude in Section 8.
2. Epidemiological modeling approaches
The search for an understanding of the behavior of
infectious diseases spread has resulted in several attempts
to model and predict the pattern of many different
communicable diseases through a population [5]. The
earliest account was carried out in 1927 by Kermack and
McKendrick [9]. Kermack and McKendrick created a
mathematical model named SIR (Susceptible-Infectious-
Recovered) based on ordinary differential equations.
Kermack and McKendrick started with the assumption
that all members of the community are initially equally
susceptible to the disease, and that a complete immunity is
conferred after the infection. The population is divided
into three distinct classes (see figure 1): the susceptible
(S) healthy individuals who can catch the disease; the
infectious (I) those who have the disease and can transmit
it; and the recovered (R) individuals who have had the
disease and are now immune to the infection (or removed
from further propagation of the disease by some other
means).
Figure 1. SIR (Susceptible–Infectious–
Recovered) Model
Let
( )tS
, and ( )tR be the number of susceptible,
, ( )tI
infected and recovered individuals, respectively, at time t,
and N is the size of the fixed population, so we have:
( )
( )tR
( )
+
+
=
tSN
tI
Upon contact with an infected a susceptible individual
contracts the disease with probabilityβ , at which time he
immediately becomes
(no
infectious
infected and
incubation period); infectious recover at an individual rate
γ per unit time. Based on mentioned assumptions;
Kermack and McKendrick derived the classic epidemic
SIR model as follows:
dS
(1)
β−=
SI
dt
dI
dt
dR
=
β −
SI
γ
I
(2)
γ=
I
dt
From equations (1) and (2), we found that SIR model is
deterministic and doesn't study the nature of population
vital dynamics
(handling newborns
and deaths).
Following Kermack and McKendrick, other physicians
contributed
to modern mathematical epidemiology;
extending SIR model with more classes and supporting
vital dynamics such as: SEIR (Susceptible–Exposed–
Infectious–Recovered),
(Immunized–
and MSEIR
Susceptible–Exposed–Infectious–Recovered) models [9].
However, mathematical models had not taken into account
spatial and temporal factors such as variable population
structure, and dynamics of daily individuals' interactions
which drive more realistic modeling results [1].
The second type of developed models is cellular
incorporate spatial
automata based models, which
parameters
to
better
reflect
the
heterogeneous
environment found in nature [13]. Cellular automata based
to using deterministic
models are an alternative
two-dimensional
differential equations, which use a
cellular
automaton
to model
location
specific
characteristics of the susceptible population together with
stochastic parameters which captures the probabilistic
nature of disease transmission [2]. Typically a cellular
automaton consists of a graph where each node is a cell.
This graph is usually in the form of a two-dimensional
lattice whose cells evolve according to a global update
function applied uniformly over all the cells [13]. Cell
state takes one of the SIR model states, and is calculated
based on cell present state and the states of the cells in its
interaction neighborhood. As the Cellular automata based
model evolves, cells states will determine how the overall
behavior of a complex system [2]. However, cellular
automata based models neglect the social behavior and
dynamics interactions among individuals in the modeled
community. Therefore, cellular automata gave way to a
new approach; Agent-based models.
Agent-based models (ABM) are similar to cellular
automata based models, but leverage extra tracking of the
effect of the social interactions of individual entities [1].
Agent-based model consists of a population of agents, an
environment, and set of rules managing agents' behavior
[12]. Each agent has two components: a state and a step
function. Agent state describes every agent attributes
values at the current state. The step function creates a new
state (usually stochastically) representing
the agent
attributes at the next time step. The great benefit of agent-
based models that these models allow epidemiological
researchers to do a preliminary "what-if" analysis with the
purpose of assessing systems' behavior under various
conditions and evaluating which alternative control
strategies to adopt in order to fight epidemics [12].
3. Multi-agent related models
Related agent-based models are Perez-Dragicevi
model, BIOWar, and EpiSims. Perez-Dragicevi [12] had
developed a multi-agent model to simulate the spread of a
communicable disease in an urban environment using
measles outbreak in an urban environment as a case study.
The model uses SEIR (Susceptible–Exposed–Infectious–
Recovered) model and makes use of census data of
Canada. The goal of this model is to depict the disease
progression based on individuals' interactions through
calculation of ratios of susceptible/infected in specific
time frames and urban environments. BIOWar [10] is a
computer model that combines computational models of
social networks, communication media, and disease
transmission with demographically resolved agent models,
urban spatial models, weather models, and a diagnostic
error model to produce a single integrated model of the
impact of a bioterrorist attack on an urban area. BIOWar
models the population of individual agents as they go
about their lives. BIOWar allows the study of various
attacks and containment policies as revealed through
indicators such as medical phone calls, insurance claims,
death rates, over-the counter pharmacy purchases, and
hospital visit rates, among others. EpiSims [8] is an agent-
based model, which combines realistic estimates of
population mobility, based on census and land-use data of
USA, with configurable parameters for modeling the
progress of a disease. EpiSims involves a way to generate
synthetic realistic social contact networks in a large urban
region.
However, the proposed agent-based model will differ
from the above models for: (i) usage of census data of
Egypt, (ii) proposed extension to SIR model, (iii) and
studying the effect of different control strategies on the
spread of the disease. This study is considered very
important which incorporate Egypt population structure in
modeling process. We have adopted Egypt census data of
2006 for creating realistic social contact networks such as
home, work and school networks. While, the proposed
extension
to SIR model encompasses new classes;
modeling the real pandemic behavior and control states
such as (in contact, quarantined, not quarantined, dead,
and immunized). In addition, we involve the study of the
effect of different control strategies on the spread of the
pandemic influenza. We plan in future work to integrate
the proposed model with different simulation tools and
models such as weather models, transportation models,
and decision support models to build a complete system
for pandemic management in Egypt.
4. Proposed model
In what follows, we propose an extension to SIR
model. Then we propose the multi-agent model based on
the proposed extension of SIR model states. Finally, we
validate the proposed multi-agent model by aligning with
the classical SIR model.
4.1. Proposed extension to SIR model
We propose an extension to SIR model by adding extra
classes to represent more realistic agent states (see figure
2). In addition, we adopt stochastic approach to traverse
among agent states using normal distribution. Agents are
grouped based on the proposed extension to SIR model
into nine classes. The first class is the (S) Susceptible
agents, who are not in contact with infectious agents and
are subject to be infected. At the start of the modeling, all
agents fall in the (S) Susceptible class. The second class is
the (C) in Contact agents, who are in direct contact with
other infectious agents. The third class is the (E) Exposed
agents, who are infected agents during the incubation time
(latent) of the disease. The fourth class is the (I) Infectious
agents, who are contagious. The fifth class is the (Q)
Quarantined agents, who are infected agents quarantined
by the health care authorities. The sixth class is the (NQ)
Not Quarantined agents, who are infected agent but not
quarantined. The seventh class is the (D) Dead agents.
The eighth class is the (R) Recovered agents. The ninth
class is the (M) Immunized agents, who are immunized
against the disease infection.
Figure 3 presents flow chart which explains in details
the sequence of the state chart of the proposed extension
to SIR model. All population members are born
susceptible then may contact contagious agents (move
into in-contact class). In-contact agents may acquire the
infection (move into the exposed class) based on given
distribution. Exposed agents remain non-contagious for
given latent time. At the end of the latent time, agents will
become contagious (move into the infectious class).
Infected agents may ask for doctor help and thus become
Figure 2. State chart of proposed extension to SIR model. (S) Susceptible, (C) in Contact, (E) Exposed,
(I) Infectious, (Q) Quarantined, (NQ) Not Quarantined, (D) Dead, (R) Recovered, and (M) Immunized.
Ind ivid ual is Sus cep tible
No
No
No
Is in co ntact
with infectio us
individual?
Yes
Individual is
in-Co ntact
Individual is
infec ted by
given
probability?
Yes
Individual is
Expos ed
Has the
expos ed period
concluded?
Individual is
Infec tious
Is ind ivid ual
quarantined ?
Yes
Individual is
Quarantined
Yes
No
No
Individual is
Not
Quarantined
Is ind ividua l
quarantined ?
No
Is individual
dead by given
probability?
Is individua l
recovered?
Individual is
Recovered
Yes
Yes
Yes
No
No
Is ind ividual
dead ?
Is individual
immunized?
Yes
No
Individual is Dead
Yes
Individual is
Immunized
Figure 3. Flow chart of the proposed extension to SIR model.
quarantined by health care authorities (move
into
quarantined class), or ignore disease symptoms (move to
non-quarantined class) based on given distribution. Non-
quarantined agents are the main source of disease in this
model. Non-quarantined agents may ask for doctor help
and thus become quarantined, or die (move to dead class)
based on given distribution. Quarantined agents may
response to disease drugs and become recovered (move to
recovered class) or die (move to dead class) based on
given distribution. Recovered agents may become
immunized (move to immunized class) based on given
distribution, or become susceptible again.
4.2. Proposed multi-agent model
The proposed multi-agent based model attempts to
realistically represent the behavior of individuals' daily
activities, and the natural biological process of the
pandemic influenza spread among individuals as a result
of individuals' interactions. Proposed agent-based model
involves (i) population agents, (ii) agents' rules which
govern the behavior of the agents, (iii) and the infection
transmission patterns following the proposed extension to
SIR model. Agents represent human population, in which
each agent is involved in a sequence of daily basis
activities according to the agent social type. These daily
activities allows agents to interact themselves in groups or
even travel and join other groups. The daily activities of
working, travelling, and public gathering are modeled,
while agents' states are calculated on discrete time steps
during agent life time.
Proposed multi-agent model has several parameters
such as: simulation parameters, disease model parameters,
agents' attributes, and population distribution based on
census data. First: simulation parameters which include (i)
number of days to be simulated, (ii) random seed for the
gaussian random number generator, (iii) population size,
(iv) and initial agents. Second: disease model parameters
which include: (i) incubation time which is the average
time of infected agent before being contagious, (ii)
percentage of recovered infected agents after treatment,
(iii) percentage of immunized agents after the recovery,
(iv) percentage of dead agents, (v) average minimum and
maximum time required to recover infected agent, and
finally (vi) percentage of quarantined infected agents (see
Table 1 for parameters default values). Third: agent
attributes which are crucial for describing the nature of the
pandemic and control the behavior of agent among time
and space. Agent attributes includes: (i) health state
(based on proposed extension to SIR Model states), (ii)
social activities level (High, Moderate, Low), (iii) daily
movement, (iv) spatial location, (v) infection time, (vi)
social type (SPOUSE, PARENT, SIBLING, CHILD,
OTHERFAMILY, COWORKER, GROUPMEMBER,
NEIGHBOR, FRIEND, ADVISOR, SCHOOLMATE,
OTHER), and (vii) agent social networks. Social activities
level controls the number of daily contacts of the agent,
which proportionally affect the number of in-contact
agents interacting with the infected agent. Increasing
number of in-contact agents adds more chance to the
reproduction number to increase which means epidemic
outbreak [4]. Reproduction number (R) can be defined as
the average number of secondary agents infected by a
primary agent case [16]. We have distributed number of
daily infected cases based on the social activities level as
following: Low: 2 agents, Moderate: 3 agents, and High 4
agents. Social type distribution is based on Egypt Census
data of 2006 [7]. Egypt census data has classified
population into five classes (see Table 2). The distribution
of social types based on census data leads to different
social network structure. All involved distributions are
assumed to be Gaussian distribution with mean of 0 and
standard deviation of 1. Distributions are subject to be
replaced in future work according to the availability of
more information about Egypt population.
At the beginning of the system startup, configurations are
loaded and user is required to set up initial agents. The
simulation runs in a loop for a pre-specified number of
days. When the simulation starts, the first step is to create
the initial agents instances. Next, agents start practicing
their natural daily activities. During each day, every
member of the agent community moves around, and
communicates with their social network agents or with
public agents. See figure 4 for the spatial representation of
agents moving and contacting each other, and their health
states represented with different icons. Daily moved
distance by agents and the number of contacted agents are
randomly determined by each agent attributes. During the
day activities, simulation keeps track of the social
networks of each agent which can be at work, home, or
school.
Table 1: Default values of model parameters
Parameter
Default Value
50
Number of simulated days
0
Random Seed
72798031*
Population Size
Agents
Initial agents
2 days**
Incubation Time (latent)
0.95**
Percentage of immunized agents
0.9**
Percentage of recovered agents
0.14**
Percentage of dead agents
5 day*s*
Min. time to be recovered
14 days**
Max. time to be recovered
Percentage of quarantined agents 0.1
* Egypt census data of 2006 [7].
** WHO [16] values for novel H1N1 pandemic.
Selection of contact agents is random and most likely
results in contacting new agents who are created at
runtime. Newly created agents are initially susceptible,
and are placed randomly across the landscape. Agent
social types are drawn from normal distribution based on
the population structure of census data of Egypt 2006.
Agent health state and disease clock are changed
according to the proposed extension to SIR model.
Infected agent will affect all agents in his social networks
to be exposed. Thus, probability of agent to be infected
increases according to the number of infected agents in his
social networks.
5. Proposed model validation
It is often very difficult to validate epidemiological
simulation models due to the lack of reliable field data,
and the lack of real geographical location of the individual
cases occurred. We have to validate the proposed multi-
agent model against other available models that have been
validated such as SIR model [14]. SIR Model has a long
history and has proved to be a plausible model for real
epidemics. The proposed model should be aligned with
Table 2. Age distribution of Egypt
Age (years) Percentage Possible social types
SIBLING, CHILD, OTHER
10.60 %
< 4
SIBLING, CHILD, OTHERFAMILY, COWORKER, GROUPMEMBER, NEIGHBOR,
05–14
21.10 %
FRIEND, SCHOOLMATE, OTHER
SPOUSE, PARENT, SIBLING, OTHERFAMILY, COWORKER, GROUPMEMBER,
NEIGHBOR, FRIEND, ADVISOR, SCHOOLMATE, OTHER
SPOUSE, PARENT, SIBLING, OTHERFAMILY, COWORKER, GROUPMEMBER,
NEIGHBOR, FRIEND, ADVISOR, OTHER
SPOUSE, PARENT, SIBLING, OTHERFAMILY, GROUPMEMBER, NEIGHBOR,
FRIEND, OTHER
49.85 %
15–44
45–59
12.36 %
> 59
6.08 %
Figure 4. Snapshot displays agents with different health states moving and contacting each others.
the SIR model at least for some simplified scenarios. In
order to align the proposed model to SIR model, we have
evaluated SIR model using Mathematica [11] based on
given parameters (basic reproduction number R0 =3,
duration of Infection = 9.5, initially immunized = 0,
initially infected = 0.01% - see Figure 5) and we have
evaluated the same model parameters in the proposed
multi-agent model (see Figure 6). Two graphs are not a
perfect match, but the proposed multi-agent model graph
match the general behavior of SIR model graph. Two
graphs differ by the magnitude and the smoothness of the
curves. The source of difference of curves behavior is
confined in the following factors: (i) the heterogeneous
structure of the population, (ii) different reproduction
numbers which
are
calculated
for
each
agent
independently, (iii) and the usage of random variable for
infection time instead of deterministic values in SIR
model.
Figure 5. Mathematica SIR Model. Parameters:
R0=3, Duration of Infection=9.5, initially
immunized = 0, initially infected = 0.01%.
population with deployed 50% of vaccination control
strategy, (iv) population with deployed 50% of social
distancing control strategy, and (v) population with
deployed 50% of quarantining control strategy. We
display susceptible, infectious and removed curves of each
scenario to be compared with each other. In the first
scenario with no deployed control strategies, we have
found that epidemic has a steep infection curve which
reaches its peak (608 infected agents – 60.8% of the
population is infected) on day 10 (see figure 7). At the end
of the simulation, we have analyzed the distribution of
health states among social types. Mortality rate is high
among schoolmates, neighbors, and advisors. Numbers of
immunized agents are close for schoolmates, and parents,
while equals zero with child agents (see figure 8).
Figure 7. Scenario 1: pandem ic peak is at day 10.
Figure 8. Scenario 1: distribution of health states
to social types.
We have a pandemic peak at day 10. Thus, we choose
to apply control strategies from day 8 to day 12 in the rest
of the scenarios to study the effect of the control strategy
on the pandemic peak. In the second scenario, we found
that the number of infected agents is reduced during the
deployment period of increase awareness control strategy
to reach it is minimum value of 67 infected agents (6.7 %
of the population) at day 12 (see figure 9.a). This is
Figure 6. Proposed multi-agent model.
Parameters: 3 infected agents – Population Size:
300 agents – Duration: 50 days.
6. Pandemic control strategies
Control strategies are useful for the development of an
action plan to control disease outbreak. Controlling
outbreak is related to the peak of infectious and the time
required reaching the peak. Time required reaching the
peak is helpful for giving time for different control
strategies to be effective. Without a properly planned
strategy, the pandemic chaos might be disastrous causing
large-scale fatalities and substantial economic damage. A
proven control strategy would incorporate increasing
awareness of population, vaccination, social distancing,
and quarantining decisions [16] [6]. The proposed multi-
agent model permits injection of control strategies to
study different scenarios for controlling the outbreak. User
can determine control strategies and
the coverage
percentage applied to the population. Control strategies
will affect the agent health states, and the agent daily
activities. Increasing awareness will increase the number
of doctors' visit and the number of quarantined infected
agents. Vaccination moves susceptible and in-contact
agents
to be
immunized. Social distancing and
quarantining reduce the number of possible contacts
among individual agents. In experiments we will run
different scenarios of applied control strategies.
7. Experiments and analysis of results
The proposed multi-agent model was programmed
using Java programming language and run on AMD
Athlon 64 X2 Dual 2.01 GHz processor with 1GB
memory. To demonstrate the model behavior, we have run
five scenarios of pandemic
in a closed
influenza
population of 1000 agents, and initially three infected
agents. Each run takes about eight minutes to be
completed. Simulated scenarios are: (i) population with no
deployed control strategies, (ii) population with deployed
increasing awareness control strategy, (iii)
50% of
because agents are asking for doctor help when they have
influenza symptoms.
a.
b.
c.
d.
Figure 9. a: Scenario 2, b: Scenario 3,
c: Scenario 4, d: Scenario 5.
In the third scenario, we found that the number of infected
agents is reduced to 320 agents (32% of the population) at
day 9 (see figure 9.b). In the fourth and the fifth scenarios,
we found that the number of infected agents is increased
to 614 (61.4% of the population) at day 10 (see figure 9.c
- d). This means that there are control strategies which not
affect the pandemic spread when applied on given
outbreak duration such as: social distancing and
quarantining through the outbreak peak. Finally, we
conclude that the aggregate attack rate is exponentially
increasing without any deployed control strategies. Attack
rate is controlled by different factors such as: the daily
travelling distance of agents, and the percentage of
vaccinated agents. As a result, proper combination of
deployed control strategies can be effective to decrease
the pandemic damage.
8. Conclusion
The field of computational epidemiology has arisen as
a new branch of epidemiology to understand epidemic
transmission patterns,
and
to help
in planning
precautionary measure. For this reason a spatially explicit
agent-based epidemiologic model of pandemic contagious
disease is proposed in this paper. The methodology for
this paper involves the development of a multi-agent
model of pandemic influenza in Egypt. The proposed
model simulates stochastic propagation of pandemic
influenza outbreaks, and the impact of the decisions made
by the healthcare authorities in population with millions of
agents. We have proposed extension to SIR model, in
which we have investigated the agent attributes. The
model can be easily customized to study the pandemic
spread of any other communicable disease by simply
adjusting the model parameters. We have simulated the
spread of novel H1N1 pandemic in Egypt. Experiments
are run in a closed population of 1000 agents, and initially
3 infected agents. Modeled novel H1N1 reaches infection
peak (608 agents) with in 10 days without deployment of
control strategies. Number of dead agents reaches its peak
at the end of the simulation with mortality of 658 dead
agents. Deployment of proper combination of control
strategies can limit the pandemic chaos and reduce the
fatalities and substantial economic damage. Further work
on proposed model includes: agents with additional
attributes that allow a better realistic model (e.g., ages,
gender, etc), as well as finding optimal combination of
control strategies to manage the pandemic outbreak
waves.
9. References
[1] Bonabeau, Eric. Agent-based modeling: methods and
techniques for simulating human systems. Proc. National
Academy of Sciences 99(3): 7280-7287, 2002.
[2] Ching Fu S, Milne G. Epidemic Modeling Using Cellular
Automata. Proceedings of the ACAL2003: The First
Australian Conference on Artificial Life; Canberra,
Australia, 2003.
[3] Chowell, G., Hyman, J.M., Eubank, S., Castillo-Chavez, C.
Scaling laws for the movement of people between locations
in a large city. Phys. Rev. E 68, 066102, 2003.
[4] Colizza V, Barrat A, Barthélemy M, Vespignani A. The
modeling of global epidemics: stochastic dynamics and
predictability. Bull Math Biol, 68:1893-1921, 2006.
[5] Daley, D. J. & Gani, J. Epidemic Modeling and
Introduction. NY: Cambridge University Press, 2005.
[6] d'Onofrio A, Manfredi P, Salinelli E. Vaccinating behavior,
information, and the dynamics of SIR vaccine preventable
diseases. Th. Pop. Biol.: 301–17, 2007.
[7] Egypt Census Data of 2006, http://www.capmas.gov.eg/.
[8] Eubank, H. Guclu, M. V. Marathe et al. Modeling disease
outbreaks in realistic urban social networks. Nature, vol.
429, no. 6988, pp. 180--184, May 2004.
[9] Hethcote. The Mathematics of Infectious Diseases. SIAM
Review. 42(4):599–653, 2000.
[10] K. M. Carley, D. B. Fridsma, E. Casman et al. Biowar:
Scalable
agent-based model of bioattacks.
IEEE
Transactions on Systems, Man, and Cybernetics, Part A,
vol. 36, no. 2, pp. 252--265, 2006.
[11] Mathematica: http://www.wolfram.com/.
[12] Perez, L., Dragicevic, S. An agent-based approach for
modeling dynamics of contagious disease
spread.
International Journal of Health Geographics, 8:50, 2009.
[13] Quan-Xing Liu, Zhen Jin and Mao-Xing Liu. Spatial
organization and evolution period of the epidemic model
using cellular automata. Phys. Rev. E. (74) 031110, 2006.
[14] Skvortsov, R..B. Connell, P. Dawson, R. Gailis. Epidemic
Modeling: Validation of Agent-based Simulation by Using
Simple Mathematical Models. Proceedings of Land
Warfare Conference, pp 221-227, 2007.
[15] Toomas Timpka, Henrik Eriksson, Elin A Gursky, James
M Nyce, Magnus Morin, Johan Jenvald, Magnus
Strömgren, Einar Holm & Joakim Ekberg. Population-
based simulations of influenza pandemics: validity and
significance for public health policy. Bull World Health
Organ.vol.87, n.4, pp. 305-311. ISSN 0042-9686, 2009.
[16] World Health Organization, http://www.who.int/en/.
|
1805.10906 | 1 | 1805 | 2018-05-28T13:15:52 | Tangramob: an agent-based simulation framework for validating urban smart mobility solutions | [
"cs.MA"
] | Estimating the effects of introducing a range of smart mobility solutions within an urban area is a crucial concern in urban planning. The lack of a Decision Support System (DSS) for the assessment of mobility initiatives, forces local public authorities and mobility service providers to base their decisions on guidelines derived from common heuristics and best practices. These approaches can help planners in shaping mobility solutions, but given the high number of variables to consider the effects are not guaranteed. Therefore, a solution conceived respecting the available guidelines can result in a failure in a different context. In particular, difficult aspects to consider are the interactions between different mobility services available in a given urban area, and the acceptance of a given mobility initiative by the inhabitants of the area. In order to fill this gap, we introduce Tangramob, an agent-based simulation framework capable of assessing the impacts of a Smart Mobility Initiative (SMI) within an urban area of interest. Tangramob simulates how urban traffic is expected to evolve as citizens start experiencing the newly offered traveling solutions. This allows decision makers to evaluate the efficacy of their initiatives taking into account the current urban system. In this paper we provide an overview of the simulation framework along with its design. To show the potential of Tangramob, 3 mobility initiatives are simulated and compared on the same scenario. This shows how it is possible to perform comparative experiments so as to align mobility initiatives to the user goals. | cs.MA | cs | INTERNAL VERSION FOR ARXIV.ORG
1
Tangramob: an agent-based simulation framework
for validating urban smart mobility solutions
Carlo Castagnari, Flavio Corradini, Francesco De Angelis, Jacopo de Berardinis, Giorgio Forcina, Andrea Polini
School of Science and Tecnology, University of Camerino, via Madonna delle Carceri 9, Camerino MC, Italy
8
1
0
2
y
a
M
8
2
]
A
M
.
s
c
[
1
v
6
0
9
0
1
.
5
0
8
1
:
v
i
X
r
a
Abstract-Estimating the effects of introducing a range of
smart mobility solutions within an urban area is a crucial
concern in urban planning. The lack of a Decision Support
System (DSS) for the assessment of mobility initiatives, forces
local public authorities and mobility service providers to base
their decisions on guidelines derived from common heuristics and
best practices. These approaches can help planners in shaping
mobility solutions, but given the high number of variables to
consider the effects are not guaranteed. Therefore, a solution
conceived respecting the available guidelines can result in a
failure in a different context. In particular, difficult aspects to
consider are the interactions between different mobility services
available in a given urban area, and the acceptance of a given
mobility initiative by the inhabitants of the area.
In order to fill this gap, we introduce Tangramob, an agent-
based simulation framework capable of assessing the impacts
of a Smart Mobility Initiative (SMI) within an urban area of
interest. Tangramob simulates how urban traffic is expected to
evolve as citizens start experiencing the newly offered traveling
solutions. This allows decision makers to evaluate the efficacy of
their initiatives taking into account the current urban system. In
this paper we provide an overview of the simulation framework
along with its design. To show the potential of Tangramob, 3
mobility initiatives are simulated and compared on the same
scenario. This shows how it is possible to perform comparative
experiments so as to align mobility initiatives to the user goals.
Index Terms-Smart City, Smart Mobility, Agent-Based Traffic
Simulations, Reinforcement Learning, Smart Urban Planning
I. INTRODUCTION
According to the United Nations [1],
in 2016, world's
population was 7.4 billions inhabitants and about 54.5% of
them lived in urban areas. Despite all the benefits histori-
cally brought by urbanization, like poverty reduction, longer
life expectancy and economic wealth, such an uncontrolled
demographic growth is pushing cities to deal with several
management problems. In particular, focusing on urban mo-
bility,
infrastructures are close to saturation and
this comes with a bunch of problems like car dependence,
spatial footprint, traffic congestion, air and noise pollution.
Novel smart mobility solutions need to be introduced, and
investments have to be carefully assessed in relation to their
effective potential to improve the mobility ecosystem.
transport
Novel mobility initiatives are generally shaped, and their
adoption assessed, considering common guidelines and best
practices. Nevertheless it is not seldom the case that the ob-
served effects, after the concrete deployment of a solution, are
not satisfactory. In particular, there are two complex aspects
Manuscript under revision.
Corresponding author: F. De Angelis ([email protected]).
that are difficult to assess when following such approaches
to planning. The first one relates to how the new mobility
solution will interact with the already available ones, whereas
the second one relates to citizens acceptance. Indeed as many
articles report ([2, 3, 4, 5, 6, 7, 8]), there are many cases
in which the adoption of a smart mobility initiative did not
bring the expected benefits. In particular, many real scenarios
from Europe ([5, 7, 2]), China ([6]) and U.S.A ([3, 4, 8])
demonstrated how proposed smart mobility services failed,
since they were not accepted by communities. As argued by
[2], the carsharing failure in London is attributed to a bad
service configuration. Similarly, [3, 4, 5] argue how initially
promising bikesharing solutions for the city of Salerno (IT)
and Seattle (USA) have not been adopted by the population.
The lack of a formal way to estimate the impacts of a range of
smart mobility services and their interactions is also reported
in [9] and [10], remarking both the importance, and the actual
absence of a "common framework" for this purpose. Indeed,
decision makers are in urgent need of innovative approaches
providing quantitative forecasts in relation to the different
aspects connected with the introduction of a novel mobility
initiative. Resulting Decision Support Systems (DSS) will
complement already available approaches in the definition and
shaping of the smart mobility solution to adopt.
This considerations motivated us in developing a novel
DSS named Tangramob. This is an agent-based simulation
framework capable of assessing the impacts of the intro-
duction of a novel smart mobility initiative (i.e. a range of
either homogeneous or heterogeneous smart mobility services)
within an urban area of interest taking into account the current
mobility ecosystem, as well as salient features of citizens in
relation to the usage of mobility services. Indeed, agent-based
approaches are considered effective for searching a solution
within huge state spaces when the domain to represent can be
easily conceived as a composition of heterogeneous entities
interacting in a distributed setting [11, 12]. This is certainly the
case of a mobility ecosystem in which many different entities
can be identified, each one with its specific characteristics
(e.g. commuters, transport means, roads, etc.), and the system
behaviour and its features emerge from the interactions among
such entities. Tangramob will simulate one day of mobility
starting from a description of the population of interest and
of the mobility resources available in the considered area,
including the ones related to the initiative to evaluate. The
day will be simulated many times over many iterations to
derive a final configuration and the corresponding output. The
iterations are needed since learning mechanisms are applied in
INTERNAL VERSION FOR ARXIV.ORG
2
Tangramob to let commuter agents try out the different avail-
able mobility solutions. After each iteration each commuter
will provide a score for the travel experience according to its
own profile, and taking into account quantitative parameters,
such as travel times. In this way, a commuter agent will
learn which are the transportation solutions that better fit
its profile. Notably, the possibility to change transportation
means in relation to the different segments of a travel allows
intermodality and multimodality within a simulation. Clearly,
the more the provided data input are effectively representative
of the reality of interest, the more the returned data set will
approximate the possible effects introduced by the mobility
solution under evaluation. From the output data it will be
possible to derive quantitative analysis in relation to changes
in emissions, costs etc, as detailed in the following sections.
Summarizing, Tangramob is a DSS that helps decision
makers in planning SMIs. The DSS is distinguishable from
other proposals in relation to two main aspects: (i) it supports
the simulation of intermodal and multimodal transport ser-
vices; and (ii) it makes it possible to reflect the diversities of
commuters with respect to their personal characteristics (e.g.
gender, age, travel demand).
The simulator is built over MATSim, a powerful traffic
simulator [13]. Tangramob is aimed at all people involved in
defining and planning new mobility services: urban planners,
who are in charge of improving urban mobility; transport com-
panies, which need to ponder their investments; researchers,
aiming at testing and validating new mobility solutions.
The rest of the paper is organized as follows: Section II
outlines the idea behind the Tangramob simulator and how it
is expected to address the research problem. In section III we
provide an overview of the agent-based model of Tangramob
and section IV describes its architecture. Finally, section V
proves the effectiveness and the potentialities of Tangramob
by reporting an example of use. Section VI shows the current
attempts in supporting urban planners and mobility service
providers in urban mobility planning. and section VII reports
some conclusions and opportunities for future work.
II. THE TANGRAMOB SIMULATOR
Tangramob is an agent-based simulation framework that
intends to support public and private decision makers in the
task of shaping smart mobility initiatives for a specific urban
area of interest. It can be considered as a Decision Support
System (DSS) for smart mobility validation, focusing on the
ability to capture and reproduce the mobility behaviour of each
single commuter belonging to the selected sample population.
For this purpose, Tangramob is organized as a simulation
environment that the urban planner can easily use in order
to understand if introducing a smart mobility initiative, i.e.
a collection of mobility services, can improve the traveling
experience of citizens as well as the performance of the
urban transport system. Since the simulator is based on an
Agent-Based Model (ABM), for each person in the sample
population, represented as an autonomous reasoning agent,
we can observe whether or not it will make use of the new
mobility services. These fine-grained results also provide users
with a measure concerning the expected adoption rate of the
simulated mobility initiative, so as to figure out beforehand if
the initiative can potentially succeed or not.
Technically, a Tangramob simulation requires four inputs:
• the urban road network of the area under study,
• a representative population of the area with the mobility
agendas of people. An agenda summarizes what a person
does during an ordinary working day (i.e. activities) and
how he moves from one place to the next one (i.e. legs).
• the description of the mobility services already offered
by the city: public transport timetable, etc.;
• the smart mobility initiative to evaluate, that is a list of
geographically located containers (called tangrhubs) of
one or more smart mobility services. Each smart mobility
service belongs to a tangrhub and it comes with a number
of mobility resources (e.g. vehicles), as well as a service
charge (i.e. cost per km and cost per hour).
It is worth mentioning that the definition of an agent popula-
tion is certainly the most complex and critical information to
supply, in particular in relation to profiles, and details on daily
travels. Obviously, the more the population is representative
of the reality of interest the more the results of the simulation
can be considered a good approximation. Strategies for the
derivation of a population are out of scope for this paper,
nevertheless different sources are available to define a repre-
sentative synthetic population. Relevant data can be certainly
collected from periodic census or questionnaires distributed to
a sample population. Particularly effective nowadays is Mobile
Crowd Sensing (MCS) [14] that uses mobile apps developed
for large scale sensing, and involve the contribution from a
crowd of people that behave as probes of the dynamics of the
mobility in the city [15]. GPS data produced by the crowd are
an excellent source of planning and transport information, and
they are widely used in mobility project (e.g., the community
based GPS navigator Waze (www.waze.com) that tracks users
to understand roadway congestion). Activity recognition of
travel demand models can also be derived using Input-Output
Hidden Markov Models (IOHMMs) to infer travelers' activity
patterns from call detail records as suggested in [16].
Starting from the provided information, the execution of
Tangramob can be thought of as performing a comparative
experiment. The experiment consists in introducing the smart
mobility initiative (i.e. applying the treatment) into the urban
area of interest (i.e. the treated system) while observing the
same reality as it is today, namely with no smart mobility
initiative (i.e. control system). In the end, we can observe how
these systems differ with respect to the following measures:
• travel distance, expressed in metres and referred to the
distance traveled by each commuter;
• travel time, expressed in seconds and referred to the time
spent traveling for each commuter;
• CO2 emissions, expressed in grams and referred to the
quantity of CO2 produced by each commuter according
to the used means of transport;
• cost of mobility, expressed in euros and referred to the
cost of mobility for each commuter;
• urban traffic levels, expressed as the number of traveling
INTERNAL VERSION FOR ARXIV.ORG
3
vehicles on each road at a given moment in time band.
This statistic is a picture of the road infrastructure under
study and it is useful when one needs to understand which
are the most congested roads within a time slot.
Such a comparison would allow the user to understand if the
proposed mobility initiative is in line with their expectations.
In case they are not satisfied with the achieved results, the user
can change the configuration of the mobility initiative (e.g.
relocating tangrhubs, adding/removing one or more tangrhubs,
modifying the parameters of a mobility service and so forth)
in order to repeat the experiment as before.
A. The tangrhub
In Tangramob,
the actual placement of smart mobility
services within the urban area under study is made possible
by tangrhubs. A tangrhub can be defined as a geo-located
entity providing citizens with one or more mobility services.
A tangrhub collects one or more smart mobility services, each
of which is offered by either private or public providers. For
instance, a carsharing service provided by two different com-
panies, results in two different characterizations of resources
and their usage deployed within the tangrhubs of interest.
Considering the typical urban conformation, such a flexible
and modular abstraction allows urban planners to represent
all the existing transport facilities like railway stations, bus
stops and so forth, and to introduce intermodality among the
mobility services. Indeed, a bus stop could be represented as
a tangrhub where only the bus service is available.
Examples of smart mobility services that the user can add
to a tangrhub are: dynamic public transport, shared transport
services (e.g. carsharing, bikesharing), dynamic ridesharing,
autonomous taxis and so forth [17]. However, each smart
mobility service mi provided by a tangrhub th j must belong
to only one of the following service types:
• intra-hub services, used for moving people to and from
th j thereby serving first mile trips, e.g. from a commuter's
home-place (departure) to th j, as well as last mile trips,
e.g. from th j to a commuter's workplace (destination).
• inter-hub services, for moving commuters from th j to
another tangrhub thk supporting the service type of mi.
From the simulator's perspective, we can think of a tan-
grhub as an entity with which people interact every time they
need to travel. As a result of such interactions, tangrhubs are
expected to collaborate with each other in order to provide
commuters with a list of valid traveling solutions. Thus, it is
up to each person to evaluate and choose the most suitable
solution according to their own needs and preferences.
B. Smart Mobility Initiative (SMI)
According to the concept of tangrhub seen before, shaping
a Smart Mobility Initiative (SMI) is about placing a number
of tangrhubs within the urban area of interest, adding one or
more smart mobility services to each of them, and providing a
specific characterization for the added mobility services. Thus,
it is up to the user (e.g. an urban planner) to design a list
of candidate SMIs according to the goals and the available
financial resources of his local authority.
To define a smart mobility service for a tangrhub, such
as carsharing, the user has to specify the service type (i.e.
intra-hub or inter-hub), the initial number of vehicles and the
service charge (i.e. cost per km, per hour, and fixed), the
CO2 footprint, and other parameters depending of the type of
vehicles. Therefore, in Tangramob, a mobility service provided
by an organization is represented as a whole as the sum of all
the services made available by the same organization within
the selected tangrhubs. It is worth noticing that the cost of a
mobility service does not need to correspond to a real currency.
In fact, we can consider cost in terms of "points" since such
an approach fits the idea of mobility as a service [18, 19].
These cost-related parameters are expected to affect the
mobility decisions of commuters. More precisely, commuters
are more inclined to choose the most convenient services,
i.e. the ones with the greatest efficiency/cost trade-off. Thus,
leveraging the cost of mobility services allows urban planners
to achieve a mobility policy, thereby promoting some services
against others. For instance, a cheap bikesharing service would
hopefully be more preferable for commuters than an expensive
carsharing service in case of short trips.
C. Tangramob commuting patterns
Fig. 1: 3-trip path
Fig. 2: 2-trip path I
Fig. 3: 2-trip path II
Fig. 4: Direct path
As new mobility opportunities are introduced, commuters
are expected to change their daily commuting patterns. A
commuting pattern is the intermodal representation of how
a person moves from one place to another. Such a trip
can be either simple (e.g. by car) or more complex (e.g.
by a combination of travel modes). A clear example of a
commuting pattern is a route provided by Google Maps.
In Tangramob, the complexity of commuting patterns can be
limited thanks to the direct interconnection of tangrhubs via
their inter-hub mobility services. Commuters are never offered
with traveling solutions made up of more than three sub-trips.
Letting T Ho, T Hd be tangrhubs respectively close to origin
o and destination d of a trip, we can group all the possible
commuting patterns in four classes. In the first class (Figure
1) a person p is expected to reach T Ho either by walk or using
an intra-hub mobility service provided by the same tangrhub;
once arrived at T Ho, an inter-hub mobility service will bring
p at T Hd; finally, the commuter will be able to reach his
destination d either by walk or using an intra-hub mobility
service offered by T Hd. Analogously, the second and the third
classes (Figures 2 and 3) represent a combination of two modal
trips performed either by intra-hub services or by personal
traveling modes. Finally, the last class (Figures 4) corresponds
either to a direct trip (e.g. by car, bike, walk), or to the case
a single inter-hub service is used.
INTERNAL VERSION FOR ARXIV.ORG
4
III. TANGRAMOB AGENT-BASED MODEL OVERVIEW
Starting from the idea of Tangramob, we present the Agent-
Based Model (ABM) on which it is conceived. The Tangramob
ABM is composed of two agent types: commuter and tan-
grhub. A commuter agent is the computational representation
of a single person that
is part of the sample population
under study. Every commuter agent comes with some relevant
personal characteristics, like gender and age, affecting the
outcomes of the actions taken during the simulation. These
effects also impact on the behavior of commuters. For instance,
an elderly person will be less prone to travel by bicycle for
long trips, since this would take too long for him. More
importantly, every commuter has a personal mobility agenda,
i.e. a sequence of daily activities (e.g. home, work, etc.)
interleaved by mobility segments that tells how the agent
manages to get from one activity location to the next one.
On the other hand, a tangrhub agent can be defined as a
local mobility service provider with the ability to improve its
services as the simulation iterates; in the real-world, this active
behaviour might corresponds to a daily enhancement.
Both agents live and operate, albeit with different per-
ceptions, in a composite environment that is made of three
different spaces: the temporal space, the geographical space
and the smart mobility services' state space. Specifically, the
temporal space reflects the passage of time in seconds. The ge-
ographical space can be defined as the directed weighted graph
resulting from the road network infrastructure of the urban
area under study; in particular, nodes represent intersections
and edges denote streets. Such a space is the actual core of
the transport simulation, since the physical limitations of the
road infrastructure can create bottlenecks and delays as people
move with a certain pace. Finally, the last sub-environment is
meant to represent the status of all the smart mobility services
which are currently provided by tangrhubs. This space can be
conceived as a tuple space in which the status of each mobility
service breaks down into a number of smaller sub-states. For
instance, the status of a carsharing service can be expressed
as the combination of the states of all its vehicles.
This complex environment allows agents to perform actions
that can eventually alter the state of affairs of one or more
sub-environments. In particular, as depicted in Figure 5, every
time a commuter needs to move from one place to another,
an interaction with the surrounding tangrhubs takes place.
During this interaction, a smart mobility negotiation occurs:
the tangrhubs collaborate with each other in order to provide
the commuter with a number of traveling alternatives. A
traveling alternative can be thought of as a combination of
one or more (up to three) mobility segments, each of which
can involve a smart mobility service and it is based on the
Tangramob commuting patterns seen in section II-C. Next, the
commuter agent will perform a decision-making process so as
to select the traveling alternative that is expected to maximize
his performance criteria.
The alternative selection process is organized as follows:
first, every single travelling alternative is evaluated according
to the expected performance of each segment it is made of;
is introduced to influence such preference-
then,
the cost
ordered rank; finally, a travelling alternative is selected and
then simulated. Once the commuter agent has reached his final
destination, he is expected to assign a score to every single
commuted mobility segment to record its traveling experience
so as to make more informed decisions for the next iterations.
As soon as a traveling alternative is chosen, the involved
tangrhubs will reserve the required mobility services so that
the commuter can start his journey. Finally, once the commuter
has reached his destination, he will be asked to leave a
feedback for each smart mobility service used.
Fig. 5: Commuter-Tangrhub interaction loop
The behaviour of a commuter revolves around four actions:
(i) synchronizing his mobility agenda with the closest tan-
grhubs, to obtain a list of traveling alternatives for reaching the
location of the forthcoming activity; (ii) choosing a traveling
alternative out of the proposed ones; (iii) performing the
chosen traveling alternative; and finally (iv) leaving as many
feedback as the number of mobility services used in the course
of the day. A commuter will then try to maximize his/her
traveling experience minimizing travel time, covered distance,
emissions and the cost of mobility. More precisely, this is
done by selecting the traveling alternative that is expected to
optimize such criteria from time to time.
On the other hand, the tangrhub agent has the following two
goals: to maximize the traveling experience of commuters and
minimize the number of mobility resources for each service.
Thus, in order to achieve these objectives, a tangrhub can
perform the following actions: (i) building a list of traveling
alternatives in collaboration with other tangrhubs, (ii) provide
a commuter with a list of valid traveling alternatives, (iii)
update the status of its own mobility services, (iv) improve
and optimize its own mobility services.
The tangrhub's service adaptation process is made possible
by commuters feedback. In particular, each feedback qualifies
the traveling experience of a commuter using a specific mo-
bility service. Collecting and averaging all the feedback of a
mobility service can give a metric concerning the performance
of that service, thereby contributing to its improvement and
optimization. For instance, if all the daily-collected carsharing
feedback are negative, a tangrhub would have a valid indicator
of such an inefficiency to run for cover. Therefore, the purpose
of a feedback is twofold: on one hand, it pushes the commuter
agent to reason about the quality of the mobility services to
make more informed decisions for the next iterations; on the
INTERNAL VERSION FOR ARXIV.ORG
5
other hand, it enables tangrhubs to align to the actual mobility
needs of the population.
Tangramob simulations are thus iterative: each iteration
corresponds to a typical day in which commuters experiment
with the introduced smart mobility services in order to record
their performance, while tangrhubs can improve their services
iteration by iteration. This time-evolving behaviour, driven by
feedback, enables commuters to make more informed deci-
sions every time they are offered a list of traveling alternatives.
Therefore, commuters are modeled as proactive agents since
there is need of an iteration-persistent memory structure, i.e.
a knowledge base, to implement such an experience-based
learning capability. With that idea, the decision-making pro-
cess of commuters exploits their personal knowledge base in
order to evaluate the expected score of a traveling alternative,
thanks to the experience accumulated from past iterations. This
is achieved by updating the knowledge base, by means of
a Hebbian-like learning function. This will permit to grad-
ually accumulate scores so as to let the commuter maturate
an experience-based perception for every segment. Similarly,
tangrhubs are modeled as self-adaptive agents which can use
different strategies and optimization methodologies to enable
their travel improving behavior at the end of each iteration and
by means of feedbacks.
IV. DESIGN & IMPLEMENTATION OVERVIEW
Considering the Tangramob agent-based nature, the frame-
work has been developed on an already validated and robust
agent-based traffic simulator named MATSim [13] (Multi
Agent Transport Simulation). Such a design choice is due to
the fact that it is possible to represent the characterizations and
the behaviors of both our agent types in MATSim. Moreover,
such a simulator can be adapted to support all
the sub-
environments of the model, allowing Tangramob to evaluate
the performance criteria as outcomes from the interactions
among such spaces and agents.
A. Multi Agent Transport Simulation: MATSim
MATSim [13] is an activity-based multi-agent simulation
framework for implementing large-scale agent-based transport
scenarios. It is an open-source project implemented in Java
under the GNU public license. As in Figure 6, the framework
consists of several modules which can be combined, used
standalone, or replaced by own implementations.
Fig. 6: MATSim modules
MATSim is designed to model a single day and it
is
based on a co-evolutionary approach in order to reproduce
real-life scenarios. Every agent repeatedly optimizes its daily
activity schedule while in competition for space-time slots
with all other agents on the transportation infrastructure. This
optimization follows an iterative process and it is based on
different choice dimensions such as route selection,
time
choice and mode choice. A MATSim run consists of a number
of iterations in which the steps in Figure 6 are repeated in a
cyclical manner. The initial demand arises from the study of
the area to simulate, and it comprehends its topology (i.e. the
network), the mobility habits of its inhabitants and their per-
sonal features. Every citizen possesses a memory containing a
fixed number of daily plans, each of which is composed of a
daily activity chain and an associated score, i.e. the utility of
that plan. Once the features of the scenario's components are
acquired, and before the MATSim mobility simulation, every
agent selects a plan from its memory according to the score
associated with each of them: higher score plans are more
likely to be selected.
Afterwards, the selected plans are executed by means of the
mobility simulation module; the latter relies on the concept
of queue simulation, which was demonstrated to efficiently
approximate real-life traffic flows. In particular, MATSim
models roads as FIFO queues with a limited vehicle capacity;
every time a vehicle asks permission to access a road, the
corresponding road agent can either respond positively, in case
there still is space, or negatively, if the queue capacity is
reached. Thus, in case a commuter is not allowed to enter a
road segment a delay is produced in the system as long as the
regular flow is restored. Finally, when a commuter manages
to enter a road, he is added to the queue tail until he reaches
its head in order to move to the next segment.
The score computation is made after every mobsim run,
and it is performed on the last executed plans by means of a
scoring function. The score represents a measure about how
the traveling choices made by agents affected the execution of
their activities: the higher the score the better the day.
Once all plans have been scored, the replanning phase takes
place, as a portion of the population is allowed to modify
their plans, in accordance with the co-evolutionary approach.
Then, such plans are modified by applying a mutation operator
according to the previously mentioned choice dimensions.
Finally, in the last iteration the replanning phase is not
executed anymore and it is replaced with the analysis module.
In fact, MATSim is strongly based on events stemming from
the mobsim and this allows to record every action in the
simulation for further analysis. These events' records can be
aggregated to evaluate any measure at the desired resolution.
MATSim can be applied in large scenarios. We show an
example considering a small city in the paper, nevertheless
scalability to bigger cities should not be much a problem
since MATsim simulations of large-scale agent-based micro
simulation models are proved scalable [20]. An experiment
made by MATSim developers with 1.62 million agents and
163K links in the area of Zurich city were simulated in about
20 minutes in a machine with 128GB RAM and 8 dual-
core AMD Opteron CPUs. Also the Switzerland traffic was
modeled in about 3 hours for a single MATSim iteration: one
million roads and 7.3 million agents clearly show that large-
scale, multi-agent micro-simulation can reasonably be used.
INTERNAL VERSION FOR ARXIV.ORG
6
B. Tangramob meets MATSim
Our framework has been implemented on top of MATSim,
taking advantage of its flexible and modular architecture and
trying to maintain the same design principles. We redefined
and extended the behavior of some original MATSim modules,
whereas other remarkable contributions were introduced in
such a way to capture all the features of the Tangramob AB
model of Section III.
Fig. 7: Tangramob architecture
In particular,
the module initial demand,
in which the
simulation input data are collected and validated, is integrated
with the specification of the SMI, describing the locations of
the tangrhubs on the map as well as the list of the mobility
services available on each of them. Making this integration
possible required us to implement the concepts of tangrhub
agent, a new static but active entity that is responsible both
for managing and offering new traveling opportunities to the
nearby population, and for managing the associated mobility
services, which can be seen as services provided by private or
public companies/organizations and that overall constitute the
infrastructure of the SMIs available in the urban area.
The mobsim module, specialized in simulating the urban
traffic, has been integrated with the MATSim "multimodal"
extension, that allows to deal with different transport modes
as well as simulating the overtaking of vehicles. This way,
Tangramob can also evaluate the impact on the urban system
caused by unconventional kinds of vehicles (e.g. scooters,
bicycles, etc.). For this purpose, we redesigned the original
concept of MATSim's vehicles, and we introduced the char-
acterization of mobility services with the ability to manage
such vehicles. Furthermore, our characterization takes into
account the most relevant vehicles features like: dimension,
velocity, fuel type and consumption; all these specifications
are expected to impact on the traffic simulation, especially for
what concerns travel delays and times, and thus are relevant
information in relation to the mobility decisions of commuters.
Concerning the scoring module, Tangramob still exploits the
original Charypar-Nagel scoring function [13]. This allowed us
to validate the new learning process of Tangramob, exploiting
the existing MATSim's validation work.
The replanning phase designed for Tangramob is completely
different from that followed by MATSim. Whereas MATSim
adopts a co-evolutionary algorithm, our framework is based on
a reinforcement learning approach, allowing each commuter
to evaluate his past traveling experience in order to improve
their daily personal mobility. This is made possible by the
implementation of iteration-persistent memory structures, that
every commuter can exploit as knowledge base, in order to
accumulate the score given for each mobility service used
during the simulation. Thus, the score of a service acts as
a reward for the action of choosing that service for a certain
trip. Such a different approach allows commuters to maximize
the expected utility of their mobility decisions. In particular,
during the last iteration of the simulation, each commuter will
decide to either use the new mobility services, or not to accept
the mobility initiative, according to the collected knowledge.
Finally, the analysis module has been integrated with new
to compare
statistical collectors to gather all data useful
the legacy urban mobility with the one emerging after the
introduction of a SMI. Some stats correspond to the agents'
performance criteria described in section III, and others are
focused on the urban system as a whole. In particular, we
aim at collecting the following statistical data: (i) urban traffic
levels, (ii) CO2 emissions, (iii) traveled distances, (iv) travel
times, (v) land use levels, (vi) cost of mobility, (vii) number
of adopters, and (viii) resource usage level.
As depicted in Figure 7, all these redefined modules form
the core of Tangramob, sitting on the top of MATSim and
some of its well-known extensions. Moreover, to make the
simulator more accessible and user-friendly, we have designed
and implemented the following features:
• a census data converter, namely a tool for translating
Italian census data into suitable input mobility agendas;
• a population generator, for synthesizing a sample popu-
lation from some statistical data about an urban area for
which neither census nor plans are available;
• tangrhub aided placement, a tool that analyzes the loca-
tions of people daily activities in order to spot the most
populated urban areas. By using clustering algorithms,
this tool supports users in the task of placing tangrhubs
in a more rational way, since keeping them close to the
population is expected to minimize micro-mobility.
• a web-based Graphical User Interface (GUI), which
allows users to select the geographical area of interest
to automatically retrieve the urban road network of the
selected area from OpenStreetMap. In case the user
cannot generate the data of the sample population under
study from the census data converter, the user is expected
to load the mobility agendas manually or to generate
a synthetic but realistic population from other sources.
Finally, once the geographical context is defined, the user
can shape a smart mobility initiative by geographically
placing a number of tangrhubs and configuring each of
them with one or more mobility services.
The resulting architecture is fully extensible in every layer,
providing the possibility to develop extensions over both the
MATSim layer and the Tangramob codebase.
V. TANGRAMOB: AN EXAMPLE OF USE
In order to show an example of use of Tangramob, we report
some experiments performed on a real scenario in the city of
INTERNAL VERSION FOR ARXIV.ORG
7
Ascoli Piceno (Italy). Ascoli Piceno is a small city with about
50.000 inhabitants over 158 km2 and several other thousand
of people who live in near places outside the city perimeter.
the city center can be served by the same hub. Each hub is
characterized as in Table II, for readability we use the same
resources for each tangrhub in this example.
As shown in Table II, each SMI shares the same number
of tangrhubs, each of which is provided with the same choice
set of mobility services. Even the geographical location of
tangrhubs is the same for all the initiatives, and it is denoted
by the triangles depicted in Figure 8. For each tangrhub we
specify the dimension of the fleet the hub manage at the start
of the simulation, its total capacity to store vehicles is set
a 25% more than the initial fleet. What differs among these
initiatives is just the number of mobility resources, which in
this case, correspond to the vehicle fleet of each service.
TH
Total
Mobility Services
SMI-1
bikesharing
carsharing
e-scootersharing
bikesharing
carsharing
e-scootersharing
10
–
–
110
–
–
Fleet
SMI-2
SMI-3
10
10
10
110
110
110
50
50
50
550
550
550
TABLE II: Grid network: tangrhubs experimental setup
For each mobility service we specify costs. For this experi-
ment, we set the costs of the chosen mobility services, which
in turn were set according to the actual average service charges
in Europe, as summarized in Table III.
Cost per h
Cost per km Fixed Cost
Bikesharing
Carsharing
e-scootersharing
0.5 e
13 e
2,5 e
0 e
0.1 e
0.1 e
0.01 e
0.01 e
0.01 e
TABLE III: Grid network: mobility services' costs
As argued in Section II, understanding how the proposed
SMIs impact both commuters and the transport system re-
quires a comparative experiment. In particular, considering that
commuters in this scenario are used to move by private cars,
first we simulate the current urban mobility (i.e. the pre-SMI
simulation), then we simulate each SMI separately (i.e. SMI-1,
SMI-2 and SMI-3 simulations). Afterwards, we compare these
simulations with respect to the following variables: traveled
distance, travel time, CO2 emissions, cost of mobility and
urban traffic levels; all these values are collected during the last
iteration of each simulation. The first 4 variables are intended
as per-capita indicators (averaged values) and summarize the
traveling experience of commuters, whereas the last one can
be seen as a performance measure of the urban system.
A. Experimental results
In this section, we show and discuss the results obtained
from our simulations to compare them according to the
just mentioned indicators. We also provide some interesting
insights concerning the impact of the 3 SMIs on people
acceptance and on mobility resource usage levels.
1) Number of tangrhubs' subscribers: A subscriber is a
person who uses the mobility services provided by tangrhubs.
This value may measure the success of a SMI in terms of
people acceptance. As noticeable in Figure 9, the number of
Fig. 8: Ascoli Piceno network with tangrubs positioning for SMI's
As depicted in Figure 8 the network represents all the city
roads and infrastructures including the city center and the
roads that connect the city with other places. Starting from
available statistics, we identified 15 areas and for each one
we identified inhabitants and jobs as described in Table I.
Areas
P.ta Solest´a
P.ta Romana
Centro
Piazzarola
C. Parignano
P.ta Maggiore
Monticelli
Brecciarolo
People
5009
1839
7740
409
3368
11500
10633
645
Jobs
3170
700
6760
250
2170
10900
8000
1300
Areas
P. di Bretta
Battente
Marino
Villa Pigna
Z. Industriale
C. di Lama
Frazioni
People
1694
103
576
3000
500
3000
6242
Jobs
500
3000
1400
2000
6500
5000
6000
TABLE I: Population and jobs in the city areas
In the experiment, we modeled the whole population con-
sidering the suburbs with 56.000 agents. Using the statistics
of the municipality we have built a normally distributed
population age with a 45% in range 25-49. Female are 52%
and male 48%. These parameters are expected to affect the
act of travelling of commuters, thus impacting on their score.
Basically, mobility agendas has been organized with three
daily activities in the following order: home, work, home.
Thus, a commuter moves from its home to a workplace in
another area and viceversa. For sake of clarity, we consider in
this experiment as work each kind of activity different to stay
at home. We also do not consider multi-trip commutes.
As a typical real-case scenario, peak activity hours can
be split into two different moments: 8:00 a.m. commuters
start moving towards the workplaces; while at 16:00 p.m.
commuters come back home from work. The first activity in
the morning is distributed in the 5:00 a.m. - 13:00 a.m. time
slot with 45% included in the 07:00 a.m. - 08:00 a.m. hour.
The homecoming happens at the end of the work activity. That
time is modeled using a Gaussian distribution centered over 6
hour of duration.
In this scenario, we aim at investigating the impacts of
three different smart mobility initiatives that integrate trans-
port services: a bikeshare, an electric carshare, and an e-
scootersharing service. All vehicles used are zero emissions.
We use 11 tangrhubs in the city areas as several locations in
INTERNAL VERSION FOR ARXIV.ORG
8
Fig. 9: Tangrhubs subscriptions
Fig. 10: Traveled distances comparison
Fig. 11: Travel times comparison
Fig. 12: Emissions comparison
Fig. 13: Mobility costs comparison
Fig. 14: Mobility resources usage
subscribers increase as the SMI has more mobility resources,
where the horizontal line denotes the entire population.
2) Commuters' performance measures: The first variable
involved in the comparison is the traveled distance of com-
muters shown in Figure 10. In the 3 SMI simulations, com-
muters are expected to travel shorter distances than the pre-
SMI ones even if the differences are not so marked. For what
concerns travel times (Figure 11), in the simulations of SMI-
1 and SMI-2 commuters spend less time traveling. This is a
good indicator of the effectiveness of the SMIs. Conversely,
in SMI-3 commuters spend much more time traveling than
before. This indicates that SMI-3 has some problems either in
the configuration or in moving people respect to the pre-SMI.
Concerning the comparison of CO2 emissions produced by
commuters, we found that the carbon footprint of the 3 SMIs
simulations tends to decrease in directly proportional way with
the number of subscribers. Therefore, we can affirm that the
more the SMI satisfies a large section of the population, the
more the simulation becomes eco-friendly if we use green
vehicles. SMI-1, SMI-2 and SMI-3 reduce CO2 respectively
of 20%, 25% and 35%. Besides the environmental impact, we
also found that there exists an inverse relationship between the
number of subscriptions and the daily costs of mobility. As can
be seen in Figure 13, a commuter in the pre-SMI simulation
spends on average e13.5 a day for traveling, whereas a
commuter can satisfy his/her needs with a lower expense of
e10, e9 and e8 in the SMI-1, SMI-2 and SMI-3 respectively.
3) Mobility resources usage: Tangramob can provide stats
concerning the level of mobility resources usage of the SMIs
(Figure 14). The analysis of these data allows to understand if
a SMI is efficiently configured, so as to refine it for obtaining
similar results with fewer mobility resources. In our case, it
turns out that SMI-1 and SMI-2 are properly configured and
they resources are used. SMI-3 have a large number of unused
vehicles, we can reduce its fleets in other simulation attempt.
A closer look to the resource usage of SMI-3 in Figure 15
shows the incorrect sizing for the car and scooter services
highlighting however the right usage of bikes.
Fig. 15: Mobility resource usage SMI-3
Gathering together all the results, we can conclude that
a properly configured SMI helps reducing several urban
problems, like traffic congestion levels and consequently air
pollution. The experiment conclude that SMI-2 shortly reduced
distances by 20% mantaining substantially the same travel
times but lowering significantly emissions and costs. Moreover
its application actually use all the resources associated with
services. The same conclusion can be made for SMI-1 also
if the benefit is less noticeable. SMI-1 and SMI-2 could be
evaluated in relation to their implementation costs by a urban
planner and a decision maker. SMI-3 on the contrary increases
the travel time while maintaining important benefits in travel
distances, emission and costs. However, its implementation
requires many resources, many of which are left unused.
The 3 simulations of this experiment took about two and a
half hours each, with 110 iterations on a Linux machine with
an i7-7700k CPU @ 4.8GHz and 16GB RAM. We used in the
test the whole 56000 agents population. Running a mid-sized
INTERNAL VERSION FOR ARXIV.ORG
9
scenario with a real population of 500K inhabitants should
be done scaling the population to 10% as suggested by [13]
thus such data makes us confident on feasibility of Tangramob
simulations also for scenarios larger.
VI. RELATED WORK
To the best of our knowledge, the state of the art does
not provide any easy-to-use tool for assessing the impacts of
smart mobility initiatives, in particular when several mobility
services are considered. This section thus aims at presenting
and discussing the main relevant studies sharing our intent.
We present simulation-based studies that can be defined as
in-silico experiments in which scientists investigate how the
introduction of a certain smart mobility service could improve
the urban transport system of a chosen area of interest. These
studies are performed by extending traffic simulators, like
MATSim [13] and SUMO [21], and the rigorous nature of
such experiments can guarantee the reliability of their results.
For instance, in [22], Bischoff et al. extended MATSim in
order to investigate a city-wide replacement of private cars
with variously sized Autonomous Taxicab (AT) fleets. Results
showed that a fleet of 100,000 AT vehicles could satisfy all
the inner city trip demand with: an average waiting time for
a vehicle of under three minutes at most times of the day,
and under five minutes during peak hours. In [23], Balac
et al. simulated the introduction of two different carsharing
initiatives: round-based and one-way. The first one allows
users to pick-up a car from a nearby station and return it later
to the same location. The second one is similar to the round-
based type, but it allows one to park the car at the closest
station from the destination. The simulation was performed
by means of a MATSim extension, limiting the scenario to
the centre of Zurich. Results proved that the round-based
carsharing is mostly chosen for shopping and leisure trips,
whilst one-way for commuting ones.
All these studies bring insightful results with regard to
how a transportation system is expected to improve after
the introduction of a smart mobility service. This is made
possible by the scientific nature of such traffic simulations,
as showed by the MATSim team in [13]. However, each of
these studies deals with a single solution, so there still is no
common framework to assess the impacts and performance
of a range of heterogeneous smart mobility services [10] as
well as an holistic and interrelated vision of these actions [9].
Furthermore, each study pertains to a specific geographical
area (e.g. Berlin for AT, Zurich for carsharing) and the
technical organization of the corresponding extension is not
flexible enough to be adapted to other areas. Finally, it is
worth mentioning that only IT-skilled users would be able to
customise such extensions because of their in-code nature.
Other than MATSim and SUMO, traffic simulators like Sim-
Mobility [24], Vissim [25] and SMART (Scalable Microscopic
Adaptive Road Traffic Simulator) [26] do not currently provide
relevant studies on the impacts of smart mobility solutions.
crucial task in the Urban Planning field. Local public author-
ities and mobility service providers currently design mobility
initiatives according to common heuristics and best practices;
these approaches cannot be expected to generalize to every
geographical context due to the complexity and the diversity of
urban systems. Thus, deploying a mobility initiative is the only
way to get measure ex post, but this also comes with potential
risks since a failing initiative would result in a considerable
waste of resources and trust.
To address this problem we introduced Tangramob, an
agent-based simulation framework that allows users to assess
impacts and performances of a mobility initiative within
an urban area of interest. Tangramob performs comparative
experiments between, before, and after the introduction of a
mobility initiative, approximating real-world urban dynamics
by adopting reinforcement learning techniques. The computa-
tional nature of these experiments makes it easy to support
urban mobility decisions permitting to reduce costs and risks.
Tangramob is still under active development and improve-
ment, nonetheless the current version already permitted to
run meaningful experiments that provided positive results
on the usefulness and the potentialities of the simulator. In
particular, users can measure the impacts of a simulated smart
mobility initiative with respect to: urban traffic levels, CO2
emissions, traveled distances, travel times, land use levels, cost
of mobility, number of adopters and resources usage level.
Thus, it is up to the user evaluate which variables are more
relevant for understanding whether or not an initiative is in
line with his objectives. The experiment we shown help urban
planner to consider future initiatives and policies. SMI-1 is the
cost effective solution impacting significantly CO2 emissions
and personal costs. SMI-2 is the most powerful initiative
able to lower again that values while offering a variety of
services to the commuters. SMI-3 is clearly oversized and the
improvements made possible by its use are not justified by
implementation costs and resource unused rate. From these,
planners could refine SMI-1 and SMI-2 to arrive at a simulated
city planning useful to the decision makers.
Planned future work includes the extension of the current
scoring function with additional traveling comfort criteria to
measure the comfort of a traveling experience with a certain
vehicle to let the commuter agents evaluate a mobility service
as a whole. We are also working on additional evaluations,
taking into account more complex scenarios.
REFERENCES
[1] UN, "The world's cities in 2016," New York, 2016.
[2] C. Pitas and D. Goodman. Bmw tries to succeed where
daimler failed with london car-share scheme. [Online].
Available: https://goo.gl/19SsS5
[3] L. Alter.
succeed
others
https://goo.gl/DM3KvQ
(2013) Why do some bike-share systems
and
[Online]. Available:
fail?
VII. CONCLUSIONS AND FUTURE WORK
[4] D. Westneat. (2016) Bike shares failure deflates seattles
self-image. [Online]. Available: https://goo.gl/fCliOa
Understanding how urban mobility is expected to evolve
after the introduction of new smart mobility services is a
[5] E. Negri. (2016) Salerno, un fallimento il servizio di bici
condivise. [Online]. Available: https://goo.gl/RWQNmC
INTERNAL VERSION FOR ARXIV.ORG
10
[6] A. Mamiit. (2016) Why the ride-sharing company failed
to conquer china and what it means for everyone else.
[Online]. Available: https://goo.gl/cHuC9n
[7] O. Sulopuisto.
on-demand bus
https://goo.gl/7FOhzy
(2016) Why helsinki's
service failed.
innovative
[Online]. Available:
[8] K. Fehrenbacher. Another failed attempt to make ride
sharing work in the u.s., ridejoy to shut down. [Online].
Available: https://goo.gl/3ITYce
[9] L. McGrath, "Smart Mobility in Smart City: Action
Taxonomy, ICT Intensity and Public Benefits Clara,"
Lecture Notes in Information Systems and Organisation,
2016.
[10] K. Zavitsas, I. Kaparias, and M. G. H. Bell, "Transport
problems in cities," Transport, 2011.
[11] P. Davidsson, J. A. Persson, and J. Holmgren, "On the
integration of agent-based and mathematical optimization
techniques," in Agent and Multi-Agent Systems: Tech-
nologies and Applications. Berlin, Heidelberg: Springer
Berlin Heidelberg, 2007, pp. 1–10.
[12] M. Barbati, G. Bruno, and A. Genovese, "Applications of
agent-based models for optimization problems: A litera-
ture review," Expert Systems with Applications, vol. 39,
no. 5, pp. 6020–6028, 2012.
[13] A. Horni, K. Nagel, and K. W. Axhausen, The
London:
Multi-Agent Transport Simulation MATSim.
Ubiquity-Press, 2016.
[14] B. Guo, Z. Wang, Z. Yu, Y. Wang, N. Y. Yen,
R. Huang, and X. Zhou, "Mobile crowd sensing and
computing: The review of an emerging human-powered
sensing paradigm," ACM Comput. Surv., vol. 48,
no. 1, pp. 7:1–7:31, Aug. 2015. [Online]. Available:
http://doi.acm.org/10.1145/2794400
[15] H. Ma, D. Zhao, and P. Yuan, "Opportunities in mo-
bile crowd sensing," IEEE Communications Magazine,
vol. 52, no. 8, pp. 29–35, 2014.
[16] M. Yin, M. Sheehan, S. Feygin, J.-F. Paiement, and
A. Pozdnoukhov, "A generative model of urban activities
from cellular data," IEEE Transactions on Intelligent
Transportation Systems, 2017.
[17] M. Ciuffini, C. Aneris, V. Gentili, S. Operto, L. Refrigeri,
and L. Trepiedi, "La sharing mobility in italia: numeri,
fatti e potenzialit´a," Osservatorio Nazionale Sharing Mo-
bility (Italia), Tech. Rep., 2016.
[18] E. M. Finger, N. Bert, and D. Kupfer, "From the Helsinki
experiment to a European model?" FSR Transport, 2015.
[Online]. Available: https://goo.gl/AFzcOI
[19] Catapult Transport Systems, "Exploring the opportunity
for Mobility as a Service in the UK," 2016.
[20] R. A. Waraich, D. Charypar, M. Balmer, and K. W.
Axhausen, Performance Improvements for Large-Scale
Traffic Simulation in MATSim.
Cham: Springer
International Publishing, 2015, pp. 211–233. [Online].
Available: https://doi.org/10.1007/978-3-319-11469-9 9
[21] M. Behrisch, L. Bieker, J. Erdmann, and D. Krajzewicz,
"Sumo–simulation of urban mobility: an overview," in
Proceedings of SIMUL 2011, The Third International
Conference on Advances in System Simulation, 2011.
[22] J. Bischoff and M. Maciejewski, "Autonomous taxicabs
in berlin–a spatiotemporal analysis of service perfor-
mance," Transportation Research Procedia, vol. 19, pp.
176–186, 2016.
[23] M. Balac and F. Ciari, "Modeling station-based car-
sharing in Switzerland," 14th Swiss Transport Research
Conference, no. May, 2014.
[24] M. Adnan, F. C. Pereira, C. M. L. Azevedo, K. Basak,
M. Lovric, S. Raveau, Y. Zhu, J. Ferreira, C. Zegras, and
M. Ben-Akiva, "Simmobility: A multi-scale integrated
agent-based simulation platform," in 95th Annual Meet-
ing of the Transportation Research Board Forthcoming
in Transportation Research Record, 2016.
[25] M. Fellendorf and P. Vortisch, Microscopic Traffic Flow
Simulator VISSIM. Springer New York, 2010.
[26] K. Ramamohanarao, H. Xie, L. Kulik, S. Karunasek-
era, E. Tanin, R. Zhang, and E. B. Khunayn, "Smarts:
Scalable microscopic adaptive road traffic simulator,"
ACM Transactions on Intelligent Systems and Technology
(TIST), vol. 8, no. 2, p. 26, 2017.
|
1110.0105 | 1 | 1110 | 2011-10-01T15:03:52 | Multi-Agent Programming Contest 2011 - The Python-DTU Team | [
"cs.MA"
] | We provide a brief description of the Python-DTU system, including the overall design, the tools and the algorithms that we plan to use in the agent contest. | cs.MA | cs |
Multi-Agent Programming Contest 2011
--
The Python-DTU Team
Jørgen Villadsen, Mikko Berggren Ettienne, and Steen Vester
Department of Informatics and Mathematical Modelling
Technical University of Denmark
Richard Petersens Plads, Building 321, DK-2800 Kongens Lyngby, Denmark
Abstract. We provide a brief description of the Python-DTU system,
including the overall design, the tools and the algorithms that we plan
to use in the agent contest.
Updated 1 October 2011: Appendix with comments on the contest added.
Introduction
1. The name of our team is Python-DTU. We participated in the contest in 2009
and 2010 as the Jason-DTU team [2,3], where we used the Jason platform
[1], but this year we use just the programming language Python.
We intend in a later paper to elaborate on the reasons for abandoning the
Jason platform and its agent-oriented programming language AgentSpeak.
2. The members of the team are as follows:
-- Jørgen Villadsen, PhD
-- Mikko Berggren Ettienne, MSc student (new in the team this year)
-- Steen Vester, MSc student
We are affiliated with DTU Informatics (short for Department of Informatics
and Mathematical Modelling, Technical University of Denmark, and located
in the greater Copenhagen area).
3. The main contact is associate professor Jørgen Villadsen, DTU Informatics,
email: [email protected]
4. We expect that we will have invested approximately 400 man hours when
the tournament starts.
1
System Analysis and Design
1. The competition is built on the Java MASSim platform and the Java EIS-
MASSim framework is distributed with the competition files. This framework
is based on EIS and abstracts the communication between the server and
the agents to simple Java method calls and callbacks.
To utilize this framework we started out with the Java implementation of
Python called Jython which in contrast to Python can import Java libraries
and classes. To support agent communication in our multi-agent system
we have so far used the Apache ActiveMQ as a messaging server which
offers clients for all popular programming languages. Using the EISMASSim
Java framework together with ActiveMQ clients written in Python gluing
it all together with Jython gave some performance issues when exchanging
percepts between the agents.
We found that each component performed well tested in a controlled context
which suggested that the issues were related to the interaction between the
components.
We decided to skip Jython and EISMASSim to instead follow a much cleaner
Python-only implementation. Even though some work was needed to imple-
ment the protocol specific parts which EISMASSim handled, this left us with
a more flexible implementation of which we have complete knowledge and
control of every part of the implementation. It also solved the performance
issues related to component interaction.
We also plan to skip the messaging server and instead let the agents com-
municate directly using a simple and efficient text-based protocol to further
improve the performance of the system. A simple text-based protocol allows
us to keep the distributivity and modularity. Thus it will still be posible
to use agents written in different programming languages. Furthermore, by
implementing our own message server, we can tweak the low-level features
to suit the need of our specific system.
2. We do not use any existing multi-agent system methodology.
3. We do not plan to distribute the agents on several machines.
4. We do not plan a solution with a centralization of coordination/information
on a specific agent. Rather we plan a decentralized solution where agents
share percepts through messages and coordinate actions using distributed
algorithms.
5. The team communication is based on the publisher-subscriber pattern. Our
message server has a number of topics to which agents can subscribe and
publish messages. This supports one-to-one communication and one-to-many
communication in a simple way where the agents determine which topics to
subscribe to.
2
6. To assign goals to agents we use a ring-based auction algorithm. Each round
in an auction includes n messages where n is the number of agents participat-
ing. Before participating in an auction all agents score their top n goals and
use this score to determine their bidding strategy. The algorithm terminates
when all n agents are assigned to a unique goal.
7. Each agent acts on its own behalf based on its local view of the world which
is updated through percepts and is thus autonomous and reactive. This is
implemented as an agent-control-loop in which the agents decide which ac-
tions to execute based on their current view of the world. We have considered
implementing an algorithm to determine the best way to parry a series of
attacks from a saboteur agent which would make the agents proactive.
Software Architecture
1. We implement the multi-agent system using just the programming language
Python.
Even though we all have more experience with Java, we choose Python as
our programming language, as we think it has some advantages over Java,
mainly in development speed/programmer effectiveness. Some of the reasons
being that Python in contrast to Java:
-- is dynamically typed
-- is concise
-- is compact
-- supports multiple programming paradigms (object-oriented, imperative,
functional)
-- is popular for scripting
-- does not need to be compiled before execution
2. We use Python 2.7 on Ubuntu Linux and Mac OS X as the development
platforms and GEdit, Eclipse and TextMate as code editors/IDEs.
3. As the runtime platform for the competition we plan to use a suitable Linux
system with Python 2.7.
4. We plan to use the following algorithms:
-- All-pair shortest path extended to support dynamic vertex addition
-- Custom breadth-first-searches
-- Ring-based agreement algorithm
3
References
1. Rafael H. Bordini, Jomi Fred Hubner, and Michael Wooldridge. Programming Multi-
Agent Systems in AgentSpeak Using Jason. John Wiley & Sons, 2007.
2. Niklas Skamriis Boss, Andreas Schmidt Jensen, and Jørgen Villadsen. Building
Multi-Agent Systems Using Jason. Annals of Mathematics and Artificial Intelli-
gence, Springer Online First 6 May 2010.
3. Steen Vester, Niklas Skamriis Boss, Andreas Schmidt Jensen, and Jørgen Villadsen.
Improving Multi-Agent Systems Using Jason. Annals of Mathematics and Artificial
Intelligence, Springer Online First 5 March 2011.
Acknowledgements
Thanks to Niklas Christoffer Petersen, MSc student, for comments.
More information about the Python-DTU team is available here:
http://www.imm.dtu.dk/~jv/MAS
Appendix
Our idea to skip the message server and implement our own text-based protocol
was slightly modified during the implementation phase. Instead of implementing
our own message system, we reached the pragmatic conclusion to instead let
the agents communicate trough direct reference to a shared data structure. This
allowed us to spend more time on other important issues and freed us from all
performance issues. The main reason for the message system idea was to preserve
modularity and this could easily be implemented at a later state if time allowed.
In the early development phase we suspected our dynamic all-pair shortest
path algorithm as a candidate for our performance issues. However we discovered
that the issues were I/O related, and found that we had plenty of processing time
and power to perform multiple stock graph search algorithms at each simulation
step. This made our dynamic all-pair shortest path algorithm superfluous for
this scenario. However it could be highly relevant in case of less processing time
or bigger graphs.
4
|
0806.3031 | 1 | 0806 | 2008-06-18T15:46:10 | Multi Site Coordination using a Multi-Agent System | [
"cs.MA"
] | A new approach of coordination of decisions in a multi site system is proposed. It is based this approach on a multi-agent concept and on the principle of distributed network of enterprises. For this purpose, each enterprise is defined as autonomous and performs simultaneously at the local and global levels. The basic component of our approach is a so-called Virtual Enterprise Node (VEN), where the enterprise network is represented as a set of tiers (like in a product breakdown structure). Within the network, each partner constitutes a VEN, which is in contact with several customers and suppliers. Exchanges between the VENs ensure the autonomy of decision, and guarantiee the consistency of information and material flows. Only two complementary VEN agents are necessary: one for external interactions, the Negotiator Agent (NA) and one for the planning of internal decisions, the Planner Agent (PA). If supply problems occur in the network, two other agents are defined: the Tier Negotiator Agent (TNA) working at the tier level only and the Supply Chain Mediator Agent (SCMA) working at the level of the enterprise network. These two agents are only active when the perturbation occurs. Otherwise, the VENs process the flow of information alone. With this new approach, managing enterprise network becomes much more transparent and looks like managing a simple enterprise in the network. The use of a Multi-Agent System (MAS) allows physical distribution of the decisional system, and procures a heterarchical organization structure with a decentralized control that guaranties the autonomy of each entity and the flexibility of the network. | cs.MA | cs | COMPUTERS IN
INDUSTRY
Multi Site Coordination using a Multi-Agent
System
Monteiro Thibaud(1), Roy Daniel(2), Anciaux Didier(1)
INRIA – MACSI project, LGIPM – SdP team
(1)Université Paul Verlaine, Île du Saulcy, F-57045 Metz CEDEX 01 – FRANCE
(2)École Nationale d’Ingénieurs de Metz, Île du Saulcy, F-57045 Metz CEDEX 01 – FRANCE
Email: {roy, anciaux, monteiro}@agip.sciences.univ-metz.fr
Abstract
A new approach of coordination of decisions in a multi site system is proposed. It is based this approach on a
multi-agent concept and on the principle of distributed network of enterprises. For this purpose, each enterprise
is defined as autonomous and performs simultaneously at the local and global levels.
The basic component of our approach is a so-called Virtual Enterprise Node (VEN), where the enterprise
network is represented as a set of tiers (like in a product breakdown structure). Within the network, each partner
constitutes a VEN, which is in contact with several customers and suppliers. Exchanges between the VENs
ensure the autonomy of decision, and guarantiee the consistency of information and material flows. Only two
complementary VEN agents are necessary: one for external interactions, the Negotiator Agent (NA) and one for
the planning of internal decisions, the Planner Agent (PA).
If supply problems occur in the network, two other agents are defined: the Tier Negotiator Agent (TNA)
working at the tier level only and the Supply Chain Mediator Agent (SCMA) working at the level of the
enterprise network. These two agents are only active when the perturbation occurs. Otherwise, the VENs process
the flow of information alone.
With this new approach, managing enterprise network becomes much more transparent and looks like
managing a simple enterprise in the network. The use of a Multi-Agent System (MAS) allows physical
distribution of the decisional system, and procures a heterarchical organization structure with a decentralized
control that guaranties the autonomy of each entity and the flexibility of the network.
Keywords: Multi-agent systems, decision support systems, coordination, negotiation, distributed control,
supply chain
1.
Introduction
Product design, manufacture, conditioning, or the
combination of the three do not result from isolated
and autarkical companies but from increasingly
complex corporative networks. Such networks can
take various forms and can be as complex as
clusters of several “Industrial Architectures”.
It is necessary to develop rigorous methods to
improve
the performance of
such complex
architecture development and the efficiency of
these enterprises. In a complex network, enterprise
performance is sensitive to the relationships and
behaviors of participating companies, a dimension
which does not exist for individual companies.
Being a function of product quality and costs
(including the societal cost) of the products, it is
also defined in terms of time to design or to
manufacture,
and depends mainly on
the
information and the material flows.
To improve the reactivity and the costs of a
company involved in such a complex architecture,
it is necessary to consider subcontracting and the
relationships with partners. The “make or buy”
decisions involved in this architecture occur in
various time frames. A long term decision or
strategic decision, corresponds to the whole set of
external and
internal nodes of production,
distribution and supply. A medium term decision or
tactic
decision
characterizes
the
contracts
(quantities of products to deliver, delays, costs,
penalties…) that the company is likely to have with
its internal and/or external providers in order to
carry out a production program that accomplishes
the best balance between cost and delay. Finally, a
short
term decision or operational decision,
considers for instance subcontractors, simply for
overloading capacity in order to absorb a transitory
fluctuation of demand. The paper only focuses on
operational decisions.
There is a need to better coordinate actions in
complex cooperative network, especially among the
partners (Altersohn, 1992; Rota, 1998; Kjenstad,
1998). Recent research shows a growing interest in
studying cooperation relationships among multiple
actors of industrial architectures (Axelrod, 1992;
Rapoport, 1987; Ferrarini, 2001; Monteiro and
Ladet, 2001b). It appears from these studies that
cooperation has two different forms. Cooperation
can be a form of collaboration between partners in
which each has equivalent decisional capacity and
acts with others towards a common objective, such
as in the co-design in the automotive sector
(Womack et al., 1992) or aeronautical sector
(Cauvin et al., 2003). Cooperation can also be a
form of coordination and a synchronization of
operations carried out by
independent actors
(Malone, and Crowston, 1994; Monteiro and Ladet,
2001a). In this case, each partner has a limited
decision power that corresponds to its action field
(Camalot, et al., 1997; Camalot, 2000; Huguet,
1994).
Multi site resources are considered in this paper
to study the planning management of complex
enterprise networks. A performance criterion valid
locally as well as globally is used, and a “win-win”
policy, based on costs is used, that controls the
network (OUZIZI, et al., 2003).
In addition, the distribution of the enterprise
network cannot be managed only by one and unique
data-processing and data base application. The
main reason is that the exchange of information and
the behaviors, wich are specific to operations of the
network members, are so complex that they require
2
computing
be
to
need
paradigms which
decentralized and shared. Consequently, it seems
that a multi-agent architecture can best meet this
need (Ferber, 1995; Patriti, et al., 1997).
Therefore, a new approach is proposed. It is
based on a cooperative multi-agent architecture in
order to represent and manage the operational
decisions made in complex enterprise networks.
First, relevant research work on the coordination
of supply chains is reviewed. Then, a simple case
for water tab production is described and used to
develop the approach by presenting the model
architecture and the agent specifications.
2. Literature review on coordination of the
supply chains and enterprise networks
Nowadays, there is a large and growing number
of research efforts on coordination and management
of supply chains.
The first area concerns operational research
models. The aim of these models is to coordinate
distributed planning with a collaborative approach
(Dudek and Stadler, 2005; Schneeweiss and
Zimmer, 2004). Each partner, supplier or producer,
is described in the mathematical model by an
objective function. The models are combined to
coordinate
activities, but do not
consider
information flows and decisional processes.
The
second
area
of
research
concerns
collaborative architectures, mostly viewed as a
centralized architecture. A large literature review on
this area can be found in the paper of Stadtler
(2005). Some research papers deal with distributed
architectures. Verwijmeren (2004) develops a
software component architecture dedicated
to
coordination, which
chain
supply
supports
cooperative
distributed
and
organizations.
Nevertheless, none of
these authors consider
cooperative and synchronization behaviors.
Finally, the third area concerns the Multi-Agent
Systems (MAS)
that describe coordination of
supply chains. The MAS are regarded as one of the
most promising
technologies
in supply chain
management. Fox et al. (2000) model such a system
with functional agents, which are responsible of
several activities such as order acquisition,
logistics, transport, or scheduling. The specificity of
the agents is to support distributed decision making.
A review of agent-based approaches in the supply
chain management can be found in a paper of
Parunak (1999) and a critical analysis in Caridi and
|
0806.3032 | 1 | 0806 | 2008-06-18T15:46:29 | Multi-agents architecture for supply chain management | [
"cs.MA"
] | The purpose of this paper is to propose a new approach for the supply chain management. This approach is based on the virtual enterprise paradigm and the used of multi-agent concept. Each entity (like enterprise) is autonomous and must perform local and global goals in relation with its environment. The base component of our approach is a Virtual Enterprise Node (VEN). The supply chain is viewed as a set of tiers (corresponding to the levels of production), in which each partner of the supply chain (VEN) is in relation with several customers and suppliers. Each VEN belongs to one tier. The main customer gives global objectives (quantity, cost and delay) to the supply chain. The Mediator Agent (MA) is in charge to manage the supply chain in order to respect those objectives as global level. Those objectives are taking over to Negotiator Agent at the tier level (NAT). These two agents are only active if a perturbation occurs; otherwise information flows are only exchange between VENs. This architecture allows supply chains management which is completely transparent seen from simple enterprise of the supply chain. The used of Multi-Agent System (MAS) allows physical distribution of the decisional system. Moreover, the hierarchical organizational structure with a decentralized control guaranties, in the same time, the autonomy of each entity and the whole flexibility. | cs.MA | cs | |
1904.01882 | 1 | 1904 | 2019-04-03T09:48:57 | Learning Nash Equilibria in Monotone Games | [
"cs.MA",
"cs.GT"
] | We consider multi-agent decision making where each agent's cost function depends on all agents' strategies. We propose a distributed algorithm to learn a Nash equilibrium, whereby each agent uses only obtained values of her cost function at each joint played action, lacking any information of the functional form of her cost or other agents' costs or strategy sets. In contrast to past work where convergent algorithms required strong monotonicity, we prove algorithm convergence under mere monotonicity assumption. This significantly widens algorithm's applicability, such as to games with linear coupling constraints. | cs.MA | cs | Learning Nash Equilibria in Monotone Games
Tatiana Tatarenko and Maryam Kamgarpour, IEEE Member
1
9
1
0
2
r
p
A
3
]
A
M
.
s
c
[
1
v
2
8
8
1
0
.
4
0
9
1
:
v
i
X
r
a
Abstract -- We consider multi-agent decision making where
each agent's cost function depends on all agents' strategies. We
propose a distributed algorithm to learn a Nash equilibrium,
whereby each agent uses only obtained values of her cost function
at each joint played action,
lacking any information of the
functional form of her cost or other agents' costs or strategy sets.
In contrast to past work where convergent algorithms required
strong monotonicity, we prove algorithm convergence under mere
monotonicity assumption. This significantly widens algorithm's
applicability, such as to games with linear coupling constraints.
Index Terms -- learning in games, distributed algorithms
I. INTRODUCTION
Game theory is a powerful framework for analyzing and
optimizing multi-agent decision making problems. In several
such problems, each agent (referred to also as a player) does
not have full information on her objective function, due to the
unknown interactions and other players' strategies affecting
her objective. Consider for example, a transportation network
in which an agent's objective is minimizing travel time or an
electricity network in which an agent's objective is minimizing
own's electricity prices. In these instances, the travel times
and prices, respectively, depend non-trivially on the strategies
of other agents. Motivated by this limited information setup,
we consider computing Nash equilibria given only the so-
called payoff-based information. That is, each player can only
observe the values of its objective function at a joint played
action, does not know the functional form of her or others'
objectives, nor the strategy sets and actions of other players,
and cannot communicate with other players. In this setting, we
address the question of how agents should update their actions
to converge to a Nash equilibrium strategy.
A large body of literature on learning Nash equilibria
with payoff-based information has focused on finite action
setting or potential games, see for example, [11, 12, 7] and
references therein. For games with continuous (uncountable)
action spaces, a payoff-based approach was developed based
on the extremum seeking idea in optimization [3, 13], and
assuming strongly convex objectives almost sure convergence
to the Nash equilibrium was proven. A payoff-based approach,
inspired by the logit dynamics in finite action games [2] was
extended to continuous action setting for the case of potential
games [14]. The work in [16] considered learning Nash
equilibria in continuous action games on networks. Crucially,
the work additionally assumed that each player exchanges
information with her neighbors, to facilitate estimation of the
gradient of her objective function online.
T. Tatarenko ([email protected]) is with the Control
Methods and Robotics Lab Technical University Darmstadt, Darmstadt, Ger-
many 64283, M. Kamgarpour ([email protected]) is with the Automatic
Control Laboratory, ETH Z urich, Switzerland.
M. Kamgarpour gratefully acknowledges ERC Starting Grant CONENE.
Recently, we proposed a payoff-based approach to learn
Nash equilibria in a class of convex games [15]. Our approach
hinged upon connecting Nash equilibria of a game to the
solution set of a related variational inequality problem. Our
algorithm convergence was established for the cases in which
the game mapping is strongly monotone or the game admits
a potential function. Apart from possibly limited scope of a
potential game, strong monotonicity can be too much to ask
for. In particular, if the objective function of an agent is linear
in her action or in the presence of coupling constraints of the
action sets the game mapping will not be strongly monotone.
Our goal here is to extend the existing payoff-based learning
approaches to a broader class of games characterized by mono-
tone game mappings. While algorithms for solving monotone
variational inequalities exist (see, for example, Chapter 12
in [9]),
these algorithms either consist of two timescales
(Tikhonov regularization approach) or have an extra gradient
step (extra-gradient methods). As such, they require more
coordination between players than that possible in a payoff-
based only information structure.
Our contributions are as follows. First, we propose a dis-
tributed payoff-based algorithm to learn Nash equilibria in a
monotone game, extending our past work [15] applicable to
strongly monotone games, inspired by the single timescale
algorithm for solving stochastic variational inequalities [6].
Second, despite lack of gradients in a payoff-based informa-
tion, contrary to the setup in [6], we show that our proposed
procedure can be interpreted as a stochastic gradient descent
with an additional biasL and regularization terms. Third, we
prove convergence of the proposed algorithm to Nash equi-
libria by suitably bounding the bias and noise variance terms
using established results on boundedness and convergence of
discrete-time Markov processes.
Notations. The set {1, . . . , N} is denoted by [N ]. Bold-
face is used to distinguish between vectors in a multi-
dimensional space and scalars. Given N vectors xi ∈ Rd,
i ∈ [N ], (xi)N
:=
(x1, . . . , xi−1, xi+1, . . . , xN ) ∈ R(N −1)d. Rd
+ and Z+ denote
respectively, vectors from Rd with non-negative coordinates
and non-negative whole numbers. The standard inner product
)⊤ ∈ RN d; x−i
i=1 := (x1⊤
, . . . , xN ⊤
on Rd is denoted by (·,·): Rd×Rd → R, with associated norm
kxk := p(x, x). Given some matrix A ∈ Rd×d, A (cid:23) (≻)0,
if and only if x⊤Ax ≥ (>)0 for all x 6= 0. We use the big-O
notation, that is, the function f (x) : R → R is O(g(x)) as
g(x) ≤ K for
x → a, f (x) = O(g(x)) as x → a, if limx→a
some positive constant K. We say that a function f (x) grows
not faster than a function g(x) as x → ∞, if there exists a
positive constant Q such that f (x) ≤ g(x) ∀x with kxk ≥ Q.
Definition 1: A mapping M : Rd → Rd is monotone over
X ⊆ Rd, if (M (x)− M (y), x− y) ≥ 0 for every x, y ∈ X.
f (x)
II. PROBLEM FORMULATION
Consider a game Γ(N,{Ai},{Ji}) with N players, the sets
of players' actions Ai ⊆ Rd, i ∈ [N ], and the cost (objective)
functions Ji : A → R, where A = A1 × . . .× AN denotes the
set of joint actions. We restrict the class of games as follows.
Assumption 1: The game under consideration is convex.
Namely, for all i ∈ [N ] the set Ai is convex and closed,
the cost function Ji(ai, a−i) is defined on RN d, continuously
differentiable in a and convex in ai for fixed a−i.
Assumption 2: The mapping M : RN d → RN d, referred to
as the game mapping, defined by
M (a) = (∇ai Ji(ai, a−i))N
i=1 = (M 1(a), . . . , M N (a))⊤,
where M i(a) = (Mi,1(a), . . . , Mi,d(a))⊤, and
Mi,k(a) =
∂Ji(a)
∂ai
k
, a ∈ A,
i ∈ [N ],
k ∈ [d],
(1)
is monotone on A (see Definition 1).
We consider a Nash equilibrium in game Γ(N,{Ai},{Ji})
as a stable solution outcome because it represents a joint action
from which no player has any incentive to unilaterally deviate.
Definition 2: A point a∗ ∈ A is called a Nash equilibrium
if for any i ∈ [N ] and ai ∈ Ai
Ji(ai∗, a−i∗) ≤ Ji(ai, a−i∗).
Theorem 1:
Our goal is to learn such a stable action in a game through
designing a payoff-based algorithm. We first connect existence
of Nash equilibria for Γ(N,{Ai},{Ji}) with solution set of a
corresponding variational inequality problem.
Definition 3: Consider a mapping T (·): Rd → Rd and a set
Y ⊆ Rd. A solution SOL(Y, T ) to the variational inequality
problem V I(Y, T ) is a set of vectors y∗ ∈ Y such that
(T (y∗), y − y∗) ≥ 0, ∀y ∈ Y .
(Proposition 1.4.2 in [9]) Given a game
Γ(N,{Ai},{Ji}) with game mapping M , suppose that the
action sets {Ai} are closed and convex, the cost functions {Ji}
are continuously differentiable in a and convex in ai for every
fixed a−i on the interior of A. Then, some vector a∗ ∈ A is
a Nash equilibrium in Γ, if and only if a∗ ∈ SOL(A, M ).
It follows that under Assumptions 1 and 2 for a game with
mapping M , any solution of V I(A, M ) is also a Nash equi-
librium in such games and vice versa. While Γ(N,{Ai},{Ji})
under Assumptions 1 and 2 might admit a Nash equilibrium,
these two assumptions alone do not guarantee existence of
a Nash equilibrium. To guarantee existence, one needs to
consider a more restrictive assumption, for example, strong
monotonicity of the game mapping or compactness of the
action sets [9]. Here, we do not restrict our attention to such
cases. However, to have a meaning discussion, we do assume
existence of at least one Nash equilibrium in the game.
Assumption 3: The set SOL(A, M ) is not empty.
Corollary 1: Let Γ(N,{Ai},{Ji}) be a game with game
mapping M for which Assumptions 1, 2, and 3 hold. Then,
there exists at least one Nash equilibrium in Γ. Moreover, any
Nash equilibrium in Γ belongs to the set SOL(A, M ).
2
Assumption 4: Each element M i of the game mapping
M : RN d → RN d, defined in Assumption (2) is Lipschitz
continuous on Rd with a Lipschitz constant Li.
Assumption 5: Each cost function Ji(a), i ∈ [N ], grows not
faster than a linear function of a as kak → ∞.
III. PAYOFF-BASED ALGORITHM
1, . . . , xi
Given a payoff-based information, each agent has access
to its current action, referred to as its state and denoted by
d)⊤ ∈ Rd, and the cost value Ji(t) at the
xi(t) = (xi
joint states x(t) = (x1(t), . . . , xN (t)), Ji(t) = Ji(x(t)) =
Ji(x1(t), . . . , xN (t)) at iteration t. Using this information in
the proposed algorithm each agent i "mixes" its next state
xi(t + 1). Namely, it chooses xi(t + 1) randomly according
to the multidimensional normal distribution N(µi(t + 1) =
(µi
d(t + 1))⊤, σ(t + 1)) with the density:
1(t + 1), . . . , µi
pi(xi
1, . . . , xi
d; µi(t + 1), σ(t + 1))
1
=
(√2πσ(t + 1))d
) .
The initial value of the means µi(0), i ∈ {N}, can be set to
k − µi
2σ2(t + 1)
any finite value. The successive means are updated as follows:
exp(−
k(t + 1))2
d
Xk=1
(xi
µi(t + 1) = ProjAi(cid:2)µi(t)
− γ(t)σ2(t)(cid:18) Ji(t)
xi(t) − µi(t)
σ2(t)
+ ǫ(t)µi(t)(cid:19)(cid:3).
(2)
In the above, ProjC [·] denotes the projection operator on set C,
γ(t) is a step-size parameter and ǫ(t) > 0 is a regularization
parameter. We highlight the difference between the proposed
approach and that of [15] due to the additional term ǫ(t)
in (2). In the absence of this term the algorithm would not
be convergent under a mere monotonicity assumption on the
game mapping (see counterexample provided in [4]).
Let us provide insight into the algorithm by deriving an
analogy to a regularized stochastic gradient algorithm. Given
σ > 0, for any i ∈ [N ] define Ji : RN d → R as
Ji(µ1, . . . , µN , σ) =ZRN d
Ji(x)p(µ, x, σ)dx,
(3)
1, . . . , xi
i=1 pi(xi
where p(µ, x, σ) = QN
d; µi, σ). Above, Ji,
i ∈ [N ], can be interpreted as the ith player's cost function
in mixed strategies. We can now show that the second term
inside the projection in (2) is a sample of the gradient of
this cost function Ji with respect to the mixed strategies. Let
µ(t) = (µ1(t), . . . , µN (t)).
Lemma 1: Under Assumptions 1 and 5, ∀i ∈ [N ], k ∈ [d]
∂ Ji(µ(t), σ(t))
∂µi
k
=E{Ji(x1(t), . . . , xN (t))
xi
k(t) ∼ N(µi
= Ex(t){ Ji(t)
k(t)
}
xi
k(t) − µi
σ2(t)
k(t)
xi
k(t) − µi
σ2(t)
k(t), σ(t)), i ∈ [N ], k ∈ [d]}.
(4)
The following additional assumptions are needed for con-
vergence of the proposed payoff-based algorithm to a Nash
equilibrium (see proofs of Lemma 3 and Theorem 2).
Proof: We verify that the differentiation under the integral
sign in (3) is justified. It can then readily be verified that
(4) holds, by taking the differentiation inside the integral. A
sufficient condition for differentiation under the integral is
that the integral of the formally differentiated function with
respect to µi
k converges uniformly, whereas the differentiated
function is continuous (see [17], Chapter 17). By formally
differentiating the function under the integral sign and omitting
the arguments t, we obtain
Ji(x)(xi
k)p(µ, x, σ)dx.
(5)
1
σ2 ZRN d
k − µi
k − µi
Given Assumption 1, Ji(x)(xi
k)p(µ, x, σ) is continuous.
Thus, it remains to check that the integral of this function
converges uniformly with respect to any µ ∈ RN d. To this end,
we can write the Taylor expansion of the function Ji around
the point µ(i, k) ∈ RN d with the coordinates µ(i, k)i
k = µi
k
and µ(i, k)j
m for any j 6= i, m 6= k, in the integral (5):
k − µi
[Ji(µ(i, k))
k)p(µ, x, σ)dx
Ji(x)(xi
m = xj
ZRN d
=ZRN d
+
=ZRN d
=ZRN d
∂Ji(η(x, µ))
∂xi
k
∂Ji(η(x, µ))
∂xi
k
∂Ji(η1(y, µ))
∂xi
k
k)p(µ, x, σ)dx
(xi
k − µi
(xi
k)](xi
k − µi
k)2p(µ, x, σ)dx
k − µi
k)2p(0, y, σ)dy,
(yi
where η(x, µ) = µ(i, k) + θ(x − µ(i, k)), θ ∈ (0, 1),
y = x − µ(i, k), η1(y, µ) = µ(i, k) + θy. The uniform
convergence of the integral above follows from the fact1 that,
under Assumption 5, ∂Ji(η1(y,µ))
k for some positive
constant li
k and for all i ∈ [N ] and k ∈ [d]. Hence,
≤ li
∂xi
k
k)2p(0, y, σ) ≤ h(y) = l(yi
(yi
k)2p(0, y, σ),
∂Ji(η1(y, µ))
∂xi
k
where RRN d h(y)dy < ∞.
σ(t) = 1
Lemma (1) shows that the second term inside the projection
in (2) is a sample of the gradient of the cost function in
mixed strategies. Hence, algorithm (2) can be interpreted
as a regularized stochastic projection algorithms. To bound
the bias and variance terms of the stochastic projection and
consequently establish convergence of the iterates µ(t), the
parameters γ(t), σ(t), ǫ(t) need to satisfy certain assumptions.
Assumption 6: Let β(t) = γ(t)σ2(t) and choose γ(t) = 1
ta ,
Theorem 2: Let
limt→∞ ǫ(t) = 0,
tc , a, b, c > 0 respectively, such that
< ∞,
t=0 β(t)σ(t) < ∞,
t=0 β(t)ǫ(t) = ∞.
tb and ǫ(t) = 1
a) P∞
t=0 β(t) = ∞,
β(t)ǫ(t)(cid:17) ǫ(t−1)−ǫ(t)2
t=0(cid:16)1 + 1
b) P∞
c) P∞
t=0 γ2(t) < ∞, P∞
d) limt→∞ σ(t) = 0, P∞
the players in game Γ(N,{Ai},{Ji})
choose the states {xi(t)} at time t according to the normal
distribution N(µi(t), σ(t)), where the mean µi(0) is arbitrary
and µi(t) is updated as in (2). Under Assumptions 1-6, as
t → ∞, the mean vector µ(t) converges almost surely to a
Nash equilibrium µ∗ = a∗ of the game Γ and the joint state
x(t) converges in probability to a∗.
ǫ2(t)
1see the basic sufficient condition using majorant [17], Chapter 17.2.3.
Remark 1: As an example for existence of parameters to
3
satisfy Assumption 6, let a = 5
9 , b = 5
27 , c = 1
27 .
IV. ANALYSIS OF THE ALGORITHM
To prove Theorem 2 we first prove boundedness of the
iterates µ(t). Due to the regularization term ǫ(t), this is done
by analyzing distance of µ(t) from the so-called Tikhonov
trajectory. Having established this boundedness, we can read-
ily show that the limit of the iterates µ(t) exists and sat-
isfies the conditions of a Nash equilibrium of the game
Γ(N,{Ai},{Ji}). For the boundedness and the convergence
proofs, we use established results on boundedness ([8], The-
orem 2.5.2) and convergence of a sequence of stochastic
processes (Lemma 10 (page 49) in [10]), respectively. For ease
of reference, we provide the statement of ([8], Theorem 2.5.2)
and (Lemma 10 (page 49) in [10] ) in the appendix.
A. Boundedness of the Algorithm Iterates
We first show that algorithm (2) falls under the framework
of well-studied Robbins-Monro stochastic approximations pro-
cedures [1] with an additional regularization ǫ(t). Next, lever-
aging this analogy and results on stability of discrete-time
Markov processes ([8], Theorem 2.5.2) applied to the sequence
µ(t) we prove boundedness of the iterates.
Using the notation M i(·) = (Mi,1(·), . . . , Mi,d(·)), we can
rewrite the algorithm step in (2) in the following form:
µi(t + 1) = ProjAi
[µi(t) − γ(t)σ2(t)
×(cid:0)M i(µ(t)) + Qi(µ(t), σ(t)) + Ri(µ(t), x(t), σ(t))
(6)
+ ǫ(t))µi(t)(cid:1),
for all i ∈ [N ]
Qi(µ(t), σ(t)) = M i(µ(t), σ(t)) − M i(µ(t)),
Ri(x(t), µ(t), σ(t)) = F i(x(t), µ(t), σ(t)) − M i(µ(t), σ(t)),
F i(x(t), µ(t), σ(t)) = Ji(t)
xi(t) − µi(t)
,
σ2(t)
and M i(·) = ( Mi,1(·), . . . , Mi,d(·))⊤ is the d-dimensional
mapping with the following elements:
Mi,k(µ(t), σ(t)) =
∂ Ji(µ(t), σ(t))
∂µi
k
, for k ∈ [d].
(7)
The vector M (µ(t)) = (M 1(µ(t)), . . . , M N (µ(t))) cor-
responds to the gradient term in stochastic approximation
procedures, whereas
Q(µ(t), σ(t)) = (Q1(µ(t), σ(t)), . . . ,QN (µ(t), σ(t)))
is a disturbance of the gradient term. Finally,
R(x(t), µ(t), σ(t)) = (R1(x(t), µ(t), σ(t)), . . . ,
RN (x(t), µ(t), σ(t)))
is a martingale difference, namely, according to (4),
Ri(x(t), µ(t), σ(t)) = F i(x(t), µ(t), σ(t))
− Ex(t){F i(x(t), µ(t), σ(t))},
i ∈ [N ].
(8)
To ensure boundedness of µ(t) (Lemma 3) we bound the
martingale term above (see Inequality (23)). To bound the
disturbance of the gradients Q(µ(t), σ(t)) (see Equation (20)),
we observe that the mapping M i(µ(t)) evaluated at µ(t) is
equivalent to the game mapping in mixed strategies (please
see Appendix for the proof of this observation). That is,
Proof: Define V (t, µ) = kµ − y(t − 1)k2, where y(t)
is the Tikhonov sequence defined by (10). We consider the
generating operator of the Markov process µ(t)
LV (t, µ) = E[V (t + 1, µ(t + 1)) µ(t) = µ] − V (t, µ),
and aim to show that LV (t, µ) satisfies the following decay
4
M i(µ(t)) =ZRN d
M i(x)p(µ(t), x)dx.
(9)
In contrast to stochastic approximation algorithms and the
proof in [15], we have an addition term ǫ(t)µ(t) to be able
to address merely monotone game mappings. As such, to
bound µ(t) we also relate the variations of the sequence
µ(t) to those of the Tikhonov sequence defined below. Let
y(t) = (y1(t), . . . , yN (t)) denote the solution of the varia-
tional inequality V I(A, M (y) + ǫ(t)y), namely
y(t) ∈ SOL(A, M (y) + ǫ(t)y).
(10)
The sequence {y(t)} is known as the Tikhonov sequence and
enjoys the following two important properties.
Theorem 3: (Theorem 12.2.3 in [9]) Under Assumptions 2,
3, and 4, y(t) defined in (10) exists and is unique for each
t. Moreover, for ǫ(t) ↓ 0, y(t) is uniformly bounded and
converges to the least norm solution of V I(A, M ).
Lemma 2: (Lemma 3 in [6]) Under Assumption 2
ky(t) − y(t − 1)k ≤ My ǫ(t − 1) − ǫ(t)
ǫ(t)
,
∀t ≥ 1,
With the results above in place, we connect the squared
where My is a uniform bound on the norm of the Tikhonov
sequence, i.e. ky(t)k ≤ My for all t ≥ 0.
distance kµ− y(t)k2 to the squared distance kµ− y(t− 1)k2
for any µ ∈ A and t ≥ 1. Due to the triangle inequality,
kµ − y(t)k ≤ kµ − y(t − 1)k + ky(t − 1) − y(t)k
≤ kµ − y(t − 1)k + My ǫ(t − 1) − ǫ(t)
ǫ(t)
,
(11)
where in the last inequality we used Lemma 2. Hence, by
taking into account that for any a, b ∈ R and θ > 0
2ab ≤ θa2 +
b2
θ
,
we conclude from (11) that for any θ > 0
(13)
LV (t, µ) ≤ −α(t + 1)ψ(µ) + φ(t)(1 + V (t, µ)),
where ψ ≥ 0 on RN d, φ(t) > 0, ∀t, P∞
0, P∞
t=0 φ(t) < ∞, α(t) >
t=0 α(t) = ∞. This enables us to apply Theorem 2.5.2
in [8] to directly conclude almost sure boundedness of µ(t).
Let us bound the growth of V (t + 1, µ) in terms of
V (t, µ). Let θ = β(t)ǫ(t) in (12). From Assumption 6 b),
→ 0 as t → ∞. Hence, ∀µ ∈ A
(14)
β(t)ǫ(t)(cid:17) ǫ(t−1)−ǫ(t)2
(cid:16)1 + 1
V (t + 1, µ) =kµ − y(t)k2
ǫ2(t)
≤(1 + β(t)ǫ(t))kµ − y(t − 1)k2
+(cid:18)1 +
=O(1 + kµ − y(t − 1)k2) = O(1 + V (t, µ)).
β(t)ǫ(t)(cid:19) M 2
y ǫ(t − 1) − ǫ(t)2
ǫ2(t)
1
From the procedure for the update of µ(t),
the non-
expansion property of the projection operator, the fact that y(t)
belongs to SOL(A, M (y) + ǫ(t)y), namely, that ∀i ∈ [N ]
yi(t) = ProjAi [yi(t) − β(t)(M i(y(t)) + ǫ(t)yi(t)],
we obtain that for any i ∈ [N ]
kµi(t + 1) − yi(t)k2
≤ kµi(t) − yi(t) − β(t)(cid:2)ǫ(t)(µi(t) − yi(t))
+ (M i(µ(t)) − M i(y(t)) + Qi(µ(t)) + Ri(x(t), µ(t))(cid:3)k2
= kµi(t) − yi(t)k2
− 2β(t)(M i(µ(t)) − M i(y(t)), µi(t) − yi(t))
− 2β(t)ǫ(t)(µi(t) − yi(t), µi(t) − yi(t))
− 2β(t)(Qi(µ(t)) + Ri(x(t), µ(t)), µi(t) − yi(t))
+ β2(t)kGi(x(t), µ(t))k2,
(15)
where, for ease of notation, we have defined
Gi(x(t), µ(t)) =ǫ(t)(µi(t) − yi(t))
+ M i(µ(t)) − M i(y(t))
+ Qi(µ(t)) + Ri(x(t), µ(t)).
(16)
kµ − y(t)k2 ≤(1 + θ)kµ − y(t − 1)k2
+(cid:18)1 +
1
θ(cid:19) M 2
y ǫ(t − 1) − ǫ(t)2
ǫ2(t)
. (12)
Our goal is to bound E{kµi(t+1)−yi(t)k2µ(t) = µ} above,
and use this bound in constructing Inequality (13). As such,
we expand Gi as below and bound the terms in the expansion.
The above bound serves as the main new inequality in order
to show almost-sure boundedness of kµ(t)k in comparison to
non-regularized stochastic gradient procedures.
Lemma 3: Let Assumptions 2-6 hold in Γ(N,{Ai},{Ji})
and µ(t) be the vector updated in the run of the payoff-based
algorithm (6). Then, Pr{supt≥0 kµ(t)k < ∞} = 1.
In the following, for simplicity in notation, we omit the ar-
gument σ(t) in the terms M , Q, and R. In certain derivations,
for the same reason we omit the time parameter t as well.
kGi(x(t), µ(t))k2 = ǫ2(t)kµi(t) − yi(t)k2
+ kM i(µ(t)) − M i(y(t))k2
+ kQi(µ(t))k2 + kRi(x(t), µ(t))k2
+ 2(Qi(µ(t)), Ri(x(t), µ(t)))
+ 2ǫ(t)(M i(µ(t)) − M i(y(t)), µi(t) − yi(t))
+ 2(ǫ(t)(µi(t) − yi(t)) + M i(µ(t)) − M i(y(t)),
Qi(µ(t)) + Ri(x(t), µ(t))),
(17)
Due to Assumption 4, we conclude that
ikµ − y(t)k2 = O(V (t + 1, µ))
kM i(µ) − M i(y(t))k2 ≤ L2
≤ O(1 + V (t, µ)),
(M i(µ) − M i(y(t)), µi − yi(t))
≤ kM i(µ) − M i(y(t))kkµi − yi(t)k
≤ O(1 + V (t + 1, µ)) ≤ O(1 + V (t, µ)),
where in the last inequalities in (18)-(19) we used (14). Let
us analyze the terms containing the disturbance of gradient,
(18)
(19)
namely Qi, in Equation (17). Since Qi(µ(t)) = M i(µ(t)) −
M i(µ(t)), due to Assumption 2 and Equation (9), we obtain
kQi(µ)k = kZRN d
[M i(x) − M i(µ)]p(µ, x)dxk
≤ZRN d kM i(x) − M i(µ)kp(µ, x)dx
≤ZRN d
Likx − µkp(µ, x)dx
Li N
≤ZRN d
Xi=1
Xi=1
xi
k − µi
Xk=1
= O(
σ),
k! p(µ, x)dx
N
d
(20)
where the last equality is due to the fact
the first
central absolute moment of a random variable with a normal
distribution N(µ, σ) is O(σ). The estimation above and (14)
imply, in particular, that for any µ ∈ A
that
N
σ)(1 + V (t, µ))
Xi=1
kQi(µ)kkµi − yi(t)k ≤ O(
kQi(µ)kkMi(µ) − Mi(y(t))k ≤ LikQi(µ)kkµ − y(t)k
Xi=1
≤ O(
σ)(1 + V (t, µ)).
N
(21)
(22)
Finally, we bound the martingale term kRi(x(t), µ(t))k2.
d
E{kRi(x(t), µ(t))k2µ(t) = µ}
k(t))2
Xk=1ZRN d
≤
k − µi
σ4(t)
2(x)
(xi
Ji
fi(µ, σ(t))
O(1 + V (t, µ))
≤
σ4(t)
≤
σ4(t)
p(µ, x)dx
,
(23)
where the first inequality is due to the fact that E(ξ − Eξ)2 ≤
Eξ2 and taking into account (8), the second inequality is due
to Assumption 5, with fi(µ, σ(t)) being a quadratic function
of µ and σ(t), i ∈ [N ]. Bringing the inequalities (18)-(23)
in the inequality (15), taking into account (14), the Cauchi-
Schwarz inequality, and the martingale properties in (8) of Ri,
i ∈ [N ], we get
E{kµi(t + 1) − yi(t)k2µ(t) = µ}
≤ (1 − 2β(t)ǫ(t))kµi − yi(t)k2
− 2β(t)(M i(µ) − M i(y(t)), µi − yi(t))
− 2β(t)(Qi(µ), µi − yi(t))
+ β2(t)E{kGi(x(t), µ)k2µ(t) = µ}
≤ (1 − 2β(t)ǫ(t))kµi − yi(t)k2
− 2β(t)(M i(µ) − M i(y(t)), µi − yi(t))
+ 2β(t)O(
σ(t))(1 + V (t, µ))
N
Xi=1
+ O(γ2(t))(1 + V (t, µ)),
5
(24)
where in the last inequality we used the fact that ǫ(t) → 0
(Assumption 6 a)), γ(t) → 0, and σ(t) → 0 for all i ∈ [N ]
as t → ∞ (Assumption 6 c), d)). Thus, taking into account
Assumption 6 c), d) and (24), we obtain
N
E[kµi(t + 1) − yi(t)k2µ(t) = µ]
E[kµ(t + 1) − y(t)k2µ(t) = µ]
Xi=1
=
≤(1 − 2ǫ(t)β(t))kµ − y(t)k2
− 2β(t)(M (µ) − M (y(t)), µ − y(t))
+ O(β(t)σ(t) + γ2(t))(1 + V (t, µ)).
(25)
Using the first inequality in (14), we get
LV (t, µ)
= E[kµ(t + 1) − y(t)k2µ(t) = µ] − kµ − y(t − 1)k2
≤ E[kµ(t + 1) − y(t)k2µ(t) = µ]
β(t)ǫ(t)(cid:17) ǫ(t−1)−ǫ(t)2
y(cid:16)1 + 1
kµ − y(t)k2 − M 2
1 + β(t)ǫ(t)
ǫ2(t)
−
. (26)
We conclude from (26) and (25) that
LV (t, µ) ≤(cid:18)1 − 2ǫ(t)β(t) −
− 2β(t)(M (µ) − M (y(t)), µ − y(t))
1
1 + ǫ(t)β(t)(cid:19)kµ − y(t)k2
+ h(t)(1 + V (t, µ))
≤ −2β(t)(M (µ) − M (y(t)), µ − y(t))
+ h(t)(1 + V (t, µ)),
where
h(t) =O(β(t)σ(t) + γ2(t))
+ O(cid:18)(cid:18)1 +
1
β(t)ǫ(t)(cid:19) ǫ(t − 1) − ǫ(t)2
ǫ2(t)
(27)
(cid:19) , (28)
and the second inequality above is due to the fact that
(1 − 2ǫ(t)β(t))(1 + ǫ(t)β(t)) ≤ 1.
account this, (27), and monotonicity of M implying
According to Assumption 6 b)-c), P∞
more, from Assumption 6 a) P∞
(M (µ) − M (y(t)), µ − y(t)) ≥ 0, ∀t, ∀µ ∈ A,
t=0 h(t) < ∞. Further-
t=0 β(t) = ∞. Taking into
(29)
we conclude that LV (t, µ) satisfies the decay needed for the
application of Theorem 2.5.2 in [8] and consequently, µ(t) is
finite almost surely for any t ∈ Z+ irrespective of µ(0).
B. Convergence of the Algorithm
Fortunately, the derivations in the previous section in prov-
ing boundedness of the iterates can be used to also prove
convergence of the algorithm. In particular, we use Inequality
(25), which bounds the decay of the sequence E[kµ(t + 1) −
y(t)k2µ(t)] in terms of kµ − y(t)k2. We can show that this
decay satisfies the conditions for applying Lemma 10 in [10].
From this, it can readily be inferred that random variables
kµ(t)−y(t−1)k converge to zero. In essence, the approach is
similar to showing that V (t, µ) serves as a stochastic Lyapunov
function for the sequence of random variables.
Proof: (of Theorem 2) First, rewrite (25) as follows:
E[kµ(t + 1) − y(t)k2Ft]
≤(1 − 2ǫ(t)β(t))kµ(t) − y(t)k2
+ O(γ2(t) + β(t)σ(t))(1 + V (t, µ(t)))
(30)
≤(1 − 2ǫ(t)β(t))kµ(t) − y(t)k2 + O(γ2(t) + β(t)σ(t))
≤(1 − 2ǫ(t)β(t))(1 + ǫ(t)β(t))kµ(t) − y(t − 1)k2 + O(h(t))
≤(1 − ǫ(t)β(t))kµ(t) − y(t − 1)k2 + O(h(t)),
where Ft is the σ-algebra generated by the random variables
{x(k), µ(k)}t
k=0 and h(t) is defined in (28). In (30) to get
the first inequality we used (29), to get the second inequality
we used Lemma 3, namely the fact that µ(t) is almost surely
bounded for all t ∈ Z+, to get the third inequality we used
(14), and to get the last inequality we used the fact that (1 −
2ǫ(t)β(t))(1 + ǫ(t)β(t)) < (1 − ǫ(t)β(t)).
From Assumption 6, and the choices of γ(t), σ(t), ǫ(t), we
get O(h(t)) = 1
tm , with l > 1, m ≤ 1. Thus,
tl , ǫ(t)β(t) = 1
lim
t→∞
O(h(t))
ǫ(t)β(t)
= 0.
Assumption 6 d), the fact that P∞
t=0 h(t) < ∞ and the above
result in the decay (30) imply that we can apply Lemma 10
in [10] to the sequence kµ(t + 1) − y(t)k2 to conclude its
almost sure convergence to 0 as t → ∞. Next, by taking into
account Theorem 3 and Theorem 1, we obtain that
t→∞
is
µ(t) = a∗} = 1,
Pr{ lim
where a∗
the least norm Nash equilibrium in the
game Γ(N,{Ai},{Ji}). Finally, Assumption 6 implies that
limt→∞ σ(t) = 0. Taking into account
that x(t) ∼
N(µ(t), σ(t)), we conclude that x(t) converges weakly to a
Nash equilibrium a∗ = µ∗. Moreover, according to Portman-
teau Lemma [5], this convergence is also in probability.
V. SIMULATION RESULTS
As noted in the introduction,
the work [4] provides a
counterexample showing that the class of gradient-based pro-
cedures proposed in [16] and [15] fail to converge to a Nash
equilibrium, if the game mapping is merely monotone. In this
section, we demonstrate that the inclusion of the Tikhonov
regularization term in algorithm [15] rectifies this issue. In
particular, the payoff-based algorithm proposed here converges
to the Nash equilibrium in the game under consideration.
6
Following the discussion in [4], we consider the game with
2 players, whose action sets are 1-dimensional sets A1 =
A2 = [−1, 1] and the cost functions are J1(a1, a2) = a1a2
and J2(a1, a2) = −a1a2 respectively. It can be verified that
the game mapping M (a1, a2) = (a2,−a1) is monotone and
the unique Nash equilibrium in this game is a∗ = (0, 0). By
implementing the payoff-based algorithm (6) with randomly
chosen initial values µ1(0) and µ2(0) and the parameters γ(t),
σ(t), and ǫ(t) set up according to Remark 1, we obtain the
updates for the mean values µ1(t) and µ2(t) of the players,
presented in Figure 1. As we can see, the procedure ensures the
means arrive at a sufficiently small neighborhood of the Nash
equilibrium after approximately 900 iterations and continue
approaching it in its further run.
1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
0
1
2
500
1000
time
1500
2000
Fig. 1. The mean values for µ1 and µ2 based on Procedure (2).
VI. CONCLUSIONS
We proposed a payoff-based algorithm for learning Nash
equilibria in convex games with monotone game mappings.
Our algorithm relied on a suitable regularization to handle
monotonicity. The convergence proof relied on the analysis
of the Tikhonov sequence related to the regularization and
well-established results on boundedness and convergence of
stochastic processes. Our current work addresses establishing
convergence rate of the algorithm under suitable assumptions.
REFERENCES
[1] B. Bharath and V. S. Borkar. Stochastic approxima-
tion algorithms: Overview and recent trends. Sadhana,
24(4):425 -- 452, 1999.
[2] L. E. Blume. The statistical mechanics of strategic
interaction. Games and economic behavior, 5(3):387 --
424, 1993.
[3] P. Frihauf, M. Krstic, and T. Basar. Nash equilibrium
seeking in noncooperative games. IEEE Transactions on
Automatic Control, 57(5):1192 -- 1207, 2012.
[4] S. Grammatico.
Comments on "distributed robust
adaptive equilibrium computation for generalized con-
vex games [automatica 63(2016) 82-91)". Automatica,
97:186 -- 188, 2018.
[5] A. Klenke. Probability theory: a comprehensive course.
Springer, London, 2008.
[6] J. Koshal, A Nedi´c, and U. Shanbhag. Single timescale
regularized stochastic approximation schemes for mono-
tone nash games under uncertainty. In IEEE Conference
on Decision and Control, pages 231 -- 236, 2010.
[7] J. R. Marden and J. S. Shamma.
Revisiting log-
linear learning: Asynchrony, completeness and payoff-
based implementation. Games and Economic Behavior,
75(2):788 -- 808, 2012.
[8] M. B. Nevelson and R. Z. Khasminskii.
Stochastic
approximation and recursive estimation. American Math-
ematical Society, 1973.
[9] J.-S. Pang and F. Facchinei. Finite-dimensional varia-
tional inequalities and complementarity problems : vol.
1. Springer series in operations research. Springer, New
York, Berlin, Heidelberg, 2003.
[10] B. T. Poljak. Introduction to optimization. Optimization
Software, 1987.
[11] B. Pradelski and H. P. Young. Learning efficient Nash
equilibria in distributed systems. Games and Economic
behavior, 75(2):882 -- 897, 2012.
[12] J. S. Shamma and G. Arslan. Dynamic fictitious play,
dynamic gradient play, and distributed convergence to
Nash equilibria. IEEE Transactions on Automatic Con-
trol, 50(3):312 -- 327, March 2005.
[13] M. S Stankovic, K. H. Johansson, and D. M. Stipanovic.
Distributed seeking of Nash equilibria with applications
to mobile sensor networks. IEEE Transactions on Auto-
matic Control, 57(4):904 -- 919, 2012.
[14] T. Tatarenko. Stochastic payoff-based learning in multi-
agent systems modeled by means of potential games. In
IEEE Conference on Decision and Control, pages 5298 --
5303, 2016.
[15] T. Tatarenko and M. Kamgarpour. Learning generalized
IEEE
to appear.
nash equilibria in a class of convex games.
Transactions on Automatic Control, 2018.
URL: https://arxiv.org/abs/1703.04113.
[16] M. Zhu and E. Frazzoli. Distributed robust adaptive
equilibrium computation for generalized convex games.
Automatica, 63:82 -- 91, 2016.
[17] V.A. Zorich and R. Cooke. Mathematical Analysis II.
Mathematical Analysis. Springer, 2004.
A. Supporting Theorems
APPENDIX
Let {X(t)}t, t ∈ Z+, be a discrete-time Markov process on
some state space E ⊆ Rd, namely X(t) = X(t, ω) : Z+ ×
Ω → E, where Ω is the sample space of the probability space
on which the process X(t) is defined. The transition function
of this chain, namely Pr{X(t+1) ∈ ΓX(t) = X}, is denoted
by P (t, X, t + 1, Γ), Γ ⊆ E.
Definition 4: The operator L defined on the set of measur-
able functions V : Z+ × E → R, X ∈ E, by
LV (t, X) =Z P (t, X, t + 1, dy)[V (t + 1, y) − V (t, X)]
= E[V (t + 1, X(t + 1)) X(t) = X] − V (t, X),
is called a generating operator of a Markov process {X(t)}t.
7
Next, we formulate the following theorem for discrete-time
Markov processes, which is proven in [8], Theorem 2.5.2.
Theorem 4: Consider a Markov process {X(t)}t and sup-
there exists a function V (t, X) ≥ 0 such that
pose that
inf t≥0 V (t, X) → ∞ as kXk → ∞ and
LV (t, X) ≤ −α(t + 1)ψ(t, X) + f (t)(1 + V (t, X)),
where ψ ≥ 0 on R× Rd, f (t) > 0, P∞
t=0 f (t) < ∞. Let α(t)
be such that α(t) > 0, P∞
t=0 α(t) = ∞. Then, almost surely
supt≥0 kX(t, ω)k = R(ω) < ∞.
The following result related to the convergence of the
stochastic process is proven in Lemma 10 (page 49) in [10].
Theorem 5: Let v0, . . . , vk be a sequence of random vari-
ables, vk ≥ 0, Ev0 < ∞ and let
E{vk+1Fk} ≤ (1 − αk)vk + βk,
where Fk is the σ-algebra generated by the random variables
k=0 αk = ∞, βk ≥ 0,
= 0. Then vk → 0 almost surely,
{v0, . . . , vk}, 0 < αk < 1, P∞
P∞
k=0 βk < ∞, limk→∞
Evk → 0 as k → ∞.
βk
αk
B. Verification of Equation (9)
We will show that the mapping M i(µ(t), σ(t)) (see (7))
evaluated at µ(t) is equivalent to the extended game mapping:
M i(µ(t)) =ZRN d
M i(x)p(µ(t), x)dx.
1
p(µi
−k) =
−k, xi
k−1, µi
k−1, xi
1, . . . , µi
1, . . . , xi
k−1, . . . µi
k−1, . . . xi
µi
−k = (µi
xi
−k = (xi
Note that for simplicity in notation, we drop the dependence
on σ(t) and on t. Now, using the notations
d) ∈ Rd−1,
d) ∈ Rd−1,
exp
(xi
−Xj6=k
exp(−
i
(xj
k − µj
k)2
2σ2
j
(√2πσi)d−1
p(µ−i, x−i) =
(√2πσj)d
Yj6=i,j=1
Xk=1
we have that for any i ∈ [N ], k ∈ [d], Mi,k(η)
Mi,k(µ) =
∂ Ji(µ(t), σ(t))
j − µi
2σ2
) ,
j)2
1
N
d
∂µi
k
=
1
σ2
i ZRN d
= −ZRN d
Ji(x)(xi
k)p(µ, x)dx
k − µi
−k, xi
Ji(x)p(µi
−k)p(µ−i, x−i)
1
√2πσi
−
× d e
(xi
k −µi
2σ2
k )2
i ! dx−i
= −ZRN d−1 Ji(x)e
× p(µi
−k, xi
(xi
−
k −µi
2σ2
k)
∞(xi
k )2
i !(cid:12)(cid:12)(cid:12)(cid:12)
−∞(xi
k)
−k)p(µ−i, x−i)
1
√2πσi
dx−i
+ZRN d
∂Ji(x)
∂xi
k
p(µ, x)dx
=ZRN d
∂Ji(x)
∂xi
k
p(µ, x)dx.
(31)
In the above, for the second equality, we used Lemma (1)
to enable differentiation under the integral and for the last
equality, we used the fact that according to Assumption 5,
8
−
(xi
k )2
k −µi
2σ2
i
= 0,
lim
Ji(x)e
xi
k→∞(−∞)
k, x−i. Now, by definition of M i(x), we have
for any fixed µi
that
ZRN d
∂Ji(x)
∂xi
k
p(µ, x)dx =ZRN d
as desired.
M i(x)p(µ(t), x)dx,
|
1611.06413 | 1 | 1611 | 2016-11-19T18:57:05 | Formalizing Multi-Agent Systems Using Action Descriptions in Single Agent Perspective | [
"cs.MA",
"cs.LO"
] | Logic-based representations of multi-agent systems have been extensively studied. In this work, we focus on the action language BC to formalize global views of MAS domains. Methodologically, we start representing the behaviour of each agent by an action description from a single agent perspective. Then, it goes through two stages that guide the modeler in composing the global view by first designating multi-agent aspects of the domain via potential conflicts and later resolving these conflicts according to the expected behaviour of the overall system. Considering that representing single agent descriptions is relatively simpler than representing multi-agent description directly, the formalization developed here is valuable from a knowledge representation perspective. | cs.MA | cs |
Formalizing Multi-Agent Systems Using Action
Descriptions in Single Agent Perspective
Orkunt Sabuncu1,2 Torsten Schaub1 Christian Schulz-Hanke1
1 University of Potsdam, Germany
2 NOVA LINCS, Universidade Nova de Lisboa, Portugal
Abstract. Logic-based representations of multi-agent systems have been exten-
sively studied. In this work, we focus on the action language BC to formalize
global views of MAS domains. Methodologically, we start representing the be-
haviour of each agent by an action description from a single agent perspective.
Then, it goes through two stages that guide the modeler in composing the global
view by first designating multi-agent aspects of the domain via potential con-
flicts and later resolving these conflicts according to the expected behaviour of
the overall system. Considering that representing single agent descriptions is rel-
atively simpler than representing multi-agent description directly, the formaliza-
tion developed here is valuable from a knowledge representation perspective.
1 Introduction
Logic-based representations of multi-agent systems (MAS) have been extensively stud-
ied. Well-defined semantics of such representations allow for carrying out various rea-
soning tasks about MAS. Action languages [1] are among these representations of
MAS. They have been successfully applied to represent and reason not only on sin-
gle agent [2,3] but also multi-agent domains [4,5,6].
In this work, we focus on the action language BC [7] to formalize MAS domains.
More specifically, the formalization deals with the global view of MAS domains, which
captures all knowledge about environment and capabilities of agents. It is basically
modelled as a transition diagram whose vertices and edges denote states of the domain
and sets of agents' actions occurring concurrently, respectively.
Unlike earlier works [5,6,8], where the modeler is expected to encode the global
view using an action language from scratch, our formalization takes action descriptions
of each agent from a single agent perspective as input and goes through two stages
focusing on different aspects of the domain. Note that representing an action description
of an agent with a local perspective is relatively easier than representing the whole
multi-agent domain directly due to interacting concurrent actions. The first stage in
our methodology establishes an intermediate action description that covers all valid
states of the global view and identifies cases of the domain where actions of agent
interact and have effects that are invisible from a single agent perspective. To this end,
we use potential conflict to specify such cases formally. The second stage addresses
identified potential conflicts by resolving them so that corresponding concurrent actions
have the desired effect in the environment of the domain. The resolution is achieved via
defeating involved dynamic laws of the intermediate description. This is possible via a
syntactic transformation that makes dynamic laws of an action description defeasible.
Formal properties of this transformation are given and the existence of a resolution at
the second stage is proved. The resulting action description represents the transition
diagram corresponding to the global view of the MAS domain.
The action language BC can express defeasible laws. This capacity is crucial for
our formalization and has been our main motivation in selecting BC as the underlying
action language.
The formalization, which is the novel contribution of this work, guides the modeler
and structures her efforts by making multi-agent aspects of the domain explicit. To
the best of our knowledge, such a methodology has not been explicitly defined and
employed before.
While the scope of this work focuses on the global view of a MAS domain, it is
important to emphasize that real agents residing in the environment may have autonomy
via their own local view exhibiting a behaviour different than the one represented by the
agent description from a single agent perspective used in the methodology. In fact they
can utilize their own strategies on how to behave and even try to perform illegal actions.
It is up to the MAS architecture utilizing this global view to react to these illegal actions
(by simply discarding them or issuing an execution failure).
In what follows, we provide background information on the action language BC.
We also illustrate an action description describing a single Sumo agent, which is the
base of the running example used throughout the paper. In Section 3, we describe our
methodology of formalizing MAS with two stages. Finally, we conclude the paper with
discussion that also includes future lines of research in Section 4.
2 Background
We give below syntax and semantics of the Boolean fragment of slightly modified action
language BC, as introduced in [7].
We consider a signature of Boolean action and fluent symbols, denoted by A and
F, respectively. In what follows, we simply refer to them as actions and fluents. Fluents
are further divided into regular and defined fluents. A defined fluent is useful for repre-
senting a property that is statically determined in terms of other fluents. A fluent literal
is a fluent a or its negation ¬a. Similarly, an action literal is an action c or its negation
¬c. A static law is an expression of the form
a0 if a1, . . . , am ifcons am+1, . . . , an
(1)
where ai is a fluent literal for 0 ≤ m ≤ n; a dynamic law is an expression of the form
a0 after a1, . . . , am ifcons am+1, . . . , an
(2)
where a0 is a regular fluent literal, each of a1, . . . , am is a fluent or action literal, and
am+1, . . . , an are fluent literals. An action description is a finite set of of static and
dynamic laws. The ifcons part of a law is important for representing defaults. Various
abbreviations, such as impossible, nonexecutable, inertial, and default laws, are use-
ful for developing succinct action descriptions. They can be rewritten in terms of static
and dynamic laws (refer to [7] for their definitions).
The semantics of action descriptions is given in terms of transition systems induced
by a translation into logic programs under stable models semantics [9]. To be more
precise, an action description D and a horizon l yield a program Pl(D) whose stable
models represent all paths of length l in the transition system corresponding to D. The
signature of each program Pl(D) consists of labeled expressions of the form i : a,
where i ≤ l and a is a fluent literal, or i < l and a is an action literal. As defined in [7],
each program Pl(D) consists of the following rules
1. for each static law of form (1) in D and i ≤ l a rule of form
i : a0 ← i : a1, . . . , i : am, not not i : am+1, . . . , not not i : an
(3)
2. for each dynamic law of form (2) in D and i < l a rule of form
(i+1) : a0 ← i : a1, . . . , i : am, not not (i + 1) : am+1, . . . , not not (i + 1) : an
(4)
3. for each regular fluent f in D a choice rule of form
4. for each action a in D and i < l a choice rule of form
{0 : f, 0 : ¬f }
{i : a}
5. for each fluent f in D and i ≤ l an integrity constraint of form
6. for each action a in D and i < l a rule of form
← {i : f, i : ¬f } 6= 1
i : ¬a ← not i : a
(5)
(6)
(7)
(8)
For a set X of labeled expressions and i ≥ 0, define Xi = {a i : a ∈ X} The
transition system (S(D), T (D)) induced by an action description D is then defined as
follows. 3
S(D) = {X0 X is a stable model of P0(D)}
T (D) = {hX0 ∩ L, X0 ∩ A, X1i X is a stable model of P1(D)}
(9)
(10)
where L = F ∪ {¬f f ∈ F}
Note that unlike the logic program used for semantics of BC in [7] we name unper-
formed actions by explicitly generating negative action literals in (8) and consequently
allow negative action literals in the after part of a dynamic law. This is critical in a
3 Note that X1 = X1 ∩ L.
multi-agent setting to represent situations where an agent does not perform a specific
action. Such a case is illustrated using the running example in Section 3.
For illustration, consider a (single) Sumo agent in a ring divided into l horizontal
slots. A Sumo can move left or right to adjacent slots in the ring, and may drop out
at each end. We capture this through fluents at (A, L), out (A) and actions left (A),
right (A), respectively, where the variable A stands for an agent identifier4 and L, L′ ∈
{1, . . . , l}.
The behavior of our simple Sumo agent in an l slot ring is represented by the follow-
ing action description, composed of static laws (11) -- (14) and dynamic laws (15) -- (21),
respectively.
¬at (A, L) if at (A, L′)
¬out (A) if at(A, L)
¬at (A, L) if out (A)
impossible ¬at (A, 1), . . . , ¬at (A, l), ¬out (A)
at (A, L) after left (A), at (A, L′)
at (A, L) after right (A), at (A, L′)
out(A) after left (A), at (A, 1)
out(A) after right (A), at (A, l)
nonexecutable left (A) if out (A)
nonexecutable right (A) if out (A)
inertial at (A, L), out(A)
(L 6= L′)
(L = L′ − 1)
(L = L′ + 1)
(11)
(12)
(13)
(14)
(15)
(16)
(17)
(18)
(19)
(20)
(21)
Instantiating this action description with an agent a moving in a two-slot ring where
l = 2 yields a transition system with three states; s1 = {at(a, 1), ¬at (a, 2), ¬out (a)},
s2 = {¬at (a, 1), at (a, 2), ¬out (a)}, and s3 = {¬at(a, 1), ¬at (a, 2), out (a)}, along
with four transitions hs1, {right (a)}, s2i, hs2, {left (a)}, s1i, hs1, {left (a)}, s3i, and
hs2, {right (a)}, s3i.
3 Foundations
We consider a set A of agents whose actions are governed by an environment. We
formalize such multi-agent systems by means of action descriptions in BC, one for each
agent in A and descriptions capturing their interplay.
Our formalization has two stages, which will be described in the following two
subsections. The first one focuses on designating aspects of the domain that do not
show up from a single-agent perspective but arise once there are multiple agents. In the
second stage, all these aspects are handled so that we get a global view of the overall
multi-agent system.
To be more precise, the signature of a multi-agent system (A, c, r) consists of ac-
tions A and fluents F. Each agent a ∈ A is represented by an action description Da
4 Strictly speaking, agent identifiers are obsolete in a single-agent environment but their intro-
duction paves the way for the multi-agent setting in the next section.
over Aa ⊆ A and Fa ⊆ F such that Aa ∩ Aa′ = ∅ for all a 6= a′. That is, while
agents may share fluents, their actions are distinct. Each action description represents
the correct behavior of an agent within the environment from a single-agent perspective
-- specific agents may or may not behave accordingly. Components c and r are repre-
sented by action descriptions Dc and Dr over A and F, respectively. They correspond
to the first and second stage in our methodology respectively. While the role of the
component c is identifying potential conflicts, the role of r is resolving such conflicts.
3.1 Identifying potential conflicts (the first stage)
Let us extend our previous example with a second Sumo b. This results in two action
descriptions Da and Db, which are composed of laws (11) -- (21) only differing in the
used agent identifier, viz. a and b, respectively. Both Sumos can thus move left or right
to adjacent slots in the ring and drop out at each end. Also, there can only be one Sumo
in a slot at a time. A Sumo who moved or is pushed out of the ring cannot return. It
can, however, resist a moving opponent by moving in the reverse direction. This results
in both Sumos staying in their previous slots. Similarly, if both Sumos want to move to
the same slot at the same time, they bounce back and stay in their previous slots.
The mere union of all action descriptions capturing single agents in A may lack
some valid states of the multi-agent system. One role of component c is to rectify this
via action description Dc. In particular, it may be necessary for the action description
Dc to defeat some of the static laws of single agent descriptions. To this end, we turn
static laws of single agent descriptions into default rules by introducing abnormality
fluents.
Definition 1. Let D be an action description in BC. Then, we define τ (D) as the result
of
1. replacing each static law of form (1) in D by
a0 if a1, . . . , am ifcons am+1, . . . , an, ¬ab(a0)
2. adding for each defined fluent ab(a0) introduced in (22) a rule of form
3. copying each dynamic law without change.
default ¬ab(a0)
(22)
(23)
At the end of the first stage of our formalization of the global view of a multi-agent
system, we get the intermediate action description:
U(A,c) = [
τ (Da) ∪ Dc
a∈A
In our Sumo example, we do not need to defeat any static laws since the union of all sin-
gle agent descriptions does not lack any states of the multi-agent system. In Section 3.3
we give an example domain where the component c has to defeat some static laws to
generate some previously lacking valid states of U(A,c).
Apart from lacking states, the mere union of agent action descriptions is prone to in-
valid states and transitions from a multi-agent perspective. For instance, the union of ac-
tion descriptions Da and Db tolerates both Sumos in the same slots, as manifested by the
states {at(a, L), at (b, L)}. Also, it permits both Sumos passing through each other, ex-
hibited by the transition h{at (a, 2), at (b, 3)}, {right(a), left (b)}, {at(a, 3), at(b, 2)}i.
Such invalid states and transitions must be ruled out by appropriate laws in Dc to gov-
ern the interplay of Da and Db. Consider the action description Dc consisting of the
laws in (24) and (25).
¬at (A, L) if at (A′, L)
nonexecutable right (A), left (A′) if at (A, L), at (A′, L + 1)
(A 6= A′)
(A 6= A′)
(24)
(25)
The action description U(A,c) induces 21 states. The part of the transition system where
Sumo a is left of b is given in Figure 1.
Note that ab fluents introduced by the transformation τ are statically defined and
they are false by default due to law (23). Since there are no laws in Dc causing an ab
fluent to be true in the Sumo example, there are no states of U(A,c) in which an ab fluent
holds.
Fig. 1. A part of the transition system of the union of action descriptions. Self loops with ∅
compound action are omitted for clearance. lx and rx abbreviate left(x) and right (x).
s5
a b
ra, rb
la, lb
s2
a b
rb
la
lb
la, rb
ra, lb
ra
lb
lb
s1
la
a
b
s6
a b
la, lb
ra, rb
rb
la
ra
a
b
s3
rb
la, lb
ra
ra, rb
a
b
s4
The other role of component c, as stated at the beginning of this section, is to desig-
nate aspects of the domain regarding the interplay of multiple agents in the environment.
To this end, we first define the notion of potential conflict to cover such multi-agent as-
pects. A potential conflict is a compound action that cannot be executed in a state in
view of the laws of an action description.
Definition 2. Let D be an action description in BC, (S(D), T (D)) the corresponding
transition system, and s ∈ S(D).
We define a potential conflict in s and D as
χD(s) = {c ⊆ A there exists no s′ ∈ S(D) s.t. hs, c, s′i ∈ T (D)} .
In our running example, there are basically 3 cases of such multi-agent aspects
of the domain: (i) a Sumo can push another Sumo, (ii) Sumos can resist a push by
moving in the reverse direction, and (iii) two Sumos bounce back when they want to
move to the same slot at the same time. The description Dc composed of laws (24)
and (25) already identifies all potential conflicts corresponding to these 3 cases. For
some representative potential conflicts of (ii) and (iii) case, consider the compound
action {right (a), left (b)}. While it can be used to switch from s1 to s2, it cannot be
used to leave any other state in Figure 1. A transition with this compound action is
ruled out in states s5, s2, and s6 by (25) and in s3 and s4 by (24). Clearly, we have
{right (a), left (b)} ∈ χU({a
c) (s) for s = s2, . . . , s6. Additionally, for the first case,
we observe that {right (a)} ∈ χU({a
c) (s) for s =
s2, s5, s6.
c)(s) and {left (b)} ∈ χU({a
b},
b},
,
b},
,
,
3.2 Resolving potential conflicts (the second stage)
The first stage of our formalization ends with identifying potential conflicts. Such con-
flicts are inadmissible in a multi-agent setting. It should be allowed for the individual
agents to perform the corresponding compound action. The second stage is about pre-
venting identified conflicts that are related to the multi-agent aspects of the domain
from becoming actual conflicts. The role of the conflict resolution component r in the
multi-agent system (A, c, r) is to rectify this. To this end, the action description Dr has
to defeat some of the dynamic laws of U(A,c).
Similar to the transformation τ, we turn dynamic laws of the description U(A,c) into
default rules by introducing an abnormality fluent using the transformation β.
Definition 3. Let D be an action description in BC.
Then, we define β(D) as the result of
1. replacing each dynamic law of form (2) in D by
a0 after a1, . . . , am ifcons am+1, . . . , an, ¬ab ′(a0)
2. adding for each fluent ab ′(a0) introduced in (26) a rule of form
3. copying each static law without change.
default ¬ab ′(a0)
(26)
(27)
Note that unlike transformation τ, abnormality fluents introduced in β are regular flu-
ents. This allows the modeler to use such fluents in heads of dynamic laws in Dr. At
the end of the second stage of the formalization we get the global view of the overall
multi-agent system represented by the action description:
M(A,c,r) = β(U(A,c)) ∪ Dr
Before illustrating the resolution step for our running example, we analyze the prop-
erties of transformation β and action description Dr. Due to Lemma 1, we can be sure
that potential conflicts of U(A,c) are preserved after applying β.
Lemma 1. Let D be an action description in BC. c ∈ χD(s) iff c ∈ χ
β(D)(s∗) where
s ⊆ s∗ and s∗ \ s has either ab′ or ¬ab′ for each ab′ fluent introduced in β(D) and no
other literal.
Although the modeler knows that potential conflicts identified in the first stage of
the formalization are preserved by β, she is not guided on the structure of laws needed
in Dr in order to resolve a conflict. To remedy this situation, we define a condition of a
dynamic law being covered by a compound action at a state.
Definition 4. A dynamic law of the form (2) in D is covered by compound action c at
state s ∈ S(D) iff for all 1 ≤ i ≤ m (i) ai ∈ c when ai is a positive action literal, (ii)
ai 6∈ c when ai is a negative action literal, and (iii) ai ∈ s when ai is a fluent literal.
Furthermore, Lemma 2 shows that laws of Dr needed for resolving a potential con-
flict must be dynamic laws that are covered by the related compound action and the
state. When the compound action and state are clear from the context, we may only use
the statement that a law is covered.
Lemma 2. Let D be an action description in BC over actions A. Given a compound
action c ⊆ A and a state s ∈ S(D), hs, c, s′i ∈ T (D) iff hs, c, s′i ∈ T (D′) where D′ is
equal to D except all its dynamic laws not covered by c at s are taken out.
For resolving a potential conflict, the modeler has to encode covered dynamic laws
in Dr. Basically, some laws in Dr should first defeat dynamic laws of β(U(A,c)) caus-
ing contradictory effects w.r.t. the desired successor state . Then additional laws in Dr
generate effect fluent literals that are previously not possible. We illustrate this method-
ology using our running example.
Consider the potential conflict related to two Sumos trying to move to the same slot
at the same time for the state s4 in Figure 1, i.e., {right (a), left (b)} ∈ χU({a
c)(s4).
Considering instances of dynamic laws (15), (16), and (21), the laws (28) -- (31) in
β(U(A,c)) are covered by {right (a), left (b)} at s4.
b},
,
at (a, 3) after right (a), at (a, 2) ifcons ¬ab ′(at (a, 3))
at (b, 3) after left (b), at (b, 4) ifcons ¬ab ′(at (b, 3))
at (a, 2) after at(a, 2) ifcons at (a, 2), ¬ab ′(at (a, 2))
at (b, 4) after at (a, 4) ifcons at (b, 4), ¬ab ′(at (b, 4))
(28)
(29)
(30)
(31)
Since Sumos bounce back in this case, effect fluents at (a, 3) and at(b, 3) do not hold in
the desired successor state. Hence, we need to defeat the laws (28) and (29) by causing
abnormality fluents ab ′(at (a, 3)) and ab ′(at (b, 3)) to be true in order to resolve the
potential conflict. In the general case, laws (32) and (33) in Dr cause these abnormality
fluents and resolve all potential conflicts related to Sumos trying to move to the same
slot at the same time. Considering these potential conflicts, notice that laws (32) and
(33) are covered by respective compound actions at respective states as formally stated
in Lemma 2.
ab ′(at (A, L + 1)) after right (A), left (A′), at (A, L), at (A′, L + 2),
¬left (A), ¬right (A′)
(A 6= A′)
ab ′(at (A′, L + 1)) after right (A), left (A′), at (A, L), at (A′, L + 2),
¬left (A), ¬right (A′)
(A 6= A′)
(32)
(33)
Next, we resolve the potential conflicts related to a Sumo pushing another Sumo. For
example, consider {right (a)} ∈ χU({a
c)(s2). In such a case, at the desired successor
state the pushed Sumo moves one slot in the direction of push (and maybe pushed out of
the ring if he is at a border slot). Dynamic laws (34) -- (37) in Dr resolve such potential
conflicts.
b},
,
at (A′, L + 2) after at (A, L), at (A′, L + 1), right (A), ¬left (A′)
(A 6= A′, L + 1 < 4)
at (A, L − 1) after at(A, L), at (A′, L + 1), left (A′), ¬right (A)
(A 6= A′, L > 1)
out (A′) after at (A, L), at (A′, L + 1), right (A), ¬left (A′)
(A 6= A′, L = 3)
out (A) after at (A, L), at (A′, L + 1), left (A′), ¬right (A)
(A 6= A′, L = 1)
(34)
(35)
(36)
(37)
Note that unlike the previous case, we have not defeated some covered laws by explic-
itly causing some ab′ fluents to be true. Actually, related literals in the successor state
caused by laws (34) -- (37) defeat contradictory covered laws (in this case laws of intertia
for the pushed Sumo) via indirect effects.
The last case concerns the potential conflicts related to two Sumos trying to push
each other at the same time. In the successor state Sumos should be in their previous
slots. The laws (38) -- (40) in Dr defeat contradictory covered laws, which represent
effects of movement, by causing related ab′ fluents to be true.
ab ′(at (A, L + 1)) after right (A), left (A′), at (A, L), at (A′, L + 1),
¬left (A), ¬right (A′)
(A 6= A′)
ab ′(at (A′, L)) after right (A), left (A′), at (A, L), at (A′, L + 1),
¬left (A), ¬right (A′)
(A 6= A′)
ab ′(n(A, A′, L)) after right (A), left (A′), at (A, L), at (A′, L + 1),
¬left (A), ¬right (A′)
(A 6= A′)
(38)
(39)
(40)
The law (40) defeats the law in β(U(A,c)) that is β transformed version of (25).
In principle such nonexecutable laws are abbreviations of pairs of dynamic laws with
contradictory head fluents [7]. We assume that these head literals are unique for each
nonexecutable law in order to avoid overly defeating such laws in the second stage. For
an action description, this can be achieved by introducing some fresh fluents for each
of such dynamic law pairs. Otherwise, a slightly modified version of β may generate
a unique id in the introduced abnormality fluent for each nonexecutable law.5 For in-
stance, in our Sumo example β adds the ¬ab ′(n(A, A′, L)) abnormality fluent to the
ifcons part when transforming dynamic laws abbreviated by (25) in a uniquely identi-
fied fashion.
Proposition 1 guarantees that a potential conflict in U(A,c) can always be resolved
in M(A,c,r). The resolution used in the proposition is clearly not the only way and
may also not be the pragmatic way to resolve a potential conflict (for instance, a more
concise way has already been illustrated in the Sumo example). However, it forms a
basic guideline as a general methodology on resolving conflicts, i.e., defeating dynamic
laws causing contradictory effects and generating desired effects in the successor state.
Proposition 1. Let D be an action description in BC over actions A and c ⊆ A be
a compound action. If c ∈ χD(s), then for any state s′ ∈ S(D), hs∗, c, s′ ∪ di ∈
T (β(D) ∪ R) such that s ⊆ s∗ and s∗ \ s has either ab′ or ¬ab′ for each ab′ fluent
introduced in β(D) and no other literal; R is a set of dynamic laws covered by c at s of
the form
ab ′(a0) after a1, . . . , am
(41)
for each dynamic law in D that is covered by c at s, where ab ′(a0) is introduced in
β(D) for the covered law in D, and of the form
f after a1, . . . , am
(42)
for each literal f ∈ s′; and set d is composed of positive ab′ literals that appear in
heads of laws (41) in R and negative ab′ literals for all the rest ab′ fluent symbols
introduced in β(D).
Let Dr be composed of laws (32) -- (40). At the end of the second stage, the ac-
tion description M(A,c,r) gives the global view of the Sumo agents. Considering a
specific potential conflict, for instance, {right (a), left (b)} ∈ χU({a
c) (s4) holds
given the state s4 in Figure 1. After the resolution stage, however, it is not anymore
a potential conflict of M(A,c,r) and the transition diagram has a corresponding transi-
tion, i.e., hs4, {right (a), left (b)}, s7i ∈ T (M(A,c,r)) where the successor state s7 =
{at(a, 2), at (b, 4), ab ′(at (a, 3)), ab ′(at (b, 3))} (Sumos bounce back and stay in their
previous slots).
b},
,
Since the ab′ fluents introduced by the transformation β are regular fluents, there
may be states of M(A,c,r) that include superfluous ab′ fluents. In our Sumo exam-
ple, for instance, {at(a, 1), at (b, 4), ab ′(at (a, 3))} ∈ S(M(A,c,r)) holds. Such states,
however, are not accessible from sound initial states. We plan to address this issue for-
mally by augmenting our formalization with a query language that enables one to ex-
press initial states and reasoning tasks (see also Section 4 for the related future work).
5 This also applies to impossible laws and the transformation τ .
3.3 Defeating static laws in the first stage
Recall that in the Sumo example, we have not used ab fluents introduced by τ to defeat
some static laws, since all valid states of the multi-agent domain are generated by the
mere union of agent descriptions of single agent perspective. Consider another domain
where there are two agents next to opposite ends of a table [10]. The table is initially on
the floor. Each agent may lift the table up using its respective end. Whenever an agent
lifts up the table, it holds the table steady in its resulting state. The table is fully lifted
when it is lifted from both ends by respective agents.
Agent l is next to the left end of the table. The behaviour of this agent can be
captured through fluents table(P ) and the action lift l where P ∈ {onf loor, lef tup}.
The fluent table(lef tup) represents the situation where the left end of the table is lifted
up and the right end of it is on the floor.
The behaviour of l from a single agent perspective can be modeled by the action
description Dl that is composed of laws (43) -- (48).
¬table(lef tup) if table(onf loor)
¬table(onf loor) if table(lef tup)
impossible ¬table(onf loor), ¬table(lef tup)
table(lef tup) after lift l
nonexecutable lift l if table(lef tup)
inertial table(P )
(43)
(44)
(45)
(46)
(47)
(48)
Dl has 2 states; s1 = {table(onf loor), ¬table(lef tup)} and s2 = {table(lef tup),
¬table(onf loor)}. It has 3 transitions; hs1, {}, s1i, hs2, {}, s2i, and hs1, {lift l}, s2i.
The behaviour of agent r is similar to that of l and the description Dr is equal to Dl
except fluent table(rightup) and action lift r are used instead of table(lef tup) and
lift l respectively.
The mere union of descriptions Dl and Dr has two states; a state where the only
positive fluent literal is table(onf loor) and an invalid state where both table(lef tup)
and table(rightup) hold. Besides, it is easy to see that the state where the table is fully
lifted is invisible to agents from a single agent perspective.
At the end of the first stage of our formalization, U(A,c) must cover all valid states
of the multi-agent system. Let Dc be laws (49) -- (56). Note that Dc uses the new fluent
table(lif ted) to cover a state that is invisible from single agent perspective. However,
this needs defeating static laws (45) from descriptions Dl and Dr. This is achieved by
static laws (52) and (53) where ¬ab(imp(l)) and ¬ab(imp(r)) are abnormality fluents
introduced by τ to make the corresponding impossible laws defeasible.
¬table(P ) if table(lif ted)
¬table(lif ted) if table(P )
¬table(P ) if table(P ′)
ab(imp(l))
(P 6= lif ted)
(P 6= lif ted)
(P 6= P ′; P, P ′ ∈ {rightup, lef tup})
(49)
(50)
(51)
(52)
(53)
ab(imp(r))
impossible ¬table(onf loor), ¬table(lef tup), ¬table(rightup), ¬table(lif ted)
nonexecutable lift l if table(P )
nonexecutable lift r if table(P )
(P 6∈ {lef tup, onf loor})
(P 6∈ {rightup, onf loor})
(54)
(55)
(56)
Dynamic laws (55) and (56) eliminate some invalid transitions so that all multi-
agent aspects of the domain are identified by potential conflicts in U(A,c). There are two
cases of multi-agent aspects of the domain; the table is fully lifted when both agents lift
from their respective ends at the same time, or when an agent lifts up the table from his
end and the table is already lifted up from the opposite end.
U(A,c) has 4 states, one for each position of the table. All the states satisfy lit-
erals ab(imp(l)) and ab(imp(r)). It has 6 transitions. Among the 6 transitions, 4
are from each state to itself with no performed actions. The remaining transitions are
from {table(onf loor)} to {table(lef tup)} and {table(rightup)} by compound ac-
tions {lift l } and {lift r }, respectively.6
One important remark is that the modeler may encode U(A,c) in a more compact
way. Our formalization is independent of how it is encoded. Given that U(A,c) covers
all valid states of the multi-agent domain and identifies potential conflicts for all desired
multi-agent aspects of the domain, it can be used in the second stage, where potential
conflicts are resolved and the global view is captured by M(A,c,r).
For the second stage of our formalization, we only show resolution of potential con-
flict related to the case which is about both agents lifting from their respective ends at
c) (s) where s = {table(onf loor)}.6
the same time. Observe that {lift l , lift r } ∈ χU(A
The dynamic laws (57) -- (59) in Dr effectively resolve this potential conflict. While laws
(58) and (59) defeat the laws in β(U(A,c)) related to individual effects of actions lift l
and lift r, (57) caused the effect fluent in the desired successor state.
,
table(lif ted) after lift l , lift r , table(onf loor)
ab ′(table(lef tup)) after lift l , lift r , table(onf loor)
ab ′(table(rightup)) after lift l , lift r , table(onf loor)
(57)
(58)
(59)
4 Discussion and Future Work
We have developed a formalization for capturing global view of MAS domains. Method-
ologically, we start representing the behaviour of each agent by an action description
6 Here we show only positive fluent literals satisfied in a state and omit the literals ab(imp(l))
and ab(imp(r)).
in BC from a single agent perspective. Then, a two-stage process guides the modeler
in composing these single agent descriptions into a single description representing the
global view of the overall MAS domain. While the modeler designates multi-agent as-
pects of the domain via potential conflicts in the first stage, she resolves these conflicts
according to the expected behaviour of the overall system in the second stage. Consid-
ering that representing single agent descriptions is relatively simpler than representing
multi-agent description directly, the formalization developed here is valuable from a
knowledge representation perspective and is different from earlier works using action
languages to represent MAS domains in a monolithic way.
Our choice of BC as the used action language is backed by its clean semantics
based on ASP and ability to express defeasible laws. Our formalization, however, is not
developed with a fixed action language in mind. One can use another action language
instead and update the methodology with less effort given that it can express defeasible
laws (e.g., C [11]).
The global view of a MAS domain is useful for carrying out reasoning tasks such as
projection, planning, or postdiction. We plan to extend our formalization with a query
language so that a modeler can represent such reasoning tasks. Using the query lan-
guage, for example, an initial and goal state of the domain can be represented to facili-
tate planning for the overall system.
The formalization introduced here paves the way for a number of avenues for fu-
ture work. An important one is to utilize online solving capacity of the ASP solver
clingo [12] to develop a complete multi-agent architecture based on logical represen-
tations. The solver may control execution in the environment using the formalization
of the global view. Moreover, this can be performed without restarting the solver from
scratch every time it communicates with the real agents residing in the environment
with the help of its online solving capacity.
References
1. Gelfond, M., Lifschitz, V.: Action languages. Electronic Transactions on Artificial Intelli-
gence 3(6) (1998) 193 -- 210
2. Baral, C., Gelfond, M.: Reasoning agents in dynamic domains. In Minker, J., ed.: Logic-
Based Artificial Intelligence. Kluwer Academic Publishers, Dordrecht (2000) 257 -- 279
3. Balduccini, M., Gelfond, M.: The autonomous agent architecture: An overview. In: Working
Notes of the AAAI Spring Symposium on Architectures for Intelligent Theory-Based Agents
(AITA08), Proceedings of the National Conference on Artificial Intelligence (AAAI) (2008)
1 -- 6
4. Baral, C., Gelfond, M.: Reasoning about effects of concurrent actions. The Journal of Logic
Programming 31(1) (1997) 85 -- 117
5. Baral, C., Son, T., Pontelli, E.: Reasoning about multi-agent domains using action language
C: A preliminary study. In Dix, J., Fisher, M., Nov´ak, P., eds.: Computational Logic in Multi-
Agent Systems - 10th International Workshop, CLIMA X, Hamburg, Germany, September
9-10, 2009, Revised Selected and Invited Papers. Volume 6214 of Lecture Notes in Computer
Science., Springer-Verlag (2009) 46 -- 63
6. Gelfond, G., Watson, R.: Modeling cooperative multi-agent systems.
In Costantini, S.,
Watson, R., eds.: Proceedings of the 4th Workshop Answer Set Programming: Advances in
Theory and Implementation (ASP'07). (2007) 67 -- 81
7. Lee, J., Lifschitz, V., Yang, F.: Action language BC: Preliminary report. In Rossi, F., ed.:
Proceedings of the Twenty-third International Joint Conference on Artificial Intelligence (IJ-
CAI'13), IJCAI/AAAI Press (2013) 983 -- 989
8. Chintabathina, S., Gelfond, M., Watson, R.: Defeasible laws, parallel actions, and reason-
ing about resources. In Amir, E., Lifschitz, V., Miller, R., eds.: Proceedings of the Eighth
International Symposium on Logical Formalization of Commonsense Reasoning (COM-
MONSENSE'07), Proceedings of the National Conference on Artificial Intelligence (AAAI)
(2007)
9. Gelfond, M., Lifschitz, V.: The stable model semantics for logic programming. In Kowalski,
R., Bowen, K., eds.: Proceedings of the Fifth International Conference and Symposium of
Logic Programming (ICLP'88), MIT Press (1988) 1070 -- 1080
10. Pednault, E.: Formulating multi-agent dynamic-world problems in the classical planning
framework. In Georgeff, M., Lansky, A., eds.: Reasoning about actions and plans, Morgan
Kaufmann Publishers (1987) 47 -- 82
11. Giunchiglia, E., Lifschitz, V.: An action language based on causal explanation: Preliminary
report. In: Proceedings of the National Conference on Artificial Intelligence (AAAI). (1998)
623 -- 630
12. Gebser, M., Kaminski, R., Kaufmann, B., Schaub, T.: Clingo = ASP + control: Preliminary
report.
In Leuschel, M., Schrijvers, T., eds.: Technical Communications of the Thirtieth
International Conference on Logic Programming (ICLP'14). Volume arXiv:1405.3694v1
of Theory and Practice of Logic Programming, Online Supplement. (2014) Available at
http://arxiv.org/abs/1405.3694v1.
|
1812.07887 | 1 | 1812 | 2018-12-19T11:32:46 | Hierarchical Macro Strategy Model for MOBA Game AI | [
"cs.MA",
"cs.AI"
] | The next challenge of game AI lies in Real Time Strategy (RTS) games. RTS games provide partially observable gaming environments, where agents interact with one another in an action space much larger than that of GO. Mastering RTS games requires both strong macro strategies and delicate micro level execution. Recently, great progress has been made in micro level execution, while complete solutions for macro strategies are still lacking. In this paper, we propose a novel learning-based Hierarchical Macro Strategy model for mastering MOBA games, a sub-genre of RTS games. Trained by the Hierarchical Macro Strategy model, agents explicitly make macro strategy decisions and further guide their micro level execution. Moreover, each of the agents makes independent strategy decisions, while simultaneously communicating with the allies through leveraging a novel imitated cross-agent communication mechanism. We perform comprehensive evaluations on a popular 5v5 Multiplayer Online Battle Arena (MOBA) game. Our 5-AI team achieves a 48% winning rate against human player teams which are ranked top 1% in the player ranking system. | cs.MA | cs | Hierarchical Macro Strategy Model for
MOBA Game AI
1Bin Wu, 1Qiang Fu, 1Jing Liang, 1Peng Qu, 1Xiaoqian Li, 1Liang Wang,
2Wei Liu, 1Wei Yang, 1Yongsheng Liu
1,2Tencent AI Lab
1{benbinwu, leonfu, masonliang, pengqu, xiaoqianli, enginewang, willyang, kakarliu}@tencent.com
[email protected]
8
1
0
2
c
e
D
9
1
]
A
M
.
s
c
[
1
v
7
8
8
7
0
.
2
1
8
1
:
v
i
X
r
a
Abstract
The next challenge of game AI lies in Real Time Strategy
(RTS) games. RTS games provide partially observable gam-
ing environments, where agents interact with one another in
an action space much larger than that of GO. Mastering RTS
games requires both strong macro strategies and delicate mi-
cro level execution. Recently, great progress has been made
in micro level execution, while complete solutions for macro
strategies are still lacking. In this paper, we propose a novel
learning-based Hierarchical Macro Strategy model for mas-
tering MOBA games, a sub-genre of RTS games. Trained
by the Hierarchical Macro Strategy model, agents explicitly
make macro strategy decisions and further guide their micro
level execution. Moreover, each of the agents makes indepen-
dent strategy decisions, while simultaneously communicat-
ing with the allies through leveraging a novel imitated cross-
agent communication mechanism. We perform comprehen-
sive evaluations on a popular 5v5 Multiplayer Online Battle
Arena (MOBA) game. Our 5-AI team achieves a 48% win-
ning rate against human player teams which are ranked top
1% in the player ranking system.
Introduction
Light has been shed on artificial general intelligence after
AlphaGo defeated world GO champion Lee Seedol (Silver
et al. 2016). Since then, game AI has drawn unprecedented
attention from not only researchers but also the public. Game
AI aims much more than robots playing games. Rather,
games provide ideal environments that simulate the real
world. AI researchers can conduct experiments in games,
and transfer successful AI ability to the real world.
Although AlphaGo is a milestone to the goal of general
AI, the class of problems it represents is still simple com-
pared to the real world. Therefore, recently researchers have
put much attention to real time strategy (RTS) games such
as Defense of the Ancients (Dota) (OpenAI 2018a) and Star-
Craft (Vinyals et al. 2017; Tian et al. 2017), which represents
a class of problems with next level complexity. Dota is a fa-
mous set of science fiction 5v5 Multiplayer Online Battle
Arena (MOBA) games. Each player controls one unit and
cooperate with four allies to defend allies' turrets, attack en-
emies' turrets, collect resources by killing creeps, etc. The
goal is to destroy enemies' base.
Copyright c(cid:13) 2019, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
There are four major aspects that make RTS games much
more difficult compared to GO: 1) Computational com-
plexity. The computational complexity in terms of action
space or state space of RTS games can be up to 1020,000,
while the complexity of GO is about 10250 (OpenAI 2018b).
2) Multi-agent. Playing RTS games usually involves mul-
tiple agents. It is crucial for multiple agents to coordinate
and cooporate. 3) Imperfect information. Different to GO,
many RTS games make use of fog of war (Vinyals et al.
2017) to increase game uncertainty. When the game map
is not fully observable, it is essential to consider gaming
among one another. 4) Sparse and delayed rewards. Learn-
ing upon game rewards in GO is challenging because the
rewards are usually sparse and delayed. RTS game length
could often be larger than 20,000 frames, while each GO
game is usually no more than 361 steps.
To master RTS games, players need to have strong skills
in both macro strategy operation and micro level execu-
tion. In recent study, much attention and attempts have
been put to micro level execution (Vinyals et al. 2017;
Tian et al. 2017; Synnaeve and Bessiere 2011; Wender and
Watson 2012). So far, Dota2 AI developed by OpenAI us-
ing reinforcement learning, i.e., OpenAI Five, has made the
most advanced progress (OpenAI 2018a). OpenAI Five was
trained directly on micro level action space using proxi-
mal policy optimization algorithms along with team rewards
(Schulman et al. 2017). OpenAI Five has shown strong
teamfights skills and coordination comparable to top pro-
fessional Dota2 teams during a demonstration match held
in The International 2018 (DOTA2 2018). OpenAI's ap-
proach did not explicitly model macro strategy and tried
to learn the entire game using micro level play. However,
OpenAI Five was not able to defeat professional teams due
to weakness in macro strategy management (Vincent 2018;
Simonite 2018).
Related work has also been done in explicit macro strat-
egy operation, mostly focused on navigation. Navigation
aims to provide reasonable destination spots and efficient
routes for agents. Most related work in navigation used in-
fluence maps or potential fields (DeLoura 2001; Hagelbäck
and Johansson 2008; do Nascimento Silva and Chaimow-
icz 2015). Influence maps quantify units using handcrafted
equations. Then, multiple influence maps are fused using
rules to provide a single-value output to navigate agents.
Providing destination is the most important purpose of nav-
igation in terms of macro strategy operation. The ability to
get to the right spots at right time makes essential differ-
ence between high level players and the others. Planning has
also been used in macro strategy operation. Ontanon et al.
proposed Adversarial Hierarchical-Task Network (AHTN)
Planning (Ontanón and Buro 2015) to search hierarchical
tasks in RTS game playing. Although AHTN shows promis-
ing results in a mini-RTS game, it suffers from efficiency
issue which makes it difficult to apply to full MOBA games
directly.
Despite of the rich and promising literature, previous
work in macro strategy failed to provide complete solution:
First, reasoning macro strategy implicitly by learning
upon micro level action space may be too difficult. OpenAI
Five's ability gap between micro level execution and macro
strategy operation was obvious. It might be over-optimistic
to leave models to figure out high level strategies by sim-
ply looking at micro level actions and rewards. We consider
explicit macro strategy level modeling to be necessary.
Second, previous work on explicit macro strategy heavily
relied on handcrafted equations for influence maps/potential
fields computation and fusion. In practice, there are usu-
ally thousands of numerical parameters to manually decide,
which makes it nearly impossible to achieve good perfor-
mance. Planning methods on the other hand cannot meet ef-
ficiency requirement of full MOBA games.
Third, one of the most challenging problems in RTS game
macro strategy operation is coordination among multiple
agents. Nevertheless, to the best of our knowledge, previ-
ous work did not consider it in an explicit way. OpenAI Five
considers multi-agent coordination using team rewards on
micro level modeling. However, each agent of OpenAI Five
makes decision without being aware of allies' macro strat-
egy decisions, making it difficult to develop top coordination
ability in macro strategy level.
Finally, we have found that modeling strategic phase is
crucial for MOBA game AI performance. However, to the
best of our knowledge, previous work did not consider this.
Teaching agents to learn macro strategy operation, how-
ever, is challenging. Mathematically defining macro strat-
egy, e.g., besiege and split push, is difficult in the first place.
Also, incorporating macro strategy on top of OpenAI Five's
reinforcement learning framework (OpenAI 2018a) requires
corresponding execution to gain rewards, while macro strat-
egy execution is a complex ability to learn by itself. There-
fore, we consider supervised learning to be a better scheme
because high quality game replays can be fully leveraged
to learn macro strategy along with corresponding execution
samples. Note that macro strategy and execution learned us-
ing supervised learning can further act as an initial policy for
reinforcement learning.
In this paper, we propose Hierarchical Macro Strat-
egy (HMS) model - a general supervised learning frame-
work for MOBA games such as Dota. HMS directly tack-
les with computational complexity and multi-agent chal-
lenges of MOBA games. More specifically, HMS is a hi-
erarchical model which conducts macro strategy operation
by predicting attention on the game map under guidance
of game phase modeling. Thereby, HMS reduces computa-
tional complexity by incorporating game knowledge. More-
over, each HMS agent conducts learning with a novel mech-
anism of communication with teammates agents to cope
with multi-agent challenge. Finally, we have conducted ex-
tensive experiments in a popular MOBA game to evaluate
our AI ability. We matched with hundreds of human player
teams that ranked above 99% of players in the ranked system
and achieved 48% winning rate.
The rest of this paper is organized as follows: First, we
briefly introduce Multiplayer Online Battle Arena (MOBA)
games and compare the computational complexity with
GO. Second, we illustrate our proposed Hierarchical Macro
Strategy model. Then, we present experimental results in the
fourth section. Finally, we conclude and discuss future work.
Multiplayer Online Battle Arena (MOBA)
Games
Game Description
MOBA is currently the most popular sub-genre of the RTS
games. MOBA games are responsible for more than 30%
of the online gameplay all over the world, with titles such
as Dota, League of Legends, and Honour of Kings (Mur-
phy 2015). According to a worldwide digital games market
report in February 2018, MOBA games ranked first in gross-
ing in both PC and mobile games (SuperData 2018).
In MOBA, the standard game mode requires two 5-player
teams play against each other. Each player controls one unit,
i.e., hero. There are numerous of heroes in MOBA, e.g.,
more than 80 in Honour of Kings. Each hero is uniquely
designed with special characteristics and skills. Players con-
trol movement and skill releasing of heroes via the game
interface.
As shown in Figure. 1a, Honour of Kings players use left
bottom steer button to control movements, while right bot-
tom set of buttons to control skills. Surroundings are ob-
servable via the main screen. Players can also learn full map
situation via the left top corner mini-map, where observable
turrets, creeps, and heroes are displayed as thumbnails. Units
are only observable either if they are allies' units or if they
are within a certain distance to allies' units.
There are three lanes of turrets for each team to defend,
three turrets in each lane. There are also four jungle areas on
the map, where creep resources can be collected to increase
gold and experience. Each hero starts with minimum gold
and level 1. Each team tries to leverage resources to obtain
as much gold and experience as possible to purchase items
and upgrade levels. The final goal is to destroy enemy's base.
A conceptual map of MOBA is shown in Figure. 1b.
To master MOBA games, players need to have both excel-
lent macro strategy operation and proficient micro level exe-
cution. Common macro strategies consist of opening, laning,
ganking, ambushing, etc. Proficient micro level execution re-
quires high accuracy of control and deep understanding of
damage and effects of skills. Both macro strategy operation
and micro level execution require mastery of timing to excel,
which makes it extremely challenging and interesting. More
(a)
(b)
Figure 1: (a) Game UI of Honour of Kings. Players use left bottom steer button to control movements, while right bottom set of
buttons to control skills. Players can observe surroundings via the screen and view the mini full map using the left top corner.
(b) An example map of MOBA. The two teams are colored in blue and red, each possesses nine turrets (circled in rounds) and
one base (circled in squares). The four jungle areas are numbered from 1 to 4.
Table 1: Computational complexity comparison between
GO and MOBA.
Action
Space
State
Space
GO
250150 ≈ 10360 (250
pos available, 150
decisions per game
in average)
3360 ≈ 10170 (361
pos, 3 states each)
MOBA
101500 (10 options,
1500 actions per
game)
1020000 (10 heroes,
2000+pos *
10+states)
discussion of MOBA can be found in (Silva and Chaimow-
icz 2017).
Next, we will quantify the computational complexity of
MOBA using Honour of Kings as an example.
Computational Complexity
The normal game length of Honour of Kings is about
20 minutes, i.e., approximately 20,000 frames in terms of
gamecore. At each frame, players make decision with tens
of options, including movement button with 24 directions,
and a few skill buttons with corresponding releasing posi-
tion/directions. Even with significant discretization and sim-
plification, as well as reaction time increased to 200ms, the
action space is at magnitude of 101,500.
As for state space, the resolution of Honour of Kings map
is 130,000 by 130,000 pixels, and the diameter of each unit
is 1,000. At each frame, each unit may have different status
such as hit points, levels, gold. Again, the state space is at
magnitude of 1020,000 with significant simplification.
Comparison of action space and state space between
MOBA and GO is listed in Table. 1.
MOBA AI Macro Strategy Architecture
Our motivation of designing MOBA AI macro strategy
model was inspired from how human players make strategic
decisions. During MOBA games, experienced human play-
ers are fully aware of game phases, e.g., opening phase, lan-
ing phase, mid game phase, and late game phase (Silva and
Chaimowicz 2017). During each phase, players pay atten-
tion to the game map and make corresponding decision on
where to dispatch the heroes. For example, during the laning
phase players tend to focus more on their own lanes rather
than backing up allies, while during mid to late phases, play-
ers pay more attention to teamfight spots and pushing ene-
mies' base.
To sum up, we formulate the macro strategy operation
process as "phase recognition -> attention prediction -> ex-
ecution". To model this process, we propose a two-layer
macro strategy architecture, i.e., phase and attention:
• Phase layer aims to recognize current game phase so that
attention layer can have better sense about where to pay
attention to.
• Attention layer aims to predict the best region on game
maps to dispatch heroes.
Phase and Attention layers act as high level guidance for
micro level execution. We will describe details of model-
ing in the next section. The network structure of micro level
model is almost identical to the one used in OpenAI Five1
(OpenAI 2018a), but in a supervised learning manner. We
did minor modification to adapt it to Honour of Kings, such
as deleting Teleport.
Hierarchical Macro Strategy Model
We propose a Hierarchical Macro Strategy (HMS) model to
consider both phase and attention layers in a unified neural
network. We will first present the unified network architec-
ture. Then, we illustrate how we construct each of the phase
and attention layers.
1https://d4mucfpksywv.cloudfront.
net/research-covers/openai-five/
network-architecture.pdf
Model Overview
We propose a Hierarchical Macro Strategy model (HMS) to
model both attention and phase layers as a multi-task model.
It takes game features as input. The output consists of two
tasks, i.e., attention layer as the main task and phase layer as
an auxiliary task. The output of attention layer directly con-
veys macro strategy embedding to micro level models, while
resource layer acts as an axillary task which help refine the
shared layers between attention and phase tasks.
The illustrating network structure of HMS is listed in Fig-
ure. 2. HMS takes both image and vector features as in-
put, carrying visual features and global features respectively.
In image part, we use convolutional layers. In vector part,
we use fully connected layers. The image and vector parts
merge in two separate tasks, i.e., attention and phase. Ul-
timately, attention and phase tasks take input from shared
layers through their own layers and output to compute loss.
Attention Layer
Similar to how players make decisions according to the
game map, attention layer predicts the best region for agents
to move to. However, it is tricky to tell from data that where
is a player's destination. We observe that regions where at-
tack takes place can be indicator of players' destination, be-
cause otherwise players would not have spent time on such
spots. According to this observation, we define ground-truth
regions as the regions where players conduct their next at-
tack. An illustrating example is shown in Figure. 3.
Let s to be one session in a game which contains several
frames, and s − 1 indicates the session right before s. In
Figure. 3, s − 1 is the first session in the game. Let ts to
be the starting frame of s. Note that a session ends along
with attack behavior, therefore there exists a region ys in ts
where the hero conducts attack. As shown in Figure. 3, label
for s−1 is ys, while label for s is ys+1. Intuitively, by setting
up labels in this way, we expect agents to learn to move to
ys at the beginning of game. Similarly, agents are supposed
to move to appropriate regions given game situation.
Phase layer
Phase layer aims to recognize the current phase. Extracting
game phases ground-truth is difficult because phase defini-
tion used by human players is abstract. Although roughly
correlated to time, phases such as opening, laning, and late
game depend on complicated judgment based on current
game situation, which makes it difficult to extract ground-
truth of game phases from replays. Fortunately, we observe
clear correlation between game phases with major resources.
For example, during the opening phase players usually aim
at taking outer turrets and baron, while for late game, players
operate to destroy enemies' base.
Therefore, we propose to model phases with respect to
major resources. More specifically, major resources indicate
turrets, baron, dragon, and base. We marked the major re-
sources on the map in Figure. 4a. Label definition of phase
layer is similar to attention layer. The only difference is that
ys in phase layer indicates attack behavior on turrets, baron,
dragon, and base instead of in regions. Intuitively, phase
layer modeling splits the entire game into several phases via
modeling which macro resource to take in current phase.
We do not consider other resources such as lane creeps,
heroes, and neutral creeps as major objectives because usu-
ally these resources are for bigger goal, such as destroying
turrets or base with higher chance. Figure. 4b shows a series
of attack behavior during the bottom outer turret strategy.
The player killed two neutral creeps in the nearby jungle
and several lane creeps in the bottom lane before attacking
the bottom outer turret.
We expect the model to learn when and what major re-
sources to take given game situation, and in the meanwhile
learn attention distribution that serve each of the major re-
sources.
Imitated Cross-agents Communication
Cross-agents communication is essential for a team of
agents to cooperate. There is rich literature of cross-agent
communication on multi-agent reinforcement learning re-
search (Sukhbaatar, Fergus, and others 2016; Foerster et al.
2016). However, it is challenging to learn communication
using training data in supervised learning because the actual
communication is unknown.
To enable agents to communicate in supervised learning
setting, we have designed a novel cross-agents communi-
cation mechanism. During training phase, we put attention
labels of allies as features for training. During testing phase,
we put attention prediction of allies as features and make
decision correspondingly. In this way, our agents can "com-
municate" with one another and learn to cooperate upon al-
lies' decisions. We name this mechanism as Imitated Cross-
agents Communication due to its supervised nature.
Experiments
In this section, we evaluate our model performance. We first
describe the experimental setup, including data preparation
and model setup. Then, we present qualitative results such
as attention distribution under different phase. Finally, we
list the statistics of matches with human player teams and
evaluate improvement brought by our proposed model.
Experimental Setup
Data Preparation To train a model, we collect around 300
thousand game replays made of King Professional League
competition and training records. Finally, 250 million in-
stances were used for training. We consider both visual and
attributes features. On visual side, we extract 85 features
such as position and hit points of all units and then blur the
visual features into 12*12 resolution. On attributes side, we
extract 181 features such as roles of heroes, time period of
game, hero ID, heroes' gold and level status and Kill-Death-
Assistance statistics.
Model Setup We use a mixture of convolutional and
fully-connected layers to take inputs from visual and at-
tributes features respectively. On convolutional side, we set
five shared convolutional layers, each with 512 channels,
padding = 1, and one RELU. Each of the tasks has two con-
volutional layers with exactly the same configuration with
Figure 2: Network Architecture of Hierarchical Macro Strategy Model
Figure 3: Illustrating example for label extraction in atten-
tion layer.
shared layers. On fully-connected layers side, we set two
shared fully-connected layers with 512 nodes. Each of the
tasks has two fully-connected layers with exactly the same
configuration with shared layers. Then, we use one concate-
nation layer and two fully-connected layers to fuse results
of convolutional layers and fully-connected layers. We use
ADAM as the optimizer with base learning rate at 10e-6.
Batch size was set at 128. The loss weights of both phase
and attention tasks are set at 1. We used CAFFE (Jia et al.
2014) with eight GPU cards. The duration to train an HMS
model was about 12 hours.
Finally, the output for attention layer corresponds to 144
regions of the map, resolution of which is exactly the same
as the visual inputs. The output of the phase task corresponds
to 14 major resources circled in Figure. 4a.
Experimental Results
Opening Attention Opening is one of the most important
strategies in MOBA. We show one opening attention of dif-
ferent heroes learned by our model in Figure. 5. In Figure.
5, each subfigure consists of two square images. The left-
hand-side square image indicates the attention distribution
of the right-hand-side MOBA mini-map. The hottest region
is highlighted with red circle. We list attention prediction of
four heroes, i.e., Diaochan, Hanxin, Arthur, and Houyi. The
four heroes belong to master, assasin, warrior, and archer re-
spectively. According to the attention prediction, Diaochan
is dispatched to middle lane, Hanxin will move to left jun-
gle area, and Authur and Houyi will guard the bottom jungle
area. The fifth hero Miyamoto Musashi, which was not plot-
ted, will guard the top outer turret. This opening is consid-
ered safe and efficient, and widely used in Honour of Kings
games.
Attention Distribution Affected by Phase Layer We vi-
sualize attention distribution of different phases in Figure.
6a and 6b. We can see that attention distributes around the
major resource of each phase. For example, for upper outer
turret phase in Figure. 6a, the attention distributes around
upper outer region, as well as nearby jungle area. Also, as
shown in Figure. 6b, attention distributes mainly in the mid-
dle lane, especially area in front of the base. These examples
show that our phase layer modeling affects attention distri-
bution in practice. To further examine how phase layer cor-
relates with game phases, we conduct t-Distributed Stochas-
tic Neighbor Embedding (t-SNE) on phase layer output. As
shown in Figure. 7, samples are coloured with respect to dif-
ferent time stages. We can observe that samples are clearly
separable with respect to time stages. For example, blue, or-
ange and green (0-10 minuets) samples place close to one
another, while red and purple samples (more than 10 min-
uets) form another group.
Macro Strategy Embedding We evaluate how important
is the macro strategy modeling. We removed the macro strat-
egy embedding and trained the model using micro level ac-
tions from the replays. The micro level model design is sim-
ilar to OpenAI Five (OpenAI 2018a). Detail description of
the micro level modeling is out of the scope of this paper.
The result is listed in Table. 2, column AI Without Macro
Strategy. As the result shows, HMS outperformed AI With-
out Macro Strategy with 75% winning rates. HMS per-
formed much better than AI Without Macro Strategy in
terms of number of kills, turrets destruction, and gold. The
most obvious performance change is that AI Without Macro
Strategy mainly focused on nearby targets. Agents did not
(a)
(b)
Figure 4: (a) Major resources (circled, i.e., turrets, base, dragon, and baron) modeled in phase layer. (b) Illustrating example for
label extraction in phase layer.
Figure 5: One of the opening strategies learned for different hero roles. The hottest regions are highlighted with red circle.
(a)
(b)
Figure 6: Attention distribution of different strategies. The
two attention figures describe attention distribution of the
two major resources, i.e., upper outer turret and base respec-
tively.
care much about backing up teammates and pushing lane
creeps in relatively large distance. They spent most of the
time on killing neutral creeps and nearby lane creeps. The
performance change can be observed from the compari-
son of engagement rate and number of turrets in Table. 2.
This phenomenon may reflect how important macro strategy
modeling is to highlight important spots.
Match against Human Players To evaluate our AI per-
formance more accurately, we conduct matches between our
AI and human players. We invited 250 human player teams
whose average ranking is King in Honour of Kings rank sys-
tem (above 1% of human players). Following the standard
procedure of ranked match in Honour of Kings, we obey
ban-pick rules to pick and ban heroes before each match.
The ban-pick module was implemented using simple rules.
Note that gamecores of Honour of Kings limit commands
frequency to a level similar with human.
The overall statistics are listed in Table. 2, column Human
Teams. Our AI achieved 48% winning rate in the 250 games.
The statistics show that our AI team did not have advantage
on teamfight over human teams. The number of kills made
by AI is about 15% less than human teams. Other items such
as turrets destruction, engagement rate, and gold per minute
were similar between AI and human. We have further ob-
served that our AI destroyed 2.5 more turrets than human on
average in the first 10 minutes. After 10 minutes, turrets dif-
ference shrank due to weaker teamfight ability compared to
human teams. Arguably, our AI's macro strategy operation
ability is close to or above our human opponents.
Imitated Cross-agents Communication To evaluate how
important the cross-agents communication mechanism is to
the AI ability, we conduct matches between HMS and HMS
trained without cross-agents communication. The result is
listed in Table. 2, column AI Without Communication. HMS
achieved a 62.5% winning rate over the version without
communication. We have observed obvious cross-agents co-
operation learned when cross-agents communication was in-
troduced. For example, rate of reasonable opening increased
from 22% to 83% according to experts' evaluation.
Figure 7: t-Distributed Stochastic Neighbor Embedding on phase layer output. Embedded data samples are coloured with
respect to different time stages.
Table 2: Match statistics. 250 games were played against Human Teams, while 40 games were played against Without Macro
Strategy, Without Communication, and Without Phase Layer, respectively.
Opponents
Winning rate
Kill
Game Length
Gold/Min
Engagement Rate
Turrets
Dragons
Barons
Dark Barons
AI Without Macro
Strategy
75% - 25%
26.0 - 21.1
16.1 min
2399 - 2287
49% - 42%
6.1 - 3.2
1.22 - 0.2
0.62 - 0.31
0.41 - 0.22
Human Teams
48.3% - 51.7%
22.6 - 26.3
16.1 min
2603 - 2616
48% - 48%
6.1 - 6.2
0.55 - 0.55
0.64 - 0.61
0.36 - 0.38
AI Without
Communication
62.5% - 37.5%
19.9 - 19.4
18.2 min
2633 - 2554
49% - 47%
6.21 - 5.26
0.65 - 0.49
0.45 - 0.41
0.35 - 0.32
AI Without Phase Layer
65% - 35%
25.6 - 22.8
18.2 min
2500 - 2333
50% - 49%
6.73 - 5.42
1 - 0.41
0.71 - 0.2
0.49 - 0.04
Phase layer We evaluate how phase layer affects the per-
formance of HMS. We removed the phase layer and com-
pared it with the full version of HMS. The result is listed in
Table. 2, column AI Without phase layer. The result shows
that phase layer modeling improved HMS significantly with
65% winning rate. We have also observed obvious AI abil-
ity downgrade when phase layer was removed. For example,
agents were no longer accurate about timing when baron first
appears, while the full version HMS agents got ready at 2:00
to gain baron as soon as possible.
Conclusion and Future Work
In this paper, we proposed a novel Hierarchical Macro
Strategy model which models macro strategy operation for
MOBA games. HMS explicitly models agents' attention
on game maps and considers game phase modeling. We
also proposed a novel imitated cross-agent communication
mechanism which enables agents to cooperate.
We used Honour of Kings as an example of MOBA games
to implement and evaluate HMS. We conducted matches
between our AI and top 1% human player teams. Our AI
achieves a 48% winning rate. To the best of our knowledge,
our proposed HMS model is the first learning based model
that explicitly models macro strategy for MOBA games.
HMS used supervised learning to learn macro strategy op-
eration and corresponding micro level execution from high
quality replays. A trained HMS model can be further used
as an initial policy for reinforcement learning framework.
Our proposed HMS model exhibits a strong potential in
MOBA games. It may be generalized to more RTS games
with appropriate adaptations. For example, the attention
layer modeling may be applicable to StarCraft, where the
definition of attention can be extended to more meaningful
behaviors such as building operation. Also, Imitated Cross-
agents Communication can be used to learn to cooperate.
Phase layer modeling is more game-specific. The resource
collection procedure in StarCraft is different from that of
MOBA, where gold is mined near the base. Therefore, phase
layer modeling may require game-specific design for differ-
ent games. However, the underlying idea to capture game
phases can be generalized to Starcraft as well.
HMS may also inspire macro strategy modeling in do-
mains where multiple agents cooperate on a map and histor-
ical data is available. For example, in robot soccer, attention
layer modeling and Imitated Cross-agents Communication
may help robots position and cooperate given parsed soccer
off
digital
games
market:
2018.
2018.
ai
bots -- for
February
World-
2018.
elon musk-backed
Pro gamers
now.
[Simonite 2018] Simonite, T.
fend
https://www.wired.com/story/pro-gamers-fend-off-elon-
musks-ai-bots/ (Aug 23, 2018).
[Sukhbaatar, Fergus, and others 2016] Sukhbaatar, S.; Fer-
gus, R.; et al. 2016. Learning multiagent communication
with backpropagation. In Advances in Neural Information
Processing Systems, 2244 -- 2252.
[SuperData 2018] SuperData.
wide
https://www.superdataresearch.com/us-digital-games-
market/.
[Synnaeve and Bessiere 2011] Synnaeve, G., and Bessiere,
P. 2011. A bayesian model for rts units control applied to
starcraft. In Computational Intelligence and Games (CIG),
2011 IEEE Conference on, 190 -- 196. IEEE.
[Tian et al. 2017] Tian, Y.; Gong, Q.; Shang, W.; Wu, Y.; and
Zitnick, C. L. 2017. Elf: An extensive, lightweight and
flexible research platform for real-time strategy games. In
Advances in Neural Information Processing Systems, 2656 --
2666.
[Vincent 2018] Vincent, J.
Humans grab vic-
tory in first of three dota 2 matches against openai.
https://www.theverge.com/2018/8/23/17772376/openai-
dota-2-pain-game-human-victory-ai (Aug 23, 2018).
[Vinyals et al. 2017] Vinyals, O.; Ewalds, T.; Bartunov, S.;
Georgiev, P.; Vezhnevets, A. S.; Yeo, M.; Makhzani, A.; Küt-
tler, H.; Agapiou, J.; Schrittwieser, J.; et al. 2017. Star-
craft ii: a new challenge for reinforcement learning. arXiv
preprint arXiv:1708.04782.
[Wender and Watson 2012] Wender, S., and Watson, I. 2012.
Applying reinforcement learning to small scale combat in
the real-time strategy game starcraft: Broodwar. In Compu-
tational Intelligence and Games (CIG), 2012 IEEE Confer-
ence on, 402 -- 408. IEEE.
2018.
recordings.
In the future, we will incorporate planning based on HMS.
Planning by MCTS roll-outs in Go has been proven essen-
tial to outperform top human players (Silver et al. 2016). We
expect planning can be essential for RTS games as well, be-
cause it may not only be useful for imperfect information
gaming but also be crucial to bringing in expected rewards
which supervised learning fails to consider.
References
2018. The international 2018.
[DeLoura 2001] DeLoura, M. A. 2001. Game programming
gems 2. Cengage learning.
Nasci-
[do Nascimento Silva and Chaimowicz 2015] do
2015. On the
mento Silva, V., and Chaimowicz, L.
development of intelligent agents for moba games.
In
Computer Games and Digital Entertainment (SBGames),
2015 14th Brazilian Symposium on, 142 -- 151. IEEE.
[DOTA2 2018] DOTA2.
https://www.dota2.com/international/announcement/.
[Foerster et al. 2016] Foerster, J. N.; Assael, Y. M.; de Fre-
itas, N.; and Whiteson, S. 2016. Learning to communicate
to solve riddles with deep distributed recurrent q-networks.
arXiv preprint arXiv:1602.02672.
[Hagelbäck and Johansson 2008] Hagelbäck, J., and Johans-
son, S. J. 2008. The rise of potential fields in real time
strategy bots. In Fourth Artificial Intelligence and Interac-
tive Digital Entertainment Conference. Stanford University.
[Jia et al. 2014] Jia, Y.; Shelhamer, E.; Donahue, J.; Karayev,
S.; Long, J.; Girshick, R.; Guadarrama, S.; and Darrell, T.
2014. Caffe: Convolutional architecture for fast feature em-
bedding. arXiv preprint arXiv:1408.5093.
2015. Most played games:
[Murphy 2015] Murphy, M.
November 2015 -- fallout 4 and black ops iii arise while
starcraft ii shines. http://caas.raptr.com/most-played-games-
november-2015-fallout-4-andblack-ops-iii-arise-while-
starcraft-ii-shines/.
[Ontanón and Buro 2015] Ontanón, S., and Buro, M. 2015.
Adversarial hierarchical-task network planning for complex
real-time games. In Twenty-Fourth International Joint Con-
ference on Artificial Intelligence.
[OpenAI 2018a] OpenAI.
https://blog.openai.com/dota-2/ (17 Apr 2018).
[OpenAI 2018b] OpenAI.
https://blog.openai.com/openai-five/ (25 Jun 2018).
[Schulman et al. 2017] Schulman, J.; Wolski, F.; Dhariwal,
P.; Radford, A.; and Klimov, O. 2017. Proximal policy op-
timization algorithms. arXiv preprint arXiv:1707.06347.
[Silva and Chaimowicz 2017] Silva, V. D. N., and Chaimow-
icz, L. 2017. Moba: a new arena for game ai. arXiv preprint
arXiv:1705.10443.
[Silver et al. 2016] Silver, D.; Huang, A.; Maddison, C. J.;
Guez, A.; Sifre, L.; Van Den Driessche, G.; Schrittwieser,
J.; Antonoglou, I.; Panneershelvam, V.; Lanctot, M.; et al.
2016. Mastering the game of go with deep neural networks
and tree search. nature 529(7587):484 -- 489.
2018a. Openai blog: Dota 2.
Openai five.
2018b.
|
1708.01925 | 1 | 1708 | 2017-08-06T19:24:00 | Designing Autonomous Vehicles: Evaluating the Role of Human Emotions and Social Norms | [
"cs.MA",
"cs.AI",
"cs.CY",
"cs.RO",
"eess.SY"
] | Humans are going to delegate the rights of driving to the autonomous vehicles in near future. However, to fulfill this complicated task, there is a need for a mechanism, which enforces the autonomous vehicles to obey the road and social rules that have been practiced by well-behaved drivers. This task can be achieved by introducing social norms compliance mechanism in the autonomous vehicles. This research paper is proposing an artificial society of autonomous vehicles as an analogy of human social society. Each AV has been assigned a social personality having different social influence. Social norms have been introduced which help the AVs in making the decisions, influenced by emotions, regarding road collision avoidance. Furthermore, social norms compliance mechanism, by artificial social AVs, has been proposed using prospect based emotion i.e. fear, which is conceived from OCC model. Fuzzy logic has been employed to compute the emotions quantitatively. Then, using SimConnect approach, fuzzy values of fear has been provided to the Netlogo simulation environment to simulate artificial society of AVs. Extensive testing has been performed using the behavior space tool to find out the performance of the proposed approach in terms of the number of collisions. For comparison, the random-walk model based artificial society of AVs has been proposed as well. A comparative study with a random walk, prove that proposed approach provides a better option to tailor the autopilots of future AVS, Which will be more socially acceptable and trustworthy by their riders in terms of safe road travel. | cs.MA | cs | Designing Autonomous Vehicles: Evaluating the Role of Human Emotions and
Social Norms
Faisal Riaz1, Muaz Niazi 2,*
1Dept. Of Computing-Iqra University, Islamabad, Pakistan
2Dept. Of Computer Sciences-COMSATS, Islamabad, Pakistan
[email protected]
*[email protected]
Abstract
Humans are going to delegate the rights of driving to the autonomous vehicles in near future.
However, to fulfill this complicated task, there is a need for a mechanism, which enforces the
autonomous vehicles to obey the road and social rules that have been practiced by well-behaved
drivers. This task can be achieved by introducing social norms compliance mechanism in the
autonomous vehicles. This research paper is proposing an artificial society of autonomous
vehicles as an analogy of human social society. Each AV has been assigned a social personality
having different social influence. Social norms have been introduced which help the AVs in
making the decisions, influenced by emotions, regarding road collision avoidance. Furthermore,
social norms compliance mechanism, by artificial social AVs, has been proposed using prospect
based emotion i.e. fear, which is conceived from OCC model. Fuzzy logic has been employed to
compute the emotions quantitatively. Then, using SimConnect approach, fuzzy values of fear has
been provided to the Netlogo simulation environment to simulate artificial society of AVs.
Extensive testing has been performed using the behavior space tool to find out the performance
of the proposed approach in terms of the number of collisions. For comparison, the random-walk
model based artificial society of AVs has been proposed as well. A comparative study with a
random walk, prove that proposed approach provides a better option to tailor the autopilots of
1
future AVS, Which will be more socially acceptable and trustworthy by their riders in terms of
safe road travel.
Key Word: - Autonomous Vehicles, Social Norms, OCC Model, Prospect Based Emotions, Fuzzy Logic, Netlogo
1. Introduction
Autonomous road vehicles (ARVs) have been considered better than human driven vehicles in
road safety and traffic management. According to the investigation of Riaz and Niazi [1],
autonomous vehicles have been found helpful in decreasing road accidents as compared to the
human-driven vehicles. Furthermore, the issue of road jams can be solved by replacing human
drivers with fully connected autonomous cars, as noted by Litman [2]. In addition, Mersky and
Samaras [3] have illustrated that AVs are very helpful in decreasing the road pollution and
making the environment green. From these benefits, it might be possible that the law agencies
delegate the driving task to AVs by issuing them the driving license. However, to fulfill this
complicated task of driving, there is a need for a mechanism, which enforces the autonomous
vehicles, which is a robot, to obey the road and social rules that have been practiced by well-
behaved drivers.
The Role of ethics and social norms have been considered important in making robots social,
well behaved and more compatible with humans. According to Malle [4], robots can serve as
competent social agents by integrating moral norms in their basic architecture. A. Rakotonirainy
et al. [5] have proven that social norms can be utilized to design human compatible social AVs
robots. According to Kummer et al. [6], social norms can be used in tailoring crash free AVs
robot by operating on roads wisely. From the above discussion, it is implied that social norms
with some norms compliance mechanism can be used to tailor the next generation of more
trustworthy social AVs.
2
Emotions can be used as norms compliance mechanism as it is already proven that emotions help
in sustaining the social norms in human society. According to Elster [7], self-attribution
emotions like shame helps the human to avoid the violation due to fear of losing their social
status. According to N. Criado [8], prospect based emotions like fear enforce the human to
follow the social norms in order to avoid the punishment from the law enforcement agencies.
Inspired by the role of emotions in social norm compliance in the human society, researchers
have used emotions to enforce the artificial agents' norm compliance. According to staller et al.
[9]emotions act as an important factor in the sustainability of social norms. Hence, it is implied
that we can use emotions as the norms compliance mechanism to design social norms enabled
AVs. Gerdes and Thornton [10] have suggested the mathematical model of social norms for
designing the control algorithms of AVs but still their works lack the simulation or proof of
concept of proposed mathematical models.
Problem statement- However, to the best of our knowledge the existing literature has not
proposed such procedures that allow AVs to configure their autopilots to make collision
avoidance decisions about norm compliance using emotional motivations as human drivers
would do. For example, Amitai and Oren[11], just suggested the use of social norms in AVs in
the
theoretical aspect without discussing any working mathematical model and
its
implementation aspects. A. Rakotonirainy et al. [5] have been proposed a novel concept of
measuring the emotional state of a driver using the HUD-UP technology and transmitting the
social norms from driver to driver to modify the behavior behind the steering of AV. However,
the concept of social norms has not been integrated into the autopilot of the AV, which helps
them to make the collision avoidance decisions by their own. The major challenge for AVs is
that how it will take decisions at the time of the crashes and this issue has been addressed by the
3
Kumfer and Burgess in [6]. The authors have used social norms as a decision mechanism to
choose a less harming crash among possible collision options. However, this paper does not
provide any collision avoidance strategy using some social norms compliance mechanism, which
avoids the collision situations.
Contribution - The existing research work is proposing a set of contributions in building the
norm compliance collision free artificial society of Autonomous vehicles inspired by human
society social norms and related emotions. Our aim is to provide humans with reliable AVs to
which they can delegate driving tasks that are regulated by legal and social norms. The main
contributions of the paper are given as.
• Viewpoint of incorporating the social norms and emotions in Autonomous vehicles and
conceiving them as artificial social entities
• Modeling of social norms inspired artificial society of Autonomous Vehicles
• Modeling of prospected based emotions to make AVs emotions enabled
• Simulation of social norms compliance artificial society of AVs using the Net logo
• Detailed experiment design
• Rigor analysis ,in terms of number of collisions, of the proposed approach in the
comparison with random walk travelling strategy
This article is organized as follows: Section 2 illustrates the literature review; Section 3 describes
the method; Section 4 provides the description of the proposed model; section 5 illustrates the
experiments; section 6 elaborates the results and discussion and section 7 contains conclusion.
2.-Literature Review
In this section, detailed literature review related to the proposed scheme has been performed. The
literature review has been divided into three main categories. The first category addresses the
4
literature supporting the role of ethics in robots using theoretical debate. The second category
discusses the literature that supports the role of using ethics or norms in the design of AVs but
only theoretically. Then the third category discusses the state of the art literature, which has used
the social norms in autonomous vehicles using a simulation approach.
According to Voort et al. [12] computers are getting autonomous day by day and are capable of
making decisions of their own. Intelligent computer systems can get information from human,
analyze it, take decisions and store that information or provide it to third parties. There is a need
to check the moral values of computer decisions. Authors have suggested that there is a need to
add ethics in technology, which is still lacking behind. Malle1 [4] summarized that from 1995 to
2015 very little efforts have been made on the implementation of ethics in robots. Past studies
provide a thought that whether a robot could be a moral agent or not. In addition, researchers
found that robot could be treated as a living thing that can take actions on its own decisions, and
it can decide what is right and what is wrong with humans.
According to Amitai and Oren [13], the latest smart machines like AVs are getting
smarter due to the incorporation of artificial intelligence (AI) algorithms. Furthermore, these
AVs are becoming more and more autonomous in the sense that they are now taking decisions on
their own using these AI algorithms. The authors suggest that as these AVs basic purpose is to
serve humans, and then there is a need to equip them with ethical and social rules so that the
autonomous devices like AVs can take decisions of their own that could not harm the passengers
and other road commuters [7]. According to Gerdes and Thornton [10], it is the responsibility of
researchers and programmers to devise ethics enabled control algorithms for AVs that make
them more acceptable to human society. The authors argue that the incorporation of ethics of the
society in which AVs are operating will help the court of law to decide the responsibility level of
5
AV in the case of an accident. In this regard, they have proposed a mathematical model of
ethical frameworks to incorporate them in the control algorithms of Autonomous vehicles. The
proposed model can read the error rate for the actual and desired path of the car based on
different constraints. However, the authors have not mentioned any case study that implements
any of the proposed mathematical model using simulation or real field tests.
Social and autonomous robots are the motivation to build social cars so that road
accidents can be eliminated. Vehicle-to-vehicle (V2V) is a sub-part of intelligent transportation
system (ITS) equipped with sensing technologies and wireless communication system is helpful
in road accident prevention. A. Rakotonirainy et al. [14] have been proposed a novel concept that
HUDs , Human-Computer-interaction, HCI and communicating social information between cars
can provide social awareness and he named it as 'social car'. This social car can sense the
driving behavior of driver by capturing the facial expression, gesture and eye contacts of the
driver. Further, the author has argued that self-efficacy and social norms can change the driver's
behavior. Social norms can be transmitted in V2V using social networks and most of the time in
the form of non-verbal communication. Hence, the combination of driver and car (machine)
become the cyborg so the driver of one can treat the other driver as a machine. He also added
that the "social pressure is particularly suitable to influence human driving behaviors for the
better and that this aspect is still relevant in the age of looming autonomous cars".
A complete autonomous vehicle (AV) was introduced first time in response to Defense
Advanced Research Projects Agency Grand Challenges. The major challenge for AVs is that
how it will take decisions at the time of the crashes. This is the key point where ethics and social
norms are required for AV's development. To address this requirement the Kumfer and Burgess
[15]evaluated three ethical theories, i.e. utilitarianism, respect for persons, and virtue ethics,
6
which help AVs to make least harming collision decision when the collision become
unavoidable. They performed the experiments using MATLAB their results revealed that the
utilitarian system produced the lowest number of death while on the other hand, the virtue ethics
system resulted in a supreme number of losses. However, the virtue ethics if fully integrated with
good AI techniques can be the best ethical solution. It is suggested that these ethical theories can
be implemented in AVs in different scenarios and complex environments.
3-Method
This section presents the method that has been used to propose the social norms and emotions
inspired artificial society of AVs. Figure 1 is the pictorial representation of our proposed method.
To introduce the emotions, a suitable appraisal model was required. According to [25], OCC
model is a best emotion appraisal model. Hence, the OCC model has studied thoroughly and
Prospect-based emotions have been selected. Further, emotion Fear has been selected to devise
the mechanism, which enforces the agents to obey the social norms in different collision leading
road scenarios. Afterward Fuzzy logic has been employed to compute the quantitative values of
different intensities of fear. Then social norms and emotions based rules have been designed,
which define the code of conduct for the artificial society of AVs. In addition, artificial social
actors along with different characteristics have been defined. Now to test the behavior of non-
social norms and social norms based artificial society, standard agent-based modeling tool
NetLogo has been used. Using, Sim-connector approach the numeric values of Fear emotion
have been provided in the simulation of artificial society. Then extensive experiments have been
performed, which helps to perform the comparison between non-social norms and social norms
7
Study OCC Model
Select Prospect Based
Emotions
Explore the variables of
prospect based emotion i.e.
Fear (likelihood,
undesirability, Ig)
Model Fear variables
(likelihood, undesirability, Ig)
according to autonomous
vehicles real road operations
Quantitative Computation of
Fear variables using Fuzzy
Logic
Design Norms obedience
Willingness function Fw
Based on the different
Intensities of Fear
Provide quantitative values of
Fear emotion to Netlogo
simulation using SimConnect
approach
Build Autonomous Vehicles
artificial social society
actors/agents and design
social characters Pre-crash
Scenario
Explore different types of
social norms
𝑭𝒆𝒂𝒓 − 𝑷𝒐𝒕𝒆𝒏𝒕𝒊𝒂𝒍(𝑨𝒗𝒊, 𝒆𝒊, 𝒕𝒊)
= 𝒇𝒇[ 𝑫𝒆𝒔𝒊𝒓𝒆 (𝑨𝒗𝒊,𝒆𝒊,𝒕𝒊), 𝑳𝒊𝒌𝒍𝒊𝒉𝒐𝒐𝒅 (𝑨𝒗𝒊,𝒆𝒊,𝒕𝒊), 𝑰𝒈(𝑨𝒗𝒊,𝒆𝒊,𝒕𝒊)]
Design the operational rules
for interaction between
𝑭𝒆𝒂𝒓 − 𝑰𝒏𝒕𝒆𝒏𝒔𝒊𝒕𝒚 (𝑨𝒗𝒊, 𝒆𝒊, 𝒕𝒊)
members of artificial society
= 𝑭𝒆𝒂𝒓 − 𝑷𝒐𝒕𝒆𝒏𝒕𝒊𝒂𝒍 (𝑨𝒗𝒊, 𝒆𝒊, 𝒕𝒊)
− 𝑭𝒆𝒂𝒓𝒕𝒉𝒓𝒆𝒔𝒉𝒐𝒍𝒅 (𝑨𝒗𝒊, 𝒆𝒊, 𝒕𝒊)
of AVs based on social norms
and emotions
Simulate the Artificial society
of AVs using Netlogo based on
defined social interaction rules
Compare with existing work
Perform experiments using
behavior space
Fig 1. Proposed Method
compliance artificial society of AVs. Furthermore, the method of using prospect-based emotion
for generating emotions in agents has been compared with existing work in literature.
8
4-Description of the Novel Solution (A view point)
In this section, the detailed description of the proposed viewpoint and model has been
presented.
4.1 Humanizing the AVs: A Possible Design Decision?
In order to include the social norms and emotions in the AVs, their fundamental design needs to
be changed because it is important since it can lead to more human-like behavior on the future
roads. It is important to note that emotional action is a social action that helps to regulate and
adapt other actors' emotions and emotional expressions according to valid norms and rules [16].
The emulation of emotions and social norms in the design of AVs can help in building a possibly
more comfortable, trustworthy and collision free AVs. However, the exact mode of
implementation is still debatable. The supposition guiding this method would be that the rules
and norms guiding human-human interaction/communication might also be pertinent in AV
design. Thus, the design principle would have a goal to introduce anthropomorphism capabilities
in AVs (Fig. 2).
4.2 How emotions enforce social Norms in Autonomous Vehicle: A scenario
To analyze the above discussion further, let us examine a conjectural interaction situation
of two autonomous vehicles as shown in figure 3. Suppose "A" is a heavy autonomous truck
followed by a smaller-sized autonomous vehicle "B" considerably less in weight but occupying
the same lane. Here, one way of representing "A" could be to consider it as a strong influencing
person having a strong social status, whereas "B" is a less influencing person having a weak
social character. For safe driving, both actors have to follow social norms and rules. In different
social societies of the world, weaker feels the emotion of fear from stronger, whereas stronger
9
Fig 2. Autonomous vehicles exhibiting different emotions based on their size and situation
Fig 3. Collision avoidance scenario using emotions enforced social norms
can have multiple emotions for the weaker like sympathy and pride, etc. Actor "B" should
maintain a safe distance from actor "A" or actor "A" should practice sympathetic emotion for
"B" to avoid the collision. Let us suppose," A" decreases its speed without taking care of in-
10
between distance from "B", which might lead to the collision. Suddenly another actor "C"
(having strong social status) appears on the road. "A" starts feeling the fear that if it collides with
"B", then "C" will have the evidence of its cruelty. According to a social norm, "You will be
punished for your act of crime "generates an emotion of fear of losing its status in social society
and get punished from the law enforcement institutes. Consequently, A maintains a safe distance
from B and avoids the collision.
What has happened here? The primary event of fretting the weaker autonomous vehicle can be
defined by some appraisal theory. This is the social norm that "Any actor of any status will be
punished for performing evil deed". Therefore, actor "A" avoids or feels hesitation in the
execution of collision scenarios in the presence of witnesses. This social norm generates the fear
of being punished. It means fear forces the actor to follow the norm.
4.3- Artificial Social Society of AVs
To further elaborate the idea presented in the previous section, we have given a concept
of an artificial society of AVs, consists of different actors, and having each one different
characteristic, which depicts their different social personalities. These different actors having
different personality characteristics along their short abbreviated nicknames are presented in
table 1. These characteristics have been assigned to these AVs from the real life behavior of
human drivers driving these types of non- autonomous vehicles. It has been observed that
drivers of heavy vehicles act dominantly, do not let the lighter vehicles to overtake or treat them
harshly, in the real road traffic.
11
DIFFERENT ACTORS OF ARTIFICIAL SOCIETY OF AVS
TABLE 1
Actor
AV_Truck
Symbol
Actor
T
of
Personality Type Weight
Very Dominating
>5000
AV_Bus
B
Very Dominating
4500<<50000
AV_Toyota_Small
TY
Dominating
3000 <4000
truck
AV_Carry
C
Weak Dominating
2000 <2800
AV_Car 3000cc
CB
Weak Dominating
2000 <2500
AV_Car 2000cc
CS
Weak Dominating
1500 <1700
AV_Rickshaw
AV_Ambulance
AV_Motorbike
R
A
M
Weak Dominating
1200 <1400
Weak Dominating
1200^<1400
Very
weak
800 <1100
Dominating
AV_Cycle
CL
Very
weak
400
Dominating
4.3.1 -Operational rules for artificial society of AVs
Humans live socially following some rules, which help them to live peacefully, avoiding
any possible conflicts, if follow them properly. In the analogy of these social rules, we have
proposed some rules as well for the proposed artificial society of AVs. These rules have been
designed in the light of different possible road scenarios, though all cannot be mentioned here,
and further, the social norms have been presented as well along with the best-suited emotions.
Furthermore, to check weather following social norm, according to the given condition, does not
lead to the violation of the road traffic rules, we have applied a check to assure that the social
norm will be followed only when the social norm and road norm both are in the compliance of
each other. The last column of table 2 presents the action that the actor has to be taken to avoid
the road collision.
12
We have used both formal and informal social norms, in our proposed model, and these have
been rewritten in the context of AVs collision avoidance strategies. For example:
• You ought to maintain a safe distance from stronger vehicles to avoid the collision.
• You ought to be kind to the weak vehicles to avoid the collisions
• Keep on your Lane
• Help in executing safe overtaking maneuver
SOCIAL NORMS AND EMOTIONS BASED ROAD INTERACTION RULES
TABLE 2
Road Scenarios
T is leading the CB and CB
Social norm
(1=Self-norm
2=Religious norm
3=Enforced Norms)
Help the weaker
Emotion
Road norm
(P) (p &q)
Action
Sympathy,
Guilt,
(Should
assist
P & q -> social norm
wants to overtake the T
Shame
overtaking vehicle to
p & q !-> social
get past swiftly and
norm
securely) &
(Keep
your lane).
CB is leading the T and T
Maintain
distance
Fear
(Must not drive in a
P & q -> social norm
wants to overtake the CB
from stronger one
bus lane) & (Should
p & q !-> social
assist
overtaking
norm
vehicle to get past
swiftly and securely)
CB–CB following scenario
Give the right or Tit
Guilt, shame,
(Must not drive in a
P & q->social norm
and the following CB wants
for Tat
Sympathy
bus lane) & (Should
p & q !-> social
to overtake the leading CB.
assist
overtaking
norm
vehicle to get past
swiftly and securely)
T is following the B and T
Give the right or Tit
Fear ,guilt ,shame
(Increase your speed
P&q -> social norm
wants to overtake B
for Tat
to
keep
a
safe
p&q !-> social norm
distance from other
vehicles) &(Should
13
assist
overtaking
vehicle to get past
swiftly and securely)
If CB following T in bright
Keep
the distance
Fear
Keep
the
safe
P -> social norm
conditions
from stronger one
distance and use the
p!->social norm
two-second rule.
If CB following T in rainy
Keep
the distance
Fear
Keep
the
safe
P -> social norm p
conditions
from stronger one
distance and use the
!-> social norm
two-second rule.
If CB
is
involved
in
Keep
the distance
Fear
(Increase your speed
P -> social norm
tailgating with T
from stronger one
to
keep
a
safe
p!->social norm
distance from other
vehicles )
If T is involved in tailgating
Keep
the distance
Fear
(Increase your speed
P & q -> social norm
with CB
from stronger one,
to
keep
a
safe
p & q !-> social
Abide the rule
distance from other
norm
vehicles) & (must not
drive in a bus lane)
4.3.2-Generate Prospect-based Emotions using OCC Model
A very popular model of emotions was developed by Ortony, Clore and Collins also
known as OCC model [17]. The reason for choosing the OCC model is they are presenting 22
basic emotions along with the concept of computing the emotions as well. In OCC model, the
authors answer the question that what concludes the strength of the emotions. They commence a
number of variables in enjoining to answer this problem. To address this problem, they introduce
three types of variables as shown in Table 3. The OCC model is shown in the figure below. The
variables given in Table 3 help to compute the strength of the emotions. However, these values
14
are still computed in qualitative manners. However, we will compute their quantitative values
using fuzzy logic first and then use them to evaluate proposed schemes given in coming sections.
4.4 Overall functionality of proposed approach
The overall functionality of the proposed approach is presented in figure 4 and its description is
given as under.
1. In EventPart1, CB is following T and it requests the T to give safe passage for
performing an overtaking maneuver. Belief will depict the situation awareness of CB and
T. The current values of Belief are passed to the Prospect Based Emotion Generation
module.
2. Prospect Based Emotion Generation module computes the emotion fear based on
equation 1 and 2 provided by OCC model [25].
𝐹𝑒𝑎𝑟 − 𝑃𝑜𝑡𝑒𝑛𝑡𝑖𝑎𝑙(𝐴𝑣𝑖, 𝑒𝑖, 𝑡𝑖) = 𝑓𝑓[ 𝐷𝑒𝑠𝑖𝑟𝑒 (𝐴𝑣𝑖,𝑒𝑖,𝑡𝑖), 𝐿𝑖𝑘𝑙𝑖ℎ𝑜𝑜𝑑 (𝐴𝑣𝑖,𝑒𝑖,𝑡𝑖), 𝐼𝑔(𝐴𝑣𝑖,𝑒𝑖,𝑡𝑖)] ..(1)
𝐹𝑒𝑎𝑟 − 𝐼𝑛𝑡𝑒𝑛𝑠𝑖𝑡𝑦 (𝐴𝑣𝑖, 𝑒𝑖, 𝑡𝑖) = 𝐹𝑒𝑎𝑟 − 𝑃𝑜𝑡𝑒𝑛𝑡𝑖𝑎𝑙 (𝐴𝑣𝑖, 𝑒𝑖, 𝑡𝑖) −
𝐹𝑒𝑎𝑟𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑 (𝐴𝑣𝑖, 𝑒𝑖, 𝑡𝑖) .. (2)
3. The computed Intensity of fear will update Belief of the agent.
4. Based on Intensity of fear Fw is computed. The Fw function is computed using equation
3. Equation 3 shows that the willing function of T, which allows the overtaking, depends
on the intensity of fear.
𝐹𝑤 (𝑜𝑏𝑒𝑦−𝑛𝑜𝑟𝑚)= 𝐹𝑒𝑎𝑟 − 𝐼𝑛𝑡𝑒𝑛𝑠𝑖𝑡𝑦 (𝐴𝑣𝑖, 𝑒𝑖, 𝑡𝑖) .. (3)
In the first iteration, the value of the emotion will be zero. In the Willing Function part, it
will be checked whether the value of Willing Function of T is higher than λ or it is less
than λ. Here two scenarios exist; first, if the Fw of T is less than λ, and second if the Fw is
greater than λ.
5. If the value of Fw is smaller than the egoist value of agent then it disobeys the Norm.
15
6. In the case of disobeying the Norm, T will be entered in the pre-crash scenario. For the
pre-crash scenario we have considered the variables defined by [18]
7. The event of pre-crash scenario will contribute in the shape of the High likelihood of an
accident and it will increase the intensity of fear. Again, the belief of agent will be
updated and Fw will be computed. If the Fw is still smaller than λ then the emotion
generation center will be consulted again to depict a highly dangerous situation. If the
value of Fw is greater than λ then the egoist agent will change its mind and turn into the
emotional agent and the emotions act as a norm compliance mechanism. In next step
agent will check the road norm that accepting the preceding AV (CB) request is not
against the road norm, then it will load possible solutions and execute the maneuver,
which will ultimately help in CB in performing an overtaking maneuver by avoiding rear
end collision.
5-Experiments
This section describes the two types of experiments: Quantitative Computation of Prospect-based
Emotion using Fuzzy Logic and validation of EEC_Agent .
5.1 Experiment 1
Since human emotions are fuzzy and complex in nature, using fuzzy sets for modeling the human
emotions can be a suitable choice [19]. For modeling the emotions, fuzzy logic has been
extensively utilized [20]. A computational model of emotions has been proposed that can be
included in any cognitive agent or program. In [21], the Takagi-Sugeno (TS) fuzzy model has
been utilized to develop an online learning system of emotions. The purpose of designing this
system was to investigate that how multi-model actions can be generated and understand by
cognitive robots. In [22], a novel method is proposed which helps to model the emotions using
16
different types of physiological data. For this purpose two fuzzy logic models are employed: one
model is used for converting the signals into valence and arousal and the second model is used
Event Part 1
11
1
Belief
2
(Initialize Likelihood, Desirability,
Ig)
3
Fw > = λ
Yes
4b
No
4a
Generate Emotions using Prospect Based
Emotion defined by OCC Model
𝑭𝒆𝒂𝒓 − 𝑷𝒐𝒕𝒆𝒏𝒕𝒊𝒂𝒍(𝑨𝒗𝒊, 𝒆𝒊, 𝒕𝒊)
= 𝒇𝒇[ 𝑫𝒆𝒔𝒊𝒓𝒆 (𝑨𝒗𝒊,𝒆𝒊,𝒕𝒊), 𝑳𝒊𝒌𝒍𝒊𝒉𝒐𝒐𝒅 (𝑨𝒗𝒊,𝒆𝒊,𝒕𝒊), 𝑰𝒈(𝑨𝒗𝒊,𝒆𝒊,𝒕𝒊)]
𝑭𝒆𝒂𝒓 − 𝑰𝒏𝒕𝒆𝒏𝒔𝒊𝒕𝒚 (𝑨𝒗𝒊, 𝒆𝒊, 𝒕𝒊)
= 𝑭𝒆𝒂𝒓 − 𝑷𝒐𝒕𝒆𝒏𝒕𝒊𝒂𝒍 (𝑨𝒗𝒊, 𝒆𝒊, 𝒕𝒊)
− 𝑭𝒆𝒂𝒓𝒕𝒉𝒓𝒆𝒔𝒉𝒐𝒍𝒅 (𝑨𝒗𝒊, 𝒆𝒊, 𝒕𝒊)
Disobey Norms
6
5
Leads
Road Norms
Pre-crash Scenario
Execute Collision
Avoidance Maneuver
10
Load Possible Collision
Avoidance Solutions
Controls
9
Obey Norms
Yes
8b
7
Comply
No
8a
Fig 4. Overall functionality of proposed approach
for converting this valence and arousal to five emotional states related to the computer games
e.g. boredom, excitement, challenge, frustration, and fun. In our case, the OCC model provides a
computation traceability algorithm shown below for computing the intensity of fear in [23].
17
However, it is useless if the numeric values of linguistic variables like likelihood, desire and Ig
variables are not known.
If Prospect (v, e, t) and Undesirable (v, e, t) < 0
Then set Fear-Potential (v, e, t) = ff [Desire (v, e, t) , Likelihood (v, e, t), Ig (v, e, t)]
If Fear-Potential (v, e, t) > Fear-Threshold (v, t)
Then set Fear-Intensity (v, e, t) = Fear-Potential (v, e, t) - Fear-Threshold (v, t)
Else set Fear-Intensity (v, e, t) =0
5.1.1 Implementation details of fuzzy logic to compute the numeric values of Fear Emotion
In order to compute the Fear-Potential as given in part 1 of the computation traceability
algorithm, we have to calculate the values of Desirability, Likelihood, and Intensity of a global
variable. In the context of the state of the art given above, we used fuzzy logic to compute the
numeric values of Desirability, Likelihood, and Intensity of global variable.
a) Likelihood:
For the variable of Likelihood the five linguistic tokens VLLH, LLH, MLH, HLH and VHLH
were defined which represent Very low likelihood, Low likelihood, Medium likelihood, High
likelihood and Very High likelihood respectively shown in table 3.
LIKELIHOOD LINGUISTIC TOKENS AND THEIR DESCRIPTION
TABLE 3
Linguistic Tokens
Description
VHLH
Very High Likelihood
HLH
MLH
LLH
High Likelihood
Medium Likelihood
Low Likelihood
VLLH
Very Low Likelihood
The variable of likelihood is affected by Distance and Speed variables. Twenty-five rules were
defined to obtain the value of the variable likelihood; these rules are given in the following table.
18
TABLE 4
LIKELIHOOD FUZZY INFERENCE RULES
If Distance is And Speed is
VHD
VHS
Then Likelihood is
MLH
VHD
VHD
VHD
VHD
HD
HD
HD
HD
HD
MD
MD
MD
MD
MD
LD
LD
LD
LD
LD
V LD
V LD
V LD
V LD
V LD
HS
MS
LS
VLS
VHS
HS
MS
LS
VLS
VHS
HS
MS
LS
VLS
VHS
HS
MS
LS
VLS
VHS
HS
MS
LS
VLS
LLH
VLLH
VLLH
VLLH
HLH
MLH
VLLH
VLLH
VLLH
VHLH
VHLH
MLH
LLH
VLLH
VHLH
VHLH
HLH
MLH
VLLH
VHLH
VHLH
VHLH
HLH
MLH
The remaining details are given in appendix A.
19
5.2 -Experiment 2
The purpose of the second main experiment is to simulate the concept of an artificial
society of AVs, which consists of different actors having different characteristics.
Another reason of simulation is to study the behavior of these actors according to the
defined social rules during autonomous driving. For this purpose, Netlogo 5.3 has been
utilized which is a standard agent-based simulation environment. The NetLogo
environment consists of patches, links, and turtles [24]. The algorithms used in this
experiment have already been given in section 3.1.3. Figure 5 presents the experimental
environment along with input and output parameters. The left side of the simulation
world contains input sliders for providing fuzzy logic based numeric values of prospect
based emotion variables (Undesirability, Likelihood, Ig). It is important to recall here that
these numeric values of prospect-based emotions were computed through experiments a,
b and c using fuzzy logic and then provided to the agent based simulation by following
proposed SimConnector approach.
Fig. 5 Main Simulation Screen of Social Norms and Emotions inspired artificial society of AVs
5.2.1- Simulation Parameters Description
20
The simulated world consists of different types of input and output parameters. To provide the
inputs, different sliders have been used, whereas to get the outputs, monitors and plots are used.
The description of each input and output object along with defined range is presented in table 5.
Simulation
Parameters
Number of Autonomous
General
Range
[1-30]
SIMULATION PARAMETERS AND THEIR DESCRIPTION
TABLE 5
Description
This slider helps in defining the maximum members of the artificial
Vehicles Agents
society of AVs.
Vehicles Ratio
[2:1, 3:1, 4:1]
It defines the ratio of AV Trucks and cars within a total number of
vehicles set by the Number of Autonomous Vehicles Agents slider.
Maximum Velocity
[0 -1; with increment
This slider helps in defining the maximum velocity that can be achieved
of 0.01]
by all actors of artificial society
Minimum Velocity
[0 -1; with increment
This slider helps in defining the lower boundary of velocity achieved by
of 0.01]
all actors of artificial society
Acceleration-rate
[0 -1; with increment
This slider helps in defining the maximum acceleration rate that can be
of 0.05]
used by all actors of artificial society
Declaration-rate
[0 -1; with increment
This slider helps in defining the minimum declaration- rate that can be
of 0.05]
used by all actors of artificial society
Safety Distance
[2 -10-; with increment
This slider helps in defining the safety distance between each actor.
of 1]
Sonar Range
[1 -10-; with increment
This slider helps in defining the sonar range of each AV to find out the
of 1]
position and distance between neighboring Avs.
Metacognition
On/Off
This switch helps in defining that the simulation is in Random walk or
social norms mode.
Prospect Based Emotion
Range
Description
i.e. Fear Generation
Parameters
Likelihood
[0 -1; with increment
This slider helps in defining the likelihood of accident perceived by AV
of 0.1]
21
Desirability
[0 -1; with increment
This slider helps in defining the current desirability value of AV.
of 0.1]
Ig
[0 -1; with increment
This slider helps in defining the current Ig value of AV.
of 0.1]
5.2.2 Experimental design
In this section, further experimental design has been proposed to perform the experiment 2 in
proper manners.
a) Experiments_TypeA
In this category of experiments, total five sets of experiments have been designed to test the non-
social norms random walk based artificial society of AVs. In the first set of experiments, the
Max Velocity Range parameter has been set to 0.8, which represents a high velocity of AVs. The
Acceleration/ Deceleration Rate parameters are set to 0.1 along with the Safety Distance equal to
3. In the second set of experiments, the Max Velocity Range parameter has been set to 0.5,
which represents a medium velocity of AVs. The Acceleration Rate parameter is set to 0.2 along
with the Safety Distance equal to 2. In the third set of experiments, the Max velocity range is set
to 0.3, which represent the low velocity of AVs. The Acceleration Rate and Declaration Rate
parameters are both set to 0.1. In the fourth set of experiments, Acceleration and Deceleration
Rates are set to 0.3 and the values of safety Distance and Sonar Range are set to 3. This set of
experiment helps in measuring the performance of non-social norms random walk based artificial
society of autonomous vehicles having equal safety distance and sonar range. In the fifth set of
experiments, safety distance, and sonar range parameters are set to 1. This set of experiment
helps in testing the behavior of AVs having equal low safety distance and sonar range. All of
these sets of experiments have been executed using the behaviour space tool within the Netlogo
22
5.3 environment. Furthermore, each set of experiments has been repeated seven times and the
total number of collisions along with their mean and standard deviation has been computed. The
details of these 5 sets of experiments are presented in table 6, table 7, table 8, table 9, and table
10 respectively.
Experiment
No
Number
of AVs
1
2
3
4
5
10
15
20
25
30
Experiment
No
Number
of AVs
1
2
3
4
5
10
15
20
25
30
Experiment
No
Number
of AVs
10
15
20
25
30
1
2
3
4
5
EXPERIMENT TYPE_A SET 1: PARAMETERS AND THEIR VALUES
TABLE 6
Min
Velocity
Range
0.14
Max
Velocity
Range
0.8
0.14
0.14
0.14
0.14
0.8
0.8
0.8
0.8
Acceleration
Rate
Deceleration
Rate
Safety
Distance
Sonar
Range
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
3
3
3
3
3
2, 5
2, 5
2, 5
2, 5
2, 5
EXPERIMENT TYPE_A SET 2: PARAMETERS AND THEIR VALUES
TABLE 7
Min
Velocity
Range
0.14
Max
Velocity
Range
0.5
0.14
0.14
0.14
0.14
0.5
0.5
0.5
0.5
Acceleration
Rate
Deceleration
Rate
Safety
Distance
Sonar
Range
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
2
2
2
2
2
2, 5
2, 5
2, 5
2, 5
2, 5
EXPERIMENT TYPE_A SET 3: PARAMETERS AND THEIR VALUES
TABLE 8
Min
Velocity
Range
0.14
Max
Velocity
Range
0.3
0.14
0.14
0.14
0.14
0.3
0.3
0.3
0.3
Acceleration
Rate
Deceleration
Rate
Safety
Distance
Sonar
Range
0.1
0.1
0.1
0.1
0.1
2
2
2
2
2
2
2
2
2
2
0.1
0.1
0.1
0.1
0.1
23
Experiment
No
Number
of AVs
1
2
3
4
5
10
15
20
25
30
Experiment
No
Number
of AVs
1
2
3
4
5
10
15
20
25
30
EXPERIMENT TYPE_A SET 4: PARAMETERS AND THEIR VALUES
TABLE 9
Min
Velocity
Range
0.14
Max
Velocity
Range
0.3
0.14
0.14
0.14
0.14
0.3
0.3
0.3
0.3
Acceleration
Rate
Deceleration
Rate
Safety
Distance
Sonar
Range
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
3
3
3
3
3
3
3
3
3
3
EXPERIMENT TYPE_A SET 5: PARAMETERS AND THEIR VALUES
TABLE 10
Min
Velocity
Range
0.14
Max
Velocity
Range
0.3
0.14
0.14
0.14
0.14
0.3
0.3
0.3
0.3
Acceleration
Rate
Deceleration
Rate
Safety
Distance
Sonar
Range
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
1
1
1
1
1
1
1
1
1
1
b) Experiments_TypeB
In this category of experiments, total five sets of experiments in parallel the Experiments_TypeA
have been designed to test and compare the social norms and emotions inspired artificial society
of AVs with non-social norms random walk based artificial society of AVs. These five sets of
experiments are designed in parallel to the Type_A experiments. These sets of experiments have
the same values of parameters as type_A experiments have. The additional parameter added in
these experiments is the variables of fear, which help in computing the intensity of fear. Table
11 through 15 presents these 5 sets of experiments respectively.
24
EXPERIMENT TYPE_B SET 1: PARAMETERS AND THEIR VALUES
TABLE 11
Experiment
No
Number
of AVs
1
2
3
4
5
10
15
20
25
30
Min
Velocity
Range
0.14
Max
Velocity
Range
0.8
0.14
0.14
0.14
0.14
0.8
0.8
0.8
0.8
Acceleration
Rate
Deceleration
Rate
Safety
Distance
Sonar
Range
LI, UD,
Ig
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
3
3
3
3
3
2 , 5
0.1-0.1-1
2 , 5
0.1-0.1-1
2 , 5
0.1-0.1-1
2, 5
0.1-0.1-1
2 , 5
0.1-0.1-1
EXPERIMENT TYPE_B SET 2: PARAMETERS AND THEIR VALUES
TABLE 12
Experiment
No
Number
of AVs
1
2
3
4
5
10
15
20
25
30
Min
Velocity
Range
0.14
Max
Velocity
Range
0.5
0.14
0.14
0.14
0.14
0.5
0.5
0.5
0.5
Acceleration
Rate
Deceleration
Rate
Safety
Distance
Sonar
Range
LI, UD,
Ig
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
2
2
2
2
2
2, 5
0.1-0.1-1
2, 5
2, 5
2, 5
2, 5
0.1-0.1-1
0.1-0.1-1
0.1-0.1-1
0.1-0.1-1
EXPERIMENT TYPE_B SET 3: PARAMETERS AND THEIR VALUES
TABLE 13
Experiment
No
Number
of AVs
1
2
3
4
5
10
15
20
25
30
Min
Velocity
Range
0.14
Max
Velocity
Range
0.3
0.14
0.14
0.14
0.14
0.3
0.3
0.3
0.3
Acceleration
Rate
Deceleration
Rate
Safety
Distance
Sonar
Range
LI, UD,
Ig
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
2
2
2
2
2
2
2
2
2
2
0.1-0.1-1
0.1-0.1-1
0.1-0.1-1
0.1-0.1-1
0.1-0.1-1
EXPERIMENT TYPE_B SET 4: PARAMETERS AND THEIR VALUES
TABLE 14
Min
Velocity
Range
0.14
Max
Velocity
Range
0.3
0.14
0.3
Acceleration
Rate
Deceleration
Rate
Safety
Distance
Sonar
Range
LI, UD,
Ig
0.3
0.3
0.3
0.3
3
3
3
3
0.1-0.1-1
0.1-0.1-1
25
Experiment
No
Number
of AVs
10
15
1
2
3
4
5
20
25
30
0.14
0.14
0.14
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
3
3
3
3
3
3
0.1-0.1-1
0.1-0.1-1
0.1-0.1-1
EXPERIMENT TYPE_B SET 5: PARAMETERS AND THEIR VALUES
TABLE 15
Min
Velocity
Range
0.14
Max
Velocity
Range
0.3
0.14
0.14
0.14
0.14
0.3
0.3
0.3
0.3
Acceleration
Rate
Deceleration
Rate
Safety
Distance
Sonar
Range
LI, UD,
Ig
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.1
1
1
1
1
1
1
1
1
1
1
0.1-0.1-1
0.1-0.1-1
0.1-0.1-1
0.1-0.1-1
0.1-0.1-1
Experiment
No
Number
of AVs
10
15
20
25
30
1
2
3
4
5
6-Results and Discussion
This section elaborates the detailed discussion according to the results achieved for
experiment1 and experiment 2.
6.1 Experiment 1
Criado et al. [8], have utilized prospect based emotions defined by the OCC model to enforce the
agents to obey the social norms. However, the authors have not proposed any proper mechanism,
which helps to quantify the different intensities of fear. Furthermore, the authors have considered
only Desirability and likelihood variables, whereas ignoring Ig variable. In addition, the values of
desirability and Likelihood variable are just supposed between [-1 1] without providing any
justification. Our approach is better than [8] in this regard that we have used fuzzy logic to
compute the numeric values of Fear variables (Likelihood, Desirability, Ig) to computer Fear
Potential and then Fear Intensity has been computed using the proper algorithm defined by the
inventors of the OCC model [25].
Table 1 shows the quantitative values of undesirability from very low (VL) to very high (VH).
The terms VLD, LD, MD, HD, and VHD are the acronyms of very low desirability, low
26
desirability, medium desirability, high desirability and very high desirability respectively. If the
agent has a value between 0-0.24 for its undesirability of an event, then it can be interpreted as
the very low undesirability. However, from an abstract analysis, it can be noted that due to the
fuzzy nature of the emotion fear the boundary of one intensity level mixes in the boundary of
another intensity level. Hence, the intensity levels lie between 0.24 and 0.5 will be interpreted as
low undesirability and lower than these values as the very low undesirability. In the same
way, the other intensity levels of undesirability variable can be interpreted.
In the same way, Table 2 and table 3 are showing the five quantitative values for finding
the different intensity levels of likelihood and Ig variables.
These quantitative values of Desirability, Likelihood and Ig are presented in table 16, 17
and 18 respectively. These values are then provided to the EEC_Agent for computing
different intensities of fear in the next section by following the proposed SimConnector
design.
VLD
0-0.24
VLL
0-0.24
VLIg
0-0.24
Quantitative Values of Five Intensity levels of Desirable Variable
TABLE 16
LD
MD
HD
0.1-0.5
0.25-0.73
0.51-0.9
QUANTITATIVE VALUES OF FIVE INTENSITY LEVELS OF LIKELIHOOD VARIABLE
TABLE 17
LL
ML
HL
0.1-0.5
0.25-0.73
0.51-0.9
VHD
0.76-1
VHL
0.76-1
Quantitative Values of Five Intensity levels of Global Variable (Ig)
TABLE 18
LIg
0.1-0.5
MIg
HIg
0.25-0.73
0.51-0.9
VIg
0.76-1
27
6.2 Experiment 2
7.2.1 The Results: Experiment_TypeA set 1 Vs Experiment_TypeB set 1 and
Experiment_TypeA set 2 Vs Experiment_TypeB set 2
The results of both experiments_TypeA set 1 and experiments_Type B set1 are presented
in the form of average accidents along with standard deviation in table 19. From the
results, it can be seen that there is a high average of accidents in case of non-social norms
random walk based artificial society of AVs as compared to the social norms and emotions
based artificial society of AVs. For example, the average accidents performed by non-
social norms random walk based are 48.63 for 10 AVs. Comparatively, 2.57 are the
average accidents performed by social norms and emotions based technique. In the same
way, for 30 AVs total average accidents by non-social norms random walk are 305. 43 and
59. 35 by the social norms and emotions based technique. From the results, another
interesting phenomenon can be observed that the average accidents in both techniques are
gradually increasing as the number of AVs is increasing. Figure 6 (A) is representing the
graphical representation of the results of table 19.
In comparison to the table 19, table 20 has been presented. Table 20 presents the results of
both experiment_TypeA set2 and experiment_TypeB set 2 in the form of average accidents
along with standard deviation. Before discussing the results of table 20 it would be
interesting to perform the comparison of table 19 and table 20. Experiments_TypeA set 1
and experiments_TypeB set1 have a high maximum velocity range, i.e. 0.8 with
acceleration and deceleration
rate 0.1. Whereas, experiments_TypeA set2 and
experiments_TypeB set 2 have medium maximum velocity range, i.e. 0.5 with acceleration
and deceleration rates 0.2. From table 20, it can be seen that average accidents have been
decreased due to medium maximum velocity range as compared to the average accidents
28
presented in table 19 having a high maximum velocity range. For example, in table 19 and
20 for social norms and emotions based technique, the average accidents performed by 10
AVs are 2.57 and 44.35 respectively. In the same way, in the case of 30 AVs, the average
accidents performed by social norms and emotions based technique are 59.35 and 24.065
respectively.
From the table 20, it can be seen that social norms and emotions based technique have less
number of collisions as compared to the non-social norms random walk based technique.
For example, the average accidents performed by social norms and emotions based
artificial society are 6.35 for 20 AVs. Whereas, for the same number of AVs the average
accidents performed by non-social norms based artificial society are 147.03. Figure 6 (B)
is representing the graphical representation of the results of table 20.
TABLE 19
Set 1 Type A & B Experiments Results
TABLE 20
Set 2 Type A & B Experiments Results
No. of AVs
Social norms and
Emotions Based
Nonsocial norms
Random walk based
10 AVs
Mean
2.573837
Stdev
1.597861
Mean
48.63886
Stdev
10.05186
No. of
AVs
10 AVs
Social norms and
Emotions Based
Non-social norms
Random walk based
Mean
0.83308
Stdev
0.75219454
Mean
44.35492
Stdev
10.41238
15 AVs
7.317401
3.12966
95.57575
13.87058
15 AVs
2.688164
1.50878434
92.14185
12.76792
20 AVs
16.17935
6.106475
152.584
18.13361
20 AVs
6.358672
2.40534634
147.0383
15.39402
25 AVs
32.20731
11.09385
222.3648
21.98538
25 AVs
13.37643
3.34040996
207.6984
17.56855
30 AVs
59.35412
17.14856
305.439
25.01929
30 AVs
24.06544
4.45493639
276.0833
19.05506
29
Social norms and
Emotions Based
Non social norms
Random walk based
330
300
270
240
210
180
150
120
90
60
30
0
10 AVs
15 AVs
20 AVs
25 AVs
30 AVs
Social norms and
Emotions Based
Non social norms
Random walk
based
300
270
240
210
180
150
120
90
60
30
0
10 AVs
15 AVs
20 AVs
25 AVs
30 AVs
(A)
(B)
Fig. 6 Graphical representation of the results (A) Experiment_TypeA set 1 Vs Experiment_TypeB set 1 (B) Experiment_TypeA set 2 Vs Experiment_TypeB set
2
In the same way, for the 10, 20, 25, and 30 number of AVs, average number of collisions
by the social norms and emotions based artificial society of AVs are less than non-social
norms and random walk based artificial society of AVs.
It would be interesting to present here the analysis of TypeB set 1 and set 2 experiments
with Type B set 3 experiments. From the comparative study of table 19, 20 and 21 it can
be seen that Type B set 3 results are better than Type B set1 and set 2 experiments. For
example, for 30 AVs, the average collisions performed by set1 and set 2 are 59.35 and
24.06. Whereas there are only 14.69 collisions on average by TypeB set 3. Hence, it can
be concluded that social norms and emotions based artificial society of AVs can have less
number of collisions by adapting low maximum velocity range i.e. 0.3 and both safety and
sonar distances equal to 2. Figure 7 (C) and (D) are the graphical representations of table
21 and 22 respectively.
30
TABLE 21
Set 3 Type A & B Experiments Results
TABLE 22
Set 4 Type A & B Experiments Results
No. of AVs
10 AVs
Social norms and
Emotions Based
Nonsocial norms Random
walk based
No. of AVs
Mean
0.785513
Stdev
0.916095
Mean
22.66486
Stdev
10.09421
10 AVs
Social norms and
Emotions Based
Stdev
2.067026
Mean
1.889218
Non-social norms
Random walk based
Mean
25.12701
Stdev
9.707391
15 AVs
2.268417
1.664854
55.82655
13.73423
15 AVs
6.042479
3.723965
60.30161
13.30325
20 AVs
4.767293
2.38503
101.9607
17.59697
20 AVs
14.54727
5.010924
106.1374
14.68821
25 AVs
8.728624
3.078323
158.7121
20.06934
25 AVs
28.01678
6.281523
158.8475
17.78358
30 AVs
14.69935
3.936857
222.0574
22.18164
30 AVs
44.53196
7.017011
219.4805
18.46683
Social norms and
Emotions Based
Non social norms
Random walk based
10 AVs
15 AVs
20 AVs
25 AVs
30 AVs
Social norms and
Emotions Based
Non social norms
Random walk
based
270
240
210
180
150
120
90
60
30
0
Number of AVs
(C)
10 AVs
15 AVs
20 AVs
Number of AVs
(D)
25 AVs
30 AVs
Fig. 7 Graphical representation of the results (C)_TypeA set 3 Vs Experiment_TypeB set 3 (D) Experiment_TypeA set 4 Vs Experiment_TypeB set 4
270
240
210
180
150
120
90
60
30
0
6.2.3 The Results: Experiment_TypeA and TypeB set 3 Vs Experiment_TypeA set 5 Vs
Experiment_TypeA and Type_B set 5
Both sets 3 and 5 of Experiments_Type A and B have the same values of parameters
expect safety distance and sonar range. In the set 3, both safety and sonar range are set to
3. Whereas for the set 5, both of these parameters are set to 1. From the comparative
analysis, it can be seen that the set 3 has less number of collisions for 10, 15, 20 and 25
AVs. However, for 30 AVs, set 5 haS less number of accidents. From the trend line shown
31
in figure 8 (E) for experiments_TypeA set 3, it can be seen that the number of collisions is
increasing gradually as the number of AVs is increasing. In contrast to the trend line
shown in figure 8 (F) for experiments_TypeA set 3 presents a gradual increase for 10, 15,
20, and 25 AVs but then suddenly drop down for 30 AVs. Hence, it means that the set 3
provides optimal operational parameters for the artificial society of AVs within the range 1
to 25. Whereas, for 30 AVs experiment no 5 of set 5 is the optimal option. From these
results, we can also deduce that for higher AVs, small and equal safety distance and sonar
range parameters are most optimal one.
If we see the results of table 23 then it is obvious that social norms and emotions based
artificial society of AVs have less number of collisions for all numbers of AVs as
compared to the non-social norms and emotions based artificial society of AVs.
6.2.4 Analysis of most optimal Sonar Range Vs. Safety Distance for less number of
collisions in Social norms and Emotions Based artificial society of AVs
Table 24 presents the number of collisions for different number of AVs according to
different safety distances and sonar ranges in experiments_TypeA set1 to set5. From the
results, it can be seen that when safety distance and sonar range parameters having values
(1, 1) and (2, 2) respectively, the average number of collisions is lesser. For example, for
10 AVs with safety distance and sonar range parameters set to (1, 1) and (1, 2) the average
number of collisions are 0.79 and 0.78 respectively as compared to the 3.01, 1.88 and 2.12
for sonar range and safety distance parameters set to (3, 2), (3, 3), and (3, 5) respectively.
From the set4, experiments_TypeA with safety distance and sonar range parameters set to
(3, 2) the average number of collisions is higher than all other experiments. Its reason is
smaller safety distance than the sonar range. It means that when the AV has higher safety
32
TABLE 21
Set 3 Type A & B Experiments Results
Social norms and
Emotions Based
Nonsocial norms
Random walk based
TABLE 23
Set 5 Type A & B Experiments Results
Social norms and
Emotions Based
Nonsocial norms
Random walk based
No. of AVs
No. of AVs
10 AVs
Mean
0.785513
Stdev
0.916095
Mean
22.66486
Stdev
10.09421
10 AVs
Mean
0.791342
Stdev
0.576955
Mean
37.75169
Stdev
11.27174
15 AVs
2.268417
1.664854
55.82655
13.73423
15 AVs
2.084906
0.999392
85.15609
14.30291
20 AVs
4.767293
2.38503
101.9607
17.59697
20 AVs
7.152462
1.965884
145.9167
16.50659
25 AVs
8.728624
3.078323
158.7121
20.06934
25 AVs
11.63587
2.745418
215.4169
18.68783
30 AVs
14.69935
3.936857
222.0574
22.18164
30 AVs
4.069908
1.37926
294.7733
20.14137
Social norms and
Emotions Based
Non social norms
Random walk based
270
240
210
180
150
120
90
60
30
0
Social norms and
Emotions Based
Non social norms
Random walk
based
300
270
240
210
180
150
120
90
60
30
0
10 AVs
15 AVs
20 AVs
25 AVs
30 AVs
10 AVs
15 AVs
20 AVs
25 AVs
30 AVs
Number of AVs
(E)
Number of AVs
(F)
Fig. 8 Graphical representation of the results (E) Experiment_TypeA set 3 Vs Experiment_TypeB set 3 (F) Experiment_TypeA set 5 Vs Experiment_TypeB set 5
distance and low capability of detecting its neighbors then the number of collisions
increases. Figure 9 presents the graphical representation of the results of table 24.
Experimental results of TypeA set1- set 5 regarding different sonar ranges
TABLE 24
range,
Safety
Sonar distance
1, 1
10 AVs1
0.791342
15 AVs
2.084906
20 AVs
7.152462
25 AVs
11.63587
30 AVs
4.069908
2, 2
0.785513
2.268417
4.767293
8.728624
14.69935
33
2, 5
3, 2
3, 3
3, 5
0.814496
2.711523
6.390049
13.36398
24.04687
3.01972
8.936398
20.14434
40.31569
72.42839
1.889218
6.042479
14.54727
28.01678
44.53196
2.127955
5.698404
12.21435
24.09892
46.27984
Comparison of different safety distances and sonar ranges
settings in experiments_TypeA set 1 to set 5
80
70
60
50
40
30
20
10
0
10 AVs1
15 AVs
20 AVs
25 AVs
30 AVs
1, 1
2, 2
2, 5
3, 2
3, 3
3, 5
FIG. 9 Graphical representation of Results regarding different sonar ranges given in Experiments: TypeA set1- set 5
7-Conclusion
The paper has been written in the context of proposing a novel collision avoidance solution
for the AVs, when they will be the main players of the road traffic. In the near future, it has
been assumed that AVs will be very common and people will delegate their driving powers
to them. To answer the question that how AVs will be able to fulfill the expectations of
humans in terms of safer road operations with less number of collisions and harmless
interactions with each other , especially when human drivers have no role in their
operations, this research work has been done. The answer has been provided through the
human social life protocol, which lies in their core, humans, to interact with each other,
avoiding the conflicts, and keeping the social society in equilibrium. The key is following
34
social norms under the influence of primary emotions. Furthermore, the simulation results
have provided optimal parameters, like optimal sonar range and different optimal speeds
suitable for avoiding the road collisions in different road traffic situations. This research
work might be suitable for AV vendors to reinvent the autopilot design, in terms of
including social norms, emotions and optimal operating parameters. Hopefully it will make
AVs capable to cope with the current dilemma that how the AVs make themselves more
trustworthy in terms of safe travelling.
References
F. Riaz and M. A. Niazi, "Road collisions avoidance using vehicular cyber-physical systems: a taxonomy and review,"
Complex Adaptive Systems Modeling, vol. 4, p. 1, 2016.
T. Litman, "Autonomous Vehicle Implementation Predictions," Victoria Transport Policy Institute, vol. 28, 2014.
A. C. Mersky and C. Samaras, "Fuel economy testing of autonomous vehicles," Transportation Research Part C: Emerging
Technologies, vol. 65, pp. 31-48, 2016.
B. F. Malle, "Integrating robot ethics and machine morality: the study and design of moral competence in robots," Ethics and
Information Technology, pp. 1-14, 2015.
A. Rakotonirainy, R. Schroeter, and A. Soro, "Three social car visions to improve driver behaviour," Pervasive and Mobile
Computing, vol. 14, pp. 147-160, 2014.
W. Kumfer and R. Burgess, "Investigation into the Role of Rational Ethics in Crashes of Automated Vehicles,"
Transportation Research Record: Journal of the Transportation Research Board, pp. 130-136, 2015.
J. Elster, "Social norms and economic theory," in Culture and Politics, ed: Springer, 2000, pp. 363-380.
N. Criado, E. Argente, P. Noriega, and V. Botti, "Human-inspired model for norm compliance decision making," Information
Sciences, vol. 245, pp. 218-239, 2013.
A. Staller and P. Petta, "Introducing emotions into the computational study of social norms: A first evaluation," Journal of
artificial societies and social simulation, vol. 4, pp. U27-U60, 2001.
J. C. Gerdes and S. M. Thornton, "Implementable Ethics for Autonomous Vehicles," in Autonomous Driving, ed: Springer,
2016, pp. 87-102.
H. T. Tavani, Ethics and technology: Controversies, questions, and strategies for ethical computing: John Wiley & Sons,
2011.
M. Van de Voort, W. Pieters, and L. Consoli, "Refining the ethics of computer-made decisions: a classification of moral
mediation by ubiquitous machines," Ethics and Information Technology, vol. 17, pp. 41-56, 2015.
A. Etzioni and O. Etzioni, "AI assisted ethics," Ethics and Information Technology, vol. 18, pp. 149-156, 2016.
T. Schmidt, R. Philipsen, and M. Ziefle, "User Diverse Privacy Requirements for V2X-Technology," 2016.
J. Gogoll and J. F. Müller, "Autonomous cars: in favor of a mandatory ethics setting," Science and Engineering Ethics, pp. 1-
20, 2016.
A. R. Hochschild, "Emotion work, feeling rules, and social structure," American journal of sociology, pp. 551-575, 1979.
A. Ortony, G. L. Clore, and A. Collins, The cognitive structure of emotions: Cambridge university press, 1990.
W. G. Najm, J. D. Smith, and M. Yanagisawa, "Pre-crash scenario typology for crash avoidance research," in DOT HS, 2007.
M. S. El-Nasr and J. Yen, "Agents, emotional intelligence and fuzzy logic," in Fuzzy Information Processing Society-
NAFIPS, 1998 Conference of the North American, 1998, pp. 301-305.
M. S. El-Nasr, J. Yen, and T. R. Ioerger, "Flame-fuzzy logic adaptive model of emotions," Autonomous Agents and Multi-
agent systems, vol. 3, pp. 219-257, 2000.
A. Aly and A. Tapus, "An Online Fuzzy-Based Approach for Human Emotions Detection: An Overview on the Human
Cognitive Model of Understanding and Generating Multimodal Actions," in Intelligent Assistive Robots, ed: Springer, 2015,
pp. 185-212.
R. L. Mandryk and M. S. Atkins, "A fuzzy physiological approach for continuously modeling emotion during interaction
with play technologies," International journal of human-computer studies, vol. 65, pp. 329-347, 2007.
C. E. Izard, Human emotions: Springer Science & Business Media, 2013.
M. A. Niazi, "Emergence of a Snake-Like Structure in Mobile Distributed Agents: An Exploratory Agent-Based Modeling
Approach," The Scientific World Journal, vol. 2014, 2014.
35
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
Appendix A
The Intensity of global variables (Ig):
To compute the intensity of Ig variable, five linguistic tokens VLIG, LIG, MIG, HIG and
VHIG were defined, which represent the very low intensity of Low likelihood, medium
likelihood, high likelihood and the very high likelihood respectively. The linguistic tokens
of Ig are presented in table 25.
IG - LINGUISTIC TOKENS AND THEIR DESCRIPTION
TABLE 25
Linguistic Tokens Description
VLIG
LIG
MIG
HIG
VHIG
Very low intensity of global variable
Low intensity of global variable
Medium intensity of global variable
High intensity of global variable
Very high intensity of global variable
The intensity of global variable depends on proximity and a sense of reality variables.
Twenty-five rules were defined to obtain the value of the variable likelihood; these rules
are given in table 26.
TABLE 26
IG - LIKELIHOOD FUZZY INFERENCE RULES
If Sense of Reality
is
VLSOR
VLSOR
VLSOR
VLSOR
VLSOR
LSOR
LSOR
LSOR
LSOR
LSOR
MSOR
MSOR
MSOR
MSOR
MSOR
HSOR
HSOR
HSOR
HSOR
HSOR
VHSOR
VHSOR
And Proximity
is
About to
Going to
MChance
LChance
NChance
About to
Going to
MChance
LChance
NChance
About to
Going to
MChance
LChance
NChance
About to
Going to
MChance
LChance
NChance
About to
Going to
Then Intensity of Goal will
be
MIG
MIG
LIG
VLIG
VLIG
HIG
MIG
MIG
LIG
VLIG
HIG
HIG
MIG
LIG
VLIG
VHIG
HIG
MIG
LIG
VLIG
VHIG
VHIG
36
VHSOR
VHSOR
VHSOR
MChance
LChance
NChance
HIG
HIG
MIG
Experiment A. Computing Undesirability
According to the OCC model, desirability is a local variable, which affects only event, and
agent-based emotions. The desirability variable further comprises two sub-variables: First,
one is the importance of the goal and the second one is the achievement of the goal.
The effects of these two sub-variables in computing the desirability can be seen in the
following scenario. Suppose that the goal of AV is reaching its destination on time.
Suddenly the battery of the AV gets down. Now here the undesirability of the event can
have more than one values. We are just representing here two cases. If the importance of
goal is very high and it has traveled only 30 % of the distance towards its destination, then
undesirability of the said event will be very high. In the second case, if the importance of
goal is very low and it has achieved 100 % of an assigned task (battery gets down after
reaching its destination) then the undesirability of the event will be very low. The main
simulation screen of computing desirability (undesirability in the case of fear) is shown in
figure 10. The screen is showing two input variables and one output variable. The input
variables are the importance of Goal (ImpGoal), achievement of the goal (AchGoal) and
the output variable is Undesirability.
37
Fig. 10 Main simulation screen for Desirability computation
To compute undesirability trigonometric function (trimf) has been used with linguistic
tokens VLUD, LUD, MUD, HUD and VHUD which represent Very low Undesirable, Low
Undesirable, Medium Undesirable, High Undesirable and Very High Undesirable
respectively to represent different intensity levels of undesirability. Twenty-five rules were
defined to obtain the value of the variable undesirability; these rules are given in table 27.
Validation of fuzzy logic rules for computing the Undesirability
The validation of undesirability fuzzy rules has been performed in rule view of FIS editor.
Rule viewer was provided random values for different linguistic tokens and in the result,
fuzzy inference system computed different intensities of undesirability. To cross check the
38
outcomes hand trace mechanism has been adopted, which further validated the outcomes
of different undesirability values shown in table 28. In test 1, it can be seen that the input
variables ImpGoal, AchGoal have values 0.1 and 0.5, which lies in the very low range and
medium range respectively. In result, the FIS system computes low undesirability i.e.
0.25, which is correct. In the same way in test 7 linguistic tokens medium importance of
goal i.e. MImpG, medium achieved goal MAG has values 0.56 and 0.5, which lies in the
medium range. In a result, the FIS system computes medium intensity of undesirability i.e.
DESIRABILITY- LIKELIHOOD FUZZY INFERENCE RULES
TABLE 27
If Importance of Goal is
VLImpG
VLImpG
VLImpG
VLImpG
VLImpG
LImpG
LImpG
LImpG
LImpG
LImpG
MImpG
MImpG
MImpG
MImpG
MImpG
HImpG
HImpG
HImpG
HImpG
HImpG
VHImpG
VHImpG
VHImpG
VHImpG
And Achievement of goal is
NAG
LAG
MAG
HAG
HFAG
NAG
LAG
MAG
HAG
VHFAG
NAG
LAG
MAG
HAG
VHFAG
NAG
LAG
MAG
HAG
VHFAG
NAG
LAG
MAG
HAG
Then undesirability will be
MUD
LUD
LUD
VLUD
VLUD
MUD
MUD
LUD
VLUD
VLUD
HUD
MUD
MUD
LUD
LUD
VHUD
HUD
HUD
MUD
VHUD
VHUD
HUD
HUD
HUD
VHImpG
VHFAG
MUD
0.567, which is correct. In the same way, other validation results can be cross-checked
using hand tracing mechanism.
The following table shows that different values for ImpGoal and AchGoal were entered as
input and each time output value of the Undesirability variable is according to the rules.
39
VALIDATION OF FUZZY LOGIC RULES FOR COMPUTING THE UNDESIRABILITY
TABLE 28
No.
Tests
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Of
ImpGoal
AchGoal
Undesirability
0.1(VLImpG)
0.2(VLImpG)
0.27(LImpG)
O.30(LImpG)
0.4(LImpG)
0.5(MImpG)
0.56(MImpG)
0.6(MImpG)
0.8(HImpG)
0.85(HImpG)
0.79(HImpG)
0.96(VHImpG)
0.98(VHImpG)
1.0(VHImpG)
0.5(MAG)
1.0(VHAG)
0(NAG)
0.5(MAG)
1.0(VHAG)
0(NAG)
0.5(MAG)
1.0(VHAG)
0(NAG)
0.5(MAG)
1.0(VHAG)
0(NAG)
0.5(MAG)
1.0(VHAG)
0.25(LUD)
0.08(VLUD)
0.52(MUD)
0.31(LUD)
0.09(VLUD)
0.74(HUD)
0.567(MUD)
0.09(VLUD)
0.91(VHUD)
0.746(HUD)
0.085(VLUD)
0.917(VHUD)
0.747(HUD)
0.08(VLUD)
Experiment B. Computing Likelihood
The Likelihood of the event depends on the Distance and speed of the following and
leading AVs. In our case, the Likelihood is representing TTA (Time To Avoid). For
example, if the distance between two vehicles is low and their speed is in high range then it
leads to the higher Likelihood of collision between these two vehicles. Therefore, the two
variables, which affect the likelihood of an event, are; the first one is the distance between
both AVs and the second one is the speed of Bullet AV.
The figure 11 is representing the main simulation screen utilized to compute the
Likelihood variable. The screen is showing two input variables and one output variable.
The input variables are Speed and Distance and the output variable is Likelihood as
discussed above.
40
Fig. 11 Main simulation screen for Likelihood computation
Experiment C. Intensity of Global Variable
The intensity of global variable further depends on proximity and the sense of reality
variables. The sense of reality is the scene interpretation by the sensing module of AV or
the reality of the event on which AV believes or not. This variable has the global influence
on the intensity of emotions. Proximity is the distance between the AVs. The proximity
influences the intensity of emotions that can involve future situations. We have taken
proximity here in spatial terms.
Figure 12 depicts the main simulation screen regarding the quantitative computation of Ig
variable. Here the sense of reality and proximity are acting as two input variables to
compute Ig.
41
Fig.12 Main Simulation Screen of the Intensity of global variable
42
|
1312.4048 | 1 | 1312 | 2013-12-14T14:15:28 | Toward an agent based distillation approach for protesting crowd simulation | [
"cs.MA"
] | This paper investigates the problem of protesting crowd simulation. It considers CROCADILE, an agent based distillation system, for this purpose. A model of protesting crowd was determined and then a CROCADILE model of protesting crowd was engineered and demonstrated. We validated the model by using two scenarios where protesters are varied with different personalities. The results indicated that CROCADILE served well as the platform for protesting crowd modeling simulation | cs.MA | cs | TOWARD AN AGENT BASED DISTILLATION APPROACH FOR
PROTESTING CROWD SIMULATION*
Lam Thu BUI and Van Vien MAC
Abstract— This paper investigates the problem of protesting
crowd simulation. It considers CROCADILE, an agent based
distillation system, for this purpose. A model of protesting crowd
was determined and then a CROCADILE model of protesting
crowd was engineered and demonstrated. We validated the model
by using two scenarios where protesters are varied with different
personalities. The results indicated that CROCADILE served well
as the platform for protesting crowd modeling simulation.
I. INTRODUCTION
Crowd control poses a great challenge for authorities.
We are usually expecting a well-managed crowd; however,
it also usually has a surprise factor. If a disruption happened
to the crowd, disorder will appear and hence cause a mess
and casualties accordingly. The research question is how to
predict, manage and control crowd behaviors? A part of the
solution is modeling and simulation [4, 6, 7].
Crowd simulation has been a significant research topic.
There have been several methods proposed for modeling and
simulating the crowd including System Dynamics, Particle
Systems, Cellular Automata and Multi-Agent Systems as
i.e Massive,
well as commercial software systems
AI.implant, Age of Empires 3, and Crowd – MAGS. Among
these, multi agent systems have emerged as a promising
technology and play as the foundation of crowd simulation.
The modeling and simulation methods can be classified
into three types of crowd simulation: (1) Flow-based
Approach, modeling a crowd as continuous flow of fluid [8,
9]; (2) Entity-based Approach, being suitable for medium-
sized and homogeneous crowds. The movement of
individuals is affected by some global/local laws that are
introduced
to
represent
various
physical/
social/
psychological influences, such as flocking; and (3) Agent-
based Approach, being for both medium and small sized
heterogeneous crowds.
In this paper, we propose to investigate the CROCADILE
system, a well-known multi-agent based system for
battlefield simulation and apply it for protesting crowd
simulation. Our research focuses on how to design the crowd
model and map it to CROCADILE. We presented a case
study where all aspects of protesting crowd were quantified
and installed to CROCADILE. Through a case study
adopted from [5], we demonstrated a protesting crowd
model and showed that CROCADILE can be appropriately
extended towards civilian crowd simulation.
The paper structure is organized as follow: the second
section describes CROCADILE system. The third and fourth
ones are for modeling and simulation of protesting crowds
with CROCADILE. The paper is concluded in Section V.
*Lam Thu BUI is with Le Quy Don Technical University, 236 Hoang
Quoc Viet, Ha Noi, Vietnam (e-mail: [email protected]).
Van Vien MAC is with Institute of Information Technology, Ha Noi,
Vietnam (e-mail: [email protected]).
II. CROCADILE SYSTEM
We overview the CROCADILE system, which is used for
modelling the protesting crowd in this paper. CROCADILE
is basically an agent-based simulation system [1]. This
simulation system concerns on emergent behaviour of the
system; it represents the environment at an abstract level and
therefore provides a broad perspective. The main focus of
CROCADILE is that agents are working and interacting
based on the behaviour weight system. The outcome of the
simulations emerges from the interaction among the agents.
That is why CROCADILE is called as an agent-based
“distillation”. Note that the concept of “distillations” is
understood as any agent-based approach that concentrates
on interaction between agents and does not concentrate as
much on individual agents (see also [2, 3, 5]). CROCADILE
applies quite simple techniques to control the motion of
agents on the field, for example calculating the movement
vector for agents based on the behaviour weight vectors.
Under the control of these techniques, the motion of the
agents can often be quite different from the reality of
entities. It has been widely investigated in the area of
military simulation where agents, representing soldiers, are
hard-wired strict rules. There has been a question of how to
extend these systems toward civilian crowd modelling and
simulation? It is the main topic reported in the paper.
The agents in CROCADILE are provided with a wide
range of capabilities, such as sensor, weapon, command,
communication, and movement. In addition, the agents can
belong to different teams, so they can be friendly, neutral, or
hostile with each other. Further, they can communicate or
command their subordinates to do specific tasks (missions).
Agents have different behaviours and a number of triggers
that help them deal with current, arbitrary events.
III. CROWD SIMULATION WITH CROCADILE
A. Protesting crowd characterization
We propose a model of the following characteristics for all
agents and mapping to CROCADILE:
Characteristics Description
CROCADILE’s
capacity
Number of agents,
Position, speed,
sensor,
communications,
health
Weapon, Armor
Strength
Red-teaming structure
1.Commander-
subordinate structure
2.Behavior: Caring the
Size, location,
speed, sensor,
communication,
strength
Physically
describe the
crowd
l
a
c
i
s
y
h
P
Lethality
Organization
l
a
i
c
o
S
Leadership
Level of violence
and property
damage
Likelihood of
injuries and
Deaths
How organized is
the group?
How established
is the leadership?
l
a
c
i
g
o
l
o
h
c
y
s
P
Mission
Aggressiveness
Cohesiveness
Emotion (panic/
angry)
How much
aggression an
agent has?
Have members of
the crowd bonded
with each other?
Is the crowd goal-
oriented?
Is the crowd
emotionally
intense?
commanders
1.Command
considerations
(Formation)
2.Design to maintain
spacing
Mission assignment
Caring behaviors:
1.Caring the allied
2.Caring the friend
3.Caring the opponent
1.Behavior: Caring the
opponent
2.Behavior:Caring the
mission
3.Capability
considerations
Note that our selection of CROCADILE is just because it is
a complete system allowing customization of the agents and
environment; no need to engineer agents from the scratch.
B. Scenario formulation
We followed a similar formulation in [5]. A scenario for
protesting crowd simulation has two sides: the police force
and the protesting crowd. For the police, they are assigned to
protect the objective, keep the formation, and prevent the
protesters penetrating to the area and avoid clashing with the
protesters at the minimum. The police only use the weapon
when they are under an attack or are being approached
closely by the protesters. They have two types of behaviors:
default and being-hit. Their weapon includes the short guns,
tear gas, water cannon and plastic bullets.
For the protesters, they can be divided into several groups
and have the leaders. The groups can be classified based on
the aggressiveness level [11]. We considered the following
types of protester groups:
Passive group: the one with tendency to avoid the police.
They always follow the leaders and maintain the distance
with the partners. If being hit or collided, they will change to
the behavior of running away from the police and cannot
maintain the distance with the partners. They have two types
of behaviors: default and being hit. The passive group
usually does not have the capability of the weapon.
Moderate group: They are moderate people. They are
organized, follow the leaders, and maintain the distance.
They will not change their behavior during the protesting.
Aggressive group: this is for aggressive people. They might
not follow the leader’s orders and also might separate from
their crowd. Their action might include attacking the police
and might be more aggressive if being hit. They might be
equipped with lethal and non-lethal weapon.
IV. A CASE STUDY
We setup two teams (meaning sides): the police and
protesting; each team might have several groups. The values
were adjusted reflecting relative relationship among entities.
A. Behavioral settings:
- The police’s default behavior: The police agents will be
initialized with this behavioral setting (values are ranged
from 0 to 1). There are a lot of behavioral parameters in
CROCADILE; here we described only the ones that we
setup different values from the default value (zero). When
the value is equal zero, it means that behavior is neutral and
will not have effects on the movement of the agent.
Here we setup only two behaviors reflecting this default
behavior: “Caring the opponent” and “Caring the allied”
with the weight values of 0.2 and 0.8 respectively. Also the
“Likelihood of using movement ability” is set at 0.1. Note that
weight is different from the likelihood concept.
- The police’s being-hit behavior: The police will change
their behavior adapting to the new situation as being
attacked. The likelihoods of using movement ability and
using firing ability are increased to maximum (1.0). The
“Caring the allied” is unchanged, however the behavior of
“Caring the opponent” are changed with more details using
sub-behavioral settings.
Note that the sub-behavior’s value ranges from -1 to 1; the
negative value means the sub-behavior will have the
opposite effect. For example instead of getting close to the
opponent, the agent will run away.
Behavior’s
Sub-behavior’s name
name
Value
0.1
0.1
0.1
0.1
0.1
Caring the
opponent
Getting close to the opponent
Getting close to the opponent in the
weapon range
Getting close to the opponent in the
sensor range
Getting close to the wounded opponent
Getting close to the wounded opponent
in the weapon range
Getting close to the wounded opponent
in the sensor range
- The passive protester’s default behavior: We use only
three behaviors “Caring the opponent”, “Caring the allied” and
“Caring the leader” with the values of 0.05, 0.45, and 0.5
respectively. This means they tend to follow the leaders and avoid
the police.
- The passive protester’s being-hit behavior: This
behavior setting is to be used once the agent is hit. In more
details, the agent will run away from the police and does not
care its leaders or partners. The changes are applied only to
all sub-behaviors of “Caring the opponent” with the value of
-0.5
- The moderate protester’s default behavior: This type of
protesters has only one behavioral setting with behaviors
Caring the opponent, Caring the allied, Caring the leaders, Caring
the terrain are set 0.15, 0.35, 0.35, and 0.15 respectively
- The aggressive protester’s default behavior:
Sub-behavior’s name
Value
Behavior’s name
Caring
the
0,4
opponent
Caring the allied
Caring the leaders
Keep the leaders in the
sensor range
Keep the leaders in the
communication range
- The aggressive protester’s being-hit behavior: This
behavior will be triggered whenever the agents got hit. They
will be more aggressive and lawless. The values of
behaviors Caring the opponent, Caring the allied, and Caring the
leaders are 0.5, 0.25, and 0.25 respectively.
value
0,3
0,3
0,5
0.1
0,5
With Short gun
5
10
5
Amount
5
5
5
Amount
25
20
20
B. Other major capability settings
Weapon: The weapon capabilities are included Plastic
Bullet, Water Cannon, Tear Gas, Short gun, Baton, and
Stone where baton and stone are for the protesters while
others are for the police.
Health: A police agent is setup as twice stronger than a
protester (200 units in comparison to that of 100 units for a
protester) reflecting that the police is more well-trained than
the protesters
C. Simulation Scenarios
Our scenarios are assumed in a flat urban terrain (i.e the
park, or city square). A group of protesters gathered for a
while and they are facing the police. To counter the
protesters, the police setups several obstacles (05 fence
objects) separating the protesting and the protected area.
They will act if the protesters break the fence and enter the
area. The numbers of police and protesters are as follows:
Protesters
Police
Items
Items
Passive agents
With plastic bullets
Moderate agents with Stone
With Tear Gas
Aggressive
With Water Canon
agents with
stone
Moderate agents with Baton
Aggressive
agents with
Baton
1
Leader agents with Stone
Since the protesting crowd is very much depending on the
leaders, we propose to investigate in two cases being
equivalent with two types of protesting leaders:
‐ Case 1: Moderate leaders
With this case, the protesting leaders were quite moderate in
which they tend to avoid the police and pay the care to their
movement and formation of the crowd.
Behavior’s name
Caring the opponent
Caring the terrain
‐ Case 2: Aggressive leaders
The leaders are different. They are more aggressive and are
keen to confrontation with the police, meaning the sub-
behaviors of “Caring the opponent” are increased from 0 to
0.1
D. Results and discussion
Based on the results logged during the simulation run, we
reported the summary as in the following table:
Simulated results
Criteria
Case 2
Case 1
Achieved the goal?
1
No
Yes
2
Dead protesters
0
0
3 Wounded protesters
0
22/106
4
Dead Police
0
0
5 Wounded police
8/20
0
Destroyed obstacles
5
8/12
5/12
7
Protester Health Damage
0/10600
220/10600
8
Police Health Damage
0/4000
20/4000
It is quite clear that in the first case, the police was able to
contain the crowd. Although the protesters damaged several
obstacles, they cannot penetrate further after the fence.
Hence, the police successfully achieved their goal. However,
Value
0,45
0,55
the first case demonstrated quite peaceful scenario. There
was no use of the weapon during the time of simulation.
In the second case, the situation is quite similar except the
protester leaders where they were more aggressive. The
dynamic of the scenario was changed accordingly in which
the police failed to prevent the protesters. It again showed
the importance of the leadership in the protesting crowd. In
this case, 22 protesters were wounded in comparison to 8
police. There were lots of violent confrontation between the
police and protesters. Note that both police and aggressive
protesters are triggered when being hit for adapting to the
new situations; during the simulation these triggers were all
activated. For the protesters, once being hit, they become
more aggressive, hence the confrontation is more serious.
In summary, the results supported the usability of
CROCADILE for crowd modeling and simulation. The
agents in CROCADILE are equipped with behaviors and
triggers, which are capable of modeling all protesting crowd
characteristics
V. CONCLUSION
This paper overviewed research on protesting crowd and the
usage of agent-based distillation for crowd simulation. For
the protesting crowd modeling and simulation, we
summarized its characteristics and classification. Based on
this knowledge, we proposed to design a mapping of the
protesting crowd to the model supported by CROCADILE,
an agent based distillation system. To validate it, we carried
out a case study on different scenarios. The results indicated
the appropriateness of CROCADILE for crowd simulation.
REFERENCES
[1] Barlow M. and Easton A., CROCADILE - An Open,
Extensible Agent-Based Distillation Engine, Information &
Security, Vol. 8, No 1, 2002.
[2] Bent N., Socrates v1.1 UserManual, Emergent Information
Technologies inc, USA, 2001.
Stephen R., Agent-Based Combat Model, Project Albert 3rd
International Workshop, Auckland, New Zealand, 2001.
[4] Challenger R., Clegg .C, Robinson M., Understanding Crowd
Behaviors: Supporting Evidence, ISBN 978-1-874321-24-8,
2009
[5] LAROCHELLE B., Multi-Agent Geo-Simulation of Crowds
and Control Forces
in Conflict Situations: Models,
Application, and Analysis, the Master’s Thesis, University of
Laval Quebec, Canada, 2009
[6] Zhou S., Chen D., Cai W., Luo L., Low W., Tian F., Tay V.,
Ong D., and Hamilton B., Crowd modeling and simulation
technologies. ACM Trans. Model. Comp. Sim. 20, 4. 2010
[7] Brown R.. The Handbook of Social Psychology, G. Lindzey
(ed.), Mass Phenomena, pages 833–876. Addison-Wesley,
Cambridge, MA, 1954.
[8] Chenney, S. Flow tiles. Proc. of the ACM SIGGRAPH
symposium on computer animation, 233–242, 2004.
[9] Kisko, T. M., Francis, R. L., and Nobel, C. R. 1998. Evacnet4
user’sguide,
1998.
http://www.ise.ufl.edu/kisko/files/evacnet/EVAC4UG.HTM
[10] Helbing, D., Farkas I ., and Vicsek, T. Simulating dynamical
features of escape panic. Letters to Nature407, 487–490, 2000
[11] McKenzie, F. D., Xu, Q., Nguyen, Q.-A. H., and Petty, M. D.
Designing physical layer components in a reconfigurable
crowd federate. In Proceedings of the Spring 2005 Simulation
Interoperability Workshop, 2005
[3]
|
1301.6400 | 1 | 1301 | 2013-01-27T20:54:07 | Achieving Fully Proportional Representation is Easy in Practice | [
"cs.MA",
"cs.GT"
] | We provide experimental evaluation of a number of known and new algorithms for approximate computation of Monroe's and Chamberlin-Courant's rules. Our experiments, conducted both on real-life preference-aggregation data and on synthetic data, show that even very simple and fast algorithms can in many cases find near-perfect solutions. Our results confirm and complement very recent theoretical analysis of Skowron et al., who have shown good lower bounds on the quality of (some of) the algorithms that we study. | cs.MA | cs |
Achieving Fully Proportional Representation is Easy in
Practice
Piotr Skowron
University of Warsaw
Warsaw, Poland
Piotr Faliszewski
AGH University
Krakow, Poland
Arkadii Slinko
University of Auckland
Auckland, New Zealand
September 7, 2018
Abstract
We provide experimental evaluation of a number of known and new algorithms for approx-
imate computation of Monroe's and Chamberlin-Courant's rules. Our experiments, conducted
both on real-life preference-aggregation data and on synthetic data, show that even very simple
and fast algorithms can in many cases find near-perfect solutions. Our results confirm and com-
plement very recent theoretical analysis of Skowron et al., who have shown good lower bounds
on the quality of (some of) the algorithms that we study.
1 Introduction
Many countries are governed using indirect democracy, where the people do not make decisions
directly, but rather select representatives (e.g., a parliament, a senate, a congress) who rule in their
interest. Unfortunately, relatively little effort was so far invested in the algorithmic study of proce-
dures for electing committees of representatives (few exceptions include papers [2,5,9,14]). Here,
we consider two particularly appealing rules for electing a set of representatives, namely those of
Monroe and of Chamberlin and Courant, and we argue that while these rules in the worst case
scenario may be difficult to compute [5,14], in practice, very simple and efficient algorithms find
almost-perfect approximate results.
There are several ways in which countries can choose their parliaments (or, more generally, in
which societies can choose committees of representatives). Often, voters are divided into districts
and in each district we hold a local election. For the case of single-representative districts, in
each district we have a single-winner election held according to one of the standard, well-known,
rules such as the Plurality rule or Borda's rule.
In particular, if the Plurality rule is used then
this system is known as First-Past-the-Post (FPP): In each district the candidate supported by the
largest number of voters is elected. However, FPP has a number of drawbacks. For example, it
is possible that in a country with two major parties, A and B, even if 49% of the citizens support
party B, only members of party A enter the parliament (this happens if in each district party A has a
slight advantage over party B). Indeed, under FPP the election organizers are particularly tempted
1
to tamper with the partition of voters into districts. To circumvent this problem one might use
multi-representative districts, where elections are held using some multi-winner voting rule (e.g.,
using Single Transferable Vote (STV),1 or using a voting rule that assigns scores to the candidates
and picks a group of those with highest scores). However, this approach only partially solves the
problem. Further, compared to single-representative districts, multi-representative districts loosen
the connection between the candidates and the voters that have elected them.
Fortunately, there is a very attractive way to avoid the problems mentioned above: Instead
of using fixed districts, we may partition the voters dynamically, based on the votes that they cast.
Indeed, this is exactly the idea behind the fully proportional representation rules of Monroe [10] and
of Chamberlin and Courant [3]. If we seek a parliament of K representatives, then Monroe's rule
says that we should pick a set of K candidates for whom there is an assignment of these candidates
to the voters such that: (a) each candidate is assigned to roughly the same number of voters, and
(b) the total satisfaction of the voters (measured in one of the ways introduced later) is maximal.
Chamberlin-Courant's rule is similar except that it allows each selected candidate to be matched to
a different number of voters. (Thus if one were to elect a parliament using Chamberlin-Courant's
rule then one should use weighted voting within the parliament, weighted by the number of voters
matched to each representative.)
In the above description we focus on political elections, but we mention that both Monroe's rule
and Chamberlin-Courant's rule have many different applications as well. For example, Skowron
et al. [15] have presented several (multi-agent) resource allocation settings that can be modeled
using these rules and Lu and Boutilier [5] have proposed to use Chamberlin-Courant's rule for
constructing recommendations for groups of agents.
Unfortunately, computing Monroe's and Chamberlin-Courant's rules is both NP-hard [5,14]
and difficult in the parametrized sense [2]. Thus using them in practice might simply be impossible.
The goal of this paper is to show that not all is lost. We provide experimental evaluation of a number
of known and new algorithms for approximate computation of Monroe's and Chamberlin-Courant's
rules. Our experiments, conducted both on real-life preference-aggregation data and on synthetic
data, show that even very simple and fast algorithms can in many cases find near-perfect solutions.
Our results confirm and complement very recent theoretical analysis of Skowron et al. [15], who
have shown good lower bounds on the quality of (some of) our algorithms. While for single-winner
rules using approximate algorithms may be debatable, for the case of electing a large body of repre-
sentatives, e.g., a parliament, using approximation algorithms seems far easier to justify. Indeed, a
good approximate solution for Monroe's or Chamberlin-Courant's rule represents the society almost
as well as a perfect solution would.
The paper is organized as follows. In Section 2 we formally define Monroe's and Chamberlin-
Courant's rules. In Section 3 we give an overview of the algorithms that we evaluate and in Section 4
we describe the data sets that we use in our experiments. Section 5 contains our main results. We
conclude in Section 6.
1STV for more than one winner is sometimes referred to as "Alternative Vote" (AV). In a recent referendum Great
Britain rejected AV as a method for choosing its members of parliament.
2
2 Preliminaries
In this section we briefly review basic notions regarding social choice theory and we define Mon-
roe's [10] and Chamberlin and Courant's [3] proportional representation systems. We assume the
reader is familiar with standard notions regarding algorithms. For each positive integer n, by [n] we
mean the set {1, . . . , n}.
Elections. We consider elections over a given set A = {a1, . . . , am} of alternatives. We have a set
N = [n] of agents (the voters), where each voter i, 1 ≤ i ≤ n, has a preference order ≻i over A. A
preference order of an agent i is a linear order over the set A; the maximal element is this agent's
most preferred alternative, the minimal element is this agent's least preferred alternative, and the
alternatives in the middle represent the agent's spectrum of preference. We refer to the collection
V = (≻1, . . . , ≻n) as the preference profile for a given election.
Let us fix an agent i, 1 ≤ i ≤ n, and an alternative a ∈ A. By posi(a) we mean the position a
has in i's preference order. If a is i's most preferred candidate then posi(a) = 1, and if a is i's least
preferred candidate then posi(a) = kAk = m.
Positional Scoring Functions. Let m be the number of candidates in eleciton. A positional scor-
ing function (PSF) is any function α : [m] → N that satisfies the following two conditions: (a)
α(m) = 0, and (b) for each i, j, 1 ≤ i < j ≤ m, α(i) ≥ α(j). In Monroe's and in Chamberlin-
Courant's proportional representation rules we will match agents to the alternatives that represent
them. Intuitively, α(i) is the amount of satsifaction that an agent derives from being represented by
an alternative that this agent ranks on the i'th position. In this paper we focus on Borda count PSF,
Borda(i) = m − i. However, occasionally we will consider
which for m alternatives is defined as αm
other PSFs as well.
In our algorithms we assume that the PSF α to be used is given explcitly, as a vector
(α1, . . . , αm) of integers such that for each i, 1 ≤ i ≤ m, α(i) = αi. We will implcitly assume that
the number of alternatives matches the domain of the given PSF.
Proportional Representation. Let A = {a1, . . . , am} be the set of alternatives and N = [n] be the
set of agents (with preference orders over A). A representation function is any function Φ : N → A.
For an m-candidate PSF α and a representation function Φ, Φ's satisfaction is defined as:
α(Φ) =
n
X
i=1
α(posi(Φ(i))).
Let K be a positive integer. A K-CC-representation function is any representation function
Φ such that kΦ−1(N )k ≤ K (that is, any representation function that matches voters to at
most K alternatives). A K-Monroe-representation function Φ is any K-CC-representation func-
tion that additionally satisfies the following requirement: For each a ∈ A it holds that either
K ⌋ ≤ kΦ−1(a)k ≤ ⌈ n
⌊ n
K ⌉ or kΦ−1(a)k = 0 (that is, each alternative represents either roughly
n
K agents or none of them).
We will also consider partial representation functions. A partial CC-representation function is
defined in the same way as a regular one, except that it may assign a null alternative, ⊥, to some
of the agents. By convention, we take that for each agent i we have posi(⊥) = m. A partial
3
Monroe-representation function is defined analogously: It may assign the null alternative to some
voters (there are no constraints on the number of agents to whom the null alternative is assigned)
but it must be possible to extend it to a regular Monroe-representation function by replacing the
occurrences of the null alternative with the real ones.
We now define Monroe's and Chamberlin-Courant's (CC) rules.
Definition 1. Let R be a member of {Monroe, CC}. Let A = {a1, . . . , am} be a set of alter-
natives, N = [n] be a set of agents, and α be an m-candidate PSF. Let K be the size of the set
of representatives that we seek (K ≤ m). We say that a K-element set W , W ⊆ A, is a set of
α-R winners if there exists a K-R-representation function Φ : N → W such that for every other
K-R-representation function Ψ it holds that α(Φ) ≥ α(Ψ).
We point out that for both Monroe's and Chamberlin-Courant's rule there may be several dif-
ferent winner sets and that some form of tie-breaking should be applied in these settings. Here we
disregard tie-breaking and simply are interested in some winner set (and, not being able to compute
that, in any set with as high a satsifaction as possible).
It is well-known that for many natural families of PSFs, both for Monroe's rule and for
Chamberlin-Courant's rule, it is NP-complete to decide if there exists a winner set that achieves
a given satisfaction [2,5,14]. However, for each R in {Monroe, CC}, for each PSF α (with the
domain matching the number of alternatives in the election), and for each set S of up to K alterna-
tives we can compute in polynomial time a (possibly partial) K-R-representation function ΦS
R that
maximizes the agent satisfaction under the condition that agents are matched to the alternatives in
S only. Indeed, it is easy to see that for α-CC this function is:
ΦS
CC(i) = argmina∈S posi(a)
and that it is never a partial representation function. For the case of α-Monroe, computing ΦS
Monroe
is more involved and requires solving a certain min-cost/max-flow problem (see the work of Betzler
et al. [2]; here if kSk < K then ΦS
Monroe is a partial Monroe-representation function). One can
see that for a given set S, there may be many different (partial) K-Monroe-representation functions
that achieve optimal satisfaction; when we write ΦS
Monroe, we mean, w.l.o.g., the particular one
computed by the algorithm of Betzler et al. [2].
3 Algorithms
Let us now describe the algorithms that we will consider in this work. Some of our algorihtms can
be applied both to Monroe's rule and to Chamberlin-Courtant's rule, while some are specific to only
one of them. For each algorithm we will exactly specify for which rules it is applicable and, if it is
applicable to both, what are the differences.
While most of the algorithms described below are based on ones already given in the literature,
in a number of cases we added heuristics on top of existing algorithms (which proved to be quite
effective, as we will see later) and, in one case, provided a completely new theortical analysis.
For each algorithm we will carefully describe what was already known in the literature, and which
additions are due to this paper.
4
Throughout this section we assume we are given the following setting. A = {a1, . . . , am} is a
set of alternatives, α is an m-candidate PSF, N = [n] is a set of agents, each with a preference order
over A, and K is a positive integer, K ≤ m (the size of the committee we want to elect).
3.1 ILP Formulation (Monroe and CC)
To measure the quality of our approximation algorithms, we compare their results against op-
timal solutions that we obtain using integer linear programs (ILPs) that describe Monroe's and
Chamberlin-Courant's rules. An ILP for Chamberlin-Courant's rule, for arbitraty PSF α, was pro-
vided by Lu and Boutilier [5]; the analogous formulation for Monroe's rule was provided by Potthoff
and Brams [13]. We used the GLPK 4.47 package (GNU Linear Programming Kit, version 4.47) to
solve these ILPs, whenever it was possible to do so in reasonable time.
3.2 Algorithms A, B, and C (Monroe)
Skowron et al. [15] have suggested and studied the following algorithm for Monroe's rule, which
we will call Algorithm A. We start with an empty partial Monroe-representation function Φ and we
execute K iterations. In each iteration we do the following:
1. For each alternative a ∈ A that does not yet represent any agents, we compute the maximal
satisfaction that some not-yet-represented ⌈ n
K ⌉ agents derive from being represented by a (we call
this number score(a) and we refer to these agents as bests(a)).
2. We pick an alternative a with maximum score(a) and extend Φ by assigning a to represent
agents in bests(a).
Borda it
This algorithm clearly works in polynomial time. Skowron et al. [15] have shown that for αm
finds a solution whose satisfaction is at least a (1− K−1
K ) fraction of a (possibly nonexistent)
perfect solution, where each agent is represented by his or her top preference (HK is the K'th
1
harmonic number, i.e., HK = Pi=1
i = Θ(log K)). This suggests that the algorithm performs
best in elections where the size of the committee we seek is relatively small with respect to the
number of alternatives.
2(m−1) − HK
Based on Algorithm A we have derived Algorithm B. The only difference is that after complet-
ing the operation of Algorithm A, we take the set S of alternatives that were assigned to represent
some agents by Algorithm A, and replace function Φ with function ΦS
Monroe, that optimally reas-
signs the alternatives to the voters. This very simple heuristic turned out to noticeably improve the
results of the algorithm in practice (and, of course, the approximation guarantees carry over from
Algorithm A to Algorithm B).
Algorithm C is a further heuristic improvement over Algorithm B. This time the idea is that
instead of keeping only one partial function Φ, we keep a list of up to d partial representation func-
tions, where d is a parameter of the algorithm. At each iteration, given these d partial representation
functions, for each Φ of them and for each alternative a that does not yet have agents assigned to by
this Φ, we compute an optimal extension of this Φ that assigns agents to a. As a result we obtain
possibly more than d (partial) representation functions. For the next iteration we keep those d of
them that have highest satisfaction.
5
Figure 1: The pseudocode for Algorithm C.
Notation: Φ ← a map defining a (partial) representation function, iteratively built by the algorithm.
Φ← ← the set of agents already represented by some alternative
Φ→ ← the set of alternatives already used in the representation function.
P ar ← a list of partial representation functions
P ar = []
P ar.push({})
for i ← 1 to K do
newP ar = []
for Φ ∈ P ar do
score ← {}
bests ← {}
foreach ai ∈ A \ Φ→ do
agents ← sort N \ Φ← so that j ≺ k in agents
=⇒ posj(ai) ≤ posk(ai)
K ⌉ elements of agents
bests[ai] ← chose first ⌈ N
Φ′ ← Φ
foreach j ∈ bests[ai] do
Φ′[j] ← ai
newP ar.push(Φ′)
sort newP ar according to descending order of the total satisfaction of the assigned agents
P ar ← chose first d elements of newP ar
for Φ ∈ P ar do
Φ ← compute the optimal representative function using an algorithm of Betzler et al. [2] for the
set of winners Φ→
return the best representative function from P ar
We provide pseudocode for Algorithm C in Figure 1. If we take d = 1, we obtain Algorithm B.
If we also disregard the last two lines prior to returning solution, we obtain Algorithm A.
3.3 Algorithm GM (Monroe and CC)
Algorithm GM (greedy marginal improvement) was introduced by Lu and Boutilier for the case
of Chamberlin-Courant's rule. Here we generalize it to apply to Monroe's rule as well, and we
show that it is a 1 − 1
e approximation algorithm for α-Monroe. We point out that this is the first
approximation result for Monroe rule that applies to all PSFs α (approximability results of Lu and
Boutilier [5] did not apply to α-Monroe, and results of Skowron et al. [15] applied to Monroe with
Borda count PSF only). For the case of Monroe, the algorithm can also be viewed as an extension
of Algorithm B.
Let R be one of Monroe and CC. The algorithm proceeds as follows. We start with an emtpy
set S. Then we execute K iterations. In each iteration we find an alternative a that is not assigned to
agents yet, and maximizes the value ΦS∪{a}
. (A certain disadvantage of this algorithm for the case
of Monroe is that it requires a large number of computations of ΦS
Monroe, which is a slow process
based on min-cost/max-flow algorithm.) We provide the pseudocode for Algorithm GM in Figure 2.
R
Theorem 1. Algorithm GM is an (1 − 1/e)-approximation algorithm for the Monroe'e election
6
Figure 2: Pseudocode for Algorithm GM.
Notation: R is either Monroe or CC.
S ← ∅
for i ← 1 to K do
a ← argmaxa∈A\Sα(ΦS∪{a}
S ← S ∪ {a}
R
return ΦS
M
)
problem for arbitrary positional scoring functions.
Proof. The proof is based on the powerful result of Nemhauser et al. [11], who have shown that
greedy algorithms achieve 1 − 1
e approximation ratio when used to optimize submodular functions.
Let A be a set of alternatives, N = [n] a set of agents with preferences over A, α an kAk-candidate
PSF, and K ≤ kAk the number of representatives that we want to elect.
We consider function z : 2A → N defined, for each set S, S ⊆ A and kSk ≤ K, as z(S) =
Monroe). Clearly, z(S) is monotonic (that is, for each two sets A and B, if A ⊆ B and kBk ≤ K
α(ΦS
then z(A) ≤ z(B). The main part of the proof below is to show that z is submodular (we provide
the definition below).
Since argmaxS⊂A,kSk=Kz(S) is the set of winners of our election (under α-Monroe) and since
Algorithm GM builds the solution iteratively by greedily extending initially empty set S so that
each iteration increases the value of z(S) maximally, by the results of Nemhauser et al. [11] we get
that Algorithm GM is a (1 − 1
e )-approximation algorithm.
Let us now prove that z is submodular. That is, our goal is to show that for each two sets S and
T , S ⊂ T , and each alternative a /∈ T it holds that z(S ∪ {a}) − z(S) ≥ z(T ∪ {a}) − z(T ). First,
we introduce a notion that generalizes the notion of a partial set of winners S. Let s : A → N denote
a function that assigns a capacity to each alternative (i.e., s gives a bound on the number of agents
that a given alternative can represent). Intuitively, each set S, S ⊆ A, corresponds to the capacity
function that assigns ⌈ n
k ⌉ to each alternative a ∈ S and 0 to each a /∈ S. Given a capacity function
Monroe to be one that maximizes the total satisfaction of the agents
s, we define a partial solution Φs
Monroe)−1(a)k ≤ s(a). To simplify notation, we
and that satisfies the capacity constraints: ∀a∈Sk(Φs
write s∪{a} to denote the function such that (s∪{a})(a) = s(a)+1 and ∀a′∈S(s∪{a})(a′) = s(a′).
(Analogously, we interpret s\{a} as subtracting one from the capacity for a; provided it is nonzero.)
Also, by s ≤ t we mean that ∀a∈As(a) ≤ t(a). We extend our function z to allow us to consider
a subset of the agents only. For each subset N ′ of the agents and each capacity function s, we
define zN ′(s) to be the satisfaction of the agents in N ′ obtained under Φs
Monroe. We will now prove
a stronger variant of submodularity for our extended z. That is, we will show that for each two
capacity functions s and t it holds that:
s ≤ t ⇒ zN (s ∪ {a}) − zN (s) ≥ zN (t ∪ {a}) − zN (t)
(1)
Our proof is by induction on N. Clearly, Equation (1) holds for N ′ = ∅. Now, assuming that
Equation (1) holds for every N ′ ⊂ N we will prove its correctness for N. Let i denote an agent such
7
Monroe(i) = a (if there is no such agent then clearly the equation holds). Let as = Φs
Monroe(i)
that Φt∪{a}
and at = Φt
Monroe(i). We have:
zN (t ∪ {a}) − zN (t) = α(posi(a)) + zN \{i}(t) − α(posi(at)) − zN \{i}(t \ {at}).
We also have:
zN (s ∪ {a}) − zN (s) ≥ α(posi(a)) + zN \{i}(s) − α(posi(as)) − zN \{i}(s \ {as}).
Since Φt
have that:
Monroe describes an optimal representation function under the capacity restrictions t, we
α(posi(at)) + zN \{i}(t \ at) ≥ α(posi(as)) + zN \{i}(t \ {as})
Finally, from the inductive hypothesis for N ′ = N \ {i} we have:
zN \{i}(s) − zN \{i}(s \ {as}) ≥ zN \{i}(t) − zN \{i}(t \ {as})
By combining these inequalities we get:
zN (s ∪ {a}) − zN (s) ≥ α(posi(a)) + zN \{i}(s) − (α(posi(as)) + zN \{i}(s \ {as}))
≥ α(posi(a)) − α(posi(as)) + zN \{i}(t) − zN \{i}(t \ {as})
≥ α(posi(a)) + zN \{i}(t) − α(posi(at)) − zN \{i}(t \ {at})
= zN (t ∪ {a}) − zN (t)
This completes the proof.
3.4 Algorithm C (CC)
This algorithm, introduced in this paper, proceeds like Algorithm GM for Chamberlin-Courant's
rule, but in each iteration it keeps up to d (partial) CC-representation functions ΦS
CC, for distinct
subsets S of alternatives (d is a parameter of the algorithm). In each iteration the algorithm extends
each function ΦS
CC by every possible alternative (obtaining O(dm) new representation functions)
and stores up to d of them, that obtain highest satisfaction.
3.5 Algorithm P (CC)
Algorithm P (position restriction) was introduced and studied by Skowron et al. [15]. The algorithm
proceeds as follows. First, we consider a certain number x (specifically, x = ⌈ mW(K)
K ⌉, where W(x)
is Lambert's W function, defined as the solution of equality x = W(x)eW(x)). Then, the algorithm
tries to greedily find a cover of as many agents as possible with K alternatives (an alternative is
said to cover a given agent if this agent ranks this alternative among top x positions). Skowron et
al. [15] have shown that for αm
times
worse than a perfect (possibly nonexistent) solution, where every agent is represented by his or her
top-preferred alternative. The pseudocode for Algorithm P is presented in Figure 3.
Borda this algorithm finds a solution that is at most 1 − 2W(K)
K
8
Figure 3: Pseudocode for Algorithm P.
Notation: We use the same notation as in Algorithm 1
num posx(a) ← k{i ∈ [n] \ Φ← : posi(a) ≤ x}k
(the number of not-yet assigned agents that rank alternative a in one of their first x positions)
Φ = {}
x = ⌈ mW(K)
K ⌉
for i ← 1 to K do
ai ← argmaxa∈A\Φ→num posx(a)
foreach j ∈ [n] \ Φ← do
if posj(ai) < x then
Φ[j] ← ai
foreach j ∈ A \ Φ← do
a ← such server from Φ→ that ∀a′∈Φ→ posj(a) ≤ posj(a′)
Φ[j] ← a
3.6 Algorithm R (Monroe and CC)
2 (1 + K
m − K 2
m2−m + K 3
Kǫ2
Algorithm R (random sampling) is based on picking the set of winners randomly and matching them
optimally to the agents. Skowron et al. [15] have shown that if one chooses a set S of K alterna-
tives uniformly at random, then for αm
Monroe)
is 1
times, to
reach probability λ of achieving this satisfaction. For example, for λ = 0.99 and ǫ = 0.1 this algo-
rithm would require to repeat the sampling process 340000/K times (each time executing a costly
matching algorithm). This makes the algorithm impractical, especially for small instances (where
K is low). Thus in our experimental evaluation we will consider the modification of the algorithm
that repeats the sampling process only 100 times.
m3−m2 ) − ǫ, and that one has to repeat this process −512 log(1−λ)
Borda-Monroe, the expected satisfaction of αm
Borda(ΦS
Oren [12] has shown an analogous result for the case of Chamberlin-Courant's rule.
3.7 Summary of the Algorithms
We summarize the algorithms that we use in Table 1. In particular, the table clearly shows that for
the case of Monroe, Algorithms B and C are not much slower than Algorithm A but offer a chance
of improved peformance. Algorithm GM is intuitively even more appealing, but achieves this at
the cost of high time complexity. For the case of Chamberlin-Courant's rule, it is unclear which
of the algorithms to expect to be superior. One of the main goals of this paper is to establish if
either of the presented algorithms clearly dominates the others. Our implementations are available
at http://mimuw.edu.pl/ps219737/monroe/experiments.tar.gz.
4 Experimental Data
We have considered both real-life preference-aggregation data and synthetic data, generated accord-
ing to a number of election models.
9
Algorithm Approximation ratio for Borda PSF
A
B
C
GM
R
P
GM
C
R
2(m−1) − HK
K
1 − K−1
as in Alg. A
as in Alg. A
as in Alg. A for Borda PSF; 1 − 1
1
2 (1 + K
1 − 2W(K)
1 − 1
e
as in Alg. GM
(1 − 1
m − K 2m−K 3
m3−m2 )
K+1 )(1 + 1
m )
K
e for others KmO(ΦS)
Runtime
Kmn
Kmn+O(ΦS)
dKmn+dO(ΦS)
log(1−λ)
Kǫ2 O(ΦS)
nmW(K)
Kmn
dKm(n+log dm)
log(1−λ)
n
ǫ2
Reference
Skowron et al. [15]
(this paper)
(this paper)
(this paper)
Skowron et al. [15]
Skowron et al. [15]
Lu and Boutilier [5]
(this paper)
Oren [12]
Table 1: A summary of the algorithms studied in this paper. The top of the table presents algorithms
for Monroe's rule and the bottom for Chamberlin-Courant's rule. In column "Approx." we give cur-
rently known approximation ratio for the algorithm under Borda PSF, on profiles with m candidates
and where the goal is to select a committee of size K. Here, O(ΦS) = O(n2(K + logn)) is the
complexity of finding a partial representation function with the algorithm of Betzler et al. [2].
4.1 Real-Life Data
We have used real-life data regarding people's preference on sushi types, movies, college courses,
and competitors' performance in figure-skating competitions. One of the major problems regarding
real-life preference data is that either people express preferences over a very limited set of alterna-
tives, or their preference orders are partial. To address the latter issue, for each such data set we
complemented the partial orders to be total orders using the technique of Kamishima [4]. (The idea
is to complete each preference order based on those reported preference orders that appear to be
similar.)
Some of our data sets contain a single profile, whereas the others contain multiple profiles.
When preparing data for a given number m of candidates and a given number n of voters from a
given data set, we used the following method: We first uniformly at random chose a profile within
the data set, and then we randomly selected n voters and m candidates. We used preference orders
of these n voters restricted to these m candidates.
Sushi Preferneces. We used the set of preferences regarding sushi types collected by Kamishima
[4].2 Kamishima has collected two sets of preferences, which we call S1 and S2. Data set S1
contains complete rankings of 10 alternatives collected from 5000 voters. S2 contains partial
rankings of 5000 voters over a set of 100 alternatives (each vote ranks 10 alternatives). We used
Kamishima [4] technique to obtain total rankings.
Movie Preferences. Following Mattei et al. [8], we have used the NetFlix data set3 of movie
preferences (we call it MV). NetFlix data set contains ratings collected from about 480 thousand
2The sushi data set is available under the following url: http://www.kamishima.net/sushi/
3http://www.netflixprize.com/
10
distinct users regarding 18 thousand movies. The users rated movies by giving them a score between
1 (bad) and 5 (good). The set contains about 100 million ratings. We have generated 50 profiles
using the following method: For each profile we have randomly selected 300 movies, picked 10000
users that ranked the highest number of the selected movies, and for each user we have extended his
or her ratings to a complete preference order using the method of Kamishima [4].
Course Preferences. Each year the students at the AGH University choose courses that they
would like to attend. The students are offered a choice of six courses of which they have to at-
tend three. Thus the students are asked to give an unordered set of their three top-preferred courses
and a ranking of the remaining ones (in case too many students select a course, those with the
highest GPA are enrolled and the remaining ones are moved to their less-preferred courses). In
this data set, which we call CR, we have 120 voters (students) and 6 alternatives (courses). How-
ever, due to the nature of the data, instead of using Borda count PSF as the satisfaction measure,
we have used the vector (3, 3, 3, 2, 1, 0). We made this data set publicly available under the url:
http://mimuw.edu.pl/ps219737/monroe/registration.tar.gz.
Figure Skating. This data set, which we call SK, contains preferences of the judges over the
performances in a figure-skating competitions. The data set contains 48 profiles, each describing a
single competition. Each profile contains preference orders of 9 judges over about 20 participants.
The competitions include European skating championships, Olympics, World Junior, and World
Championships, all from 19984. (Note that while in figure skating judges provide numerical scores,
this data set is preprocessed to contain preference orders.)
4.2 Synthetic Data
For our tests, we have also used profiles generated using three well-known distributions of prefer-
ence orders.
Impartial Culture. Under impartial culture model of preferences (which we denote IC), for a
given set A of alternatives, each voter's preference order is drawn uniformly at random from the
set of all possible total orders over A. While not very realistic, profiles generated using impartial
culture model are a standard testbed of election-related algorithms.
Polya-Eggenberger Urn Model. Following Walsh [16], we have used the Polya-Eggenberger urn
model [1] (which we denote UR). In this model we generate votes as follows. We have a set A of
m alternatives and an urn that initially contains all m! preference orders over A. To generate a
vote, we simply randomly pick one from the urn (this is our generated vote), and then -- to simulate
correlation between voters -- we return a copies of this vote to the urn. When generating an election
with m candidates using the urn model, we have set the parameter a so that a
m! = 0.05 (Walsh [16]
calls this parameter b; we mention that other authors use much higher values of b but we felt that
too high a value of b leads to a much too strong a correlation between votes).
Generalized Mallow's Model. We refer to this data set as ML. Let ≻ and ≻′ be two preference
orders over some alternative set A. Kendal-Tau distance between ≻ and ≻′, denoted dK(≻, ≻′),
4This data set is available under the following url: http://rangevoting.org/SkateData1998.txt.
11
is defined as the number of pairs of candidates x, y ∈ A such that either x ≻ y ∧ y ≻′ x or
y ≻ x ∧ x ≻′ y.
Under Mallow's distribution of preferences [7] we are given two parameters: A center prefer-
ence order ≻ and a number φ between 0 and 1. The model says that the probability of generating
preference order ≻′ is proportional to the value φdK (≻,≻′). To generate preference orders following
Mallow's distribution, we use the algorithm given by Lu and Boutilier [6].
In our experiments, we have used a mixture of Mallow's models. Let A be a set of al-
ternatives and let a be a positive integer. This mixture model is parametrized by three vec-
tors, Λ = (λ1, . . . , λa) (where each λi, 1 ≤ i ≤ a is between 0 and 1, and Pa
i=1 λ1 = 1),
Φ = (φ1, . . . , φa) (where each φi, 1 ≤ i ≤ a, is a number between 0 and 1), and Π = (≻1, . . . , ≻a)
(where each ≻i, 1 ≤ i ≤ a, is a preference order over A). To generate a vote, we pick a random in-
teger i, 1 ≤ i ≤ a (each i is chosen with probability λi), and then generate the vote using Mallow's
model with parameters (≻i, φi).
For our experiments we have used a = 5, and we have generated vectors Λ, Φ, and Π uniformly
at random.
5 Experiments
In this section we present the results of the evaluation of algorithms from Section 3 on the data sets
from Section 4. In all cases, except for the college courses data set, we have used Borda PSF to
measure voter satisfaction. For the case of the courses data set, we have used vector (3, 3, 3, 2, 1, 0).
We have conducted three sets of experiments. First, we have tested all our algorithms on rela-
tively small elections (up to 10 candidates, up to 100 agents). In this case we were able to compare
our algorithms' solutions with the optimal ones. (To obtain the optimal solutions we were using the
ILP formulations and GLPK's ILP solver.) Thus we report the quality of our algorithms as the av-
erage of fractions C/Copt, where C is the satisfaction obtained by a respective algorithm and Copt
is the satisfaction in the optimal solution. For each algorithm and data set, we also report the aver-
age fraction C/Cideal, where Cideal is the satisfaction that the voters would have obtained if each
of them were matched to his or her most preferred alternative. In our further experiments, where
we consider larger elections, we were not able to compute optimal solutions, but fraction C/Cideal
gives a lower bound for C/Copt. We report this value for small elections so that we see an example
of relation between C/Copt and C/Cideal and so that we can compare the results for small elections
with the results for the larger ones. Further, for the case of Borda PSF the C/Cideal fraction has
a very natural interpretation: If its value is α (for a given solution), then, on the average, in this
solution each voter is matched to an alternative that he or she prefers to (m − 1)α alternatives.
In our second set of experiments we have run our algorithms on large elections (thousands
of agents, hundreds of alternatives), coming either from the NetFlix data set or generated by us
using one of our models. Here we reported the average fraction C/Cideal only. We have analyzed
the quality of the solutions as a function of the number of agents, the number of candidates, and
the relative number of winners (fraction K/m). (This last set of results is particularly interesting
because in addition to measuring the quality of our algorithms, it allows one to asses the size of a
committee one should seek if a given average agent satisfaction is to be obtained).
12
Monroe
C
A
0.94
0.95
B
S1
0.99 ≈ 1.0
S2
0.99
MV 0.96 ≈ 1.0
CR
0.98
0.99
SK
0.99 ≈ 1.0
IC
0.94
ML
0.94
UR
0.95
0.99 ≈ 1.0
0.99
0.99 ≈ 1.0
1.0
1.0
1.0
1.0
1.0
GM
0.99
≈ 1.0
≈ 1.0
≈ 1.0
≈ 1.0
0.99
0.99
0.99
R
0.99
0.99
0.98
0.99
0.94
0.99
0.99
0.99
CC
GM
C
1.0 ≈ 1.0
1.0 ≈ 1.0
1.0 ≈ 1.0
1.0 ≈ 1.0
1.0 ≈ 1.0
1.0 ≈ 1.0
1.0 ≈ 1.0
0.99
1.0
P
R
0.99
0.99
0.98
0.99
0.96 ≈ 1.0
1.0 ≈ 1.0
0.99
0.85
0.99
0.99
0.99
0.99
0.99
0.97
Table 2: The average quality of the algorithms compared with the optimal solution (C/Copt) for the
small instances of data and for K = 3.
Monroe
C
1.0
B
A
S1
0.95 ≈ 1.0
S2
0.94
MV 0.95
CR
SK
IC
ML
UR
0.99 ≈ 1.0
0.99
0.96 ≈ 1.0
0.99 ≈ 1.0
0.95
0.95
0.96
0.99 ≈ 1.0
0.99 ≈ 1.0
0.99 ≈ 1.0
1.0
1.0
1.0
R
GM
0.99
0.99
0.99
0.99
≈ 1.0
0.98
≈ 1.0
0.99
≈ 1.0
0.88
0.99
0.99
0.99
0.99
0.99 ≈ 1.0
CC
C
GM
1.0 ≈ 1.0
1.0 ≈ 1.0
1.0 ≈ 1.0
1.0
1.0
1.0 ≈ 1.0
1.0 ≈ 1.0
1.0 ≈ 1.0
1.0
1.0
P
R
0.97
0.99
0.98 ≈ 1.0
0.97 ≈ 1.0
1.0
1.00
0.91 ≈ 1.0
0.99
0.99
0.99
0.98
0.96
0.99
Table 3: The average quality of the algorithms compared with the optimal solution (C/Copt) for the
small instances of data and for K = 6.
In the third set of experiments we have measured running times of our algorithms and of the
ILP solver.
5.1 Evaluation on Small Instances
We now present the results of our experiments on small elections. For each data set, we generated
elections with the number of agents n = 100 (n = 9 for data set SK because there are only 9
voters there) and with the number of alternatives m = 10 (m = 6 for data set CR because there
are only 6 alternatives there) using the method described in Section 4.1 for the real-life data sets,
and in the natural obvious way for synthetic data. For each algorithm and for each data set we ran
500 experiments on different instances for K = 3 (for the CR data set we used K = 2) and 500
experiments for K = 6 (for CR we set K = 4). For Algorithms C we set the parameter d = 15.
The results (average fractions C/Copt and C/Cideal) for K = 3 are given in Tables 2 and 4; the
results for K = 6 are given in Tables 3 and 5 (they are almost identical as for K = 3). For each
13
Monroe
CC
A
S1
0.85
S2
0.85
MV 0.88
CR
0.94
SK
0.96
IC
0.8
ML
0.83
UR
0.8
B
0.89
0.89
0.92
0.97
0.96
0.84
0.88
0.85
C
0.9
0.89
0.92
0.96
0.97
0.85
0.88
0.86
GM R
0.89
0.89
0.89
0.89
0.91
0.92
0.96
0.96
0.91
0.97
0.84
0.84
0.9
0.88
0.85
0.87
C
0.92
0.93
0.97
0.97
1.0
0.85
0.92
0.9
GM
0.89
0.9
0.92
0.97
0.97
0.83
0.90
0.87
P
0.91
0.91
0.93
0.97
0.82
0.84
0.89
0.87
R
0.92
0.92
0.97
0.97
0.99
0.85
0.94
0.89
Table 4: The average quality of the algorithms compared with the simple lower bound (C/Cideal)
for the small instances of data and for K = 3.
A
S1
0.91
S2
0.88
Mv
0.85
Cr
0.95
Sk
0.91
IC 0.91
Ml
0.89
Ur
0.91
Monroe
B
0.96
0.93
0.89
0.98
0.92
0.95
0.94
0.95
C
0.96
0.93
0.89
0.99
0.92
0.95
0.94
0.95
GM R
0.95
0.95
0.93
0.93
0.88
0.89
0.98
0.99
0.92
0.81
0.95
0.94
0.93
0.94
0.94
0.95
C
0.98
0.98
0.99
1.0
1.0
0.96
0.97
0.98
GM
0.98
0.98
0.99
1.0
1.0
0.96
0.98
0.98
CC
P
R
0.98
0.98
0.99
1.0
0.96
0.96
0.97
1.0
0.91 ≈ 1.0
0.95
0.95
0.98
0.95
0.94
0.97
Table 5: The average quality of the algorithms compared with the simple lower bound (C/Cideal)
for the small instances of data and for K = 6.
experiment in this section we also computed the standard deviation; it was always on the order of
0.01. The results lead to the following conclusions:
1. Even Algorithm A obtains very good results, but nonetheless Algorithms B and C improve no-
ticeably upon Algorithm A. In particular, Algorithm C (for d = 15) obtains the highest satisfaction
on all data sets and in almost all cases was able to find an optimal solution.
2. Algorithm R gives slightly worse solutions than Algorithm C.
3. The quality of the algorithms does not depend on the data set used for verification (the only
exception is Algorithm R for Monroe's system on data set SK; however SK has only 9 voters so it
can be viewed as a border case).
14
5.2 Evaluation on Larger Instances
For experiments on larger instances we needed data sets with at least n = 10000 agents. Thus we
used the NetFlix data set and synthetic data. (Additionally, we run the subset of experiments (for
n ≤ 5000) also for the S2 data set.) For Monroe's rule we present results for Algorithm A, Algo-
rithm C, and Algorithm R, and for Chamberlin-Courant's rule we present results for Algorithm C
and Algorithm R. We limit the set of algorithms for the sake of the clarity of the presentation. For
Monroe we chose Algorithm A because it is the simplest and the fastest one, Algorithm C because
it is the best generalization of Algorithm A that we were able to run in reasonable time, and Algo-
rithm R to compare a randomized algorithm to deterministic ones. For Chamberlin-Courant's rule
we chose Algorithm C because it is, intuitively, the best one, and we chose Algorithm R for the
same reason as in the case of Monroe. Further, we present results for the NetFlix data set and for
the urn model only. We chose these data sets because the urn model results turned out to be the
worst ones among the synthetic data sets, and the NetFlix data set is our only large real-life data set.
First, for each data set and for each algorithm we fixed the value of m and K and for each
n ranging from 1000 to 10000 with the step of 1000 we run 50 experiments. We repeated this
procedure for 4 different combinations of m and K: (m = 10, K = 3), (m = 10, K = 6),
(m = 100, K = 30) and (m = 100, K = 60). We measured the statistical correlation between
the number of voters and the quality of the algorithms C/Cideal. The ANOVA test in most cases
showed that there is no such correlation. The only exception was S2 data set, for which we obtained
an almost negligible correlation. For example, for (m = 10, K = 3) Algorithm C under data set
S2 for Monroe's system for n = 5000 gave C/Cideal = 0.88, while for n = 100 (in the previous
section) we got C/Cideal = 0.89. Thus we conclude that in practice the number of agents has
almost no influence on the quality of the results provided by our algorithms.
Next, we fixed the number of voters n = 1000 and the ratio K/m = 0.3, and for each m
ranging from 30 to 300 with the step of 30 (naturally, as m changed, so did K to maintain the ratio
K/m), we run 50 experiments. We repeated this procedure for K/m = 0.6. The relation between
m and C/Cideal for MV and UR, under both Monroe's rule and Chamberlin-Courant's rule, is given
in Figures 4 and 5 (the results for K/m = 0.6 look similar).
Finally, we fixed n = 1000 and m = 100, and for each K/m ranging from 0.1 and 0.5 with the
step of 0.1 we run 50 experiments. The relation between the ratio K/m and the quality C/Cideal is
presented in Figures 6 and 7.
For the case of Chamberlin-Courant's system increasing the size of the committee we elect
improves agent satisfaction: Since there are no constraints on the number of agents matched to a
given alternative, larger committees mean more opportunities to satisfy the agents. For Monroe,
larger committees may lead to lower total satisfaction. This happens if many agents like a particular
alternative a lot, but only some of them can be matched to this alternative and others have to be
matched to their less-preferred ones. Nonetheless, we see that Algorithm C achieves C/Cideal =
0.925 even for K/m = 0.5 for the NetFlix data set.
Our conclusions from these experiments are the following. For Monroe's rule, even Algorithm A
achieves very good results. However, Algorithm C consistently achieves better ones (indeed, almost
perfect ones). Randomized algorithms consistently do worse than our deterministic ones.
15
Figure 4: The relation between the number of alternatives m and the quality of the algorithms
C/Cideal for the Monroe's system.
Figure 5: The relation between the number of alternatives m and the quality of the algorithms
C/Cideal for the Chamberlin-Courant's system.
5.3 Running time
In our final set of experiments we have measured running times of our algorithms on the data set
MV. We have used a machine with Intel Pentium Dual T2310 1.46GHz processor and 1.5GB of
16
Figure 6: The relation between the ratio K/m and the quality of the algorithms C/Cideal for the
Monroe's system.
Figure 7: The relation between the ratio K/m and the quality of the algorithms C/Cideal for the
Chamberlin-Courant's system.
RAM. In Figure 8 we show the running time of GLPK ILP solver for Monroe's and for Chamberlin-
Courant's rules. These running times are already large for small instances and they are increasing
exponentially with the number of voters. For Monroe's rule, even for K = 9, m = 30, n = 100
some of the experiments timed out after 1 hour, and for K = 9, m = 30, n = 200 none of the
17
Figure 8: The running time of the standard ILP solver for the Monroe's and for the Chamberlin-
Courant's systems. For the Monroe's system, for K = 9, m = 30, and for n ≥ 200 none of the
single algorithm execution finished within 1 day.
experiments finished within one day. Thus we conclude that the real application of the ILP algorithm
is very limited. Example running times of the other algorithms for some combinations of n, m, and
K are presented in Table 6.
6 Conclusions
We have provided experimental evaluation of a number of algorithms (both known ones and their
extensions) for computing the winners under Monroe's rule and under Chamberlin-Courant's rule.
While finding winners under these rules is NP-hard [2,5,14], it turned out that in practice we can
obtain very high quality solutions using simple algorithms. Indeed, both for Monroe's rule and for
Chamberlin-Courant's rule we recommend using Algorithm C (or Algorithm A on very large Mon-
roe elections). We believe that our results mean that (approximations of) Monroe's and Chamberlin-
Courant's rules can be used in practice.
Acknowledgements The authors were supported in part by AGH Univ. grant 11.11.120.865, by the
Foundation for Polish Science's Homing/Powroty program, by Poland's National Science Center
grant DEC-2011/03/B/ST6/01393, and by EU's Human Capital Program "National PhD Programme
in Mathematical Sciences" carried out at the University of Warsaw.
18
m = 10, K = 3
m = 10, K = 6
n = 2000
A
0.01
B
0.08
C
1.1
GM 0.8
R
7.6
C
0.02
GM 0.003
P
0.009
R
0.014
6000
0.03
0.9
8
7.3
50
0.07
0.009
0.032
0.04
10000
0.05
2.3
22
20
180
0.12
0.015
0.05
0.065
2000
0.01
0.2
2.1
1.9
6.5
0.05
0.003
0.008
0.02
6000
0.04
1.4
16
13
52
0.14
0.01
0.02
0.06
10000
0.07
3.6
37
52
140
0.26
0.018
0.05
0.11
m = 100, K = 30
m = 100, K = 60
n = 2000
A
0.5
B
0.8
C
38
GM 343
R
41
C
4.3
GM 0.06
P
0.03
R
0.06
6000
1.6
4
140
2172
329
11
0.2
0.1
0.24
10000
2.8
9.5
299
5313
830
19
0.4
0.26
0.45
2000
0.9
1.7
64
929
88
7.5
0.09
0.03
0.1
6000
2.8
8
221
5107
608
19
0.3
0.1
0.4
10000
4.9
18
419
13420
1661
31
0.7
0.2
0.8
e
o
r
n
o
M
C
C
e
o
r
n
o
M
C
C
Table 6: Example running times of the algorithms [in seconds].
References
[1] S. Berg. Paradox of voting under an urn model: The effect of homogeneity. Public Choice,
47:377 -- 387, 1985.
[2] N. Betzler, A. Slinko, and J. Uhlmann. On the computation of fully proportional representa-
tion. Technical report, U. of Auckland, November 2011.
[3] B. Chamberlin and P. Courant. Representative deliberations and representative decisions: Pro-
portional representation and the Borda rule. American Political Science Review, 77(3):718 --
733, 1983.
[4] T. Kamishima. Nantonac collaborative filtering: recommendation based on order responses.
In Proceedings of KDD-03, pages 583 -- 588, 2003.
[5] T. Lu and C. Boutilier. Budgeted social choice: From consensus to personalized decision
making. In Proceedings of IJCAI-2011, pages 280 -- 286, 2011.
19
[6] T. Lu and C. Boutilier. Learning Mallows models with pairwise preferences. In Proceedings
of ICML-11, pages 145 -- 152, June 2011.
[7] C. L. Mallows. Non-null ranking models. i. Biometrika, 44(1-2):114 -- 130, June 1957.
[8] N. Mattei, J. Forshee, and J. Goldsmith. An empirical study of voting rules and manipulation
with large datasets. In COMSOC, 2012.
[9] R. Meir, A. Procaccia, J. Rosenschein, and A. Zohar. The complexity of strategic behavior in
multi-winner elections. JAIR, 33:149 -- 178, 2008.
[10] B. Monroe. Fully proportional representation. American Political Science Review, 89(4):925 --
940, 1995.
[11] G. Nemhauser, L. Wolsey, and M. Fisher. An analysis of approximations for maximizing
submodular set functions. Mathematical Programming, 14(1):265 -- 294, 1978.
[12] J. Oren. Personal communication, 2012.
[13] R. Potthoff and S. Brams. Proportional representation: Broadening the options. Journal of
Theoretical Politics, 10(2):147 -- 178, 1998.
[14] A. Procaccia, J. Rosenschein, and A. Zohar. On the complexity of achieving proportional
representation. Social Choice and Welfare, 30(3):353 -- 362, 2008.
[15] P. Skowron, P. Faliszewski, and A. Slinko. Fully proportional representation as resource allo-
cation: Approximability results. Technical Report arXiv:0809.4484 [cs.GT], arXiv.org, Aug.
2012.
[16] T. Walsh. Where are the hard manipulation problems? JAIR, 42:1 -- 29, 2011.
20
|
1712.01648 | 1 | 1712 | 2017-12-05T14:23:35 | The Role of Compliance in Heterogeneous Interacting Agents: Data from Observations | [
"cs.MA"
] | The dynamics of agent-based systems provide a framework to face the complexity of pedestrian-vehicle interactions in future cities, in which the compliance to traffic norms plays a fundamental role. The data of an observation performed at a non-signalized intersection are presented to provide useful insights for supporting the future development of agent-based models. Results focus on drivers' compliance to crossing pedestrians, describing potentially conflictual interactions among heterogeneous agents. The discussion closes with the potential applications of the collected data set for modelling the phenomenon. | cs.MA | cs | The Role of Compliance in Heterogeneous
Interacting Agents: Data from Observations
Stefania Bandini1, 2, Luca Crociani1, Giuseppe Vizzari1,
Flavio Soares Correa da Silva3, and Andrea Gorrini1
1 Complex Systems and Artificial Intelligence research center, Univ. Milano - Bicocca (Italy)
2 Research Center for Advance Science and Technology, The University of Tokyo (Japan)
3 Institute of Mathematics and Statistics, University of Sao Paulo (Brazil)
Abstract. The dynamics of agent-based systems provide a framework to face
the complexity of pedestrian/vehicle interactions in future cities, in which the
compliance to traffic norms plays a fundamental role. The data of an observa-
tion performed at a non-signalized intersection are presented to provide useful
insights for supporting the future development of agent-based models. Results
focus on drivers' compliance to crossing pedestrians, describing potentially con-
flictual interactions among heterogeneous agents. The discussion closes with the
potential applications of the collected data set for modelling the phenomenon.
Keywords: Agent-based Modelling, Compliance, Pedestrian, Vehicular Traffic
1
Introduction and Related Work
Agent-based modelling and simulations of pedestrian and crowd dynamics have been
increasingly reported in the technical and scientific specialized literature. Scientific
communities started to incorporate agent-based systems to improve the expressiveness
of traditional approaches and to simulate the complex behaviour of people and traf-
fic in outdoor and indoor urban scenarios. The intrinsically dynamical properties of
agent-based models offer a research framework to face the complexity of the future
cities [7], offering new possibilities to incorporate and integrate the growing presence
of autonomous entities/artefacts both physical (e.g. autonomous vehicles) and virtual
(e.g. data coming from heterogeneous sources: social media, distributed sensors etc.).
The development of agent-based models and systems requires to check the quality,
robustness and plausibility of the obtained simulations against real data, in order to
tackle decision making problems related to urban mobility. Recent literature contains a
wide range of methods and study cases supporting this view [1]. The aim of this paper
is to present a real case of data collection performed to collect useful insights about
pedestrian-vehicles interactions at non-signalized intersections, supporting the future
development of a heterogeneous agent-based system to simulate the phenomenon.
From pioneering works, several models have been developed and applied for the
simulation of pedestrian and vehicular dynamics, including both CA and particles mod-
els [3]. These two approaches have, separately and independently, produced a signifi-
cant impact, yet efforts characterized by an integrated model considering the simulta-
neous presence of vehicles and pedestrians are not as frequent or advanced. With the
7
1
0
2
c
e
D
5
]
A
M
.
s
c
[
1
v
8
4
6
1
0
.
2
1
7
1
:
v
i
X
r
a
2
Bandini, S., Crociani, L., Vizzari, G., Correa da Silva, F., Gorrini, A.
Fig. 1. The work flow for selecting the sample of crossing episodes from the video images.
notable exception of [6], most efforts in this direction are relatively simplistic, narrow
(i.e. targeting extremely specific situations), homogeneous for the simulated entities,
and they are often not validated against real data [4, 10]. In this framework, we con-
sider the possibility to model and simulate the complex aspect of drivers' compliance
to pedestrian yielding rules in the context of interactions among heterogeneous agents,
and its implications on self-organization dynamics.
2 Observation Results
A video recorded observation [5] has been performed at a non-signalized intersection
in Milan (from noon, 73 minutes), characterized by a high number of pedestrian acci-
dents in the past years. We selected from the video a sample of 812 crossing episodes
considering only the cases in which one vehicle directly interacted with one or more
pedestrians (see Fig. 1). At non-signalized intersections traffic laws require drivers to
yield those pedestrians who are already occupying the zebra, but also those who are
localized nearby or in correspondence of the curb waiting to cross. The level of compli-
ance of drivers to pedestrians' right-of-way have been estimated considering the posi-
tion of crossing pedestrians with respect to the curb (i.e. pedestrian about 1.5 meter far
from the curb, waiting at the curb, crossing on the zebra) and to the direction of travel
of vehicles (i.e. pedestrian from the near or the far side-walk).
Preliminary analyses on results (see Tab. 1) showed that 48% of the total number
of crossing episodes was characterized by non-compliant drivers with crossing pedes-
trians from the two side-walks. A multiple linear regression was calculated to predict
the percentage of non-compliant drivers per minute based on: (i) number of vehicles
per minute (18.89 veh/min in average; p = 0.007, significant predictor) and (ii) num-
ber of crossing pedestrian per minute (8.01 ped/min in average; p < 0.001, significant
predictor). A significant regression equation was found [F(2,70) = 14.526, p < 0.001],
with R2 of 0.293. This demonstrates that the non-compliance of drivers is negatively
determined by traffic conditions and positively determined by pedestrian flows on ze-
bra. Despite the low level of drivers' compliance, no accidents or risky situations have
been observed, thanks to the self-organization of the system based on pedestrians' yield-
ing/collaborative behaviour to approaching cars.
The Role of Compliance in Heterogeneous Interacting Agents
3
Table 1. Results about the drivers' compliance to the right-of-way of crossing pedestrians.
Compliant Non-compliant
Types of pedestrian/vehicle interaction
Ped. approaching/waiting/crossing from the near side-walk 191 (46.14%) 223 (53.86%)
Ped. approaching/waiting/crossing from the far side-walk 230 (57.69%) 168 (42.21%)
Fig. 2. The relation between non-compliant drivers and the number of vehicles (left) and crossing
pedestrians (right) in the observed intersection. Each data refers to one minute of the video.
3 Discussion
The results showed in the previous section represent applicable insights towards the
extension of a model for the analysis of pedestrian crossings [2], considering poten-
tially conflictual interactions among heterogeneous agents. As shown in Figure 3(a),
the model is based on the integration of two independent models for the simulation of
vehicles, moving in continuous lanes, and pedestrians, moving in a 2-dimensional dis-
crete environment. The two environments are superimposed, and car-agents perceive
pedestrian-agents while they are crossing or in the nearby of the curb (grey cells in
Fig. 3(a)) and vice-versa. The interactions between them are described in Figure 3(b).
Pedestrian-agents consider the speed and distance of cars to avoid collisions, giving
way to non-compliant vehicles. The compliance of car-agents is mainly influenced by
the necessary braking distance. On the other hand, according to a fixed probability that
will be set on the collected data on compliance, car-agents can deliberately avoid to
stop even if the braking distance is sufficient, requiring pedestrian-agents to yield.
The potential applications of such research are related to the possibility to test the
effect of non-compliance on emergent observables like: near accident situations, ex-
posure to accidents, traffic capacity of the road and Level of Service [8]. In addition,
this kind of research is also potentially relevant to complement studies on autonomous
vehicles [9] to evaluate future transportation scenarios in Smart Cities.
References
[1] Campanella, M., Hoogendoorn, S., Daamen, W.: Quantitative and qualitative validation
procedure for general use of pedestrian models. In: Pedestrian and Evacuation Dynamics
4
Bandini, S., Crociani, L., Vizzari, G., Correa da Silva, F., Gorrini, A.
(a)
(b)
Fig. 3. A description of the environment and the two classes of agents (a). Car-agents life-cycle
considering their possible non-compliance with respect to the presence of pedestrian-agents (b).
2012, 891 -- 905. Springer (2014)
[2] Crociani, L., Vizzari, G.: An integrated model for the simulation of pedestrian crossings.
In: Cellular Automata for Research and Industry, 670 -- 679. Springer (2014)
[3] Duives, D.C., Daamen, W., Hoogendoorn, S.P.: State-of-the-art crowd motion simulation
models. Transportation research part C: emerging technologies 37, 193 -- 209 (2013)
[4] Godara, A., Lassarre, S., Banos, A.: Simulating Pedestrian-Vehicle Interaction in an Urban
Network Using Cellular Automata and Multi-Agent Models. Traffic and Granular Flow'05,
411 -- 418 (2007)
[5] Gorrini, A., Vizzari, G., Bandini, S.: Towards modelling pedestrian-vehicle interactions:
Empirical study on urban unsignalized intersection. In: Pedestrian and Evacuation Dynam-
ics, 25 -- 33 (2016)
[6] Helbing, D., Jiang, R., Treiber, M.: Analytical investigation of oscillations in intersecting
flows of pedestrian and vehicle traffic. Physical Review E - Statistical, Nonlinear, and Soft
Matter Physics 72(4) (2005)
[7] Masthoff, J., Vasconcelos, W.W., Aitken, C., Correa da Silva, F.: Agent-based group mod-
elling for ambient intelligence. In: Symposium on Affective Smart Environments (2007)
[8] Milazzo II, J.S., Rouphail, N.M., Hummer, J.E., Allen, D.P.: Highway capacity manual
2000. Journal of the Transportation Research Board 1678(1), 25 -- 31 (2000)
[9] Tachet, R., Santi, P., Sobolevsky, S., Reyes-Castro, L.I., Frazzoli, E., Helbing, D., Ratti, C.:
Revisiting street intersections using slot-based systems. PLoS ONE 11(3), 1 -- 9 (2016)
[10] Zeng, W., Nakamura, H., Chen, P.: A Modified Social Force Model for Pedestrian Behavior
Simulation at Signalized Crosswalks. Social and Behavioral Sciences 138, 521 -- 530 (2014)
|
1909.08345 | 1 | 1909 | 2019-09-18T10:43:09 | Limited-budget output consensus for descriptor multiagent systems with energy constraints | [
"cs.MA",
"eess.SY",
"eess.SY"
] | The current paper deals with limited-budget output consensus for descriptor multiagent systems with two types of switching communication topologies; that is, switching connected ones and jointly connected ones. Firstly, a singular dynamic output feedback control protocol with switching communication topologies is proposed on the basis of the observable decomposition, where an energy constraint is involved and protocol states of neighboring agents are utilized to derive a new two-step design approach of gain matrices. Then, limited-budget output consensus problems are transformed into asymptotic stability ones and a valid candidate of the output consensus function is determined. Furthermore, sufficient conditions for limited-budget output consensus design for two types of switching communication topologies are proposed, respectively. Finally, two numerical simulations are shown to demonstrate theoretical conclusions. | cs.MA | cs |
IEEE TRANSACTIONS ON CYBERNETICS
1
Limited-budget output consensus for descriptor
multiagent systems with energy constraints
Jianxiang Xi, Cheng Wang, Xiaojun Yang, Bailong Yang
Abstract -- The current paper deals with limited-budget output
consensus for descriptor multiagent systems with two types of
switching communication topologies; that is, switching connected
ones and jointly connected ones. Firstly, a singular dynamic
output feedback control protocol with switching communication
topologies is proposed on the basis of the observable decom-
position, where an energy constraint is involved and protocol
states of neighboring agents are utilized to derive a new two-
step design approach of gain matrices. Then,
limited-budget
output consensus problems are transformed into asymptotic
stability ones and a valid candidate of the output consensus
function is determined. Furthermore, sufficient conditions for
limited-budget output consensus design for two types of switching
communication topologies are proposed, respectively. Finally,
two numerical simulations are shown to demonstrate theoretical
conclusions.
Index Terms -- Multiagent system, descriptor system, output
consensus, limited budget, switching topology.
I. INTRODUCTION
D URING the last decade, consensus of multiagent sys-
tems receives considerable attention, which designs a
distributed control protocol to drive multiple agents to achieve
an agreement about some interested variables such as time,
position, velocity and temperature, et al., as shown in [1]-
[6]. Consensus has potential practical applications in formation
control [7]-[10], target tracking [11]-[12], network synchroni-
sation [13]-[14] and multiple source data analysis [15]-[16], et
al.
The communication topologies are critically important for
multiagent systems to achieve consensus, which can usually
be categorized into the fixed ones and switching ones. For
fixed communication topologies, the neighboring relationship
and communication weights are time-invariant, as discussed
in [17] and [18]. For switching communication topologies, the
neighboring relationships may be time-varying, but communi-
cation weights are time-invariant. In this case, the associated
Laplacian matrices of the communication topologies are piece-
wise continuous, as shown in [19]-[21], where switching com-
munication topologies are divided into switching connected
ones and jointly connected ones. For switching connected
communication topologies, each topology in the switching set
is connected. For jointly connected communication topologies,
the union of topologies in certain time interval is connected,
but each topology in the switching set can be unconnected.
It is well-known that consensus for jointly connected com-
munication topology cases is more complex than the one for
switching connected communication topology cases.
Jianxiang Xi, Cheng Wang, Bailong Yang are with Rocket Force University
of Engineering, Xi'an, 710025, P.R. China, and Xiaojun Yang is with
School of Information Engineering, Guangdong University of Technology,
Guangzhou, 510006, P.R. China. (Corresponding author: Xiaojun Yang,
[email protected])
According to the dynamics of each agent, multiagent sys-
tems can be categorized into normal ones and descriptor
ones. For normal multiagent systems, each agent is modeled
as differential equations and only owns dynamic modes, as
addressed in [22]-[24]. For descriptor multiagent systems, the
dynamics of each agent may contain algebraic constrains and
cannot be modeled by differential equations. In practical appli-
cations, descriptor multiagent systems can be used in complex
networks and multiagent supporting systems, as discussed in
[25]. Moreover, the motion of each agent may be derived by
impulsive modes, static modes and dynamic modes, where
impulsive modes should be eliminated since they destroy the
operation of the whole system, and the motions associated with
both static modes and dynamic modes should achieve con-
sensus. Yang and Liu [26] proposed necessary and sufficient
consensus conditions of descriptor multiagent systems with
fixed communication topologies by a static output feedback
control protocol, where a specific rank constrain is required.
In [27], the impacts of communication delays on consensus
of descriptor multiagent systems were analyzed, where it was
also supposed that communication topologies are fixed. Wang
and Huang [28] dealt with consensus analysis and design
problems for descriptor multiagent systems with switching
communication topologies.
Based on different interested variables, consensus can be
categorized into state consensus and output consensus, where
it is only required that partial states or state combinations
achieve agreement for output consensus. Chopra and Spong
[29] first
introduced the concept of output consensus and
proposed the associated consensus criteria for normal non-
linear multiagent systems with relative degree one. Xiao et
al. [30] modeled the dynamics of each agent in multiagent
systems as a normal second-order integrator and gave sufficient
conditions for output consensus. Liu and Jia [31] proposed
an H∞ analysis approach to investigate output consensus for
normal multiagent systems with external disturbances, where
each agent was modeled as a type of special controllability
canonical form. In [32], output consensus for normal single-
input-single-output high-order linear multiagent systems was
addressed by the output regulation theory. Xi et al. [33]
focused general normal high-order multiagent systems and
proposed output consensus analysis and design criteria by a
partial stability method. It should be pointed that the dynamics
of each agent is normal in [29]-[33], where it was shown that
output consensus is more challenging than state consensus.
In the above literatures, the energy limitation and/or the
consensus performance were not considered, which can be
modeled as certain optimization consensus problems. Cao
and Ren [34] gave optimal consensus criteria by the linear
quadratic regulator, where it was required that the communi-
IEEE TRANSACTIONS ON CYBERNETICS
2
cation topology is complete, which means that the nonzero
eigenvalues of the corresponding Laplacian matrix are iden-
tical. Actually,
this requirement of optimal consensus on
the communication topology is conservatism, and suboptimal
consensus was extensively studied. Guaranteed-performance
consensus for normal multiagent systems was dealt with in
[35] and [36], where it was supposed that communication
topologies are fixed. Linear matrix inequality criteria for
guaranteed-cost consensus with communication delays were
presented in [37] and [38]. In [39] and [40], guaranteed-
cost consensus for normal multiagent systems with switching
topologies was discussed and the associated guaranteed-cost
consensus analysis and design criteria were proposed. In [35]-
[40], state consensus instead of output consensus for normal
multiagent systems was investigated and the guaranteed cost
cannot be given previously. To the best of our knowledge,
limited-budget output consensus for descriptor multiagent sys-
tems with switching communication topologies is still open
and the following three challenging problems should be dealt
with: (i) How to eliminate impulsive modes and guarantee
that the motions associated with static modes achieve output
consensus; (ii) How to introduce the given budget into consen-
sus criteria and construct the relationship between the given
budget and initial outputs instead of initial states; (iii) How to
determine the impacts of switching topologies and ensure the
checkable feature of the associated consensus criteria.
The current paper focuses on limited-budget output con-
sensus for descriptor multiagent systems and addresses the
impacts of two types of switching communication topologies.
Based on the observable decomposition of the dynamics of
each agent, a singular dynamic output feedback consensus
protocol with an energy constraint and topology switching
is presented, which can make the dynamics of the whole
multiagent system satisfy some separation feature; that is, its
gain matrices lies on diagonal boxes of an upper triangular
matrix. Furthermore, using the characteristics that the row sum
of the Laplaican matrix of the communication topology is zero,
the disagreement and consensus dynamics of a descriptor mul-
tiagent system are determined and they are independent with
each other. For switching connected communication topology
cases, a new two-step design approach is proposed to design
gain matrices of consensus protocols, the relationship between
the given energy budget and the matrix variable is constructed,
and limited-budget output consensus design criteria are given,
respectively. For jointly connected communication topology
cases, by the Cauchy convergence criterion and Barbalat's
lemma, sufficient conditions for limited-budget output consen-
sus design are presented, respectively.
II. PROBLEM DESCRIPTION
The dynamics of the mth agent of a high-order homoge-
neous descriptor multiagent system is modeled by
(cid:26) E xm(t) = Axm(t) + Bum(t),
ym(t) = Cxm(t),
(1)
where m = 1, 2,··· , N, E ∈ Rn×n, A ∈ Rn×n, B ∈
Rn×k, C ∈ Rl×n and xm(t), um(t) and ym(t) denote
the system state, the control input and the system output,
respectively. Because the matrix E may be singular; that is,
rank(E) ≤ n , multiagent system (1) is called the descriptor
multiagent system. Compared with normal multiagent systems
modeled by differential equations, multiagent system (1) may
contain static modes and impulsive modes besides dynamic
modes. Impulsive modes must be eliminated by designing
proper consensus protocols since they destroy the operation
of the whole system. For multiagent system (1), all the states
associated with both dynamic modes and static modes should
achieve consensus.
This autonomous system of multiagent system (1) can
also be represented by the pair (E, A), whose some specific
properties compared with normal systems are listed as follows.
Definition 1: [41] The pair (E, A) is regular if σE−A (cid:54)= 0
for some σ ∈ C and impulse-free if deg(σE− A) = rank(E)
for ∀σ ∈ C.
Lemma 1: [41] The pair (E, A) is regular and impulse-free
if and only if rank
= n + rank(E).
(cid:20)E 0
(cid:21)
A E
Lemma 2: [41] The pair (E, A) is regular, impulse-free and
asymptotically stable if and only if there exists a matrix R
such that ET R = RT E ≥ 0 and AT R + RT A < 0.
We first introduce the observable decomposition of the tripe
(E, A, C). Let Uo be an invertible matrix such that
(cid:21)
(cid:21)
(cid:20) Eo
(cid:20) Ao 0
0
Eo Eo
Ao Ao
U−1
o EUo =
U−1
o AUo =
(cid:20) Bo
Bo
(cid:21)
,
, U−1
o B =
, CUo = [Co, 0] ,
where Eo ∈ Rh×h, Ao ∈ Rh×h, Bo ∈ Rh×k, Co ∈
Rl×h and the triple (Eo, Ao, Co) is observable. Then, the
following consensus protocol with switching topologies and
an energy constraint is proposed:
(cid:80)
Eo zm(t) = (Ao + BoKu) zm(t)
− KzCo
(cid:80)
j∈Nm,κ(t)
wmj,κ(t) (zj(t) − zm(t))
wmj,κ(t) (yj(t) − ym(t)) ,
(2)
+ Kz
j∈Nm,κ(t)
um(t) = Kuzm(t),
N(cid:80)
(cid:82) +∞
0
m=1
m(t)M um(t)dt,
uT
Je =
where m = 1, 2,··· , N, Ku ∈ Rk×h, Kz ∈ Rh×l, M T =
M > 0, Nm,κ(t) is the neighbor set of the mth agent at time
t, zm(t) with zm(0) = 0 stands for the protocol state, and
Je represents the energy consumption of the whole multiagent
system.
Let J∗
e > 0 be a given energy budget; that is, the whole
energy consumption of multiagent system (1) with consensus
protocol (2) must be less than J∗
e . In the following, we give the
definition of the limited-budget output consensus of descriptor
multiagent systems.
Definition 2: For any given J∗
e > 0, multiagent system (1) is
said to be limited-budget output consensualizable by consensus
protocol (2) if it is regular and impulse-free and there exist
Ku and Kz such that limt→+∞ (ym(t) − co(t)) = 0 (m =
1, 2,··· , N ) and Je ≤ J∗
for any bounded disagreement
e
IEEE TRANSACTIONS ON CYBERNETICS
initial outputs, where co(t) is said to be the output consensus
function.
The current paper gives two design approaches of gain
matrices Ku and Kz such that multiagent system (1) with con-
sensus protocol (2) achieves limited-budget output consensus
for switching connected communication topologies and jointly
connected communication topologies, respectively.
III. PROBLEM TRANSFORMATION
By the observable decomposition and the separation prin-
ciple, consensus problems for multiagent system (1) with
consensus protocol (2) are transformed into asymptotic sta-
bility ones for a reduced-order subsystem, and the consensus
dynamics is also determined.
Let U−1
om(t), xT
tem (1) can be transformed into
o xm(t) =(cid:2)xT
(cid:21)(cid:20) xom(t)
(cid:21)
(cid:20) Eo
0
Eo Eo
xom(t)
=
om(t)(cid:3)T , then multiagent sys-
(cid:20) Ao 0
(cid:20) Bo(t)
(cid:21)(cid:20) xom(t)
(cid:21)
Ao Ao
xom(t)
(cid:21)
(3)
+
Bo(t)
um(t),
IN ⊗ (Ao + BoKu) + Lκ(t) ⊗ KzCo
where xo(t) = (cid:2)xT
(cid:2)zT
2 (t),··· , zT
1 (t), zT
ym(t) = Coxom(t),
where m = 1, 2,··· , N. It can be found that the system output
ym(t)(m ∈ 1, 2,··· , N ) depends on the observable compo-
nent xom(t), but it is not associated with the unobservable
component xom(t). Hence, when the limited budget is not
considered, multiagent system (1) achieves output consensus
if and only if the observable component of each agent achieves
consensus. In this case, the unobservable component of each
agent can be neglected and the observable component of
multiagent system (1) with consensus protocol (2) can be
written in the following Kronecker form as
IN ⊗
(cid:21)(cid:20) xo(t)
(cid:20) Eo
IN ⊗ Ao
(cid:21)
(cid:20)
Eo
−Lκ(t) ⊗ KzCo
(cid:21)(cid:20) xo(t)
=
zo(t)
IN ⊗ BoKu
(cid:21)
N (t)(cid:3)T and Lκ(t) is the Laplacian matrix
,
(4)
zo(t) =
oN (t)(cid:3)T
o2(t),··· , xT
o1(t), xT
zo(t)
,
of the switching communication topology, which is piecewise
continuous since the switching signal κ(t) is piecewise con-
tinuous.
√
(cid:104)
Since each communication topology in the switching set
is undirected, the associated Laplacian matrix Lκ(t) is sym-
metric, where zero is its eigenvalue with an eigenvector
there exists an orthonormal matrix
1/
Uκ =
(cid:105)
κ Lκ(t)Uκ = diag(cid:8)0, ∆κ(t)
N. In this case,
√
1/
such that
(cid:9) ,
N , Uκ
U T
κ Lκ(t)
Uκ ∈ R(N−1)×(N−1) and Uκ is
where ∆κ(t) = U T
set to be time-varying for switching connected communica-
tion topology cases and time-invariant for jointly connected
communication topology cases. We introduce the following
nonsingular transformation
xo(t) =(cid:0)U T
κ ⊗ Ih
(cid:1) xo(t) =(cid:2)xT
oN (t)(cid:3)T
,
o1(t), xT
o2(t),··· , xT
,
(5)
3
.
(cid:21)
(6)
and
,
=
=
Eo
Eo
Ao
×
(cid:21)
(cid:21)
BoKu
xo1(t)
xo∆(t)
xo1(t)
xo∆(t)
o1(t), zT
κ ⊗ Ih
IN−1⊗
zo(t) =(cid:0)U T
(cid:20) Eo
oN (t)(cid:3)T
(cid:21)
(cid:21)(cid:20) zo1(t)
(cid:1) zo(t) =(cid:2)zT
(cid:21)
(cid:21)(cid:20) zo∆(t)
IN−1⊗Ao
IN−1⊗BoKu
(cid:20) Ao + BoKu 0
o2(t),··· , zT
Thus, multiagent system (4) can be transformed into
(cid:21)(cid:20) zo1(t)
(cid:20) Eo
(cid:20) IN−1⊗(Ao + BoKu)+∆κ(t)⊗KzCo −∆κ(t) ⊗ KzCo
(cid:20) zo∆(t)
where xo∆(t) = (cid:2) xT
zo∆(t) =(cid:2) zT
(cid:20) I(N−1)h
(cid:21)
(cid:20) zo∆(t)
(cid:21)(cid:20) zo∆(t)
(cid:20) Eo
(cid:20)IN−1⊗(Ao + BoKu)
(cid:21)
(cid:20) zo∆(t)
xo∆(t)
then subsystem (6) can be transformed into
IN−1⊗
o3(t), ··· , xT
(cid:21)
oN (t) ]T . Let
(cid:21)
IN−1⊗Ao+∆κ(t)⊗KzCo
(cid:20) zo∆(t)
I(N−1)h −I(N−1)h
o2(t), xT
o3(t), ··· , zT
∆κ(t) ⊗ KzCo
o2(t), zT
oN (t) ]T
xo∆(t)
xo∆(t)
Us =
= Us
(cid:21)
(cid:21)
Eo
,
,
=
0
0
(7)
×
xo∆(t)
.
Let the minimum and maximum nonzero eigenvalues of
the Laplacian matrices of all the communication topologies
in the switching set denote as λmin = min{λm,min,∀m ∈ χ}
and λmax = max{λm,min,∀m ∈ χ}, where λm,min and λm,max
represent the minimum and maximum nonzero eigenvalues of
the Laplacian matrix of the mth communication topology in
the switching set, respectively. These two notations are used
for both switching connected communication topology cases
and jointly connected communication topology cases.
Moreover, subsystems (5) and (6) determine consensus
and disagreement dynamics of multiagent system (1). The
following theorem converts the consensus problem into the
asymptotic stability one and determines a candidate of the
output consensus function.
tocol (2) is regular and impulse-free and limt→+∞(cid:2)zT
o∆(t)(cid:3)T = 0,
Theorem 1: If multiagent system (1) with consensus pro-
o∆(t),
then it achieves output consensus and
(cid:17)
√
(cid:16)
co(t) − Co xo1(t)/
N
= 0.
Proof: Let em(m ∈ 1, 2,··· , N ) denote the N-dimensional
column vector with the mth component 1 and 0 elsewhere.
Due to
xT
limt→∞
Uκe1 =
one can show that
1√
N
1,
Uκe1 ⊗ xo1(t) =
1 ⊗ xo1(t).
1√
N
(8)
IEEE TRANSACTIONS ON CYBERNETICS
4
o∆(t))T(cid:3)T with xm
where ET
··· , xN
it can be deduced by (7) that
o Rx=RT
x
m=2
(xm
x Eo≥0. Let xo∆(t)=(cid:2)(x2
N(cid:88)
o∆(t))T(cid:0)RT
+(cid:0)Ao + λκ(t),mKzCo
o∆(t))T(cid:0)RT
o∆(t))T(cid:0)RT
x Ao + AT
minR−T
x Ao+AT
x C T
Vx(t) =
Let Kz = −λ−1
Vx(t)=
(xm
N(cid:88)
≤ N(cid:88)
m=2
m=2
(xm
o∆(t))T ,
o∆(t) ∈ Rh (m = 2, 3,··· , N ), then
o∆(t))T , x3
(cid:1)
(cid:0)Ao + λκ(t),mKzCo
(cid:1)T
o∆(t).
o /2, then one can obtain that
Rx ) xm
o Rx−λκ(t),mλ−1
minC T
o Co
(cid:1) xm
o∆(t)
(cid:1) xm
(cid:16)
(11)
o Co
o∆(t).
o Rx − C T
(cid:17)
(cid:104)(cid:0)z2
o∆(t)(cid:1)T
,(cid:0)z3
o∆(t)(cid:1)T
zo∆(t),
Rz
,
o∆(t) ∈ Rh (m = 2, 3,··· , N ),
Furthermore, construct the following Lyapunov function can-
didate
From the property of the Kronecker product, one can find that
Uκe1 ⊗ xo1(t) = (Uκ ⊗ Ih)
Uκem ⊗ xom(t) = (Uκ ⊗ Ih)
N(cid:88)
Because Us is invertible, limt→∞(cid:2)zT
limt→∞(cid:2)zT
o∆(t)(cid:3)T
o∆(t), xT
= 0. Due to
m=2
(cid:21)
(cid:20) xo1(t)
(cid:20)
0
0
xo∆(t)
o∆(t), xT
,
(cid:21)
o∆(t)(cid:3)T
.
(9)
(10)
= 0 if
xo(t) = (Uκ ⊗ Ih) xo(t),
the conclusion can be obtained from (8) to (10). The proof of
Theorem 1 is completed.(cid:4)
IV. SWITCHING CONNECTED COMMUNICATION TOPOLOGY
CASES
For multiagent system (1) with switching connected com-
munication topologies, this section gives sufficient conditions
for limited-budget output consensualization and consensus in
terms of matrix inequality techniques.
κ(t)∆κ(t)
Since each communication topology in the switching set
is connected, zero is its simple eigenvalue of the Laplacian
matrix Lκ(t) and all other eigenvalues are positive. With-
out loss of generality, we can set that Uκ = Uκ(t) such
that U T
with 0 < λκ(t),2 ≤ λκ(t),3 ≤ ··· ≤ λκ(t),N . In this
case, Uκ is piecewise continuous, λmin = min{λm,2,∀m ∈
χ} and λmax = max{λm,N ,∀m ∈ χ}. Let yo(0) =
1 (0), yT
(cid:9)
Uκ(t) = diag(cid:8)λκ(t),2, λκ(t),3,··· , λκ(t),N
N (0)(cid:3)T . The following theorem proposes
2 (0),··· , yT
(cid:2)yT
a limited-budget output consensualization criterion by matrix
inequality techniques.
Theorem 2: For any given J∗
e > 0, multiagent system (1)
with switching connected cummnication topologies is limited-
budget output consensualizable by consensus protocol (2) if
there exist Rx and Rz such that
= h + rank(Eo),
(II)
(I)rank
(cid:21)
Ao Eo
(cid:20) Eo 0
ET
o (0)(cid:0)(cid:0)IN−N−111T(cid:1)⊗Ih
Θ11 Θm
x Eo ≥ 0,
o Rx − C T
o = EoRz ≥ 0,
o Rx = RT
x Ao + AT
RT
yT
z ET
RT
(III)
Θm =
12
Θ22
∗
where λm = λmin, λmax and
∗
∗
o Co < 0,
0.5BoM
0
−M
(cid:1) yo(0)ET
< 0,
o Rx≤J∗
e C T
o Co,
Θ11 = AoRz + RT
12 = −0.5λmR−T
Θm
x Ao + AT
Θ22 = RT
In this case, Ku=−BT
o R−1
o /2.
Proof: Consider a Lyapunov function candidate as follows
o Co.
z /2 and Kz=−λ−1
o − BoBT
z AT
o ,
x C T
o Co,
o Rx − C T
minR−T
x C T
o∆(t)(cid:0)IN−1 ⊗ ET
o Rx
(cid:1) xo∆(t),
Vx(t) = xT
where ET
o
··· ,(cid:0)zN
o
o∆(t)
Rz= RT
Vz(t) = zT
IN−1 ⊗ ET
o∆(t)(cid:1)T(cid:105)T
z Eo≥0. Let zo∆(t)=
N(cid:88)
o∆(t))T(cid:16) RT
with zm
Vz(t) =
(zm
then it can be deduced by (7) that
z (Ao + BoKu)
(cid:17)
m=2
+ (Ao + BoKu)T Rz
zm
o∆(t)
o∆(t))T RT
z KzCo xm
+ 2λκ(t),m (zm
z /2 and Rz = R−1
o R−1
o∆(t).
o∆(t)(cid:1)
Let Ku = −BT
that
z , then it can be derived
o
m=2
N(cid:88)
z AT
o∆(t).
z zm
z zm
z zm
Vz(t)=
KzCo xm
+ 2λκ(t),m
o−BoBT
(cid:0)R−1
(cid:1)(cid:0)R−1
o∆(t)(cid:1)T(cid:0)AoRz+RT
o∆(t)(cid:1)T
(cid:0)R−1
(cid:1) (m = 2, 3,··· , N ) and (Eo, Ao +
(cid:0)Eo, Ao + λκ(t),mKzCo
By Lemma 2, (cid:0)Eo, Ao + λκ(t),mKzCo
(cid:1) (m = 2, 3,··· , N )
(12)
Because Uκ(t)⊗Ih and Us are nonsingular, multiagent system
(4) is regular and impulse-free if and only if (Eo, Ao),
BoKu) are regular and impulse-free. From Lemma 1, the pair
(Eo, Ao) is regular and impulse-free if Condition (I) holds.
x Eo ≥ 0 and
are regular and impulse-free if ET
o Rx = RT
o Co < 0, and (Eo, Ao + BoKu) is
RT
o = EoRz ≥ 0 and
regular and impulse-free if RT
o < 0. Thus, by (11) and (12), according
AoRz+RT
to Theorem 1, multiagent system (1) with consensus protocol
(2) achieves output consensus if Conditions (I)-(III) hold.
o Rx − C T
o −BoBT
x Ao + AT
z ET
z AT
In the sequel, the limited-budget output consensus is ad-
dressed. It can be shown by (2) that
(cid:90) +∞
o (t)(cid:0)IN ⊗ K T
zT
Je =
(cid:16)
(13)
Due to zm(0) = 0 (m = 1, 2,··· , N ) and zo(t) =
o = EoRz ≥ 0
κ(t) ⊗ Ih
U T
zo(t), one has zo1 = 0. If RT
u M Ku
z ET
(cid:17)
0
(cid:1) zo(t)dt.
IEEE TRANSACTIONS ON CYBERNETICS
5
and AoRz + RT
Hence, it can be deduced by (13) that
z AT
o − BoBT
o < 0, then limt→∞ zo1(t) = 0.
z zom(t)(cid:1)T
(cid:0)R−1
z zom(t)(cid:1)dt.(14)
Je=
For any ≥ 0, it can be shown that
0
BoMBT
o
dt = Vx()−Vx(0)+Vz()−Vz(0).
(15)
Due to zo∆(t) = zo∆(t), from (14) and (15), it can be deduced
that
0
m=2
(cid:90) +∞
N(cid:88)
0.25(cid:0)R−1
(cid:90)
(cid:16) Vx(t)+ Vz(t)
(cid:17)
(cid:90)
o (t)(cid:0)IN ⊗ K T
(cid:90)
N(cid:88)
0.25(cid:0)R−1
z zom(t)(cid:1)T
(cid:90)
(cid:16) Vx(t)+ Vz(t)
(cid:17)
(cid:20) R−1
N(cid:88)
zT
(cid:44)
e
+
0
0
0
m=2
J
=
u M Ku
(cid:1) zo(t)dt
(cid:0)R−1
z zom(t)(cid:1) dt
BoM BT
o
dt+Vx(0)+Vz(0)−Vx()−Vz(). (16)
By (11) and (12), one can find that
Vx(t)+ Vz(t)=
o∆(t)
z zm
xm
o∆(t)
minR−T
x C T
o Rx − C T
o Co
o Co
m=2
−0.5λκ(t),mλ−1
RT
x Ao + AT
(cid:21)T(cid:20) AoRz+RT
(cid:21)(cid:20) R−1
o∆(t)
z zm
xm
o∆(t)
z AT
∗
o
o−BoBT
(cid:21)
.
(17)
From (16) and (17), based the convex property of the linear
matrix inequality,
then it can be
derived as → +∞ that
if Condition (III) holds,
Je ≤ Vx(0) + Vz(0).
(cid:16)
(18)
Since it is assumed that zm(0) = 0 (m = 1, 2,··· , N ), one
has zo(0) =
zo(0) = 0; which means that
zo∆(0) = 0. According to the structure of Us, it can be
obtained that
κ(0) ⊗ Ih
U T
(cid:17)
zo∆(0) = zo∆(0) = 0,
xo∆(0) = zo∆(0) − xo∆(0) = −xo∆(0).
o∆(0)(cid:0)IN−1 ⊗ ET
(cid:1)(cid:20)
0T
o Rx
(cid:1) xo∆(0).
(cid:21)(cid:0)IN−1 ⊗ ET
I(N−1)h
o Rx
(19)
(20)
(cid:1)
Thus, one can see that
From (18) to (20), it can be derived that
Vx(0) + Vz(0) = xT
Je ≤ xT
o (0)(cid:0)Uκ(0) ⊗ Ih
(cid:3)(cid:16)
(cid:2)0, I(N−1)h
κ(0) ⊗ Ih
U T
κ(0) = IN , one has
Due to Uκ(0)U T
(cid:17)
xo(0).
(21)
Uκ(0)
(cid:3)(cid:16)
Due to(cid:2)0, I(N−1)h
(cid:18)(cid:18)
by (21) and (22) that
(cid:17)
κ(0) = IN − 1
U T
N
κ(0) ⊗ Ih
= U T
U T
(cid:19)
Je ≤ xT
o (0)
IN − 1
N
11T
⊗ ET
o Rx
(cid:19)
(22)
11T .
κ(0) ⊗ Ih, one can find
xo(0).
(23)
0
11T
(cid:21)
o Co
o Co
= xT
xo(0)
(cid:19)
xT
o (0)
(cid:18)(cid:18)
IN − 1
N
Furthermore, one can obtain that
⊗ C T
(cid:19)
o∆(0)(cid:0)IN−1 ⊗ C T
(cid:1) xo∆(0). (24)
Due to xo(0) =(cid:0)Uκ(0) ⊗ Ih
(cid:1) xo(0), it can be obtained that
yo(0) =(cid:0)Uκ(0) ⊗ Co
(cid:1) xo(0).
(cid:1)(cid:20)
1⊗Co(0)xo1(0)=(cid:0)Uκ(0)⊗Co
(cid:1)(cid:20)
(cid:21)(cid:19)T(cid:0)Uκ(0) ⊗ Co
(cid:1) xo∆(0) > 0.
(cid:19)
(cid:19)
By (9) and (10), the disagreement output component at t = 0
is
y(0)− 1√
. (25)
N
Because it is supposed that ym(0) (m = 1, 2,··· , N ) are not
agreement completely, it can be shown by (25) that
(cid:18)(cid:0)Uκ(0) ⊗ Co
o∆(0)(cid:0)IN−1 ⊗ C T
(cid:18)(cid:18)
(cid:18)(cid:18)
By (24) and (26), one has
IN − 1
N
o Co
In this case, the given energy budget J∗
⊗ ζC T
(27)
where ζ > 0. For the matrix IN − N−111T , zero eigenvalue
is simple and all the other eigenvalues are positive. By (23)
e . Due to
and (27), ET
e can be denoted by
IN − 1
N
J∗
e = xT
(cid:1)(cid:20)
xo(0) > 0.
⊗ C T
xT
o (0)
xo∆(0)
xo∆(0)
xo∆(0)
xo(0),
(cid:19)
(cid:19)
= xT
o (0)
o Co
o Co
(26)
(cid:21)
11T
11T
0
0
11T
(cid:19)
⊗ C T
(cid:18)(cid:18)
o Rx ≤ ζC T
IN − 1
N
(cid:19)
o Co can ensure that Je ≤ J∗
(cid:19)
(cid:18)(cid:18)
o (0)(cid:0)(cid:0)IN − N−111T(cid:1) ⊗ Ih
(cid:1) yo(0)ET
o Rx ≤ J∗
IN − 1
N
⊗ Ih
(cid:19)
xo(0)
= yT
o (0)
o Co
11T
one can obtain that
e C T
o Co.
xT
o (0)
yo(0),
yT
Based on the above analysis, the conclusion of Theorem 2 can
be obtained.(cid:4)
V. JOINTLY CONNECTED COMMUNICATION TOPOLOGY
CASES
This section investigates limited-budget output consensus
design problems for multiagent systems with jointly connected
communication topologies and gives the corresponding output
consensus criteria in terms of the Cauchy convergence criterion
and Barbalat's lemma.
It is assumed that the time interval [tm, tm+1) consists of a
series of non-overlapping contiguous subintervals as follows
(cid:1),(cid:2)t1
(cid:2)ti
at
m, t1
m
m, t2
m
m − ti
nonnegative
t0
m, t1
m, ti+1
m
where ti+1
a
switch
during
m
, tkm
m
(cid:1), t0
(cid:1),···,(cid:2)tkm−1
(cid:1) (i = 0, 1,··· , km − 1).
m = tm+1,
m ≥ Td (i = 0, 1,···, km − 1) and m is
topologies
integer. The
m,··· , tkm−1
time-invariant
jointly
For
communication
m = tm, tkm
and
is
m
(cid:2)t0
(cid:98)RT
o (cid:98)Rx ≤ 0 and (cid:98)RT
o∆(t)(cid:0)−λ−1
x Ao + AT
o Co < 0, one
o (cid:98)Rx − C T
(cid:1) xo∆(t) ≤ 0,
6
(29)
Vx(t) ≤ xT
min∆κ(t) ⊗ C T
o Co
x Ao + AT
can find by (28) that
which means that Vx(t) converges to a finite value as time
tends to infinity.
In the following, it is shown that Vx(t) converges to zero
as time tends to infinity. By the Cauchy convergence criterion,
for the infinite sequence V (tm) (m = 0, 1,··· ) and any
δ > 0, there exists an integer M > 0 such that the following
inequality holds for ∀m > M : V (tm) − V (tm+1) < δ; that
is,
V (t)dt < δ.
tm
Thus, one can obtain that
(cid:90) t1
m
−
(cid:90) t2
m
−
V (t)dt−
V (t)dt−···−
t0
m
t1
m
Since Td
{0, 1,··· , km − 1}, it can be found by (29) that
the minimum dwell
time,
is
(cid:90) ti
−
m+Td
V (t)dt ≤ −
ti
m
ti
m
(cid:90) ti+1
m
(cid:90) tkm
m
tkm−1
m
V (t)dt < δ. (30)
for any i ∈
(cid:90) tm+1
e C T
o Co,
m+Td
xT
min∆κ(ti
m) ⊗ C T
From (30) to (31), one has
(cid:90) ti
km−1(cid:88)
i=0
ti
m
o∆(t)(cid:0)λ−1
o∆(t)(cid:0)λ−1
Hence, one can show that
(cid:90) t+Td
km−1(cid:88)
IEEE TRANSACTIONS ON CYBERNETICS
,
m
m
m
m
m
m
m
i=0 Lκ(ti
m), Lκ(t1
, Gt1
Lκ(tm)=(cid:80)km−1
connected communication topology cases,
,··· , Gtkm−1
, Gt1
Gt0
connected. Let Lκ(t0
union is
be the Laplacian matrices of Gt0
the topologies
their
can be unconnected, but
m),··· , Lκ(tkm−1
)
,··· , Gtkm−1
then
m) is the Laplacian matrix of a connected
(cid:9)
topology if the communication topologies during [tm, tm+1)
are jointly connected. Here, we set that the transformation
Uκ is time-invariant. Let U T
with ∆κ(ti
can be obtain directly.
m)Uκ = diag(cid:8)0, ∆κ(ti
κ Lκ(ti
m)
Uκ, then the following conclusion
Lemma 3: ∆κ(tm) = (cid:80)km−1
m) is symmetric and
positive definite and its eigenvalues are nonzero eigenvalues
of Lκ(tm).
m) = U T
i=0 ∆κ(ti
κ Lκ(ti
In the following, a limited-budget output consensus design
criterion for jointly connected communication topology cases
is proposed.
Theorem 3: For any given J∗
e > 0, multiagent system (1)
with jointly connected cummnication topologies is limited-
budget output consensualizable by consensus protocol (2) if
(Eo, Ao, Co) is detectable, (Eo, Ao, Bo) is stabilizable, and
m)
12
yT
(II)
(III)
z ET
(I) rank
o Co < 0,
0
Ao Eo
= h + rank (Eo) ,
∗
∗
z AT
0.5BoM
0
−M
ET
x Ao + AT
x Ao + AT
(cid:1) yo(0)ET
o (cid:98)Rx≤J∗
< 0,
there exist (cid:98)Rx and (cid:98)Rz such that
(cid:20) Eo
(cid:21)
o (cid:98)Rx = (cid:98)RT
o (cid:98)Rx ≤ 0,
(cid:98)RT
x Eo ≥ 0,
o (cid:98)Rx − C T
(cid:98)RT
o (0)(cid:0)(cid:0)IN−N−111T(cid:1)⊗Ih
(cid:98)RT
o = Eo(cid:98)Rz ≥ 0,
(cid:98)Θ11
(cid:98)Θm
(cid:98)Θm =
(cid:98)Θ22
Ao(cid:98)Rz + (cid:98)RT
∗
o ≤ 0,
(cid:98)Θ11 = Ao Rz + (cid:98)RT
where λm = λmin, λmax and
(cid:98)Θm
12 = −0.5λm(cid:98)R−T
(cid:98)Θ22 = (cid:98)RT
o (cid:98)R−1
o Co.
z /2 and Kz=−λ−1
(cid:16)
o − BoBT
o ,
o (cid:98)Rx − C T
o Co,
(cid:17)
o (cid:98)Rx
IN−1 ⊗(cid:16)(cid:98)RT
min(cid:98)R−T
min(cid:98)R−T
o∆(t)
x Eo ≥ 0. Let Kz = −λ−1
(cid:17)
(cid:16)
o (cid:98)Rx
o∆(t)
min∆κ(t) ⊗ C T
− λ−1
o (cid:98)Rx = (cid:98)RT
IN−1 ⊗ ET
z AT
x C T
Vx(t) = xT
x Ao + AT
x Ao + AT
Vx(t) = xT
xo∆(t).
xo∆(t),
(cid:17)
o Co
where ET
o /2,
then by taking the time derivative of Vx(t), one can see by
(7) that
x C T
In this case, Ku=−BT
o /2.
Proof: Construct the following Lyapunov function candidate
x C T
(28)
κ Lκ(t)
Since ∆κ(t) = U T
is the Laplacian
matrix of an undirected communication topology, ∆κ(t) is
symmetric and positive definite and its eigenvalues are the
ones of Lκ(t) except one zero eigenvalue. Furthermore, since
Uκ and Lκ(t)
V (t)dt.
(31)
o Co
(cid:1) xo∆(t)dt < δ.
(cid:1)xo∆(t)dt = 0. (32)
o∆(t)(cid:0)λ−1
(cid:1) xo∆(t) = 0.
(33)
m)⊗C T
xT
t
i=0
o Co
min∆κ(ti
lim
t→∞
Due to V (t) ≤ 0, xo∆(t) is bounded. By (7),
bounded. Hence, one can obtain that xT
⊗C T
lemma in [43], one can find that
xo∆(t) is
min∆κ(tm)
o Co ) xo∆(t) is uniformly continuous. From Barbalat's
o∆(t)(cid:0)λ−1
t→∞ xT
lim
min∆κ(tm) ⊗ C T
o Co
By Lemma 3, there exists an orthonormal matrix Utm such
∆κ(tm)Utm = diag {λtm,2, λtm,3,··· , λtm,N} > 0.
that U T
tm
From (32) and (33), one can see that
t→∞ λ−1
minλtm,i
o Co
lim
tm
xi
C T
(cid:0)U T
o∆(t)(cid:1)T
×(cid:0)U T
o (cid:98)Rx = (cid:98)RT
x Eo ≥ 0, (cid:98)RT
o (cid:98)Rx − C T
xi
tm
o∆(t)(cid:1) = 0 (i = 2, 3,··· , N ).
o (cid:98)Rx ≤ 0 and
x Ao + AT
(cid:98)RT
Because (Eo, Ao, Co) is detectable and Us is nonsingu-
lar, ET
x Ao + AT
o Co < 0 in Condition (I) can guarantee
that
limt→∞ xo∆(t) = 0. Similarly, one can show that
limt→∞ zo∆(t) = 0 if Condition (II) holds. Furthermore, the
proofs of the regular, impulse-free and limited-budget property
are similar to Theorem 2. Thus, the conclusions of Theorem
3 can be obtained.(cid:4)
IEEE TRANSACTIONS ON CYBERNETICS
7
VI. NUMERICAL SIMULATION
In this section, we present
two simulation examples to
illustrate the validity of the theoretical results given in the
above section and consider a singular system with five agents.
)
t
(
κ
Fig. 2: Switching signal for switching connected case.
t / s
)
5
,
·
·
·
,
2
,
1
=
m
(
)
t
(
1
m
y
)
5
,
·
·
·
,
2
,
1
=
m
(
)
t
(
2
m
y
t / s
t / s
Fig. 3: Output trajectories for switching connected case.
J∗
e
Je (t)
Fig. 4: Trajectories of Je (t) and J∗
e .
t / s
Fig. 1: Switching connected topologies.
Example 1: (Switching connected case) The system matrices
of each agent in multiagent system (1) are set as
E =
A =
0
1.5
0
1.6
1.6
1.0
0
1
3
2
0
4
0.3556
0.7111
−2.4
2.4
0.4667 −0.8
−2.4
−0.6
−0.8
−2.4
0.6
0.4667
−2.4
1.2
1.2 −1.0667 −1.6 −0.0667
0
2.4 −0.1333 −3.2 −1.1333
3.4 4.5333 −3.2
0.8
2.4 6.8667
−2.6
4.2
2
0
3
3
3.8 5.7333 −2.4 −8.7667 −0.4889
2.5333 −3.0222
0
1.8667 −3.2444 −3
−2.4
−3.9
3
−4
0
2
−3.6667 −6
4
6
6.6
2
7
0.4222
(cid:21)T
(cid:20) −35 −12 −14 −4 −33 9
(cid:20) −1 −4.3333 3 −2.3333 2.8889
18
25
14
12
51
7
B =
C =
−1
−2.0
3
−3.5
1.3333
(cid:21)
,
,
0
0
,
,
and the system can be decomposed into an observable system
by the invertible matrix Uo as
Uo =
.
4
0
3
0
0
0
0 3 0
4 0 1
0 1 3
2 0 2
9 0 0
0 0 4
0
2
0
0
3
0
0
0
0
0
0
1
Fig. 1 shows four different undirected connected topologies,
which are chosen as the switching topologies set κ. For the
convenience of analysis, the communication weights among
five agents of each topology are assumed to be 0-1. Fig. 2
indicates that the switching signal is random.
Let
(cid:20)1 0
(cid:21)
0 1
,
M =
IEEE TRANSACTIONS ON CYBERNETICS
8
e = 50000, then one can
and the energy budget is set as J∗
obtain from Theorem 2 that
−2.6807
14.8810
Rx =
,
,
0
0
−2.6807
−18.1199 −12.2003 −10.0022
20.8005
1.0833
(cid:20) 7.3118 −5.1748 −0.0027
(cid:20)−0.3686 −0.4183
(cid:21)T
−909.7727 −910.3096 −911.3356
11.0466 −8.1502
1.5299
2.2577
1.5299
0.0011
(cid:21)
0
0
Ku =
Rz =
Kz =
−0.2873
0.2121
0
0
.
We set the initial state value of each agent as follows:
x1(0) = [60,−26, 81,−6,−102, 38]T ,
x2(0) = [64,−81, 75,−16,−186, 24]T ,
x3(0) = [−60, 103,−6, 62, 189, 37]T ,
x4(0) = [−68, 49,−30, 48, 114, 19]T ,
x5(0) = [52, 38, 87, 54, 66, 52]T .
In Fig. 3, the output trajectories of the descriptor multiagent
system are depicted, where the red circle markers describe the
trajectory of the output consensus function in Theorem 4. Fig.
4 shows the trajectory of the energy cost function Je(t) with
= 5. One can see that output trajectories of all agents converge
to the curve formed by circle markers and Je(t) < J∗
e , which
means that the descriptor multiagent system achieves limited-
budget output consensus.
Example 2: (Jointly connected case) In this case, the system
matrices of each agent in multiagent system (1) are set as
,
1.5
12
11
22
4
21
−8
9
−5
9
−12 14.5
−10
14
3
0
−10
14
−9 −17
6
−6 −14
4
9 21.25 −12 −28
−12 −28
8
−4
1
0
−12 −27
8
13 −11.5 −8 −15.75
11
22
−4
−1 −2
1.75
31 −38.5 −21 −55.75
33
75
−20
−22 −56
43.5
−64
84
34
26
−24
50.25 −26 −65
(cid:20) −11 −11 10 −6 17 −11
(cid:20) −32
15 −5
58 −2
−32 −76
58
24
29.75 −20 −41
−20 22.5 15
27
16
−41 −19
19
31.5
−3
36
(cid:21)T
30
4
,
(cid:21)
,
,
E =
A =
B =
C =
and the system can be decomposed into an observable system
by the invertible matrix Uo as
Uo =
.
1
0
3
0
1
0
0 3 0
3 0 1
0 4 0
2 0 2
0 0 0
3 0 2
0
1
2
0
1
0
0
0
3 −1
0
0
,
Fig. 5: Jointly connected topologies.
Fig. 5 shows two different jointly connected topologies and
each topology is composed of two unconnected topologies.
For the convenience of analysis, the communication weights
among these agents of each topology are assumed to be 0-1.
The switching order of four topologies is designed as G1 →
G2 → G3 → G4 → G1 → G2 ··· ; that is, the switching
process is repeated every four times as shown in Fig. 6.
Let
M =
(cid:20)1
0
(cid:21)
0
1
,
and the energy budget is given as J∗
Theorem 3, one can show that
e = 10000. Then, by
−6.7489
0
0
0
0
,
,
(cid:21)
,
0
0
0
0
0
26.9957
26.9957
56.8087
(cid:20)−0.1320
(cid:20) 0.2856 −0.1428
−0.1056
56.8087
−0.1785 −0.1428
0
0
−14.2022
0
0.0440 −0.3169
0.1760
0.0264
(cid:21)T
0
0
.
(cid:98)Rx =
(cid:98)Rz =
Ku =
Kz =
We set the initial state value of each agent as follows:
x1(0) = [2,−7,−9,−6, 6,−14]T ,
x2(0) = [14, 41, 15, 42,−4, 52]T ,
x3(0) = [11,−29, 18,−18, 15,−34]T ,
x4(0) = [2, 19,−18, 2,−7, 8]T ,
x5(0) = [16,−3, 33,−14, 27,−21]T .
Fig. 7 depicts the output
trajectories of this descriptor
multiagent system, where the red circle markers depict the
curves of the output consensus function shown in Theorem 4.
Fig. 8 shows the trajectory of the energy cost function Je(t)
with = 14. From these figures, two output trajectories of all
agents converge to the curves formed by the red circle markers
and Je(t) < J∗
e , which means that this descriptor multiagent
system achieves limited-budget output consensus.
IEEE TRANSACTIONS ON CYBERNETICS
9
)
t
(
κ
Fig. 6: Switching signal for jointly connected cases.
t / s
)
5
,
·
·
·
,
2
,
1
=
m
(
)
t
(
1
m
y
)
5
,
·
·
·
,
2
,
1
=
m
(
)
t
(
2
m
y
t / s
t / s
Fig. 7: Output trajectories for jointly connected case.
J∗
e
Je (t)
t / s
Fig. 8: Trajectories of Je (t) and J∗
e .
VII. CONCLUSIONS
For descriptor multiagent systems with both switching con-
nected communication topologies and jointly connected com-
munication topologies, a descriptor dynamic output feedback
consensus protocol with an energy constraint was proposed
to realize limited-budget output consensus. By the matrix
inequality tool, sufficient conditions for limited-budget output
consensus design of multiagent systems with switching con-
nected communication topologies were presented, where a new
two-step design approach was given to deal with the nonlinear
coupled design problems of two gain matrices and those
sufficient conditions are checkable since they are independent
of the number of agents. Furthermore, by combining the
Cauchy convergence criterion and Barbalat's lemma, limited-
budget output consensus design criteria for jointly connected
communication topology cases were proposed, where it was
required that each agent is Lyapunov stable.
REFERENCES
[1] T. Vicsek, A. Czir´ok, E. Ben-Jacob, I. Cohen, and O. Shochet, "Novel
type of phase transition in a system of self-driven particles," Physic
Review Letters, vol. 75, no. 6, pp. 1226-1229, Aug. 1995.
[2] A. Jadbabaie, J. Lin, and A. S. Morse, "Coordination of groups of mobile
autonomous agents using nearest neighbor rules," IEEE Transactions on
Automatic Control, vol. 48, no. 6, pp. 988-1001, Jun. 2003.
[3] R. Olfati-Saber, "Flocking for multi-agent dynamic systems: Algorithms
and theory," IEEE Transactions on Automatic Control, vol. 51, no. 3,
pp. 401-420, Mar. 2006.
[4] L. Tian, Z. Ji, T. Hou, and K. Liu, "Bipartite consensus on coopetition
networks with time-varying delays," IEEE Access, vol. 6, pp. 10169-
10178, Feb. 2018.
[5] Y. Zhang, J. Sun, H. Liang, and H. Li, "Event-triggered adaptive
tracking control for multi-agent systems with unknown disturbances,"
IEEE Transactions on Cybernetics, DOI: 10.1109/TCYB.2018.2869084.
[6] J. Qu, Z. Ji, C. Lin, and H. Yu, "Fast consensus seeking on net-
works with antagonistic interactions," Complexity, 7831317, 2018. (DOI:
10.1155/2018/7831317)
[7] L. Consolini, F. Morbidi, D. Prattichizzo, and M. Tosques, "Leader-
follower formation control of nonholonomic mobile robots with input
constraints," Automatica, vol. 44, no. 5, pp. 1343-1349, May 2008.
[8] K. K. Oh and H. S. Ahn, "Formation control of mobile agents based
on distributed position estimation," IEEE Transactions on Automatic
Control, vol. 58, no. 3, pp. 737-742, Mar. 2013.
[9] X. Dong, Y. Zhou, Z. Ren, and Y. Zhong, "Time-varying formation
tracking for second-order multi-agent systems subjected to switching
topologies with application to quadrotor formation flying," IEEE Trans-
actions on Industrial Electronics, vol. 64, no. 6, pp. 5014-5024, Jun.
2017.
[10] H. Liu, T. Ma, F. L. Lewis, and Y. Wan, "Robust formation control
for multiple quadrotors with nonlinearities and disturbances," IEEE
Transactions on Cybernetics, DOI: 10.1109/TCYB.2018.2875559.
[11] N. Ili´c, M. S. Stankovi´c, and S. S. Stankovi´c, "Adaptive consensus-
based distributed target tracking in sensor networks with limited sensing
range," IEEE Transactions on Control Systems Technology, vol. 22, no.
2, pp. 778-785, Mar. 2014.
[12] Q. Zhou, S. Zhao, H. Li, R. Lu, and C. Wu, "Adaptive neural
network tracking control for robotic manipulators with dead-zone,"
IEEE Transactions on Neural Networks and Learning Systems, DOI:
10.1109/TNNLS.2018.286937.
Jia, N. Cai,
"Adaptive guaranteed-
performance consensus control for multi-agent systems with an ad-
justable convergence speed," Discrete Dynamics in Nature and Society,
https://arxiv.org/abs/1905.03404.
[13] H. Ma, X.
and J. Xi,
[14] T. Zheng, J. Xi, M. Yuan, and G. Liu, "Guaranteed-performance con-
sensus design for Lipschitz nonlinear multi-agent systems with jointly
connected topologies," International Journal of Robust and Nonlinear
Control, to be published. DOI: 10.1002/rnc.4572.
[15] Z. Y. Tan, N. Cai, J. Zhou, and S. G. Zhang, "On performance of peer
review for academic journals: Analysis based on distributed parallel
system," IEEE Access, 2019, DOI: 10.1109/ACCESS.2019.2896978.
IEEE TRANSACTIONS ON CYBERNETICS
10
[16] W. Ren and R. W. Beard, "Consensus seeking in multiagent systems
under dynamically changing interaction topologies," IEEE Transactions
on Automatic Control, vol. 50, no. 5, pp. 655-661, May. 2005.
[17] J. Qin, X. Z. Wei, and H. Gao, "Coordination of multiple agents with
double-integrator dynamics under generalized interaction topologies,"
IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cyber-
netics), vol. 42, no. 1, pp. 44-57, Feb. 2012.
[18] Z. Wu, Y. Xu, Y. Pan, H. Su, and Y. Tang, "Event-triggered control for
consensus problem in multi-agent systems with quantized relative state
measurement and external disturbance," IEEE Transactions on Circuits
and Systems I: Regular paper, DOI: 10.1109/TCSI.2017.2777504.
[19] Z. Wu, Y. Xu, R. Lu, Y. Wu, and T. Huang, "Event-triggered control
for consensus of multi-agent systems with fixed/switching topologies,"
IEEE Transactions on Systems, Man and Cybernetics: Systems, DOI:
10.1109/TSMC.2017.2744671.
[38] J. Xu, G. Zhang, J. Zeng, J. Xi, and B. Du, "Robust guaranteed
cost consensus for high-order discrete-time multi-agent systems with
parameter uncertainties and time-varying delays," IET Control Theory
& Applications, vol. 11, no. 5, pp. 647-667, May. 2017.
[39] J. Xi, C. Wang, H. Liu, and L. Wang, "Completely distributed
guaranteed-performance consensualization for high-order multiagent
systems with switching topologies," IEEE Transactions on Systems,
Man, and Cybernetics: Systems, DOI: 10.1109/TSMC.2018.2852277.
[40] Y. Zhao, G. Guo, and L. Ding, "Guaranteed cost control of mobile sensor
networks with Markov switching topologies," ISA Transactions, vol. 58,
no. 9, pp. 206-213, Sept. 2015.
[41] L. Dai, Singular Control Systems, Springer, Berlin, 1989.
[42] I. Masubuchi, Y. Kamitane, A. Ohara, and N. Suda, "H∞ control for
descriptor systems: a matrix inequalities approach," Automatica, vol. 33,
no. 4, pp. 669-673, Apr. 1997.
[20] B. Cheng, X. Wang, and Z. Li, "Event-triggered consensus of ho-
mogeneous and heterogeneous multi-agent systems with jointly con-
nected switching topologies," IEEE Transactions on Cybernetics, DOI:
10.1109/TCYB.2018.2864974.
[21] J. Xi, Z. Fan, H. Liu, and T. Zheng, "Guaranteed-cost consensus for
multiagent networks with Lipschitz nonlinear dynamics and switching
topologies," International Journal of Robust and Nonlinear Control, vol.
28, no. 7, pp. 2841-2852, May. 2018.
[22] Y. Zheng, J. Ma, and L. Wang, "Consensus of hybrid multi-agent
systems," IEEE Transactions on Neural Networks and Learning Systems,
to be published. DOI: 10.1109/TNNL-S.2017.2651402.
[23] F. Xiao, L. Wang, and T. Chen, "Finite-time consensus in networks
of integrator-like dynamic agents with directional link failure," IEEE
Transactions on Automatic Control, vol. 59, no. 3, pp. 756-762, Mar.
2014.
[24] X. Lin and Y. Zheng, "Finite-time consensus of switched multiagent sys-
tems," IEEE Transactions on Systems, Man, and Cybernetics: Systems,
vol. 47, no. 7, pp. 1535-1545, Jul. 2017.
[25] J. Xi, Z. Shi, and Y. Zhong, "Admissible consensus and consensualiza-
tion of high-order linear time-invariant singular swarm systems," Physica
A: Statistical Mechanics and its Applications, vol. 391, no. 23, pp. 5839-
5849, Nov. 2012.
[26] X. R. Yang and G. P. Liu, "Necessary and sufficient consensus conditions
of descriptor multi-agent systems," IEEE Transactions on Circuits and
Systems-I: Regular Papers, vol. 59, no. 11, pp. 2669-2677, Nov. 2012.
[27] J. Xi, M. He, H. Liu, and J. Zheng, "Admissible output consensualization
control for singular multi-agent systems with time delays," Journal of
the Franklin Institute, vol. 353, no. 16, pp. 4074-4090, Nov. 2016.
[28] S. Wang and J. Huang, "Cooperative output regulation of singular multi-
agent systems under switching networks by standard reduction," IEEE
Transactions on Circuits and Systems-I: Regular Papers, vol. 65, no. 4,
pp. 1377-1385, Apr. 2018.
[29] N. Chopra and W. Spong, "Output synchronization of nonlinear systems
with relative degree one," Recent Advances in Learning Control, vol.
371, no. 1, pp. 51-64, Jan. 2008.
[30] F. Xiao, L. Wang, and J. Chen, "Partial state consensus for networks of
second-order dynamic agents," Systems & Control Letters, vol. 59, no.
12, pp. 775-781, Dec. 2010.
[31] Y. Liu and Y. Jia, "Consensus problem of high-order multi-agent systems
with external disturbances: An H∞ analysis approach," International
Journal of Robust and Nonlinear Control, vol. 20, no. 9, pp. 1579-1593,
Sep. 2009.
[32] H. Kim, H. Shim, and J. H. Seo, "Output consensus of heterogeneous
uncertain linear multi-agent systems," IEEE Transactions on Automatic
Control, vol. 56, no. 1, pp. 200-206, Jan. 2011.
[33] J. Xi, Z. Shi, and Y. Zhong, "Output consensus analysis and design for
high-order linear swarm systems: Partial stability method," Automatica,
2012, vol. 48, no. 9, pp. 2335-2343, Sept. 2012.
[34] Y. C. Cao and W. Ren, "Optimal linear-consensus algorithms: An LQR
perspective," IEEE Transactions on Systems, Man, and Cybernetics -- Part
B Cybernetics, vol. 40, no. 3, pp. 819-829, Mar. 2010.
[35] Z. H. Guan, B. Hu, M. Chi, D. X. He, and X. M. Cheng, "Guaranteed
performance consensus in second-order multi-agent systems with hybrid
impulsive control," Automatica, vol. 50, no. 9, pp. 2415-2418, Sept.
2014.
[36] X. Zhou, P. Shi, C. Lim, C. Yang, and W. Gui, "Event based guaranteed-
the
cost consensus for distributed multi-agent systems," Journal of
Franklin Institute, vol. 352, no. 9, pp. 3546-3563, Mar. 2015.
[37] Y. Zhao and W. Zhang, "Guaranteed cost consensus protocol design
for linear multi-agent systems with sampled-data information: An input
delay approach," ISA Transactions, vol. 67, pp. 87-97, Mar. 2017.
|
1807.02856 | 5 | 1807 | 2019-05-08T16:04:30 | Resilient Synchronization of Distributed Multi-agent Systems under Attacks | [
"cs.MA"
] | In this paper, we first address adverse effects of cyber-physical attacks on distributed synchronization of multi-agent systems, by providing conditions under which an attacker can destabilize the underlying network, as well as another set of conditions under which local neighborhood tracking errors of intact agents converge to zero. Based on this analysis, we propose a Kullback-Liebler divergence based criterion in view of which each agent detects its neighbors' misbehavior and, consequently, forms a self-belief about the trustworthiness of the information it receives. Agents continuously update their self-beliefs and communicate them with their neighbors to inform them of the significance of their outgoing information. Moreover, if the self-belief of an agent is low, it forms trust on its neighbors. Agents incorporate their neighbors' self-beliefs and their own trust values on their control protocols to slow down and mitigate attacks. We show that using the proposed resilient approach, an agent discards the information it receives from a neighbor only if its neighbor is compromised, and not solely based on the discrepancy among neighbors' information, which might be caused by legitimate changes, and not attacks. The proposed approach is guaranteed to work under mild connectivity assumptions. | cs.MA | cs | Resilient Synchronization of Distributed Multi-agent
Systems under Attacks
Aquib Mustafa, Student Member, IEEE, Rohollah Moghadam, Student Member, IEEE, Hamidreza
Modares, Senior Member, IEEE
1
9
1
0
2
y
a
M
8
]
A
M
.
s
c
[
5
v
6
5
8
2
0
.
7
0
8
1
:
v
i
X
r
a
Abstract -- In this paper, we first address adverse effects of
cyber-physical attacks on distributed synchronization of multi-
agent systems, by providing conditions under which an attacker
can destabilize the underlying network, as well as another set
of conditions under which local neighborhood tracking errors
of intact agents converge to zero. Based on this analysis, we
propose a Kullback-Liebler divergence based criterion in view
of which each agent detects its neighbors' misbehavior and,
consequently, forms a self-belief about the trustworthiness of the
information it receives. Agents continuously update their self-
beliefs and communicate them with their neighbors to inform
them of the significance of their outgoing information. Moreover,
if the self-belief of an agent is low, it forms trust on its neighbors.
Agents incorporate their neighbors' self-beliefs and their own
trust values on their control protocols to slow down and mitigate
attacks. We show that using the proposed resilient approach, an
agent discards the information it receives from a neighbor only
if its neighbor is compromised, and not solely based on the dis-
crepancy among neighbors' information, which might be caused
by legitimate changes, and not attacks. The proposed approach
is guaranteed to work under mild connectivity assumptions.
Index Terms -- Distributed control, Resilient Control, Attack
Analysis, Multi-agent systems.
I. Introduction
A Distributed Multi-Agent System (DMAS) is collection of
dynamical systems or agents that interact with each other over
a communication network to achieve coordinated operations
and behaviors [1] -- [4]. In the case of synchronization of
DMASs, the objective is to guarantee that all agents reach
agreement on a common value or trajectory of interest. Despite
their numerous applications in a variety of disciplines, DMASs
are vulnerable to attacks, which is one of the main bottleneck
that arise in their wide deployment. In contrast
to other
undesirable inputs, such as disturbances and noises, cyber-
physical attacks are intentionally planned to maximize the
damage to the overall system or even destabilize it.
There has been extensive research progress in developing
attack detection/identification and mitigation approaches for
both spatially distributed systems [5] -- [14] and DMASs [15] --
[30]. Despite tremendous and welcoming progress, most of
the mentioned mitigation approaches for DMASs use the
discrepancy among agents and their neighbors to detect and
mitigate the effect of an attack. However, as shown in this
paper, a stealthy attack can make all agents become un-
stable simultaneously, and thus misguide existing mitigation
approaches. Moreover, this discrepancy could be caused by a
legitimate change in the state of an agent, and rejecting this
useful information can decrease the speed of convergence to
the desired consensus and harm connectivity of the network.
In this paper, we present attack analysis, detection, and
mitigation mechanisms for DMASs with linear structures.
We show that local neighborhood tracking errors of intact
agents converge to zero, regardless of the attack, if the set of
eigenvalues of the attacker signal generator dynamics matrix
is a subset of the set of eigenvalues of the system dynamics
matrix. We call these types of attacks internal model principle
(IMP)-based attacks. In spite of convergence to zero of local
neighborhood tracking errors, the overall network could be
destabilized, and we provide sufficient conditions for this
to happen. We then develop attack detectors that
identify
both IMP-based and non-IMP-based attacks. To detect IMP-
based attacks, two local error sequences with folded Gaussian
distributions are introduced based on the relative information
of the agents. We show that they diverge under an IMP-
based attack. A Kullback-Liebler (KL) divergence criterion
is then introduced to measure the divergence between these
two univariate folded Gaussian distributions, and consequently
capture IMP-based attacks. Similarly, since non-IMP based
attacks change the statistical properties of the local neighbor-
hood tracking error, to detect non-IMP-based attacks, the KL
divergence is employed to measure the discrepancy between
the Gaussian distributions of the actual and nominal expected
local neighborhood tracking errors. Then, a self-belief value,
as a metric capturing the probability of the presence of attacks
directly on sensors or actuators of the agent itself or on its
neighbors, is presented for each agent by combining these
two KL-based detectors. The self-belief indicates the level of
trustworthiness of the agent's own outgoing information, and
is transmitted to its neighbors. Furthermore, when the self-
belief of an agent is low, the trustworthiness of its incoming
information from its neighbors is estimated using a particular
notion of trust. Trust for each individual neighbor is developed
based on the relative entropy between the neighbor's informa-
tion and agent's own information. Finally, by incorporating
neighbor's self-belief and trust values, we propose modified
weighted control protocols to ensure mitigation of both types
of attacks. Simulation results included in the paper validate
the effectiveness of the approach.
II. Preliminaries
A directed graph (digraph) G consists of a pair (V, E) in
which V = {v1,··· , vN} is a set of nodes and E ⊆ V ×V is a
set of edges. We denote the directed link (edge) from v j to vi
by the ordered pair (v j, vi). The adjacency matrix is defined as
A = [ai j], with ai j > 0 if (v j, vi) ∈ E, and ai j = 0 otherwise.
We assume there are no repeated edges and no self loops, i.e.,
in-degree matrix, with di = (cid:80)
aii = 0 ∀i ∈ N with N = {1, . . . , N} . The nodes in the set
Ni = {v j : (v j, vi) ∈ E} are said to be neighbors of node νi.
The in-degree of vi is the number of edges having vi as a head.
The out-degree of a node vi is the number of edges having vi
as a tail. If the in-degree equals the out-degree for all nodes
vi ∈ V the graph is said to be balanced. The graph Laplacian
matrix is defined as L = D − A, where D = diag(di) is the
ai j as the weighted in-degree
of node νi. A node is called as a root node if it can reach all
other nodes of the digraph G through a directed path. A leader
is a root node with no incoming link. A (directed) tree is a
connected digraph where every node except one, called the
root, has in-degree equal to one. A spanning tree of a digraph
is a directed tree formed by graph edges, which connects all
the nodes of the graph.
j∈Ni
Throughout the paper, we denote the set of integers by
Z. The set of integers greater than or equal to some integer
q ∈ Z is denoted Z(cid:62)q. The cardinality of a set S is denoted
by S. λ(A) and tr(A) denote, respectively, the eigenvalues
and trace of the matrix A. Furthermore, λmin(A) represents
minimum eigenvalue of matrix A. The Kronecker product of
matrices A and B is denoted by A ⊗ B, and diag (A1, . . . , An)
represents a block diagonal matrix with matrices Ai, ∀ i ∈ N
as its diagonal entries. 1N is the N-vector of ones and IN
is the N × N identity matrix. Im(R) and ker(R) represent,
respectively, the range space and the null space of R, and
span(a1, . . . , an) is the set of all linear combinations of the
vectors a1, . . . , an. A Gaussian distribution with mean µ and
covariance Σ is denoted by N (µ, Σ). Moreover, F N(cid:16)
¯µ, ¯σ2(cid:17)
represents univariate folded Gaussian distribution with ¯µ and
¯σ2 as mean and variance, respectively [?]. E[.] denotes the
expectation operator.
Assumption 1. The communication graph G is directed and
has a spanning tree.
Definition 1 [31]- [32]. A square matrix A ∈ Rn×n is called
a singular M-matrix, if all its off-diagonal elements are non-
positive and all its eigenvalues have non-negative real parts. (cid:3)
Definition 2 [31]- [32]. A square matrix A ∈ Rn×n is called
a non-singular M-matrix, if all its off-diagonal elements are
non-positive and all its eigenvalues have positive real parts. (cid:3)
Lemma 1 [31]- [32]. The graph Laplacian matrix L of a
directed graph G has at least one zero eigenvalue, and all its
nonzero eigenvalues have positive real parts. Zero is a simple
eigenvalue of L, if and only if Assumption 1 is satisfied.
III. Overview of Consensus in DMASs
In this section, we provide an overview of the consensus
problem for leaderless DMAS. Consider a group of N homo-
geneous agents with linear identical dynamics described by
∀ i ∈ N
xi(t) = Axi(t) + Bui(t)
(1)
where xi ∈ Rn and ui ∈ Rm denote, respectively, the state
and the control input of agent i. The matrices A ∈ Rn×n and
B ∈ Rn×m are, respectively, the drift dynamics and the input
matrix.
2
Problem 1. Design local control protocols ui for all agents
∀i ∈ N in (1) such that all agents reach consensus or synchro-
nization (agreement) on some common value or trajectory of
interest, i.e.,
t→∞x j(t) − xi(t) = 0 ∀i, j ∈ N
lim
(2)
Assumption 2. The system dynamics matrix A in (1) is
assumed to be marginally stable.
Remark 1. Based on Assumption 2, all the eigenvalues of A
have non-positive real part. This is a standard assumption in
the literature for consensus or synchronization problems [33].
Note that if A is Hurwitz, the synchronization problem has a
trivial solution and can be solved by making the dynamics of
each agent stable independently. Moreover, stable eigenvalues
of A, if there are any, can be ignored by reducing the dimension
of A, because they only contribute to the transient response of
the consensus trajectories [33].
Consider the distributed control protocol for each agent i as
[31]- [32]
ui = cKηi
∀ i ∈ N
where
ηi =
ai j(x j − xi)
N(cid:88)
(3)
(4)
j=1
represents the local neighborhood tracking error for the agent i
with ai j as the (i, j)-th entry of the graph adjacency matrix A.
Moreover, c and K ∈ Rm×n denote, respectively, scalar coupling
gain and feedback control gain matrix. The control design (3)-
(4) is distributed in the sense that each agent seeks to make the
difference between its state and those of its neighbors equal
to zero using only relative state information of its neighbors
provided by (4).
1
Several approaches are presented to design c and K locally
the gains K
to solve Problem 1 [31]- [34]. To this end,
and c are designed such that A − cλiBK is Hurwitz for all
i = 2, . . . , N [31]- [34]. Specifically, it is shown in [31], [34]
that under Assumption 1, if K is designed locally for each
agent by solving an algebraic Riccati equation and c >
2λmin(L),
then A − cλiBK is Hurwitz and Problem 1 is solved. In the
subsequent sections, we assume that c and K are designed
appropriately so that in the absence of attacks Problem 1 is
solved. We then analyze the effect of attacks and propose
mitigation approaches.
Remark 2. Note that the presented results subsume the leader-
follower synchronization problem and the average consensus
as special cases. For the leader-follower case, the leader is
only root node in the graph and thus the desired trajectory is
dictated by the leader, whereas for the average consensus case,
the graph is assumed to be balanced and A = 0 and B = Im.
IV. Attack Modelling and Analysis for DMASs
In this section, attacks on agents are modelled and a
complete attack analysis is provided.
A. Attack Modelling
In this subsection, attacks on DMASs are modelled. Attacks
on actuators of agent i can be modelled as
uc
i
= ui + βiud
i
(5)
i and uc
where ui, ud
i denote, respectively, the nominal value
of the control protocol for agent i in (1), the disrupted signal
directly injected into actuators of agent i, and the corrupted
control protocol of agent i. If agent i is under actuator attack,
then βi = 1, otherwise βi = 0. Similarly, one can model attacks
on sensors of agent i as
xc
i
= xi + αixd
i
(6)
i and xc
where xi, xd
i denote, respectively, the nominal value of
the state of agent i in (4), the disrupted signal directly injected
into sensors of agent i, and the corrupted state of agent i. If
agent i is under sensor attack, then αi = 1, otherwise αi = 0.
Using the corrupted state (6) in the controller (3)-(4) with the
corrupted control input (5) in (1), the system dynamics under
attack becomes
xi = Axi + Bui + B fi
(7)
where fi denotes the overall attack affecting the agent i which
can be written as
fi = βiud
i
+ cK
α jxd
j − αixd
i
ai j
(8)
(cid:16)
(cid:88)
j∈Ni
(cid:17)
with xd
j as the disruption in the received state of the jth
neighbor due to injected attack signal either into sensors or
actuators of agent j or into the incoming communication link
from agent j to agent i.
The following definition categorizes all attacks into two
categories. The first type of attack exploits the knowledge of
the system dynamics A and use it in the design of its attack
signal. That is, for the first type of attack for fi in (8), one has
Define
fi = Ψ fi
(9)
where Ψ ∈ Rm×m depends on the knowledge of the system
dynamics A as discussed in Definition 3. On the other hand,
for the second type of attack, the attacker has no knowledge
of the system dynamics A and this can cover all other attacks
that are not in the form of (9).
(cid:40) EΨ = {λ1(Ψ), . . . , λm(Ψ)}
EA = {λ1(A), . . . , λn(A)}
(10)
where λi(Ψ) ∀i = 1, . . . , m and λi(A) ∀i = 1, . . . , N are, re-
spectively, the set of eigenvalues of the attack signal generator
dynamics matrix Ψ and the system dynamics matrix A. Define
a set of eigenvalues of the system dynamics matrix A which lie
A ⊆ EA.
on imaginary axis as Em
A
Definition 3 (IMP-based and non-IMP-based Attacks).
fi in (7) is generated by (9), then the
If the attack signal
attack signal is called the internal model principle (IMP)-based
attack, if EΨ ⊆ EA. Otherwise, i.e., EΨ (cid:49) EA or if the attacker
has no dynamics (e.g. a random signal), it is called a non-IMP
(cid:3)
based attack.
= {λ1(A), . . . , λl(A)} where Em
3
Remark 3. Note that we do not limit attacks to the IMP-based
attacks given by (9). Attacks are placed into two classes in
Definition 3 based on their impact on the system performance,
as to be shown in the subsequent sections. The non-IMP based
attacks cover a broad range of attacks.
Definition 4 (Compromised and Intact Agent). We call an
agent that is directly under attack as a compromised agent.
An agent is called intact if it is not compromised. We denote
the set of intact agents as NInt, i.e.,NInt = N\NComp where
NComp denotes the set of compromised agents.
(cid:3)
Using (3)-(4), the global form of control input, i.e., u =
1 , . . . , uT
N]T can be written as
[uT
u = (−cL ⊗ K)x
(11)
where L denotes the graph Laplacian matrix.
By using (11) in (7), the global dynamics of agents under
attack becomes
x(t) = (IN ⊗ A) x(t) + (IN ⊗ B) uc(t)
where
uc(t) = u(t) + f (t) ∆
= (−cL ⊗ K)x(t) + f (t)
(12)
(13)
1 (t), . . . , f T
with f (t) = [ f T
N(t)]T
denote, respectively, the overall vector of attacks on agents
and the global vector of the states of agents.
N (t)]T and x(t) = [xT
1 (t), . . . , xT
If agents are not under attack, i.e., f (t) = 0, then, the control
input (11) eventually compensates for the difference between
the agents' initial conditions and becomes zero once they reach
an agreement. That is, in the absence of attack, uc(t) = u(t)
goes to zero (i.e., uc(t) → 0), and, on consensus, the global
dynamics of agents become
xss(t) = (IN ⊗ A) xss(t)
(14)
t→∞x(t) is called the global steady state of agents.
where xss = lim
Throughout the paper, regardless of whether agents are under
attack or not, we say that agents reach a steady state and their
steady state is generated by (14) if uc(t) → 0. Otherwise, if
uc(t) (cid:54)→ 0, we say agents never reach a steady state, and thus
(14) does not hold true.
Remark 4. In the presence of attack, whether agents reach
a steady state or not, i.e., whether uc(t) → 0 or uc(t) (cid:54)→ 0,
plays an important role in the attack analysis and mitigation
to follow. Reaching a steady state is necessary for agents to
achieve consensus. However, we show that even if agents reach
a steady state, they will not achieve consensus if the system is
under attack. More specifically, we show that under a non-
IMP based attack, agents do not reach a steady state and
their local neighborhood tracking errors also do not converge
to zero. For an IMP-based attack, the attacker can either 1)
make all agents reach a steady state, but agents are still far
from synchronization or consensus, or 2) destabilize the entire
network by assuring that agents do not reach a steady state.
B. Attack Analysis
In this subsection, a graph theoretic-based approach is
utilized to analyze the effect of attacks on DMASs. To this
end, the following notation and lemmas are used.
Let the graph Laplacian matrix L be partitioned as
(cid:34) Lr×r
L =
0r×nr
Lnr×r Lnr×nr
(cid:35)
(15)
where r and nr in (15) denote, respectively, the number of
root nodes and non-root nodes. Moreover, Lr×r and Lnr×nr
are, respectively, the sub-graph matrices corresponding to the
sub-graphs of root nodes and non-root nodes.
Lemma 2. Consider the partitioned graph Laplacian matrix
(15). Then, Lr×r is a singular M-matrix and Lnr×nr is a non-
singular M-matrix.
Proof. We first prove that
the subgraph of root nodes is
strongly connected. According to the definition of a root node,
there always exists a directed path from a root node to all other
nodes of the graph G, including other root nodes. Therefore, in
the graph G, there always exists a path from each root node
to all other root nodes. We now show that removing non-
root nodes from the graph G does not affect the connectivity
of the subgraph comprised of only root nodes. In the graph
G, if a non-root node is not an incoming neighbor of a root
node, then its removal does not harm the connectivity of the
subgraph of the root nodes. Suppose that removing a non-root
node affects the connectivity of the subgraph of root nodes.
This requires the non-root node to be an incoming neighbor
of a root node. However, this makes the removed node a root
node, as it can now access all other nodes through the root
node it is connected to. Hence, this argument shows that the
subgraph of root nodes is always strongly connected. Then,
based on Lemma 1, Lr×r has zero as one of its eigenvalues,
which implies that Lr×r is a singular M-matrix according to
Definition 1. On the other hand, from (15), since L is a lower
triangular matrix, the eigenvalues of L are the union of the
eigenvalues of Lr×r and Lnr×nr. Moreover, as stated in Lemma
1, L has a simple zero eigenvalue and, as shown above, zero
is the eigenvalue of Lr×r. Therefore, all eigenvalues of Lnr×nr
have positive real parts only, and thus based on Definition 2,
Lnr×nr is a non-singular M-matrix.
(cid:3)
In the following Lemmas 3-4 and Theorem 1, we now
provide the conditions under which the agents can reach a
steady state.
Lemma 3. Consider the global dynamics of DMAS (12) under
attack. Let the attack signal f (t) be a non-IMP based attack.
Then, agents never reach a steady state, i.e., uc(t) (cid:54)→ 0.
Proof. We prove this result by contradiction. Assume that the
f (t) is a non-IMP based attack, i.e., EΨ (cid:49) EA,
attack signal
but uc(t) → 0 in (12), which implies xi → Axi for all i ∈ N.
Using the modal decomposition, one has
(r jxi(0))eλ j(A)tm j
(16)
xi(t) → n(cid:88)
j=1
4
ai j(x j(t) − xi(t))
fi(t) →(cid:88)
where r j and m j denote, respectively,
the left and right
eigenvectors associated with the eigenvalue λ j(A). On the other
hand, based on (13) uc(t) → 0 implies f (t) → (cL ⊗ K)x(t) or
equivalently
j∈Ni
(17)
for all i ∈ N. As shown in (16), the right-hand side of (17)
is generated by the natural modes of the system dynamics
whereas the left-hand side is generated by the natural modes
of the attack signal generator dynamics in (9). By the prior
assumption, EΨ (cid:49) EA, the attacker's natural modes are differ-
ent from those of the system dynamics. Therefore, (17) cannot
be satisfied which contradicts the assumption. This completes
(cid:3)
the proof.
Equation (17) in Lemma 3 also shows that for non-IMP
based attacks, the local neighborhood tracking error is nonzero
for a compromised agent. The following results show that
under IMP-based attack, either agents' state diverge, or they
reach a steady state while their local neighborhood tracking
errors converge to zero, despite attack. The following lemma
is needed in Theorem 1, which gives conditions under which
agents reach a steady state under IMP-based attack. Then,
Theorem 2 shows that under what conditions an IMP-based
attack makes the entire network of agents unstable.
(cid:40) S A(t) = [eλA1 t, . . . , eλAn t]
Define
S ψ(t) = [eλΨ1 t, . . . , eλΨn t]
(18)
where eλAi t ∀i = 1, . . . , n and eλΨi t ∀i = 1, . . . , m are, respec-
tively, the set of natural modes of agent dynamics A in (1) and
the attacker dynamics Ψ in (9).
Lemma 4. Consider the global dynamics of DMAS (12) under
attack on non-root nodes. Then, for an IMP-based attack,
agents reach a steady state, i.e., uc(t) → 0.
Proof. According to (14), in steady state, one has xss(t) →
(IN ⊗ A) xss(t) since uc(t) → 0. This implies that xss(t) ∈
span(S A) where S A is defined in (18). On the other hand, if
agents reach a steady state, then based on (13), one has
(cL ⊗ K)xss(t) = f (t)
Define the global steady state vector xss(t) = [ ¯xT
(19)
nrs]T ,
where ¯xrs and ¯xnrs are, respectively, the global steady states
of root nodes and non-root nodes. Since attack is only on
non-root nodes, f (t) can be written as f (t) = [0r, ¯f T
nr]T , where
¯fnr = [ f T
N ]T represents the attack vector on non-root
nodes.
r+1, . . . , f T
rs, ¯xT
Then, using (15) and (19), one has
(cLr×r ⊗ K) ¯xrs = 0
(cLnr×r ⊗ K) ¯xrs + (cLnr×nr ⊗ K) ¯xnrs = ¯fnr
(20)
As stated in Lemma 2, Lr×r is a singular M-matrix with zero
as an eigenvalue and 1r is its corresponding right eigenvector
and, thus, the solution to the first equation of (20) becomes
¯xrs = c11r for some positive scalar c1. Using ¯xrs = c11r in the
second equation of (20), the global steady states of non-root
nodes becomes
¯xnrs = (cLnr×nr ⊗ K)−1(cid:104)−(cLnr×r ⊗ K)c11r + ¯fnr
(21)
(cid:105)
(cid:40)
Equation (21) shows that the steady states of non-root nodes
f (t). If EΨ (cid:49) EA, it results
are affected by the attack signal
in ¯xnrs ∈ span(S A, S Ψ) where S A and S Ψ are defined in (18)
which contradicts xss(t) ∈ span(S A). Therefore, condition EΨ ⊂
EA is necessary to conclude that for any f = [0r, ¯f T
nr]T , there
exists a steady state solution xss(t), i.e., uc(t) → 0 holds true.
(cid:3)
This completes the proof.
The following theorem provides necessary and sufficient
conditions for IMP-based attacks to assure uc(t) → 0.
Theorem 1. Consider the global dynamics of DMAS (12)
with the control protocol (13), where the attack signal f (t) is
generated based on an IMP-based attack. Then, agents reach
a steady state, i.e., uc(t) → 0 if and only if the attack signals
satisfy
pk fk = 0
(22)
N(cid:88)
k=1
where pk are the nonzero elements of the left eigenvector of the
graph Laplacian matrix L associated with its zero eigenvalue.
Proof. It was shown in the Lemma 4 that for the IMP-
based attack on non-root nodes, agents reach a steady state,
i.e., uc(t) → 0. Therefore, whether agents reach a steady
state or not depends solely upon attacks on root nodes. Let
f (t) = [ ¯fr, ¯fnr] where ¯fr represents the vector of attacks for
root nodes given by ¯fr = [ f T
r ]T . Now, we first prove
the necessary condition for root nodes. If uc(t) → 0, then,
using (15) and (19) , there exists a nonzero vector ¯xrs for root
nodes such that
1 , . . . , f T
(cLr×r ⊗ K) ¯xrs = ¯fr
(23)
where ¯xrs can be considered as the global steady state of the
root nodes. Moreover, based on Lemma 3, (23) does not hold,
if EΨ (cid:49) EA which implies that (23) is true only for EΨ ⊆ EA.
As stated in Lemma 2, Lr×r is a strongly connected graph
of root nodes and, therefore, it is a singular M-matrix. Let
¯wT = [p1, . . . , pr] be the left eigenvector associated with the
zero eigenvalue of Lr×r. Now, pre-multiplying both sides of
(23) by ¯wT and using the fact that ¯wTLr×r = 0 yield
¯wT (cLr×r ⊗ K) ¯xrs = ¯wT ¯fr = 0
(24)
k=1
This states that IMP-based attacks on root nodes have to
pk fk = 0 to ensure agents reach a steady state, i.e.,
satisfy
uc(t) → 0. Note that pk = 0 for k = r + 1, . . . , N, i.e., the
elements of the left eigenvector of the graph Laplacian matrix
L, corresponding to its zero eigenvalue, are zero for non-root
nodes [31]- [32]. This proves the necessity part.
N(cid:80)
Now, we prove the sufficient part by contradiction for root
nodes. Assume agents reach a steady state, i.e., uc(t) → 0, but
pk fk (cid:44) 0. Note that, agents reach a steady state implies that
k=1
there exists a nonzero vector ¯xrs such that (23) holds. Using
pk fk (cid:44) 0, one can conclude that ¯wT (cLr×r⊗K) ¯xrs (cid:44)
(24) and
0. This can happen only when Lr×r does not have any zero
eigenvalue, which violates the fact in Lemma 2 that Lr×r is
a strongly connected graph. Therefore, ¯wT (cLr×r ⊗ K) ¯xrs = 0
N(cid:80)
k=1
N(cid:80)
N(cid:80)
5
which results in
made. This completes the proof.
pk fk = 0 and contradicts the assumption
(cid:3)
k=1
A
Theorem 2. Consider the global dynamics of DMAS (12) with
the control protocol (13) under IMP-based attack. If (22) is
not satisfied and EΨ ∩ Em
(cid:44) ∅, then the dynamics of agents
become unstable.
Proof. Since it is assumed that the condition in (22) is not
satisfied, then based on Theorem 1, uc(t) (cid:54)→ 0 even under
IMP-based attack. Thus, the attack signal f (t) does not vanish
over time and eventually acts as an input to the system in
(12). Assume that there exists at least one common marginal
eigenvalue between the system dynamics matrix A in (1) and
the attacker dynamics matrix Ψ in (9), i.e., EΨ∩Em
(cid:44) ∅. Then,
the multiplicity of at least one marginally stable pole becomes
greater than 1. Therefore, the attacker destabilizes the state of
the agent in (12). Moreover, since (22) is not satisfied, then
the attack is on root nodes, and root nodes have a path to all
other nodes in the network, the state of the all agents become
(cid:3)
unstable. This completes the proof.
A
Theorem 3 below now shows that despite IMP-based at-
tacks, if uc(t) → 0, the local neighborhood tracking error (4)
converges to zero for intact agents that have a path to the
compromised agent, while they do not synchronize.
Theorem 3. Consider the global dynamics of DMAS (12)
under attack f (t). Then, the local neighborhood tracking error
(4) converges to zero for all intact agents if uc(t) → 0. More-
over, intact agents that are reachable from the compromised
agents do not converge to the desired consensus trajectory.
Proof. In the presence of attacks, the global dynamics of the
DMAS (12) with (13) can be written as
x(t) = (IN ⊗ A) x(t) + (IN ⊗ B)((−cL ⊗ K)x(t) + f (t))
(25)
1 (t), . . . , xT
xT
N(t)
f T
1 (t), . . . , f T
where x(t) =
is the global vector of the state
of agents and f (t) =
denotes the global
vector of attacks. As shown in (14) that if uc(t) → 0, agents
reach a steady state. That is,
N (t)
cKηi → − fi
∀ i ∈ N
(26)
where ηi denotes the local neighborhood tracking error of
agent i defined in (4). For the intact agent, by definition one
has fi = 0, and thus (26) implies that the local neighborhood
tracking error (4) converges to zero. Now, we show that intact
agents which are reachable from the compromised agent do
not synchronize to the desired consensus behavior. To do this,
j be under attack. Assuming that all intact agents
let agent
synchronize, one has xk = xi ∀i, k ∈ N−{ j}. Now, consider the
intact agent i as an immediate neighbor of the compromised
agent j. Then using (13), if uc(t) → 0, for intact agent i, i.e.,
fi = 0, one has(cid:88)
ai j(xk − xi) + (x j − xi) → 0
(27)
(cid:105)T
(cid:104)
(cid:105)T
(cid:104)
k∈Ni−{ j}
where xk denotes the state of the all
intact neighbors of
agent i. On the other hand, (7) shows that the state of the
j, i.e., x j, is deviated from the desired
compromised agent
consensus value with a value proportional to f j. Therefore,
(27) results in deviating the state of the immediate neighbor of
the compromised agent j from the desired consensus behavior,
which contradicts the assumption. Consequently, intact agents
that have a path to the compromised agent do not reach
consensus, while their local neighborhood tracking error is
(cid:3)
zero. This completes the proof.
Remark 5. The effects of an attacker on a network of agents
depend upon the dynamics of the attack signal. As stated in
Theorem 2, to destabilize the entire network, the attack signal
requires access to at least one common marginal eigenvalue
with the system dynamics. To this end, an attacker can exploit
the security of the network by eavesdropping and monitoring
the transmitted data to identify at least one of the marginal
eigenvalues of the agent dynamics, and then launch a signal
with the same frequency to a root node to make the agents
state go to infinity.
Remark 6. Although, for the sake of simplicity, we consider
DMASs with identical dynamics, the presented result can be
extended to heterogeneous MASs. This is briefly discussed
in the following formulation. The dynamics for a linear
heterogeneous MASs is given by
(cid:40) xi(t) = Aixi(t) + Biui(t)
∀ i ∈ N
yi(t) = Cixi(t)
(28)
where xi ∈ Rni , ui ∈ Rmi and yi ∈ Rp denote, respectively, the
state, the control input and the output of agent i. The matrices
Ai ∈ Rni×ni , Bi ∈ Rni×mi and Ci ∈ Rp×ni are, respectively, the
drift dynamics, the input matrix and the output matrix.
For heterogeneous MASs, the consensus trajectory is usu-
ally generated by a virtual exosystem dynamics given by [35]-
[36]
(cid:40) xc(t) = S xc(t)
yc(t) = Rxc(t)
(29)
where xc ∈ Rq and yc ∈ Rp are, respectively, the state and
output of the desired consensus trajectory. For heterogeneous
MASs the distributed control protocol ui in (28) is designed
such that all agents synchronize to the output of virtual exosys-
tem trajectory [35]- [36]. The attacker can design IMP-based
attacks by exploiting the knowledge of consensus dynamics
S in (29), instead of agents' dynamics and all the analysis
results presented in Section IV are valid for the heterogeneous
MASs. In this case, to launch an IMP-based attack, the attacker
should satisfy EΨ ⊆ ES where ES = {λ1(S ), . . . , λq(S )} with
λi(S ) ∀i = 1, . . . , q as the set of eigenvalues of the virtual
exosystem drift dynamics matrix S .
Up to now, the presented analysis has been under the as-
sumption that the communication is noise free. We now briefly
discuss what changes if the communication noise is present,
and propose attack detection and mitigation in the presence of
communication noise. In the presence of Gaussian distributed
communication noise, the local neighborhood tracking error
in (4) becomes
¯ηi = ηi + ωi,
(30)
6
where ωi ∼ N(0, Σωi) denotes the aggregate Gaussian noise
affecting the incoming information to agent i and is given as
(31)
(cid:88)
ai jωi j,
ωi =
j∈Ni
with ωi j the incoming communication noise from agent j to
agent i. In such situations, the DMAS consensus problem
defined in Problem 1 changes to the mean square consensus
problem. In the presence of Gaussian noise, based on (30), the
control protocol in (3)-(4) becomes
∀ i ∈ N,
ai j(x j(t) − xi(t)) + ωi
j=1
ui(t) = cK
(32)
where ωi ∼ N(0, Σωi) is defined in (31). Based on mean square
consensus, one has
t→∞ E[ui(t)] →0
lim
∀ i ∈ N,
(33)
N(cid:88)
ss = (IN ⊗ A) xm
xm
ss,
and thus, based on (1), the steady state of agents converge to a
consensus trajectory in mean square sense and its global form
in (14) becomes
(34)
where xm
ss = limt→∞E[x(t)] denotes the global steady state
of agents in mean square sense. Then, following the same
procedure as Lemmas 3-4 and Theorems 1-3, one can show
that an IMP-based attack does not change the statistical
properties of the local neighborhood tracking error, while a
non-IMP based attack does. Moreover, the local neighborhood
tracking error converges to zero in mean for an IMP-based
attack, and it does not converges to zero in mean for a non-
IMP based attack.
In the next section, attack detection and mitigation mech-
anisms are proposed for both IMP-based and non-IMP based
attacks. To this end, it is assumed that the communication
network is noisy.
V. An Attack Detection Mechanism
In this section, Kullback-Liebler (KL)-based attack detec-
tion and mitigation approaches are developed for both IMP-
based and non-IMP-based attacks.
The KL divergence is a non-negative measure of the relative
entropy between two probability distributions [37], [38] which
is defined as follows.
Definition 5 (KL divergence) [37], [38]. Let X and Z be two
random sequences with probability density functions PX and
PZ, respectively. The KL divergence measure between PX and
PZ is defined as
(cid:90)
(cid:32) PX(θ)
(cid:33)
PZ(θ)
DKL(XZ) =
PX(θ) log
dθ
(35)
with the following properties [37]:
1) DKL(PXPz) (cid:62) 0
2) DKL(PXPz) = 0 if and only if, PX = Pz
3) DKL(PXPz) (cid:44) DKL(PzPX)
In the following subsections, KL-divergence is used to
detect IMP-based and non-IMP-based attacks on DMASs.
A. Attack detection for IMP-based attacks
In this subsection, an attack detector is designed to identify
IMP-based attacks. To this end, two error sequences τi and
ϕi are defined based on only local exchanged information for
agent i as
(36)
(37)
and
τi =
ϕi =
ai jdi j
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:88)
(cid:88)
(cid:13)(cid:13)(cid:13)ai jdi j
j∈Ni
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)
j∈Ni
where the measured discrepancy di j between agent i(cid:48)s state
and its neighbor j(cid:48)s state under attack becomes
j − xc
i
+ ωi j ∀ j ∈ Ni
di j = xc
(38)
where ωi j ∼ N(0, Σωi j) denotes the Gaussian incoming com-
munication noise from agent j to agent i. Moreover, xc
i is the
measured state of agent i under attack and xc
j is the possibly
corrupted information it receives from its jth neighbor. If agent
i is not compromised, then xc
= xi, and, similarly, if agent j
i
is not compromised, then xc
= x j. In fact, (36) is the norm
j
of the summation of the measured discrepancy of agent i and
all its neighbors, and (37) is the summation of norms of those
measured discrepancies. In the absence of attack, these two
signals show the same behavior in the sense that their means
converge to zero.
In the presence of an IMP-based attack and in the absence
of noise, based on Theorem 3, τi goes to zero for intact
agents, despite attack. However, it is obvious that ϕi does
not converge to zero in the presence of an attack. In the
presence of noise, the statistical properties of τi converge to
the statistical properties of the noise. In contrast, the statistical
properties of ϕi depend upon not only the statistical properties
of the noise signal, but also of the attack signal. Therefore,
the behavior of these two signals significantly diverges in the
presence of attacks and can be captured by KL-divergence
methods. Note that one can measure τi and ϕi based on the
exchanged information among agents, which might be cor-
rupted by the attack signal. Existing KL-divergence methods
are, nevertheless, developed for Gaussian signals. However,
while the communication noise is assumed to be Gaussian,
error sequences (36) and (37) are norms of some variable
with Gaussian distributions, thus, they have univariate folded
Gaussian distributions given by [39] ϕi ∼ F N(µ1i, σ2
1i) and
τi ∼ F N(µ2i, σ2
2i) . That is,
1√
2πσ1i e
1√
2πσ2i e
− (qi−µ1i)2
2σ2
1i +
− (qi−µ2i)2
2σ2
2i +
Pϕi(qi, µ1i, σ1i) =
Pτi(qi, µ2i, σ2i) =
− (qi +µ1i)2
2σ2
1i
− (qi +µ2i)2
2σ2
2i
1√
2πσ1i e
1√
2πσ2i e
(39)
where µ1i and σ1i are the mean and variance of the error
sequences ϕi and µ2i and σ2i are the mean and variance of the
7
Pτi(qi)
Pϕi(qi)
Pτi(qi)
error sequences τi. Using (35), the KL divergence in terms of
the local error sequences ϕi and τi can be defined as
DKL(ϕiτi) =
(cid:32) Pϕi(qi)
dqi = E1
Pϕi(qi) log
(cid:90)
log
(cid:33)
(cid:32)
(cid:33)
(40)
where E1[.] represents the expectation value with respect to
the distribution of the first sequence [39].
A KL divergence formula for the folded Gaussian distribu-
tions is now developed in the following lemma.
Lemma 5. Consider the error sequences τi and ϕi in (36)-
(37) with folded Gaussian distributions Pϕi and Pτi in (39).
Then, the KL divergence between error sequences τi and ϕi,
i.e., DKL(ϕiτi), becomes
σ2
2i
σ2
1i
− µ2
1i
2σ2
1i
DKL(ϕiτi) ≈ 1
2
8µ2
1i
σ2
1i
1
2i (µ2i − µ1i)2 + 1
−2
2 σ
− 1 + (σ
−2
2i σ2
1i)
ρ2
4
2σ2
1i
ρ2
2
2σ2
1i
4µ2
1i
σ2
1i
(cid:33)
ρ2
3
2σ2
1i + e
ρ2
1
2σ2
1i + e
+
e
log
(cid:32)
+ e
1 − e
−
e
1
2
e
1
2
for some ρ1 = (µ1i − 2µ2iσ2
ρ3 = (µ1i − 4µ2iσ2
Proof. See Appendix A.
1iσ−2
1iσ−2
2i ), ρ2 = (µ1i + 2µ2iσ2
2i ) and ρ4 = (µ1i + 4µ2iσ2
1iσ−2
2i ).
(41)
1iσ−2
2i ),
(cid:3)
In the following theorem, we show that the effect of IMP-
based attacks can be captured using the KL divergence defined
in (41).
Theorem 4. Consider the DMAS (1) along with the controller
(13), and under the IMP-based attacks. Assume that
the
communication noise sequences are i.i.d. Then, for a reachable
intact agent i,
DKL(ϕiτi)dk > γi
(42)
(cid:90) k+T−1
1
T
k
where ϕi and τi are defined in (36) and (37), respectively,
and T and γi represent the window size and the predesigned
threshold parameter.
Proof. According to Theorem 3, the local neighborhood track-
ing error goes to zero for intact agents in the presence
of an IMP-based attack when there is no communication
noise. In the presence of communication noise with Gaussian
distribution, i.e., ωi j ∼ (0, Σωi j) and IMP-based attack, the
expectation value of the local neighborhood tracking error for
intact agent i becomes
E[ηi] = E[
(43)
where the measured discrepancy di j is defined in (38). Using
(43), one can write (36) as
j∈Ni
ai jdi j] → 0
(cid:88)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∼ F N(0, ¯υ2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:88)
j∈Ni
τi =
ai jdi j
ωi)
(44)
which represents a folded Gaussian distribution with mean
zero and variance ¯υ2
ωi. Note that the mean and variance of
the distribution Pτi in (39) become µ2i = 0 and σ2
2i
= ¯υ2
ωi.
Since noise signals are independent and identically dis-
tributed, from (37), one can infer that the folded Gaussian
distribution Pϕi in (39) has the following statistical properties
ϕi ∼ F N(µ f d
i
, ¯υ2
ωi
+ υ2
ωi
+ ¯υ2
f d
i
)
(45)
i
+ υ2
ωi
+ ¯υ2
f d
i
and ¯υ2
ωi
where µ f d
represent the overall mean and
covariance due to the communication noise and overall devi-
ation from the desired behavior in intact neighbors reachable
from the compromised agent.
sponding to sequences τi and ϕi become F N(0, ¯υ2
F N(0, ¯υ2
divergence in (41) becomes
In the absence of attack, the statistical properties corre-
ωi) and
ωi) , respectively, and the corresponding KL
+ υ2
ωi
KL(ϕiτi) ≈ 1
Dwa
2
¯υ2
ωi
+ υ2
ωi
¯υ2
ωi
+ ¯υ
−2
ωi υ2
ωi))
(46)
where υ2
depends on the communication noise.
ωi represents additional variance in sequence ϕi, which
Note that τi
in (36) is the norm of the summation of
the measured discrepancy of agent i and all its neighbors
whereas ϕi
in (37) is the summation of norms of those
measured discrepancies. Even in the absence of attack, they
represent folded Gaussian distributions with zero means and
different covariances due to application of norm on measured
discrepancies.
Now, in the presence of IMP-based attacks, using the de-
rived form of KL divergence for folded Gaussian distributions
from Lemma 5, one can simplify (41) using (44)-(45) as
log
log
(cid:90) k+T−1
1
T
k
DKL(ϕiτi) ≈ 1
2
¯υ2
ωi
+ υ2
ωi
¯υ2
ωi
+
1
2
−2
¯υ
ωi (µ f d
i
)2 +
1
2
e
4(µ f d
i
+υ2
ωi
¯υ2
ωi
)2
+¯υ2
f d
i
+ ¯υ2
f d
i
1 − e
8(µ f d
i
+υ2
ωi
¯υ2
ωi
)2
+¯υ2
f d
i
+ ¯υ
−2
ωi (¯υ2
f d
i
+ υ2
ωi)
Then, one can design the threshold parameter γi such that
DKL(ϕiτi)dk > γi
(48)
8
B. Attack detection for non-IMP-based attacks
This subsection presents the design of a KL-based attack
detector for non-IMP based attacks.
It was shown in Theorem 3 that the local neighborhood
tracking error goes to zero if agents are under IMP-based
attacks. Therefore, for the case of non-IMP-based attacks,
one can identify these types of attacks using the changes in
the statistical properties of the local neighborhood tracking
error. In the absence of attack, since the Gaussian noise, i.e.,
ωi ∼ N(0, Σωi), is considered in the communication link, the
local neighborhood tracking error ηi in (30) has the following
statistical properties
ηi ∼ N(0, Σωi)
(50)
and it represents the nominal behavior of the system.
In the presence of attacks, using (30), the local neighbor-
hood tracking error ηa
i can be written as
ηa
i
=
ai jdi j
(51)
(cid:88)
j∈Ni
where measured discrepancy under attacks di j is defined (38).
From (51), one has
i ∼ N(µ fi , Σ fi
ηa
+ Σωi)
(52)
where µ fi and Σ fi are, respectively, mean and covariance of
the overall deviation due to corrupted states under attacks as
given in (38).
Now, since both ηa
the KL divergence in the terms of ηa
i and ηi have normal Gaussian distri-
i and ηi as
butions,
DKL(ηa
i ηi) can be written as [40]
log
1
2
1
(µηi − µηa
2
i
(cid:12)(cid:12)(cid:12)Σηi
(cid:12)(cid:12)(cid:12)Σηa
(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12) − n + tr(Σ−1
ηi
i )
Σηa
ηi (µηi − µηa
i )
i )T Σ−1
(53)
(47)
DKL(ηa
i ηi) =
+
i and Σηa
i denote the mean and covariance of ηa
where µηi and Σηi denote the mean and covariance of ηi and
i . Moreover,
µηa
n denotes the dimension of the error sequence. Define the
average of KL divergence over a window T as
i ηi)dk
(cid:90) k+T−1
DKL(ηa
¯Di =
(54)
1
T
k
The following theorem says that the effect of non-IMP based
attacks can be detected using the KL divergence between the
two error sequences ηa
Theorem 5. Consider the DMAS (1) along with the controller
(13). Then,
i and ηi.
1) in the absence of attack, ¯Di defined in (54) tends to zero.
2) in the presence of a non-IMP-based attack, ¯Di defined
in (54) is greater than a predefined threshold γi.
Proof. In the absence of attacks, the statistical properties of
sequences ηi and ηa
i are the same as in (50). Therefore, the KL
i ηi) in (53) becomes zero and this makes
divergence DKL(ηa
¯Di in (54) zero. This completes the proof of part 1.
where T denotes the sliding window size. This completes the
(cid:3)
proof.
Based on Theorem 4, one can use the following conditions
for attack detection.
1
T
1
T
(cid:90) k+T−1
(cid:90) k+T−1
k
k
DKL(ϕiτi)dk < γi
DKL(ϕiτi)dk > γi
: H0
: H1
(49)
where γi denotes the designed threshold for detection, the null
hypotheses H0 represents the intact mode and H1 denotes the
compromised mode of an agent.
To prove Part 2, using (50)-(52) in (53) and the fact that
ωi (Σ fi
Σ fi), one can write the KL divergence
(Σ−1
between ηa
DKL(ηa
(cid:12)(cid:12)(cid:12)
ωi
1
2
(log
+Σωi)−n = tr(Σ−1
i and ηi as
i ηi) =
k+T−1(cid:90)
(cid:12)(cid:12)(cid:12)Σωi
(cid:12)(cid:12)(cid:12)Σ fi
(cid:12)(cid:12)(cid:12)Σωi
(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)Σ fi
log
1
2
1
T
+ Σωi
+ Σωi
k
ωi
(cid:12)(cid:12)(cid:12) + tr(Σ−1
(cid:12)(cid:12)(cid:12) + tr(Σ−1
ωi
Then, using (54), one has
¯Di =
Σ fi) + µT
fi
Σ−1
ωi µ fi) (55)
and consequently makes the value of c1
i (t) close to zero. On the
other hand, without an attack, DKL(ϕiτi) tends to zero, making
i (t) approach 1 and, consequently, c1
χ1
i (t) becomes close to 1.
The larger the value of c1
i (t) is, the more confident the agent
is about the trustworthiness of its broadcasted information.
Similarly,
for
the non-IMP-based attacks, using the
DKL(ηa
i ηi) from Theorem 5, we define c2
i (t) as
9
(60)
(61)
t(cid:90)
eκ2(τ−t)χ2
i (τ)dτ
0
∆i
∆i + DKL(ηa
i ηi)
Σ fi) + µT
fi
Σ−1
ωi µ fi) > γi
c2
i (t) = κ2
(56)
where T and γi denote the sliding window size and the
predefined design threshold, respectively. This completes the
(cid:3)
proof.
where 0 (cid:54) c2
i (t) (cid:54) 1 with
χ2
i (t) =
Based on Theorem 5, one can use the following conditions
for attack detection:
¯Di < γi : H0
¯Di > γi
: H1
(57)
where γi denotes the designed threshold for detection, the null
hypotheses H0 represents the intact mode of the system and
H1 denotes the compromised mode of the system.
In the next section, Theorems 4 and 5 are employed to
propose an attack mitigation approach which enables us to
mitigate both IMP-based attacks and non-IMP-based attacks.
VI. An Attack Mitigation Mechanism
In this section, both IMP-based and non-IMP-based attacks
are mitigated using the presented detection mechanisms in the
previous section.
A. Self-belief of agents about their outgoing information
To determine the level of trustworthiness of each agent
about its own information, a self-belief value is presented.
If an agent detects an attack, it reduces its level of trustwor-
thiness about its own understanding of the environment and
communicates it with its neighbors to inform them about the
significance of its outgoing information and thus slow down
the attack propagation.
For the IMP-based attacks, using the DKL(ϕiτi) from The-
orem 4, we define c1
i (t) as
t(cid:90)
c1
i (t) = κ1
eκ1(τ−t)χ1
i (τ)dτ
where 0 (cid:54) c1
i (t) (cid:54) 1 with
0
χ1
i (t) =
∆i
∆i + DKL(ϕiτi)
(58)
(59)
where ∆i represents the threshold to account for the channel
fading and other uncertainties and κ1 > 0 denotes the discount
factor. Equation (58) can be implemented by the following
differential equation
c1
i (t) + κ1c1
i (t)
attacks, DKL(ϕiτi) increases, which makes χ1
According to Theorem 4, in the presence of IMP-based
i (t) approach zero
i (t) = κ1χ1
where ∆i represents the threshold to account for the channel
fading and other uncertainties, and κ2 > 0 denotes the discount
factor. Expression (60) can be generated by
i (t)
c2
i (t) + κ2c2
i (t) = κ2χ2
i (t), one can show that c2
Using Theorem 6 and the same argument as we employed
for c1
i (t) is close to 1 in the absence
of an attack, and close to zero in the presence of a non-IMP
based attack.
i (t) and c2
Then, using c1
i (t) defined in (58) and (60), the self-
belief of an agent i for both IMP and non-IMP-based attacks
is defined as
ξi(t) = min{c1
i (t), c2
i (t)}
(62)
If an agent i is under direct attack or receives corrupted
information from its neighbors, then the self-belief of the
agent i tends to zero. In such a situation, it transmits the
low self-belief value to its neighbor to put less weight on
the information they receive from it and this prevents attack
propagation in the distributed network.
B. Trust of agents about their incoming information
The trust value represents the level of confidence of an
agent on its neighbors' information. If the self-belief value
of an agent is low, it forms beliefs on its neighbors (either
intact or compromised) and updates its trust value which
depends on the beliefs on each of its neighbors using only
local information. Therefore, agents identify the compromised
neighbor and discard its information.
Using the KL divergence between exchanged information
of agent i and its neighbor, one can define ηi j(t) as
t(cid:90)
ηi j(t) = κ3
eκ3(τ−t)Li j(τ)dτ
where 0 (cid:54) ηi j(t) (cid:54) 1 with
0
Li j(t) = 1 −
with mi = (cid:80)
(cid:18)
Λ1
DKL(x jmi)
−Λ2
(cid:19)
∀ j ∈ Ni
Λ1 + e
x j and Λ1, Λ2 > 0 represent the threshold to
account for channel fading and other uncertainties, and κ3 > 0
denotes the discount factor. For the compromised neighbor, the
j∈Ni
(63)
(64)
KL divergence DKL(x jmi) tends to zero, which makes Li j(t)
close to zero. Consequently, this makes the value of ηi j(t)
close to zero. On the other hand, if the incoming neighbor
is not compromised, then DKL(x jmi) increases and makes
ηi j(t) approach 1. Equation (63) can be implemented using
the following differential equation
ηi j(t) + κ3ηi j(t) = κ3Li j(t)
Now, we define the trust value of an agent on its neighbors
Ωi j(t) = max(ξi(t), ηi j(t))
(65)
as
with 0 (cid:54) Ωi j(t) (cid:54) 1.
In the absence of attacks, the state of agents converge to the
consensus trajectory and the KL divergence DKL(x jmi), ∀ j ∈
Ni tends to zero which results in Ωi j(t) being 1 ∀ j ∈ Ni. In the
presence of attacks, ηi j(t) corresponding to the compromised
agents tends to zero.
C. The mitigation mechanism using trust and self-belief values
In this subsection, the trust and self-belief values are utilized
to design the mitigation algorithm. To achieve resiliency, both
self-belief and trust values are incorporated into the exchange
information among agents. Consequently, the resilient form of
local neighborhood tracking error (30) is presented as
ηi =
Ωi j(t)ξ j(t)ai j
+ ωi
(66)
(cid:88)
j∈Ni
(cid:16)
(cid:17)
x j − xi
where Ωi j(t) and ξ j(t) denote, respectively, the trust value and
the self-belief of neighboring agents. Using (3) and (66), the
resilient control protocol becomes
ui = cK ηi,
∀ i ∈ N
(67)
According to (66), the topology of the graph changes over
time due to incorporation of the trust and the self-belief values
of agents, thus we denote the time-varying graph as G(t) =
(V, E(t)) with E(t) ⊆ V × V representing the set of time-
varying edges.
Now, based on following definitions and lemma, we for-
mally present Theorem 6 to illustrate that the trust and self-
belief based proposed resilient control protocol (67) solves
Problem 1 and all intact agents, i.e., NInt = N\NComp as
defined in Definition 4 achieve the final desired consensus
value for DMAS in (1), despite attacks.
Definition 6 (r-reachable set) [41]. Given a directed graph
G and a nonempty subset Vs ⊂ V, the set Vs is r-reachable
if there exists a node i ∈ Vs such that Ni\Vs (cid:62) r, where
r ∈ Z(cid:62)0.
(cid:3)
Definition 7 (r-robust graph) [41]. A directed graph G is
called an r-robust graph with r ∈ Z(cid:62)0 if for every pair of
nonempty, disjoint subsets of V, at least one of the subsets is
(cid:3)
r-reachable.
Assumption 3. If at most q neighbors of each intact agents is
under attack, at least (q + 1) neighbors of each intact agents
are intact.
Lemma 6. [41] Consider an r-robust time-varying directed
graph G(t). Then, the graph has a directed spanning tree, if
and only if G(t) is 1-robust.
10
Theorem 6. Consider the DMAS (1) under attack with
the proposed resilient control protocol ui
in (67). Let the
time-varying graph G(t) be such that at each time in-
stant t, Assumption 1 and Assumption 3 are satisfied. Then,
t→∞x j(t) − xi(t) = 0 ∀i, j ∈ NInt.
lim
Proof. The DMAS (1) with the proposed resilient control
(cid:88)
protocol ui in (67) without noise can be written as
(cid:16)
(cid:17)
xi = Axi + cBK
ai j(t)
x j − xi
(68)
j∈Ni
where ai j(t) = Ψi j(t)C j(t)ai j. The global form of resilient
system dynamics in (68) becomes
x = (IN ⊗ A − cL(t) ⊗ BK)x
(69)
where L(t) denotes the time-varying graph Laplacian matrix
of the directed graph G(t). Based on Assumption 3, even
if q neighbors of an intact agent are attacked and collude
to send the corrupted value to misguide it, there still exists
q + 1 intact neighbors that communicate values different from
least q + 1 of
the compromised ones. Moreover, since at
the intact agent's neighbors are intact, it can update its trust
values to remove the compromised neighbors. Furthermore,
since the time varying graph G(t) resulting from isolating the
compromised agents is 1-robust, based on Definition 7 and
Lemma 6, the entire network is still connected to the intact
agents. Therefore, there exists a spanning tree in the graph
associated with all intact agents NInt. Hence, it is shown in
[42] that the solutions of DMAS in (69) reach consensus
on desired behavior if the time-varying graph G(t) jointly
contains a spanning tree as the network evolves with time. This
t→∞x j(t) − xi(t) = 0 ∀i, j ∈ NInt assymptotically.
results in lim
This completes the proof.
(cid:3)
Remark 7. The proposed approach discards the compromised
agent only when an attack is detected, in contrast to most of
the existing methods that are based on solely the discrepancy
among agents. Note that discrepancy can be the result of a
legitimate change in the state of one agent. Moreover, in the
beginning of synchronization, there could be a huge discrep-
ancy between agents' states that should not be discarded.
VII. Simulation Results
In this section, an example is provided to illustrate the effec-
tiveness of the proposed detection and mitigation approaches.
Consider a group of 5 homogeneous agents with the dynamics
defined as
xk = Axk + Buk
k = 1, . . . , 5
(70)
where
A =
(cid:35)
(cid:34) 0 −1
1
0
, B =
(cid:35)
.
(cid:34) 1
0
The communication graph is shown in Fig. 1. In the absence
of an attack, agents reach desired synchronization and there
emerges the healthy behavior of the system with noisy com-
munication as shown in Fig. 2.
11
Fig. 1: Communication topology.
Fig. 3: The state of agents when Agent 1 is under an IMP-based attack.
(a)
(b)
(a)
(b)
Fig. 2: Desired synchronization in the absence of attack. (a) The state of agents. (b) The
local neighborhood tracking error of agents.
Fig. 4: Agent 5 is under IMP-based attack. (a) The state of agents. (b) The local
neighborhood tracking error of agents.
A. IMP-based attacks
This subsection analyzes the effects of IMP-based attacks
and illustrates our attack detection and mitigation scheme. The
attack signal is assumed to be f = 20 sin(t). This is an IMP-
based attack and is assumed to be launched on Agent 1 (root
node) at time t=20. The results are shown in Fig. 3. It can
be seen that the compromised agent destabilizes the entire
network. This result is consistent with Theorem 2. It is shown
in Fig. 4 that the same IMP-based attack on Agent 5 (noon-root
node) cannot destabilize the entire network. However, Agent
4, which is the only agent reachable from Agent 5, does not
synchronize to the desired consensus trajectory. Moreover, one
can see that the local neighborhood tracking error converges
to zero for all intact agents except the compromised Agent
5. These results are in line with Theorem 3. Fig.5 shows the
divergence in the presence of non-IMP based attack on Agent 5
based on Theorem 4. Then, the effect of attack is rejected using
the presented belief-based detection and mitigation approach
in Theorem 4 and Theorem 6. Fig.6 shows that reachable
agents follow the desired consensus trajectory, even in the
presence of the attack.
B. Non-IMP-based attacks
This subsection analyzes the effects of non-IMP-based at-
tacks and validates our attack detection and mitigation ap-
proach. The attack signal is assumed to be f = 10 + 5 sin(2t).
Fig. 5: Divergence for state of agents when Agent 5 is under a IMP-based attack.
Fig. 6: The state of agents using the proposed attack detection and mitigation approach
for IMP-based attack.
The effect of this non-IMP-based attack on Agent 5 (non-root
node) is shown in Fig.7. It can be seen that this non-IMP-based
010203040Time (s)-2-1012Agent 1Agent 2Agent 3Agent 4Agent 5010203040Time (s)-50510Agent 1Agent 2Agent 3Agent 4Agent 5510152025303540Time (s)-1000-50005001000Agent 1Agent 2Agent 3Agent 4Agent 501020304050Time (s)-505Agent 1Agent 2Agent 3Agent 4Agent 501020304050Time (s)-10-50510Agent 1Agent 2Agent 3Agent 4Agent 501020304050Time (s)0200400600800Agent 1Agent 2Agent 3Agent 4Agent 501020304050Time (s)-10-50510Agent 1Agent 2Agent 3Agent 4Agent 5 attack on Agent 5 only affects the reachable Agent 4. Then,
Fig.8 shows the divergence in the presence of non-IMP based
attack on Agent 5 based on Theorem 5. It is shown in Fig.9 that
the effect of the attack is removed for the intact Agent 4 using
belief-based detection and mitigation approaches presented in
Theorems 5 and 6.
A possible direction for future work is to extend these
results to synchronization of DMASs with nonlinear dynam-
ics. Since nonlinear systems can exhibit finite-time escape
behavior, a problem of interest
is to find the conditions
under which the attacker can make the trajectories of agents
become unbounded in finite time, and to obtain detection and
mitigation mechanisms to counteract such attacks fast and thus
avoid instability.
12
Appendix A
Proof of Lemma 5
Using (41), the KL divergence between error sequences ϕi
and τi can be written as
DKL(ϕiτi) = E1[log Pϕi − log Pτi]
(71)
Fig. 7: The state of agents when Agent 5 is under a non-IMP-based attack.
Fig. 8: Divergence for state of agents when Agent 5 is under a non-IMP based attack.
Fig. 9: The state of agents after attack detection and mitigation for non-IMP based attack.
VIII. Conclusion
A resilient control framework has been introduced for
DMASs. First, the effects of IMP-based and non-IMP-based
attacks on DMASs have been analyzed using a graph-theoretic
approach. Then, a KL divergence based criterion, using only
the observed local information of agents, has been employed to
detect attacks. Each agent detects its neighbors' misbehaviors,
consequently forming a self-belief about the correctness of its
own information, and continuously updates its self-belief and
communicates it with its neighbors to inform them about the
significance of its outgoing information. Additionally, if the
self-belief value of an agent is low, it forms beliefs on the type
of its neighbors (intact or compromised) and, consequently,
updates its trust of its neighbors. Finally, agents incorporate
their neighbors' self-beliefs and their own trust values in their
control protocols to slow down and mitigate attacks.
2σ2
1i
]
where probability density functions Pϕi and Pτi are defined in
(39). Using (39), (71) becomes
DKL(ϕiτi)
= E1[log
− (qi−µ1i)2
2σ2
1i +
1√
2πσ1i e
1√
2πσ1i e
− (qi +µ1i)2
(72)
− log
1√
2πσ2i e
− (qi−µ2i)2
2σ2
2i +
− (qi +µ2i)2
2σ2
2i
1√
2πσ2i e
By the aid of the logarithm property as log(a +b) = log(a) +
log(1 + b/a), (72) turns into
= E1[log
DKL(ϕiτi) =
− (qi−µ1i)2
]
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
1√
1√
2πσ2i e
2πσ1i e
(cid:33)
(cid:32)
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
− log
(cid:33)
+ E1[log
− log
(cid:32)
− 2qi µ2i
σ2
2i
− 2qi µ1i
σ2
1i
− (qi−µ2i)2
1 + e
1 + e
2σ2
1i
2σ2
2i
T1
]
T2
The first
(73)
term in (73) is a KL divergence formula for
statistical sequences with normal Gaussian distribution which
is given in [40] as
σ2
2i
σ2
1i
− 1 + (σ
−2
2i σ2
1i)
2i (µ2i − µ1i)2
−2
1
2 σ
(74)
The second term T2 in (73), using power series expansion
and ignoring higher order
(n + 1)
log
log(1 + a) = (cid:80)
T1 =
1
2
n(cid:62)0
(−1)nan+1(cid:46)
(cid:16)
+
(cid:17)
terms, can be approximated as
− 2qi µ1i
σ2
1i − (e
T2 ≈ E1[e
which can be expressed as
2
− 2qi µ1i
σ2
1i )
2
− 2qi µ2i
σ2
2i − (e
2
− 2qi µ2i
σ2
2i )
2
]
(75)
∞(cid:90)
−∞
T2 ≈
∞(cid:90)
−
−∞
] − E1[e
∞(cid:90)
1i dqi − 1
∞(cid:90)
2
−∞
1
2
−∞
− 2qi µ1i
σ2
Pϕie
− 4qi µ1i
σ2
1i dqi
Pϕie
− 2qi µ2i
σ2
2i dqi +
Pϕie
− 4qi µ2i
σ2
2i dqi
Pϕie
(76)
01020304050Time (s)-505Agent 1Agent 2Agent 3Agent 4Agent 501020304050Time (s)01020304050Agent 1Agent 2Agent 3Agent 4Agent 501020304050Time (s)-50510Agent 1Agent 2Agent 3Agent 4Agent 5 (78)
(79)
(80)
(81)
(82)
where ρ3 = (µ1i − 4µ2iσ2
2i ) and ρ4 = (µ1i + 4µ2iσ2
Adding (78), (80), (83) and (84), T2 can be written as
1iσ−2
1iσ−2
2i ) .
−
e
ρ2
3
2σ2
1i + e
ρ2
4
2σ2
1i
ρ2
1
2σ2
1i + e
ρ2
2
2σ2
1i
e
1
1 − e
2
8µ2
1i
σ2
1i
T2 ≈ e
− µ2
1i
2σ2
1i
+1 +
4µ2
1i
σ2
1i
e
1
2
13
(85)
Now, using (74)-(75) and (85), one gets (41). This completes
the proof.
References
[1] R. Olfati-Saber, J. A. Fax, and R. M. Murray, "Consensus and coop-
eration in networked multi-agent systems," Proceedings of the IEEE,
vol. 95, pp. 215 -- 233, Jan 2007.
[2] F. Bullo, J. Cortes, and S. Martinez, Distributed Control of Robotic Net-
works: A Mathematical Approach to Motion Coordination Algorithms.
Princeton University Press, 2009.
[3] A. Khanafer and T. Bas¸ar, "Robust distributed averaging: When are
potential-theoretic strategies optimal?," IEEE Transactions on Automatic
Control, vol. 61, pp. 1767 -- 1779, July 2016.
[4] Q. Zhu and T. Bas¸ar, "Game-theoretic methods for robustness, security,
and resilience of cyberphysical control systems: Games-in-games prin-
ciple for optimal cross-layer resilient control systems," IEEE Control
Systems, vol. 35, pp. 46 -- 65, Feb 2015.
[5] Y. Mo, R. Chabukswar, and B. Sinopoli, "Detecting integrity attacks
on scada systems," IEEE Transactions on Control Systems Technology,
vol. 22, pp. 1396 -- 1407, July 2014.
[6] Y. Shoukry, P. Martin, P. Tabuada, and M. Srivastava, "Non-invasive
spoofing attacks for anti-lock braking systems," in International Work-
shop on Cryptographic Hardware and Embedded Systems, pp. 55 -- 72,
Springer, 2013.
[7] H. Fawzi, P. Tabuada, and S. Diggavi, "Secure estimation and control for
cyber-physical systems under adversarial attacks," IEEE Transactions on
Automatic Control, vol. 59, pp. 1454 -- 1467, June 2014.
[8] M. Pajic, J. Weimer, N. Bezzo, P. Tabuada, O. Sokolsky, I. Lee,
and G. J. Pappas, "Robustness of attack-resilient state estimators," in
2014 ACM/IEEE International Conference on Cyber-Physical Systems
(ICCPS), pp. 163 -- 174, April 2014.
[9] F. Pasqualetti, F. Drfler, and F. Bullo, "Attack detection and identification
in cyber-physical systems," IEEE Transactions on Automatic Control,
vol. 58, pp. 2715 -- 2729, Nov 2013.
[10] Y. Shoukry and P. Tabuada, "Event-triggered state observers for sparse
sensor noise/attacks," IEEE Transactions on Automatic Control, vol. 61,
pp. 2079 -- 2091, Aug 2016.
[11] Y. Mo and B. Sinopoli, "Secure estimation in the presence of integrity
attacks," IEEE Transactions on Automatic Control, vol. 60, pp. 1145 --
1151, April 2015.
[12] A. Kanellopoulos and K. G. Vamvoudakis, "Non-equilibrium dynamic
games and cyber -- physical security: A cognitive hierarchy approach,"
Systems & Control Letters, vol. 125, pp. 59 -- 66, 2019.
[13] C. De Persis and P. Tesi, "Input-to-state stabilizing control under denial-
of-service," IEEE Transactions on Automatic Control, vol. 60, no. 11,
pp. 2930 -- 2944, 2015.
[14] Y. Yan, P. Antsaklis, and V. Gupta, "A resilient design for cyber physical
systems under attack," in 2017 American Control Conference (ACC),
pp. 4418 -- 4423, May 2017.
[15] F. Pasqualetti, A. Bicchi, and F. Bullo, "Consensus computation in
unreliable networks: A system theoretic approach," IEEE Transactions
on Automatic Control, vol. 57, pp. 90 -- 104, Jan 2012.
[16] A. Teixeira, H. Sandberg, and K. H. Johansson, "Networked control
systems under cyber attacks with applications to power networks," in
Proceedings of the 2010 American Control Conference, pp. 3690 -- 3696,
June 2010.
[17] S. Amin, X. Litrico, S. S. Sastry, and A. M. Bayen, "Cyber security
of water scada systems-part ii: Attack detection using enhanced hydro-
dynamic models," IEEE Transactions on Control Systems Technology,
vol. 21, pp. 1679 -- 1693, Sept 2013.
[18] S. Weerakkody, X. Liu, S. H. Son, and B. Sinopoli, "A graph-theoretic
characterization of perfect attackability for secure design of distributed
control systems," IEEE Transactions on Control of Network Systems,
vol. 4, pp. 60 -- 70, March 2017.
− (qi +5µ1i)2−24µ2
1i
2σ2
1i
dqi + e
dqi
− (qi +3µ1i)2−8µ2
1i
2σ2
1i
e
4µ2
1i
σ2
1i + e
12µ2
1i
σ2
1i
Pϕie
− 4qi µ1i
σ2
1i dqi = −1
2
∞(cid:90)
e
− (qi−µ1i)2
2σ2
1i e
− 2qi µ2i
σ2
2i + e
− (qi +µ1i)2
2σ2
1i e
dqi
− 2qi µ2i
σ2
2i
∞(cid:90)
−1
2
−∞
= −
√
2
− 4qi µ1i
σ2
1i dqi
Pϕie
−∞
e
∞(cid:90)
−∞
1
2πσ1i
∞(cid:90)
− 1
2
−∞
which yields
The third term of T2 is
∞(cid:90)
−
−∞
= −
− 2qi µ1i
σ2
1i dqi
Pϕie
−∞
1√
2πσ1i
∞(cid:90)
e
−
−∞
= −
− 2qi µ1i
σ2
1i dqi
Pϕie
∞(cid:90)
1i−ρ2
− µ2
1
2σ2
1i
∞(cid:90)
1i−ρ2
− µ2
2
2σ2
1i
+e
−∞
where ρ1 = (µ1i − 2µ2iσ2
which becomes
which can be written in the form
Now, the first term of T2 can be written as
∞(cid:90)
∞(cid:90)
−∞
=
−∞
− 2qi µ1i
σ2
1i dqi
Pϕie
1√
2πσ1i e
∞(cid:90)
− (qi +µ1i)2
2σ2
1i dqi +
∞(cid:90)
−∞
− (qi +µ1i)2 +4qi µ1i
2σ2
1i
1√
2πσ1i e
dqi
(77)
Using the fact that density integrates to 1, (77) becomes
− 2qi µ1i
σ2
1i dqi = 1 + e
4µ2
1i
σ2
1i
Pϕie
Similarly, second term of T2 can be written as
1√
2πσ1i e
− (qi−ρ1)2
2σ2
1i dqi
− (qi−ρ2)2
2σ2
1i dqi
−∞
1√
2πσ1i e
1iσ−2
2i ) and ρ2 = (µ1i + 2µ2iσ2
∞(cid:90)
−
−∞
∞(cid:90)
−∞
1
2
− 2qi µ1i
σ2
1i dqi = −
Pϕie
1i−ρ2
− µ2
1
2σ2
1i + e
1i−ρ2
− µ2
2
2σ2
1i
Similarly, the last term of T2 can be simplified as
− 2qi µ1i
σ2
1i dqi =
Pϕie
1
2
1i−ρ2
− µ2
3
2σ2
1i + e
1i−ρ2
− µ2
4
2σ2
1i
e
e
1iσ−2
2i )
(83)
(84)
[19] Z. Feng and G. Hu, "Distributed secure average consensus for linear
multi-agent systems under dos attacks," in 2017 American Control
Conference (ACC), pp. 2261 -- 2266, May 2017.
[20] S. Sundaram and C. N. Hadjicostis, "Distributed function calculation
via linear iterative strategies in the presence of malicious agents," IEEE
Transactions on Automatic Control, vol. 56, pp. 1495 -- 1508, July 2011.
[21] W. Zeng and M. Y. Chow, "Resilient distributed control in the presence
of misbehaving agents in networked control systems," IEEE Transac-
tions on Cybernetics, vol. 44, pp. 2038 -- 2049, Nov 2014.
[22] X. Jin, W. M. Haddad, and T. Yucelen, "An adaptive control architecture
for mitigating sensor and actuator attacks in cyber-physical systems,"
IEEE Transactions on Automatic Control, vol. 62, pp. 6058 -- 6064, Nov
2017.
[23] H. J. LeBlanc and X. Koutsoukos, "Resilient first-order consensus and
weakly stable, higher order synchronization of continuous-time net-
worked multi-agent systems," IEEE Transactions on Control of Network
Systems, pp. 1 -- 1, 2017.
[24] S. Bolouki, D. G. Dobakhshari, T. Bas¸ar, V. Gupta, and A. Nedi´c,
"Applications of group testing to security decision-making in networks,"
in 2017 IEEE 56th Annual Conference on Decision and Control (CDC),
pp. 2929 -- 2934, Dec 2017.
[25] E. Akyol, T. Bas¸ar, and C. Langbort, "Signaling games in networked
cyber-physical systems with strategic elements," in 2017 IEEE 56th
Annual Conference on Decision and Control (CDC), pp. 4576 -- 4581,
Dec 2017.
[26] M. O. Sayin and T. Bas¸ar, "Secure sensor design for cyber-physical
systems against advanced persistent threats," in Decision and Game
Theory for Security (S. Rass, B. An, C. Kiekintveld, F. Fang, and
S. Schauer, eds.), (Cham), pp. 91 -- 111, Springer International Publishing,
2017.
[27] S. M. Dibaji, H. Ishii, and R. Tempo, "Resilient randomized quantized
consensus," IEEE Transactions on Automatic Control, pp. 1 -- 1, 2017.
[28] S. M. Dibaji, M. Pirani, D. B. Flamholz, A. M. Annaswamy, K. H.
Johansson, and A. Chakrabortty, "A systems and control perspective of
cps security," 2019.
[29] R. Moghadam and H. Modares, "An internal model principle for the
attacker in distributed control systems," in 2017 IEEE 56th Annual
Conference on Decision and Control (CDC), pp. 6604 -- 6609, Dec 2017.
[30] Z. Feng and G. Hu, "Secure cooperative event-triggered control of linear
multiagent systems under dos attacks," IEEE Transactions on Control
Systems Technology, pp. 1 -- 12, 2019.
[31] Z. Li and Z. Duan, Cooperative Control of Multi-Agent Systems: A
Consensus Region Approach. Automation and Control Engineering,
Taylor & Francis, 2014.
[32] F. Lewis, H. Zhang, K. Hengster-Movric, and A. Das, Cooperative Con-
trol of Multi-Agent Systems: Optimal and Adaptive Design Approaches.
Communications and Control Engineering, Springer London, 2013.
[33] Y. Su and J. Huang, "Stability of a class of linear switching systems
with applications to two consensus problems," IEEE Transactions on
Automatic Control, vol. 57, pp. 1420 -- 1430, June 2012.
[34] H. Zhang, F. L. Lewis, and A. Das, "Optimal design for synchronization
of cooperative systems: State feedback, observer and output feedback,"
IEEE Transactions on Automatic Control, vol. 56, pp. 1948 -- 1952, Aug
2011.
[35] P. Wieland, R. Sepulchre, and F. Allgower, "An internal model principle
is necessary and sufficient for linear output synchronization," Automat-
ica, vol. 47, no. 5, pp. 1068 -- 1074, 2011.
[36] J. Lunze, "Synchronization of heterogeneous agents," IEEE Transactions
on Automatic Control, vol. 57, pp. 2885 -- 2890, Nov 2012.
[37] S. Kullback and R. A. Leibler, "On information and sufficiency," The
Annals of Mathematical Statistics, vol. 22, no. 1, pp. 79 -- 86, 1951.
[38] M. Basseville and I. V. Nikiforov, Detection of Abrupt Changes: Theory
and Application, vol. 104. Prentice Hall Englewood Cliffs, 1993.
[39] S. Kotz and N. L. Johnson, Process Capability Indices. New York, USA:
Chapman & Hall/CRC, 1993.
[40] T. M. Cover and J. A. Thomas, Elements of Information Theory. New
York, NY, USA: Wiley-Interscience, 2006.
[41] H. J. LeBlanc, H. Zhang, X. Koutsoukos, and S. Sundaram, "Resilient
asymptotic consensus in robust networks," IEEE Journal on Selected
Areas in Communications, vol. 31, pp. 766 -- 781, April 2013.
[42] W. Ren and R. W. Beard, "Consensus seeking in multiagent systems
under dynamically changing interaction topologies," IEEE Transactions
on Automatic Control, vol. 50, pp. 655 -- 661, May 2005.
14
Aquib Mustafa (S17) received the B. Tech. degree
from the Aligarh Muslim University, Aligarh, India,
in 2013, and the Masters degree from the Indian
Institute of Technology Kanpur, Kanpur, India, in
2016. He is currently pursuing the Ph.D. degree in
the Department of Mechanical Engineering, Michi-
gan State University, East Lansing, USA. His pri-
mary research interests include Resilient control,
Multi-agent systems, and sensor networks.
Rohollah Moghadam (S17) received the M.S. de-
gree from the Shahrood University of Technology,
Shahrood, Iran, in 2007, in electrical engineering.
He was a visiting scholar with the University of
Texas at Arlington Research Institute, Fort Worth,
TX, USA in 2016. He is currently pursuing the Ph.D.
degree in the Department of Electrical and Computer
Engineering, Missouri University of Science and
Technology, Rolla, USA. His current research inter-
ests include cyber-physical systems, reinforcement
learning, neural network, network control systems,
and distributed control of multi-agent systems.
Hamidreza Modares (M15) received the B.Sc. de-
gree from Tehran University, Tehran, Iran, in 2004,
the M.Sc. degree from the Shahrood University of
Technology, Shahrood, Iran, in 2006, and the Ph.D.
degree from the University of Texas at Arlington
(UTA), Arlington, TX, USA, in 2015. From 2006 to
2009, he was with the Shahrood University of Tech-
nology as a Senior Lecturer. From 2015 to 2016, he
was a Faculty Research Associate with UTA. From
2016 to 2018, he was an Assistant Professor with the
Department of Electrical and Computer Engineering,
Missouri University of Science and Technology, Rolla, USA.
He is currently an Assistant Professor with the Department of Mechanical
Engineering, Michigan State University, East Lansing, USA. He has authored
several journal and conference papers on the design of optimal controllers
using reinforcement learning. His current research interests include cyber-
physical systems, machine learning, distributed control, robotics, and renew-
able energy microgrids.
Dr. Modares was a recipient of the Best Paper Award from the 2015
IEEE International Symposium on Resilient Control Systems, the Stelmakh
Outstanding Student Research Award from the Department of Electrical
Engineering, UTA, in 2015, and the Summer Dissertation Fellowship from
UTA, in 2015. He is an Associate Editor of the IEEE TRANSACTIONS ON
NEURAL NETWORKS AND LEARNING SYSTEMS.
|
1410.6516 | 2 | 1410 | 2018-08-23T11:53:11 | Coalition Structure Generation on Graphs | [
"cs.MA"
] | Two fundamental algorithm-design paradigms are Tree Search and Dynamic Programming. The techniques used therein have been shown to complement one another when solving the complete set partitioning problem, also known as the coalition structure generation problem [5]. Inspired by this observation, we develop in this paper an algorithm to solve the coalition structure generation problem on graphs, where the goal is to identifying an optimal partition of a graph into connected subgraphs. More specifically, we develop a new depth-first search algorithm, and combine it with an existing dynamic programming algorithm due to Vinyals et al. [9]. The resulting hybrid algorithm is empirically shown to significantly outperform both its constituent parts when the subset-evaluation function happens to have certain intuitive properties. | cs.MA | cs |
Coalition Structure Generation on Graphs
Talal Rahwana, Tomasz Michalakb
aMasdar Institute of Science and Technology, UAE
bUniversity of Oxford, UK, and University of Warsaw, Poland
Abstract
Two fundamental algorithm-design paradigms are Tree Search and Dynamic Pro-
gramming. The techniques used therein have been shown to complement one another
when solving the complete set partitioning problem, also known as the coalition
structure generation problem [5]. Inspired by this observation, we develop in this
paper an algorithm to solve the coalition structure generation problem on graphs,
where the goal is to identifying an optimal partition of a graph into connected sub-
graphs. More specifically, we develop a new depth-first search algorithm, and com-
bine it with an existing dynamic programming algorithm due to Vinyals et al. [10].
The resulting hybrid algorithm is empirically shown to significantly outperform both
its constituent parts when the subset-evaluation function happens to have certain
intuitive properties.
1. Introduction
The coalition structure generation problem is a fundamental problem in multi-agent
systems research. It involves partitioning the set of agents into mutually disjoint
coalitions so that the total reward from the resulting coalitions is maximized. Most
of the literature on this topic assumes that the agents can split into teams (or
coalitions) in any way they like [8]. In practice, however, some coalition structures
may be inadmissible due to various constraints present in the problem domain.
This paper considers one such class of problems, known as graph-restricted games
[3]. Here, we are given a graph in which every node represents an agent, and every
edge can be interpreted as a communication channel, or a trust relationship, which
facilitates the cooperation between its two ends. A coalition is then feasible if and
only if it induces a connected subgraph of G. The intuition here is that any two
agents cannot belong to the same coalition unless they are able to communicate
with one another, either directly through an edge, or indirectly through (some of)
the other members of the coalition who collectively form a path between the two
agents. Following convention, we will assume that G is connected.1
1If G is not connected, the coalition structure generation problem can be decomposed into smaller
Preprint submitted to Artificial Intelligence
July 18, 2021
To the best of our knowledge, the two state-of-the-art algorithms for solving
the coalition structure generation problem in graph-restricted games are: (i) a tree-
search algorithm called CFSS [1], and (ii) a dynamic-programming algorithm called
DyPE [10]. Each algorithm has its relative strengths and weakness compared to the
other. In particular,
• CFSS is superior in that it is an anytime algorithm -- its solution quality im-
proves monotonically as computation time increases. As such, it can return a
valid solution even if it was unable to run to completion, e.g., due to a fail-
ure or due to time constraints. DyPE , on the other hand, is not an anytime
algorithm, and so does not return interim solutions. Another advantage of
CFSS is that it applies a branch-and-bound technique, which enables it to
exploit the specifics of any given problem instance, resulting in (possibly sig-
nificant) speedups. This is not possible with DyPE due to the absence of any
such a branch-and-bound technique. Finally, CFSS uses very little memory
compared to DyPE ; the latter requires storing in memory the solutions of dif-
ferent sub-problems, the number of which may be exponential (depending on
the density of the graph).
• On the other hand, DyPE is superior in terms of the computational complexity.
For instance, given a complete graph of n agents, DyPE runs in O(3n), while
CFSS runs in O(nn). This is because the latter depends heavily on the branch-
and-bound technique which, in the worst case, may fail to prune even a single
solution, resulting in a brute-force search.
Since DyPE and CFSS have their own strengths and weaknesses relative to each
other, it is desirable to develop an algorithm that has the best of both. Perhaps a
promising direction is to combine DyPE with CFSS , following the steps of Rahwan
et al. [6], who handled the general (not graph-restricted) coalition structure genera-
tion problem by combining a dynamic programming algorithm [4] with a depth-first
search algorithm [9], thus obtaining the best of both. In our case, however, CFSS
and DyPE are built on entirely different search-space representations (see Section 3),
which makes it hard to combine the two algorithms elegantly and efficiently.
With this in mind, the contribution of this paper are as follows:
• We develop TSP -- a new depth-first search algorithm specifically designed to
be compatible with DyPE .
• We show how to modify both TSP and DyPE such that they complement one
another when merged into a single hybrid algorithm, called D-TSP .
independent sub-problems, each having a connected graph.
2
• We empirically evaluate D-TSP in randomly-generated super-subadditive set-
tings, and show that it significantly outperforms its constituent parts.
The remainder of the paper is structured as follows. The main notation is in-
troduced in Section 2. The existing dynamic-programming algorithm -- DyPE -- and
the existing tree-search algorithm -- CFSS -- are described in Section 3. Our new
tree-search algorithm -- TSP -- is introduced in Section 4. The hybrid algorithm --
D-TSP -- is introduced in Section 5. Finally, Section 7 concludes the paper and
discusses future directions.
2. Preliminaries
A graph-restricted game is a tuple, (A, v, G), where A is the set of agents, v : 2A → R
is a characteristic function that evaluates each coalition of agents, and G = (A, E)
is a graph whose set of nodes is A, and whose set of edges, E, specifies which agents
are connected to each other. The number of agents in A will be denoted by n.
For every coalition, C ⊆ A, let P C denote the set of partitions of C, also known
as coalition structures over C.2 Given our focus on characteristic function games,
the value of a partition is simply the sum of the values of the coalitions therein.3
More formally, for every C ⊆ A, and every P ∈ P C, the value of partition P is:
(cid:88)
p∈P
V (P ) =
v(p).
We will denote by opt(C) an optimal partition of C, and by v∗(C) the value of such
a partition. More formally,
opt(C) ∈ arg max
P∈P C
V (P )
and
v∗(C) = max
P∈P C
V (P ).
In a graph-restricted game, (A, v, G), we say that a coalition, C ⊆ A, is connected in
G if and only if C induces a connected subgraph of G. Moreover, for every C ⊆ A,
we will denote by ConnectedSubsets(C, G) the set of all non-empty subsets of C that
are each connected in G. Similarly, we will denote by P C
G the set of every partition
G = {P ∈ P C :
in P C whose coalitions are connected in G. More formally, P C
P ∩ ConnectedSubsets(C, G) = P}. Then, given (A, v, G), the coalition structure
generation problem is to find an optimal partition of A, defined as follows:
CS∗ ∈ arg max
P∈P A
G
V (P ).
2The terms "partition" and "coalition structure" will be used interchangeably throughout the
paper, as common practice in the literature.
3The coalition structure generation problem has also been studied in games where partitions are
evaluated differently, e.g., due to the presence of externalities [2, 7]. However, these are out of the
scope of this paper, and are the focus of future work.
3
Next, we define weakly super-subadditive games. To this end, recall that a game
(A, v) is weakly superadditive if: v(C ∪ C(cid:48)) ≥ v(C) + v(C(cid:48)) for any two disjoint
coalitions C, C(cid:48) (i.e., merging any two coalitions is never harmful). Conversely, a
game (A, v) is weakly subadditive if: v(C ∪ C(cid:48)) ≤ v(C) + v(C(cid:48)) for any two disjoint
coalitions C, C(cid:48) (i.e., merging any two coalitions is never beneficial). Finally, recall
that a game (A, v) is the sum of two games, (A, v1) and (A, v2), if v(C) = v1(C) +
v2(C) for all C ⊆ A. In this case, we write: (A, v) = (A, v2) + (A, v2). Now, we are
ready to introduce the following definition.
Definition 1. A game (A, v) is (weakly) super-subadditive if it is the sum of two
games: a (weakly) superadditive game, denoted by (A, vsup), and a (weakly) subad-
ditive game, denoted by (A, vsub).
The intuition is that (A, vsup) represents the rewards from cooperation, which
is assumed to increase (weakly) with the size of the coalition. On the other hand,
(A, vsub) represents the coordination costs, which are also assumed to increase (weakly)
with the size of the coalition.
3. Related Work
This section is divided into two subsections. The first describes the CFSS algorithm
of Bistaffa et al. [1], while the second describes the DyPE algorithm of Vinyals et al.
[10].
3.1. The CFSS Algorithm
Bistaffa et al. [1] proposed the CFSS algorithm. It is based on edge contraction -- a
basic operation in graph theory which involves: (i) removing an edge from a graph,
and (ii) merging the two nodes that were previously joined by that edge. In our
context of graph-restricted games, since every node represents an agent (i.e., a sin-
gleton coalition), "merging the two nodes" corresponds to merging the two coalitions
that were represented by those nodes. An example is illustrated in Figure 1(A).
Taking the entire graph into consideration, the contraction of an edge can be
interpreted as a transition from one coalition structure to another. For instance,
the contraction of the edge ({a1},{a3}) in Figure 1(B) corresponds to the transition
from {{a1},{a2},{a3},{a4},{a5}} to {{a1, a3},{a2},{a4},{a5}}. Based on this ob-
servation, the algorithm repeats the process of contracting different edges, in order
to eventually visit all coalition structures. During this process, to ensure that each
coalition structure is visited no more than once, the algorithm marks all previously-
contracted edges to avoid contracting them again in the future. In Figure 1, the
marked edges are illustrated as dashed lines. Here, it is important to note that
the contraction of an edge may result in merging other edges. In Figure 1(B) for
example, contracting ({a1},{a3}) results in merging ({a1},{a4}) with ({a3},{a4}),
4
as well as merging ({a1},{a2}) with ({a3},{a2}). Whenever this happens, if one of
the merged edges happens to be dashed, the edge that results from the merger must
also be dashed, again see Figure 1(B). This ensures that the agents appearing at the
two ends of a dashed edge never appear together in the same coalition.
Figure 1(C) illustrates the sequence in which the algorithm visits all possible
coalition structures given the graph G = (A, E) where A = {a1, a2, a3, a4} and
E = {(a1, a2), (a1, a4), (a3, a2), (a3, a4)}. Each coalition structure is represented as
a node in the illustrated search tree, and the numbers on the edges represent the
order in which the algorithm visits the different coalition structures. Consider the
root for example:
in its second child
we make ({a1},{a2}) dashed, and contract ({a2},{a3}); in its third child we make
({a1},{a2}), ({a2},{a3}) dashed, and contract ({a3},{a4}); finally in its fourth child
we make ({a1},{a2}), ({a2},{a3}), ({a3},{a4}) dashed, and contract ({a1},{a4}).
in its first child we contract ({a1},{a2});
Figure 1: Illustration of how Bistaffa et al.'s algorithm works.
To speed up the search, a branch-and-bound technique is used whenever the
algorithm visits a node -- i.e., a partition, P -- in the search tree. The purpose of
this technique is to determine whether it is worthwhile to search T P -- the sub-tree
rooted at P . The general idea is to compute an upper bound, denoted UB (T P ), on
5
{a1}{a2}{a3}{a1,a3}{a2}{a1}{a2}{a3}{a4}{a5}{a1,a3}{a2}{a4}{a5}contract the edge ({a1},{a3})contract the edge ({a1},{a3}){a2}{a3}{a4}{a1}{a3}{a4}{a1,a2}{a4}{a1}{a2,a3}{a2}{a1}{a3,a4}{a4}{a1,a2,a3,a4}{a1,a2,a3}{a1,a2}{a3,a4}{a3}{a1,a2,a4}{a2}{a1,a3,a4}{a2,a3}{a1}{a1,a4}{a2,a3,a4}(A)(B)(C)1234569111078{a2}{a3}{a1,a4}a previously-contracted edge (it cannot be contracted again).an edge whose contraction is still a possibility. the values of all partitions in T P . Then, if this upper bound was not greater than
the value of the best partition found so far, then the algorithm avoids searching
T P . Bistaffa et al. proposed a way of computing UB (T P ) for cases where the game
under consideration is weakly super-subadditive (see Section 2 for more details). In
particular, it is possible to compute an upper bound U B(T P ) based on the following
observations:
• Every coalition structure in T P is the result of merging some (if not all) of the
coalitions in P that are connected via solid edges. Here, the only constraint
is that agents appearing at the two ends of a dashed edge must not appear
together in the same coalition.
• Merging coalitions in P can never improve solution quality in a weakly subad-
ditive game. Thus, V sub(P ) = maxP∈T P V sub(P ).
• Merging coalitions in P can never reduce solution quality in a weakly superad-
ditive game. Thus, no solution in T P can be better than the solution obtained
by: (i) removing all dashed edges, and (ii) merging all coalitions in P that
are connected via solid edges. Let us denote this solution as P merge. Then,
V sup(P merge) ≥ maxP∈T P V sup(P ).
Based on the above observations, we can establish the following upper bound on
solution quality: U B(T P ) = V sub(P )+V sup(P merge). This concludes our description
of Bistaffa et al.'s algorithm. More details can be found in [1].
3.2. The DyPE Algorithm
Vinyals et al. [10] proposed a dynamic-programming algorithm called DyPE . Be-
fore explaining how this algorithm works, let us first briefly describe how dynamic
programming works for general games, rather than graph-restricted games. Here is
the main idea: to compute an optimal partition of the set of agents, A:
• First, compute an optimal partition of each strict subset of A.
• After that, examine all the possible ways of splitting A into two halves, and
replace one of the halves with its optimal partition. More specifically, for every
non-empty subset S ⊆ A : S (cid:54)= ∅, split A into two halves, S and A \ S, and
replace A\ S with opt(A\ S). Clearly, the union {S}∪ opt(A\ S) is a partition
of A, and the value of this union is v(C) + v∗(A \ S). Furthermore, the best
such union (i.e., the one with the largest value) is an optimal partition of A.
Importantly, the above process can be carried out recursively, as captured by the
following formula:
(cid:0)v(S) + v∗(C \ S)(cid:1)
(1)
v∗(C) = max
S⊆C:S(cid:54)=∅
6
Having described a general dynamic programming formula, we now explain how
DyPE speeds up this formula when the game is restricted by a graph. The main
idea is to use a pseudotree. Basically, given a graph G = (A, E), the pseudotree of G,
denoted by PT G, is a rooted tree such that: (i) the set of nodes of PT G is the set of
agents, and (ii) any two agents who share an edge in G appear on the same branch
in PT G (an example is illustrated in Figure 2). Let us now explain how DyPE uses
the pseudotree to speed up the formula. To this end, let bi denote the agent at the
ith position of the breadth-first order of nodes in PT G. In Figure 2(B) for example,
that order is: (a3, a1, a4, a2, a5), and so b1 = a3 while b4 = a2. Now, the broad idea
behind DyPE is to start with the last agent in the breadth-first order, bn, and then
move to bn−1, then bn−2 and so on until it reaches b1. Let bDyPE denote the agent
at which DyPE has reached in the breadth-first order at any point in time during
execution. Then, for each bDyPE , the algorithm solves the following sub-problems:
(cid:27)
(2)
(cid:26)
(C, v, G)(cid:12)(cid:12) (C∈ConnectedSubsets({bDyPE ,··· ,bn},G))
∧(bDyPE∈C)
∧(A\C∈ConnectedSubsets(A,G))
The pseudo code of DyPE is shown in Algorithm 1. For a proof of the correctness
of this algorithm, see [10].
Figure 2: A sample graph G and its corresponding pseudotree PT G.
4. Our Tree-Search Algorithm -- TSP
As mentioned earlier in the introduction, Rahwan et al. [6] developed an algorithm
for general coalition structure generation problems, which combined a tree-search
algorithm with a dynamic-programming algorithm, resulting in a combination that
is superior to both its constituent parts. So why not develop a similar hybrid algo-
rithm for graph-restricted games? Perhaps the most natural starting point would
be to try and combine CFSS -- an existing tree-search algorithm -- with DyPE -- an
existing dynamic-programming algorithm. Unfortunately, however, as we have seen
in the above section, both algorithms are based on very different ideas; one is based
on edge contraction, while the other is based on a pseudo tree. As such, the two seem
incompatible, or at least hard to combine smoothly. With this in mind, our goal in
this section is to develop a tree-search algorithm that can be combined with DyPE .
7
a1a3a2(A)(B)a4a5a5a3a2a1a4graph 𝐺pseudotree 𝑃𝑇𝐺 Algorithm 1: DyPE (A, v, G, PT G).
Input: A graph-restricted game (A, v, G), and a pseudotree, PT G.
Output: An optimal coalition structure over A.
1 for bDyPE = bn to b2 do
// iterate over all sub-problems in (2):
foreach C ∈ ConnectedSubsets({bDyPE ,··· , bn}, G) such that
(bDyPE ∈ C) ∧ (A \ C ∈ ConnectedSubsets(A, G)) do
// Compute v∗(C) and bestSubset(C) (in lines 3 to 8):
v∗(C) ← −∞;
foreach S ∈ ConnectedSubsets(C, G) : bDyPE ∈ S do // iterate over
T∈connectedComponents(C\S) v∗(T ); // Compute the
value ← v(S) +(cid:80)
all non-empty subsets of C that are each connected in G and contain bDyPE .
value of {S} ∪ opt(C \ S), i.e., compute v(S) + v∗(C \ S).
if v∗(C) < value then
v∗(C) ← value;
bestSubset(C) ← S;
// Compute v∗(A) and bestSubset(A) (in lines 9 to 14):
9 v∗(A) ← −∞;
10 foreach S ∈ ConnectedSubsets(A, G) : b1 ∈ S do // iterate over all non-empty
subsets of A that are each connected in G and contain b1.
T∈connectedComponents(A\S) v∗(T );
value ← v(S) +(cid:80)
if v∗(A) < value then
v∗(A) ← value;
bestSubset(A) ← S;
2
3
4
5
6
7
8
11
12
13
14
17
// Compute an optimal coalition structure over A (in lines 15 to 17):
15 opt(A) ← {A};
16 while ∃C ∈ opt(A) : C (cid:54)= bestSubset(C) do
replace every C ∈ opt(A) with bestSubset(C), C \ bestSubset(C);
18 return opt(A);
8
We build our algorithm around the pseudotree representation used by DyPE ; the
hypothesis here is that if the two algorithms were built around the same represen-
tation, it should be possible to combine the two smoothly and effectively. Based on
this, we call our algorithm TSP , where TS stands for Tree-Search, and P stands for
Pseudotree.
The pseudo code of TSP can be found in Algorithm 2.
In more detail, the
algorithm takes as input a graph-restricted game, (A, v, G), and a pseudotree PT G.
First, in lines 1 to 4, it initializes CS† -- the current best solution -- to either be equal
to {A} or {{a1}, . . . ,{an}}, whichever has higher value. After that, in line 5, it
uses the parameter bTSP to iterate over the agents in a breadth-first order in PT G,
starting with b2, and ending with bn.4 Let us denote by bTSP−1 the agents who
is just before bTSP in the breadth-first order. Now, for every bTSP , the algorithm
enumerates all the coalitions that are each connected in G, and contain every agent
in {b1,··· , bTSP−1} but do not contain bTSP (line 6). For every such coalition, C,
the algorithm sets the current partition, P †, to be equal to {C} (line 7). Finally, in
†
G, P †, CS†) to search through the coalition
line 8, it used the function search(PT
structures that are supersets of P †, i.e., the coalition structure that contain C.
Basically, this recursive function generates different partitions while trying to avoid
the unpromising ones using a branch-and-bound technique. Next, we explain how
this function works.
†
G, P †, CS†) is given in Algorithm 3. Here, C†
denotes the agents that are not in P † (line 1 of Algorithm 3), while a† denotes
the first agent in the breadth-first order who is not in P † (line 2). Then, out of
all the connected coalitions that can be added to P †, the algorithm always starts
by adding to P † a coalition containing a† (lines 3 and 4). Now if the new P † is
a coalition structure over A, then the algorithm updates CS† -- the best solution
found so far (lines 5 to 7). Otherwise, it computes an upper bound on the value
of every partition of A that is a superset of the new P † (line 9). Based on this
upper bound, the algorithm determines whether it is worthwhile to consider adding
more coalitions to the new P † (line 10). If so, then it makes a recursive call with
the new P † (line 11). The function computeUpperBound(P †) -- which computes
the aforementioned upper bound -- can be specified based on any additional domain
knowledge. For instance, if the game is known to be super-subadditive, then this
function may return: V (P †) + vsup(C†) +(cid:80)
The pseudo code of search(PT
ai∈C† vsub({ai}).
5. Our Hybrid Algorithm -- D-TSP
In this section, we present D-TSP -- a hybrid algorithm that combines DyPE with
TSP in a way that obtains the best features of both. First, we introduce the neces-
4See Section 3.2 for more details on the breadth-first order of agents in PT G.
9
Algorithm 2: TSP (A, v, G, PT G).
Input: A graph-restricted game (A, v, G), and a pseudotree PT G.
Output: An optimal coalition structure over A.
// initialize CS†---the current best solution (lines 1 to 4).
1 if V ({A}) > V ({{a1}, . . . ,{an}}) then
CS† ← {A};
2
3 else
4
CS† ← {{a1}, . . .{an}};
// Search through different coalition structures (lines 5 to 8).
5 for bTSP = b2 to bn do
// iterate over all non-empty subsets of A that are each connected in G and do not
contain bTSP but contain every agent before bTSP in the breadth-first order.
foreach
C ∈ ConnectedSubsets(A, G) : {b1,··· , bTSP−1} ⊆ C ⊆ A \ {bTSP} do
†
G, P †, CS†); // updated CS† by searching through the
P † ← {C}; // initialize P †---the current partition.
CS† ← search(PT
partitions of A that are supersets of P †.
6
7
8
9 return CS†;
sary modifications of each algorithm (Subsections 5.1 and 5.2), and then show how
to combine the modified versions (Subsection 5.3).
5.1. DyPE∗ -- a Modified Version of DyPE
In this subsection, we modify DyPE such that it becomes an anytime algorihtm,
i.e., it does not only return a solution after termination, but also returns interim
solutions during execution. This clearly adds more resilience against failure. For
example, if the algorithm runs out of memory during execution, then instead of
wasting all the effort that the algorithm has put before the failure, it would at least
return a valid solution using all the sub-problems that it has already solved.
Before introducing our modifications, let us first revisit DyPE and analyze the
way it works. Looking at Algorithm 1, one can see that DyPE ultimately boils down
to the following main steps:
• Step 1: with bDyPE running from bn to b2, solve the following sub-problems:
(C, v, G) such that C ⊆ {bDyPE , . . . , bn} and bDyPE ∈ C and A\ C is connected
(lines 1 to 8 of Algorithm 1).
• Step 2: for each subset S ∈ ConnectedSubsets(A, G) : b1 ∈ S, compute the
value of the best coalition structure containing S (lines 10 and 11 of Algo-
rithm 1).
10
Algorithm 3: search(PT G, P †, CS†) -- a function used in TSP .
Input: PT G -- the pseudotree, P † -- the current partition (which does not yet
Output: The best partition of A that is a superset of P †.
1 C† ← A \(cid:83) P †;// i.e., C† consists of all agents not in P †.
contain all agents in A), and CS† -- the current best solution.
2 a† ← bi : (bi ∈ C†) ∧ ({b1,··· , bi−1} ∩ C† = ∅);// i.e., a† is the first agent in
3 foreach C ∈ ConnectedSubsets(C†, G) : a† ∈ C do // iterate over all non-empty
the breadth-first order who is not in P †.
subsets of C† that are each connected in G and contain a†.
P † ← P † ∪ {C}; C† ← A \ C; // add C to P † and remove it from C†.
if C† = ∅ then // if P † is a coalition structure over A.
if V (CS†) < V (P †) then
CS† ← P †;
4
5
6
7
8
9
10
11
12
else
UB ← computeUpperBound(P †) // compute an upper bound on the
value of every partition of A that is a superset of P †
if V (CS†) < UB then // apply the branch-and-bound technique.
search(PT
†
G, P †, CS†); // recursive call.
P † ← P † \ {C}; C† ← A ∪ C; // remove C from P † and add it to C†.
13 return CS†;
The problem with the above process is that DyPE does not examine a single
coalition structure over A until it has finished Step 1 -- a step which involves solving
sub-problems the number of which may be exponential (depending on the topology
of the graph). Let us now consider a sample subset that the algorithm encounters
during Step 2, given a problem of 7 agents. Let this subset be S = {b1, b2, b3, b6}.
When the algorithm encounters this particular S, it will compute the value of the
best coalition structure containing {b1, b2, b3, b6}, using the already-computed solu-
tions to the following sub-problems: (T, v, G) where T is a connected component in
the sub-graph induced by {b4, b5, b7} (see line 11 of Algorithm 1). Our critical obser-
vation is that the solutions to the aforementioned sub-problems were all computed
when DyPE finished dealing with bDyPE = b4 in Step 1. More specifically, at that
moment, DyPE has already solved the following sub-problems:
• (C, v, G) where A \ C is connected and b7 ∈ C and C ⊆ {b7};
• (C, v, G) where A \ C is connected and b6 ∈ C and C ⊆ {b6, b7};
• (C, v, G) where A \ C is connected and b5 ∈ C and C ⊆ {b5, b6, b7};
11
Figure 3: The illustration.
• (C, v, G) where A \ C is connected and b4 ∈ C and C ⊆ {b4, b5, b6, b7}.
The above sub-problems surely include every (T, v, G) where T is a connected com-
ponent in the sub-graph induced by {b4, b5, b7}. In other words, after solving the
above sub-problems, DyPE had all the information needed to compute the value of
the best coalition structure containing {b1, b2, b3, b6}. This suggests that DyPE can
be modified such that it examines certain coalition structures during Step 1, not
after Step 1.
Based on the above observation, we modify DyPE such that, instead of following
the above two steps, it follows txhis one:
• Step 1: with bDyPE running from bn to b2:
12
(A)(B)𝑃𝑇𝐺𝑎9𝑎1𝑎8𝑎5𝑎10𝑎7𝑎14𝑎6𝑎4𝑎15𝑎2𝑎3𝑎12𝑎13𝑃𝑇𝐺ϯ. Here, 𝑃ϯis still empty𝑎11𝑎9,𝑎10,𝑎11𝑎1𝑎2,𝑎3𝑎12,𝑎13(C)The remaining agents, after adding {𝑎1,𝑎3,𝑎4,𝑎8}to 𝑃ϯ. Now, Nϯ=𝑎2𝑎14𝑎15𝑎2𝑎14𝑎15The remaining agents after adding {𝑎2,𝑎5,𝑎6,𝑎7}to 𝑃ϯ. Now, Nϯ={𝑎9,𝑎10,𝑎11}(D)𝑎8𝑎5𝑎7𝑎14𝑎6𝑎4𝑎15𝑎5𝑎7𝑎6𝑎9,𝑎10,𝑎11𝑎12,𝑎13𝑎12,𝑎13𝑎9,𝑎10,𝑎11after adding {a9,a10}to 𝑃ϯ,we have Nϯ={𝑎10}(E)𝑎14𝑎15𝑎12,𝑎13𝑎10after adding {a10}to 𝑃ϯ, we have Nϯ= {𝑎12,𝑎13}(F)𝑎14𝑎15𝑎12,𝑎13after adding {a12,a15}to 𝑃ϯ,Nϯ={𝑎13}(G)𝑎14𝑎13after adding {a13,a14}to 𝑃ϯ,the function ''search'' would have generated the partition 𝑃(H) -- Step 1.1: solve the following sub-problems: (C, v, G) such that C ⊆
{bDyPE , . . . , bn} and bDyPE ∈ C and A \ C is connected.
-- Step 1.2: for each subset S ∈ ConnectedSubsets(A, G) such that bDyPE /∈
S and {b1, . . . , bDyPE−1} ⊆ S, compute the value of the best coalition
structure containing S.
One can easily see that, with the above steps, the algorithm will never consider the
same S more than once. Moreover, whenever a certain S is encountered, all relevant
sub-problems of A \ S have already been solved, including every (T, v, G) where T
is a connected component in the sub-graph induced by A \ S.
We call the modified version DyPE∗. The pseudo code is provided in Algo-
rithm 4. As can be seen, this an anytime algorithm, unlike DyPE .
5.2. TSP∗ -- a Modified Version of TSP
Our goal in this subsection is to modify TSP such that it can take advantage of any
solutions to sub-problems that were already computed by DyPE∗. To this end, let
us first analyze how TSP works. Looking at Algorithm 2, one can see that TSP
ultimately boils down to the following main steps:
• Step 1: with bTSP running from b2 to bn, set the current partition P † to be
equal to some {C}, where C is a connected coalition that does not contain
bTSP , but contains all of: b1, . . . , bTSP−1 (lines 5 to 7 of Algorithm 2).5
-- Step 1.1: keep adding different coalitions to P †, thus obtaining different
coalition structures over A (line 8 of Algorithm 2). Any coalition added
to P † must contain a† -- the first agent in the breadth-first order who is
not already in P † (see line 2 of Algorithm 3). Every time a new coalition
is added to P †, a branch-and-bound technique is used to check whether
the coalitions that are in P † are promising (lines 9 to 11 of Algorithm 3).
During the above process, for any given P †, the algorithm will try all possible coali-
tion structures that are supersets of P †, except those that are deemed unpromising
by the branch-and-bound technique. In other words, it will try adding to P † every
promising partition of A \(cid:83) P †. Importantly, however, if we were to run DyPE∗ in
partition of A \(cid:83) P † easily using the partial results of the former. This is based on
• A \(cid:83) P † ⊆ {a†, . . . , bn}. This is simply because a† is by definition the first
parallel with TSP , then the latter algorithm may be able to construct an optimal
the following two observations:
agent in the breadth-first order who is not in P †.
5Recall that bTSP−1 denotes the agent just before bTSP in the breadth-first order of agents in
the pseudo tree PT G.
13
Algorithm 4: DyPE∗(A, v, G, PT G).
Input: A graph-restricted game (A, v, G), and a pseudotree, PT G.
Output: An optimal coalition structure over A.
1 CS† ← {A};// initialize CS†---the current best solution, which is needed in this
anytime version of DyPE.
2 for bDyPE = bn to b2 do
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
// iterate over all sub-problems in (2):
foreach C ∈ ConnectedSubsets({bDyPE ,··· , bn}, G) such that
(bDyPE ∈ C) ∧ (A \ C ∈ ConnectedSubsets(A, G)) do
// Compute v∗(C) and bestSubset(C) (in lines 4 to 8):
v∗(C) ← −∞;
foreach S ∈ ConnectedSubsets(C, G) : bDyPE ∈ S do // iterate over
T∈connectedComponents(C\S) v∗(T );// Compute the
value ← v(S) +(cid:80)
all non-empty subsets of C that are each connected in G and contain bDyPE .
value of {S} ∪ opt(C \ S), i.e., compute v(S) + v∗(C \ S).
if v∗(C) < value then
v∗(C) ← value;
bestSubset(C) ← S;
// Search every coalition structure containing a connected coalition whose members
include b1, · · · , bDyPE−1, but not bDyPE :
foreach
S ∈ ConnectedSubsets(A, G) : {b1,··· , bDyPE−1} ⊆ S ⊆ A \ {bDyPE} do
value ← v(S) +(cid:80)
T∈connectedComponents(A\S) v∗(T );
if v∗(A) < value then
v∗(A) ← value;
bestSubset(A) ← S;
CS† ← {S} ∪ connectedComponents(A \ S);
while ∃C ∈ CS† : C (cid:54)= bestSubset(C) do
replace every C ∈ CS† with bestSubset(C) and
C \ bestSubset(C);
18 return CS†;
14
• if the current bDyPE happens to be before a† in the breadth-first order, then
DyPE∗ has already computed all relevant sub-problems (C, v, G) such that
C ⊆ {a†, . . . , bn} (see Sectino 5.1 for more details).
Based on the above observations, we propose a modified version of TSP , called
TSP∗, which works as follows. Whenever bDyPE happens to be before a† in the
breadth-first order, TSP∗ does not try the different partitions of A \(cid:83) P †, but
(cid:88)
instead computes the value of an optimal such partition as follows:
v∗(T ).
P †) =
T∈connectedComponents(A\(cid:83) P )
V ∗(A \(cid:91)
Now, if V (P ) + V ∗(A \(cid:83) P †) happens to be greater than V (CS†) -- the value of the
(cid:83) P †) because it is better than CS†. This computation can be done as follows. First,
the algorithm sets CS† to be equal to P †∪ connectedComponents(A\(cid:83) P ), and then
current best solution, then TSP∗ needs to compute a coalition structure P †∪opt(A\
iteratively replaces every C ∈ CS† with bestSubset(C) and A \ bestSubset(C). This
is done until C = bestSubset(C) for all C ∈ CS†.
5.3. Combining DyPE∗ with TSP∗
In this subsection, we introduce D-TSP , an algorithm that runs both DyPE∗ and
TSP∗ in parallel, such that they aid each other during the search. Basically, D-TSP
is based on the following observations:
• DyPE∗ solves sub-problems in the following sequence (see Sectino 5.1 for more
details). With bDyPE running from bn to b2
-- it solves the sub-problems: (C, v, G) such that C ⊆ {bDyPE , . . . , bn} and
bDyPE ∈ C and A \ C is connected (see lines 2 and 3 of Algorithm 4).
-- it searches all coalition structures containing a connected coalition C
where:
{b1, . . . , bDyPE−1} ⊆ C ⊆ A \ {bDyPE}.
• We deliberately designed TSP∗ such that it searches coalition structures in the
following sequence. With bTSP running from b2 to bn, it searches all coalition
structures containing a connected coalition C where:
{b1,··· , bTSP−1} ⊆ C ⊆ A \ {bTSP}
.
Note that bDyPE runs from bn to b2, while bTSP runs from b2 to bn. Thus, based
on the above observations, when the position of bDyPE becomes smaller than that
of bTSP , the algorithms DyPE∗ and TSP∗ would have jointly searched the entire
space, at which case D-TSP terminates.
15
Figure 4: Very preliminary simulation results. The number of agents runs from 10 to 22 (x-axis.
6. Performance Evaluation
Figure 4 presents very preliminary simulation results.
7. Conclusions and Future Work
Our aim was to develop a coalition structure generation problem for graph-restricted
games. Our inspiration came from an algorithm for general coalition structure gen-
eration problems, which combined a dynamic-programming algorithm with a tree-
search algorithm, resulting in a combination that is superior to both its constituent
parts [6]. Following these guidelines, we developed a tree-search algorithm, called
TSP to be compatible with an existing dynamic-programming algorithm, called
DyPE [10]. After that, we showed how to modify the two algorithms such that
they are compatible with each other. Specifically, we modified DyPE to make an
anytime algorithm that returns interim solutions, and modified TSP such that it
solutions to sub-problems that were computed by DyPE at any point in time. Af-
ter that, we showed that the modified version of DyPE gradually covers the search
space from a certain direction, while the modified version of TSP gradually covers
the search space from the opposite direction; the two algorithms terminate when
they meet each other somewhere in the middle. This way, the portion searched by
each algorithm will naturally reflect its relative strength on the problem instance at
hand. Our future work involves evaluating D-TSP empirically on a wider range of
graph-restricted games.
16
Acknowledgements
This work was supported by the Polish National Science Centre grant number
2014/13/B/ST6/01807. Tomasz P. Michalak was also supported by the European
Research Council under Advanced Grant 291528 ("RACE").
References
[1] Bistaffa, F., Farinelli, A., Cerquides, J., Rodr´ıguez-Aguilar, J., and Ram-
churn, S. D. 2014. Anytime coalition structure generation on synergy graphs. In Pro-
ceedings of the 13th International Conference on Autonomous Agents and Multi-agent
Systems. AAMAS '14. 13 -- 20.
[2] Michalak, T., Rahwan, T., Sroka, J., Dowell, A., Wooldridge, M., McBur-
ney, P., and Jennings, N. R. 2009. On representing coalitional games with externali-
ties. In ACM EC '09: Tenth ACM Conference on Electronic Commerce. 11 -- 20.
[3] Myerson, R. 1977. Graphs and cooperation in games. Mathematics of Operations
Research 2, 3, 225 -- 229.
[4] Rahwan, T. and Jennings, N. R. 2008. An improved dynamic programming algo-
rithm for coalition structure generation. In AAMAS'08: Seventh International Conference
on Autonomous Agents and Multi-Agent Systems. 1417 -- 1420.
[5] Rahwan, T., Michalak, T., Elkind, E., Wooldridge, M., and Jennings, N. R.
2014. An exact algorithm for coalition structure generation and complete set partitioning.
http://www.cs.ox.ac.uk/publications/publication6962-abstract.html .
[6] Rahwan, T., Michalak, T., and Jennings, N. R. 2012. A hybrid algorithm for
coalition structure generation. In Proceedings of the 26th AAAI Conference on Artificial
Intelligence (AAAI-2012).
[7] Rahwan, T., Michalak, T., Wooldridge, M., and Jennings, N. R. 2012. Anytime
coalition structure generation in multi-agent systems with positive or negative externali-
ties. Artificial Intelligence 186, 0, 95 -- 122.
[8] Rahwan, T., Michalak, T., Wooldridge, M., and Jennings, N. R. 2015. Coali-
tion structure generation: A survey. Artificial Intelligence 229, 0, 139 -- 174.
[9] Rahwan, T., Ramchurn, S. D., Giovannucci, A., and Jennings, N. R. 2009.
An anytime algorithm for optimal coalition structure generation. Journal of Artificial
Intelligence Research (JAIR) 34, 521 -- 567.
[10] Vinyals, M., Voice, T., Ramchurn, S., and Jennings, N. R. 2013. A hi-
erarchical dynamic programming algorithm for optimal coalition structure generation.
http://arxiv.org/abs/1310.6704 .
17
|
1803.04927 | 1 | 1803 | 2018-03-13T16:36:10 | An Agent-Based Simulation of Residential Location Choice of Tenants in Tehran, Iran | [
"cs.MA",
"cs.AI",
"cs.CY"
] | Residential location choice modeling is one of the substantial components of land use and transportation models. While numerous aggregated mathematical and statistical approaches have been developed to model the residence choice behavior of households, disaggregated approaches such as the agent-based modeling have shown interesting capabilities. In this article, a novel agent-based approach is developed to simulate the residential location choice of tenants in Tehran, the capital of Iran. Tenants are considered as agents who select their desired residential alternatives according to their characteristics and preferences for various criteria such as the rent, accessibility to different services and facilities, environmental pollution, and distance from their workplace and former residence. The choice set of agents is limited to their desired residential alternatives by applying a constrained NSGA-II algorithm. Then, agents compete with each other to select their final residence among their alternatives. Results of the proposed approach are validated by comparing simulated and actual residences of a sample of tenants. Results show that the proposed approach is able to accurately simulate the residence of 59.3% of tenants at the traffic analysis zone level. | cs.MA | cs | An Agent-Based Simulation of Residential Location Choice of
Tenants in Tehran, Iran
Ali Shirzadi Babakan and Abbas Alimohammadi
This is an Author's Original Manuscript (Preprint) of an Article Published by John Wiley & Sons in
Transactions in GIS, 2016, Vol. 20, No. 1, 101-125.
DOI: 10.1111/tgis.12144
To link to this article: http://onlinelibrary.wiley.com/doi/10.1111/tgis.12144/full
Abstract
Residential location choice modeling is one of the substantial components of land use and
transportation models. While numerous aggregated mathematical and statistical approaches have
been developed to model the residence choice behavior of households, disaggregated approaches
such as the agent-based modeling have shown interesting capabilities. In this paper, a novel agent-
based approach is developed to simulate the residential location choice of tenants in Tehran, the
capital of Iran. Tenants are considered as agents who select their desired residential alternatives
according to their characteristics and preferences to various criteria such as the housing rent,
accessibility to different services and facilities, environmental pollutions, and distance from their
workplace and former residence. The choice set of agents is limited to their desired residential
alternatives by applying a constrained NSGA-II algorithm. Then, agents compete with each other
to select their final residence among their alternatives. Results of the proposed approach are
validated by comparing simulated and actual residences of a sample of tenants. Results show that
the proposed approach is able to accurately simulate the residence of 59.3% of tenants at the
traffic analysis zones level.
Keywords
Residential location choice, Agent-based modeling, Multi-objective decision making, Tenant household,
Tehran.
1
1. Introduction
There are strong interactions between land use and transportation. For studying these interactions,
numerous land use and transportation models have been developed. These models have been reviewed by
many researchers (e.g. (Chang, 2006 ; Iacono et al., 2008)). One of the most important components of
land use and transportation models is residential location choice modeling (Sener et al., 2011 ; Chang &
Mackett, 2006). Many activities and urban travels of individuals are influenced by their residence.
Thereby, residential location choice process of households can directly or indirectly affect various aspects
of a city including transportation system, land uses, utilities, and socio-economic structures. Thus, this
process has received high attention from many researchers in different fields such as urban planning,
transportation, geography, and geosciences.
Residential location choice modeling has been initiated by Alonso (1960) and Lowry (1964) who applied
economic and spatial interaction principles in their models. A decade later, Lerman (1976) and Mcfadden
(1978) pioneered the use of discrete choice models in this area. In discrete choice models, still widely
used in recent researches, utilities of a finite number of alternatives are calculated and the one with the
maximum utility is selected. Within this class of models, multinomial logit (MNL) and nested logit (NL)
are the most commonly used models (Sener et al., 2011 ; Rashidi et al., 2012). However, conventional
models generally use zone-based aggregated characteristics of households and are insensitive to inherent
heterogeneities among individual households. This issue has been recognized as one of the main sources
of error in these models (Arentze et al., 2010 ; Benenson, 2004). Therefore, disaggregated models such as
the microsimulation and agent-based models have found wider application.
In this paper, a novel agent-based approach has been developed to simulate the residential location choice
of tenants in Tehran, Iran. Agent-based modeling have opened new ways to theoretically and
experimentally model complex phenomena such as the urban system (Barros, 2004). An Agent-based
model is composed of multiple interacting elements (agents) with some level of autonomy which can
perceive their environment and act to change the environment according to their desires and objectives.
These models almost have no limitations for directly representing and simulating behavior of urban
elements including individuals and households (Pagliara & Wilson, 2010). They are "bottom-up"
approaches, in which the behavior of system is emerged from the aggregation of agents' behavior. In
agent-based residential location choice modeling, households can be represented as agents who decide to
move and choose new dwellings and thereby affect behavior of other agents and urban components. In
fact, from the perspective of agent-based modeling, regional or urban patterns of residential locations are
outcomes of agents' residence choice behavior (Benenson, 2004).
There are many agent-based studies to model the housing market and residential location choice of
households (e.g. Otter et al., 2001 ; Benenson, 2004 ; Jordan et al., 2012 ; Ettema, 2011 ; Rosenfield et
2
al., 2013 ; Devisch et al., 2009 ; Haase et al., 2010 ; Gaube & Remesch, 2013 ; Jackson et al., 2008). In
addition to these studies, a new generation of comprehensive urban models such as ALBATROSS1
(Arentze & Timmermans, 2004), RAMBLAS2 (Veldhuisen et al., 2000), MUSSA3 (Martinez & Donoso,
2010 ; Martinez, 1996), ILUTE4 (Salvini & Miller, 2005 ; Miller & Salvini, 2001), UrbanSim5 (Waddell,
2002 ; Waddell et al., 2003), and ILUMASS6 (Strauch et al., 2005) have been developed using
microsimulation, cellular automata, and agent-based models. A detailed review of agent-based residential
location choice models can be found in (Huang et al., 2014).
In almost all previous agent-based models, residential location choice is based on a utility-measuring or
suitability-measuring function (Huang et al., 2014). There are rare studies (e.g. Jackson et al., 2008)
which have utilized different heuristic or non-heuristic approaches in this area. Therefore, in this study, a
novel two-step agent-based approach is proposed for residential location choice modeling which is the
main contribution of this paper. In the first step of the proposed approach, a multi-objective decision
making method, non-dominated sorting genetic algorithm II (NSGA-II), is introduced for evaluating
residential alternatives by certain criteria and restricting the choice set of agents to a finite number of their
desired residential alternatives. As far as the authors know, multi-objective decision making methods
have not been used for determining a set of residential alternatives. For this purpose, two general
approaches including consideration of all alternatives and random selection of some alternatives have
been used by researchers. However, both approaches can raise some concerns. For example, the former
unrealistically assumes that households search all alternatives, and the latter may result in inaccurate
estimations (Rashidi et al., 2012). In addition to these general approaches, some researchers such as
Rashidi et al. (2012) have proposed heuristic approaches to form smaller and more manageable choice
sets. But, these approaches also have some limitations, for example, Rashidi et al. (2012) only considered
average work distance for choice set formation, while there are several criteria such as property value and
neighborhood characteristics that clearly affect the selection of residential alternatives. However, the
proposed multi-objective decision making method in this paper allows modelers to use various criteria
and objectives for determining a set of residential alternatives which leads to more realistic results.
In the second step, a heuristic competition mechanism is suggested in which agents compete with each
other to select their final residence among their desired residential alternatives. This means that the
residence choice of each agent is influenced by choices of the other agents. This is a critical issue which is
not supported by the conventional models due to their inherent aggregated nature. As a result, in this
1 A Learning-Based Transportation Oriented Simulation System
2 Regional planning model based on the microsimulation of daily activity patterns
3 A Land Use Model for Santiago City
4 Integrated Land Use, Transportation, Environment
5 Urban Simulation model
6 Integrated Land-Use Modeling and Transportation System Simulation
3
paper, agents may not reside in their best residential option, because it may be previously occupied by
another agent looking for a residence. This is more similar to the actual process of residential choice of
households in the real world.
The main objective of this paper is the spatially explicit simulation of residential location choice of
individual tenants. The proposed approach is a useful tool for simulating the residential location choice
behavior of individual households and spatially explicit distribution of different socio-economic
categories of population. The approach then can be used by urban planners and policy makers to
investigate effects of different plans and policies on these concerns.
As a case study, the proposed approach is implemented in Tehran, the capital of Iran. Although more than
50 years have passed since the beginning of residential location choice modeling in the world, there is
little research conducted in this area in Tehran. Tehran contains a large number of tenants who change
their residence every year, but the residential location choice behavior of this population has been never
studied. Awareness of spatial distribution of different socio-economic categories of tenants and their
residential choice behavior is a basic requirement for effective urban planning. In order to address this
requirement, a practical microsimulation approach is proposed in this paper. The proposed approach can
greatly help urban planners and policy makers in Tehran to simulate different land use and transportation
scenarios and predict their impacts on residential location choice patterns of different socio-economic
groups of population.
The rest of the paper is organized as follow. A background of rental residence choice in Tehran is
presented in Section 2. Section 3 provides a detailed explanation of the proposed agent-based approach.
In section 4, the proposed approach is implemented in Tehran metropolis and results are presented.
Section 5 provides validation results of the proposed approach followed by the discussion and conclusion
in section 6.
2. Background: Rental residence choice in Tehran, Iran
Tehran, the capital of Iran, with an area of about 750 km2 is located at longitude of 51 8(cid:1) to 51 37(cid:1) and
latitude of 35 34(cid:1) to 35 50(cid:1). According to the census conducted by the Statistical Center of Iran in 2011,
Tehran's population is 7,803,883, composed of 2,245,601 households, of which about 950,000 are tenants
(Tehran Municipality, 2013b). Tehran consists of 560 traffic analysis zones (TAZs) of which 532 zones
include residential areas (Figure 1). In this paper, these zones are used as the spatial units for simulating
residential location choice of tenants.
4
Figure (1): Spatial distribution of 560 traffic analysis zones (TAZs) of Tehran
The process of renting a residence in Iran is substantially different from those of the other countries. In
Iran, landlords determine the rent of their properties by considering neighborhood characteristics and
other attributes of the property. Then, usually the first tenant who affords and likes the property, rents it
(Habibi & Ahari, 2005). Study of the process of renting a house in Tehran and consultation with real
estate agencies show that landlords usually don't like to rent their properties to singles or households with
many members. Therefore, if a number of households simultaneously ask to rent a property, couple
households have a greater chance of success. The period of renting a house usually is one year in Iran.
After one year, if the landlord and tenant don't reach an agreement for extension of the renting contract
for the next year, the tenant has to leave his residence and to look for a new residence (Habibi & Ahari,
2005). According to the available statistics, about 30% of tenants in Tehran change their residence every
year (Tehran Municipality, 2013b). This research is an attempt to simulate residential location choice
process of these households. Therefore, in this paper, it has been assumed that agents want or have to
change their residence and attempt to rent the best possible residence according to their characteristics
and preferences.
It should be noted that due to some reasons such as having low incomes, low prices of fuel and public
transit fares, high housing rents and existence of more employment opportunities in Tehran, some tenants
prefer to reside in the surrounding cities of Tehran such as Karaj, Shahriar, Robat Karim, Eslamshahr, and
Pardis, while their workplaces are located in Tehran. These cities are located in approximately short
distances, less than 40 km, from Tehran. Therefore, these tenants commute between these cities and
Tehran every day by their private car or public transit including bus, taxi and train. These tenants are not
considered in this study. However, it is assumed that tenants who cannot reside in any residential zone of
Tehran may have to move to one of these cities.
5
3. Proposed approach
In the proposed agent-based simulation of rental residence choice, tenants are simulated as agents who
look for a preferred residence. They search among residential zones and select their appropriate
residential alternatives by considering some criteria such as housing rent, environmental pollutions,
distance from their workplaces and former residence, and accessibility to various services. Finally, to
select their preferred residence, agents compete with each other. General framework of the proposed
approach is shown in Figure (2). This framework consists of three basic modules including; I) generation
of tenants (agents) using the Monte Carlo simulation, II) determination of desirable residential
alternatives of agents using NSGA-II, and III) competition between agents to determine their final
residence. These modules are briefly explained below.
Agent generation (Monte Carlo Simulation)
Number of
Members
Age of Members
Monthly Income
Required
Residential Area
Number of Cars
Number and
Workplaces of
Employees
Criteria and
Preferences
Agents
Residential Alternatives selection
Accessibility to
transportation
Environmental
pollutions
Restrictions of
the traffic plans
Housing Rent
Multi-Objective
Decision Making
(NSGA II)
Preferred
Alternatives
Accessibility to
public facilities
Distance from
the workplace
Distance from the
former residence
Competition and Residence choice
Number of
Members
Monthly Income
Age of the
Members
Agent-based Modeling
Residence
Figure (2): General framework of the proposed approach
6
3.1. Agent generation
The first module of the proposed approach is generation of tenants (agents) with required attributes for
simulating their residential location choice behavior. Demographic and socio-economic attributes of
agents are generated using the Monte Carlo simulation such that their aggregated average and standard
deviation match with those of the available aggregated data in every zone (Tehran Municipality, 2013b).
For this purpose, a sequential approach is used in which some attributes of households are simulated
based on the previously determined attributes. This approach has been applied by Miller et al. (2004) in
ILUTE model. In this approach, initially the number and age of members and monthly income of agents
are generated using the Monte Carlo simulation such that the aggregated average and standard deviation
of these attributes in each zone closely match to the available zone-based data of mean income, household
size and percentages of different age groups provided by Tehran Municipality (2013b). Then, the number
of cars and employees and the required residential area of agents are generated based on their previous
attributes using the Monte Carlo simulation. Subsequently, using the available records of employment
rate in different zones and general patterns of home-to-work travel distances (Tehran Municipality,
2013a), workplaces of employed members of agents are randomly allocated to zones. The employment
capacity (EC) of each zone is computed by Eq. (1):
(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:9)(cid:10)(cid:11) (cid:12)(cid:13)(cid:14)(cid:15) (cid:12)(cid:13)(cid:14)
(cid:16)(cid:14)(cid:17)(cid:18)
(cid:19)(cid:12)(cid:20)
where Nei is the total number of employments in zone i and Na is the total number of agents.
Finally, residential criteria and preferences of agents are generated using the Monte Carlo simulation
based on their previously simulated attributes. Depending on their demographic and socio-economic
characteristics, agents use different criteria and preferences to select their residential zone. In order to
simulate these criteria and preferences, stated preferences of 330 sample tenants with different
characteristics were surveyed by filling questionnaires. Because of the lack of suitable sample data and
limitations of collecting a large data set, a sample data consisting of 330 tenants were surveyed. Although
this sample seems relatively small to represent the whole target population, due to surveying from
different residential areas of Tehran, it covers tenants with various demographic, socio-economic and
cultural characteristics. Statistical distribution of the sample data shows that it properly represents
different categories of the target population (Table (1)). Sample tenants stated their preferences in three
linguistic levels of importance including 'very important', 'important', and 'not important' in the
questionnaires. It should be noted that tenants were quite justified to state 'important' and 'very
important' preferences only for criteria which are actually considered by them in their residential location
choice process. A summary of characteristics, preferred criteria, and percentage of tenants who stated
'very important' or 'important' preferences for each criterion is presented in Table (1).
7
Table (1): A summary of characteristics and preferred criteria of surveyed households
Attribute Category Number Percentage Average
Residential
Area (m2)
Housing
Rent
Accessibility
to
Educational
Locations
Accessibility
to shopping
Locations
Accessibility
to Green
and
Recreational
Locations
Preferred Criteria (%)
Accessibility
to Cultural
Locations
Accessibility
to health
Locations
Accessibility
to Highways
Accessibility
to Subway
Stations
Accessibility
to Bus
Stations
Air and
Noise
Pollutions
Distance
from the
Workplace
Distance
from the
Former
Residence
Without
Traffic
Restrictions
Size
Monthly
income
(million
IRR)*
Number of
Cars
Total
Single
Couple
3-4
> 4
< 7.5
7.5-30
> 30
0
1
> 1
28
97
171
34
69
204
57
34
231
65
330
8.5
29.4
51.8
10.3
20.9
61.8
17.3
10.3
70.0
19.7
100
55
74
86
97
67
81
102
71
82
98
78
100
100
100
100
100
100
100
100
100
100
100
17.9
10.3
71.3
88.2
42.0
58.3
33.3
55.9
55.8
29.2
50.6
* IRR (Iranian Rial); At the time of this study 1USD is about 36000 IRR
14.3
40.2
33.9
29.4
18.8
34.8
47.4
38.2
34.6
27.7
33.6
7.1
50.5
57.3
64.7
34.8
53.9
64.9
44.1
55.8
41.5
51.8
10.7
7.2
8.8
5.9
5.8
8.8
8.8
8.8
8.2
7.7
8.2
3.6
26.8
34.5
44.1
21.7
33.3
31.6
26.5
32.9
24.6
30.6
71.4
83.5
87.1
85.3
65.2
87.7
96.5
26.5
88.7
100
84.5
35.7
43.3
52.0
52.9
53.6
47.1
45.6
91.2
47.2
29.2
48.2
17.9
20.6
22.8
20.6
29.0
21.1
14.0
73.5
18.2
6.2
21.5
32.1
43.3
41.5
41.2
20.3
40.2
70.2
26.5
37.7
61.5
41.2
92.9
90.7
84.2
79.4
92.8
85.8
80.7
97.1
86.1
81.5
86.4
46.4
52.6
62.6
61.8
68.1
56.4
52.6
55.9
58.4
58.5
58.2
46.4
56.7
53.8
52.9
39.1
50.0
86.0
2.9
49.4
96.9
53.9
3.2. Residential alternatives selection
In this step, agents freely select their desired residential alternatives using a multi-objective decision
making method without considering the choice set of other agents and residential capacity of zones. They
may have multiple and conflicting objectives. For example, they may want to reside in a zone with low
rent and high accessibility to different services. But these objectives usually conflict with each other,
because a high accessibility is generally coincided with a higher rent. Thus, each agent attempts to meet
his objectives and accordingly selects the most suitable alternatives. In other words, each agent faces with
a multi-objective decision making problem. In the proposed approach, a constrained NSGA-II is
developed to determine a finite number of the best possible residential alternatives (up to ten alternatives
in this case study) for agents in accordance to their residential criteria and preferences.
A number of different evolutionary algorithms such as MOGA7 (Fonseca & Fleming, 1993), NSGA8
(Srinivas & Deb, 1994), NPGA9 (Horn et al., 1994), SPEA10 (Zitzler & Thiele, 1998), PAES11 (Knowles
& Corne, 1999) and NSGA-II (Deb et al., 2002) have been developed to solve multi-objective
optimization problems. Because of using elitism, SPEA, PAES and NSGA-II have attracted more interest
of researchers. Zitzler et al. (2000) showed that elitism results in enhancing the convergence of a multi-
objective evolutionary algorithm. NSGA-II is one of the fast and most efficient of elitist algorithms which
has been widely used for solving multi-objective optimization problems in various applications (Li et al.,
2010 ; Iniestra & Gutierrez, 2009 ; Huang et al., 2010). Deb et al. (2002) compared the convergence and
spread of solutions of NSGA-II, PAES and SPEA on difficult test problems and found that NSGA-II is
able to find better solutions for most problems.
7 Multi-Objective Genetic Algorithm
8 Non-dominated Sorting Genetic Algorithm
9 Niched Pareto Genetic Algorithm
10 Strength-Pareto Evolutionary Algorithm
11 Pareto-Archived Evolution Strategy
8
In multi-objective optimization problems, there is not a global optimal solution with respect to all
objectives. But there is a set of non-dominated solutions, generally known as the Pareto-optimal solutions.
Since no one of the non-dominated solutions is better than the other, each of them can be accepted as
optimal solution (Luh et al., 2003). NSGA-II uses the concept of non-domination to distinguish optimal
solutions. For a multi-objective optimization problem as expressed in Eq. (2), it is said that solution a
dominates solution b or b is dominated by a or a is not dominated by b if Eq. (3) is satisfied (Van
Veldhuizen & Lamont, 2000):
(cid:9)(cid:21)(cid:3)(cid:4)(cid:22)(cid:6)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:23)(cid:24)(cid:25)(cid:24)(cid:23)(cid:24)(cid:26)(cid:27)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:28)(cid:4)(cid:29)(cid:6)(cid:11)(cid:30)(cid:28)(cid:18)(cid:4)(cid:29)(cid:6)(cid:31)(cid:28) (cid:4)(cid:29)(cid:6)(cid:31)!(cid:31)(cid:28)(cid:16)(cid:4)(cid:29)(cid:6)"
(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)#$%&(cid:27)'((cid:8)()(cid:8)(cid:8)(cid:8)*(cid:4)(cid:29)(cid:6)(cid:11)(cid:30)*(cid:18)(cid:4)(cid:29)(cid:6)(cid:31)* (cid:4)(cid:29)(cid:6)(cid:31)!(cid:31)*(cid:16)(cid:4)(cid:29)(cid:6)"
(cid:9)(cid:21)(cid:3)(cid:4)+(cid:6)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:24)(cid:28)(cid:8),(cid:24)-(cid:30)(cid:5)(cid:31)(cid:22)(cid:31)!(cid:31)(cid:25)"(cid:7)(cid:8)(cid:8)(cid:28)(cid:14)(cid:4).(cid:6)/(cid:28)(cid:14)(cid:4)%(cid:6)(cid:8)(cid:8)(cid:8)(cid:8).(cid:25)0
(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)1(cid:24)-(cid:30)(cid:5)(cid:31)(cid:22)(cid:31)!(cid:31)(cid:25)"(cid:7)(cid:8)(cid:8)(cid:28)(cid:14)(cid:4).(cid:6)2(cid:28)(cid:14)(cid:4)%(cid:6)
where:
f(x) is the set of objective functions,
g(x) is the set of constraints,
a and b are the possible solutions.
NSGA-II starts with an initial random generation of parent population, P0 of size N. This population
contains possible solutions of the multi-objective optimization problem. In this research, P0 contains
possible residential zones for each agent. Afterwards, an offspring population, Q0 of size N, is generated
by applying genetic operators including binary tournament selection, crossover and mutation on P0. Since
the procedure is repetitive after the initial generation, the tth generation is described at the following. A
combined population, Rt of size 2N, is generated by Rt=Pt345(cid:3)(cid:8) The solutions in Rt are ranked and
assigned to different fronts according to the non-dominated sorting. In this sorting process, solutions
which do not dominate each other and dominate all the other solutions are assigned to the first (best)
front. This process is continued until all solutions of Rt are assigned to the non-dominated fronts. Then,
the parent population in the next generation, Pt+1 of size N, is generated from the solutions belonging to
the first (best) fronts of Rt. The new offspring population, Qt+1 of size N, is generated by applying binary
tournament selection, crossover and mutation operators on Pt+1. In the binary tournament selection, the
winner (better) solution is selected using the crowded-comparison operator. In this operator, a solution
with lower rank is selected as the winner. If two solutions have the same rank, a solution with higher
crowding distance is declared as the winner. The crowding distance of a solution is the perimeter of the
cuboid formed by its nearest neighboring solutions in the objective space. In fact, this distance estimates
the density of solutions surrounding a particular solution. The algorithm is continued until the
convergence criterion such as the maximum number of generations is satisfied and Pt is returned as the
9
output (Deb et al., 2002). Pt contains the optimal residential alternative zones of each agent. More details
of this algorithm and its implementation procedure can be found in Deb et al. (2002). .
In the following of this section, the most important criteria for residential location choice of tenants
derived from the survey of stated preferences of sample tenants in Tehran (Table 1) are described in
detail. It should be noted that depending on their preferences, agents may use one, some, or all of these
criteria for selecting their desired residential alternatives.
3.2.1. Housing rent
Housing price is one of the most important factors affecting the residential location choice of households
(Hunt, 2010 ; Ettema, 2011 ; Devisch et al., 2009 ; Jackson et al., 2008 ; Wu et al., 2013 ; Lee &
Waddell, 2010 ; Waddell et al., 2003 ; Sener et al., 2011). In this research, it is assumed that the average
housing rent per square meter in every zone is exogenously known and fixed during the simulation.
Tenants only select zones in which the housing rent is compatible with their affordability and required
residential area. For this purpose, a condition is considered to limit the search space of agents to zones in
which the housing rent of their required residential area is between the specified minimum and maximum
percentages of their monthly income. Because of limitations caused by the income level of agents, the
maximum limit seems applicable for all agents. But, the minimum limit is considered for some agents,
especially for agents with high monthly incomes, because, for cultural and social reasons, they usually
prefer to reside in rich neighborhoods (Habibi & Ahari, 2005). As a result, these limits lead to
consideration of socio-economic composition of residents in the residential location choice process of
agents. Also, by applying this condition, the search space of agents and required computational time are
significantly decreased. The following objective function is considered for each agent:
(cid:1)(cid:2)(cid:3)(cid:4)6(cid:6)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8) 789,(cid:8)(cid:20)-:;
,(cid:8)(cid:14)-:<(cid:4)=>(cid:20)(cid:19)=(cid:14)(cid:6)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)?@(cid:27)A(cid:27)(cid:7)(cid:8)(cid:8)B(cid:23)(cid:24)(cid:25)(cid:20)(cid:19)C(cid:20)/(cid:8)=>(cid:20)(cid:19)=(cid:14)/B(cid:23).(cid:29)(cid:20)(cid:19)C(cid:20)
where:
RAa is the required residential area of agent a;
Ri is the average housing rent per square meter in zone i;
Ia is the agent's monthly income;
Pmina and Pmaxa respectively are the minimum and maximum percentages of monthly income which are
considered by agent a for renting a residence. These percentages are defined using the Monte Carlo
simulation;
Na is the set of agents that the criterion is important for them; and
Nz is the set of residential zones.
10
3.2.2. Accessibility to public facilities
Households generally prefer to live in zones with high accessibilities to public facilities including
educational (Hunt, 2010 ; Jackson et al., 2008 ; Myers & Gearin, 2001 ; Wu et al., 2013 ; Lee &
Waddell, 2010), shopping (Hunt, 2010 ; Lee & Waddell, 2010 ; Chen et al., 2008 ; Lee et al., 2010 ;
Srour et al., 2002 ; Sener et al., 2011), green and recreational (Wu et al., 2013 ; Chen et al., 2008 ; Srour
et al., 2002 ; Sener et al., 2011), cultural (Sener et al., 2011), and health locations (Wu et al., 2013).
However their preferences for accessibility to various facilities are different depending on their socio-
economic and demographic characteristics. For example, while accessibility to schools may be very
important for a tenant with student members, it may not be important for the others. Overall accessibility
of zones to public facilities for each agent who this criterion is important for him is calculated by Eq. (5)
which is an extension of the accessibility index developed by Tsou et al. (2005). Also, accessibility of
zones to each public facility type is calculated using this equation, where value of pka for the public
facility type of interest is set to 1 and values of pka for the other public facility types are set to 0. This
index is normalized using (x-xmin)/(xmax-xmin).
(cid:1)(cid:2)(cid:3)(cid:4)D(cid:6)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8) 7EF,(cid:8)(cid:20)-:;
J
,(cid:8)(cid:14)-:<GHHIJ(cid:20)(cid:19)?K(cid:4)J(cid:6)(cid:19)0(cid:14)KL
K(cid:4)J(cid:6)
M
where:
k is the type of public facility including educational, shopping, green and recreational, cultural, and health
locations;
j(k) is the jth case of the kth type of public facility;
pka is the agent's preference to the public facility type k which is determined in the agent generation
module, where (cid:1)pka=1;
wj(k) is the relative effect of j(k) which is calculated by wj(k)=Aj(k)/max(Aj(k)), where Aj(k) is the area of j(k);
dij is the distance between zone i and j(k).
3.2.3. Accessibility to transportation services
Like the accessibility to public facilities, households usually prefer to reside in zones with high
accessibilities to transportation services including highways and/or public transit (Hunt, 2010 ; Myers &
Gearin, 2001 ; Wu et al., 2013 ; Sener et al., 2011). Overall accessibility of zones to transportation
services for each agent who this criterion is important for him is measured by Eq. (6) which is derived
from the studies of Currie (2010) and Delbosc and Currie (2011). Also, this equation is used to calculate
accessibility of zones to each transportation service type, where value of Pka for the transportation service
type of interest is set to 1 and values of Pka for the other transportation service types are set to 0. Finally,
accessibilities are normalized using (x-xmin)/(xmax-xmin).
11
(cid:1)(cid:2)(cid:3)(cid:4)N(cid:6)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8) 7EF,(cid:8)(cid:20)-:;
J
,(cid:8)(cid:14)-:<GHH>O(cid:4)J(cid:6)>(cid:14) (cid:19)BJ(cid:20)
O(cid:4)J(cid:6)
M
where:
k is the type of transportation service including highways, bus stops and subway stations;
t(k) is the tth case of the kth type of transportation service;
At(k) is the area of service range of t(k) which is inside zone i;
Ai is the area of zone i; and
Pka is the agent's preference to the transportation service type k which is determined in the agent
generation module, where (cid:1)Pka=1;
A fundamental component of accessibility to transportation services is the service range or access
distance. Researchers have typically used various walking distances ranging from 300 to 800 m for this
distance. Validity of these heuristic distances has been investigated using the travel survey data (Mavoa et
al., 2012). However, these distances are longer in Tehran, because in addition to walking, people often
use local taxis with low fares to access transit services. Therefore, according to the comprehensive
transportation and traffic studies of Tehran (Tehran Municipality, 2013a), the access distances to bus
stops, subway stations, and highways are defined as 1.5 km, 1.9 km, and 2 km, respectively.
3.2.4. Air and noise pollutions
Air and noise pollutions can have important influences on the residential location choice of households
(Hunt, 2010). By using the annual mean noise and air pollution records of Tehran (Tehran Municipality,
2013a), residential zones are classified in five categories varying from clean to highly polluted zones.
Then, residential alternatives of agents having very important preferences for air and noise pollutions are
restricted to medium to clean pollution classes. Also, the following objective functions are applied for
agents who these criteria are recognized as their important criteria.
(cid:1)(cid:2)(cid:3)(cid:4)P(cid:6)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8) 789,(cid:8)(cid:20)-:;
(cid:1)(cid:2)(cid:3)(cid:4)Q(cid:6)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8) 789,(cid:8)(cid:20)-:;
,(cid:8)(cid:14)-:<(cid:8)(cid:4)>B(cid:14)(cid:6)
,(cid:8)(cid:14)-:<(cid:8)(cid:4)(cid:12)B(cid:14)(cid:6)(cid:8)
where APi and NPi respectively are the annual mean air and noise pollutions of zone i.
3.2.5. Distance from the workplace
Distance of residence from the workplace is an important factor for many households (Hunt, 2010 ;
Jackson et al., 2008 ; Myers & Gearin, 2001 ; Wu et al., 2013 ; Rashidi et al., 2012 ; Lee & Waddell,
2010 ; Chen et al., 2008 ; Lee et al., 2010 ; Srour et al., 2002 ; Waddell et al., 2003). The following
12
objective function is used for agents who prefer to minimize distance of their residence from the
workplace(s) of their members.
(cid:1)(cid:2)(cid:3)(cid:4)R(cid:6)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8) 789,(cid:8)(cid:20)-:;
,(cid:8)(cid:14)(cid:31)S-:<H(cid:8)0(cid:14)S(cid:4)T(cid:6)
(cid:16)
T(cid:17)(cid:18)
where:
diw(m) is the distance between zone i and workplace of employed member m of agent a.
n is the number of employees in agent a.
3.2.6. Distance from the former residence
Another important factor in residential location choice of some households is distance from their former
residence (Jackson et al., 2008 ; Chen et al., 2008). Because of familiarity, various dependencies, and
meeting their requirements in the former residential area, some agents prefer to live nearby their former
residential area. Objective function of these agents is defined as:
(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)U(cid:6)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8) 789,(cid:8)(cid:20)-:;
,(cid:8)(cid:14)(cid:31)VW-:<(cid:8)(cid:4)0(cid:14)VW(cid:4)(cid:20)(cid:6)(cid:6)
where diFr(a) is the distance between zone i and the former residence of agent a.
3.2.7. Traffic restrictions
Households usually consider type of streets and traffic situations in their residential area (Hunt, 2010 ;
Martinez & Viegas, 2008). Because of traffic restrictions such as the odd-even car restrictions and
restrictions on private cars in central areas of Tehran, some households prefer to reside outside these
areas. For this purpose, agents with very important preferences to traffic restrictions look for their
residence in non-restricted areas. Also, the following objective function is used to minimize selection of
traffic restricted zones by agents having important preferences to these restrictions:
(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:5)(cid:6)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8) 789,(cid:8)(cid:20)-:;
,(cid:8)(cid:14)-:<(cid:4)X=(cid:14)(cid:6)
where TRi is the traffic restriction code of zone i with the values of 0 for no traffic restriction, 1 for odd-
even car restriction, and 2 for restriction on all private cars.
3.3. Competition and residence choice
After selection of desired residential alternatives, agents compete with each other to select their final
residence. In reality, tenants relocate in different months of the year. A monthly time step is considered in
this paper, because rent is usually paid each month and tenants commonly begin to look for a new
residence from one month before their lease deadline. In other words, households usually find a new
13
residence during one month. For simulation of this process, the available statistical information of the
residential relocation rates in different months of the year in Tehran (Tehran Municipality, 2013a) is used
to randomly define the relocation month of each agent. At the other hand, residential capacity of zones is
limited in each month and estimated by Eq. (12):
(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:22)(cid:6)(cid:7)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)(cid:8)=(cid:10)(cid:14)O(cid:11) =>(cid:14)(cid:15) =>(cid:14)
(cid:14)
(cid:19)(cid:4)YO(cid:19)(cid:12)(cid:20)O(cid:6)
where:
RCit is the residential capacity of zone i at month t;
RAi is the total residential area inside zone i;
Nat is the number of agents who look for residence at month t; and
(cid:2)t is a balancing parameter that presents an equilibrium between the residential supply and demand at
month t. For example, value of more than one for this parameter means that the residential supply is more
than the demand. This parameter is calibrated using the available statistical information in each month
(Tehran Municipality, 2013a).
In each month, a number of agents compete with each other for selecting a residence among their
residential alternatives. They search among their desired alternative zones according to the distance of
these zones from their former residence and reside in the first zone having the enough residential
capacity. If demand for a zone is more than its capacity, then agents compete with each other. As
mentioned in section 2, study of the process of renting a house and consultation with real estate agencies
in Tehran show that landlords commonly prefer to rent their properties to households with few members
(except singles), without child, and with high incomes. Therefore, in the proposed residential competition,
agents with fewer members (except singles), with higher incomes, and without child have a higher chance
of success, respectively. In the case of equity of the above mentioned conditions, winners are randomly
selected. Defeated agents try to reside in their next alternatives and the process is repeated until all agents
either reside in one of their residential alternatives or evaluation of all alternatives of defeated agents is
completed. Agents who cannot reside in any of their alternatives are moved to the next month. It is
assumed that these agents can take additional time (up to one month) from the landlords to find a new
residence. Landlords usually agree with this request, because the legal process of expelling a tenant from
occupying a property generally is too long and expensive. Therefore, these agents would compete again
with agents of the next month and if they still cannot reside in any zone, no residence is allocated for
them. In fact, they are agents with many members or low incomes who cannot reside in any zone of
Tehran and possibly they have to move to the surrounding cities of Tehran.
14
4. Empirical results
In an attempt to cover all socio-economic categories of tenants in Tehran, 100,000 agents were simulated
using the agent generation module developed in MATLAB 7.9.0 software. These agents were randomly
classified to 12 months according to the percentage of residential relocation in different month of the year
in Tehran (Tehran Municipality, 2013a). Table (2) shows the number of agents who look for a residence
in each month. As shown in this table, due to school holidays, about 65% of the annual residential
relocation in Tehran occur in three months of the summer (Tehran Municipality, 2013a).
Table (2): Distribution of residential relocation and agents in different months of the year
Time
Period
(Month)
April
May
June
July
August
September
October
November
December
January
February
March
1 Year
Percentage
Of Relocation
1%
8%
7%
19%
22%
24%
8%
5%
2%
1%
2%
1%
100%
Number
Of
Agents
1000
8000
7000
19000
22000
24000
8000
5000
2000
1000
2000
1000
100000
Agents then selected up to 10 residential alternatives by the developed NSGA-II. For determining the
maximum number of residential alternatives of agents, the proposed approach was run with different
maximum numbers of alternatives varying from 2 to 15 for sample tenants presented in Table (1).
Simulation accuracy of the actual residential zone of these tenants by different maximum numbers of
alternatives has been shown in Figure (3). As seen in this Figure, the simulation run with the maximum
number of ten residential alternatives shows the best performance. Also, because of using relatively large
zones in this case study, it seems incredible that agents be able to search more than ten zones.
In order to test the repeatability and stability, the proposed simulation approach was executed 5 times and
aggregated results of these runs were compared with each other. Results showed that aggregated
residential location choice behaviors of different demographic and socio-economic categories of agents in
different simulation runs are very close to each other with differences of less than 2%. This suggests
acceptable repeatability and stability of the proposed approach. Therefore, in this paper, empirical results
of one simulation run are presented. The average run time of each simulation is about 7.3 hours in a
platform with Core i7 2.00 GHz of CPU and 6 GB of RAM. The simulation stops either all agents reside
in one of their residential alternatives or evaluation of all alternatives of agents is completed.
15
70
60
50
40
30
20
10
0
)
%
(
n
o
i
t
a
l
i
u
m
S
f
o
y
c
a
r
u
c
c
A
2
3
4
5
6
7
8
9
10 11 12 13 14 15
Maximum number of residential alternatives
Figure (3): Accuracy of simulation of the actual residential zone of 330 sample tenants by the maximum number of residential
alternatives selected by them
In the following of this section, major patterns derived from simulation of the residential location choice
of individual agents are analyzed. For different demographic and socio-economic categories of agents,
these aggregated patterns are described by various criteria considered in the simulation design including
housing rent, accessibility to transportation services, accessibility to public facilities, distance from the
former residence and workplace, air and noise pollutions and traffic restrictions. In fact, this analysis
shows that how well the proposed approach can simulate the observed or expected residential location
choice behaviors of tenants in Tehran. This highlights the performance of the simulation in generating
expected and near-reality results.
Spatial distribution of residential alternatives of agents is shown in Figure 4(a). As expected, because of
low housing rents and high accessibilities to employment opportunities, public facilities and public transit
services, central areas of Tehran are more attractive than the other areas. However, relatively higher air
and noise pollutions and traffic restrictions in these areas lead to reducing the residential attractiveness of
these areas for some agents, especially for agents with high incomes that is consistent with the
observations. Also, spatial distribution of final residence of agents is shown in Figure 4(b). As illustrated
in this Figure, the population density is higher in central areas which is consistent with the existing
population distribution in Tehran (Tehran Municipality, 2013b). Due to higher residential demand and
consequently more intense competition with other agents in central areas, 64.2% of residents of these
areas cannot reside in their first three alternatives. On the other hand, because of low residential demands
in southern and northwestern areas, 69.5% of residents of these areas have resided in one of their first
three alternatives.
16
(a)
(b)
Figure (4): Distribution of: (a) the residential alternatives; and (b) the final residence of agents in TAZs
Distribution of agents by the number of their residential alternatives has been represented in Figure 5(a).
As can be seen, 83.3% of agents have selected the maximum possible number of residential alternatives,
i.e. ten alternatives. This suggests that the constraints applied in the proposed approach are reasonable so
that the choice set of agents is limited to a few residential alternatives. However, the choice set of 1.8% of
agents has been limited to less than 4 residential alternatives. These agents generally have low incomes or
strict preferences for choosing their residential location. Also, distribution of agents by the rank of their
final residence among their residential alternatives has been represented in Figure 5(b). As indicated in
this histogram, 19.7% of agents have resided in their first residential alternative. These agents generally
have a few members and high incomes or have selected their residence from southern and northwestern
areas in which the residential demand is less than those of the other areas. An observable decreasing trend
in this histogram is consistent with the expectations. Because, agents attempt to reside in their first
alternatives and the number of agents who have resided in their last alternatives has sequentially
decreased. Finally, only 0.7% of agents cannot reside in any zone and may have to move to the
surrounding cities of Tehran. It seems that low income level and high household size of these agents are
the main reasons for this situation. Mean income and household size of these agents are 7,020,000 IRR12
and 4.3, respectively.
Distribution of agents by distance of their residence from their workplace(s) is shown in Figure 6(a).
Although proximity to the workplace is important for 86.4% of agents, results of the simulation show that
21.2% of agents have resided in distances of more than 10 kilometers from their workplace(s). This
unexpected result suggests that though a great number of agents prefer to reside in proximities of their
workplace(s), a majority of them cannot practically reside close to their workplace(s) due to other
important factors such as their socio-economic characteristics, housing rent, and not meeting the other
12 Iranian Rial. At the time of this study 1 USD is about 36,000 IRR.
17
requirements. In addition, low prices of fuel and transit fares in Tehran may be regarded as the other
important reason for residing of a considerable number of agents far from their workplace(s). However,
because of reduction of their commuting costs, low-income agents and who have two or more employees
with near workplaces have shown higher interests to reside close to their workplace(s) which is consistent
with the expectations. As an example, 65.7% of agents who have two or more employees with near
workplaces have resided in distances of less than 5 km from their workplaces, whereas only 8.1% of these
agents have resided in distances of more than 10 km from their workplaces.
Figure 6(b) shows distribution of agents by distance between their current and former residences. As
expected, due to some reasons such as forced relocation because of lack of agreement with the landlord
for extension of the lease, familiarity and dependency to the former residential area, and high financial
and psychological costs of moving to far areas, a great number of agents have preferred to reside in
proximities of their former residence. For instance, only 4.7% of agents have moved to distances of more
than 10 km from their former residence.
)
%
(
s
t
n
e
g
A
90
80
70
60
50
40
30
20
10
0
< 4
4
5
(a)
7
6
9
Number of alternatives
8
21
18
15
)
%
12
(
s
t
n
e
g
A
9
6
3
0
10
(b)
1 2 3 4 5 6 7 8 9
Rank of alternative
0
1
…
o
h
t
i
W
Figure (5): Distribution of agents by: (a) the number of their residential alternatives; (b) the rank of their final residence among
their alternatives
)
%
(
s
t
n
e
g
A
25
20
15
10
5
0
)
%
(
s
t
n
e
g
A
25
20
15
10
5
0
(a)
Distance from the workplace (km)
Distance from the former residence (km)
Figure (6): Distribution of agents by: (a) distance from their workplace; (b) distance from their former residence
(b)
18
Spatial distribution of the mean income of agents exhibits a clear south-north pattern (Figure (7)). This
pattern shows considerable similarities with the existing distribution of mean income of residents in
Tehran (Tehran Municipality, 2013b). As hypothesized, agents have resided in zones in which the mean
income of residents is compatible with their income level. For example, although high-income agents
could easily reside in zones with low housing rents, they have preferred to reside in zones with higher
housing rents compatible with their income level.
Figure (7): Distribution of the mean income of residing agents in TAZs
As expected, heterogeneous distributions of highways and public transit in Tehran lead to high
importance of these factors in residential location choice of agents. Simulation results show that agents
with cars particularly with more than one car have more resided in proximities of highways. For example,
71.8% of agents having more than one car have resided in zones with accessibilities of more than 70% to
highways. On the other hand, as anticipated, agents without car have considerably resided in central zones
with higher accessibilities to public transit services. In addition, an unanticipated weak association is
observed between residing in zones with high accessibilities to public transit and the number of members
and employees of agents. Probably, because of higher demand for public transit use by their members,
agents with more than one employee or three members slightly have more resided in proximities of public
transit services.
As hypothesized, results show the importance of accessibility to various public facilities in residence
choice of different socio-economic categories of agents. A majority of agents have selected their
residential alternatives from zones with high accessibilities to shopping and recreational locations which
is consistent with the expectations. For example, the number of residential alternatives selected by agents
in zones with accessibilities of higher than 70% to shopping and recreational locations is 10.2% and 5.9%
19
more than the average number of residential alternatives in all zones, respectively. In addition, as
anticipated, agents with student members have tended to reside in proximities of educational locations.
For example, 54.1% of agents having more than one student have resided in zones with accessibilities of
higher than 70% to educational locations. Finally, results do not show any meaningful relationships
between residential location choice and accessibility to cultural and health centers. This suggests the
minor role of these criteria in rental residence choice in Tehran.
As expected, air pollution and especially noise pollution play an important role in the residence choice of
agents.. An anticipated strong relationship between the income level of agents and their preferences for
less polluted areas is observed so that high-income agents have mostly resided in zones with low air and
noise pollutions. For example, only 13.8% and 7.7% of high-income agents have resided in zones with
high levels of air and noise pollutions, respectively. On the other hand, low-income agents have been
forced to reside in zones with higher pollutions due to their limited affordability and lower housing rents
in these zones. Importance of the noise pollution can be clearly shown in residential areas around the
Mehrabad airport. Although these areas have high accessibilities to highways, public transit services and
public facilities, the number of residential alternatives selected by agents in these areas is 44.3% less than
the average number of residential alternatives in all zones. Also, only 8.7% of residents of these areas
have high income levels.
As hypothesized, results show the significant influences of traffic restrictions (e.g. restrictions for all and
odd-even cars) on the residence choice of agents, especially on car-owning agents. The number of
residents having one and more than one car in traffic-restricted areas respectively is 10.6% and 52.2% less
than the average number of this class of residents in all zones. On the other hand, due to central location
of these areas and high accessibilities to employment opportunities and public transit services, agents
without car have shown great interests for residing in these areas which is consistent with the
observations. Number of residents without car in these areas is 57.8% more than the average number of
this class of residents in all zones. Finally, it should be noted that simulated distribution of the mean car
ownership is considerably consistent with the existing mean car ownership distribution in Tehran (Tehran
Municipality, 2013b).
In general, it can be said that major residential location choice patterns of different demographic and
socio-economic categories of agents are considerably consistent with the observations and expectations.
A summary of residential location choice and selected residential zones by agents have been presented in
Tables (3) and (4).
20
Table (3): A summary of residential location choice of agents
Category of Agents
Selected all 10 alternatives
Selected more than 7 alternatives
Selected less than 4 alternatives
Resided in the first alternative
Resided in one of the first three alternatives
Resided in one of the last three alternatives
Moved to the next time period
unable to reside in any zone
Resided in a zone with distance of less than 5 km from the former residence
Resided in a zone with distance of more than 10 km from the former residence
Resided in a zone with distance of less than 5 km from the workplace
Resided in a zone with distance of more than 10 km from the workplace
Percentage
of Agents
(%)
83.3
92.8
1.8
19.7
47.3
15.5
5.7
0.7
69.2
4.7
42.5
21.2
Average
Income
(million
IRR)
14.2
11.4
7.0
15.9
13.7
10.3
7.1
6.7
12.6
13.5
14.1
14.9
Average
Number of
Members
3.43
3.49
3.86
3.02
3.31
3.78
4.05
4.24
3.48
3.46
3.47
3.50
Table (4): A summary of selected residential zones by agents
Average
Car
Ownership
1.02
0.97
0.82
1.14
1.00
0.88
0.84
0.76
0.97
1.02
0.90
1.05
Category of Residential
Zones
The most selected zones as
residential
by
agents
alternatives
The least selected zones as
residential
by
agents
alternatives
Zones with the most residents
Zones with the lowest residents
Zones which were occupied
earlier than other zones
Zone
Number
Number
of Agents
*Housing
Rent per
m2 (%)
Air
Pollution
Noise
Pollution
531
311
528
2
16
17
531
560
360
2
14
15
311
508
221
5639
5303
5058
764
809
855
828
814
790
64
71
74
595
237
280
62.7
58.5
68.3
40.6
44.2
43.8
62.7
67.4
38.6
40.6
39.7
41.3
58.5
60.9
30.2
Clean
Relatively Clean
Clean
Clean
Clean
Clean
Highly Polluted Highly Polluted
Highly Polluted Highly Polluted
Highly Polluted Highly Polluted
Clean
Clean
Medium
Clean
Clean
Relatively Clean
Highly Polluted Highly Polluted
Highly Polluted Highly Polluted
Highly Polluted Highly Polluted
Relatively Clean
Relatively Clean
Clean
Medium
Medium
Medium
**Overall
Access to
Public
Facilities
(%)
56.4
60.1
43.2
68.3
62.8
65.0
56.4
58.9
41.2
68.3
56.1
59.0
65.1
73.3
58.4
***Overall
Access to
Highway
Network
(%)
70.6
63.8
65.2
28.4
33.5
32.1
70.6
66.4
58.0
28.4
25.6
30.7
63.8
68.9
43.5
****Overall
Access to
Public
Transit
(%)
37.3
46.5
29.6
78.9
81.3
79.5
37.3
41.2
33.4
78.9
84.4
79.7
46.5
61.5
48.7
* The housing rent per m2 has been normalized using (x-xmin)/(xmax-xmin).
** Overall access to public facilities is calculated using Eq. (5) where all public facilities have the same preferences (pka). This index has been
normalized using (x-xmin)/(xmax-xmin).
*** Overall access to highway network is calculated using Eq. (6) where Pka is equal to 1 and k is highway. This index has been normalized
using (x-xmin)/(xmax-xmin).
**** Overall access to public transit is calculated using Eq. (6) where Pka is equal to 0.5 and k is the bus and subway. This index has been
normalized using (x-xmin)/(xmax-xmin).
5. Validation
One of the greatest challenges of utilizing agent-based models is their validation (Crooks & Heppenstall,
2012). In this research, 1350 tenant households with various socio-economic characteristics residing in
different zones of Tehran were sampled for validation of the proposed approach. Socio-economic
characteristics, residence and workplace(s) of these households were collected, but their residential
21
criteria and preferences were simulated using the proposed approach. For this purpose, households were
classified to 12 months according to the month in which they rented their residence. Then, residence of
these households was simulated by the proposed approach and results were compared with those of the
actual residence.
As shown in Figure (8), spatial distribution of simulated and actual residences of these households is
significantly compatible. Distribution of households by distance of their simulated residence from their
actual residence demonstrates validity of the simulation (Figure 9(a)). As indicated in this diagram,
simulated residence of 60.2% of households is located in distance of less than one kilometer from their
actual residence. Also, simulated residence of only 5.4% of households is located in more than ten
kilometers from their actual residence. In addition, distribution of the actual and simulated residences of
households by housing rent, proximity to their workplace(s) and distance from their former residence
show high similarities (Figures 9(b), 10(a) and 10(b)).
Comparison of the actual and simulated residences of households in different spatial and attribute
categories has been presented in Table (5). As shown in this table, the proposed approach has exactly
simulated the residential zone of 59.3% of households. Also, simulated residential zone of 67.5% of
households has been located in the adjacency of their actual residential zone, composed of the actual
residential zone and its neighboring zones. In addition, the actual residence of 72.8% of households has
been simulated as one of their residential alternatives. The approach is also able to simulate different
attributes of the actual residence of more than 70% of households including the housing rent, air and noise
pollutions and accessibilities to public facilities and transportation services with errors of less than 15%.
For example, the housing rent of 84.0% of households has been simulated with errors of less than 15%.
Therefore, it seems that the proposed approach shows promising performance.
(a)
(b)
Figure (8): Distribution of (a) the actual; and (b) simulated residences of households in TAZs
22
100
95
90
85
80
75
70
65
60
55
50
)
%
(
s
d
l
o
h
e
s
u
o
H
(a)
)
%
(
s
d
l
o
h
e
s
u
o
H
(b)
1
2
3
4
5
6
7
8
9
10 > 10
Distance (km)
40
35
30
25
20
15
10
5
0
Actual
Simulated
<5
5-10
10-15
15-20
>20
Housing Rent (million IRR)
Figure (9): Distribution of households by: (a) distance between their simulated and actual residences; (b) housing rent of their
actual and simulated residences
)
%
(
s
t
n
e
g
A
Actual
Simulated
25
20
15
10
5
0
Actual
Simulated
25
20
15
10
5
0
)
%
(
s
t
n
e
g
A
(a)
Distance from the workplace (km)
(b)
Distance from the former residence (km)
Figure (10): Distribution of households by: (a) distance from their workplace; (b) distance from their former residence
Table (5): Comparison of the actual and simulated residences of households
Simulated residence is identical to the actual residence
The actual residence is one of the simulated residential alternatives
Simulated residence is located in the adjacency of the actual residence
Simulated residence is located in distance of less than 5 km from the actual residence
Simulated residence is located in distance of more than 10 km from the actual residence
Rent of simulated residence is in the range of 85% to 115% of the rent of actual residence
Accessibility of simulated residence to the public facilities is in the range of 85% to 115% of accessibility of the actual residence
Accessibility of simulated residence to the highway network is in the range of 85% to 115% of accessibility of the actual residence
Accessibility of simulated residence to the public transit is in the range of 85% to 115% of accessibility of the actual residence
Air pollution level of simulated residence is identical to the actual residence
Noise pollution level of simulated residence is identical to the actual residence
Households
%
59.3
72.8
67.5
81.4
5.4
84.0
78.1
84.5
70.6
79.2
76.3
6. Discussion and Conclusion
In this paper, a novel two-step agent-based approach has been developed to simulate residential location
choice of tenants. Tenants are considered as agents who compete with each other to reside in one of their
preferred residential alternatives. They first select their desired residential alternatives according to their
residential criteria and preferences using NSGA-II. In this step, they do not consider residential capacity
23
of zones and residential alternatives of the other agents. Then, they select their final residential zone in a
competition with other agents who look for the same zone at the same time period. Results show that the
proposed approach has considerably simulated major residential patterns of different socio-economic
categories of tenants in Tehran. For example, simulated residential choice behavior of agents with respect
to accessibility to various public facilities and transportation services, air and noise pollutions and traffic
restrictions are significantly consistent with the expectations. Also, validation results of the proposed
approach with a sample of tenants shows that the approach has correctly simulated the actual residential
zone of 59.3% of tenants. In addition, for more than 70% of sample households, various attributes of their
actual residential zone including accessibility to public facilities and transportation services, air and noise
pollutions and housing rent has simulated with errors of less than 15%. All of these results suggest the
remarkable performance of the proposed approach.
The proposed approach has several advantages which lead to more closeness of the simulation of
residential location choice of agents to the reality. First, due to using a multi-objective decision making
method, NSGA-II, various conflicting objectives can be considered in the residence choice process of
agents. There is no limitation on the number and type of objectives used in this method. The method also
allows modeler to consider different objectives according to agents' preferences. In addition, the method
can be used with a huge number of residential options. Second, In contrast with the conventional
residential location choice models, agents do not essentially select the best residential option (the global
optimum). They select a finite number of their desired residential alternatives with respect to their criteria
and preferences and finally select their residence among these alternatives by competition with other
agents.
The proposed residential location choice approach can be used in different urban applications. It can be
considered as a main component of the land use and transportation models. It remarkably helps urban
planners to investigate the residence choice behavior of individual households and consequently the
spatial distribution of different socio-economic categories of households. Therefore, effects of various
urban policies on these concerns can be investigated. Also, the proposed approach can be used for
determining tenants who are unable to reside in any zone and possibly have to move to the surrounding
cities of Tehran. Moreover, socio-economic structure of the population resided in different zones can be
revealed and used for different purposes such as the housing subsidies to specific groups of population.
Although the proposed approach has been implemented in Tehran as a case study, it is a general approach
which can be adapted with the residential location choice process in other developing and developed
countries. The proposed multi-objective decision making method, NSGA-II, allows modelers to define
different preferred criteria and objectives in accordance with the residential location choice in their study
area. Also, although the criteria including number of members, income level, and number of children are
24
considered in the proposed residential competition mechanism in Tehran, other criteria such as ethnicity,
professional status, age and education level of the head of household can be considered in other areas in
conformity with their cultural and socio-economic conditions.
Various aspects of the proposed approach need further considerations and developments. In this paper, the
approach has been used by relatively large spatial units, TAZs, which may leads to modifiable areal unit
problem (MAUP). This means that changes in the size or configuration of zones can affect the results
(Openshaw & Taylor, 1979 ; Martinez et al., 2009). Therefore, MAUP effects must be carefully
evaluated by using smaller spatial units such as parcels, census blocks or larger units resulting from their
aggregations. Also, in this study, Euclidian distance has been used for measuring the proximity to
different opportunities. Due to use of relatively large spatial units, TAZs, in this study, the use of
Euclidian distance seems to be appropriate. It seems that people perceive proximity as a straight line
(Euclidian distance) at the large spatial levels such as regions and TAZs. Also, the Euclidean distance is
strongly correlated with the network distance at the census tract level in metropolitan areas (Apparicio et
al., 2008). It is recommended that the network distance be used in future studies, especially when smaller
spatial units such as census blocks or parcels are used. This measure may more accurately correspond to
human perceptions of accessibility to different opportunities at the parcel level, because it measures the
road distance between the parcel and the opportunity. However, calculation of the network distance is
computationally intensive and requires greater user effort and knowledge for data input and preparation
(Sander et al., 2010). The use of network distance in metropolitan areas such as Tehran which is
composed of very dense transport networks may significantly reduce the computational performance of
the proposed approach. In addition, only residential location choice of tenant households is simulated in
this research, but the approach can be developed to include all categories of households. Finally, modules
for determination of housing rent in a bid-rent framework and simulation of housing supply can be added
to the approach in the future.
References
Alonso W 1960 A theory of the urban land market Paper and Proceedings of the Regional Science Association 6
149-157
Apparicio P, Abdelmajid M, Riva M and Shearmur R 2008 Comparing alternative approaches to measuring the
geographical accessibility of urban health services: Distance types and aggregation-error issues
International Journal of Health Geographics 7(7)
Arentze T, Timmermans H and Veldhuisen J 2010 The Residential Choice Module in the Albatross and Ramblas
Model Systems in Pagliara F, Preston J and Simmonds D eds. Residential Location Choice: Models and
Applications, Springer-Verlag, Berlin Heidelberg 209-222
Arentze T and Timmermans HJP 2004 A learning-based transportation-oriented simulation system Transportation
Research B 38(7) 613-633
25
Barros JX 2004 Urban Growth in Latin American Cities: Exploring urban dynamics through agent-based simulation
in Urban Growth in Latin American Cities: Exploring urban dynamics through agent-based simulation,
University College London (UCL), 285
Benenson I 2004 Agent-Based Modeling: From Individual Residential Choice to Urban Residential Dynamics in
Goodchild MF and Janelle DG eds. Spatially Integrated Social Science: Examples in Best Practice, Oxford
University Press, New York 67-95
Chang JS 2006 Models of the relationship between transport and land-use: a review Transport Reviews 26 (3) 325-
350
Chang JS and Mackett RL 2006 A bi-level model of the relationship between transport and residential location
Transportation Research Part B 40 123-146
Chen J, Chen C and Timmermans H 2008 Accessibility trade-offs in household residential location decisions
Transportation Research Record 2077 71-79
Crooks AT and Heppenstall AJ 2012 Introduction to Agent-Based Modelling in Heppenstall AJ, Crooks AT, See
LM and Batty M eds. Agent-Based Models of Geographical Systems, Springer, Dordrecht 85-105
Currie G 2010 Quantifying spatial gaps in public transport supply based on social needs Journal of Transport
Geography 18 31-41
Deb K, Pratap A, Agarwal S and Meyarivan T 2002 A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II
IEEE Transactions on Evolutionary Computation 6 (2) 182-197
Delbosc A and Currie G 2011 Using Lorenz curves to assess public transport equity Journal of Transport
Geography 19 1252-1259
Devisch OTJ, Timmermans HJP, Arentze TA and Borgers AWJ 2009 An agent-based model of residential choice
dynamics in nonstationary housing markets Environment and Planning A 41(8) 1997-2013
Ettema D 2011 A multi-agent model of urban processes: Modelling relocation processes and price setting in housing
markets Computers, Environment and Urban Systems 35 1-11
Fonseca CM and Fleming PJ 1993 Genetic algorithms for multiobjective optimization: Formulation, discussion and
generalization in Genetic algorithms for multiobjective optimization: Formulation, discussion and
generalization, Morgan Kauffman, 416-423
Gaube V and Remesch A 2013 Impact of urban planning on household's residential decisions: An agent-based
simulation model for Vienna Environmental Modelling & Software 45 92-103
Haase D, Lautenbach S and Seppelt R 2010 Modeling and simulating residential mobility in a shrinking city using
an agent-based approach Environmental Modelling & Software 25 1225-1240
Habibi M and Ahari Z 2005 Study of qualitative aspects of housing in Iran in Study of qualitative aspects of housing
in Iran, Report in Persian, Ministry of Roads & Urban Development, Deputy of Housing & Building City
Horn J, Nafploitis N and Goldberg DE 1994 A niched Pareto genetic algorithm for multiobjective optimization in A
niched Pareto genetic algorithm for multiobjective optimization, IEEE Press, 82-87
Huang B, Buckley B and Kechadi T-M 2010 Multi-objective feature selection by using NSGA-II for customer churn
prediction in telecommunications Expert Systems with Applications 37 3638-3646
Huang Q, Parker DC, Filatova T and Sun S 2014 A review of urban residential choice models using agent-based
modeling Environment and Planning B: Planning and Design 41(4) 661-689
Hunt JD 2010 Stated Preference Examination of Factors Influencing Residential Attraction in Pagliara F, Preston J
and Simmonds D eds. Residential Location Choice: Models and Applications, Springer, Verlag Berlin
Heidelberg
Iacono M, Levinson D and El-Geneidy A 2008 Models of transportation and land use change: A guide to the
territory Journal of Planning Literature 22 (4) 323-340
26
Iniestra JG and Gutierrez JG 2009 Multicriteria decisions on interdependent infrastructure transportation projects
using an evolutionary-based framework Applied Soft Computing 9 512-526
Jackson J, Forest B and Sengupta R 2008 Agent-based simulation of urban residential dynamics and land rent
change in a gentrifying area of Boston Transactions in GIS 12 475-491
Jordan R, Birkin M and Evans A 2012 Agent-Based Modelling of Residential Mobility, Housing Choice and
Regeneration in Heppenstall AJ, Crooks AT, See LM and Batty M eds. Agent-Based Models of
Geographical Systems, Springer, 511-524
Knowles J and Corne D 1999 The Pareto archived evolution strategy: A new baseline algorithm for multiobjective
optimization in The Pareto archived evolution strategy: A new baseline algorithm for multiobjective
optimization, IEEE Press, 98-105
Lee BHY and Waddell P 2010 Residential mobility and location choice: a nested logit model with sampling of
alternatives Transportation 37 587-601
Lee BHY, Waddell P, Wang L and Pendyala RM 2010 Reexamining the influence of work and nonwork
accessibility on residential location choices with a microanalytic framework Environment and Planning A
42 913-930
Lerman SR 1976 Location, housing, automobile ownership, and mode to work: a joint choice model Transportation
Research Record 610 6-11
Li M, Lin D and Wang S 2010 Solving a type of biobjective bilevel programming problem using NSGA-II
Computers and Mathematics with Applications 59 706-715
Lowry IS 1964 A Model of Metropolis in A Model of Metropolis, RAND Corporation
Luh GC, Chueh CH and Liu WW 2003 MOIA: Multi-objective immune algorithm Engineering Optimization 35 (2)
Martinez F and Donoso P 2010 The MUSSA II Land Use Auction Equilibrium Model in Pagliara F, Preston J and
Simmonds D eds. Residential Location Choice: Models and Applications, Springer, Verlag Berlin
Heidelberg 99-113
Martinez FJ 1996 MUSSA: Land Use Model for Santiago City Transportation Research Record 1552 126-134
Martinez LM, Viegas JM and Silva EA 2009 A traffic analysis zone definition: a new methodology and algorithm
Transportation 36 581–599
Martinez LMG and Viegas JMC 2008 Activities, Transportation Networks and Land Prices as Key Factors of
Location Choices: An Agent-Based Model for Lisbon Metropolitan Area (LMA) in Activities,
Transportation Networks and Land Prices as Key Factors of Location Choices: An Agent-Based Model for
Lisbon Metropolitan Area (LMA), MIT-Portugal Program
Mavoa S, Witten K, McCreanor T and O'Sullivan D 2012 GIS based destination accessibility via public transit and
walking in Auckland, New Zealand Journal of Transport Geography 20 15-22
Mcfadden D 1978 Modelling the choice of residential location in Karlqvist ALL, Lundqvist L, Snickars F and
Weibull J eds. Spatial Interaction Theory and Planning Models, North-Holland, Amsterdam 75-96
Miller EJ, Hunt JD, Abraham JE and Salvini PA 2004 Microsimulating urban systems Computers, Environment and
Urban Systems 28 9-44
Miller EJ and Salvini PA 2001 The integrated land use transportation environment (ILUTE) microsimulation
modelling system: description and current status in Hensher DA ed. Travel Behaviour Research The
Leading Edge, Pergamon Press, Amsterdam 711-724
Myers D and Gearin E 2001 Current preferences and future demand for denser residential environments Housing
Policy Debate 12(4) 633-659
Openshaw S and Taylor PJ 1979 A Million or so Correlation Coefficients: Three Experiments on the Modifiable
Areal Unit Problem in Wrigley N ed. Statistical Applications in the Spatial Sciences, Pion, London 127-
144
27
Otter HS, van der Veen A and de Vriend HJ 2001 ABLOoM: Location behaviour, spatial patterns, and agent-based
modelling Journal of Artificial Societies and Social Simulation 4 (4)
Pagliara F and Wilson A 2010 The State-of-the-Art in Building Residential Location Models in Pagliara F, Preston J
and Simmonds D eds. Residential Location Choice: Models and Applications, Springer, Verlag Berlin
Heidelberg 1-20
Rashidi TH, Auld J and Mohammadian A 2012 A behavioral housing search model: Two-stage hazard-based and
multinomial logit approach to choice-set formation and location selection Transportation Research Part A
46 1097–1107
Rosenfield A, Chingcuanco F and Miller EJ 2013 Agent-based housing market microsimulation for integrated land
use, transportation, environment model system Procedia Computer Science 19 841-846
Salvini PA and Miller EJ 2005 ILUTE: An Operational Prototype of Comprehensive Microsimulation Model of
Urban Systems Network and Spatial Economics 5 217-234
Sander HA, Ghosh D, van Riper D and Manson SM 2010 How do you measure distance in spatial models? An
example using open-space valuation Environment and Planning B: Planning and Design 37(5) 874-894
Sener IN, Pendyala RM and Bhat CR 2011 Accommodating spatial correlation across choice alternatives in discrete
choice models: an application to modeling residential location choice behavior Journal of Transport
Geography 19 294-303
Srinivas N and Deb K 1994 Muiltiobjective Optimization Using Nondominated Sorting in Genetic Algorithms
Evolutionary Computation 2 (3) 221-248
Srour IM, Kockelman KM and Dunn TP 2002 Accessibility indices: a connection to residential land prices and
location choices Transportation Research Record 1805 25-34
Strauch D, Moeckel R, Wegener M, Grafe J, Muhlhans H and Rindsfuser G 2005 Linking transport and land use
planning: the microscopic dynamic simulation model ILUMASS in Atkinson PM, Foody GM, Darby SE
and Wu F, et al. eds. Geodynamics, CRC Press, Boca Raton 295-311
Tehran Municipality 2013a Tehran municipality statistical yearbook 2012-2013 ICT Organization of Tehran
Municipality Tehran, Iran
Tehran Municipality 2013b Tehran statistical yearbook 2012-2013 ICT Organization of Tehran Municipality
Tehran, Iran
Tsou KW, Hung YT and Chang YL 2005 An accessibility-based integrated measure of relative spatial equity in
urban public facilities Cities 22 (6) 424-435
Van Veldhuizen DA and Lamont GB 2000 Multiobjective evolutionary algorithms: analyzing the state-of-the-art
Evolutionary Computation 8 (2) 125-147
Veldhuisen KJ, Timmermans HJP and Kapoen LL 2000 RAMBLAS: a regional planning model based on the
microsimulation of daily activity travel patterns Environment and Planning A 32 427-443
Waddell P 2002 UrbanSim: Modeling urban development for land use, transportation, and environmental planning
Journal of the American Planning Association 68 (3) 297-314
Waddell P, Borning A, Noth M, Freier N, Becke M and Ulfarsson GF 2003 Microsimulation of urban development
and location choices: Design and implementation of UrbanSim Networks and Spatial Economics 3(1) 43-67
Wu W, Zhang W and Dong G 2013 Determinant of residential location choice in a transitional housing market:
Evidence based on micro survey from Beijing Habitat International 39 16-24
Zitzler E, Deb K and Thiele L 2000 Comparison of multiobjective evolutionary algorithms: Empirical results
Evolutionary Computation 8 (2) 173-195
Zitzler E and Thiele L 1998 Multiobjective optimization using evolutionary algorithms-A comparative case study
in Eiben AE, Back T, Schoenauer M and Schwefel HP eds. Parallel Problem Solving From Nature,
Springer-Verlag, Berlin 292-301
28
|
1906.11452 | 1 | 1906 | 2019-06-27T06:31:29 | Traffic Management Strategies for Multi-Robotic Rigid Payload Transport Systems | [
"cs.MA",
"cs.RO"
] | In this work, we address traffic management of multiple payload transport systems comprising of non-holonomic robots. We consider loosely coupled rigid robot formations carrying a payload from one place to another. Each payload transport system (PTS) moves in various kinds of environments with obstacles. We ensure each PTS completes its given task by avoiding collisions with other payload systems and obstacles as well. Each PTS has one leader and multiple followers and the followers maintain a desired distance and angle with respect to the leader using a decentralized leader-follower control architecture while moving in the traffic. We showcase, through simulations the time taken by each PTS to traverse its respective trajectory with and without other PTS and obstacles. We show that our strategies help manage the traffic for a large number of PTS moving from one place to another. | cs.MA | cs | Traffic Management Strategies for Multi-Robotic Rigid Payload
Transport Systems
Yahnit Sirineni, Pulkit Verma, Kamalakar Karlapalem
9
1
0
2
n
u
J
7
2
]
A
M
.
s
c
[
1
v
2
5
4
1
1
.
6
0
9
1
:
v
i
X
r
a
Abstract -- In this work, we address traffic management
of multiple payload transport systems comprising of non-
holonomic robots. We consider loosely coupled rigid robot
formations carrying a payload from one place to another. Each
payload transport system (PTS) moves in various kinds of
environments with obstacles. We ensure each PTS completes
its given task by avoiding collisions with other payload systems
and obstacles as well. Each PTS has one leader and multiple
followers and the followers maintain a desired distance and
angle with respect to the leader using a decentralized leader-
follower control architecture while moving in the traffic. We
showcase, through simulations the time taken by each PTS to
traverse its respective trajectory with and without other PTS
and obstacles. We show that our strategies help manage the
traffic for a large number of PTS moving from one place to
another.
I. INTRODUCTION
Multi-robot systems are a well studied problem in robotics
where applications like payload transportation [1][2][3], traf-
fic management [4][5], and area exploration [6][7] are most
commonly used. A system with multiple robots/agents offers
several advantages like low power consumption, increased
redundancy, and makes the system modular [8][9][10]. Han-
dling more number of robots poses different challenges on
the functionality of the system. Collision avoidance of a
system with other neighbourhood transport systems and other
obstacles in the environment is an important challenge. The
basic idea of collision avoidance is that a system derives
the control inputs based on its surroundings such that it
does not collide with any other system or obstacles while
performing its task. Work has been carried out on avoiding
the obstacles for a single or multiple robots [11], further
research is needed on obstacle avoidance for multiple robots
moving in a formation(s) of different shapes.
For payload carrying tasks in industries, assembly lines,
warehouses, etc., where payloads need to be transported from
one place to another without affecting the functionality of
other systems, is a challenging problem. Such applications
require smooth and collision free movement of these systems
such that
• A PTS does not collide with the other payload systems
in the surroundings.
• Avoids collisions with static obstacles.
Yahnit Sirineni and Pulkit Verma are students of Agents and Applied
Robotics Group(AARG), Kohli Center on Intelligent Systems(KCIS), In-
ternational Institute of Information Technology, Hyderabad (IIIT-H), India
[email protected]
Kamalakar Karlapalem is a faculty of Computer Science with
India
Information Technology, Hyderabad,
International
Institute of
[email protected]
• Reaches the destination by following the desired trajec-
tory.
In our work, we simulate multiple PTS carrying payloads
of different shapes (like triangular, square, circle, etc.) from
one point to another. Each system comprises of a group
of non-holonomic robots moving in a formation using a
decentralized leader-follower approach in [12]. One leader
and multiple followers are present
in each system. The
leader uses the trajectory information provided by our control
architecture which is derived from source to destination.
The followers generate their velocities using the leader
information which makes the formation intact all the time.
We show, through simulations, the time taken by a PTS to
move from one point to another in an environment with
(a) Only one payload transport system. (b) Multiple PTS
with a possibility of having a collision (c) Multiple PTS and
static obstacles in the environment.
We propose a modified version of Optimal Reciprocal
Collision Avoidance for non-holonomic robots (nh-ORCA)
[13] algorithm to identify if a PTS is prone to collide
with other formations and obstacles in the environment.
Our method provides the command velocities to the robots
in the formation such that no collision happens while the
transport system is approaching its goal and also ensures least
deviation possible to the existing path of the formation. We
present the simulation results for different formation shapes
with and without obstacles.
We discuss the related work done in this domain in Section
II. Our contribution in this paper is discussed in Section III.
The modeling and control of non-holonomic robots is carried
out in Section IV. Traffic Management strategies for multiple
PTS are discussed in Section V. To make our solution
more concrete, we display simulation results in Section VI.
Conclusion is discussed in Section VII
II. RELATED WORK
Over the years, various methods have been developed for
formation control and navigation of formations for payload
transportation. [14], [15] propose methods for path planning
of a single formation without laying much emphasis on the
presence of obstacles. [16], [17], [18], [19] shows navigation
of single formation in environments which consists of static
obstacles.Path planning for 3D formations is proposed in
[20] and [21]. Method for navigating a single formation
using the leader-follower approach is shown in [18] and
[22]. [23],[24] shows some promising results in aerial vehicle
formation control. A slung payload transportation method
is demonstrated in [25] where as [26] devised wheeled
locomotion for payload carrying with modular robot. [27]
proposes path planning of a single formation in environment
with dynamic obstacles but is a computationally expensive
centralized method.
To the best of our knowledge, no work has been done on
path planning and collision avoidance of multiple formations
of robots in environments with obstacles. We cannot consider
other formations as dynamic obstacles and use the approach
presented in [27] because the formations have a specific goal
of reaching the destination while dynamic obstacles do not
have a proper goal and hence show random behaviour. This
aspect makes our task even more challenging .
III. CONTRIBUTION
We propose a Leader-Follower-ORCA-RRT* framework
which is a decentralized method for navigation of multi-
ple formations of robots carrying payloads to the desired
destination while avoiding collisions with other formations
and obstacles. Given an environment
in which multiple
payloads are to be transported from one place to another
through formation of robots, we first compute the path of
the leader of the formation using RRT* [28] from respective
source to destination considering the static obstacles and
interpolate the path to get multiple way points. We modify
nh-ORCA [13] to compute collision-free linear and angular
velocity for the leader of the formation. A leader-follower
based decentralized control law is incorporated to compute
the corresponding command velocities of the followers in
the formation. Finally, We present simulation results of the
implemented control architecture to validate our method.
IV. BACKGROUND
A. Non-holonomic Robot Kinematics
Consider a non-holonomic mobile robot positioned at
(x, y) and oriented at and angle θ. Assuming the pseudo-
command velocities be v, ω, the robot kinematics is given
as
x
y
θ
=
cosθ −dsinθ
dcosθ
sinθ
0
1
(cid:20) v
ω
(cid:21)
where d is the distance of the robot's center to the center of
mass of the robot.
B. Leader-Follower Formation Control
Each PTS comprises of a number of non-holonomic
wheeled mobile robots which move on a trajectory while
maintaining a rigid formation. A decentralized leader-
follower approach control algorithm [12] is used which
include a leader and multiple followers. The followers use the
command velocities and odometery information of the leader
to derive their velocities such that the followers maintain
a specified distance and angle to remain in a formation.
Fig. 1 shows different shapes of the formation used in the
paper. The formation control law is given by the following
equations
Fig. 1: Leader follower formation with different shapes
vj = (k1αj + vicosθij − ρd
ijωisin(ψd
ijωicos(ψd
ij + θij) + k2βj + k3θje
d
visinθij + ρd
ij − θij)
ωj =
(1)
where αj and βj is the error in longitudinal and vertical
direction respectively. Constants k1, k2, k3, k4, k5, k6 > 0,
ij and ψd
ij are the desired distance and orientation to
ρd
maintain between the leader and follower robot, vi and ωi are
command velocities of the leader, vj and ωj are the generated
command velocities of the jth follower, d is distance from
the robot's center to the robot's center of mass. θij is the
orientation error and xje, yje, θje are the positional tracking
errors between the leader and follower.
C. Optimal Reciprocal Collision Avoidance
holonomic robots (nh-ORCA)
for non-
nh-ORCA [13] is a collision avoidance algorithm for
non-holonomic robots which is an extension to ORCA[11]
which deals with only holonomic robots. Using the algorithm
presented in [13], each robot independently computes its
velocity such that it is collision free for at least τ time
assuming that the other robots also compute their velocity
using the same method. There is no explicit communication
amongst the robots to choose their velocities, hence making
this is a robust decentralized system. Each robot i constructs
ij ∀ j (cid:54)= i which in turn is computed using velocity
ORCAτ
obstacles presented in [29].
Let pi, pj be the positions of two robots i and j and ri,
ij is the set
rj be the radius of the robots respectively, V Oτ
of velocities of robot i w.r.t robot j such that:
V Oτ
ij = {¯v ∃t ∈ [0, τ ] , t · ¯v ∈ D (pj − pi, ri + rj)} (2)
where
D (p, r) = {q (cid:107) q − p (cid:107)< r}
Then ORCAτ
ij is computed from Eq (3):
ij =(cid:8)vHi (cid:0)vHi −(cid:0)vopt
i + 0.5 ∗ u(cid:1)(cid:1) .n ≥ 0(cid:9) (3)
ORCAτ
V Oij at (vopt
where n denotes the outward normal of the boundary of
) + u, u is computed as shown in [11].
Now, the set of collision free velocities for robot i is given
-vopt
j
i
ORCAτ
ORCAτ
ij
(4)
by Eq (4).
i = SAHVi ∩(cid:92)
j(cid:54)=i
LF1L LF1L : LeaderF : FollowerF2F3F4F5F6F1F2F2F3F3F4 with SAHVi being the set of allowed velocities. The final
optimal velocity of the robot is given by equation (5).
v∗
Hi
= arg min
vHi∈ORCAτ
i
(cid:107) vHi − vpref
i
(cid:107)
(5)
We extend this approach to our problem for avoiding
collisions among other formations and obstacles.
V. TRAFFIC MANAGEMENT
Given an environment in which multiple payloads are
to be transported from one place to another with multi-
robot payload transport systems, we propose an efficient
traffic management strategies for path planning and collision
avoidance of formations carrying payloads. A bio inspired
neurodynamic based leader follower-ORCA-RRT* frame-
work is discussed, which computes a path to the respective
destination for all the leaders of PTS considering the static
obstacles using RRT* [28], then uses a modified version of
nh-ORCA [13] for collision avoidance amongst other forma-
tions and finally incorporates a bio inspired neurodynamic
based leader-follower formation control law for computing
the command velocities for the followers in the formation.
The architecture diagram has been shown in Fig.2
Algorithm 1: Leader Follower-ORCA-RRT*
1 Input : environment representing space and obstacles,
source and destination of the formations,
Configuration of each formation ;
2 for i = 1, i ≤ N o. of F ormations, i + + do
// Compute path from src to dest
for ith formation using RRT*
3
4
5
pathi = RRT*(sourcei, destinationi) ;
// Interpolate the path returned
pathi = Interpolate(pathi) ;
// Initialize nextDesti of formation
nextDesti = pathi[1]
// Initialize vpref
vpref
= (nextDesti - pi)
i
of formation
i
6
7 end
8 while All Formations have'nt reached their destinations
do
for i = 1, i ≤ N o. of F ormations, i + + do
timestep ++;;
if fi has not yet reached then
// Compute command velocities
of the leader
vi, ωi = Algorithm2(i);
// Compute command velocites
of the followers using
leader-follower control law
LeaderFollower(vi, ωi);
Apply corresponding controls to the leader
and followers of the formation;
9
10
11
12
13
14
end
15
16 end
i
i
, vmax
, neighboursi;
Algorithm 2: Modified nh-ORCA
1 Input : pathi, pi, vi, ri, vpref
2 for j in neighboursi do
3
4
5
6 end
7 Compute ORCAτ
8 Compute the optimal holonomic velocity of the leader
Compute V Oτ
Compute u from [11];
Compute ORCAτ
ij from Eq 3
i from Eq 2;
i from Eq 4;
Hi
v∗
from Eq(5);
leader from v∗
Hi
nh-ORCA [13];
9 Compute the optimal linear and angular velocity of the
for non-holonomic robot using
10 if distance(pi, nextDesti) ≤ δ then
// If the formation has almost
reached its nextDest, then update
nextDesti and vpref
i
11
Update nextDesti to next point of interpolated
path;
vpref
i
12
13 return non-holonomic command velocities of the
= ( nextDesti - pi );
Leader.
Fig. 2: Architecture Diagram of the framework
Multiple PTS are considered in this work. Each PTS is
a set of robots with a leader and its followers working
together to perform a particular task. Follower j of each
formation (fi) maintains a certain angle and distance relative
to its leader. Each formation fi has a current position
pi, radius ri, current velocity vi, number of followers
f ollowersi, source srci, destination desti, next immedi-
ate destination nextDesti, preferred velocity vpref
i.e the
velocity with which it would traverse had there been no
obstacles in its path, maximum velocity vmax
, neighbours
neighboursi which indicates the formations in its vicinity,
neighbourDisti which refers to the formations that are at a
distance less than this value.
i
i
We use the RRT* to find the preferred path from srci
to desti for each formation fi and interpolate the path
returned by RRT* to get multiple way points in the path. The
nextDesti for each formation i is set to the first point of
the interpolated path indicating that the formation's tentative
Formation Control(Formation 1)Refer Section (ⅣB)Formation Control(Formation 2)Refer Section (ⅣB)Formation Control(Formation 3)Refer Section (ⅣB)Modified Version of ORCA (Algorithm 2) Gives collision avoidance Velocity for the leader of the formation(Command Velocities of followers are computed using leader's velocity) RRT* andInterpolationTrajectory Information target is to reach the nextDesti and vpref
is set to (nextDesti - pi).
i
for each formation
For each timestep dt, new collision avoiding velocities
are computed for each of the leaders of the formations
through Algorithm 2. To compute the collision avoidance
velocities for each formation fi, consider all the neighbours
j in neighboursi for formation fi and compute V Oτ
ij and
ij from Eq.(2) and Eq(3) for all the neighbours j
ORCAτ
in neighboursi and compute u from [11]. Now, compute
i and v∗
from Eq(4) and Eq(5) to get the holonomic
ORCAτ
velocity of the leader of the formation fi. Now, map the
holonomic velocity to non-holonomic velocities as illustrated
in [13]. If the formation fi at a distance less than δ from
nextDesti where δ is a parameter to be set based on the
environment, then update the nextDesti to the next point of
the interpolated path and the preferred velocity of formation
vpref
to (nextDesti - pi). This modified version of nh-
i
ORCA is given in Algorithm 2.
Hi
Once the leader velocities of the formations are computed
using Eq.(2), command velocities of all the followers are
computed using Eq. (1) which are then applied to all the
robots. The complete algorithm is illustrated in Algorithm 1
VI. RESULTS
TABLE I: List of parameters for each robot
Parameter Name
k1
k2
k3
k4
k5
k6
δ
τ
dt
τobstacle
maxspeed (m/s)
neighbourDist
Value
1.5
1.0
0.025
15.0
1.0
1.0
1.3
11
5
4
0.0167
0.03
We showcase the simulation results of the presented al-
gorithm and discuss various aspects of the system. In the
simulations,we consider a Baseline scenario in which a pay-
load transport system is carrying a payload from one place
to another while there are no other transport systems and
static obstacles in the surroundings. We compare the baseline
results with (a) Multiple PTS moving in the environment
(b) Multiple PTS and static obstacles in the environment. We
compare the time taken, velocity variations and scalability
aspect of this approach with the baseline result.
A. Multiple Payload Transport Systems
In this case, we consider four payload transport systems
carrying a payload to their respective destinations. Each PTS
has to avoid the traffic as posed by the other transport sys-
tems while heading towards the destination simultaneously.
Different shapes (circle, square, triangle) are considered for
different PTS to make the problem more challenging. Each
formation has different number of robots ranging from 3−10.
However, the number of formations in an environment and
number of robot
limited and will
in a formation is not
Fig. 3: Four PTS are running avoiding collision with the
other systems and simultaneously reaching towards the goal
discussed in latter part of the section. Fig. 3 shows the
initial set of all the Payload transport systems present and
their respective goal points (denoted by the same color).
The simulation is symmetric and hence all formations travel
almost the same distance.
Fig. 4: Comparison of time taken by our system with and
without any traffic in the environment
We compare the time taken by the system in this envi-
ronment with a system having no traffic in the environment.
Fig. 4 shows that the time taken by the system in presence
of traffic is approximately equal to a system moving in no
traffic environment. This time is nearly equal, as all the PTS
create a symmetric trajectory and hence takes nearly equal
time to reach their destination.
It is important to note that all the payload transport sys-
tems maintain a minimum distance from each other ensuring
no collision while in motion. Fig. 5 shows the distance
maintained by each PTS to its closest neighbourhood. The
figure shows that all the PTS are maintaining a minimum
threshold and hence it is clear that the payload systems are
not colliding with each other. The change in the velocities
of each PTS when coming close to each other is shown in
Fig. 6. We can see that whenever the PTS are coming close
to each other, the velocities of the followers in the formation
experiences a changes such that the collision can be avoided.
Points A, B, C, D are marked in Fig. 5 and Fig. 6 which
can be used to relate the two graphs. We can observe that at
point B, all the formation are close to each other (see Fig.
(a)(b)(c)(d)(e)(f) figure shows that the distance of each followers from the
leader is approximately equal to 0.35, even when a system
avoids any collision with its neighbourhood.
B. Multiple Payload Transport Systems with Static Obstacles
To make our results more concrete, we consider an en-
vironment where there are multiple PTS are moving along
with the static obstacles between the start point and the desti-
nation. We consider four payload transport systems carrying
Fig. 5: Distance of each PTS with its closest neighbourhood.
3(f)) and hence suffers a velocity change in the robots of
all the PTS. Similarly, point A, C are the points where the
PTS started changing the trajectory to successfully avoid the
obstacles.
Fig. 8: Four PTS are running avoiding collision with static
obstacles and other transport systems
a payload. Five static obstacles are considered at random
places in the arena. Each payload transport system receives
the trajectory information using Algorithm 1 such that there
is no collision amongst the PTS and static obstacles. Fig.
8 shows a initial set up of all formations and the obstacles
present in the surroundings. We observe that all the systems
successfully avoid collision with the obstacles.
Fig. 6: Velocities of all the robots in each PTS.
Fig. 7: Distance of all the robots in each formation
While avoiding traffic in the surroundings, we ensure
that each payload transport system maintains the robots
formation using the leader-follower based approach. Fig. 7
depicts that the followers in each formation nearly maintains
the desired distance and orientation from the leader. For
simulation, we have considered this distance constant for all
the followers in all the PTS, which is equal to 0.35m. The
Fig. 9: Comparison of time taken by our system with and
without any traffic and static obstacles in the environment
We compare the time taken by the PTS to reach to their
respective destinations with and without obstacles. Fig. 9
shows that the time taken by the PTS in presence of obstacles
is comparable to the time taken by the PTS in no obstacles
scenario. We conclude that an optimal collision avoiding path
is provided by our control algorithm.
In this case, we plot the minimum distance of a PTS which
is to be maintained wrt. neighbourhood PTS (similar to the
0100200300400500Time steps (in seconds)0.02.55.07.510.012.515.0Distance (in meters)ABCD1 meter(a) Formation 1Formation 2Formation 3Formation 4Threshold0100200300400500Time steps (in seconds)0.02.55.07.510.012.515.0Distance (in meters)ABCD1 meter(b) Formation 2Formation 1Formation 3Formation 4Threshold0100200300400500Time steps (in seconds)0.02.55.07.510.012.515.0Distance (in meters)ABCD1 meter(c) Formation 3Formation 1Formation 2Formation 4Threshold0100200300400500Time steps (in seconds)0.02.55.07.510.012.515.0Distance (in meters)ABCD1 meter(d) Formation 4Formation 1Formation 2Formation 3ThresholdDistance of each formation to its closest formation0100200300400500Time steps (in seconds)0.20.10.00.10.2Velocity (m/s)ABCD(a) Formation 1leaderFollower 1Follower 2Follower 3Follower 40100200300400500Time steps (in seconds)0.20.10.00.10.2Velocity (m/s)ABCD(b) Formation 2leaderFollower 1Follower 2Follower 3Follower 40100200300400500Time steps (in seconds)0.20.10.00.10.2Velocity (m/s)ABCD(c) Formation 3leaderFollower 1Follower 2Follower 3Follower 40100200300400500Time steps (in seconds)0.20.10.00.10.2Velocity (m/s)ABCD(d) Formation 4leaderFollower 1Follower 2Follower 3Follower 4Velocities of all the followers in each Formation0100200300400500Time steps(in seconds)0.00.10.20.30.40.5Distance from Leader(a) Formation 1Follower 1Follower 2Follower 3Follower 40100200300400500Time steps(in seconds)0.00.10.20.30.40.5Distance from Leader(b) Formation 2Follower 1Follower 2Follower 3Follower 40100200300400500Time steps(in seconds)0.00.10.20.30.40.5Distance from Leader(c) Formation 3Follower 1Follower 2Follower 3Follower 40100200300400500Time steps(in seconds)0.00.10.20.30.40.5Distance from Leader(d) Formation 4Follower 1Follower 2Follower 3Follower 4Desired Distance to maintain = 0.35m(a)(b)(c)(d)(e)(f) Fig. 10: Distance of each PTS with its closest neighbourhood
in presence of static obstacles.
Fig. 12: Distance of all the robots in each formation in
presence of static obstacles
previous case with no static obstacle). This will make sure
that the formations do not collide with each other. Fig. 10
shows the minimum distance maintained by each PTS with
its closest neighbourhood PTS. We present a change in the
velocities of each PTS when coming close to each other Fig.
11. We observe, a PTS changes its velocities (as generated
by the nh-ORCA) such that it avoids any collision with other
obstacles. As discussed in the above case, we mark points
in Fig. 10 and Fig. 11 to relate the change in velocities of
the formation in the vicinity of the other formations or static
obstacles. For example, formation 2 and formation 4 come
close to each other at Point B and the velocity of both the
formation also changes at the same to avoid any obstacle
(see Fig. 11).
through simulations, that our system works well with 30
different payload transport systems with variable number of
robots in each of them. But, in general, it can be extended to
any number of PTS. Fig. 13 shows that all the PTS reach their
destination to complete a given task without colliding with
any of the neighbourhood PTS. Thus our system is more
robust and reliable to work with large number of PTS in
an industrial/warehouse environment. We have added more
detailed results and simulations of the scalability in the video
uploaded.
Fig. 11: Robots velocities of all PTS with static obstacles.
It
is even more challenging in this case to maintain
the formation for each PTS as the static obstacles in the
surroundings make it more difficult to maintain the desired
distance and angles of the follower with the leader robot. Fig.
12 highlights that the followers in each PTS is maintaining
the specified parameters with the leader when avoiding the
obstacles. For simulation of this scenario also, we have taken
the distance of followers w.r.t to leader to be 0.35m.
C. Scalability analysis of the system
Our system scales well with the increase in number of
PTS, hence making it more efficient and robust. We showcase
Fig. 13: Thirty different PTS are moving to carry a payload
to their respective destination without any collision.
VII. CONCLUSION
We simulate the traffic management strategies for multiple
loosely coupled payload transport system. Each PTS com-
prises of a number of non-holonomic mobile robots which
are moving in a formation and carrying a payload from
one place to another. The formation control is done using
a decentralized leader-follower based algorithm. A modified
version of decentralized collision avoidance algorithm nh-
ORCA is used to compute the collision-free velocities for the
leader of the formation. The followers follow the leader such
that a rigid formation is maintained while simultaneously
moving towards a goal. Various environments are considered
for the simulation which includes (a) Multiple PTS are
running with no other obstacles (b) Multiple PTS are running
with static obstacles. A comparison is made with a baseline
scenario where individual PTS are moving to reach a goal.
0100200300400500600Time steps (in seconds)0.02.55.07.510.012.515.0Distance (in meters)ABCD1 meter(a) Formation 1Formation 2Formation 3Formation 4Threshold0100200300400500600Time steps (in seconds)0.02.55.07.510.012.515.0Distance (in meters)ABC1 meter(b) Formation 2Formation 1Formation 3Formation 4Threshold0100200300400500600Time steps (in seconds)0.02.55.07.510.012.515.0Distance (in meters)ABC1 meter(c) Formation 3Formation 1Formation 2Formation 4Threshold0100200300400500600Time steps (in seconds)0.02.55.07.510.012.515.0Distance (in meters)ABCD1 meter(d) Formation 4Formation 1Formation 2Formation 3ThresholdDistance of each formation to its closest Formation0100200300400500600Time steps (in seconds)0.20.10.00.10.2Velocity (m/s)ABCD(a) Formation 1leaderFollower 1Follower 2Follower 3Follower 4Follower 5Follower 6Follower 70100200300400500600Time steps (in seconds)0.20.10.00.10.2Velocity (m/s)ABC(b) Formation 2leaderFollower 1Follower 2Follower 3Follower 40100200300400500600Time steps (in seconds)0.20.10.00.10.2Velocity (m/s)ABC(c) Formation 3leaderFollower 1Follower 2Follower 3Follower 40100200300400500600Time steps (in seconds)0.20.10.00.10.2Velocity (m/s)ABCD(d) Formation 4leaderFollower 1Follower 2Follower 3Follower 4Follower 5Follower 6Velocities of all the followers in each Formation0100200300400500600Time steps0.00.10.20.30.40.5Distance from Leader(a) Formation 1Follower 1Follower 2Follower 3Follower 4Follower 5Follower 6Follower 70100200300400500600Time steps0.00.10.20.30.40.5Distance from Leader(b) Formation 2Follower 1Follower 2Follower 3Follower 40100200300400500600Time steps0.00.10.20.30.40.5Distance from Leader(c) Formation 3Follower 1Follower 2Follower 3Follower 40100200300400500600Time steps0.00.10.20.30.40.5Distance from Leader(d) Formation 4Follower 1Follower 2Follower 3Follower 4Follower 5Follower 6Desired Distance to maintain = 0.35m(a)(b)(c)(d) REFERENCES
[1] P. Verma, R. Tallamraju, A. Rawat, S. Chand, and K. Karlapalem,
"Loosely coupled payload transport system with robot replacement,"
in Autonomous Robots and Multi-Robot Systems (ARMS) with AAMAS,
2019.
[2] J. Alonso-Mora, S. Baker, and D. Rus, "Multi-robot formation control
and object transport in dynamic environments via constrained opti-
mization," The International Journal of Robotics Research, vol. 36,
no. 9, pp. 1000 -- 1021, 2017.
[3] Z. Wang and M. Schwager, "Multi-robot manipulation without com-
munication," in Distributed Autonomous Robotic Systems, pp. 135 --
149, Springer, 2016.
[4] H.-z. Zhuang, S.-x. Du, and T.-j. Wu, "Research on path planning and
related algorithms for robots [j]," Bulletin of Science and Technology,
vol. 3, 2004.
[5] M. Vestal, M. Lafary, and P. Stopera, "Job management system for a
fleet of autonomous mobile robots," Dec. 11 2014. US Patent App.
14/370,383.
[6] A. Baranzadeh and A. V. Savkin, "A distributed control algorithm for
area search by a multi-robot team," Robotica, vol. 35, no. 6, pp. 1452 --
1472, 2017.
[7] S. Sharma, A. Shukla, and R. Tiwari, "Multi robot area exploration
using nature inspired algorithm," Biologically Inspired Cognitive Ar-
chitectures, vol. 18, pp. 80 -- 94, 2016.
[8] Y. U. Cao, A. S. Fukunaga, and A. Kahng, "Cooperative mobile
robotics: Antecedents and directions," Autonomous robots, vol. 4,
no. 1, pp. 7 -- 27, 1997.
[9] L. E. Parker and F. Tang, "Building multirobot coalitions through
automated task solution synthesis," Proceedings of the IEEE, vol. 94,
no. 7, pp. 1289 -- 1305, 2006.
[22] A. N. Asl, M. B. Menhaj, and A. Sajedin, "Control of leader --
follower formation and path planning of mobile robots using asexual
reproduction optimization (aro)," Applied Soft Computing, vol. 14,
pp. 563 -- 576, 2014.
[23] J. Alonso-Mora, E. Montijano, M. Schwager, and D. Rus, "Distributed
multi-robot formation control among obstacles: A geometric and
optimization approach with consensus," in 2016 IEEE international
conference on robotics and automation (ICRA), pp. 5356 -- 5363, IEEE,
2016.
[24] D. Zhou, Z. Wang, and M. Schwager, "Agile coordination and assistive
collision avoidance for quadrotor swarms using virtual structures,"
IEEE Transactions on Robotics, vol. 34, no. 4, pp. 916 -- 923, 2018.
[25] M. Bernard and K. Kondak, "Generic slung load transportation system
using small size helicopters," in 2009 IEEE International Conference
on Robotics and Automation, pp. 3258 -- 3264, IEEE, 2009.
[26] F. Hou, N. Ranasinghe, B. Salemi, and W.-M. Shen, "Wheeled loco-
motion for payload carrying with modular robot," in 2008 IEEE/RSJ
International Conference on Intelligent Robots and Systems, pp. 1331 --
1337, IEEE, 2008.
[27] J. Alonso-Mora, S. Baker, and D. Rus, "Multi-robot formation control
and object transport in dynamic environments via constrained opti-
mization," The International Journal of Robotics Research, vol. 36,
no. 9, pp. 1000 -- 1021, 2017.
[28] S. Karaman and E. Frazzoli, "Sampling-based algorithms for optimal
journal of robotics research,
motion planning," The international
vol. 30, no. 7, pp. 846 -- 894, 2011.
[29] P. Fiorini and Z. Shiller, "Motion planning in dynamic environments
using velocity obstacles," The International Journal of Robotics Re-
search, vol. 17, no. 7, pp. 760 -- 772, 1998.
[10] X. Yang, K. Watanabe*, K. Izumi, and K. Kiguchi, "A decentral-
ized control system for cooperative transportation by multiple non-
holonomic mobile robots," International Journal of Control, vol. 77,
no. 10, pp. 949 -- 963, 2004.
[11] J. Van Den Berg, S. J. Guy, M. Lin, and D. Manocha, "Reciprocal
n-body collision avoidance," in Robotics research, pp. 3 -- 19, Springer,
2011.
[12] Z. Peng, G. Wen, A. Rahmani, and Y. Yu, "Leader -- follower formation
control of nonholonomic mobile robots based on a bioinspired neuro-
dynamic based approach," Robotics and autonomous systems, vol. 61,
no. 9, pp. 988 -- 996, 2013.
[13] J. Alonso-Mora, A. Breitenmoser, M. Rufli, P. Beardsley, and
R. Siegwart, "Optimal reciprocal collision avoidance for multiple
non-holonomic robots," in Distributed Autonomous Robotic Systems,
pp. 203 -- 216, Springer, 2013.
[14] S. Liu, D. Sun, and C. Zhu, "Coordinated motion planning for multiple
mobile robots along designed paths with formation requirement,"
IEEE/ASME transactions on mechatronics, vol. 16, no. 6, pp. 1021 --
1031, 2011.
[15] E. Pereyra, G. Aragu´as, and M. Kulich, "Path planning for a formation
of mobile robots with split and merge," in International Conference
on Modelling and Simulation for Autonomous Systems, pp. 59 -- 71,
Springer, 2017.
[16] Y. Shapira and N. Agmon, "Path planning for optimizing survivabil-
ity of multi-robot formation in adversarial environments," in 2015
IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS), pp. 4544 -- 4549, IEEE, 2015.
[17] L. Polkowski and P. Osmialowski, "Navigation for mobile autonomous
robots and their formations: An application of spatial reasoning
induced from rough mereological geometry," in Mobile robots nav-
igation, InTech, 2010.
[18] S. Garrido, L. Moreno, J. V. Gomez, and P. U. Lima, "General
path planning methodology for leader-follower robot formations,"
International Journal of Advanced Robotic Systems, vol. 10, no. 1,
p. 64, 2013.
[19] R. Gautam, R. Kala, et al., "Motion planning for a chain of mobile
robots using a* and potential field," Robotics, vol. 7, no. 2, p. 20,
2018.
´Alvarez, J. V. G´omez, S. Garrido, and L. Moreno, "3d robot
formations path planning with fast marching square," Journal of
Intelligent & Robotic Systems, vol. 80, no. 3-4, pp. 507 -- 523, 2015.
[20] D.
[21] Q. Zuo, M. Chu, Y. Ding, L. Ma, and H. Sun, "A new path planning
method for multi-robot formation in three-dimensional space," in
2016 4th International Conference on Sensors, Mechatronics and
Automation (ICSMA 2016), Atlantis Press, 2016.
|
1711.07951 | 1 | 1711 | 2017-11-21T18:22:05 | Cellular Automata Simulation on FPGA for Training Neural Networks with Virtual World Imagery | [
"cs.MA"
] | We present ongoing work on a tool that consists of two parts: (i) A raw micro-level abstract world simulator with an interface to (ii) a 3D game engine, translator of raw abstract simulator data to photorealistic graphics. Part (i) implements a dedicated cellular automata (CA) on reconfigurable hardware (FPGA) and part (ii) interfaces with a deep learning framework for training neural networks. The bottleneck of such an architecture usually lies in the fact that transferring the state of the whole CA significantly slows down the simulation. We bypass this by sending only a small subset of the general state, which we call a 'locus of visibility', akin to a torchlight in a darkened 3D space, into the simulation. The torchlight concept exists in many games but these games generally only simulate what is in or near the locus. Our chosen architecture will enable us to simulate on a micro level outside the locus. This will give us the advantage of being able to create a larger and more fine-grained simulation which can be used to train neural networks for use in games. | cs.MA | cs | Cellular Automata Simulation on FPGA for Training
Neural Networks with Virtual World Imagery
Olivier Van Acker and Oded Lachish
Department of Computer Science and Information Systems
Birkbeck, University of London
London, United Kingdom
[email protected], [email protected]
Graeme Burnett
Enhyper Ltd.
London, United Kingdom
[email protected]
7
1
0
2
v
o
N
1
2
]
A
M
.
s
c
[
1
v
1
5
9
7
0
.
1
1
7
1
:
v
i
X
r
a
Abstract-We present ongoing work on a tool that consists
of two parts: (i) A raw micro-level abstract world simulator
with an interface to (ii) a 3D game engine,
translator of
raw abstract simulator data to photorealistic graphics. Part (i)
implements a dedicated cellular automata (CA) on reconfigurable
hardware (FPGA) and part (ii) interfaces with a deep learning
framework for training neural networks. The bottleneck of such
an architecture usually lies in the fact that transferring the state
of the whole CA significantly slows down the simulation. We
bypass this by sending only a small subset of the general state,
which we call a 'locus of visibility', akin to a torchlight in a
darkened 3D space, into the simulation. The torchlight concept
exists in many games but these games generally only simulate
what is in or near the locus. Our chosen architecture will enable
us to simulate on a micro level outside the locus. This will give
us the advantage of being able to create a larger and more fine-
grained simulation which can be used to train neural networks
for use in games.
Index Terms-Cellular Automata; FPGA; Simulation; Ma-
chine learning; Neural networks; Unreal Engine;
I. BACKGROUND
There have been exciting new results of training neural
networks with photo realistic imagery from virtual worlds [1].
The training of these neural networks uses rendered images
from virtual worlds instead of real world data, the two biggest
advantages of this approach being, firstly, fewer limitations
in executing potentially difficult or dangerous scenarios, and
secondly, the ability to accelerate the speed of the simulation
means faster training of the neural network.
In recent years, the game industry has spent a lot of effort
on creating game engines which can output near-photorealistic
imagery in real
time, making it possible to train neural
networks for real world scenarios using this output. Several
projects are being developed to make it easier for neural
network frameworks to interface with these engines [2], [3].
Cellular Automata (CA) is an effective technique for simu-
lating, on a micro level, complex behavior such as pedestrian
traffic, moving agents [4] or, as in our proof of concept, the
traffic of narrowboats1 on a system of canals. With a simple
set of rules, contained in each cell, describing the behavior of
passing agents, it is possible to get an extraordinarily complex
macroscopic view of the flow of traffic [4].
1Narrowboats were the main transportation system for goods at the start of
the industrial revolution in the UK
CA, because of its inherent massive spatial parallelism,
locality and discrete nature, is a perfect candidate for imple-
mentation on programmable FPGA (field-programmable gate
array) technology [5]. FPGAs are reconfigurable hardware
devices, where the set of rules contained in each cell can be
described in hardware via lookup tables and flip-flops (for
storing state), and with every clock cycle the state of all cells
can be updated in parallel.
The remainder of this paper describes our current work, in
which we are implementing a CA for simulating traffic on an
FPGS, to train neural networks. These trained neural networks
can be used to create game environments with 'real world'-like
behavior.
II. IMPLEMENTATION OF CA MICROSIMULATION ON FPGA
We are developing a tool which microsimulates traffic in a
virtual world and gives a game engine a limited view of certain
areas of the world a 'locus of visibility'. By limiting the
amount of data made available for rendering and subsequent
learning of a neural network, we can increase the size and
granularity of the simulation, which will make the macro view
more realistic.
The data of this locus of view, akin to a torchlight in a
darkened 3D space, or a traffic camera used to monitor traffic
on a busy crossing in a city, will be passed on to a game
engine which will generate a photorealistic video feed of the
exposed area. This video feed will be used for training a neural
network. The simulation will run several times faster than it
would normally do when a game is played, to speed up the
training of the neural network.
The simulation uses the cellular automata (CA) method and
a 'locus of visibility controller' extracts the localized data from
the simulation and exposes it over the network to one or more
consumers.
The CA will be implemented on a FPGA and communicate
over PCIe to a network card (NIC), exposing it over the
network. The Unreal Engine captures this data and a neural
network interfaces with the game engine via UnrealCV [2], an
©2017 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including
reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or
reuse of any copyrighted component of this work in other works. DOI: 10.1109/CIG.2017.8080450
A. Proof of concept: Narrowboat Simulator
In our first proof of concept we will simulate the traffic of
a large number of narrowboats on an intricate canal system,
transporting resources (for example, coal or grain) between
supply points and delivery points in competing economic
areas. The flow and density of traffic can be regulated via
locks (chambered gates used to raise or lower water level,
allowing boats to move to higher or lower levels of a canal),
distributed throughout the canal system.
The neural network first trains itself by observing the traffic
at several locks. Once trained, the neural network will be
used to play a game in which it competes against a (human)
opponent, to bring as many resources as possible to its own
delivery points, manipulating the traffic by operating the locks.
FUTURE WORK
The implementation of the narrowboat simulator will be a
starting point from which to build more complex simulations
of growing cities with different transportation systems, inter-
acting with each other. This city transport simulator can then
be used to train neural networks to operate different aspects
of the simulation, for example, resource management. Both
trained neural networks and simulator can be used to create
games in which the environments and elements within them
exhibit more complex, 'real world'-like behavior.
REFERENCES
[1] "Artificial
intelligence: Why
video
The
[Online].
http://www.economist.com/news/science-and-technology/
games
Available:
21721890-games-help-them-understand-reality-why-ai-researchers-video-games
researchers
May
Economist,,"
AI
like
2017.
[2] W. Qiu and A. Yuille, "UnrealCV: Connecting Computer Vision to
Unreal Engine," arXiv:1609.01326 [cs], Sep. 2016, arXiv: 1609.01326.
[Online]. Available: http://arxiv.org/abs/1609.01326
[3] H. Kinsley, "Explorations of Using Python to play Grand Theft Auto 5."
2017.
[4] V.
J. Blue and J. L. Adler, "Cellular automata microsimulation
for modeling bi-directional pedestrian walkways," Transportation
Research Part B: Methodological, vol. 35, no. 3, pp. 293–312, Mar.
2001. [Online]. Available: http://www.sciencedirect.com/science/article/
pii/S0191261599000521
[5] M. Halbach and R. Hoffmann, "Implementing cellular automata in
FPGA logic," in 18th International Parallel and Distributed Processing
Symposium, 2004. Proceedings., Apr. 2004, pp. 258–.
[6] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick,
S. Guadarrama, and T. Darrell, "Caffe: Convolutional Architecture
for Fast Feature Embedding," in Proceedings of
the 22Nd ACM
International Conference on Multimedia, ser. MM '14. New York,
NY, USA: ACM, 2014, pp. 675–678.
[Online]. Available: http:
//doi.acm.org/10.1145/2647868.2654889
Fig. 1. Architecture of the tool
Unreal Engine plugin for interfacing with deep learning
frameworks such as Caffe [6].
The simulated world is split up into hexagons and each in-
dividual area has a fixed locality to adjacent hexagons and can
hold multiple agents. Each hexagon is represented by a cell
in the CA. The movement of agents and the state of the cell is
determined by the rule set contained in every individual cell.
|
1003.0617 | 1 | 1003 | 2010-03-02T15:38:48 | Agent Based Approaches to Engineering Autonomous Space Software | [
"cs.MA",
"cs.AI"
] | Current approaches to the engineering of space software such as satellite control systems are based around the development of feedback controllers using packages such as MatLab's Simulink toolbox. These provide powerful tools for engineering real time systems that adapt to changes in the environment but are limited when the controller itself needs to be adapted.
We are investigating ways in which ideas from temporal logics and agent programming can be integrated with the use of such control systems to provide a more powerful layer of autonomous decision making. This paper will discuss our initial approaches to the engineering of such systems. | cs.MA | cs | Agent Based Approaches to Engineering Autonomous Space
Software∗
Louise A. Dennis
Department of Computer Science, University of Liverpool, UK
[email protected]
Michael Fisher
Department of Computer Science, University of Liverpool, UK
Nicholas Lincoln
Department of Engineering, University of Southampton, UK
Alexei Lisitsa
Department of Computer Science, University of Liverpool, UK
Sandor M. Veres
Department of Engineering, University of Southampton, UK
Current approaches to the engineering of space software such as satellite control systems are based
around the development of feedback controllers using packages such as MatLab's Simulink tool-
box. These provide powerful tools for engineering real time systems that adapt to changes in the
environment but are limited when the controller itself needs to be adapted.
We are investigating ways in which ideas from temporal logics and agent programming can be
integrated with the use of such control systems to provide a more powerful layer of autonomous
decision making. This paper will discuss our initial approaches to the engineering of such systems.
1 Introduction
Modern control systems are limited in their ability to react flexibly and autonomously to changing situ-
ations. The limiting factor is the complexity inherent in analysing situations where many variables are
present. There are many complex, real-world, control systems but we are primarily interested in the
(autonomous) control of satellite systems.
Consider the problem of a single satellite attempting to maintain a geostationary orbit. Current
satellite control systems maintain orbits using feedback controllers. These implicitly assume that any
errors in the orbit will be minor and easily corrected. In situations where more significant errors occur,
for example caused by a thruster malfunction, it is desirable to modify or change the controller. The
complexity of the decision task is a challenge to standard approaches, and has led, for example, to
complex, evolutionary control systems. These become very difficult to understand.
We approach the problem from the perspective of rational agents [6]. We consider a satellite to be
an agent which consists of a discrete (rational decision making) part and a continuous (calculation) part.
The discrete part uses the Belief-Desire-Intention (BDI) theory of agency [5] and governs high level
decisions about when to generate new feedback controllers. The continuous, calculational part is used to
∗Work funded by EPSRC grants EP/F037201/1 and EP/F037570/1
M. Bujorianu and M. Fisher (Eds.):
Workshop on Formal Methods for Aerospace (FMA)
EPTCS 20, 2010, pp. 63–67, doi:10.4204/EPTCS.20.6
c(cid:13) L. A. Dennis et al.
This work is licensed under the
Creative Commons Attribution License.
64
Agent Based Approaches to Engineering Autonomous Space Software
derive controllers and to calculate information from continuous data which can be used in the decision
making process; this part can be viewed as a hybrid system.
2 Architecture
Figure 1: Implemented Hybrid Agent Architecture
Our prototype system is shown in Fig. 1. We have implemented a simulated environment and real
time satellite control system in MatLab using the Simulink tool kit. The continuous agent part is also
implemented in MatLab. MatLab has no easy provision for threaded execution which forces us to use
separate instances for the Real Time aspects (i.e. the feedback controller and simulated environment)
and for the Continuous Agent part. The agent also contains a discrete agent part which is currently
implemented in the Gwendolen agent programming language1. Gwendolen [2] is implemented on top of
JAVA.
The real time control system sends information (which may be pre-processed) to the agent part of the
system. When it acts, the discrete part of the agent may either cause the continuous agent part to perform
some calculation (and wait for the results) or it may send an instruction to the real time control system to
alter its controller. Since the new controller has been created "on the fly" by the continuous part, some
aspects of this controller are stored in the shared file system (accessed by both MatLab processes).
The discrete agent part is divided into an abstraction engine which takes continuous data supplied
by the satellite simulation and transforms this data into discrete shared beliefs which are accessed by a
1The choice of language was dictated entirely by convenience. It is a subject for further work to examine more widely used
BDI-languages and evaluate which is most appropriate for the system.
AgentPartMATLABJAVAMODEL OF SYTEMRational Agent PartThin Environment LayerAGENTFile SystemReal Time ControlSend DataActData FlowControl FlowAbstraction EngineShared BeliefsReasoningReasonActSenseContinuous Calculate L. A. Dennis et al.
65
reasoning engine which makes decisions about how to behave. The discrete part is split in two because
reasoning is comparatively slow compared to the flow of data coming in from the simulation. It can
become "clogged" up with the need to react to changing information if it tries to perform both the
abstraction tasks and the reasoning tasks at once. The separation of abstraction and reasoning is both
theoretically clean and practical at an implementational level.
3 BDI Programming Aspects
The architecture lets us represent the high-level decision making aspects of the program in terms of the
beliefs and goals of the agent and the events it observes. So, for instance, when the agent observes the
event that the satellite is in a new position (information relayed to it by the real time controller) it can call
on the continuous part to calculate whether this position is within acceptable bounds of the desired orbit
(i.e. whether the existing real-time controller is capable of maintaining the orbit). If, as a result of this, it
gains a belief that the satellite has strayed from the orbit it can request the continuous part to calculate a
new path for the satellite to follow using techniques described in [4].
Similarly, if the satellite has strayed from its bounds, the discrete agent part can examine its beliefs
about the current status of the thrusters and, if necessary, instruct the continuous part to generate a new
feedback controller which takes into account any malfunctions or inefficiencies in the thrusters.
Such programs can be expressed compactly in the BDI-programming style without the need for
programming large decision trees to consider all possible combinations of thruster status and satellite
positions. This should then reduce the probability of error in the decision-making parts of the program
and opens the possibility that existing techniques for model checking such programs [1] can be adapted
to verify this part.
3.1 Geostationary Orbit Case Study
The agent code for the geostationary orbit is shown in code fragments 3.1 and 3.2. Fragment 3.1 shows
the code for the abstraction engine. Every time it "perceives" the satellite position (stateinfo) it calls
upon MatLab to calculate whether or not this position is within bounds (comp_distance) and then
asserts and removes shared beliefs appropriately.
The code is shown as a series of plans of the form trigger:{guard} ← deeds where the trigger is
some event observed by the agent, the guard is a set of facts that must be true before the plan is activated
and the deeds are a stack of deeds to be executed. +b is the addition of a belief, b, and -b is the removal
of the belief, b. In a guard B b means that b is believed.
Code fragment 3.1 Geostationary Orbit Control (Abstraction Engine)
+stateinfo(L1, L2, L3, L4, L5, L6) :
{B proximity_to_centre(V1)} ←
comp_distance(L1, L2, L3, L4, L5, L6, Val),
+proximity_to_centre(Val);
+proximity_to_centre(in) : {B proximity_to_centre(out)} ←
-proximity_to_center(out),
remove_shared(proximity_to_centre(out)),
assert_shared(proximity_to_centre(in));
+proximity_to_centre(out) :
{B proximity_to_centre(in),
1
2
3
4
5
6
7
8
9
10
11
12
66
Agent Based Approaches to Engineering Autonomous Space Software
B stateinfo(L1, L2, L3, L4, L5, L6)} ←
-proximity_to_centre(in),
remove_shared(stateinfo(A1, A2, A3, A4, A5, A6)),
assert_shared(stateinfo(L1, L2, L3, L4, L5, L6)),
remove_shared(proximity_to_centre(in)),
assert_shared(proximity_to_centre(out));
13
14
15
16
17
18
Fragment 3.2 reacts to the dynamic information about whether the satellite is within bounds or not.
It may call a MatLab function, plan_approach_to_centre which returns the name of a plan to
move a satellite back within bounds. apply_controls and maintain_path are actions applied to the
simulation of the satellite which apply a named plan, or continue normal operation as appropriate. The
syntax +!g indicates the acquisition of a goal.
Code fragment 3.2 Geostationary Orbit Control
+proximity_to_centre(out) : {(cid:62)} ←
-proximity_to_centre(in),
+!get_to_centre;
+proximity_to_centre(in) : {(cid:62)} ←
-proximity_to_centre(out),
maintain_path;
+!get_to_centre :
{B proximity_to_centre(out),
B stateinfo(L1, L2, L3, L4, L5, L6)} ←
plan_approach_to_centre(P, locn(L1, L2, L3, L4, L5, L6)),
+!try_execute(P) ;
+!try_execute(P) : {B proximity_to_centre(out)} ←
apply_controls(P);
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
3.2 Decision and Control
The important aspect of both the above example and the architecture in general is that the (MatLab) con-
trol systems take care of the detailed calculation of continuous functions (paths, etc), while the rational
agent takes care of high-level decisions about targets and plans. This separation of concerns simplifies
both parts and avoids the problems associated with large, opaque, complex, adaptive and evolutionary
control systems.
4 Future Work
We are currently working on our prototype system and case study which will allow us to make com-
parisons of this agent approach to autonomous decision-making in satellite systems to approaches based
on finite state machines and standard control. We also are interested in investigating the use of tempo-
ral logic and model checking to generate forward planning capabilities for the agent along the lines of
those investigated by Kloetzer and Belta [3]. We aim to explore the possibility of using model checking
to verify aspects of the agent's behaviour. Given that we already have a formal verification system for
Gwendolen agents [1], there is a strong possibility that we can extend this to cope with (abstractions
of) the continuous part. As the diagram below shows, we already have model-checking tools for the
L. A. Dennis et al.
67
discrete/finite parts. Our interest now is how far such techniques can be extended to account for other
aspects of the agent's behaviour.
References
[1] R. H. Bordini, L. A. Dennis, B. Farwer, and M. Fisher. Automated verification of multi-agent programs.
In Proceedings of the 23rd IEEE/ACM International Conference on Automated Software Engineering (ASE
2008), pages 69–78, L'Aquila, Italy, September 2008.
[2] L. A. Dennis and B. Farwer. Gwendolen: A BDI Language for Verifiable Agents. In B. Lowe, editor, Proc.
AISB'08 Workshop on Logic and the Simulation of Interaction and Reasoning, Aberdeen, 2008. AISB.
[3] M. Kloetzer and C. Belta. A Fully Automated Framework for Control of Linear Systems From Temporal
Logic Specifications. IEEE Transactions on Automatic Control, 53(1):287–297, 2008.
[4] N. Lincoln and S. Veres. Components of a Vision Assisted Constrained Autonomous Satellite Formation
Flying Control System. International Journal of Adaptive Control and Signal Processing, 21(2-3):237–264,
October 2006.
[5] A. S. Rao and M. Georgeff. BDI Agents: From Theory to Practice. In Proc. 1st International Conference on
Multi-Agent Systems (ICMAS), pages 312–319, June 1995.
[6] M. Wooldridge and A. Rao, editors. Foundations of Rational Agency. Applied Logic Series. Kluwer Academic
Publishers, Mar. 1999.
Discrete and FiniteAgentPartMATLAB"RealWorld"JAVARational Agent PartThin Environment LayerAbstraction EngineShared BeliefsReasoningContinuous |
0909.3475 | 1 | 0909 | 2009-09-18T15:58:43 | Multi-agent Coordination in Directed Moving Neighborhood Random Networks | [
"cs.MA",
"cs.IT",
"cs.IT"
] | In this paper, we consider the consensus problem of dynamical multiple agents that communicate via a directed moving neighborhood random network. Each agent performs random walk on a weighted directed network. Agents interact with each other through random unidirectional information flow when they coincide in the underlying network at a given instant. For such a framework, we present sufficient conditions for almost sure asymptotic consensus. Some existed consensus schemes are shown to be reduced versions of the current model. | cs.MA | cs |
Multi-agent Coordination in Directed
Moving Neighborhood Random Networks
Yilun Shang1
Abstract
In this paper, we consider the consensus problem of dynamical multiple agents
that communicate via a directed moving neighborhood random network. Each agent
performs random walk on a weighted directed network. Agents interact with each other
through random unidirectional information flow when they coincide in the underlying
network at a given instant. For such a framework, we present sufficient conditions for
almost sure asymptotic consensus. Some existed consensus schemes are shown to be
reduced versions of the current model.
Keywords: consensus problem; random graph; stochastic stability.
1. Introduction
Synchronization and consensus problems in multi-agent systems have a long history
[3]. Over the last few years, considerable research attention has been paid to cooperative
group coordination, which has broad applications in a variety of areas including vehicle
formation, distributed sensor networks, coupled oscillator systems, animal aggregation
and epidemics; see the survey papers [10, 14, 19] and references therein. In a consensus
problem, a group of autonomous agents seeks to agree on some quantities of interest
via a process of distributed decision making based upon local information interactions.
A consensus is reached when all agents in the system tend to attain agreement on the
quantities of interest as time evolves, that is, they converge to a common value.
The ability to achieve coherent behavior (or consensus) of multi-agent system is inti-
mately related with the coupling topology of the underlying communication graph associ-
ated with the agents' positions. For example, Ali Jadbabaie et al. [6] analyzed a simplified
model of flocking by Vicsek et al. [20] showing that all agents will synchronize eventually,
provided the communication graph switching deterministically over time is periodically
jointly connected. More recently, there has been some interest in stochastic consensus
problems. [4] deals with the agreement problem among a group of static agents with the
communication graph presented as Erdos-R´enyi random graph. The results are further
extended by [12] to include weighted and directed information flows. In [18], the authors
reveal this problem can be reduced to weak ergodicity of a sequence of stochastic matrices.
However, the agents are static in these works, in other words, the switching of communi-
cation graph is merely due to the adjunctive random mechanism rather than the motion
of agents.
The moving neighborhood network proposed in [16] is meant to treat stochastic con-
sensus among dynamical agents (in social interaction and epidemics originally), where
each agent carries an oscillator and diffuses in the environment. The computer simula-
tion in the aforementioned paper shows that synchronization is possible even when the
communication network is spatially disconnected in general at any frozen time instant.
This is in sharp contrast with the behavior of Vicsek model mentioned above. Analytical
formalizations are provided by authors of [17, 13, 9] with different consensus protocols
(continuous-time with dwell time or discrete-time) through fast switching techniques [17]
1Department of Mathematics, Shanghai Jiao Tong University, Shanghai 200240, CHINA. email:
[email protected]
1
and stochastic stability theory [7]. In particular, they specify the moving neighborhood
network model as a group of n identical agents implementing simple random walks on a
fixed finite connected graph. The vertices of the moving neighborhood network are rep-
resented by the agents, and the edges are determined by their locations in the underlying
graph, i.e. a link between two agents appears if and only if they reside in the same node
simultaneously. The time evolution of the moving neighborhood graph governed by the
random walk of agents is called network dynamics and the time evolution of the corre-
sponding oscillators system is called system dynamics, which is coupled together with the
network dynamics.
The purpose of this note is to generalize the above results and go a further step in
the direction of [9]. Our extension is threefold. Firstly, we consider the underlying fixed
graph as a weighted directed graph. In other words, random walks on weighted directed
graph is tackled. Secondly, further randomness is involved in the moving neighborhood
graph. A probability is associated (not necessarily independent) to the link between two
agents when they occupy the same node, which can be interpreted as communication links
may be unreliable due to disturbance. We allow different probabilities and unidirectional
communications. Therefore the topology of the moving neighborhood graph is no longer
strictly separated clique-like. Finally, we study a discrete-time consensus protocol with
inhomogeneous weights, which quantify the way the agents influence each other. By these
modifications, some pertinent known results can be recovered (see the remarks in Section
2 & 4), and a number of introduced parameters enhance more flexibility.
The rest of this paper is organized as follows. In Section 2, we elaborate on our formal
model under consideration. Section 3 contains some mathematical preliminaries and the
statement of main result. Proofs are given in Section 4.
2. Problem setup
Let ¯G = ( ¯V , ¯E, W ) be a weighted digraph with vertex set ¯V = [m], here [m] :=
{1, 2,··· , m}. The elements in ¯V will be referred to as nodes. ¯E is a set of ordered pair
of elements of ¯V called arcs. W = (wij) is the weight matrix, where wij > 0 if (i, j) ∈ ¯E,
and wij = 0 otherwise. Notice that we do not exclude loops. We impose the following
condition on the graph ¯G.
Assumption 1. The underlying graph ¯G is strongly connected and the gcd (greatest
common divisor) of all cycle lengths in ¯G is 1.
This assumption will be elucidated in the next section. Let t ∈ N∪{0} be the discrete
time step. We consider n identical agents signified by {v1, v2,··· , vn} as random walkers
meandering on ¯G, moving randomly to a neighbor of their current location in ¯G at each
time step. For each agent, the random neighbor that is chosen is not affected by the
agent's previous trajectory. The n random walk processes are mutually independent. A
time-varying moving neighborhood graph, G = (V, E, P ), is constructed as follows. Let
V = V (t) = {v1, v2,··· , vn} be the vertex set. P = (pij) is an n × n matrix with entries
0 < pij ≤ 1 for i 6= j, and pii = 0 for 1 ≤ i ≤ n. P is dubbed as the linkage probability
matrix. Given vi and vj meet at the same node simultaneously, an arc (vi, vj) originating
from vi and terminating in vj appears with probability pij. The arc set is denoted by
E = E(t). We emphasize here that the selections of arcs are independent with the random
walk processes, but are not required to be independent with each other. By definition,
G = G(t) has no loops with probability 1.
Let Xi(t) ∈ R be the state (attitude, heading, opinion etc.) of agent vi at time t. The
2
consensus protocol can be expressed as
Xi(t + 1) = Xi(t) + ε Xj∈Ni(t)
bij(Xj(t) − Xi(t))
(1)
where Ni(t) is the index set of neighbors of agent vi at instant t, i.e. Ni(t) := {j(vi, vj) ∈
E(t)}. Here, weighting factor bij > 0 for i 6= j, and bii = 0 for 1 ≤ i ≤ n. Let
A(t) = (aij(t)) be the adjacency matrix of the moving neighborhood graph G(t), whose
entries are given by,
aij(t) = (cid:26) bij,
0,
(vi, vj) ∈ E(t)
otherwise
for 1 ≤ i, j ≤ n. Moreover, we assume the following
Assumption 2. Suppose that
aij(t) =
n
Pj=1
n
Pj=1
aji(t), for 1 ≤ i ≤ n.
The implication of Assumption 2 is the out-degrees are equal to in-degrees for every
v ∈ V (t). A similar definition used in [11] is referred to as "balanced " digraph. Note that
in the context of [11], the underlying communication network is static, while in the current
case, we dictate each trajectory of G(t) should be "balanced". However, this requirement
is not very stringent after all. For example, if we postulate G(t) can only have bidirectional
(or undirected) edges, that is, bij = bji and pij = pji for 1 ≤ i, j ≤ n, then Assumption
2 is satisfied automatically. Also, if bij ≡ b for i 6= j and suppose for each v ∈ V (t), the
numbers of arcs going out and coming in are equal, then Assumption 2 is filled, too. It is
worth noting that since the selections of arcs can be arbitrarily dependent with each other
as mentioned before, the latter of the two examples above can actually occur.
sumption is natural; see Section 4.
Next, let △ := max1≤i≤n(Pn
j=1 bij), and we further assume ε ∈ (0, 1/△). This as-
As usual, denote the linear subspace M := {x = (x1, x2,··· , xn)T ∈ Rn xi = xj ∀ 1 ≤
i, j ≤ n} as the synchronization manifold. A consensus is said to be reached if the states
of agents converge into M as t approaches infinity. We aim to show that the n agents in
V reach consensus almost surely.
We end this section by two remarks.
Remark 1. Suppose that the linkage probabilities pij(t) and the weighting factors bij(t)
are time-dependent, then the main result in this paper holds verbatim as long as pij(t)
and bij(t) converge.
Remark 2. If we take the underlying graph ¯G to be a single node, then our framework
somewhat reduces to that of [4, 18, 12].
3. Preliminaries and main result
Let {Yi(t), t ∈ N ∪ 0} denote the random walk process performed by agent vi such
that Yi(t) ∈ ¯V designates the position of vi in the underlying graph ¯G at instant t. For
1 ≤ i ≤ n, let πi(t) = (πi1(t),··· , πim(t))T be the probability distribution of Yi(t) at time
t, that is, P (Yi(t) = j) = πij(t) for j ∈ ¯V . By the construction, {Yi(t), 1 ≤ i ≤ n} are
n independent homogeneous Markov chains with finite state space ¯V , sharing the same
transition probability matrix Q = (qij), whose entry is given by
qij =
wij
di
for 1 ≤ i, j ≤ m
(2)
3
[15] Thm. 4.2).
irreducible and aperiodic (see e.g.
j=1 wij represents the out-degree of node i. Also recall that qij = P (Y1(t +
where di = Pm
1) = j Y1(t) = i). Notice that by Assumption 1 di 6= 0, so (2) is well-defined.
As is known, a finite Markov chain converges to a unique stationary distribution if it
is ergodic, i.e.
It is clear that the
transition probability matrix Q is irreducible if and only if ¯G is strongly connected. Also,
it's easy to see that the gcd of all cycle lengths in ¯G is 1 if and only if all eigenvalues of
Q other than 1 have modules strictly less than 1. Actually, the above assertions may be
proven by employing Perron-Frobenius Theorem [15] and some other equivalent conditions
in terms of the ergodicity of random walk are given in [2]. Consequently, in view of
Assumption 1, Yi(t) converges to a unique stationary distribution π = (π1,··· , πm)T for
1 ≤ i ≤ n, that is, πT = πT Q. For instance, if Pm
j=1 wji for 1 ≤ i ≤ m, then
we have πi = di/Pm
i,j=1 wij for 1 ≤ i ≤ m, see e.g. [8]. As for the rate of convergence, we
have the following lemma, (see [1] Thm. 8.9 for a proof).
Lemma 1.([1]) For a finite ergodic Markov chain with transition probability matrix P ,
let p(t)
ji be the (j, i) entry of P t, t ∈ N ∪ 0, then there is a unique stationary distribution
{pi} and
j=1 wij = Pm
p(t)
ji − pi ≤ λρt,
j=1 b1jp1j,··· ,Pn
Xj=1
E∗(L) = πT π(cid:16)diag(cid:0)
n
n
b1jp1j,··· ,
Xj=1
4
where λ ≥ 0, 0 ≤ ρ < 1.
moving neighborhood graph G(t). To be precise, dii(t) = Pn
Next, let D(t) = diag(d11(t),··· , dnn(t)) be the out-degree diagonal matrix of the
j=1 aij(t), for 1 ≤ i ≤ n,
where aij(t) are the entries of A(t) defined in Section 2. Notice that A(t) and D(t) are
non-negative time-dependent random matrices. Since aij(t) is a two-point distribution, its
expectation is shown to be given by
E(aij(t)) = bijpij
m
Xk=1
πik(t)πjk(t),
for i 6= j,
(3)
and aii(t) = 0. The diagonal entries of D(t) can be written as
m
n
dii(t) =
bij1[Yi(t)=k, Yj(t)=k, (vi,vj)∈E(t)].
For a sequence of random elements Z(t), the ergodic limit, E∗(Z),
Xk=1
Xj=1,j6=i
Thereby, E(dii(t)) = Pm
k=1Pn
j=1 bijpijπik(t)πjk(t), for 1 ≤ i ≤ n. Now let L(t) = D(t) −
A(t) denote the Laplacian matrix [2] of G(t), and it's expectation at time t is EL(t) =
ED(t) − EA(t).
is defined by
E∗(Z) = limt→∞ E(Z(t)), whenever the limit exists. E∗(Z) describes the long-run av-
erage respect to its stationary distribution [13]. By the above discussion, the ergodic limit
of (3) is E∗(aij) = bijpij Pm
k for i 6= j and E∗(aii) = 0 for 1 ≤ i ≤ n. Likewise,
E∗(dii) = Pm
Now we make a reference to Schur product of two real matrices, which is also known
as Hadamard product [5]. Let C = (cij), E = (eij) be two n × m real matrices, then the
Schur product, C ◦ E = (cijeij), is simply the product of corresponding entries of C and
E. Obviously, C ◦ E = E ◦ C. Utilizing this notation, we obtain E∗(A) = πT πB ◦ P and
E∗(D) = πT πdiag(Pn
j=1 bijpij for 1 ≤ i ≤ n.
j=1 bnjpnj). Hereby, we have
kPn
k=1 π2
k=1 π2
bnjpnj(cid:1) − B ◦ P(cid:17).
(4)
E∗(L) is positive semi-definite and have an eigenvalue 0 since E∗(L)1 = 0, where 1 is the
all-1 column vector. We need the following lemma regarding Schur product.
Lemma 2. Suppose C, E are two n× n real matrices and x, y are two n× 1 real vectors,
then
yT (C ◦ E)x = tr(cid:0)diag(yT )T Cdiag(xT )ET(cid:1).
The proof will be given in the next section. An asymptotic stability result in [7] is
restated below for convenience, which is an analogy of the deterministic Lyapunov stability
theorem.
Lemma 3.([7], pp.195) Let {Xn} be a Markov chain on state space S. Suppose that
there is a non-negative function ξ(x) satisfying
E(ξ(X1)X0 = x) − ξ(x) = −η(x),
where η(x) ≥ 0 on the set Qβ := {x : ξ(x) < β}. Then
P(cid:0) sup
n≥0
ξ(Xn) ≥ βX0 = x(cid:1) ≤ ξ(x)/β,
and accordingly, η(Xn) → 0 almost surely for paths which remain in Qβ.
It is at this stage, we state our main result as follows.
Theorem 1. Under the circumstances and assumptions presented above, the stochastic
system expressed by (1) reaches consensus almost surely.
4. Proofs
This section includes the proofs of Lemma 2 and Theorem 1.
Proof of Lemma 2. Let 1 = (1, 1,··· , 1)T be the n× 1 vector. Note that diag(yT )1 = y.
Therefore, we get
yT (C ◦ E)x = 1T diag(yT )T (C ◦ E)x
= 1T (diag(yT )T C ◦ E)x
= tr(cid:0)diag(yT )T Cdiag(xT )ET(cid:1)
We have exploited a basic property: (Adiag(xT )BT )(i, i) = ((A ◦ B)x)(i) in the last
equality above. Here notation A(i, i) means the (i, i) entry of matrix A, and x(i) the ith
element of vector x. ✷
Proof of Theorem 1. We rewrite the protocol (1) in a compact matrix form as
X(t + 1) = F (t)X(t),
(5)
where X(t) = (X1(t),··· , Xn(t))T is the state vector of agents at instant t, and F (t) =
In − εL(t). The spectrum of F (t) satisfies 1 = λn(F (t)) ≥ ··· ≥ λ1(F (t)) ≥ 1− 2ε△, since
the eigenvalues of Lapacian L(t) are 2△ ≥ λn(L(t)) ≥ ··· ≥ λ1(L(t)) = 0; see [2] for more
about Laplacian spectrum.
Recall we assume ε ∈ (0, 1/△) in Section 2, which yields λi(F (t)) < 1 for 1 ≤ i < n
[10]
and causes the state transition matrix F (t) to be stable at each time step, (see e.g.
Lemma 3). By the Assumption 2 and iterative equation (5), we obtain
1T X(t + 1) = 1T F (t)X(t) = (F (t)X(t))T 1 = X(t)T F (t)T 1 = (X(t)T 1)T = 1T X(t).
5
From this, it is clear that the projection of the state vector X(t) on synchronization
manifold M is a constant. In fact, let α := 1
i=1 Xi(0), then 1T X(t) = nα for all t.
Now we may decompose X(t) as
n Pn
X(t) = X pa(t) + X pe(t),
(6)
where X pa(t) ∈ M and X pe(t)⊥M. The superscript "pa" stands for "parallel", while
"pe" for "perpendicular". We thereby get X pa(t) = 1√n 1T X(t) = √nα. Here and in
the sequel, we take · as 2-norm for vectors and induced 2-norm for matrices [5]. The
disagreement among the agents now can be described by X pe(t)2 = X(t)2 − nα2. To
prove the theorem, it suffices to show X pe(t) → 0 almost surely as t → ∞.
We follow [9] to introduce a Lyapunov function ξ(x) = xT x − α2n, and let
ξ(t) := ξ(X(t)) = X(t)T X(t) − α2n.
Our plan is to use Lemma 3 to show ξ(t) → 0 almost surely, as t → ∞. Note that X(t) is
indeed a Markov chain. With this in mind, we evaluate by employing (5),
E(ξ(t + 1)X(t)) = X(t)T E(F (t)T F (t))X(t) − α2n.
Therefore we get
E(ξ(t + 1)X(t)) − ξ(t) = X(t)T EH(t)X(t)
(7)
where, H(t) := F (t)T F (t) − I = ε2L(t)T L(t) − ε(L(t)T + L(t)).
Let Li be the ith possible realization of L(t), with probability pi(t) at instant t, and
denote Ω as the collection of all Li. We hereby may write the expectation of L(t) as
Ω is the cardinality of Ω. Since L(t) inherited the property of
the random walks of the agents, it is also an ergodic Markov chain with state space Ω.
Let pi be the ith component of the unique stationary probability distribution. Then we
may apply Lemma 1 and interpret the corresponding meaning of notations in the current
situation. By doing so, we derive
EL(t) = PΩi=1 Lipi(t).
pi(t) − pi = (cid:12)(cid:12)(cid:12)
p(t)
(p(t)
Ω
Xj=1
Ω
Xj=1(cid:12)(cid:12)p(t)
Ω
= (cid:12)(cid:12)(cid:12)
ji − pi)pj(0)(cid:12)(cid:12)(cid:12)
ji pj(0) − pi(cid:12)(cid:12)(cid:12)
Xj=1
Ω
Xj=1
ji − pi(cid:12)(cid:12)pj(0) ≤ λρt
pj(0) = λρt.
≤
Now going back to the equation (7), we have
X(t)T EH(t)X(t) =
Ω
Xi=1
pi(t)X(t)T HiX(t)
(8)
(9)
where Hi = ε2LT
i + Li) is the ith possible realization of H(t) and is negative
semi-definite with eigenvalues 0 = λn(Hi) ≥ ··· ≥ λ1(Hi). Using the decomposition (6),
we get
i Li − ε(LT
X(t)T HiX(t) = X pe(t)T HiX pe(t) + X pa(t)T HiX pe(t)
+X pe(t)T HiX pa(t) + X pa(t)T HiX pa(t)
6
There are four terms on the right hand side of the above equation, the last three of which
are actually zero. To see why let's take the second term as an example,
X pa(t)T HiX pe(t) = ε2X pa(t)T LT
i LiX pe(t) − εX pa(t)T LT
i X pe(t) − εX pa(t)T LiX pe(t)
= ε2(LiX pa(t))T LiX pe(t) − ε(LiX pa(t))T X pe(t)
−ε(LT
i X pa(t))T X pe(t)
= 0
since LiX pa(t) = LT
Plugging this equation into (9), we get
i X pa(t) = 0. Thereby we have X(t)T HiX(t) = X pe(t)T HiX pe(t).
X(t)T EH(t)X(t) = X pe(t)T(cid:16)
Ω
Xi=1
pi(t)Hi(cid:17)X pe(t) = X pe(t)T EH(t)X pe(t).
We have the ergodic limit, defined in Section 3, of H(t) as
E∗H = lim
t→∞
EH(t) = lim
t→∞
E(cid:0)ε2L(t)T L(t) − ε(L(t)T + L(t))(cid:1)
= ε2E∗(LT L) − ε(E∗LT + E∗L)
(10)
since we know L(t) is an ergodic Markov chain.
Take R(t) := EH(t) − E∗H as the remainder matrix, then we have
X pe(t)T EH(t)X pe(t) = X pe(t)T (E∗H)X pe(t) + X pe(t)T R(t)X pe(t).
(11)
The second term on the right hand side of (11) can be calculated, by noting (8), as follows
X pe(t)T R(t)X pe(t) = (cid:12)(cid:12)(cid:12)(cid:12)X pe(t)T R(t)X pe(t)(cid:12)(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)(cid:12)X pe(t)T(cid:16)
Hi(pi(t) − pi)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2 ·(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤ (cid:12)(cid:12)(cid:12)(cid:12)X pe(t)(cid:12)(cid:12)(cid:12)(cid:12)
≤ ξ(t)Ωγλρt,
Ω
Xi=1
Ω
Xi=1
Hi(pi(t) − pi)(cid:17)X pe(t)(cid:12)(cid:12)(cid:12)(cid:12)
(12)
where γ := max1≤i≤Ω Hi. Wherefore, (12) tends to 0, since 0 ≤ ρ < 1 and ξ(t) is
bounded by (1).
Next, we turn to the first term on the RHS of (11). In view of (10),
X pe(t)T (E∗H)X pe(t) = ε2X pe(t)T (E∗LT )(E∗L)X pe(t) − εX pe(t)T (E∗LT )X pe(t)
−εX pe(t)T (E∗L)X pe(t).
(13)
By using Lemma 2, we have
X pe(t)T (BT ◦ P T )X pe(t) = tr(cid:0)diag(X pe(t)T )BT diag(X pe(t)T )P(cid:1)
= tr(cid:0)diag(X pe(t)T )Bdiag(X pe(t)T )P T(cid:1)
= X pe(t)T (B ◦ P )X pe(t).
7
Then it's easy to see the last two terms on the RHS of (13) is equal, by noting equation
(4). Thereby we obtain
X pe(t)T (E∗H)X pe(t) = ε2X pe(t)T (E∗L)T (E∗L)X pe(t) − 2εX pe(t)T (E∗L)X pe(t)
≤ (2△ε2 − 2ε)X pe(t)T (E∗L)X pe(t)
= 2επT π(ε△ − 1)X pe(t)T
b1jp1j,··· ,
≤ 2επT π△(ε△ − 1)ξ(t)
·(cid:0)diag(cid:0)
Xj=1
n
n
Xj=1
bnjpnj(cid:1) − B ◦ P(cid:1)X pe(t)
(14)
The second inequality above holds because there is some N (B) > 0 depends only on B,
such that (E∗L)T (E∗L) ≤ 2△(E∗L) + N (B)11T and
X pe(t)T N (B)11T X pe(t) = N (B)(1T X pe(t))T (1T X pe(t)) = 0.
The fourth inequality holds by noting that ε△ − 1 < 0 and
n
Xj=1
(cid:0)diag(cid:0)
b1jp1j,··· ,
n
Xj=1
bnjpnj(cid:1) − B ◦ P(cid:1) ≥ △I − M (B)11T
for some M (B) > 0 depending only on B. The similar argument as above may be applied.
Combing (12) and (14), we finally get
E(ξ(t + 1)X(t)) − ξ(t) ≤ ξ(t)(cid:0)Ωγλρt + 2επT π(ε△ − 1)△(cid:1)
:= ξ(t)δ(t) := −η(t)
There exists t0 ∈ N such that δ(t) < 0 for t ≥ t0. Furthermore, δ(t) → 2επT π(ε△ − 1)△,
as t → ∞. Also, note that ξ(t) is a Markov chain, since X(t) is. By applying Lemma 3,
where β can take any positive number, we obtain
η(t) → 0
as
t → ∞
almost surely, which yields ξ(t) → 0 almost surely, as t → ∞. The proof now is completed.
✷
Remark 3. Theorem 1 shows an average consensus which may be regarded as a kind of
stochastic version of Theorem 9 in [11].
References
[1] P. Billingsley, Probability and Measure. John Wiley & Sons, New York, 1995
[2] F. R. K. Chung, Spectral Graph Theory. American Mathematical Society, Providence,
1997
[3] M. H. DeGroot, Reaching a consensus, J. Am. Statistical Assoc., 69(1974) pp.118 -- 121
[4] Y. Hatano, M. Mesbahi, Agreement over random networks. IEEE Trans. Autom.
Control, 50(2005) pp.1867 -- 1872
8
[5] R. A. Horn, C. R. Johnson, Matrix Analysis. Cambridge University Press, Cambridge,
1985
[6] A. Jadbabaie, J. Lin, A. S. Morse, Coordination of groups of mobile autonomous
agents using nearest neighbor rules. IEEE Trans. Autom. Control, 48(2003) pp.988 --
1001
[7] H. Kushner, Introduction to Stochastic Control. Holt, Rinehart and Winston, New
York, 1971
[8] L. Lov´asz, Combinatorics, Paul Erdos is Eighty, vol 2. Bolyai Society Mathematical
Studies, 1996 ch. Random walks on graphs: A survey.
[9] F. Mudasir, M. Porfiri, V. Kapila, Agreement over networks of mobile agents. Proc.
IEEE Conference on Decision and Control, 2007 pp.4239 -- 4244
[10] R. Olfati-Saber, J. A. Fax, R. M. Murray, Consensus and cooperation in networked
multi-agent systems. Proceedings of the IEEE, 95(2007) pp.215 -- 233
[11] R. Olfati-Saber, R. M. Murray, Consensus problems in networks of agents with switch-
ing topology and time-delays. IEEE Trans. Autom. Control, 49(2004) pp.1520 -- 1533
[12] M. Porfiri, D. J. Stilwell, Consensus seeking over random weighed directed graphs.
IEEE Trans. Autom. Control, 52(2007) pp.1767 -- 1773
[13] M. Porfiri, D. J. Stilwell, E. M. Bollt, J. D. Skufca, Random talk: Random walk and
synchronizability in a moving neighborhood network. Physica D, 224(2006) pp.102 --
113
[14] W. Ren, R. W. Beard, E. M. Atkins, A survey of consensus problems in multi-agent
coordination, Proc. American Control Conference, 2005 pp.1859 -- 1864
[15] E. Seneta, Non-negative Matrices and Markov Chains. Springer, New York, 2006
[16] J. D. Skufca, E. M. Bollt, Communication and synchronization in disconnected net-
works with dynamic topology: Moving neighborhood networks. Math. Biosci. Eng.,
1(2004) pp.347 -- 359
[17] D. J. Stilwell, E. M. Bollt, D. G. Roberson, Sufficient conditions for fast switching
synchronization in time-varying network topologies. SIAM J. Appl. Dynam. Syst.,
6(2006) pp.140 -- 156
[18] A. Tahbaz-Salehi, A. Jadbabaie, A necessary and sufficient condition for consensus
over random networks. IEEE Trans. Autom. Control, 53(2008) pp.791 -- 795
[19] F.-X. Tan, X.-P. Guan, D.-R. Liu, Consensus protocols for multi-agent continuous
systems. Chinese Physics B, 17(2008) pp.3531-3535
[20] T. Vicsek, A. Czir´ok, E. Ben-Jacob, I. Cohen, O. Shochet, Novel type of phase tran-
sition in a system of self-driven particles. Phys. Rev. Lett., 75(1995) pp.1226 -- 1229
9
|
1204.0163 | 3 | 1204 | 2012-10-18T23:33:06 | Fashion, Cooperation, and Social Interactions | [
"cs.MA",
"cs.SI",
"physics.soc-ph"
] | Fashion plays such a crucial rule in the evolution of culture and society that it is regarded as a second nature to the human being. Also, its impact on economy is quite nontrivial. On what is fashionable, interestingly, there are two viewpoints that are both extremely widespread but almost opposite: conformists think that what is popular is fashionable, while rebels believe that being different is the essence. Fashion color is fashionable in the first sense, and Lady Gaga in the second. We investigate a model where the population consists of the afore-mentioned two groups of people that are located on social networks (a spatial cellular automata network and small-world networks). This model captures two fundamental kinds of social interactions (coordination and anti-coordination) simultaneously, and also has its own interest to game theory: it is a hybrid model of pure competition and pure cooperation. This is true because when a conformist meets a rebel, they play the zero sum matching pennies game, which is pure competition. When two conformists (rebels) meet, they play the (anti-) coordination game, which is pure cooperation. Simulation shows that simple social interactions greatly promote cooperation: in most cases people can reach an extraordinarily high level of cooperation, through a selfish, myopic, naive, and local interacting dynamic (the best response dynamic). We find that degree of synchronization also plays a critical role, but mostly on the negative side. Four indices, namely cooperation degree, average satisfaction degree, equilibrium ratio and complete ratio, are defined and applied to measure people's cooperation levels from various angles. Phase transition, as well as emergence of many interesting geographic patterns in the cellular automata network, is also observed. | cs.MA | cs |
1
Fashion, Cooperation, and Social Interactions
Zhigang Cao1,∗, Haoyu Gao1, Xinglong Qu1, Mingmin Yang1, Xiaoguang Yang1,
1 Key Laboratory of Management, Decision & Information Systems, Academy of
Mathematics and Systems Science, Chinese Academy of Sciences, Beijing, China
∗ E-mail: [email protected], [email protected],
[email protected], [email protected], [email protected]
Abstract
Fashion plays such a crucial rule in the evolution of culture and society that it is regarded as a second
nature to the human being. Also, its impact on economy is quite nontrivial. On what is fashionable, inter-
estingly, there are two viewpoints that are both extremely widespread but almost opposite: conformists
think that what is popular is fashionable, while rebels believe that being different is the essence. Fashion
color is fashionable in the first sense, and Lady Gaga in the second. We investigate a model where the
population consists of the afore-mentioned two groups of people that are located on social networks (a
spatial cellular automata network and small-world networks). This model captures two fundamental kinds
of social interactions (coordination and anti-coordination) simultaneously, and also has its own interest to
game theory: it is a hybrid model of pure competition and pure cooperation. This is true because when
a conformist meets a rebel, they play the zero sum matching pennies game, which is pure competition.
When two conformists (rebels) meet, they play the (anti-) coordination game, which is pure cooperation.
Simulation shows that simple social interactions greatly promote cooperation: in most cases people can
reach an extraordinarily high level of cooperation, through a selfish, myopic, naive, and local interacting
dynamic (the best response dynamic). We find that degree of synchronization also plays a critical role,
but mostly on the negative side. Four indices, namely cooperation degree, average satisfaction degree,
equilibrium ratio and complete ratio, are defined and applied to measure people's cooperation levels from
various angles. Phase transition, as well as emergence of many interesting geographic patterns in the
cellular automata network, is also observed.
Introduction
Fashion is a very interesting phenomenon that plays a critical role in economy. Firstly, fashion is a
huge industry. Despite the complication to give a satisfactory definition of the fashion industry, its
global market size is estimated at some two hundred billion US dollars [1, 2]. The importance of the
fashion industry can be further justified by the following facts [1]: fashion industry "is one of the few
industrial segments that have remained a constant world economy contributor with an annual growth
rate of approximately 20 percent. [...] The luxury fashion sector is the fourth largest revenue generator
in France; and one of the most prominent sectors in Italy, Spain, the USA and the emerging markets of
China and India. The sector is currently one of the highest employers in France and Italy." Secondly,
fashion serves as a constant and efficient consumption stimulator. It is well known that the purchases of
most consumer durables are replacement ones [3]. For instance, the shoes that we buy are usually not our
first ones, the aim that we buy them is to replace our old ones. However, a not so environmentally friendly
fact is that a significant percentage of replaced products still function very well [3]. It is becoming more
and more popular that people discard their old possessions not because they fail to meet their physical
needs, but because they are not fashionable any longer. If all people had purchased new goods only when
their old ones were completely broken, then the world economy would have grown much slower.
Not only in economy, but also behind many phenomena in society, education, politics, arts and
academics, fashion is a factor that cannot be neglected. A recent research shows that even charitable
donation is a highly subject of fashion [4]. Lars Svendsen, a famous philosopher, argues that [5] "Fashion
2
has been one of the most influential phenomena in Western civilization since the Renaissance. It has
conquered an increasing number of modern man's fields of activity and has become almost 'second
nature' to us." He believes that fashion deserves serious studies from philosophers. We are sure that
as the post industrial society [6] is coming to reality for more and more countries, the practical functions
of commodities and human behaviors are mattering less and less compared with their social functions.
Consequently, fashion is playing an increasingly crucial role.
On what is fashionable, interestingly, there are two almost opposite viewpoints that are both extremely
popular. One point of view thinks that fashion is a distinctive or peculiar manner or way. Lady Gaga is
regarded as fashionable in this sense. The other takes fashion as a prevailing custom or style. Fashion
color is fashionable in this latter sense. This difference has a very deep root in psychology, and reveals
that people have various desiring or enduring levels of to be how different with the others. Following
Jackson [7], we call the former type of people rebels, and the latter conformists.
The phenomenon of fashion has so far attracted some attention from the academical realm, mostly in
economics [1, 8 -- 12]. However, compared with its great importance, this is far from sufficient.
We shall study fashion through a game-theoretical model, which is called the fashion game. Formally,
each fashion game is represented by a triple I = (N, E, T ), where N = {1, 2, · · · , n} is the set of agents,
E ⊆ N × N the set of edges, and T = (τ1, τ2, · · · , τn) ∈ {C, R}N the configuration of types. For each
agent i ∈ N , τi ∈ {C, R} is her type: τi = C means that i is a conformist, and τi = R a rebel. For agents
i, j ∈ N , they are neighbor to each other if and only if ij ∈ E. If ij ∈ E, then ji ∈ E, i.e. the network
is undirected. Ni is the neighbor set of player i, and {0, 1} the identical (pure) action set of all players.
We use xi ∈ {0, 1} to denote the action of player i. Given a pure action profile X = (x1, x2, · · · , xn),
Li(X) ⊆ Ni is the set of neighboring agents that i likes (w.r.t. X), i.e.
Li(X) = (cid:26) {j ∈ Ni : xj = xi} if τi = C
{j ∈ Ni : xj 6= xi}
if τi = R
.
Similarly, Hi(X) ⊆ Ni is the set of neighboring agents that i hates, i.e. Hi(X) = Ni \ Li(X). Using
· to denote the cardinality of a set, the utility function of player i can be defined naturally as follows:
ui(X) = Li(X) − Hi(X).
Utilities of mixed action profiles can be extended as usual. If the utility of an agent is nonnegative,
we say that she is satisfied. An action profile that all agents are satisfied corresponds clearly to a Nash
equilibrium.
As far as we know, the above fashion game is proposed by Young ( [13], 2001, p.38) and Jackson ( [7],
2008, p.271), independently. Obviously, this game is an extension of the famous matching pennies game,
which is stated below.
H
T
H (1, −1)
T (−1, 1)
.
(−1, 1)
(1, −1)
In fact, in matching pennies, the row player is a conformist, and the column player a rebel. A special
case of the fashion game, a dyad with one conformist and one rebel, is exactly the matching pennies
game. Just like matching pennies, the fashion game always has a mixed Nash equilibrium: all agents
play half 0 and half 1. The existence of pure Nash equilibrium, however, cannot be guaranteed, and
it is NP-hard to check whether a fashion game on a general network has a pure Nash equilibrium or
not [14]. Consequently, it is impossible to compute a pure Nash equilibrium efficiently when it does exist,
unless P=NP (i.e. the set of problems admitting deterministic polynomial time algorithms equals that
of problems admitting non-deterministic polynomial time algorithms, a statement widely conjectured
as impossible to be true. Extensive discussion of this conjecture can be found in any textbook about
computational complexity).
3
It is valuable to note that the fashion game, though very simple, is a typical heterogeneous model.
There are two types of players, and actually three base games played: When a conformist faces a
conformist, they play the pure coordination game; When two rebels meet, they play the pure anti-
coordination game; And when a conformist confronts a rebel, they are in the exact game of matching
pennies.
It is widely accepted that competition and cooperation are the two eternal topics in game theory.
Zero sum games (or more generally, constant sum games) are polar examples for competition. This is
why they are also called strictly competitive games (cf. [15], p.21). Common interest games (a.k.a. team
games) are polar examples for cooperation, where the preference rankings of pure action profiles for all
players are the same. Among the three base games of the fashion game, the pure coordination game and
the pure anti-coordination game are both common interest games, while matching pennies is a zero sum
game. For general normal form games, competition and cooperation are both embodied. This is a kind
of vertical hybrid, and is the key observation of [16]. The fashion game, on the other hand, is a kind of
horizontal hybrid of competition and cooperation. This feature determines that the fashion game has a
very special interest to game theory.
The fashion game can be safely classified into network games, a typical multi-disciplinary field that
rests at the intersection of social economics, social physics, theoretical biology, and algorithmic game
theory [7,17 -- 20]. Studying closely related models, it is a pity that researchers from different fields seldom
cite each other so far. Although the exact model of the fashion game does not draw much attention
today, there has been a lot of related work.
In social economics, the coordination game, one side of
the fashion game, has been extensively studied [13, 21 -- 29]. The other side of the fashion game, the anti-
coordination game, however, attracts very little attention [30 -- 33]. In the field of social physics, for models
where agents are homogenously conformists or homogenously rebels, there has been a large number of
references, which are impossible to fully survey here. They are called "majority game" (cf. [34, 35])
and "minority game" (cf. [36, 37]), respectively.
It is well known that the minority game has been
used to study the financial and stock market since very soon after its birth. Marsili [38] noticed early
enough that people do not necessarily play the minority game in stock market, because except for a few
"market fundamentalists", most people are "trend followers". And thus instead, the minority-majority
game is much more appropriate. Following Marsili, there are a dozen of papers [35, 39 -- 46]. Another
extensively studied model that is closely related with the anti-coordination game is the snowdrift game
(a.k.a. the hawk-dove game, see the excellent survey [18]). Related work in the fields of social learning
and opinion dynamics includes [47 -- 51]. In statistical physics, another similar model is the generalization
of the Kuramoto model with conformist oscillators and contrarian oscillators ( [52 -- 54]). It is valuable to
remark that, in the afore mentioned papers, neither the relation of their models with the phenomenon
of fashion nor the relation with the matching pennies game is noticed. For recent work on repeated
matching pennies from the perspective of behavioral science, please refer [55]. A brand new perspective
on matching pennies is that it can be interpreted as a (symmetric) Predator-Prey game [56, 57]. A quite
complete literature review of "Social Influence, Binary Decisions and Collective Dynamics", a classical
topic in sociology and also closely related with this paper, can be found in [58].
The main concern of this paper is, in a world where each agent is fashionable (in one of the possible
two ways), selfish (cares only about the welfare of her own), extremely naive and myopic (has only one
step memory and does not look forward), and has very limited information (that of herself and her
neighbors), to what extent can cooperation be reached through social interaction? Will things be even
worse?
Our finding is, in general, quite encouraging: an extraordinarily high level of cooperation can be
reached through a simple updating rule, the best response dynamic. That is, social interactions generally
promote cooperation (compared with the initial settings that are uniformly random, i.e. each agent takes
action 1 with probability 0.5 and action 0 with probability 0.5). In very rare cases, the discouraging result
that social interactions prohibit cooperation can also be observed. Degree of synchronization, captured by
4
updating probability in the best response dynamic that will be introduced in the next section, also plays
a critical role, but mostly on the negative side. We remark that the negative effect of synchronization is
observed in [58] too, for the anti-coordination game on a complete network.
Methods
Most of our conclusions are derived through a two dimensional (stochastic) cellular automaton. This
spatial structure is simple enough and can serve very well for a first step study. After discussion of this
particular case, more extensive analyses are done for the more general and more realistic small-world
networks [60], where most of our conclusions are confirmed. Since it is notoriously hard to rigorously
analyze a two dimensional cellular automaton (even a deterministic one), and the negative result of [14]
tells us that there is no good characterization of the fashion game, simulation is our natural choice.
Three things need to be clearly stated: the underlying network, the initial settings, and the updating
In this section, we shall only introduce the particular cellular automata network. Small-world
rule.
networks and the corresponding simulation settings will be introduced in a later section. The updating
rule, as well as the four indices that will be introduced soon, is universal.
Primary simulation settings
The cellular automata network is a special 8-degreed regular graph. It is a sheet of grids, where each grid
stands for an agent. For each agent, the eight grids that are touching it are her neighbors (that is, we
are taking the Moore neighborhood). The sheet is finite but unbounded: each leftmost agent has three
neighbors on the rightmost, and each rightmost agent has three neighbors on the leftmost. Likewise, each
uppermost agent has three neighbors on the lowermost, and each lowermost agent has three neighbors on
the uppermost. Intuitively, we can imagine this world as a torus. In the next section, all the simulations
(except for the ones in Fig. 6) are done on such a torus of size 41 × 41 = 1681.
To characterize the initial settings and the updating rule, we need two parameters: the rebel ratio r
and the updating probability p. They are the only parameters for simulations on the cellular automata
network.
For each initialization, each agent has a probability of r to be a rebel, and 1 − r to be a conformist.
This is done for all agents independently. r = 0 is the case of all conformists, and r = 1 that of all rebels.
Since the network is quite large, r can be roughly taken as the percentage of rebels. It is found that the
percentage of rebels matters a lot to the fashion game. For a special instance, the all-conformist case is
completely different from the all-rebel one.
Actions are initialized uniformly, that is, each agent takes an action of 1 and 0 equally likely (with
probability 0.5). Time elapses discretely. At each time step, each agent checks if she is satisfied with
the previous action profile.
If so, her action keeps unchanged. Otherwise, she switches to the other
action with a probability p. Intuitively, p measures the degree of synchronization. p = 1 is exactly the
synchronous best response dynamic. When p is infinitely small, we know that at each step there is at
most one agent switching her action. To be precise, the probability that two or more agents switch actions
is second-order infinitely small (w.r.t. p). This can be roughly taken as the asynchronous best response
dynamic. The introduction of parameter p allows us to compare the cases of synchronous updating,
asynchronous updating, and all the middle cases. It turns out that p also matters a lot, usually on the
negative side.
Four indices
We use cooperation degree as a main index to measure the level of cooperation between agents. Formally,
the cooperation degree for any configuration of actions in a fashion game is defined as the percentage
5
of satisfied agents. Clearly, Nash equilibrium corresponds to an action configuration with cooperation
degree of 1. This index can be roughly taken as an approximation of Nash equilibrium. Since Nash
equilibrium is not guaranteed, this choice is quite natural. Three other indices are also used, namely
average satisfaction degree, equilibrium ratio, and complete ratio.
The satisfaction degree of any agent is the percentage of neighbors that she likes. Obviously, satisfac-
tion degree is an extension of the concept of satisfaction: an agent is satisfied if and only if her satisfaction
is no less than 0.5. The average satisfaction degree for an action configuration is simply defined as the
average of all agents' satisfaction degrees.
The equilibrium ratio of a group of simulations is defined as the percentage of simulations that pure
Nash equilibrium is reached. Recall that pure Nash equilibrium is not guaranteed, and even if it exists,
it may not be reached by the best response dynamic. The equilibrium ratio also measures the degree
of cooperation, because Nash equilibrium can be treated as stable cooperation. The other advantage of
this index is that when we investigate phase transition, it is much sharper than cooperation degree and
average satisfaction degree.
Given an action configuration, an agent is said to be completely satisfied if and only if her satisfaction
degree is 1. The complete ratio of an action configuration is defined as the percentage of completely
satisfied agents.
Simulation size
All results (except for Fig. 4 and Fig. 5) in the next section are based on 420 groups of simulations: the
rebel ratio r takes 21 values, 0, 0.05, 0.1, · · · , 1, and the updating ratio p takes 20 values, 0.05, 0.1, · · · , 1.
For each of the 420 combination of parameters, 10 simulations are done. For each simulation, it is stopped
at the 500-th step. For each combination of parameters, we display the average of the corresponding 10
final values.
It may seem at the first sight that stopping at the 500-th step and taking an average of only 10
simulations are not sufficient for such a large network. However, it turns out that this is enough, because
the dynamic is surprisingly both fast and robust.
Main Results
Cooperation degree
First of all, let's calculate the expected cooperation degree of the initial configuration. This value is
equivalent to the probability that an agent is initially satisfied. Since each agent has 8 neighbors in
total, and she is satisfied if and only if the number of agents who take the same action as she does
is at least 4 when she is a conformist, and at most 4 when she is a rebel, and each agent takes an
action equally likely from {0, 1}, we know that this probability is (1/2)8(C 8
8 ) =
(1/2)8(C 0
8 ) = 163/256
.
= 0.64.
8 + C 7
8 + C 6
8 + C 5
8 + C 4
8 + C 4
8 + C 1
8 + C 2
8 + C 3
The average of the 420 cooperation degrees in our simulations is calculated to be 0.97. The 420 values
are displayed in the color map of Fig. 1, where different colors represent different values, as shown in the
right bar.
The most prominent observations from Fig. 1 are as follows.
(1) For most parameter combinations, stable cooperation, i.e. a pure Nash equilibrium, can be reached.
This tells us that although pure Nash equilibrium cannot be guaranteed generally, it does exist for most
cases in the cellular automaton world, and it can be reached by simple adaptive dynamics: best response
dynamics. We remind the reader that, for general games, even if pure Nash equilibrium exists, there is
no general simple adaptive dynamic that always leads to one [61]. The only theoretical result we know is
that best response dynamic (in fact, better response dynamic) always leads a potential game to a pure
6
Figure 1. Cooperation degree. The fashion game can reach high degree of cooperation through best
response dynamics.
Nash equilibrium [62]. However, the fashion game is not a potential game, even in the cellular automaton
structure. This is because pure Nash equilibrium is guaranteed in potential games, which is not true in
the fashion game.
It is also valuable to remark that for the synchronous best response dynamic (i.e. p = 1), it may not
lead the fashion game to a pure Nash equilibrium, even if it exists. In fact, this always occurs when all
agents are rebels (a case which is a potential game and thus guarantees the existence of a pure Nash
equilibrium, cf. [30]) and they take the same action initially. It is obvious that at each step, no agent is
satisfied, and thus all agents switch to the other action simultaneously, leading to the other state where no
agent is satisfied either. The configuration will oscillate between the two extreme states, and no agent is
ever satisfied at all. This illustrates the most terrible situation that no one wants to see. Synchronization,
of course, plays a critical role. In fact, asynchronous best response dynamic (i.e. p is infinitely small), as
shown in Fig. 1, leads to a pure Nash equilibrium almost surely, and for any p < 1, the corresponding
best response dynamic leads to a pure Nash equilibrium with a positive probability. This will be explored
more extensively in later subsections.
(2) Taking into account the fact that the network size is 1681 and in most cases the best response
dynamics converge within 500 steps, the convergence is remarkably fast.
(3) Bad cooperation occurs only when the percentage of rebels r and the updating probability p are
both high. To investigate this more clearly, we put the upper-right corner of Fig. 1 in another way, as
shown in Fig. 2.
It can be observed that the cooperation degree is always a decreasing function of the rebel ratio, which
means that when the degree of synchronization is high, rebels, in general, are a block for cooperation.
This is consistent with the theoretical results that when all agents are conformists, stable cooperation is
always possible (i.e. pure Nash equilibrium exists), but when a portion of rebels are added, this may not
be true. It can also be observed that higher probability of updating in general (but not always) means
worse cooperation. This is also consistent with our argument in (1), i.e. asynchronization is better than
synchronization. These observations will be further justified by more extensive simulations of the next
7
Figure 2. Bad cooperation. Cooperation degree is a decreasing function of rebel ratio r when p is
large.
subsection.
(4) For some combinations of parameters (upper-right corner of Fig. 1), the final cooperation degrees
can be even worse than the initial ones. This tells us that best response dynamics with high updating
probabilities may not be always good for the whole system, even if updating itself is costless (i.e. agents
don't pay for switching actions).
Average satisfaction degree
Since the initializations of all agents are independent, it is obvious that the initial expected average
satisfaction degree is 0.5. Fig. 3 is also based on the same 4200 simulations, where each simulation is
terminated at the 500-th step. However, only 11 × 21 =231 values are displayed, each as an average
satisfaction degree of 10 corresponding simulations.
Figure 3. Average satisfaction degree as a function of rebel ratio.
Below is the analysis of this figure.
(1) In most cases, the average satisfaction degree is rather promising. In fact, the overall average
8
satisfaction degree (i.e. average of the 231 values) is 0.73, much higher than the initial expected value
of 0.5. This is consistent with our observations in the last subsection, and justifies our claim once again
that, in general, various best response dynamics can promote cooperation.
(2) The overall average satisfaction degree, 0.73, is not that high compared with the overall cooperation
degree, 0.97. This is not surprising, because even if the cooperation degree is 1, it is very unlikely for the
corresponding average satisfaction degree to be 1, in which case it must be that all agents are completely
satisfied.
If the above perfect situation occurs, we say that perfect cooperation is obtained. Perfect cooperation
can only possibly be obtained in one extreme case: the all-conformist case (r = 0). This deepens
our observation in the last subsection in a way that in most cases, almost all agents are satisfied, not
completely satisfied, but satisfied to some degree. We remark that in the all-rebel case (r = 1), the
configuration where every agent gets complete satisfaction in general does not exist, although pure Nash
equilibrium is guaranteed. To be precise, in the all-rebel case, the configuration where every agent is
completely satisfied exists if and only if the graph is bipartite. Unfortunately, the two dimensional
cellular automaton with Moor neighborhood is not bipartite (there are cycles of length three). However,
the two dimensional cellular automaton with von Neumann neighborhood, i.e. each agent has exactly
four neighbors, is indeed bipartite.
(3) In the all-conformist case, perfect cooperation is not attainable through best response dynamics.
In this case, there are two focal equilibria, i.e. all agents taking action 1 and all agents taking action 0,
and they are both perfectly good. However, it is very unlikely for simple dynamics to lead the system to
them. It is imaginable that perfect cooperation is hugely hard, and the main obstacle comes from the
underlying network. When the network is a complete graph, it may be not that hard to reach perfect
cooperation.
(4) Fortunately, in the very case that all agents are conformists, the real equilibria reached through
best response dynamics are good enough. The average satisfaction degree is shown to be 0.96. In fact,
we will show later that the equilibrium states when all agents are conformists are usually composed of
impressively large"continents", and the agents that are not completely satisfied all sit on the coastlines.
(5) In most cases, the average satisfaction degree is not sensitive to the updating probability p, unless
p and r are both large (p > 0.8, r > 0.5).
In fact, the eight curves for r = 0.1, · · · , 0.8 are almost
identical. This is consistent with the observations of the last subsection. However, cooperation degrees
are monotonic there, always approximately 1. Here, the average satisfaction degree, as a function of
rebel ratio, is richer and more interesting.
In this sense, the fact that (in most cases) the average
satisfaction degree is insensitive to updating probability is much more striking than that (in most cases)
the cooperation degree is insensitive to updating probability.
(6) In most cases (p < 0.8), average satisfaction degree is a convex function of the rebel ratio r, and
reaches its lowest value 0.68 at 0.5. This says that the case where there are equally number of conformists
and rebels is the most difficult situation to cooperate.
This is intuitively reasonable, because an edge between a conformist and a rebel always contributes
0 to exactly one of the agents. The number of these inter-type edges is expected to reach its maximum
when r = 0.5. In fact, the very special case where every agent has four conformist neighbors and four
rebel ones has exactly an average satisfaction degree of 0.5. Having this in mind, and considering that the
expected initial average satisfaction degree is also 0.5, the final value of 0.68 is really not disappointing
at all. The property that average satisfaction degree is decreasing when r < 0.5 and increasing when
r > 0.5 is easy to understand. The convexity of this function, however, is not that intuitive. As to why
the satisfaction degree of the all-rebel case, 0.71, is much lower than 0.96, the satisfaction degree of the
all-conformist case, remember the remark we give in (2). The particular value of 0.71 will be revisited
later.
(7) For large p and large r, the average satisfaction degree can be rather low, and may be even much
lower than the expected initial value, 0.5. This is consistent with our observations in the last subsection.
9
(8) When r > 0.5, the average satisfaction degree is an increasing function of r for small p, and a
decreasing function of r for large p. This reflects that there may be some phase transition for p, which
will be explored more extensively in the next subsection.
Phase transition
To explore the possible phase transition, we did more careful simulations. Our main worries are that
(i) 500 steps may not be enough for the cases where p and r are both large to reach a relatively stable
state, and (ii) 0.05, the difference between any two adjacent ps we take, is too rough for observing phase
transition. Based on the above two worries, we change the simulation settings to allow each simulation
run 5000 steps, and ps are taken more densely. To be precise, we investigate 41 ps: 0.8, 0.805, · · · , 1, and
5 rs: 0.5, 0.6, · · · , 1. For each of the 41 × 5 = 205 parameter combinations, we take 10 simulations and
calculate the average cooperation degree. The results are displayed in Fig. 4.
Figure 4. Phase transition of cooperation degree in p.
Figure 5. Phase transition of equilibrium ratio in p. r=1.
Fig. 4 shows that phase transition does exist, especially for large rs: cooperation degree drops from
1 to a very low level within a small change of p. We put the all-rebel case in another way as in Fig. 5,
where the index of cooperation degree is replaced by the equilibrium ratio.
Fig. 5 shows that when all agents are rebels, Nash equilibrium can always be reached through best
response dynamics if the updating probability p < 0.85. However, once p exceeds 0.9, it can never be
reached. Within a 5 percent fluctuation of p, the equilibrium ratio drops down sharply from 1 to 0. This
is really striking, and we believe that it can be treated as a significant phase transition. The all-rebel
10
case will be explored more deeply in the next subsection, where we shall see that regular patterns emerge
steadily regardless of whether Nash equilibrium can be reached or not.
Pattern emergence
We show first the all-conformist case and then the all-rebel case.
In the figures below, each agent is
represented by a small square (for conformists, it is simply a square; for rebels, there is a dark triangle
in this square), red means that the corresponding agent chooses an action of 1 and green of 0.
Size=41X41
Size=101X101
Size=201X201
Figure 6. Emergence of continents. r=0, p=0.5. The upper three are initial configurations, and
the lower three are the corresponding stable ones.
We call the set of agents that are connected and share the same action a "continent". A torus with any
configuration can be decomposed into continents. Fig. 6 shows that toruses with initial configurations
have only relatively small continents. However, continents in the stable configurations arrived through
best response dynamics are in general impressively large. When the torus is small, the largest continent
can even cover more than a half of it. Obviously, an agent is completely satisfied if and only if it is in
the inner part of some continent, and incompletely satisfied, or unsatisfied, if and only if it is on the
"coastal line", i.e. the boundary of some continent. It can be proved trivially that complete ratio, i.e.
the percentage of completely satisfied agents, for any initialization is expectedly (0.5)8 .
= 0.0039. Large
continents in the stable configurations imply that the final complete ratio is high. In fact, calculation
shows that this value in our simulations (r=0) is in general larger than 0.5. This tells us that if we
consider the index of complete ratio as the coordinating ability of best response dynamics, the result is
still rather optimistic. "Lakes", i.e. small continents resting in bigger ones, can also be observed.
11
Initialization
P=0.2
P=0.4
P=0.6
P=0.8
P=0.85
P=0.86
P=0.87
P=0.88
P=0.89
Figure 7. Emergence of mazes. r=1,size=41×41.
Next, let's turn our attention to the all-rebel case, i.e. r = 1. An amazing property of this case is
that "mazes" frequently emerge. This is shown clearly in Fig. 7. If an agent is on the "street" of a maze,
rather than on the corner, then she has two neighbors taking the same action as she does, and six ones
taking the opposite action. Thus, her satisfaction degree is 6/8 = 0.75. Since almost all agents are on
streets, and only a few ones on corners, this explains the overall satisfaction degree 0.71 for r = 1 and
p ≤ 0.8 in Fig. 3 very well.
It should also be noted that every agent in the maze is satisfied, and thus a maze corresponds to a
Nash equilibrium. When r = 1, mazes emerge definitely for p ≤ 0.85, and never occur for p > 0.9. This is
also implied in Fig. 5. For updating probabilities between 0.85 and 0.9, mazes occur with some positive
probability strictly less than 1. It can also be observed that the lengths of streets in mazes generally
increase as p increases.
The last part of this subsection is devoted to the emergence for the cases that r = 1 and Nash
equilibrium is not attainable through best response dynamics, i.e. p > 0.9.
As shown in Fig. 8, regular patterns such as strips (straight ones as well as declining ones) and
continents can be frequently observed. Since rebels don't like her neighbors to take the same action as
she does, this is not a good thing. And this explains also why the cases that rebel ratios and updating
probabilities are both high behave so badly in Fig. 1-5. An extremely terrible situation is the one showed
at the upper left corner of Fig. 8, where almost all agents, except for very few ones (this number is
expectedly equal to (1 − p) × s, where s is the size of the torus), are completely unsatisfied. The patterns
we display in Fig. 8 are the only ones. We note that small wheels can also be frequently observed when
r = 1 and p = 1.
It you check really carefully the lower three pictures in Fig. 8, you will find that there are only two
types of rebels, i.e. the ones that keep switching their actions and the ones that never switch. This means
that the lengths of limit cycles of synchronous best response dynamics in our model are always 2. To put
it another way, synchronous best response dynamics will eventually lead the system to oscillate between
12
Figure 8. Emergence of strips, continents, and wheels. r=1. p=0.98 for the upper three
pictures, and p=1 for the lower three.
two states. The lower left state and the lower middle one in Fig. 8, in fact, compose such a pair. This
is consistent with the theoretical observation of Cannings [35] that synchronous best response dynamic
always lead a system with all rebels to a limit cycle of length 1 or 2, regardless of the structure. Limit
cycles of length 1, i.e. Nash equilibria, however, are never observed by us. This does not mean that they
do not exist at all, but that they are really rare. Initial configurations that will eventually reach a Nash
equilibrium, can be easily designed. For the simplest instance, we can let the initial configuration be an
equilibrium.
Up to now, we have discussed the emergence of regular patterns in the two extreme cases either with
all rebels or with all conformists. For the cases in between, similar patterns can be observed too. In a
word, there exists a spectrum and the patterns change over r continuously. When the rebel ratio r is
close to 0, the patterns are more like those in the all-conformist case, and as r tends to 1, the patterns
are more similar to those in the all-rebel case. But when r is near 0.5, the patterns are quite blurred
(we can also say that in these cases there is no pattern at all). What's definite is that when r < 0.2 or
r > 0.8, the patterns are rather clear and steady.
All the results of this subsection are obtained with the assistance of the excellent multi-agent pro-
grammable software NetLogo [59], and our program can be mailed at request.
13
Results on Small-World Networks
To check that whether the results discussed in the last section, which are derived from a rather special
network, are still valid in more realistic networks, we did more simulations on small-world networks
proposed by Watts and Strogatz ( [60]). The conclusion is quite promising: to a great extent, our results
are rather robust, they are still valid for the more realistic small-world networks. Needless to say, we are
unable to discuss the pattern emergence results, because they rely on geographic locations that do not
exist in small-world networks. And not surprisingly, we are not able to give convincing explanations to
several results as to why they are like that. For instance, we cannot give a similar explanation as to why
the average satisfaction degree is 0.71 for the all-rebel case.
Small-world networks are widely accepted as excellent mimics of the real social networks. The idea
of Watts and Strogatz's algorithm to generate an arbitrary small-world network is to derive it from a
regular network through "randomly rewiring" some edges. Roughly speaking, there is a uniform rewiring
probability, which is denoted as q in our paper (p in the paper of Watts and Strogatz), for each of the
original edge to be severed and replaced by a random edge (however, one endpoint of the new edge
must be taken from the two old ones). The advantage of introducing q is that it captures the degree of
randomness of the corresponding network: Larger q implies more randomness. When q = 0, the network
is regular, as in the last section of our paper, and when q = 1, the network is completely random (an
approximation of the classical ER random network). This gives us a new angle to study the fashion game,
i.e. how randomness affects the cooperation level of agents.
While there are only two parameters in the simulations of the previous section, namely rebel ratio
r and updating probability p, there are four in this section. In addition to the two old ones, there are
two new ones, network density (i.e. average degree) k and rewiring probability q. In our simulations, the
network size is fixed at 200, and (1) the network density k takes six values: 8, 18, 28, 38, 48 and 58, (2)
the rewiring probability q takes 11 values evenly from 0 to 1, (3) the rebel ratio r takes 11 values evenly
from 0 to 1, (4) the updating probability p takes 10 values evenly from 0.1 to 1. For each combination
of parameters, the result is based on the average of 10 simulations, and each simulation is terminated at
the 500-th step, as set in the previous section.
Calculation shows that the overall cooperation degree is 0.80. This value, though much lower than
the corresponding one (0.97) for the cellular automata network, is still promisingly high, confirming our
main result that, in fairly general and realistic scenarios, agents in the fashion game can reach high level
of cooperation through the simple best response dynamic.
It can also be observed from Fig. 9 that, for all the six cases with different network densities, rewiring
probability q always plays a negative role at small values (less than 0.1). But once exceeding the threshold
of 0.1, increasing q will be good to the cooperation degree, though the increasing speed of the right part
is much lower than the decreasing speed of the left part. The maximum values of cooperation degrees
are always obtained at q = 0, and the values at q = 1 are only a little bit smaller than the maxima.
This tells us that randomness of networks plays some very interesting role. Completely regular networks
are the best of all for reaching high degree of cooperation, completely random networks are only a little
worse, and networks with a relatively low randomness of about 0.1 are the worst of all.
Fig. 9 indicates also that high density is always an obstacle for cooperation degree. Intuitively, this
is not hard to understand: the more neighbors you have, the harder for you to cooperate.
Since the cellular automata network we study in the previous section is both completely regular and
with a low density, no wonder that we observed an amazingly high degree of cooperation.
Now let's turn our attention from cooperation degree to another index, the average satisfaction degree.
As discussed for the cellular automata network, this index is always much smaller than cooperation degree.
Calculation shows that its overall average value is 0.6, still significantly higher than the initially expected
value of 0.5, and confirms our claim again that best response dynamic promotes cooperation. As for
the index of cooperation degree, high density of networks plays a negative role for average satisfaction.
This effect is also quite robust. The effect of rewiring probability q, however, is unclear for low density
14
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
0
Figure 9. Effects of rewiring probability q and network density k on cooperation degree
and average satisfaction degree. The horizontal axis is for the rewiring probability, solid blue lines
for cooperation degree and dashed red ones for average satisfaction degree. Different densities of
k = 8, 18, 28, 38, 48, 58 are represented by ◦, ×, +, ∗, (cid:3), and ♦, respectively.
networks (k = 8, 18, 28). For high density networks (k = 38, 48, 58), its effect is steadily negative. An
interesting observation from the network with density k = 58 is that, if we let q grow from 0.1 to 1 and
compare the index of cooperation degree and that of average satisfaction degree, we can find that the
former index increases while the latter one decreases slowly. In this process, more and more agents are
satisfied, but their average satisfaction degree becomes worse. This implies that more and more agents
must have behaved compromisingly, leading the society to more equality. To put it another way, in the
case that the network is highly dense and not so regular, adding randomness (or decreasing regularity)
of people's interactions can improve equality of the society.
It is valuable to remark that if we abandon the normalization in the definition of average satisfaction
degree, i.e. define the absolute satisfaction degree of each agent as the number of neighbors that s/he
likes, then we could find that adding density to people's interaction networks would be a good thing in
general. This is not surprising, because having more neighbors means that you have more chances to get
rewarded from interacting with them (notice that in the above argument the disadvantage from neighbors
is not calculated).
As to the other two indices, i.e. equilibrium ratio and complete ratio, which are auxiliary in this
paper, their overall average values are 0.45 and 0.09, respectively. Considering their definitions, these
values are not low at all. The effects of rewiring probability q and network density k, however, are not
stable on neither of them.
Let's consider next the effects of rebel ratio r and updating probability p. The simulation results,
which excellently confirm our claims for the cellular automata network that high updating probability
is always bad for cooperation and high rebel ratio is also bad when p is large, are displayed in Fig. 10.
When p is small, the effect of r on cooperation level is roughly a pattern of "V". This pattern can
also be observed for the cellular automata network in Fig. 3. For small-world networks, however, it is
much more significant, especially if we only concentrate on the index of equilibrium ratio. This result is
plausible because in the two extreme cases with all conformists or all rebels, Nash equilibrium can always
be guaranteed, but not in other cases. And the more balanced the proportion between conformists and
rebels, the more difficult it is to reach a Nash equilibrium or a high average degree of satisfaction. It
15
Figure 10. Effects of rebel ratio r and updating probability p on cooperation degree
(Top-Left), average satisfaction degree (Top-Right), equilibrium ratio (Bottom-Left), and
complete ratio (Bottom-Right). All horizontal axises represent rebel ratio r.
is quite remarkable that updating probability p has absolutely no effect at all to the index of complete
ratio. This ratio is very high for the case with all conformists. However, once the rebel ratio r increases
a little bit, it drops down dramatically, and then, very soon, approximates zero.
Phase transition of equilibrium ratio is also confirmed in the more general settings, as displayed in Fig.
11. It should be noted that significance of this phenomenon is negatively correlated with the rewiring
probability q. For completely regular networks (q = 0), phase transition is rather significant. However,
as randomness of the network grows, this phenomenon becomes less and less significant. In fact, when q
is larger than 0.3, it is farfetched to call them phase transitions any more. Since the cellular automata
network is completely regular, no wonder that we observed a beautiful phase transition.
16
q=0
q=0.1
q=0.2
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Figure 11. Phase transition of equilibrium ratio in updating probability p in small-world
networks (k = 8, r = 1).The horizontal axis is for p.
Discussion
The fashion game is investigated in this paper through simulations. Our focus is on cooperation of agents.
To study this, various indices are used, namely the cooperation degree, the average satisfaction degree,
the equilibrium ratio, and the complete ratio. Our finding is quite promising:
in most cases, agents
can cooperate rather well through best response dynamics. It tells us also that the interaction structure
matters a lot. Considering that these dynamics are really simple, agents are selfish, myopic, naive, and
have very limited information, this finding is fairly surprising.
It is valuable to note that small-world networks used in the last section are generated by the original
algorithm of Watts and Strogatz. This algorithm, though extremely popular and widely accepted, does
not take the cellular automata network but a ring type regular network as the benchmark network
(p = 0). As suggested by one anonymous referee, it is natural to take the cellular automata network
as the benchmark network. This treatment will bridge the two kinds of networks we study seamlessly.
We call networks generated in this way the modified small-world networks. For these modified small-
world networks (with density k = 8), we did the same group of simulations as in the previous section. For
technical reasons, we didn't take the network size so large as for the cellular automata network (recall that
the size there is 1681), but a normal size of 441, about twice the size we used in the previous section for the
original small-world networks (200). We made a comparison between the original and the modified small-
world networks. It turns out that, as shown in Fig. 12, the two groups of networks make no meaningful
difference for three of the four indices. This might indicate that, compared with network density and
randomness level, details such as how they are actually connected matter very little. Of course, this
broader guess should be confirmed or falsified by more extensive simulations, and related further studies
include how degree distribution and clustering coefficient affect agents' behavior and consequently their
cooperation level in the fashion game. Network size, though does not affect cooperation degree, average
satisfaction degree, or complete ratio significantly, also as shown in Fig. 12, it may well matter a lot to
equilibrium ratio. And the effect must be on the negative side. This is intuitive, because the definition
of an equilibrium is quite restrictive. As long as there is one agent that is unsatisfied, an action profile
cannot be an equilibrium. This possibility increases as the network size grows.
In this paper, we mainly focus on the overall cooperation levels, i.e. the macro side. For further
17
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Figure 12. Comparison between original small-world networks (dashed red lines) and
modified small-world networks (solid blue lines).The horizontal axis is for rewiring probability p.
◦ is for the index of cooperation degree, ∗ for average satisfaction degree, (cid:3) for equilibrium ratio, and ♦
for complete ratio.
research, it is also meaningful to investigate the micro side of the fashion game, i.e. how the payoffs of
agents are affected by their status in the networks, measured by various centrality indices, say their degree
centralities, their betweenness centralities, their closeness centralities, and their clustering coefficients.
To study these problems, the utility function as defined in the introduction part of this paper should be
given at least equal attention as the satisfaction degree.
Another obvious direction is to study the kind of realistic network, scale-free network [63]. "Ho-
mophily", a basic observation from the real world that "birds of a feature flock together", is attracting
more and more attention recently (c.f. [64, 65]), and also deserves the discussion of its effect on the fash-
ion game. It is natural to conjecture that cooperation level is positively correlated with homophily level.
Throughout this paper, we have assumed that the initial settings are uniformly random, that is, each
agent takes action 1 with probability 0.5 and action 0 with probability 0.5. It is interesting to consider
nonuniform initializations, say action 1 is biased. Another interesting future direction is to take into
account the possibility of network formation, where forming a new link (or severing an old one, or both)
is an extra strategy of each agent. Assuming that there is a cost to this new strategy, it is very meaningful
to study the co-evolution of cooperation level and network structure.
Acknowledgements
We thank professor Xiang Li for reminding us of the connection between the anti-coordination game and
the snowdrift game, and Dr. Yuqing Qiu for technical helps. We are also indebted to two anonymous
referees for their helpful suggestions and comments.
In particular, more simulations on small-world
networks and homophily and network formation as interesting future directions were suggested by one of
them, references [28] and [58] were also informed by her or him.
References
1. Okonkwo U (2007) Luxury Fashion Branding: Trends, Tactics, Techniques. Palgrave Macmillan.
2. Euromonitor International (2011) Global luxury goods overview, June 2011.
18
3. van Nes N (2010) Understanding replacement behaviour and exploring design solutions. In Longer
Lasting Products: Alternatives to the Throwaway Society, edited by T. Cooper.
4. Schweitzer F, Mach R (2008) The epidemics of donations:
logistic growth and power-laws. PLoS
ONE 3(1): e1458. doi:10.1371/journal.pone.0001458
5. Svendsen L (2006) Fashion: A Philosophy. Reaktion Books.
6. Bell D (1973) The Coming of Post-industrial Society: A venture in Social Forecasting. New York:
Basic Books.
7. Jackson M (2008) Social and Economic Networks. Princeton University Press, Princeton, NJ.
8. Banerjee AV (1992) A simple model of herd behavior. The Quarterly Journal of Economics 107:
797-817.
9. Bikhchandani B, Hirshleifer D, Welch I (1992) A theory of fads, fashion, custom, and cultural change
as informational cascades. Journal of Political Economy 100(5): 992-1026.
10. Chai A, Earl PE, Potts J (2007) Fashion, growth and welfare: an evolutionary approach. Advances
in Austrian Economics 10: 187-207.
11. Pesendorfer W (1995) Design innovation and fashion cycles. American Economic Review 85(4):
771-792.
12. Acerbi A, Ghirlanda S, Enquist M (2012) The logic of fashion cycles. PLoS ONE, 7(3): e32541.
doi:10.1371/journal.pone.0032541
13. Young HP (2001) Individual Strategy and Social Structure: An Evolutionary Theory of Institu-
tions. Princeton, NJ: Princeton University Press.
14. Cao Z, Yang X (2012) The Fashion Game:
Matching Pennies On Social Net-
2012). Available at SSRN: http://ssrn.com/abstract=1767863 or
(September 24,
works
http://dx.doi.org/10.2139/ssrn.1767863
15. Osborne M, Rubinstein A (1994) A Course on Game Theory. MIT Press.
16. Kalai A, Kalai E (2010) Cooperation in strategic games revisited, Games and Economic Behavior.
17. Galeotti A, Goyal S et. al. (2010) Network games. The Review of Economic Studies 77(1): 218-244.
18. Szabo G, Fath G (2007) Evolutionary games on graphs. Physics Reports 446: 97-216.
19. Kearns M, Littman M, Singh S (2001) Graphical models for game theory. In Proc. Conf. on
Uncertainty in Artificial Intelligence: 253-260.
20. Jackson MO, Zenou Y (2014) Games on Networks. Handbook of Game Theory, Vol. 4, Peyton
Young and Shmuel Zamir, eds., Elsevier Science, 2014. Available at SSRN (August 25, 2012):
http://ssrn.com/abstract=2136179
21. Berninghaus SK, Schwalbe U (1996) Conventions, local interaction, and automata networks. Jour-
nal of Evolutionary Economics 6: 297-312.
22. Ellison G (1993) Learning, local interaction, and coordination. Econometrica 61(5): 1047-1071.
23. Blume L (1993) The statistical mechanics of strategic interaction. Games and Economic Behavior
5: 387-424.
19
24. Blume L (1995) The statistical mechanics of best-response strategy revision. Games and Economic
Behavior 11: 111-145.
25. Morris S (2000) Contagion. The Review of Economic Studies 67(1): 57-78.
26. Brock W, Durlauf S (2001) Discrete choice with social interactions. The Review of Economic
Studies 68(2): 235-260.
27. Jackson MO, Watts A (2002) On the formation of interaction networks in social coordination
games. Games and Economic Behavior 41: 265-291.
28. Watts DJ (2002) A simple model of global cascades on random networks. Proceedings of the
National Academy of Sciences of the United States of America, 99(9): 5766-5771.
29. Goyal S, Vega-Redondo F (2005) Network formation and social coordination. Games and Economic
Behavior 50: 178-207.
30. Bramoulle Y (2007) Anti-coordination and social interactions. Games and Economic Behavior 58:
30-49.
31. Bramoulle Y, Lopez-Pintado D, Goyal S, Vega-Redondo F (2004) Network formation and anti-
coordination games. International Journal of Game Theory 33(1): 1-19.
32. Lopez-Pintado D (2009) Network formation, cost sharing and anti-coordination. International
Game Theory Review 11(1): 53-76.
33. Cao Z, Yang X (2012). A note on anti-coordination and social interactions. Journal of Combinatorial
Optimization. online first, DOI: 10.1007/s10878-012-9486-7
34. Anderlini L, Ianni A (1993) Path dependence and learning from neighbours. Mimeo, University of
Cambridge.
35. Cannings C (2009) The majority and minority models on regular and random graphs. International
Conference on Game Theory for Networks.
36. Challet D, Zhang YC (1997) Emergence of cooperation and organization in an evolutionary game.
Physica A 246: 407-418.
37. Challet D, Marsili M, Zhang YC (2004) Minority Games: Interacting Agents in Financial Markets.
Oxford Univ. Press: Oxford.
38. Marsili M (2001) Market mechanism and expectations in minority and majority games. Physica A
299: 93-103.
39. de Martino A, Gardino I, Mosetti G (2003) Statistical mechanics of the mixed majority-minority
game with random external information. Journal of Physics A: Mathematical and General 36: 3935-
54.
40. Gou C (2006) The simulation of financial markets by an agent-based mix-game model. Journal of
Artificial Societies and Social Simulation 9(3): 321-328.
41. Gou C (2006) Agents play mix-game. Econophysics of Stock and other Markets: 123-132.
42. Gou C (2006) Deduction of initial strategy distributions of agents in mix-game models. Physica A:
Statistical Mechanics and its Applications 371(2): 633-640.
20
43. Challet D (2008) Inter-pattern speculation: beyond minority, majority and $-games. Journal of
Economic Dynamics and Control 32(1): 85-100.
44. Caporale GM, Serguieva A, Wu H (2008) A mixed-game agent-based model for simulating financial
contagion. In Proceedings of the 2008 Congress on Evolutionary Computation, IEEE Press: 3420-
3425.
45. Chen F, Gou C, Guo X, Gao J (2008) Prediction of stock markets by the evolutionary mix-game
model. Physica A 387(14): 3594-3604.
46. Liu F, Serguieva A, Date P (2010) A mixed-game and co-evolutionary genetic programming agent-
based model of financial contagion. in 2010 IEEE Congress on Evolutionary Computation (CEC):
1-7.
47. Acemoglu D, Ozdaglar A (2010) Opinion dynamics and learning in social networks. Dynamic Games
and Applications 1(1): 3-49.
48. Golub B, Jackson M (2010) Naive learning in social networks and the wisdom of crowds. American
Economic Journal: Microeconomics, 2(1): 112-149.
49. Cao Z, Yang M, Qu X, Yang X (2011) Rebels lead to the doctrine of the mean: opinion dynamic
in a heterogeneous DeGroot model. The 6th International Conference on Knowledge, Information
and Creativity Support Systems Beijing, China: 29-35.
50. Galam S (2004) Contrarian deterministic effects on opinion dynamics:"the hung elections scenario".
Physica A: Statistical and Theoretical Physics 333: 453-460.
51. Galam S (2008) Social physics: a revew of Galam models. International Journal of Modern Physics
C 19(3): 409-440.
52. Hong H, Strogatz SH (2011) Kuramoto model of coupled oscillators with positive and negative
coupling parameters: An example of conformist and contrarian oscillators. Physical Review Letters
106(5): 054102.
53. Hong H, Strogatz SH (2011) Conformists and contrarians in a Kuramoto model with identical
natural frequencies. Physical Review E 84(4): 046202.
54. Hong H, Strogatz SH (2012) Mean-field behavior in coupled oscillators with attractive and repulsive
interactions. Physical Review E 85(5): 056210.
55. Eliaza K, Rubinstein A (2011) Edgar Allan Poe's riddle: framing effects in repeated matchingpen-
niesgames. Games and Economic Behavior 71(1): 88-99.
56. Chen
S, Bao S
(2008) A game
for Game Theory
ter
http://www.gtcenter.org/Archive/2010/Conf/Chen956.pdf. Accessed 2012 October 13.
in
theory based
Economics Archive,
predation behavior model. Cen-
at
2010 Conference,
available
57. Jackson M, Shoham Y (2012) Lecture Notes in Game Theory, available at Coursera :
https://class.coursera.org/gametheory/lecture/index. Accessed 2012 October 13.
58. L´opez-Pintado D, Watts DJ (2008) Social influence, binary decisions and collective dynamics.
Rationality and Society, 20(4): 399-443.
59. Wilensky U (1999) NetLogo. http://ccl.northwestern.edu/netlogo/. Center for Connected Learning
and Computer-Based Modeling, Northwestern University, Evanston, IL.
21
60. Watts DJ, Strogatz SH (1998) Collective dynamics of small-world networks. Nature, 393: 440-442.
61. Hart S, Mas-Colell A (2003) Uncoupled dynamics do not lead to Nash equilibrium. American
Economic Review 93: 1830-1836.
62. Monderer D, Shapley L (1996) Potential games. Games and Economic Behavior 14: 124-143.
63. Barab´asi A, Albert R (1999) Emergence of scaling in random networks. Science 286: 509-512.
64. McPherson M, Smith-Lovin L, Cook JM (2001) Birds of a feather: homophily in social networks.
Annual Review of Sociology 27: 415-444.
65. Aral S, Muchnik L, Sundararajan A (2009) Distinguishing influence-based contagion from
homophily-driven diffusion in dynamic networks. Proceedings of the National Academy of Sciences
of the United States of America 106(51): 21544-21549.
|
1011.6127 | 1 | 1011 | 2010-11-29T03:57:51 | Visibility maintenance via controlled invariance for leader-follower Dubins-like vehicles | [
"cs.MA"
] | The paper studies the visibility maintenance problem (VMP) for a leader-follower pair of Dubins-like vehicles with input constraints, and proposes an original solution based on the notion of controlled invariance. The nonlinear model describing the relative dynamics of the vehicles is interpreted as linear uncertain system, with the leader robot acting as an external disturbance. The VMP is then reformulated as a linear constrained regulation problem with additive disturbances (DLCRP). Positive D-invariance conditions for linear uncertain systems with parametric disturbance matrix are introduced and used to solve the VMP when box bounds on the state, control input and disturbance are considered. The proposed design procedure is shown to be easily adaptable to more general working scenarios. Extensive simulation results are provided to illustrate the theory and show the effectiveness of our approach | cs.MA | cs | Visibility maintenance via controlled invariance
for leader-follower Dubins-like vehicles
0
1
0
2
v
o
N
9
2
]
A
M
.
s
c
[
1
v
7
2
1
6
.
1
1
0
1
:
v
i
X
r
a
Fabio Morbidi⋆ a, Francesco Bullo b, Domenico Prattichizzo c
aDept. of Computer Science and Engineering, University of Texas at Arlington, TX, 76019, USA
bDept. of Mechanical Engineering and Center for Control, Dynamical Systems and Computation, University of California at
Santa Barbara, CA, 93106-5070, USA.
cDept. of Information Engineering, University of Siena, 53100 Siena, Italy
Abstract
The paper studies the visibility maintenance problem (VMP) for a leader-follower pair of Dubins-like vehicles with input
constraints, and proposes an original solution based on the notion of controlled invariance. The nonlinear model describing
the relative dynamics of the vehicles is interpreted as linear uncertain system, with the leader robot acting as an external
disturbance. The VMP is then reformulated as a linear constrained regulation problem with additive disturbances (DLCRP).
Positive D-invariance conditions for linear uncertain systems with parametric disturbance matrix are introduced and used to
solve the VMP when box bounds on the state, control input and disturbance are considered. The proposed design procedure
is shown to be easily adaptable to more general working scenarios. Extensive simulation results are provided to illustrate the
theory and show the effectiveness of our approach.
Key words: Autonomous mobile robots; Visibility maintenance; Leader-follower; Controlled invariance; Input constraints
1 Introduction
1.1 Problem description and motivation
This paper considers a number of visibility mainte-
nance problems between autonomous vehicles. The sim-
plest formulation is a leader-follower setup, in which
both leader and follower are nonholonomic vehicles
constrained to move along planar paths of bounded
curvature, with limited positive forward speed. The
follower vehicle's goal is to maintain the leader inside
an appropriate sensing region. The theory of controlled
invariance for uncertain linear systems is shown to be
well suited for this objective as well as for achieving
more involved tasks, such as, e.g., simultaneously reject
unknown but bounded disturbances or preserve visibi-
lity in multi-vehicle chain formations.
We have two main motivations for the visibility main-
tenance problems studied in this article. First, we are
interested in surveillance and patrolling problems with
formations of robotic vehicles in aerial and ground
environments. We envision scenarios where a robot
equipped with sensors with limited sensing footprints
(such as, e.g., panoramic cameras, laser range finders,
or high resolution radars) moves in such a way as to
maintain a second moving target within its field of view.
Second, this work is motivated by the need to design
network-wide visibility and connectivity maintenance
algorithms for groups of robotic vehicles. In the multi-
agent network domain, connectivity is indeed a classic
requirement necessary to guarantee the correct comple-
tion of numerous distributed algorithms.
1.2 Literature review
⋆ Corresponding author. This material is based upon work
supported in part by ONR Award N00014-07-1-0721. This
paper was not presented at any IFAC meeting.
Email addresses: [email protected] (Fabio
Morbidi⋆), [email protected]
(Francesco Bullo), [email protected]
(Domenico Prattichizzo).
In the context of visibility maintenance between pairs of
vehicles, two distinct literature domains are relevant to
this work. First, visibility-based pursuit-evasion prob-
lems for robots in complex environments have been in-
vestigated in continuous-time in (Guibas et al., 1997;
Gerkey et al., 2006; Bhattacharya and Hutchinson, 2010)
and in discrete-time in (Isler et al., 2005): in these works
the vehicles' dynamic models are elementary and the
proposed solutions are not applicable to nonholonomic
vehicles with limited positive forward speed. Second,
the vast literature on aircraft pursuit-evasion has fo-
cused much attention to game theory, optimal control,
and numerical algorithms. A hybrid-systems and game-
theoretic approach to aircraft conflict resolution is pur-
sued in (Tomlin et al., 2000). Differential game problems
between aircraft are discussed in (Merz and Hague, 1977;
Jarmark and Hillberg, 1984; Shima and Shinar, 2002).
In (Glizer, 1999), a planar pursuit-evasion problem in
which the target set is defined by a capture radius and
constraints on the angular state variables (line-of-sight
angle) is analyzed, and a necessary and sufficient con-
dition for capture of the evader from any initial state is
established using a variational method. However, differ-
ently from the problem studied in this paper, the author
considers constant positive forward speed for the non-
holonomic vehicles and unlimited turning rate for the
evader. More recently, in (Mazo Jr et al., 2004), the prob-
lem of estimating and tracking the motion of a moving
target by a team of unicycles equipped with directional
sensors with limited range, is addressed using a hierar-
chical control scheme.
In the context of connectivity maintenance in multi-
agent networks, the literature has experienced a recent
spurt of growth (we refer to (Bullo et al., 2009; Mesbahi
and Egerstedt, 2010) for recent surveys on this topic):
two typical multi-agent tasks requiring network connec-
tivity are "consensus" (Moreau, 2005; Olfati-Saber et
al., 2007) and "rendezvous" (Cort´es et al., 2006; Lin et
al., 2007). In this active research area, robots with li-
mited communication capability are often modeled as
transmitters with disks of finite radius. As the vehicles
move to achieve a goal, it is generally hard to guaran-
tee the connectivity among the members of the group
is preserved over time. In terms of design, it is then re-
quired to constrain robots' control inputs such that the
resulting topology is always connected throughout its
course of evolution. Potential fields and geometric opti-
mization methods are the standard tools used in the lit-
erature to address the connectivity maintenance prob-
lem: a list of key references, yet far from being complete,
is (Ando et al., 1999; Spanos and Murray, 2005; Ji and
Egerstedt, 2007; Zavlanos and Pappas, 2007; Dimarog-
onas and Kyriakopoulos, 2008; Savla et al., 2009; Yang
et al., 2010). These works differ from the setup pro-
posed in this paper in at least two important ways: first,
the vehicle's dynamics is assumed to be locally con-
trollable, and second, connections between two robots
are bidirectional. In our problem, instead, we consider
nonholonomic vehicles that are not locally controllable
(they move forward with positive speed, along paths
of bounded curvature) and we deal with sensor foot-
prints such that the visibility links are not bidirectional.
We finally point out that most of the works cited above
do not explicitly account for robots' input constraints.
1.3 Original contributions
The basic setup considered in this paper consists of
two nonholonomic agents: a leader (or evader) L and
a follower (or pursuer) F. The robots can rotate with
bounded angular velocity, but similarly to Dubins'
vehicles (Dubins, 1957) can only move forward. The
follower is equipped with a sensing device characterized
by a visibility set S, a compact and convex polyhedral
region encoding both the position and angle informa-
tion. The leader moves along a given trajectory and
the follower aims at maintaining L always inside its
visibility set S, while respecting suitable bounds on
the control inputs. Inspired by (Tiwari et al., 2004),
where the concept of cone invariance is used to solve
the multiagent rendezvous problem and by the results
in (Blanchini, 1990; Blanchini, 1991), this paper ad-
dresses the visibility maintenance problem (VMP) using
the notion of controlled invariance. The key idea is to
interpret the nonlinear model describing the relative
dynamics of the leader and the follower, as a linear sys-
tem with model parameter uncertainty, with the control
input of the leader playing the role of an external di-
sturbance. The VMP can then be easily reformulated as
linear constrained regulation problem with additive dis-
turbances (DLCRP). Positive D-invariance conditions
for general linear uncertain systems with parametric
disturbance matrix are introduced and used to solve the
VMP when box bounds on the state, control inputs and
disturbances are considered. Analytical conditions for
the solution of the VMP are obtained by symbolically
solving with the Fourier-Motzkin elimination method,
the set of linear inequalities defining the polytope of all
the feasible state feedback matrices. The proposed de-
sign procedure can be easily adapted to provide the con-
trol with unknown but bounded (UBB) disturbances re-
jection capabilities. Other extensions are also discussed:
we present conditions for the solution of the VMP when
robots' desired displacement is defined through angular
parameters instead of distances, and extend the results
valid for a leader-follower pair of robots to chains of
n vehicles. Extensive simulation results illustrate the
theory in the different working scenarios. The present
paper builds upon (Morbidi et al., 2008), compared to
which we provide herein a more detailed and extended
theory, as well as a more accurate numerical validation.
1.4 Organization
The rest of the paper is organized as follows. In Sect. 2
the linear constrained regulation problem is reviewed
and new positive D-invariance conditions for linear
systems with parameter uncertainty are presented.
In Sect. 3 we introduce the VMP and prove the main
result of the paper. In Sect. 4 we investigate some exten-
sions of the basic setup of Sect. 3. In Sect. 5, simulation
results are presented. In Sect. 6 the main contributions
of the paper are summarized and possible avenues for
future research are highlighted.
2
2 The linear constrained regulation problem
to the following condition: for every s0 ∈ ∂S and q ∈ Q,
This section presents a series of results that are
instrumental in addressing the visibility maintenance
problem in Sect. 3. Our exposition will basically fol-
low (Blanchini, 1991): Theorem 10, Corollary 11,
Theorem 12 and Corollary 13 extend the corre-
sponding results in (Blanchini, 1990; Blanchini, 1991)
(see also (Blanchini and Miani, 2008, Ch. 4)), to linear
uncertain systems with parametric disturbance matrix.
Consider the following system,
s(t) = A(q(t)) s(t) + B(q(t)) u(t),
(1)
where s(t) ∈ X ⊂ Rn and u(t) ∈ U ⊂ Rm are respec-
tively the state and input vectors, q(t) ∈ Q ⊂ Rp is the
model parameter uncertainty vector, while U, X , Q are
assigned sets containing the origin, with U and Q com-
pact. We assume that A(q) and B(q) are matrices of suit-
able dimensions whose entries are continuous functions
of q. We will suppose q(t) to be a piecewise continuous
function of time.
Definition 1 (Positive invariance) The set S ⊂ Rn
is positively invariant for system (1), if and only if, for
every initial condition s(0) ∈ S and every admissible
q(t) ∈ Q, the solution obtained for u(t) ≡ 0, satisfies the
condition s(t) ∈ S for t > 0.
Definition 2 (Admissible region) A region S ⊂ Rn
is said to be admissible for the feedback control law u =
Ks, if and only if, for every s ∈ S, the condition u ∈ U
holds. If U and S are convex polyhedral sets containing
the origin, the admissibility of S is simply equivalent to
Kvi ∈ U, vi ∈ vert(S), i ∈ {1, . . . , µ},
(2)
where vert(S) denotes the set of vertices of S.
We can now introduce the linear constrained regulation
problem (LCRP), (Blanchini, 1991).
Problem 3 (LCRP) Given a system in the form (1),
find a linear feedback control law u(t) = Ks(t) and a set
S ⊂ X such that, for every initial condition s(0) ∈ S
and every admissible function q(t) ∈ Q, the conditions
s(t) ∈ X and u(t) ∈ U are fulfilled for t > 0.
Theorem 4 The LCRP has a solution if and only if
there exists a feedback matrix K and a set S ⊂ X that
is positively invariant and admissible for the closed loop
system
s(t) = F (q(t)) s(t),
(3)
where F (q(t)) = A(q(t)) + B(q(t))K.
Theorem 5 (Sub-tangentiality condition) Let
S ⊂ Rn be a compact and convex set with nonempty in-
terior. The positive invariance of S for (1) is equivalent
A(q) s0 ∈ TS(s0),
(4)
where TS(s0) is the tangent cone to S at s0 (see (Blanchini,
1999, Def. 3.1), and (Aubin and Frankowska, 1990,
Ch. 4), (Aubin and Cellina, 1984, Ch. 5)) for more
details).
The main difficulty in exploiting condition (4) to study
the positive invariance of an assigned region S is that
it has to be checked on the boundary of S. However, if
convex polyhedral sets are considered, only their vertices
must be taken into account and easy algebraic conditions
can be derived. In this respect, let us consider a system
of the form (1), with
p
p
A(q(t)) = A0 +
Al ql(t), B(q(t)) = B0 +
(5)
where Al and Bl, l ∈ {1, . . . , p}, are constant matrices
of appropriate dimension and q(t) takes values in a com-
pact and convex polyhedron Q ⊂ Rp, (ql(t) denotes the
l-th component of vector q(t)). Let the set U be compact,
convex and polyhedral as well. We consider a candidate
compact and convex polyhedral set S containing the ori-
gin in its interior and we search for a feedback matrix K
that assures the positive invariance of S for the closed
loop system (3), (note that the previous assumptions on
S will be retained throughout this section). Since S is
polyhedral, then condition (4) is fulfilled on ∂S if and
only if is fulfilled on every vertex of S.
Xl=1
Bl ql(t),
Xl=1
Theorem 6 The set S is positively invariant for system
(3) with feedback u = Ks, if and only if, for all vi ∈
vert(S) and wj ∈ vert(Q) :
F (wj ) vi ∈ TS(vi),
i ∈ {1, . . . , µ}, j ∈ {1, . . . , ν}.
The LCRP as formulated in Problem 3 does not require
the stability. However, a desirable property is the global
uniform stability of the closed loop system. The link be-
tween the stability property and the existence of pos-
itively invariant regions is established by Theorem 5.2
in (Blanchini, 1991).
Let us now turn our attention to systems in the form
s(t) = A(q(t)) s(t) + B(q(t)) u(t) + E(q(t)) δ(t),
(6)
where the unknown external disturbance δ(t) is con-
strained in a compact and convex polyhedral set D ⊂ Rl
containing the origin. Note that with respect to the sys-
tems considered in (Blanchini, 1991), the structure in (6)
is more general inasmuch as matrix E also depends on
the uncertain parameter q. As an immediate extension
of the positive invariance property introduced in Defi-
nition 1, we may require that the state s remains in S
despite the presence of the disturbance δ(t).
3
Definition 7 (Positive D-invariance) The set S ⊂
Rn is positively D-invariant for system (6), if for every
initial condition s(0) ∈ S and all admissible q(t) ∈ Q
and δ(t) ∈ D, the solution obtained for u(t) ≡ 0, satisfies
the condition s(t) ∈ S for t > 0.
We can now introduce the linear constrained regulation
problem with additive disturbances (DLCRP).
Problem 8 (DLCRP) Given a system in the form (6),
find a linear feedback control law u(t) = Ks(t) and a set
S ⊂ X such that, for every initial condition s(0) ∈ S and
every admissible q(t) ∈ Q and δ(t) ∈ D, the conditions
s(t) ∈ X and u(t) ∈ U are fulfilled for t > 0.
Theorem 9 The DLCRP has a solution if and only if
there exists a feedback matrix K and a set S ⊂ X that is
positively D-invariant and admissible for the closed loop
system s(t) = F (q(t)) s(t) + E(q(t)) δ(t).
Similarly to (5), we will henceforth suppose that
l=1 El ql(t). We are now ready to state
E(q(t)) = E0 +Pp
the main theorem of this section.
Theorem 10 (Main result) The set S is positive
D-invariant for system (6) with feedback u = Ks,
if and only if, for all vi ∈ vert(S), ωj ∈ vert(Q) and
rk ∈ vert(D),
F (wj) vi + E(wj ) rk ∈ TS(vi),
i ∈ {1, . . . , µ}, j ∈ {1, . . . , ν}, k ∈ {1, . . . , η}.
(7)
Proof: The proof is based on the same ideas as those in
(Blanchini, 1990, Th. 2.1) and (Blanchini, 1991, Th. 4.1)
(see also (Aubin and Cellina, 1984, Ch. 2, Sect. 4)) and
the references therein). For the necessity, we have to
prove that if S is a positive D-invariant region for system
(6), then condition (7) holds. The proof is straightfor-
ward, since for the sub-tangentiality condition the posi-
tive D-invariance of S for system (6) is equivalent to
Let wlj be the l-th entry of wj. We have that
F (q) vi + E(q) rk =
p
p
ν
p
Fl
Xl=1
=(cid:0)F0 +
=(cid:0)F0 +
Xj=1
Xj=1
Xl=1
Fl ql(cid:1)vi +(cid:0)E0 +
Xl=1
Xj=1
βj wlj(cid:1)vi +(cid:0)E0 +
Xj=1
Xl=1
Fl wlj(cid:1)vi(cid:3)+
βj(cid:2)(cid:0)F0 +
βj(cid:2)F (wj ) vi + E(wj) rk(cid:3),
=
=
p
ν
ν
ν
p
El
El ql(cid:1)rk
Xl=1
Xj=1
βj wlj(cid:1)rk
Xl=1
El wlj(cid:1)rk(cid:3)
βj(cid:2)(cid:0)E0 +
p
ν
i ∈ {1, . . . , µ}, k ∈ {1, . . . , η}.
From (7) we have thatPν
j=1 βj(cid:2)F (wj ) vi + E(wj ) rk(cid:3) ∈
h=1 αh vh, with gT
i vh = ξi and Pµi
s = Pµi
i (F (q)Pµi
i (cid:0)F (q)Pµi
TS(vi), i ∈ {1, . . . , µ}, k ∈ {1, . . . , η}, therefore (9)
is proved. If πi is a delimiting plane of S for s (i.e.,
such that gT
i s = ξi), we may write s as a convex
combination of the vertices of S that belong to πi:
h=1 αh = 1,
0 ≤ αh ≤ 1, h ∈ {1, . . . , µi}. Then gT
i (F (q)s + E(q) δ) =
gT
k=1 ρk rk). But from (9)
and recalling the expression of the tangent cone
when S is described in terms of
its vertices, we
have that gT
i (F (q) vh + E(q) rk) ≤ 0, that implies
gT
sidering all the planes for s, condition (8) follows
(cid:4)
immediately.
The application of Theorem 10 requires the knowledge
of all cones TS(vi), i ∈ {1, . . . , µ}. An alternative solu-
tion is given by the following corollary (whose proof is
analogous to that of (Blanchini, 1991, Corollary 4.1))
in which the Euler auxiliary system associated to (6) is
involved (cf. (Blanchini and Miani, 2008, Sect. 12.1)).
h=1 αh vh + E(q)Pη
h=1 αhvh + E(q)Pη
k=1 ρk rk(cid:1) ≤ 0. By con-
Corollary 11 The set S is positively D-invariant for
system (6), if and only if there exists τ > 0 such that, for
all vi ∈ vert(S), ωj ∈ vert(Q) and rk ∈ vert(D),
F (q) s + E(q) δ ∈ TS(s), s ∈ ∂S, q ∈ Q, δ ∈ D,
(8)
vi + τ (F (wj ) vi + E(wj ) rk) ∈ S,
(10)
that trivially implies condition (7). For sufficiency, let
us consider s arbitrary in S, q arbitrary in Q and δ
arbitrary in D. Supposing condition (7) holds, inclu-
i=1 αi,
i=1 αi = 1,
k=1 ρk = 1, for some 0 ≤ αi ≤ 1, i ∈
{1, . . . , µ}, 0 ≤ βj ≤ 1, j ∈ {1, . . . , ν} and 0 ≤ ρk ≤ 1,
k ∈ {1, . . . , η}. We first prove that
sion (8) has to be proved. We have that s = Pµ
vi, q =Pν
Pν
j=1 βj = 1, Pη
k=1 ρk rk withPµ
j=1 βj wj , δ =Pη
F (q) vi + E(q) rk ∈ TS(vi),
i ∈ {1, . . . , µ}, q ∈ Q, k ∈ {1, . . . , η}.
(9)
4
i ∈ {1, . . . , µ}, j ∈ {1, . . . , ν}, k ∈ {1, . . . , η}.
To overcome the problem of the choice of τ , we intro-
duce Theorem 12 that provides a condition equivalent
to (10). The proof is analogous to that of (Blanchini,
1990, Th. 2.3). Let Ci be the convex cone defined by the
delimiting planes of S which contain vi (see (Panik, 1993,
Ch. 4)) :
Ci = {gT
h s ≤ ξh, ξh > 0, for every gT
h
and ξh : gT
h vi = ξh, vi ∈ vert(S)}.
Theorem 12 The set S is positively D-invariant for
system (6), if and only if, for all τ > 0, vi ∈ vert(S),
ωj ∈ vert(Q) and rk ∈ vert(D):
vi + τ (F (wj ) vi + E(wj) rk) ∈ Ci,
i ∈ {1, . . . , µ}, j ∈ {1, . . . , ν}, k ∈ {1, . . . , η}.
PSfrag replacements
If the plane description of S is available, the next
corollary, whose proof directly follows from that of
Theorem 10, holds.
Corollary 13 The set S is positively D-invariant for
system (6), if and only if, for every τ > 0 and every
vi ∈ vert(S), ωj ∈ vert(Q),
y0
yL
xL
θL
PL
yF
PF
xF
θF
x0
P0
Fig. 1. Leader-follower setup.
(In + τ F (wj )) vi ∈ C⋆
i , i ∈ {1, . . . , µ}, j ∈ {1, . . . , ν},
(11)
i is the cone obtained by shifting the planes of Ci
where C⋆
as follows :
i = {gT
C⋆
h s ≤ ξh − maxjk{τ gT
h E(wj ) rk}, ωj ∈ vert(Q),
rk ∈ vert(D), for every gT
h : gT
h vi = ξh}.
Remark 14 According to Theorem 9, conditions (11)
and (2) provide us with a set of inequalities in the un-
known K defining the polytope K of all the state feedback
matrices solving the DLCRP.
3 The visibility maintenance problem
3.1 Modeling
Let Σ0 ≡ {O0 ; x0, y0} be the fixed reference frame in R2,
and ΣF ≡ {OF ; xF, yF} and ΣL ≡ {OL ; xL, yL} the
reference frames attached to a follower robot F and a
leader robot L (see Fig. 1). The robots are supposed to
have single integrator dynamics,
p F
F = σF
F ,
θ F = ωF ,
p L
L = σL
L ,
θ L = ωL ,
(12)
L[1], σL
F[1], σF
L = (σL
F[2])T , σL
F = (xF, yF)T , p L
F = (σF
where p F
L = (xL, yL)T are the posi-
tions, σF
L[2])T the lin-
ear velocities and ωF, ωL the angular velocities of robots
F and L in the frames ΣF and ΣL, respectively. We are go-
ing to derive a dynamic model describing the relative dy-
namics of the robots F and L. Referring (12) to the frame
Σ0, we obtain (Siciliano et al., 2008), p0
F(θF) σF
F,
p0
L = R0
L(θL) σL
L, where
F = R0
R0
F(θF) ="cos θF − sin θF
sin θF cos θF #, R0
sin θL cos θL #.
L(θL) ="cos θL − sin θL
The position of robot L with respect to ΣF is then
given by
pF
L = RF
0 (θF)(p0
L − p0
F),
(13)
0 (θF) = (R0
where RF
respect to time, we get
F(θF))T . Differentiating (13) with
L = RF
p F
0 (θF)(p0
L−p0
F)+RF
L(θL) σL
L−R0
F(θF) σF
Since
RF
0 (θF) =h 0
we can rewrite (14) as
0 (θF)(cid:0)R0
−ωF 0 iRF
ωF
0 (θF),
F(cid:1).
(14)
pF
L =" 0 ωF
−ωF 0 # pF
L + RF
L(βF
L ) σL
L − σF
F,
(15)
where βF
, θL−θF. Collecting equation (15) and the rel-
L
ative angular dynamics of the robots together, we obtain
the following system
pF
L
βF
L
=
−I2
pF
L[2]
−pF
L[1]
0 0
−1
where pF
we will suppose that
L = (pF
L[1], pF
F
ωF #+
" σF
RF
L(βF
L )
0 0
0
0
1
L
ωL #,
" σL
(16)
L[2])T . For the sake of simplicity,
σF
F = (1 + vF , 0)T,
σL
L = (1 + vL , 0)T,
(17)
where vF(t) < 1, vL(t) < 1, for all t ≥ 0. F and L will
then behave in a way similar to Dubins vehicles since
5
After simple matrix manipulations in (18), we obtain
3.2 Problem statement
where
they can only move forward (however, differently from
the standard Dubins model, vF and vL are not necessarily
constant in our case). Substituting (17) in (16), we finally
come up with the following system
pF
L[1]
pF
L[2]
βF
L
=
cos βF
L − 1
sin βF
L
0
−1 pF
L[2]
0 −pF
L[1]
0
−1
vF
ωF
+
cos βF
L 0
sin βF
L 0
0
1
+
vL
ωL
,
L[1], pF
L[2], βF
(18)
with state vector s = (pF
L )T ∈ X ⊂ SE(2),
input vector u = (vF, ωF)T ∈ U ⊂ (−1, 1) × R and
disturbance vector δ = (vL, ωL)T ∈ D ⊂ (−1, 1) × R.
System (18) describes the relative dynamics of the
Dubins-like vehicles F and L in the configuration
space SE(2).
In the forthcoming analysis, we will suppose that robot
F is equipped with a sensor (e.g., a panoramic camera,
a laser range finder, etc.) with limited sensing range.
We will call visibility set of robot F any compact and
convex polyhedral set S ⊂ X containing the origin in
its interior. Please note that the visibility set generalizes
the notion of sensor footprint since it encodes both the
position and angle information.
Robot L moves along a given trajectory and robot F
aims at keeping L always inside its visibility set S, while
respecting the control bounds. By referring to system
(18), we can formalize this problem as follows:
Problem 15 (Visibility maintenance problem: VMP)
Let S be the visibility set of robot F and let s(0) ∈ S. Find
a control u(t) such that for all δ(t) ∈ D, the conditions
s(t) ∈ S and u(t) ∈ U are fulfilled for t > 0.
In the following, we will refer to Problem 15 as to the
VMP with candidate positively D-invariant set S, con-
trol set U and disturbance set D.
3.3 Solution method
Next, we will transcribe system (18) into the linear para-
metric form (6): in this way, the VMP simply reduces
to the DLCRP (Problem 8) introduced in Sect. 2 and
suitable solvability conditions can be derived by means
of (11) and (2).
0 −d − ∆pF
∆ pF
L[1]
pF
L[2]
βF
L
=
0
−1
+
A(q) =
L[1]
pF
L[2]
βF
L
pF
L[2]
L
L
L
0
βF
vF
L −1
ωF
−1
∆pF
L[1]
0 0
0 0
0 0 cos βF
βF
sin βF
+
, B(q) =
E(q) =
1 + q5
q6
0
0
1
0
,
which can be written in the form (6) with
q2
0 0
0 0 1 + q1
0 0
0
q4
−1
0 −d − q3
0
−1
cos βF
L 0
sin βF
L 0
0
1
vL
ωL
,
(19)
,
(20)
(21)
q1 =
sin βF
L
βF
L
− 1, q2 =
q4 = pF
L[2],
q5 = cos βF
, q3 = ∆pF
cosβF
L − 1
βF
L
L − 1, q6 = sin βF
L .
L[1],
We made the following change of variables in (19),
(pF
L[1], pF
L[2], βF
L )T → (∆pF
L[1], pF
L[2], βF
L )T ,
L[1] = pF
where ∆pF
L[1] − d and d is a strictly positive con-
stant. Two main reasons motivated this transformation:
first of all, if robot F is able to keep L always inside a vi-
sibility set displaced of d with respect to its center (with
1
2 ks1 − s2k2), then this automati-
d > max s1,s2 ∈ vert(S)
cally guarantees robots' collision avoidance. Second, this
choice simplifies the study of the VMP with chains of
robots (see Sect. 4.3).
Notice that A0, B0 and E0 in (20) (recall the notation
used in Sect. 2) correspond to the constant matrices ob-
tained by linearizing system (18) around the equilibrium
seq = (d, 0, 0)T , ueq = (0, 0)T , δeq = (0, 0)T .
For the sake of simplicity, we will henceforth make the
following assumption:
Assumption 16 Suppose that (see Fig. 2)
U =(cid:8)(vF, ωF)T : vF ∈ [−VF, VF], ωF ∈ [−ΩF, ΩF](cid:9),
D =(cid:8)(vL, ωL)T : vL ∈ [−VL, VL], ωL ∈ [−ΩL, ΩL](cid:9),
(22)
6
2
a
xL
S
L
PSfrag replacements
yL
2
a
2
b
d
F
Fig. 2. The visibility set S in (23) and the pose of the robots
L and F for (∆pF
L )T = (0, 0, 0)T , d > a.
L[2], βF
L[1], pF
S =(cid:8)(∆pF
L )T : ∆pF
L[2], βF
L[1], pF
pF
L[2] ∈ [−a, a], βF
L[1] ∈ [−a, a],
(23)
where VF < 1, VL < 1, ΩF, ΩL, a, b, are strictly positive
constants.
L ∈ [−b, b](cid:9),
The definition (22) of control and disturbance sets is mo-
tivated by the presence of saturation bounds on the driv-
ing motors of physical robots. The candidate invariant
set S is defined as in (23) because it is computationally
simple to handle (this will allow us to provide concise
solvability conditions for the VMP in Theorem 18), and
because its horizontal section represents a reasonable
good inner approximation of a disk sensor footprint e.g.,
due to an omnidirectional camera or a 360◦ laser scanner
(the problem of precisely quantify the non-conservatism
introduced by this approximation goes beyond the scope
of this paper, and it is left as a subject of future research).
Finally, it is worth emphasizing that is not unusual in
the multi-agent systems literature to encounter rectan-
gular footprints, that are typically used, for example, to
model "push-broom" or line-scanner sensors (see, e.g.,
(Finke et al., 2005)).
We now complete our transcription of system (18) into
the linear parametric form (6), by defining an appropri-
ate polyhedral set for the model parameter uncertainty:
Q =(cid:8)(q1, . . . , q6)T : q1 ∈(cid:2) sin b
b − 1, 0(cid:3),
(cid:3) , q3 ∈ [−a, a] , q4 ∈ [−a, a] ,
b
b
, 1−cos b
q2 ∈(cid:2) cos b−1
q5 ∈(cid:2)cos b − 1, 0(cid:3), q6 ∈ [− sin b, sin b](cid:9).
It is easy to show that, if (∆pF
definition (21) immediately implies that q ∈ Q.
L[2], βF
L[1], pF
(24)
L )T ∈ S, then
Remark 17 In the previous passages, the nonlinear
system (18) has been absorbed into a linear (controlled)
differential inclusion (see (Blanchini and Miani, 2008,
Sect. 2.1.2)). This is an approximate transformation:
however, no matter how the input u is chosen, we
7
have that any trajectory of the original system (18) is
also a trajectory of the corresponding linear uncertain
system (the opposite is clearly not true in general).
As a consequence, if we are able to determine the qualita-
tive behavior of the absorbing system, we can determine
(in a conservative way) the behavior of the original sys-
tem. Some tools are available in the robust control liter-
ature to quantify this conservativeness, such as, e.g., the
recent nonlinear extensions of the gap and Vinnicombe's
ν-gap metrics (see (Bian and French, 2005; James et
al., 2005) and the references therein). However, in the
interest of brevity, we will not perform such an analysis
in this paper.
We are now ready to state the main result of this section.
Theorem 18 (Solvability of the VMP) For
the
robots F and L, consider the VMP with candidate posi-
tive D-invariant set S, control set U and disturbance set
D satisfying Assumption 16 with d > a, 0 < b ≤ π/2.
This VMP has a solution if the following conditions are
satisfied
VF ≥ VL(cid:18)1 +
a sin b
d − a(cid:19) + 1 − cos b +
a b
d − a
,
(25)
ΩL ≤
(1 − VL) sin b
d + a
,
VL sin b + b
d − a
≤ ΩF .
(26)
The state feedback matrix has the form
K =
0
k11 0
0 k22 k23
,
(27)
where k11, k22 and k23 belong to the polytope K ⊂ R3
defined by (30)-(31), (see the proof below).
Proof: Let us apply Corollary 13 to system (19). By se-
lecting τ = 1 in (11), we obtain
1 − k11 + q4k21 −k12 + q4k22
q2 − k13 + q4k23
−(d + q3)k21
1 − (d + q3)k22 1 + q1 − (d + q3)k23
−k21
−k22
1 − k23
vi ∈ C ⋆
i .
(28)
Condition (28) must be evaluated only on the vertices
v1 = (a, a, b)T , v2 = (a, a, −b)T , v3 = (a, −a, b)T , v4 =
(a, −a, −b)T since the set (23), whose plane representa-
tion is
1/a
−1/a
0
0
0
1/a
0 −1/a
0
0
0
0
0
0
0
1/b
0 −1/b
s ≤
1
1
1
1
1
1
,
s = (∆pF
L[2], βF
the origin. The cones C⋆
L[1], pF
L )T , is symmetric with respect to
1 , . . . , C⋆
4 are given by
C⋆
1 = {gT
1 s ≤ 1 − VL
a , gT
3 s ≤ 1 − VL sin b
a
C⋆
2 = {gT
1 s ≤ 1 − VL
a , gT
3 s ≤ 1 − VL sin b
a
C⋆
3 = {gT
1 s ≤ 1 − VL
a , gT
4 s ≤ 1 − VL sin b
a
C⋆
4 = {gT
1 s ≤ 1 − VL
a , gT
4 s ≤ 1 − VL sin b
a
, gT
5 s ≤ 1 − ΩL
b },
, gT
6 s ≤ 1 − ΩL
b },
, gT
5 s ≤ 1 − ΩL
b },
, gT
6 s ≤ 1 − ΩL
b }.
Condition (28) can then be rewritten as:
a(1 − k11 + q4k21 − k12 + q4k22) + b(q2 − k13 + q4k23)
a(1 − (k21 + k22)(d + q3)) + b(1 + q1 − k23(d + q3))
−a(k21 + k22) + b(1 − k23)
a(1 − k11 + q4k21 − k12 + q4k22) − b(q2 − k13 + q4k23)
a(1 − (k21 + k22)(d + q3)) − b(1 + q1 − k23(d + q3))
−a(k21 + k22) − b(1 − k23)
a(1 − k11 + q4k21 + k12 − q4k22) + b(q2 − k13 + q4k23)
a(−1 + (−k21 + k22)(d + q3)) + b(1 + q1 − k23(d + q3))
−a(k21 − k22) + b(1 − k23)
a(1 − k11 + q4k21 + k12 − q4k22) − b(q2 − k13 + q4k23)
a(−1 + (−k21 + k22)(d + q3)) − b(1 + q1 − k23(d + q3))
−a(k21 − k22) − b(1 − k23)
∈ C ⋆
1 ,
∈ C ⋆
2 ,
∈ C ⋆
3 ,
∈ C ⋆
4 .
(29)
Because of the special structure of B(q) in (20), we
can select a simplified state feedback matrix K of the
form (27): this allows to the decouple the control inputs
vF and ωF and visualize the polytope K ⊂ R3 of all the
feasible gain matrices. We can then rewrite (29) in the
following simplified form:
b
a
q4 k23 ≤ −
(d + q3) k23 ≤ −
q2 −
VL
a
,
(1 + q1) −
b
a
b
a
b
a
q4 k23 ≤
q2 −
VL
a
,
− k11 + q4k22 +
−(d + q3) k22 −
b
a
−k11 + q4 k22 −
−(d + q3) k22 +
b
a
b
a
b
a
(d + q3) k23 ≤
(1 + q1) −
−
a
b
a
b
k22 − k23 ≤ −
k22 − k23 ≤ −
b
a
b
a
q4 k23 ≤ −
q4 k23 ≤
b
a
,
,
ΩL
b
ΩL
b
b
a
q2 −
q2 −
VL
a
VL
a
.
−k11 − q4 k22 +
−k11 − q4 k22 −
VL sin b
a
,
VL sin b
a
,
,
(30)
8
The admissibility condition (2) leads to the additional
constraints
k11 ≤
VF
a
,
k11 ≥ −
VF
a
,
k22 +
k22 −
b
a
b
a
k23 ≤
ΩF
a
,
k23 ≥ −
ΩF
a
,
k22 −
k22 +
b
a
b
a
k23 ≤
ΩF
a
,
k23 ≥ −
ΩF
a
.
(31)
The Fourier-Motzkin elimination is a mathematical al-
gorithm for eliminating variables from a system of linear
inequalities. Elimination of unknown kij from the sys-
tem of inequalities, consists in creating another system
of the same kind but without kij , such that both systems
have the same solutions over the remaining variables.
If one removes all variables from a system of inequalities
with numerical coefficients, then one obtains a system
of constant inequalities, which can be trivially decided
to be true or false. This procedure can then be used to
easily check whether a given system admits solutions
or not (see Appendix A for more details).
Applying the Fourier-Motzkin elimination to the in-
equalities (30)-(31) with the assumption that d > a
(in order to fix the sign of the coefficients of k22 and
k23 in the second and fourth inequality of (30)) we end
up with the following conditions on the variables a, b,
d, VF, VL, ΩF, ΩL and uncertain parameters q1, . . . , q4
(the detailed passages are reported in Appendix B),
ΩL ≤ b(1+q1)−VL sin b
d+q3
,
b(1+q1)+VL sin b
d+q3
≤ ΩF ,
VF ≥ VL + b q2,
VF ≥ VL(cid:0)1 + q4 sin b
VF ≥ VL(cid:0)1 + q4 sin b
VF ≥ VL(cid:0)1 − q4 sin b
VF ≥ VL(cid:0)1 − q4 sin b
d+q3 (cid:1)+ b(cid:0)q2 + q4(1+q1)
d+q3 (cid:1),
d+q3 (cid:1)− b(cid:0)q2 + q4(1+q1)
d+q3 (cid:1),
d+q3 (cid:1)+ b(cid:0)q2 + q4(1+q1)
d+q3 (cid:1),
d+q3 (cid:1)− b(cid:0)q2 + q4(1+q1)
d+q3 (cid:1),
for q4 > 0 ,
for q4 > 0 ,
for q4 = 0 ,
for q4 < 0 ,
for q4 < 0.
An appropriate selection of the parameters q1, . . . , q4 on
the extremes of the intervals (24), (i.e., q1 = sin b
b − 1,
q3 = a, for the first inequality, q1 = 0, q3 = −a, for the
second and q1 = 0, q2 = 1−cos b
, q3 = −a, q4 = a, for
(cid:4)
the third), leads to (25) and (26).
b
Some remarks are in order at this point.
The inequalities (25) and (26) (which are linear in VF,
VL, ΩF, ΩL and nonlinear in a, b, d), specify the role
played by each of the parameters introduced in Assump-
tion 16, in the solvability of the VMP. In particular, they
show how the bounds on the forward and angular veloc-
ity of the follower robot are affected by the size of the
visibility set S and the velocity of the leader.
Note that conditions (25) and (26) are necessary and
sufficient for the linear uncertain system (19). Note also
that owing to (26), we have ΩF ≥ ΩL. Once fixed the
variables a, b, d, VF, VL, ΩF, ΩL according to (25) and
(26), the polytope of all the feasible state feedback matri-
ces is given by (30)-(31): by evaluating (30)-(31) on the
vertices of the polyhedron Q, we can see that K is defined
by a set of 392 inequalities, most of whom are redundant
(see, e.g., the external (green) polytope in Fig. 3, below).
3
2
k
Remark 19 (Selection of the gain matrix K)
Since the polytope K contains infinite gain matrices, one
needs an optimal criterion to select K, such as, e.g.,
minimizing any matrix norm. In the simulation experi-
ments reported in Sect. 5, we have chosen the matrix
PSfrag replacements
1
0. 9
0. 8
0. 7
0. 6
0. 5
0. 4
4
3
k22
2
1
5
5. 5
6. 5
6
k11
K =
with minimum 2-norm.
0
k11 0
0 k22 k23
,
Fig. 3. Polytopes K for a set of given parameters:
(blue, internal) with disturbances; (green, external) without
disturbances.
we can rewrite (32) as
4 Extensions and applications
In this section, we propose various extensions of the basic
setup considered in Theorem 18, and discuss a few appli-
cations. We study the VMP in the presence of unknown
but bounded disturbances, and consider the case of the
leader moving along a circular path around a stationary
target. We also extend our results to chains of robots.
4.1 Rejection of unknown but bounded disturbances
Let us consider the following system
0 0 cos βF
βF
L
L −1
0 0
sin βF
L
βF
L
0 0
0
∆ pF
L[1]
pF
L[2]
βF
L
=
−1 0
pF
L[2]
0 −1−∆pF
L[1] − d
0 0
−1
vF
hF
ωF
∆pF
L[1]
pF
L[2]
βF
L
+
+
cos βF
L − sin βF
L 0
sin βF
L cos βF
L 0
0
0
1
ωL
vL
hL
.
(32)
With respect to (19), two new components, hF and hL,
are present in the input and disturbance vectors u and δ.
They are unknown but bounded disturbances acting on
the robots F and L (e.g., lateral wind in a real setup).
Our purpose here is to solve the VMP in the presence
of the disturbances hF, hL.
Collecting together all
the perturbations acting
on the nominal system (i.e., vL, ωL, hF and hL),
∆ pF
L[1]
pF
L[2]
βF
L
−1
=
pF
L[2]
0 0 cos βF
βF
L
L −1
0 0
0 0
sin βF
L
βF
L
0
∆pF
L[1]
pF
L[2]
βF
L
+
0 −∆pF
L[1] − d
0
−1
vF
ωF
cos βF
sin βF
0 − sin βF
L 0
L
L 0 −1 cos βF
L
0
1
0
0
+
.
vL
ωL
hF
hL
(33)
Let U be given in (22), and define
D = {(vL, ωL, hF, hL)T : vL ∈ [−VL, VL],
ωL ∈ [−ΩL, ΩL], hF ∈ [−HF, HF], hL ∈ [−HL, HL]},
(34)
where HF, HL are strictly positive constants. Using the
same arguments as those in Theorem 18, we can prove
the following corollary (note that the feedback matrix
K is again of the form (27)).
Corollary 20 (VMP with disturbances) Choose U
and S as in Assumption 16, D as in (34), and let d > a,
0 < b ≤ π/2. The VMP for the robots F and L in the
presence of the unknown but bounded disturbances hF,
hL, has a solution if the following conditions are satisfied
a sin b
d − a(cid:19) + 1 − cos b
a(HF + HL + b)
+
d − a
(35)
+ HL sin b,
VF ≥ VL(cid:18)1 +
9
PSfrag replacements
xL
L
S
(cid:16) sin γ
ρ ,
1−cos γ
ρ
(cid:17)
*
yL
γ
ρ
F
Fig. 4. VMP on a circle: pose of the robots L and F
L )T = (0, 0, 0)T , vL = vF and
for (∆pF
ωL = ωF = ρ.
L[2], ∆βF
L[1], ∆pF
ΩL ≤
ΩF ≥
(1 − VL) sin b − (HF + HL)
d + a
,
VL sin b + b + (HF + HL)
d − a
.
(36)
Remark 21 Note that because of the additional terms
HF and HL, conditions (35)-(36) are stricter than (25)-
(26) and then the polytope K is smaller in this case. This
is evident in Fig. 3, where the polytope K ( blue, internal)
obtained for a = 0.15 m, b = π/3 rad, d = 1.6 m, VF =
0.95 m/s, VL = 0.1 m/s, ΩF = π/2 rad/s, ΩL = π/20
rad/s and HF = 0.2 m/s, HL = 0.1 m/s is compared with
the polytope ( green, external) corresponding to HL =
HF = 0 m/s.
4.2 VMP on a circle
In this section, we will suppose that the leader robot
moves along a circular path, around a static target. This
scenario could be of interest in several real-world ap-
plications, such as, e.g., for environmental surveillance,
patrolling or terrain and utilities inspection (Casbeer et
al., 2006; Susca et al., 2008). Differently from Sect. 3.3,
we will assume that the pose of robot L with respect to
the frame of F is defined through the angle 0 < γ < π/2
and the angular velocity ρ > 0, instead of the distance
parameter d (see Fig. 4). Let us consider the following
change of variables in system (18):
(pF
L[1], pF
L[2], βF
L )T → (∆pF
L[1], ∆pF
L[2], ∆βF
L )T ,
(vL, ωL)T → (vL, ωL)T ,
(vF, ωF)T → (vF, ∆ωF)T ,
L[1] − sin γ
L[2] − 1−cos γ
L[1] = pF
where ∆pF
,
∆βF
L = βF
L − γ and ∆ωL = ωL − ρ, ∆ωF = ωF − ρ.
Following the same procedure detailed in the previous
sections, we obtain the system
L[2] = pF
ρ , ∆pF
ρ
∆ pF
L[1]
∆ pF
L[2]
∆ βF
L
=
0
ρ q2 − sin γ
−ρ 0 q1 + cos γ
0
0
0
∆pF
L[1]
∆pF
L[2]
∆βF
L
+
+
−1 q4 + 1−cos γ
0 −q3 − sin γ
ρ
ρ
0
−1
vF
∆ωF
q5 + cos γ 0
q6 + sin γ 0
0
1
vL
∆ωL
,
(37)
where
q1 = sin(∆βF
L +γ)−sin γ
∆βF
L
q2 = cos(∆βF
L +γ)−cos γ
∆βF
L
− cos γ,
+ sin γ,
q3 = ∆pF
L[1], q5 = cos(∆βF
L + γ) − cos γ,
q4 = ∆pF
L[2], q6 = sin(∆βF
L + γ) − sin γ.
Assumption 22 Let us suppose that
U =(cid:8)(vF, ∆ωF)T : vF ∈ [−VF, VF], ∆ωF ∈ [−ΩF, ΩF](cid:9),
D =(cid:8)(vL, ∆ωL)T : vL ∈ [−VL, VL], ∆ωL ∈ [−ΩL, ΩL](cid:9),
where 0 < VF < 1, 0 < VL < 1 and 0 < ΩF, 0 < ΩL < ρ.
Let us also consider the following visibility set (see Fig. 4)
S = {(∆pF
L[1], ∆pF
L[2], ∆βF
L )T : ∆pF
L[1] ∈ [−a, a],
∆pF
L[2] ∈ [−a, a], ∆βF
L ∈ [−b, b]},
(38)
where a > 0 and b > 0.
Since the state of system (37) is constrained in (38), the
polyhedron Q ⊂ R6 is defined as follows:
q1 ∈(cid:2) sin(b+γ)−sin γ
q2 ∈ [ cos(b+γ)−cos γ
b
b
− cos γ, sin(b−γ)+sin γ
b
+ sin γ, − cos(b−γ)+cos γ
b
− cos γ(cid:3),
+ sin γ],
q3 ∈ [−a, a], q5 ∈ [cos(b + γ) − cos γ, cos(b − γ) − cos γ],
q4 ∈ [−a, a], q6 ∈ [−sin(b − γ) − sin γ, sin(b + γ) − sin γ].
The proof of the next theorem is analogous to that of
Theorem 18 and it is omitted. The feedback matrix K
has also in this case the form (27).
Theorem 23 (Solvability of the VMP on a circle)
Choose U, D and S as in Assumption 22 and let
1 − cos γ > ρa, 0 ≤ b ± γ ≤ π/2. The VMP on a circle
10
PSfrag replacements
1
2
n − 1
n
S2
S3
Sn
Fig. 5. Chain of n robots.
has a solution if the following conditions are satisfied
VF ≥ VL(cid:0)cos(b − γ) + sin(b+γ)(1−cos γ−ρ a)
sin γ+ρa
(cid:1)+ cos γ + ρa
(sin(b + γ) − sin γ + ρa) − cos(b + γ),
− 1−cos γ−ρ a
sin γ+ρa
sin γ + ρa
ΩL ≤ ρ(cid:18) (1 − VL) sin(b + γ)
ΩF ≥ ρ(cid:18) VL sin(b + γ) + sin(b − γ) + sin γ + ρa
− 1(cid:19),
sin γ − ρa
Ω1 ≤
(1 − V1) sin b2
d2 + a2
, Ωn ≥
Vn−1 sin bn + bn
dn − an
,
(41)
and for all k ∈ {2, . . . , n − 1},
Vk−1 sin bk + bk
dk − ak
≤ Ωk ≤
(1 − Vk) sin bk+1
dk+1 + ak+1
.
(42)
It is an easy task to verify that if ak = a, bk = b, dk =
d, k ∈ {2, . . . , n} and Vk = V , k ∈ {1, . . . , n}, (a >
0, d > a, 0 < b ≤ π/2, 0 < V < 1), condition (42)
(and condition (40) as well) is not satisfied. Nevertheless,
it can be proved that if Vk and at least one of the three
parameters defining the visibility sets are left free to
vary from robot to robot, then (40)-(42) can be always
fulfilled. Fig. 6 shows the progression of a feasible set
of parameters Vk, Ωk, bk for a chain of n = 15 robots,
when ak = a = 0.1 and dk = d = 7, k ∈ {2, . . . , 15}.
We conclude this section with the following corollary
and a few remarks:
(cid:19).
(39)
Corollary 24 The chain of robots can hold only a finite
number of agents and cannot be closed, i.e., robot 1 can-
not cyclically pursue robot n.
Note that differently from Theorem 18, in this case,
owing to (39) is not always true (i.e., for all values of the
parameters) that ΩF ≥ ΩL.
Proof: Let us study the connection existing between pa-
rameters V1 and Vn. After simple algebraic manipula-
tions on (40), we obtain the following inequality:
4.3 Chain of robots
Next, we consider more complex robotic networks
built by concatenating multiple leader-follower units
(see Fig. 5). When equipped with wireless sensors,
these arrays of robots could be used, for example, to
efficiently monitor the temperature and/or salinity of
the ocean or measure the average concentration of air
pollutants (Lynch et al., 2008). In what follows, the
feasibility conditions of Theorem 18 will be extended
to such robot chains in order to maintain network-wide
visibility between the agents.
Because of the leader-follower hierarchy within the
chain, vehicle k + 1 (the "follower") will aim at keeping
the vehicle ahead (robot k, the "leader"), in its visibility
set. Let ak, bk, dk, dk > ak, bk ≤ π/2, k ∈ {2, . . . , n},
n > 2, be the strictly positive parameters defining the
visibility set Sk of robot k-th and let 0 < Vk < 1, Ωk > 0,
k ∈ {1, . . . , n} be the bounds on k-th robot's linear
and angular velocities (recall Assumption 16). By prop-
agating conditions (25)-(26) of Theorem 18 starting
from robot 1 (that guides the formation), we obtain the
following set of inequalities (linear in Vk and Ωk)
PSfrag replacements
0.5
V1 ≤
n
Yi=1(cid:16)1 +
Vn
1
ai sin bi
di − ai (cid:17)
Xi=2
−
n
1 − cos bi +
i
Yk=2(cid:16)1 +
ai sin bi
di − ai
ak sin bk
dk − ak (cid:17)
(43)
.
Vk
Ωk
bk
2.5
2
1.5
1
Vk+1 ≥ Vk(cid:18)1 +
ak+1 sin bk+1
dk+1 − ak+1(cid:19) + 1 − cos bk+1
+
ak+1 bk+1
dk+1 − ak+1
, k ∈ {1, . . . , n − 1},
0
2
4
6
10
12
14
8
k
(40)
Fig. 6. Progression of a feasible set of parameters Vk, Ωk, bk
for a chain of n = 15 robots (ak = 0.1, dk = 7, for all k).
Angles are in radians.
11
]
m
[
0
y
9
8
7
6
5
4
3
2
1
0
Leader
Follower
]
m
[
]
1
[
F L
p
∆
]
m
[
]
2
[
F L
p
]
d
a
r
[
PSfrag replacements
F L
β
0
2
4
6
x
0 [m]
(a)
0.5
0
_
0.5
0.5
0
_
0.5
1
0
_
1
0
1
2
3
4
5
6
7
8
9
10
0
1
2
3
4
5
6
7
8
9
10
]
s
/
m
[
F
v
]
s
/
d
a
r
[
F
ω
PSfrag replacements
5
10
6
7
8
9
0
1
2
3
4
time [s]
(b)
1
0.5
0
_
0.5
_
1
1
0.5
0
_
0. 5
_
1
0
1
2
3
4
5
6
7
8
9
10
0
1
2
3
4
5
6
7
8
9
10
time [s]
(c)
Fig. 7. Basic scenario: (a) Trajectory of the leader and follower, and visibility set S; (b) ∆pF
±a, ±a, ±b (dash); (c) vF, ωF (solid) and bounds ±VF, ±ΩF (dash).
L[1], pF
L[2], βF
L (solid) and bounds
When the number of robots n tends towards infinity, the
first term on the right-hand side of (43) converges to
zero (for any choice of the parameters ai, bi, di satisfy-
ing the previous assumptions) asymptotically leading to
the inequality V1 ≤ 0, which contradicts the initial hy-
pothesis of a strictly positive V1. On the other hand, if
the chain of robots were closed, the following additional
inequality should be satisfied
V1 ≥ Vn(cid:18)1 +
a1 sin b1
d1 − a1(cid:19) + 1 − cos b1 +
a1 b1
d1 − a1
,
(cid:4)
but it is incompatible with condition (43).
It is easy to prove from condition (43), that once fixed
a rule for the evolution of the parameters defining the
visibility sets (i.e., ak = fa(k), bk = fb(k), dk = fd(k),
with fa, fb, fd : Z>1 → R given discrete maps), an up-
per bound on the maximum number of robots the chain
can hold is given by the maximum positive integer N
that fulfills the following inequality:
N
Xi=2(cid:16)1 − cos bi +
ai bi
di − ai(cid:17)
N
Yk = i+1(cid:16)1 +
ak sin bk
dk − ak (cid:17) < 1.
Note that once the state feedback matrices Ki, i ∈
{2, . . . , n} of the robots have been established, the im-
plementation of the control laws is totally distributed :
in fact, each agent only needs to know the relative posi-
tion and orientation of the preceding vehicle in the chain,
to execute its control action.
5 Simulation results
5.1 Basic scenario
The simulation results reported in Fig. 7 refer to the
basic scenario studied at the beginning of Sect. 3.
The leader robot moves with velocities vL(t) =
0.05 sin(t), ωL(t) = π
20 cos(0.1t). We set VL = 0.1 m/s,
ΩL = π/15 rad/s, VF = 0.9 m/s, ΩF = π/3 rad/s,
a = 0.4 m, b = π/4 rad and d = 2 m, according to the
conditions of Theorem 18 and we chose the gain matrix
in K with minimum 2-norm:
K =" 1.5173
0
0
0.3707 0.4925#.
0
Note that since K is in the interior of K, the asymptotic
stability is assured (cf. Theorem 5.2 in (Blanchini, 1991)).
System (19) has been initialized with
(∆pF
L[1](0), pF
L[2](0), βF
L (0))T
(44)
= (0.3285, −0.1626, 0.1071)T.
Fig. 7(a) reports the trajectory of robot L and F and the
visibility set S, (in order to have a temporal reference
the robots are drawn every two seconds). Fig. 7(b) shows
that ∆pF
L (solid), keep inside the respective
bounds ±a, ±a, ±b (dash), as expected. Finally, Fig. 7(c)
exposes that the control inputs vF, ωF (solid), respect
the corresponding bounds ±VF, ±ΩF (dash).
L[2], βF
L[1], pF
5.2 Rejection of unknown but bounded disturbances
Extensive simulation experiments have been performed
to illustrate the theory and assess the soundness of
the proposed approach.
In the simulation results reported in Fig. 8, the
leader robot moves with velocities vL(t) = 0.01,
12
2
0
]
m
[
0
y
−2
−4
−6
0
5
x
(a)
]
m
[
]
1
[
F L
p
∆
]
m
[
]
2
[
F L
p
Leader
Follower
10
PSfrag replacements
F L
β
15
0 [m]
]
d
a
r
[
0.5
0
−0.5
0
0.5
0
−0.5
0
1
0
−1
0
5
5
5
20
20
]
s
/
m
[
F
v
]
s
/
d
a
r
[
10
15
10
15
F
ω
PSfrag replacements
10
15
20
time [s]
(b)
1
0.5
0
−0.5
−1
0
1
0.5
0
−0.5
−1
0
20
5
10
15
5
10
15
20
time [s]
(c)
Fig. 8. Rejection of unknown but bounded disturbances: (a) Trajectory of the leader and follower, and visibility set S; (b) ∆pF
pF
L[2], βF
L (solid) and bounds ±a, ±a, ±b (dash); (c) vF, ωF (solid) and bounds ±VF, ±ΩF (dash).
L[1],
]
m
[
0
y
4
3
2
1
0
−1
Leader
Follower
]
m
[
]
1
[
F L
p
∆
]
m
[
]
2
[
F L
p
∆
]
d
a
r
[
PSfrag replacements
F L
β
∆
−3
−2
−1
0
x 0 [m]
1
2
(a)
0.5
0
−0.5
0
0.5
0
−0.5
0
1
0
−1
0
2
4
6
8
10
2
4
6
8
10
12
12
F
ω
∆
PSfrag replacements
2
4
6
8
10
12
time [s]
(b)
1
0.5
0
−0.5
]
s
/
m
[
F
v
]
s
/
d
a
r
[
−1
0
1
0.5
0
−0.5
−1
0
2
4
6
8
10
12
2
4
6
8
10
12
time [s]
(c)
Fig. 9. VMP on a circle: (a) Trajectory of the leader and follower, and visibility set S; (b) ∆pF
bounds ±a, ±a, ±b (dash); (c) vF, ∆ωF (solid) and bounds ±VF, ±ΩF (dash).
L[1], ∆pF
L[2], ∆βF
L (solid) and
ωL(t) = − π
20 sin(0.08t). Unknown but bounded dis-
turbances hL(t), hF(t) (uniform random noises in the
interval (−0.1, 0.1)) act on the leader and the follower
(recall Sect. 4.1). Owing to the conditions of Corol-
lary 20, we chose VL = 0.03 m/s, ΩL = π/18 rad/s,
VF = 0.95 m/s, ΩF = π/4 rad/s, HL = HF = 0.12 m/s,
a = 0.4 m, b = π/4 rad and d = 2 m. The initial condi-
tion of system (33) is (44) and, again, we selected the
gain matrix in K with minimum 2-norm:
5.3 VMP on a circle
In Fig. 9, the leader robot moves with velocities
vL(t) = 0.05 sin(t), ωL(t) = π/30, and the parame-
ters VL = 0.06 m/s, ΩL = π/25 rad/s, VF = 0.8 m/s,
ΩF = π/3 rad/s, a = 0.4 m, b = π/4 rad, ρ = 0.3 rad/s
and γ = π/6 rad, have been chosen according to
the conditions of Theorem 23 (recall Sect. 4.2).
The minimum 2-norm gain matrix in K is, in this case
1.6735
0
K =
0
0
0.5896 0.5326
.
Fig. 8(a) reports the trajectory of L and F and the visibi-
lity set S. From Figs. 8(b) and 8(c), we note that despite
the presence of the unknown but bounded disturbances,
∆pF
L and vF, ωF respect the relative state
and control bounds.
L[2], βF
L[1], pF
K =" 1.3812
0
0
0.6051 0.5508#,
0
L[2](0), ∆βF
and the initial condition of system (37) is (∆pF
L[1](0),
∆pF
L (0))T = (0, 0, 0.5597)T . Fig. 9(a) re-
ports the trajectory of L and F and the visibility
set S. Figs. 9(b) and 9(c) show the time history of
∆pF
L[1], ∆pF
L and vF, ∆ωF, and the correspond-
ing bounds.
L[2], ∆βF
13
14
12
10
8
6
4
2
0
]
m
[
0
y
Robot 1
Robot 2
Robot 3
Robot 4
]
m
[
0
y
S2
Robot 1
Robot 2
Robot 3
Robot 4
14
12
10
8
6
4
2
0
]
m
[
0
y
S3
PSfrag replacements
0
2
4
8
PSfrag replacements
6
x 0 [m]
10
12
0
2
4
PSfrag replacements
10
12
8
6
x 0 [m]
14
12
10
8
6
4
2
0
Robot 1
Robot 2
Robot 3
Robot 4
S4
0
2
4
8
10
12
6
x 0 [m]
(a)
(b)
(c)
]
m
[
]
1
[
2 1
p
∆
]
m
[
]
2
[
2 1
p
PSfrag replacements
]
d
a
r
[
2 1
β
0.5
0
−0.5
0
0.5
0
−0.5
0
0.2
0
−0.2
0
2
4
6
8
2
4
6
8
10
10
]
m
[
]
1
[
3 2
p
∆
]
m
[
]
2
[
3 2
p
]
d
a
r
[
3 2
β
0.5
0
−0.5
0
0.5
0
−0.5
0
0.4
0
−0.4
0
PSfrag replacements
PSfrag replacements
2
4
6
8
10
time [s]
(d)
2
4
6
8
10
time [s]
(e)
2
4
6
8
2
4
6
8
10
10
]
m
[
]
1
[
4 3
p
∆
]
m
[
]
2
[
4 3
p
]
d
a
r
[
4 3
β
0.5
0
−0.5
0
0.5
0
−0.5
0
1
0
−1
0
2
4
6
8
2
4
6
8
10
10
2
4
6
8
10
time [s]
(f)
]
s
/
m
[
2
v
0.1
0.05
0
−0.05
−0.1
0
]
s
/
d
a
r
[
0.1
0.05
0
2
4
6
8
10
]
s
/
m
[
3
v
0.2
0
−0.2
0
]
s
/
d
a
r
[
0.2
0.1
0
2
4
6
8
10
2
4
6
8
10
]
s
/
m
[
4
v
1
0.5
0
−0.5
−1
0
]
s
/
d
a
r
[
0.5
0
2
ω
PSfrag replacements
−0.05
−0.1
0
2
4
3
ω
PSfrag replacements
6
8
10
−0.1
−0.2
0
2
4
4
ω
PSfrag replacements
6
8
10
−0.5
0
2
4
6
8
10
time [s]
(g)
time [s]
(h)
time [s]
(i)
1[1], p2
Fig. 10. Chain of robots: Trajectory of the 4 robots and (a) visibility set S2, (b) visibility set S3, (c) visibility set S4; (d)
∆p2
3[2],
β 4
3 and bounds ±a4, ±a4, ±b4; (g) v2, ω2 (solid) and bounds ±V2, ±Ω2 (dash); (h) v3, ω3 and bounds ±V3, ±Ω3; (i) v4, ω4
and bounds ±V4, ±Ω4.
1 (solid) and bounds ±a2, ±a2, ±b2 (dash); (e) ∆p3
2 and bounds ±a3, ±a3, ±b3; (f) ∆p4
1[2], β 2
2[2], β 3
2[1], p3
3[1], p4
5.4 Chain of robots
In Fig. 10 a chain of 4 robots is considered (recall
Sect. 4.3). Robot 1 guides the formation with veloci-
ties, v1(t) = 0.01, ω1(t) = π/52. The following set of
parameters, satisfying conditions (40)-(42), has been
used in the simulation: V1 = 0.02 m/s, V2 = 0.085 m/s,
V3 = 0.25 m/s, V4 = 0.8 m/s, Ω1 = π/50 rad/s,
Ω2 = π/35 rad/s, Ω3 = π/21 rad/s, Ω4 = π/6 rad/s,
and a2 = a3 = a4 = 0.4 m, b2 = π/14 rad, b3 = π/9 rad,
14
b4 = π/4 rad and d2 = d3 = d4 = 3 m. The minimum
2-norm gain matrices in the polytopes K2, K3 and K4
of robots 2, 3 and 4, are respectively
presence of visibility constraints. The use of polar coor-
dinates to describe conic-like visibility sets S, is also a
subject of on-going research.
K2 =
0.2066
0
0
0
0.0315 0.3361
0
0
0.0669 0.3400
0
0.5087
, K3 =
0.2678 0.3348#,
0
0
K4 =" 1.7273
0
,
Appendix A: The Fourier-Motzkin elimination
method
The Fourier-Motzkin elimination (FME), a generaliza-
tion of Gauss elimination, is a computational method
for solving a system
3[2](0), β4
1[1](0), p2
2[2](0), β3
and the initial conditions for the three dynamic systems
1[2](0), β2
in the form (19) are, (∆p2
1(0))T =
(0, 0, 0.0374)T , (∆p3
2[1](0), p3
2(0))T = (0, 0,
3[1](0), p4
0.2244)T , (∆p4
3(0))T = (0, 0, 0.2618)T ,
where pk+1
denotes the position of robot k with re-
spect to the reference frame attached to robot k + 1
and βk+1
, θk − θk+1, k ∈ {1, 2, 3}. Figs. 10(a)-(c)
report the trajectory of the 4 robots and the visibility
sets S2 (green), S3 (cyan) and S4 (gray), respectively.
Figs. 10(d)-(f) and 10(g)-(i) show the time history
of ∆pk+1
, k ∈ {1, 2, 3} and vk, ωk,
k ∈ {2, 3, 4}, and the relative state and control bounds.
[2], βk+1
[1], pk+1
k
k
k
k
k
6 Conclusions and future work
The paper proposes an original solution to the visibi-
lity maintenance problem (VMP) for a leader-follower
pair of Dubins-like vehicles with input constraints.
By interpreting the nonlinear model describing the rel-
ative dynamics of the robots as a linear system with
parameter uncertainty, the VMP is reformulated as
a linear constrained regulation problem with additive
disturbances (DLCRP). General conditions for the
positive D-invariance of linear uncertain systems with
parametric disturbance matrix are derived and used
to study the feasibility of the VMP when box bounds
on the state,
input and disturbance are considered.
The proposed design procedure can be easily adapted
to provide the control with unknown but bounded dis-
turbances rejection capabilities. Conditions for the so-
lution of the VMP when robots' desired displacement is
defined through angular parameters are also presented,
and the extension to chains of n robots is discussed.
A drawback of the approach proposed in this paper is
that it requires a great amount of computational work
off-line, because the feedback matrix K is obtained as a
solution of a large set of inequalities. In addition all the
vertices of the visibility set S are required. In this re-
spect, the solution to the LCRP proposed in (Vassilaki
and Bitsoris, 1989) appears to be preferable to the
one in (Blanchini, 1990), even though neither distur-
bances nor model parametric uncertainty are considered
therein.
Future research lines include the extension of our re-
sults to robotic networks with arbitrary topologies and
the application of the proposed approach to the study
of consensus, rendezvous and coverage problems in the
15
A x ≤ b, A ∈ Rm×n, b ∈ Rm,
of m linear inequalities in n variables (Motzkin, 1951;
Schrijver, 1986). The key idea of the FME method is
to eliminate one variable of the system A x ≤ b at each
iteration and rewrite the resulting equations accord-
ingly. Even though the number of variables decreases at
each step, the number of inequalities in the remaining
variables grows exponentially fast: in fact, at iteration
j the number of inequalities to be evaluated is at most
2 ⌋2j
⌊ m
. Because of its double-exponential computational
complexity, the FME method can be applied efficiently
only to problems with a small number of inequalities
and it is not competitive with standard LP solvers
However, differently from this numerical approach, the
method can handle symbolic inequalities. We illustrate
this idea with a simple example. Consider the following
set of symbolic inequalities
a11 x1 + a12 x2 ≤ b1,
−a21 x1 − a22 x2 ≤ b2,
a32 x2 ≤ b3,
(45)
(46)
(47)
where a11, a12, a21, a22, a32, b1, b2, b3 ∈ R>0 are un-
known parameters. We wish to determine under which
conditions on these parameters, system (45)-(47) admits
solutions. The first step is to eliminate the variable x1.
To this end, solving (45)-(46) in x1, we get:
x1 ≤
b1 − a12 x2
a11
,
x1 ≥
−b2 − a22 x2
a21
.
Joining the two inequalities, after a few computations,
we obtain the following condition on the variable x2
(assuming that a11 a22 − a12 a21 6= 0)
x2 ≥
−a11 b2 − a21 b1
a11 a22 − a12 a21
.
(48)
If we finally combine inequalities (47) and (48), we end
up with the sought solvability condition of system (45)-
(47), in terms of the symbolic parameters a11, a12, a21,
a22, a32, b1, b2, b3:
b3
a32
≥
−a11 b2 − a21 b1
a11 a22 − a12 a21
.
Appendix B: Application of the Fourier-Motzkin
elimination to the inequalities (30)-(31)
Our first step to solve the system of
inequalities
(30)-(31), is to eliminate variable k11. Solving in k11
(under the assumption of d > a), we obtain the follow-
ing set of conditions:
Combining conditions (49)-(54) and (55)-(58), we end
up with a total set of 35 inequalities, of whom only 4 are
non-trivial:
ΩL ≤
b (1 + q1) − VL sin b
d + q3
, ΩF ≥
b (1 + q1) + VL sin b
d + q3
,
,
,
VL
a
VL
a
VL
a
VL
a
q2 +
q2 +
VF ≥ VL(cid:16)1 +
VF ≥ VL(cid:16)1 +
q4 sin b
d + q3 (cid:17)+ b(cid:16)q2 +
d + q3 (cid:17)− b(cid:16)q2 +
q4 sin b
q4(1 + q1)
d + q3 (cid:17),
d + q3 (cid:17).
q4(1 + q1)
If we now assume that q4 = 0, we obtain the unique con-
dition: VF ≥ VL + b q2. Finally, if we assume that q4 < 0,
we come up with the following inequalities
,
,
k11 ≥ q4 k22 +
k11 ≥ q4 k22 −
b
a
b
a
q4 k23 +
q4 k23 −
b
a
b
a
q2 +
q2 +
k11 ≥ −q4 k22 +
k11 ≥ −q4 k22 −
b
a
b
a
q4 k23 +
q4 k23 −
b
a
b
a
k11 ≥ −
VF
a
, k11 ≤
VF
a
.
Combining these inequalities, we get:
q4 k22 ≤
q4 k22 ≤
VF
a
VF
a
−
+
b
a
b
a
q4 k23 −
q4 k23 +
b
a
b
a
q2 −
q2 −
VL
a
VL
a
,
,
q4 k22 ≥ −
q4 k22 ≥ −
VF
a
VF
a
+
−
b
a
b
a
q4 k23 +
q4 k23 −
b
a
b
a
q2 +
q2 +
VL
a
VL
a
,
.
In order to eliminate the second variable, k22, we should
consider three cases, according to the sign of q4. If we
assume that q4 > 0, we then obtain the following set of
inequalities:
k22 ≥ −
b
a
k23 +
b (1 + q1) + VL sin b
a(d + q3)
k22 ≥
b
a
k23 +
−b (1 + q1) + VL sin b
a(d + q3)
,
,
k22 ≥ −
b
a
k23 +
ΩL
a
,
k22 ≥ −
b
a
k23 +
k22 ≥
b
a
k23 +
VF + b q2 − VL
aq4
,
k22 ≥
b
a
k23 −
,
ΩF
a
b
a
−VF + b q2 + VL
k22 ≤
k23 −
,
aq4
k22 ≥ −
k22 ≤
b
a
k22 ≤ −
k22 ≤ −
k23 +
b
a
b
a
k23 +
k23 +
,
−VF − b q2 + VL
aq4
k22 ≤
b
a
ΩF
a
,
k23 +
ΩF
a
,
VF − b q2 − VL
,
aq4
k23 −
ΩF
a
,
,
b
a
ΩL
a
k22 ≥ −
b
a
k23 +
b (1 + q1) + VL sin b
a(d + q3)
k22 ≥
b
a
k23 +
−b (1 + q1) + VL sin b
a(d + q3)
,
,
k22 ≥ −
k22 ≥
b
a
k22 ≥ −
k22 ≥
k22 ≤
b
a
b
a
k22 ≤ −
b
a
k23 +
ΩL
a
,
k23 +
−VF + b q2 + VL
aq4
b
a
k23 +
−VF − b q2 + VL
aq4
,
k23 −
k23 −
b
a
k23 +
b
a
, k22 ≥ −
ΩF
a
ΩL
a
VF − b q2 − VL
,
aq4
,
,
k23 −
ΩF
a
,
k22 ≤
b
a
k23 +
VF + b q2 − VL
aq4
,
k22 ≤ −
b
a
k23 +
ΩF
a
, k22 ≤
b
a
k23 +
ΩF
a
.
from which we deduce the following two new conditions:
VF ≥ VL(cid:16)1 −
VF ≥ VL(cid:16)1 −
References
q4 sin b
d + q3 (cid:17) + b(cid:16)q2 +
d + q3 (cid:17) − b(cid:16)q2 +
q4 sin b
q4 (1 + q1)
d + q3 (cid:17),
d + q3 (cid:17).
q4 (1 + q1)
Ando, H., Y. Oasa,
I. Suzuki and M. Yamashita (1999).
Distributed Memoryless Point Convergence Algorithm for
Mobile Robots with Limited Visibility. IEEE Trans. Robot.
Autom. 15(5), 818–828.
Aubin, J.-P. and A. Cellina (1984). Differential Inclusions:
Set-Valued Maps and Viability Theory. Springer-Verlag,
Heidelberg.
Aubin, J.-P. and H. Frankowska (1990). Set-Valued Analysis.
Birkhauser, Boston.
Bhattacharya, S. and S. Hutchinson (2010). On the Existence
of Nash Equilibrium for a Two-player Pursuit-Evasion
Game with Visibility Constraints. Int. J. Robot. Res.
29(7), 831–839.
(49)
(50)
(51)
(52)
(53)
(54)
(55)
(56)
(57)
(58)
16
Bian, W. and M. French (2005). Graph Topologies, Gap Metrics
and Robust Stability for Nonlinear Systems. SIAM J. Contr.
Optim. 44(2), 418–443.
Lynch, K.M., I.B. Schwartz, P. Yang and R.A. Freeman (2008).
Decentralized Environmental Modeling by Mobile Sensor
Networks. IEEE Trans. Robot. 24(3), 710–724.
Blanchini, F. (1990). Feedback Control for Linear Time-Invariant
Systems with State and Control Bounds in the Presence of
Disturbances. IEEE Trans. Automat. Contr. 35(11), 1231–
1234.
Blanchini, F. (1991). Constrained Control for Uncertain Linear
Systems. J. Optim. Theory Appl. 71(3), 465–484.
Blanchini, F. (1999). Set invariance in control. Automatica
35(11), 1747–1767.
Blanchini, F. and S. Miani (2008). Set-Theoretic Methods in
Control. Systems & Control: Foundations & Applications.
Birkhauser, Boston.
Bullo, F., J. Cort´es and S. Mart´ınez (2009). Distributed
Control of Robotic Networks. Applied Mathematics Series.
Princeton University Press. Available electronically at
http://www.coordinationbook.info.
Casbeer, D.W., D.B. Kingston, R.W. Beard, T.W. Mclain, S.-M.
Li and R. Mehra (2006). Cooperative forest fire surveillance
using a team of small unmanned air vehicles. Int. J. Syst.
Sci. 37(6), 351–360.
Cort´es, J., S. Mart´ınez and F. Bullo (2006). Robust rendezvous for
mobile autonomous agents via proximity graphs in arbitrary
dimensions. IEEE Trans. Automat. Contr. 51(8), 1289–1298.
Dimarogonas,
and
D.V.
K.J.
Kyriakopoulos (2008). Connectedness Preserving Distributed
Swarm Aggregation for Multiple Kinematic Robots. IEEE
Trans. Robot. 24(5), 1213–1223.
Dubins, L.E. (1957). On curves of minimal
length with a
constraint on average curvature and with prescribed initial
and terminal positions and tangents. American J. Math.
79(3), 497–516.
Finke, J., K.M. Passino, S. Ganapathy and A. Sparks (2005).
Modeling and Analysis of Cooperative Control Systems for
Uninhabited Autonomous Vehicles. In: Cooperative Control.
Vol. 309 of Lecture Notes in Control and Information
Sciences. pp. 79–102. Springer.
Gerkey, B.P., S. Thrun and G. Gordon (2006). Visibility-based
Pursuit-evasion with Limited Field of View. Int. J. Robot.
Res. 25(4), 299–315.
Glizer, V.Y. (1999). Homicidal Chauffeur Game with Target
Set in the Shape of a Circular Angular Sector: Conditions
for Existence of a Closed Barrier. J. Optim. Theory Appl.
101(3), 581–598.
Guibas, L.J., J.C. Latombe, S.M. LaValle, D. Lin and R. Motwani
(1997). Visibility-based pursuit-evasion in a polygonal
environment. In: Algorithms and Data Structures. Vol. 1272
of Lecture Notes in Computer Science. pp. 17–30. Springer.
Isler, V., S. Kannan and S. Khanna (2005). Randomized pursuit-
evasion in a polygonal environment. IEEE Trans. Robot.
5(21), 875–884.
James, M.R., M.C. Smith and G. Vinnicombe (2005). Gap
Metrics, Representations, and Nonlinear Robust Stability.
SIAM J. Contr. Optim. 43(5), 1535–1582.
Jarmark, B. and C. Hillberg (1984). Pursuit-evasion between two
realistic aircraft. AIAA J. Guid. Contr. Dynam. 7, 690–694.
Ji, M. and M. Egerstedt (2007). Distributed Coordination Control
of Multiagent Systems while Preserving Connectedness.
IEEE Trans. Robot. 23(4), 693–703.
Lin, J., A.S. Morse and B.D.O. Anderson (2007). The Multi-
Agent Rendezvous Problem. Part 1: The Synchronous Case.
SIAM J. Contr. Optim. 46(6), 2096–2119.
17
Mazo Jr, M., A. Speranzon, K.H. Johansson and X. Hu (2004).
Multi-Robot Tracking of a Moving Object using Directional
Sensors. In: Proc. IEEE Int. Conf. Robot. Automat. pp. 1103–
1108.
Merz, A.W. and D. S. Hague (1977). Coplanar Tail-Chase Aerial
Combat as a Differential Game. AIAA Journal 15, 1419–
1423.
Mesbahi, M. and M. Egerstedt (2010). Graph Theoretic Methods
in Multiagent Networks. Princeton University Press.
Morbidi, F., F. Bullo and D. Prattichizzo (2008). On Leader-
Follower Visibility Maintenance via Controlled Invariance for
Dubins-like Vehicles. In: Proc. 47th IEEE Conf. Dec. Contr.
pp. 1821–1826.
Moreau, L. (2005). Stability of multiagent systems with time-
dependent communication links. IEEE Trans. Automat.
Contr. 50(2), 169–182.
Motzkin, T.S. (1951). Two consequences of the transposition
theorem on linear inequalities. Econometrica 19(2), 184–185.
Olfati-Saber, R., J.A. Fax and R.M. Murray (2007). Consensus
and Cooperation in Networked Multi-Agent Systems. Proc.
IEEE 95(1), 215–233.
Panik, M.J. (1993). Fundamentals of convex analysis: duality,
separation, representation, and resolution. Kluwer Academic
Publishers.
Savla, K., G. Notarstefano and F. Bullo (2009). Maintaining
Limited-Range Connectivity among Second-Order Agents.
SIAM J. Contr. Optim. 48(1), 187–205.
Schrijver, A. (1986). Theory of Linear and Integer Programming.
John Wiley. New York.
Shima, T.
and J. Shinar
(2002). Time-Varying Linear
PursuitEvasion Game Models with Bounded Controls. AIAA
J. Guid. Contr. Dynam. 25(3), 425–432.
Siciliano, B., L. Sciavicco, L. Villani and G. Oriolo (2008).
Robotics: Modelling, Planning and Control. Advanced
Textbooks in Control and Signal Processing. Springer.
Spanos, D.P. and R.M. Murray (2005). Motion planning with
wireless network constraints. In: Proc. American Contr.
Conf. pp. 87–92.
Susca, S., F. Bullo and S. Mart´ınez (2008). Monitoring
Environmental Boundaries with a Robotic Sensor Network.
IEEE Trans. Contr. Syst. Tech. 16(2), 288–296.
Tiwari, A., J. Fung, R. Bhattacharya and R.M. Murray (2004).
Polyhedral Cone Invariance Applied to Rendezvous of
Multiple Agents. In: Proc. 43rd IEEE Conf. Dec. Contr.
pp. 165–170.
Tomlin, C.J., J. Lygeros and S.S. Sastry (2000). A game theoretic
approach to controller design for hybrid systems. Proc. IEEE
88(7), 949–970.
Vassilaki, M. and G. Bitsoris (1989). Constrained Regulation of
Linear Continuous-time Dynamical Systems. Syst. Contr.
Lett. 13(3), 247–252.
Yang, P., R. Freeman, G. Gordon, K. Lynch, S. Srinivasa and
R. Sukthankar (2010). Decentralized estimation and control
of graph connectivity for mobile sensor networks. Automatica
46(2), 390–396.
Zavlanos, M.M. and G.J. Pappas (2007). Potential Fields for
Maintaining Connectivity of Mobile Networks. IEEE Trans.
Robot. 23(4), 812–816.
|
1510.09115 | 1 | 1510 | 2015-10-30T15:12:02 | Continuous Time Gathering of Agents with Limited Visibility and Bearing-Only Sensing | [
"cs.MA",
"eess.SY"
] | A group of mobile agents, identical, anonymous, and oblivious (memoryless), having the capability to sense only the relative direction (bearing) to neighborhing agents within a finite visibility range, are shown to gather to a meeting point in finite time by applying a very simple rule of motion. The agents' rule of motion is : set your velocity vector to be the sum of the two unit vectors in R^2 pointing to your "extremal" neighbours determining the smallest visibility disc sector in which all your visible neighbors reside, provided it spans an angle smaller than pi, otherwise, since you are "surrounded" by visible neighbors, simply stay put (set your velocity to 0). Of course, the initial constellation of agents must have a visibility graph that is connected, and provided this we prove that the agents gather to a common meeting point in finite time, while the distances between agents that initially see each other monotically decreases. We will also prove a geometrical result, a tight lower bound on the sum of cosines of the interior angles of a convex polygon, that we will use to prove the gathering of our dynamical system. | cs.MA | cs |
Continuous Time Gathering of
Agents with Limited Visibility
and Bearing-Only Sensing
Levi-Itzhak Bellaiche and Alfred Bruckstein
Center for Intelligent Systems - MARS Lab
Autonomous Systems Program
Technion Israel Institute of Technology
32000 Haifa, Israel
Abstract
A group of mobile agents, identical, anonymous, and oblivious (memoryless),
having the capacity to sense only the relative direction (bearing) to neighborhing
agents within a finite visibility range, are shown to gather to a meeting point in
finite time by applying a very simple rule of motion. The agents' rule of motion
is : set your velocity vector to be the sum of the two unit vectors in R2 pointing
to your "extremal" neighbours determining the smallest visibility disc sector in
which all your visible neighbors reside, provided it spans an angle smaller than
π, otherwise, since you are "surrounded" by visible neighbors, simply stay put
(set your velocity to 0). Of course, the initial constellation of agents must have
a visibility graph that is connected, and provided this we prove that the agents
gather to a common meeting point in finite time, while the distances between
agents that initially see each other monotonically decreases.
1
1
Introduction
This paper studies the problem of mobile agent convergence, or robot gather-
ing under severe limitations on the capabilities of the agent-robots. We assume
that the agents move in the environment (the plane R2) according to what they
currently "see", or sense in their neighborhood. All agents are identical and
indistinguishable (i.e. they are anonimous having no i.d's) and, all of them are
performing the same "reactive" rule of motion in response to what they see.
Our assumption will be that the agents have a finite visibility range V , a dis-
tance beyond which they cannot sense the presence of other agents. The agents
within the "visibility disk" of radius V around each agent are defined as his
neighbors, and we further assume that the agent can only sense the direction
to its neighbors, i.e.
it performs a "bearing only" measurement yielding unit
vectors pointing toward its neighbor.
Therefore, in our setting, each agent senses its neighbors within the visibility
disk and sets its motion only according to the distribution of unit vectors point-
ing to its current neighbors. Figure 1 shows a constellation of agents in the
plane (R2), their "visibility graph" and the visibility disks of some of them,
each agent moves based on the set of unit vectors pointing to its neighbors.
Figure 1: A constellation of agents in the plane displaying the "visibility disks"
of agents Ak, Al, Ai, Aj, Ap and the visibility graph that they define, having
edges connecting pairs of agents that can see each other.
2
In this paper we shall prove that continuous time limited visibility sensing of di-
rections only and continuous adjustment of agents' velocities according to what
they see is enough to ensure the gathering of the agents in finite time to a point
of encounter.
The literature of robotic gathering is vast and the problem was addressed un-
der various assumptions on the sensing and motion capabilities of the agents.
Here we shall only mention papers that deal with gathering assuming contin-
uous time motion and limited visibility sensing, since these are most relevant
to our work reported herein. The paper [1] by Olfati-Saber, Fox, and Murray,
nicely surveys the work on the topic of gathering (also called consensus) for
networked multi agent systems, where the connections between agents are not
necessarily determined by their relative position or distance. This approach to
multi-agent systems was indeed the subject of much investigation and some of
the results, involving "switching connection topologies" are useful in dealing
with constellation-defined visibility-based interaction dynamics too. A lot of
work was invested in the analysis of "potential functions" based multi-agent
dynamics, where agents are sensing each other through a "distance-based" in-
fluence field, a prime example here being the very influential work of Gazi and
Passino [2] which analyses beautifully the stability of a clustering process. In-
teractions involving hard limits on the "visibility distance" in sensing neighbors
were analysed in not too many works. Ji and Eggerstedt in [3] analysed such
problems using potential functions that are "visibility-distance based barrier
functions" and proved connectedness-preservation properties at the expense of
making some agents temporarily "identifiable" and "traceable" via a hystere-
sis process. Ando, Oasa, Suzuki and Yamashita in [4] were the first to deal
with hard constraints of limited visibility and analysed the "point convergence"
or gathering issue in a discrete-time synchronized setting, assuming agents can
see and measure both distances and bearing to neighbors withing the visibility
range.
Subsequently, in a series of papers, Gordon, Wagner, and Bruckstein, in [5],
[6], [7], analysed gathering with limited visibility and bearing only sensing con-
straints imposed on the agents. Their work proved gathering or clustering results
in discrete-time settings, and also proposed dynamics for the continuous-time
settings. In the sequel we shall mention the continuous time motion model they
analysed and compare it to our dynamic rule of motion.
In our work, as well as most of the papers mentioned above one assumed that
the agents can directly control their velocity with no acceleration constraints.
We note that the literature of multi-agent systems is replete with papers assum-
ing more complex and realistic dynamics for the agents, like unicycle motions,
second order systems and double integration models relating the location to the
controls, and seek sensor based local control-laws that ensure gathering or the
achievement of some desired configuration. However we feel that it is still worth-
while exploring systems with agents directly controlling their velocity based on
very primitive sensing, in order to test the limits on what can be achieved by
agents with such simple, reactive behaviours.
3
2 The gathering problem
We consider N agents located in the plane (R2) whose positions are given by
{Pk = (xk, yk)T}k=1,2,...,N , in some absolute coordinate frame which is unknown
to the agents.We define the vectors
(cid:40) Pj−Pi
(cid:107)Pj−Pi(cid:107)
0
uij =
0 < (cid:107)Pj − Pi(cid:107) ≤ V
(cid:107)Pj − Pi(cid:107) = 0 or (cid:107)Pj − Pi(cid:107) > V
hence uij are, if not zero, the unit vectors from Pi to all Pj's which are neighbors
of Pi in the sense of being at a distance less than V from Pi, i.e. Pj's obeying
:
(cid:107)Pj − Pi(cid:107) (cid:44) [(Pj − Pi)T (Pj − Pi)]1/2 ≤ V
i and u−
Note that we have uij = −uji,∀(i, j). For each agent Pi, let us define the special
vectors, u+
i (from among the vectors uij defined above). Consider the
nonzero vectors from the set {uij}j=1,2,...,N . Anchor a moving unit vector ¯η(θ)
at Pi poiting at some arbitrary neighbor, i.e. at uik (cid:54)= 0, ¯η(0) = uik and rotate
it clockwise, sweeping a full circle about Pi. As ¯η(θ) goes from η(0) to η(2π)
it will encounter all the possible uij's and these encounters define a sequence
of angles α1, α2, . . . , αr that add to 2π = α1 + . . . + αr (αk = angle from k-th
to (k+1)-th encounter with a uij, αr = angle from last encounter to first one
again, see Figure 2). If none of the angles {α1, . . . , αr} is bigger than π, set
i = u−
u+
i = ui(n) the unit vectors
encountered when entering and exiting the angle αb > π bigger than π.
i = 0. Otherwise define u+
i = ui(m) and u−
4
Figure 2: Leftmost and rightmost visible agents of agent located at Pi.
One might call u−
i the pointer to the "leftmost visible agent" from Pi and u+
i
the pointer to the "rightmost visible agent" among the neighbors of Pi. If Pi has
nonzero right and leftmost visible agents it means that all its visible neighbors
belong to a disk sector defined by an angle less than π, and Pi will be movable.
Otherwise we call him "surrounded" by neighbors and, in this case, it will stay
in place while it remains surrounded (see Appendix 2 for an alternative way of
defining the leftmost and rightmost agents).
The dynamics of the multi-agent system will be defined as follows.
dPi
dt
= v0(u+
i + u−
i ) for i = 1, . . . , N
(1)
Note that the speed of each agent is in the span of [0, 2v0].
With this we have defined a local, distributed, reactive law of motion based on
the information gathered by each agent. Notice that the agents do not commu-
nicate directly, are all identical, and have limited sensing capabilities, yet we
shall show that, under the defined reactive law of motion, in response to what
they can "see" (which is the bearings to their neighboring agents), the agents
will all come together while decreasing the distance between all pairs of visible
agents.
Assume that we are given an initial configuration of N agents placed in the
plane in such a way that their visibility graph is connected. This just means
5
that there is a path (or a chain) of mutually visible neighbors from each agent
to any other agent.
Our first result is that while agents move according to the above described rule
of motion, the visibility graph will only be supplemented with new edges and
old "visibility connections" will never be lost.
2.1 Connectivity is never lost
We shall show that
Theorem 2.1.1. A multi agent systems under the dynamics
{ Pi = v0(u+
i + u−
i )}i=1,...,N )
ensures that pairs of neighboring agents at t = 0 (i.e. agents at a distance less
than V ) will remain neighbors forever.
Proof. To prove this result we shall consider the dynamics of distances between
pairs of agents.
We have that the distance ∆ij between Pi and Pj is
hence
or
∆ij = (cid:107)Pj − Pi(cid:107) = [(Pj − Pi)T (Pj − Pi)]1/2
d
dt
∆(t)
ij =
1
(cid:107)Pj − Pi(cid:107) (Pj − Pi)T ( Pj − Pi)
d
dt ∆(t)
ij =
ij( Pj − Pi)
uT
= −uT
Pi + uT
Pj
ij
Pi − uT
= −uT
Pj
ij
ji
ij
But we know that the dynamics (1) is
Pi = v0(u+
Pj = v0(u+
i + u−
i )
j + u−
j )
Therefore
d
dt
∆ij = −v0uT
ij(u+
i + u−
ji(u+
j + u−
j )
i ) − v0uT
i +u−
j
(cid:44) 0 if agent is surrounded, or
is in the direction of the center of the disk sector in which all neighbors
However for every agent Pi we have either u+
i + u−
u+
(including Pj) reside (see Figure 3).
i
6
Figure 3: uT
ij(u+
i + u−
i ) ≥ 0
i ) =< uij, (u+
Therefore the inner product uT
be positive (see Appendix 3 for a formal proof), hence
ij(u+
i + u−
i + u−
i ) > will necessary
ij = −(v0 ∗ positive + v0 ∗ positive) ≤ 0
∆(t)
d
dt
This shows that distances between neighbors can only decrease (or remain the
same). Hence agents never lose neighbors under the assumed dynamics.
2.2 Finite-time gathering
We have seen that the dynamics of the system (1) ensures that agents that are
neighbors at t = 0 will forever remain neighbors. We shall next prove that,
as time passes, agents acquire new neighbors and in fact will all converge to a
common point of encounter. We prove the following.
Theorem 2.2.1. A multi-agent system with dynamics given by (1) gathers all
agents to a point in R2, in finite time.
Proof. We shall rely on a Lyapunov function L(P1, . . . , PN ), a positive function
defined on the geometry of agent constellations which becomes zero if and only
if all agents' locations are identical. We shall show that, due to the dynamics
7
of the system, the function L(P1, . . . , PN ) decreases to zero at a rate bounded
away from zero, ensuring finite time convergence.
The function L will be defined as the perimeter of the convex hull of all agents'
locations, CH{Pi(t)}i=1,...,N . Indeed, consider the set of agents that are, at a
given time t, the vertices of the convex hull of the set {Pi(t)}i=1,...,N . Let us call
these agents { Pk(t)} for k = 1, . . . , K ≤ N . For every agent Pk on the convex
hull (i.e.
for every agent that is a corner of the convex polygon defining the
convex hull), we have that all other agents, are in a region (wedge) determined
by the half lines from Pk in the directions Pk Pk−1 and Pk Pk+1, a wedge with
an opening angle say θk (see Figure 4).
Figure 4: Angle at a vertex of the convex hull
k and u−
Since clearly θk ≤ π for all k we must have that agent Pk has all its visible
neighbors in a wedge of his visibility disk with an angle αk ≤ θk ≤ π hence
k vectors will not be zero, causing the motion of Pk towards the
his u+
interior of the convex hull. This will ensure the shrinking of the convex hull,
while it exists, and the rate of this shrinking will be determined by the evolution
of the constellation of agents' locations. Let us formally prove that indeed, the
convex hull will shrink to a point in finite time.
Consider the perimeter L(t) of CH{Pi(t)}i=1,...,N
K(t)(cid:88)
K(t)(cid:88)
L(t) =
∆k,k+1 =
[( Pk+1)(t) − Pk(t))T ( Pk+1(t) − Pk(t))]1/2
k=1
k=1
where the indices are considered modulo K(t).
We have, assuming that K remains the same for a while,
8
K(cid:88)
K(cid:88)
k=1
K(cid:88)
K(cid:88)
k=1
K(cid:88)
k=1
K(cid:88)
k=1
d
dt
∆k = − K(cid:88)
k=1
d
dt
L(t) =
(cid:0)v0 uT
k,k+1(u+
k + u−
k ) + v0 uT
k,k+1(u+
k+1 + u−
k+1)(cid:1)
but note that uk,k+1 does not necessarily lie between u+
in fact, Pk and Pk+1 might not even be neighbors.
k and u−
k anymore, since,
Now let us consider d
dt L(t) and rewrite it as follows
L(t) = −v0
d
dt
k,k+1(u+
uT
k + u−
k ) − v0
k+1,k(u+
uT
k+1 + u−
k+1)
k=1
k=1
Rewriting the second term above, by moving the indices k by -1 we get
L(t) = −v0
d
dt
k,k+1(u+
uT
k + u−
k ) − v0
This yields
K(cid:88)
k=1
L(t) = −v0
d
dt
< u+
k , uk,k+1 + uk,k−1 > −v0
k,k−1(u+
uT
k + u−
k )
< u−
k , uk,k+1 + uk,k−1 >
Note that we have here inner products between unit vectors, yielding the cosines
of the angles between them. Therefore, defining θk = the angle between uk,k−1
and uk,k+1 (i.e. the interior angle of the convex hull at the vertex k, see Figure
5), and the angles :
α+
k
β+
k
α−
β−
k
k
(cid:44) γ(u+
k , uk,k+1)
(cid:44) γ(uk,k−1, u+
k )
(cid:44) γ(uk,k−1, u−
k )
(cid:44) γ(u−
k , uk,k+1)
9
Figure 5: Angles at a vertex of the convex hull.
we have α+
k + β+
k = α−
k + β−
k = θk and all these angles are between 0 and π.
Using these angles we can rewrite
L(t) = − K(cid:88)
d
dt
k ) − K(cid:88)
v0(cos α+
k + cos β+
k=1
k=1
v0(cos α−
k + cos β−
k )
Now, using the inequality (proved in Apenedix 1)
we obtain that
cos α + cos β ≥ 1 + cos(α + β)
0 ≤ α, β, α + β ≤ π
− d
dt
L(t) ≥ 2v0
K(cid:88)
i=1
(1 + cos θi)
(2)
(3)
For any convex polygon we have the following result (see the detailed proof in
Appendix 1) :
Lemma 1. For any convex polygon with K vertices and interior angles θ1, . . . , θK,
with (θ1 + . . . + θK) = (K − 2)π we have that
K(cid:88)
k=1
cos(θi) ≥
1 + (K − 1) cos
(cid:16) (K−2)π
(cid:16) (K−2)π
(cid:17)
K cos
K−1
K
(cid:17)
2 ≤ K ≤ 6
K ≥ 7
(4)
10
Therefore, we obtain from (3) and (4) that
L(t) ≥ µ(K)
− d
dt
where
µ(K) = 2v0
K +
=
2v0K
(cid:17)
1 + (K − 1) cos
(cid:16) (K−2)π
(cid:16) (K−2)π
(cid:17)
(cid:110)
(cid:16)
(cid:1) , K−1
cos(cid:0) 2π
1 − max
K cos
K−1
K
(cid:16) π
2 ≤ K ≤ 6
(cid:111)(cid:17)
K ≥ 7
(cid:17) − 1
K
K
K cos
K−1
(5)
Note here that, since (1 − max{. . .}) > 0 we have that the rate of decrease in
the perimeter of the configuration is srictly positive while the convex hull of the
agent location is not a single point.
The argument outlined so far assumed that the number of agents determining
the convex hull of their constellation is a constant K. Suppose however that
in the course of evolution some agents collide and/or some agents become "ex-
posed" as vertices of the convex hull, and hence K may jump to some different
integer value. At a collision between two agents we assume that they merge and
thereafter continue to move as a single agent. Since irrespective of the value
of K the perimeter decreases at a rate which is strictly positive and bounded
away from zero we have effectively proved that in finite time the perimeter of
the convex hull will necessarily reach 0. This concludes the proof of Theorem
2.2.1.
Figure 6 shows the bound as a function of K assuming v0 = 1. Note that we
always have K ≤ N , and µ(K) is a decreasing function of K, hence for any finite
number of agents there will be a strictly positive constant µ(N ) so that
L(t) ≥ µ(N )
− d
dt
ensuring that after a finite time of Tub given by
µ(N )Tub = L(0) ⇒ Tub =
L(0)
µ(N )
we shall have that L(Tub) = 0.
Hence we have an upper bound on the time of convergence for any configuration
11
of N agents given by L(0)
µ(N ) .
Note from (3) that the rate of decrease does not depend on the perimeter of the
convex hull but only on the number of agents forming it. This was an expected
result, since the dynamics does not rely on Euclidian distances. This bound is
decreasing with K, so that the more agents form the convex hull, the smaller
will be the rate of decrease. For K = 2 and K = 3 the bound is −8v0, for
K = 4, it is −7v0, and then it keeps increasing for higher values of K, slowly
converging towards 0 form below. Note the change of curve between K = 6 and
K = 7, due to the "interresting" discontinuity in the geometric result exhibited
in equation (1).
Figure 6: Graph of the bound µ(K) of (3). The graph on the right is a zoom
on small values of K.
The inequalities of (2) and of (4) become equalities for particular configurations
of the agents (for example a regular polygon in which each pair of adjacent
neighbors are visible to each other, if K ≥ 7). In this case, the bound in (3)
will yield the exact rate of convergence of the convex-hull perimeter as long as
K remains the same.
3 Generalizations
All the above analysis can be generalized for dynamics of the form
dPi
dt
= f (P (i))(u+
i + u−
i ) for i = 1, . . . , N
(6)
f (P (i)) ≥ 0 is some positive function of the configuration of the neighbours seen
by agent i. This generalization also guarantees that the rule of motion is locally
defined and reactive, and defined in the same way for all agents.
The dynamics (1) correspond to a particular case of (6), with f (P (i)) = v0 = constant
12
for all agents.
It is easy to slightly change the proofs above in order to show that Theorem
2.1.1 (ensuring that connectivity is not lost) is still valid as long as f (P (i)) ≥ 0
for all i, and that Theorem 2.2.1 (ensuring finite time gathering) is also valid as
long as f (P (i)) ≥ > 0 for all i.
Note that in the work of Gordon et al [5], a constant speed for the agents was
considered, and this corresponds to setting f (P (i)) =
i for a mobile
−
agent i, rather than v0. Given that in this case f (P (i)) ≥ 1
2 , the conditions
for Theorem 2.1.1 and Theorem 2.2.1 are verified, and hence the dynamics
with constant speed also ensures convergence to a single point without pairs
of initially visible agents losing connectivity. We therefore also have a proof
for the convergence of the algorithm that was proposed in the above-mentioned
paper.
u+
1
i +u
However, it was pointed out in the above-mentioned paper that in the model
with constant speed agents of one has to deal with quite unpleasant chattering
effects and "Zeno"-ness in order to effectively modulate the speed of motion of
some of the agents. In fact, in certain configurations, an agent may "oscillate"
infinitely often between a position in which it is "surrounded" to a position
in which it is not. This implies alternating its speed infinitely often between
zero and a constant value. In contrast, in our model, the speed is defined to
vary smoothly in the range [0, 2v0]. Therefore our model presents some clear
advantages and natural modulation for the speed of the agents.
4 Simulations
Let us consider some simulations of the multi-agent dynamics discussed in this
paper. We will start with a randomly generated swarm and then we shall have
a look at some interesting particular configurations.
4.1 Some random generated swarm of 15 agents
First, we randomly generate a swarm of 15 agents, with a connected visibility
graph as initial configuration. We set v0 to 1 and visibility to 200. The config-
uration of the swarm at different times during the evolution is given in Figure
7.
13
Figure 7: Configuration of the swarm at different times. The convex hull of the
set of agents is also shown.
We also plot some properties of the swarm during the time of its evolution.
Figure 8 represents the perimeter of the convex hull of the set of agents. Figure
9 plots the count of indistinguishable agents (i.e. collided agents count as one)
in the convex hull. Figure 10 represents the time derivative of the convex hull
perimeter and the theoretical bound given by equation (5), as function of the
number of indistinguishable agents in the convex hull.
14
Figure 8: Perimeter of the convex hull, decreasing until it reaches zero.
Figure 9: Number of agents forming the convex hull of the set. Decreases when
two or more agents of the hull "merge" and increases when one or more agents
are "collected" by the hull and added to it.
15
Figure 10: Derivative of the convex hull perimeter (blue) and theoretical bound
(green), function of number of agents forming the convex hull, given by (5).
For a fixed number of agents forming the convex hull, one can see in Figure 10
that the derivative of the perimeter goes towards the theoretical bound. This
can be intuitively explained by the facts that far away agents evolve towards the
inside more rapidly, making the convex hull shape more regular, approaching to
the shape that yields the theoretical bound.
The discontinuity of the derivative of the perimeter of the convex hull that
occurs when there is no change in the number of agents forming the hull, for
example around t = 35 in Figure 10, is due to change in the connectivity graph
(which has not been printed for clarity). In this particular case, two agents on
the top left became visible to each other at this time and their directions and
speed changed, slowing down slightly the rate of perimeter decrease.
4.2 Regular polygon of 10 agents
The initial configuration is a regular polygon with 10 agents. Again, v0 is set
to 1 and visibility is 200. As expected, the decreasing rate of the perimeter of
the convex hull is constant, and all the agents contribute to the convex hull all
along, until the very end where they collide and merge. Simulations and the
converging parameters' plots are seen in Figures 11 and 12.
16
Figure 11: Starting as regular polygon, the swarm keeps its regular shape until
the final collision. The leftmost and rightmost agents of each agent are its two
neighbors in the polygon.
Figure 12: From left to right : the perimeter of the convex hull decreasing at a
constant rate, the number of agents forming the convex hull being constant, and
the derivative of the perimeter of the convex hull being constant and equaling
its theoretical bound all along the evolution.
4.3 Close-to-minimum-configuration with n = 4
We start at a configuration close to the configuration that reaches the minimum
of the sum of the cosinuses of the interior angles of the polygon. Results are
represented in Figures 13, 14, 15. One can see that this configuration provides
a smaller decreasing rate than the configuration of a regular polygon, in confor-
mity with our analysis, where the bound is not realized for a regular polygon
configuration for small values of n.
17
Figure 13: The cone configuration approaches the theoretical bound configu-
ration. In this case, the gathering can be divided in two stages. First, the 3
agents of one side merge, then the resulting composite agent merges with the
single agent of the other side.
Figure 14: From top to bottom : The perimeter of the convex hull, the number
of agents forming it, and the derivative of it.
18
Figure 15: A zoom on the the derivative of the perimeter of the convex hull
(in blue) at the beginning of the dynamics. A regular polygon (a square in this
case) would give a constant rate of -8, lower than the current rate approaching
the theoretical bound of -7 (in green).
5 Concluding remarks
We have shown that a very simple local control on the velocity of agents in
the plane, based on limited visibility and bearing only sensing of neighbors
ensures their finite time gathering. The motion rule is simply to set the agents'
velocity to equal the vector sum of unit vector pointers to two external neighbors
if all visible neighbors reside inside a halfplane (a half-disk) about the agent,
otherwise set the velocity to zero. This very simple rule of behavior is slightly
different from the one assumed by Gordon, Wagner, and Bruckstein in [5], where
the motion was set to have a constant velocity in the direction bisecting the disk
sector where visible neighbors reside, or zero if the agent was "surrounded".
As we showed in this paper, that model also ensures gathering. However, it
was pointed out there that their proposed model had to deal with some quite
unpleasant chattering or zeno-ness effects in order to effectively modulate the
speed of motion of agents.
In this paper, in conjunction with our model, and some generalizations too,
including the model of [5], we provided a very simple geometric proof that finite
time gathering is achieved, and provided precise bounds on the rate of decrease
of the perimeter of the agent configuration's convex hull. These bounds are
based on a geometric lower bound on the sum of cosines of the interior angles
of an arbitrary convex planar polygon, that is interesting on its own right (a
curious breakpoint occurring in the bound at 7 vertices). Our result may be
regarded as a convergence proof for a highly nonlinear autonomous dynamic
system, naturally handling dynamic changes in its dimension (the events when
two agents meet and merge).
19
References
[1] Olfati Saber R., Fox J.A., and Murray R.M. Consensus and cooperation
in networked multi-agent systems. Proceedings of the IEEE, 95-1:215–233,
2007.
[2] Gazi V. and Passimo K.M. Stability analysis of swarms. IEEE Transactions
on Automatic Control, 51-6:692–697, 2006.
[3] Ji M. and Eggerstedt M. Distributed coordination control of multiagent
systems while preserving connectedness. IEEE Transactions on Robotics,
23-4:693–703, 2007.
[4] Ando H.and Oasa Y., Suzuki I., and Yamashita. Distributed memoryless
point convergence algorithm for mobile robots with limited visibility. IEEE
Transactions on Robotics and Automation, 15-5:818–828, 1999.
[5] Gordon N., Wagner I.A., and Bruckstein A.M. Gathering multiple robotic
a(ge)nts with limited sensing capabilities. Lecture Notes in Computer Sci-
ence, 3172:142–153, 2004.
[6] Gordon N., Wagner I.A., and Bruckstein A.M. A randomized gathering
algorithm for multipe robots with limited sensing capabilities. MARS 2005
Workshop Proceedings (ICINCO 2005), Barcelona, Spain, 2005.
[7] Gordon N., Elor Y., and Bruckstein A.M. Gathering multiple robotic agents
with crude distance sensing capabilities. Lecture Notes in Computer Science,
5217:72–83, 2008.
20
Appendix 1 : Proof of Lemma 1
We will first prove the following facts :
Fact a. Let 0 ≤ a ≤ b ≤ π and 0 ≤ a + b ≤ π. Then we have
(cid:112)2(1 + cos(a + b)) = 2 cos
(cid:18) a + b
(cid:19)
2
≥ cos(a)+cos(b) ≥ 2 cos2
(cid:18) a + b
(cid:19)
2
= 1+cos(a+b)
2 ≥ b−a
2
:
Proof. The function cosine is decreasing in [0, π], and given that a+b
(cid:18) b − a
(cid:19)
2
≥ cos
(cid:18) a + b
(cid:19)
1 ≥ cos
Multiplying by 2 cos(cid:0) a+b
(cid:1) ≥ 0 :
2 cos(cid:0) a+b
2 cos(cid:0) a+b
2
2
(cid:1) ≥ 2 cos(cid:0) a+b
(cid:1) cos(cid:0) b−a
(cid:1) ≥ cos(a) + cos(b)
2
2
2
2
(cid:1) ≥ 2 cos2(cid:0) a+b
(cid:1)
≥ 1 + cos(a + b)
2
A direct consequence is the following lemma.
Fact b. Let 0 ≤ a, b ≤ π. Then
cos(a) + cos(b) ≥
(cid:26) 1 + cos(a + b)
2 cos(cid:0) a+b
(cid:1)
2
: a + b ≤ π
: a + b ≥ π
Proof. The first line is already part of Lemma a. The second line can be proven
by using the left inequality of Lemma a with π − a and π − b, noticing that
0 ≤ π − a ≤ π, 0 ≤ π − b ≤ π, and π − a + π − b ≤ π for a + b ≥ π.
Now we can prove Lemma 1. Suppose any given initial configuration of the poly-
gon with interior angles 0 ≤ x1, . . . , xn ≤ π. We then have x1+. . .+xn = (n−2)π.
Now repeat the following step : Go through all the pairs of non-zero values
(xi, xj). As long as there is still a pair verifying xi + xj ≤ π, transform it from
(xi, xj) to (0, xi + xj). When there are no such pairs, then among all the non-
zero values, take the the minimum value and the maximum value, say xi and
xj (they must verify xi + xj ≥ π due to the previously applied process) , and
transform the pair from (xi, xj) to
(cid:16) xi+xj
, xi+xj
(cid:17)
.
2
2
Repeat the above process until convergence. We prove that the process con-
verges and that we can get as close as wanted to a configuration where all
non-zero values are equal. Note that after each step, the sum of the values is
21
unchanged, (n − 2)π, and that the values of all xi's remain between 0 and π.
The number of values that the above process set to zero must be less or equal to
2 in order to have the sum of the n positive values equal to (n − 2)π. Therefore
we can be sure that after a finite number of iterations, there will be no pairs of
nonzero values whose sum will be less than π (otherwise this would allow us to
add a zero value without changing the sum).
(cid:16) xi+xj
(cid:17)
Once in this situation, all we do is replacing pairs of "farthest" non-zero values
. Let us show that all the nonzero values
(xi, xj) with the pair
will converge to the same value, specifically to their mean.
, xi+xj
2
2
Let k be the number of remaining non-zero values after the iteration t0 which
sets the "last value" to zero. Denote these values at the i-th iteration by
(x(i)
1 , . . . , x(i)
k ). Define :
m =
x(i)
1 + . . . + x(i)
k
k
(n − 2)π
k
=
Ei = (x(i)
1 − m)2 + . . . + (x(i)
k − m)2
(cid:18)
Suppose, without loss of generality, that at the i-th iteration the extreme values
were x1 and x2 and so we transformed (x(i)
So we have :
1 , x(i)
2 ) into
, x(i+1)
x(i+1)
1
= x(i)
1 +x(i)
2
2
2
= x(i)
1 +x(i)
2
2
(cid:19)
.
Ei+1 − Ei = 2( x(i)
1 +x(i)
2
2
=
− m)2 − (x(i)
2 (x(i)
1 − m)2 − (x(i)
1 − x(i)
2 )2
− 1
2 − m)2)
But x(i)
1 and x(i)
2 being the extreme values, we have for any 1 ≤ l ≤ k :
1 − x(i)
(x(i)
2 )2 ≥ (x(i)
l − m)2
and by summing over l we get that :
k(x(i)
1 − x(i)
2 )2 ≥ Ei
Hence
22
(cid:16) (n−2)π
(cid:16) (n−2)π
n−2
n−1
(cid:17)
(cid:17)
= −(n − 4) (case with 2 zeros)
(case with 1 zero)
(case with no zero)
• 2 + (n − 2) cos
• 1 + (n − 1) cos
• n cos
(cid:17)
(cid:16) (n−2)π
is increasing in (cid:2)0, π
n
(cid:3) and decreasing in (cid:2) π
Now let us compare these values. In order to do so define the function e(x) = cos(x)+x sin(x)
for 0 ≤ x ≤ π. Basic calculations give us e
e(x) ≥ e(0) = 1
In order to compare the case with 2 zeros to the case with 1 zero, define for
n ≥ 2
2 , π(cid:3). Therefore for 0 ≤ x ≤ π
(x) = x cos(x) and therefore e
2 ,
2
(cid:48)
Ei+1 − Ei = − 1
0 ≤
Ei+1
Ei
≤
≤
2 (x(i)
2 )2 ≤ − Ei
1 − x(i)
(cid:0)1 − 1
(cid:0)1 − 1
(cid:1) Ei
(cid:1)i−t0 Et0
2k
2k
2k
proving that Ei converges to zero, i.e. all the non-zero values converge to m.
At each step of the above described process, according to fact b, the sum of
cosines can only decrease. Therefore from any given configuration we can get
as close as possible to a configuration in which all non-zero values are equal,
without increasing the sum of the cosines. Hence, the minimum value must be
reached in a configuration in which all non-zero values are equal.
Remebering that there can be at most only two zero values, the minimum value
of the sum of the cosines is the minimum of the following :
(cid:16) (n−2)π
(cid:17)
(cid:16) π
(cid:17) − π
n−1
n−1 sin
n−1
+ (n − 4)
(cid:17)
(cid:16) π
n−1
(cid:17)
f (n) = 1 + (n − 1) cos
= n − 3 − (n − 1) cos
(cid:48)
f
(n) = 1 − cos
= 1 − e
n−1
(cid:16) π
(cid:16) π
(cid:17)
n−1 ≤ π
n−1
π
2
Therefore f
(cid:48)
(n) ≤ 0 for n ≥ 4 because
f (2) = 0 ≤ 0
f (3) = 0 ≤ 0
2 ≤ 0
f (4) = − 1
f (n) ≤ f (4) ≤ 0 for n ≥ 4
23
Therefore f (n) ≤ 0 and the case with 2 zeros is never exclusively the optimal so-
lution (since the case with 1 zero always has a smaller or an equal value).
In order to compare the two remaining values, define for n ≥ 2
(cid:18) 2π
(cid:19)
n
(cid:18) π
(cid:19)
n − 1
+ 1
h(n) = n cos
− (n − 1) cos
The derivative according to n is
(cid:48)
h
n sin(cid:0) 2π
(cid:1) + 2π
(n) = cos(cid:0) 2π
(cid:17)
(cid:16) π
= e(cid:0) 2π
(cid:1) − e
n−1
n ≤ π
n−1 ≤ 2π
n
n
n
(cid:1) − cos
(cid:16) π
n−1
(cid:17) − π
n−1 sin
(cid:17)
(cid:16) π
n−1
π
For n ≥ 4, we have
(n) ≥ 0. One can check
that h(n) ≤ 0 for 2 ≤ n ≤ 6, and h(7) > 0, therefore h(n) > 0 for n ≥ 7. This
allows us to conclude that for 2 ≤ n ≤ 6 the minimal configuration is the one
corresponding to 1 zero, whereas for n ≥ 7, it is the one with no zeros.
(cid:48)
2 and therefore h
24
Appendix 2
Let us understand the geometric Lemma 1 by some illustrations. Consider a
convex polygon of n vertices. The sum of its interior angles is equal to (n− 2)π
and each angle is between 0 and π. Denote by Cn the bound given by Lemma
1.
First let us notice that sometimes the minimum value of Cn corresponds to a
set of interior angles that cannot be realized by a polygon in the plane (for
example we can show that in a convex polygon, if one of the interior angles is
0, then there are exactly two interior angles of 0 and (n − 2) interior angles
of π, but for example the configuration that realizes the minimum of C4 with
n = 4 does not correspond to such a realizable configuration, see figure 19 for
an illustration). However, we can get arbitrarily close to this value of Cn with
changing the angles with value of 0 by some and substracting in other angles
(that are not 0) the added values (see figures 17 and 20 for examples).
Let us explain the intuition behind Lemma 1. If ones tries to minimize the cosi-
nus of an angle, he would open it at maximum. But the constraint of forming a
convex polygon forces one to close the loop of the polygon. This is what limits
how much one can open the angles. It is intuitive that the more agents we have,
the more freedom we have to open the angles and use the numerous agents at
our disposal to close the loop.
The abstract case when there is an infinity of agents corresponds to a circle
where all angles can be open at maximum, i.e. with an angle of π.
What this Lemma also infers is that the configuration of the polygon that reaches
the minimum value is a regular polygon for n ≥ 7, and that for n ≤ 6, the min-
imum value of the sum of the cosinuses of the interior angles can be arbitrarily
closely approached by a polygon in a shape of a cone. Figures 16, 17, and 18
illustrate this for the case n = 3, while figures 19, 20, and 21 illustrate it for
the case n = 4. The cases n = 5 and n = 6 are similar, and for n = 7 and
above, the configuration of the minimum is a regular polygon leading the value
of Cn = n cos
(see examples in figures 22,23,24 ).
n
(cid:16) (n−2)π
(cid:17)
25
Figure 16: Theoretical configuration corresponding the the minimum value of
C3 = 1
Figure 17: Practical configuration for n = 3 provinding a value that can get
potentially arbitrarily close to the theoretical minimum, with the help of .
angles is 3 cos(cid:0) π
(cid:1) = 3
Figure 18: Regular polygon with n = 3, the sum of cosinuses of the interior
2 . This value is greater than the minimum possible value.
3
26
Figure 19: Theoretical configuration corresponding the the minimum value of
C4 = − 1
2
Figure 20: Practical configuration for n = 4 provinding a value that can get
potentially arbitrarily close to the theoretical minimum, with the help of .
Figure 21: Regular polygon with n = 4, the sum of cosinuses of the interior
(cid:1) = 0. This value is greater than the minimum possible value.
angles is 4 cos(cid:0) 2π
4
27
Figure 22: n = 7 is the first value of n for which the configuration of minimum
value of the sum of the cosinuses of the interior angles corresponds to a regular
polygon. In this case, this value is 7 cos(cid:0) 5π
(cid:1) ≈ −4.36.
7
Figure 23: Regular polygon and configuration of minimum value for n = 10, with
(cid:1) ≈ −8.09, each angle contributing a value of cos(cid:0) 8π
(cid:1) ≈ −0.81.
value 10 cos(cid:0) 8π
10
10
Figure 24: Regular polygon and configuration of minimum value for n = 30,
with value 30 cos(cid:0) 28π
cos(cid:0) 28π
(cid:1) ≈ −29.34, each angle contributing a value of
(cid:1) ≈ −0.98. This last value gets closer to −1 with higher values of
30
30
n as explained.
28
Appendix 3
The dynamics of the sytem described by (1) can be defined in the following
alternative way.
We define the two following functions :
1
1
2
0
: x > 0
: x = 0
: x < 0
h+(x) =
and h−(x) = h+(−x).
Let ez be a unitary vector orthogonal to the plane (in any direction). Then we
define :
ij = 0 if(cid:80)
j p±
where w±
ij = 0.
Finally, we define wij = w+
:
ij + w−
ij and the equations of movement are given by
And the vectors u+
i and u−
i of system (1) are given by:
sijk = (uij × uik) · ez
p±
ij =
h±(sijk)
(cid:89)
ij(cid:80)
p±
j p±
ij
k
w±
ij =
xi = v0
wijuij
u±
i =
w±
ijuij
(cid:88)
j
(cid:88)
j
29
Using the fact that u−
half-plane of u+
i
i and uij are, by the way they are defined, on the same
Appendix 4
i × uij) · (u+
(u+
i × u−
i ) ≥ 0
where × is the cross product of vectors. In the same way, uij and u+
same half-plane of u−
:
i
i are on the
i × uij) · (u−
(u−
i × u+
i ) ≥ 0
But, using the fact that these vectors are unit vectors,
i × uij) · (u+
(u+
i × uij) · (u−
(u−
i × u−
i × u+
i ) = uij · u−
i ) = uij · u+
i − (u+
i − (u−
i
i
· u−
· u+
i )(uij · u+
i )
i )(uij · u−
i )
Therefore
i × uij)· (u+
(u+
i × u−
i ) + (u−
i × uij)· (u−
i × u+
i ) = (1− u+
i · u−
i ·u−
Now u+
In any other case, 1 − u+
positive, we must have
i = 1 implies that u+
· u−
i
i ) (7)
i )uij · (u+
i +u−
i + u−
i )·uij = 2 > 0.
i = uij, and in this case (u+
i = u−
i > 0, and given that the left hand side of (7) is
i + u−
i ) · uij ≥ 0
(u+
30
|
1311.5108 | 1 | 1311 | 2013-11-20T15:44:26 | A Methodology to Engineer and Validate Dynamic Multi-level Multi-agent Based Simulations | [
"cs.MA"
] | This article proposes a methodology to model and simulate complex systems, based on IRM4MLS, a generic agent-based meta-model able to deal with multi-level systems. This methodology permits the engineering of dynamic multi-level agent-based models, to represent complex systems over several scales and domains of interest. Its goal is to simulate a phenomenon using dynamically the lightest representation to save computer resources without loss of information. This methodology is based on two mechanisms: (1) the activation or deactivation of agents representing different domain parts of the same phenomenon and (2) the aggregation or disaggregation of agents representing the same phenomenon at different scales. | cs.MA | cs |
A Methodology to Engineer and Validate Dynamic
Multi-level Multi-agent Based Simulations
Jean-Baptiste Soyez1,2
Gildas Morvan1,3
Daniel Dupont1,4
Rochdi Merzouki1,2
1Univ. Lille Nord de France
2 Polytech Lille, LAGIS F-59655 Villeneuve d'Ascq, France
[email protected]
[email protected]
3 UArtois, LGI2A, F-62400, Béthune, France
[email protected]
4 HEI, UC Lille, F-59046 Lille, France
[email protected]
Abstract
This article proposes a methodology to model and simulate complex systems, based
on IRM4MLS, a generic agent-based meta-model able to deal with multi-level systems.
This methodology permits the engineering of dynamic multi-level agent-based models,
to represent complex systems over several scales and domains of interest. Its goal is to
simulate a phenomenon using dynamically the lightest representation to save computer
resources without loss of information. This methodology is based on two mechanisms:
(1) the activation or deactivation of agents representing different domain parts of the same
phenomenon and (2) the aggregation or disaggregation of agents representing the same
phenomenon at different scales.
Introduction
1
Today, more and more engineering projects try to cope with complex systems. Complexity
can come from the number of represented entities, their structure, or the fact that informa-
tion is coming from difference sources and is incomplete.
Agent-based modeling is a very powerful and intuitive framework to study such systems.
However, the limitations of this approach lead to the development of multi-level agent-based
modeling (ML-ABM). It is defined by Morvan (2012, p. 1) as: "Integrating heterogenous ABMs,
representing complementary points of view, so called levels (of organization, observation, analysis,
granularity, ...
), of the same system. Integration means, of course, these ABMs interact but also
they can share entities such as environments and agents". From an engineering point of view,
ML-ABM reduces the complexity of the problem, so it becomes easier to implement.
In complex systems simulations, it is generally necessary to find a compromise between
1
the quality of simulations (amount of information or realism) and their resource consumption
(used CPU and memory).
A way to deal with this compromise is to use different models, more or less detailed or
treating different aspects of the same phenomenon and that are (dis)activated at run-time,
according to the context. This article proposes a methodology to engineer and validate such
simulations, based on IRM4MLS, a ML-ABM meta-model proposed by Morvan and Jolly
(2012); Morvan et al. (2011).
The next section presents recent works in the domain of multi-resolution or multi-level
modeling. Section 3 introduces a generic agent-based meta-model IRM4MLS. Then, section
4 shows some possibilities offered by IR4MLS to model complex systems in which different
domains interact. Section 5 explains how to construct models with dynamic change of level
of detail (LOD), i.e., switching scales or domains of interest. Section 6 gives a tool to measure
the quality of multi-level models endowed with dynamic changes of resolution. Finally, we
expose the conclusions and perspectives of our work in section 7.
2 Related Works
In this section, multi-modeling approaches, dealing with models at different scales in an engi-
neering context, are presented.
Multi-Resolution modeling (Davis and Hillestad, 1993) is the joint execution of different
models of the same phenomenon within the same simulation or across several heterogeneous
systems. It can inspire our approach if different models can be considered as different lev-
els. Consistency represents the amount of essential information lost when crossing different
models and it is an adapted tool to test the quality of this approach.
The High Level Architecture (Simulation Interoperability Standards Comittee (SISC),
2000) (HLA) is a general purpose architecture for distributed computer simulation systems.
Using HLA, computer simulations can interact (communicate data and synchronize actions)
with other computer simulations regardless of the computing platforms. The interaction be-
tween simulations is managed by a Run-Time Infrastructure (RTI). Scerri et al. (2010) devel-
oped HLA-Repast, a unified agent-based simulation framework, in which concurrent modules
with their own temporality can use global variables through centralized services.
Holonic multi-agent systems (HMAS) can be viewed as a specific case of multi-level
multi-agent systems (MAS). The most obvious aspect being the hierarchical organization of
levels. However, from a methodological perspective, differences remain. Most of holonic
meta-models focus on organizational and methodological aspects while ML-ABM is process-
oriented. HMAS meta-models have been proposed in various domains, e.g., ASPECTS (Gaud
et al., 2008) or PROSA(Van Brussel et al., 1998). Even if ML-ABM and HMAS structures are
close, the latter is too constrained for the target application of this work.
Navarro et al. (2011) present a framework to dynamically change the level of detail in
agent-based simulation. That is to say, represent only what is needed during simulation, to
save CPU resources and keep the consistency of the simulation. But this framework is lim-
ited because levels form a merged hierarchy, without the possibility of having two levels at
the same scale and communication between levels is not explicitly defined.
The possibility for agents to exist in several levels simultaneously is a way to make simu-
lations benefit of a higher power of representation. It permits to 1) simulate nested entities, 2)
create agents with concurrent psychological trends and 3) model complex systems implying
various domains.
It is possible to model the coexistence of nested entities at different scales. Agents present
in different levels can be seen as "gate" between these levels. For example, Picault and Mathieu
(2011), give the example of cell membrane elements that are the "gates" between the inside and
the outside of the cell, i.e., between two scales and exposed to the influences of two different
environments.
2
0..n
0..n
environment
agent
0..n
1
1..n
0..n
level
Figure 1: Central Concepts of IRM4MLS (cardinalities are specifed the UML way)
An agent existing at different levels simultaneously can fulfill a global objective while
following its own goals. In Stratulat et al. (2009), authors decompose, with the MASQ model,
agents into two bodies: a physical one (individual) and a social one (collective) to do this.
Levels can have different temporal dynamics, independently of other levels. It allows to
optimize the execution of complex agents by (dis)activating their bodies at run-time to use
the lightest representation (Soyez et al., 2011).
Readers interested in a more comprehensive presentation of ML-ABM should refer to Gil-
Quijano et al. (2012); Morvan (2012).
IRM4MLS
3
IRM4MLS is a ML-ABM meta-model proposed by Morvan and Jolly (2012); Morvan et al.
(2011). It relies on the influence/reaction model (Ferber and Müller, 1996) and its extension
to temporal systems, IRM4S (Michel, 2007). An interesting aspect of IRM4MLS is that any
valid instance can be simulated by a generic algorithm. The main aspects of this meta-model
are presented in this section.
A IRM4MLS model is characterized by a set of levels, L, and relations between levels. Two
types of relations are considered: influence (agents in a level l are able to produce influences in
a level l(cid:48) (cid:54)= l) and perception (agents in a level l are able to perceive the state of a level l(cid:48) (cid:54)= l).
These relations are respectively formalized by two digraphs, 〈L, EI
〉 where EI and
EP are sets of edges, i.e., ordered pairs of elements of L. The dynamic set of agents at time t
is denoted A(t). ∀l ∈ L, the set of agents in l at t is Al
(t) ⊆ A(t). An agent acts in a level if
a subset of its external state belongs the state of this level. An agent can act in multiple levels
at the same time. Environment is also a top-class abstraction. It can be viewed as a tropistic
agent with no internal state that produces "natural" influences in the level (Fig. 2).
〉 and 〈L, EP
The scheduling of each level is independent: models with different temporalities can be
simulated without temporal bias. On an other hand, only the relevant processes are permitted
to execute during a time-step. A major application of IRM4MLS is to allow microscopic
agents (members) to aggregate and form-up lower granularity agents (organizations). It can
be useful to create multiple levels at the same scale to represent different domain parts of the
same phenomenon. In the following, we consider that two levels are at the same scale if they
have the same spatial and temporal extents.
4 Multi-level, Single Scale Simulation
In this section we give a framework to improve the integration of agents located in different
levels (not necessary at different scales) simultaneously. Then, we show how to take advantage
of this concept to simulate complex systems while optimizing the use of computer resources.
"One Mind, Several Bodies"
4.1
In our approach, inspired by Picault and Mathieu (2011), agents can be present in several lev-
els at the same time. We propose to decompose agents in a "central" unsituated part and a set
of n "peripheral" parts, each situated in a given level. Thus, we call spiritAgent the unsituated
3
conceptualAgent
environment
1
spiritAgent
1
1
1
1..n
1..n
0..n
bodyAgent
1
1
0..n
0..n
1
level
Figure 2: Class diagram of central Concepts of IRM4MLS with separation of situated-or-not
agent parts
part of the agent which contains its internal state, its decision processes and that cannot act in
a level. BodyAgents in levels l ∈ L are the situated part of the agent which contains its external
state and the possible actions in its level, like perception of the environment.
ConceptualAgents stand for common agents in classical simulation. SpiritAgents only
contain the internal state of the agent and its decision module. BodyAgents have to be situated
in one and only one level. They contain the external state of the agent specific to a level,
and an action module that indicates: 1) what are the available actions at a given time and 2)
what are their results in term of produced influences. The perception process must be in this
action module. Levels contain inactive objects that support agent actions. The only use of
Environments is to produce the natural influences of the level (like the gravity force in a
physical level).
To obtain valid simulations with such models, a spiritAgent has to be able to access the
external state of its conceptualAgent contained in its bodyAgent when it is active (during the
execution of its level). Thus, we can consider the several steps of the life cycle of agents. Each
time a bodyAgent is active, 1) it perceives its level (and others perceptible from this one), 2) it
sends a part of these perceptions and the possible actions to the spiritAgent, 3) the spiritAgent
modifies its internal state and 4) indicates the most appropriate action to be accomplished by
the bodyAgent, 5) the bodyAgent accomplishes this action which produces influences in di-
rection of its levels and others possibly influenced by this one.
4.2 Level Temporality
In this section we explain the possibility to attribute a different temporality to each level and
how to adapt it to our models. IRM4MLS uses the framework of timed event systems (Zeigler
et al., 2000). The scheduling is distributed between levels with no constraint on the scheduling
mode (step wise or discrete events). This approach seems more adapted to our problems than
the agent one (Weyns and Holvoet, 2003) or the system one (Michel et al., 2003).
Our goal to give to agents the longest possible life cycle which stay coherent with the rest
of the simulation. This is done to minimize the computer resources allocated to the agents
updating process. Morvan et al. (2011) propose an algorithm adapted to IRM4MLS which
manage the coupling between levels with different temporal dynamics. This is made to apply
easily the proposed methods above.
The Figure 3 illustrates different constraints which fix the life cycle of agents in a same
level. The frequency of a level is expressed in Hertz, indicating how many times a second,
it is necessary to execute the updating process of the dynamic state of a level. Let imagine
that all functions of an agent possess a minimal frequency beyond which their simulation is
not realistic anymore. If a level permits to its agents to dispose of functions with different
frequencies, it adopts the higher one, to keep a correct simulation of the functions with this
frequency. Therefore, in the example of Fig. 3, the frequency of the level N1 is equal to 60H z
4
Figure 3: Example of Multi-Level ML-ABM with different temporalities
because the diagnostic function of the modeled vehicles needs this minimal frequency.
The other constraint comes from the interactions between levels. If we continue with
the previous example, let say that N2 level needs a minimal frequency equal to 20H z, this
frequency could be allocated to N2. However if the N1 level is influenced by N2 and has to
calculate the reaction induced by these influences at a frequency higher than 20H z (logically
less or equal to 60 Hz), it can be necessary to allocate a higher frequency to N2. Thus, it
is necessary to dynamically modify the frequency of a level N and adapt it to the chang-
ing needs of the simulation and return it back to its minimal frequency, defined during the
implementation phase.
5 Dynamic Change of Level of Detail (LOD)
In this section we give a methodology to apply dynamic changes of LOD in a simulation.
First we present the hierarchical level graph, which indicates the links between levels and
the dis/aggregation functions attached to change the LOD of simulated entities. Finally, we
specify when and in which conditions dis/aggregation functions can be applied. In the next
part, we give a method to test the quality of the dis/aggregation mechanisms exposed here by
measuring the whole consistency of simulations.
5.1 Hierachical Level Graph
〉 where EH are sets
Relations between levels are respectively formalized by a digraph, 〈L, EH
of edges, i.e., ordered pairs of elements of L. This digraph whose vertices are levels, is called
the hierarchical level graph. This graph indicates how levels are nested and which couple of
levels treats different domain of interest of the same phenomenon.
) edge
A simple edge represents an inclusion link between two levels. For example, an (l1, l2
signifies that l2 has higher spatial or temporal extents than l1. Then the bodyAgents situated
in l1 can be aggregated and the resulting aggregate can be instantiated in l2. We note that
≺ l2.
l1
A pair of symmetric edges means there is a complementarity link between two levels. For
) edges mean that l1 and l3 are at the same scale. Thus a spiritA-
example, the (l1, l3
gent can control several bodyAgents simultaneously present and activated in l1 and l3. We
≡ l3
note that l1
A loop on a vertex indicates levels whose bodyAgents can adopt a similar behaviour. For
example, a (l1, l1
) edge means that the spiritAgent, of some bodyAgents situated in l1, can be
aggregated to form a single spiritAgent which will control these unchanged bodyAgents in l1.
These bodyAgents will have the same behaviour when confronted to similar situations, but
will keep their autonomy.
) and (l3, l1
The following rules have to be applied if we want to obtain a coherent model.
Rule 1 Inclusion and Complementarity links are transitive.
∧ l2
≺ l2
l1
Rule 2 A level cannot be included in itself by a direct or transitive way. This rule is translated
by the fact that if we delete all pairs of symmetric edges, there should not be directed cycles in the
≺ l3, l1
≡ l3.
→ l1
≡ l2
∧ l2
≺ l3
→ l1
≡ l3
5
FAg1
l1
FAg2, FAg3
l2
l3
Figure 4: An example of Hierarchical Level Graph
≺ l1
≡ l2
(cid:54)≺ l2.
→ l1
(cid:54)≡ l2, l1
∈ L∧ l1
hierarchical level graph.
(cid:64)l1
Rule 3 Two distinct levels cannot share simultaneously an inclusion and a complementarity link,
directly or by a transitive way.
≺ l2
→ l1
l1
Each edge which is not part of a symmetric pair of edges is labelled with one or more
aggregation function names. An aggregation function name can be placed on several edges.
The (l1, l1
) edge, labelled FAg1, indicates that the spiritAgents controlling some bodyA-
gents present in l1 can aggregate themselves to form a single spiritAgent controlling all these
bodyAgents, through the FAg1 function. The (l1, l2
) edge, labelled FAg2, FAg3, means that
the spiritAgents controlling some bodyAgents present in l1 can aggregate themselves to form
a single spriritAgent controlling a single aggregated bodyAgent situated in l2, through the
FAg2 or FAg3 function. These two functions concerns different combination of bodies. And
the symmetric pair of edges between l1 and l3, with no label, represents the fact that some
spiritAgents can control simultaneously bodyAgents situated in these two levels.
5.2 Dis/Aggregation Functions
5.2.1 Content
As shown before, there are two types of aggregation. The first one deals with the aggregation
of spiritAgents and the second one with the aggregation of spiritAgents and their associated
bodyAgents. The first type of aggregation is used to represent a set of agents with the same
internal state, that leads to agents which act similarly in the same situation but which can be
place in several situations. The aggregation of several bodyAgents without the aggregation of
their spiritAgent is impossible because a body cannot be controlled simultaneously by several
concurrent spirits.
Once the hierarchical level graph is fixed, the modeler has to indicate every class of bodyA-
gent that he decides to place in levels and which class of spiritAgent control these bodyAgents.
For each aggregation function the modeler has to precise how many agents have to be merged,
the class of aggregated and aggregate agents and how to generate internal and/or external state
of the aggregate agent.
In this article we don't give any indication to set the decision module or the action mod-
ule of aggregate agents or not but we focus on how to aggreagte internal and external states
of agents, respectively contained in spiritAgents and bodyAgents. Each aggregation function
can be divided into several subfunctions. These subfunctions can be of two types. First type:
a subfunction takes the same variable in each agents concerned (spiritAgents or bodyAgents)
and aggregates them to obtain a single value to place it in the aggregated agent state. For ex-
ample, a agent representing a platoon of vehicles has the mean position of all vehicle agents.
Second type: a subfunction similar to the first does an aggregation on several variables con-
tained in the agents to aggregated but produces only one value. This can be illustrated by the
platoon agent described above. It only possesses one variable in its internal state called "pri-
ority" whose value is generated with the compound of the "stamina" and "speed" variables
of each vehicle agents in the platoon. Some variables of the agents to be aggregated can be
6
ignored to construct an aggregate.
5.2.2 Notation
An aggregation function consists in creating a composite agent from several agents. Here
is the general form of an aggregation function FAg using for argument n conceptualAgent
class, c t a (class to aggregate), endowed of an interval, [mini, maxi
], indicating how many
instances of these classes are necessary to accomplish this aggregation. For each conceptualA-
gent class it is precised if the aggregation implies bodyAgents in addition of spiritAgent with
the indication of a level li where the bodyAgents are situated. The class of the agent produced
by the aggregation, AAC (Aggregate Agent Class), is the output of FAg with its level l if the
= ∅.
aggregation concerns bodyAgents. If the aggregation only concerns spiritAgents l = li
((cid:89)
〈[mini; maxi
FAg
i∈n
]c t ai, li
〉) = (AAC , l)
(1)
For example, let consider the FAg2 function described in the hierarchical graph below. Let
FAg2 aggregates one bodyAgent of class Leader and at least 4 to 9 bodyAgents of class Follower
all situated in l1 level and their linked spiritAgents to create a bodyAgent of class Platoon
situated in l2 level and its linked spiritAgent. Then:
(〈[1;1], Lead e r, l1
〉,〈[4;9], F ol l owe r, l1
FAg2
(2)
Aggregation subfunctions have quite the same notation than aggregation functions. It is
not necessary to precise the number of concerned agents anymore. But variables, in con-
cerned agents, which will be mixed together have to be known. For example the subfunction
described in the previous subsection can be noted like this:
〉) = (P lat oon, l2
)
(Lead e r.s t amina, Lead e r.s peed, l1
)
fAg2,1
(F ol l owe r.s t amina, F ol l owe r.s peed, l1
= (C r owd.p r ior i t y, l2
),
)
(3)
5.2.3 Disaggregation and Memorization Functions
Each aggregation function possesses its disaggregation function and eventually a memoriza-
tion function. A disaggregation function permits to create several instances of the aggregated
agents from the aggregate agent. A memorization function can be used to store some infor-
mation. Each memorization function is associated to a disaggregation one to generate several
agents representing the initial aggregated agents taking into account the last state of the aggre-
gated agents and the system evolution since the aggregation. Here, nbi indicates the number
of agents of each class involved in the aggregation.
FDi sa g
(AAC , l, FM e mor i zat ion
〈nbi, c t ai, li
i∈n
〈nbi, c t ai, li
〉)
i∈n
(4)
((cid:81)
〉)) = ((cid:81)
These two functions are divided in subfunctions in a similar way than the aggregation
function. Let take a platoon endowed of the two position variables, X and Y , representing
the position variable x and y of all the vehicles constituting it. The memorization function
store positions of all these vehicles. Memorization is not active during the execution of the
platoon agent. After the platoon agent have moved in (X (cid:48),Y (cid:48)) position, it can be disaggregated
by recreating the vehicles agents, calculating the value of their x and y variables with X (cid:48) and
Y (cid:48) and applying the memorized repartition.
7
High-resolution
inputs
High-resolution
model : M
High-resolution
outputs
Aggregate
function
Aggregate
function
Low-resolution
inputs
Low-resolution
model : M'
Low-resolution
outputs
Aggregated
high-resolution
outputs
The models are
consistent if these
are approximately
equal
Figure 5: Weak consistency, according to Davis and Hillestad (1993)
5.3 Dis/Aggregation Tests
Navarro et al. (2011) explains how to decide when agents should be aggregated. He uses an
affinity function which measure the similarity of internal and external states of agents. When
the similarity is more important than a given threshold he links the two agents. Linked agents
with the higher similarity value are aggregated together.
We can use a similar mechanism to decide when to use an aggregation function, but in
our case we need one utility function Af f by aggregation function FAg . If there are several
aggregation functions which concern the same spiritAgents or bodyAgents in the same levels,
it is necessary to decide when apply one instead of another. There are three possibilities. 1)
The choice of FAg is done after measuring the affinity of agent groups with all Af f and the
aggregate are instantiated each time, choosing the group with the higher affinity, until there
is no group. 2) It is also possible to impose an order to test different FAg . All groups with a
high affinity for one FAg are aggregated, then the next FAg is tested until there is no more FAg .
3) The choice of FAg can be done by a mix of the two previous methods. An partial order is
defined on FAg s space. And if there is no precedence link between different FAg , we apply the
first method to aggregate agents considering that the model FAg only contains these FAg after
that we continue following the established order.
6 Measuring Consistency
Davis and Hillestad (1993) uses the notion of consistency to measure the quality of simula-
tions dealing with models of different resolution. "Consistency between a high-resolution
model M and a low-resolution model M' is the comparison between the projected state of an
aggregate of high-resolution entities which evolved in M, and the projected state of the same
aggregate initially controlled by M' ".
It is more intuitive to base the comparison on the evolution of the more detailed model
instead of the aggregate model because it has a higher resolution and possesses more significant
information.
Before modeling the system, it is necessary to locate the significant simulation elements.
These elements can be in the internal (spiritAgent) or external (bodyAgent) states of agents
or in their environment. Once these elements are identified, several simulations are launched
8
detailled state
at time t
evolution of M
detailled state
at time t+n
Global system
is consistent if
this differents states
are equivalent
FAg M e m
FAg
FDi sa g
aggregated state
at time t
evolution of M'
aggregated state
at time t+n
Figure 6: Strong consistency, according to Davis and Hillestad (1993)
with the same parameters (initial state and execution time) using only the most detailed levels,
carrying the more information but the most expensive one. At the end of the simulations
execution a mean state of the identified elements is recorded. The same process is done with
the model using dynamic change of LOD. Then the dissimilarity is measure between these
two recording to calculate the consistency.
7 Conclusion and Perspectives
This article introduces a methodology and theoretical tools to engineer and validate multi-
level agent based simulations with dynamic change of LOD.
It is applied in the european project InTrade1. This project deals with logistic in european
container ports endowed with Autonomous Intelligent Vehicles (AIV). Partners involved in
this project work at different scales and use simulation tools adapted to it (SCANeRstudio or
Flexsim Container Terminal 2). The agent-based platform MadKit3 is used to make models
coexist in a single simulation. Results are visualized with SCANeRstudio or Flexsim CT.
An interesting perspective of this work would be to find better ways (cheaper or more
realistic) to decide when simulated entities should be (dis)aggregated.
It is closely related
to the emergence detection and reification problem (David and Courdier, 2009). Two main
approaches have been proposed to tackle this issue: a statistical one (e.g., (Caillou and Gil-
Quijano, 2012; Caillou et al., 2013; Moncion et al., 2010; Vo et al., 2012)) and a symbolic
one (Chen et al., 2010, 2009). It would be interesting to integrate them.
Another perspective is the integration of organizational concepts, such as Systems of Sys-
tems (SoS), in our methodology. It would allow to explicitly represent system or group level
properties such as goals or missions.
References
Caillou, P. and Gil-Quijano, J. (2012). Simanalyzer : Automated description of groups dynamics in agent-based
simulations. In Proc. of 11th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS 2012).
Caillou, P., Gil-Quijano, J., and Zhou, X. (2013). Automated observation of multi-agent based simulations: a
statistical analysis approach. to appear in Studia Informatica Universalis.
Chen, C., Clack, C., and Nagl, S. (2010). Identifying multi-level emergent behaviors in agent-directed simulations
using complex event type specifications. Simulation, 86(1):41 -- 51.
1http://www.intrade-nwe.eu/
2http://www.intrade-nwe.eu/ or www.flexsim.com/
3http://www.madkit.org/
9
Chen, C., Nagl, S., and Clack, C. (2009). A formalism for multi-level emergent behaviours in designed component-
based systems and agent-based simulations. In Aziz-Alaoui, M. and Bertelle, C., editors, From System Complexity
to Emergent Properties, volume 12 of Understanding Complex Systems, pages 101 -- 114. Springer.
David, D. and Courdier, R. (2009). See emergence as a metaknowledge. a way to reify emergent phenomena in
multiagent simulations? In Proceedings of ICAART'09, pages 564 -- 569, Porto, Portugal.
Davis, P. and Hillestad, R. (1993). Families of model that cross levels of resolution : Issues for design, calibration and
management. In 25th Winter Simulation Conference (WSC'93).
Ferber, J. and Müller, J.-P. (1996). Influences and reaction: a model of situated multiagent systems. In 2nd Interna-
tional Conference on Multi-agent systems (ICMAS-96), pages 72 -- 79.
Gaud, N., Galland, S., Gechter, F., Hilaire, V., and Koukam, A. (2008). Holonic multilevel simulation of complex
systems : Application to real-time pedestrians simulation in virtual urban environment. Simulation Modelling
Practice and Theory, 16:1659 -- 1676.
Gil-Quijano, J., Louail, T., and Hutzler, G. (2012). From biological to urban cells: Lessons from three multilevel
agent-based models. In Desai, N., Liu, A., and Winikoff, M., editors, Principles and Practice of Multi-Agent Systems,
volume 7057 of Lecture Notes in Computer Science, pages 620 -- 635. Springer.
Michel, F. (2007). The irm4s model: the influence/reaction principle for multiagent based simulation. In AAMAS
'07: Proceedings of the 6th international joint conference on Autonomous agents and multiagent systems, pages 1 -- 3,
New York, NY, USA. ACM.
Michel, F., Gouaïch, A., and Ferber, J. (2003). Weak interaction and strong interaction in agent based simulations.
Lecture Notes in Computer Science, 2927:43 -- 56.
Moncion, T., Amar, P., and Hutzler, G. (2010). Automatic characterization of emergent phenomena in complex
systems. Journal of Biological Physics and Chemistry, 10:16 -- 23.
Morvan, G. (2012). Multi-level agent-based modeling - bibliography. CoRR, abs/1205.0561.
Morvan, G. and Jolly, D. (2012). Multi-level agent-based modeling with the Influence Reaction principle. CoRR,
abs/1204.0634.
Morvan, G., Veremme, A., and Dupont, D. (2011). IRM4MLS: the influence reaction model for multi-level simu-
lation. In Bosse, T., Geller, A., and Jonker, C., editors, Multi-Agent-Based Simulation XI, volume 6532 of Lecture
Notes in Artificial Intelligence, pages 16 -- 27. Springer.
Navarro, L., Flacher, F., and Corruble, V. (2011). Dynamic level of detail for large scale agent-based urban simu-
lations. In Tumer, Yolum, Sonenberg, and Stone, editors, 10th Int. Conf on Autonomous Agents and Multiagent
Systems (AAMAS 2011), pages 701 -- 708.
Picault, S. and Mathieu, P. (2011). An interaction-oriented model for multi-scale simulation. In the 22nd International
Joint Conference on Artificial Intelligence (IJCAI'11).
Scerri, D., Hickmott, S., Drogoul, A., and Padgham, L. (2010). An architecture for distributed simulation with
agent-based models. In van der Hoek, Kaminka, Lespérance, Luck, and Sen, editors, Proc. of 9th Int. Conf on
Autonomous Agents and Multiagent Systems (AAMAS 2010), pages 541 -- 548, Toronto, Canada.
Simulation Interoperability Standards Comittee (SISC) (2000).
IEEE Standard for Modeling and Simulation (M&S)
High Level Architecture (HLA) - Framework and Rules. IEEE Computer Society.
Soyez, J.-B., Morvan, G., Merzouki, R., Dupont, D., and Kubiak, P. (2011). Multi-agent multi-level modeling -- a
methodology to simulate complex systems. In Proceedings of the 23rd European Modeling & Simulation Symposium.
Stratulat, T., Ferber, J., and Tranier, J. (2009). Masq : toward an integral approach to interaction. In Proceedings of
the 8t h conference on Autonomous Agents and Multiagent Systems (AAMAS 2009), pages 813 -- 820.
Van Brussel, H., Wyns, J., Valckenaers, P., Bongaerts, L., and Peeters, P. (1998). Reference architecture for holonic
manufacturing systems: Prosa. Computers in Industry, 37(3):255 -- 274.
Vo, D.-A., Drogoul, A., Zucker, J.-D., and Ho, T.-V. (2012). A modelling language to represent and specify emerging
structures in agent-based model. In Desai, N., Liu, A., and Winikoff, M., editors, Principles and Practice of Multi-
Agent Systems, volume 7057 of Lecture Notes in Computer Science, pages 212 -- 227. Springer.
Weyns, D. and Holvoet, T. (2003). Model for simultaneous in situated multi-agent systems. Lectures Notes in Artificial
Intelligence, 2831:105 -- 118.
Zeigler, B., , Kim, T., and Praehofer, H. (2000). Theory of Modeling and Simulation. Academic Press, 2nd edition.
10
|
1210.4778 | 2 | 1210 | 2013-02-22T13:16:30 | Average Consensus in the Presence of Delays and Dynamically Changing Directed Graph Topologies | [
"cs.MA",
"cs.DC"
] | Classical approaches for asymptotic convergence to the global average in a distributed fashion typically assume timely and reliable exchange of information between neighboring components of a given multi-component system. These assumptions are not necessarily valid in practical settings due to varying delays that might affect transmissions at different times, as well as possible changes in the underlying interconnection topology (e.g., due to component mobility). In this work, we propose protocols to overcome these limitations. We first consider a fixed interconnection topology (captured by a - possibly directed - graph) and propose a discrete-time protocol that can reach asymptotic average consensus in a distributed fashion, despite the presence of arbitrary (but bounded) delays in the communication links. The protocol requires that each component has knowledge of the number of its outgoing links (i.e., the number of components to which it sends information). We subsequently extend the protocol to also handle changes in the underlying interconnection topology and describe a variety of rather loose conditions under which the modified protocol allows the components to reach asymptotic average consensus. The proposed algorithms are illustrated via examples. | cs.MA | cs |
Average Consensus in the Presence of Delays and
Dynamically Changing Directed Graph Topologies
Christoforos N. Hadjicostis, Senior Member, IEEE and
Themistoklis Charalambous, Member, IEEE
Abstract
Classical approaches for asymptotic convergence to the global average in a distributed fashion
typically assume timely and reliable exchange of information between neighboring components of a
given multi-component system. These assumptions are not necessarily valid in practical settings due
to varying delays that might affect transmissions at different times, as well as possible changes in
the underlying interconnection topology (e.g., due to component mobility). In this work, we propose
protocols to overcome these limitations. We first consider a fixed interconnection topology (captured
by a possibly directed graph) and propose a discrete-time protocol that can reach asymptotic aver-
age consensus in a distributed fashion, despite the presence of arbitrary (but bounded) delays in the
communication links. The protocol requires that each component has knowledge of the number of its
out-neighbors (i.e., the number of components to which it sends information). We subsequently extend
the protocol to also handle changes in the underlying interconnection topology and describe a variety of
rather loose conditions under which the modified protocol allows the components to reach asymptotic
average consensus. The proposed algorithms are illustrated via examples.
Keywords: Average consensus, digraphs, bounded delays, changing interconnection topology, ratio
consensus, weak convergence.
Preliminary results of the work in this paper have been presented in [1] which only included the discussions on bounded
delay and did not provide all details for the proofs.
Christoforos N. Hadjicostis is with the Department of Electrical and Computer Engineering at the University of Cyprus,
Nicosia, Cyprus and also with the Department of Electrical and Computer Engineering at the University of Illinois, Urbana-
Champaign, IL, USA. E-mail: [email protected].
Themistoklis Charalambous was formerly with the Department of Electrical and Computer Engineering at the University of
Cyprus, Nicosia, Cyprus. He is currently with the Automatic Control Lab, Electrical Engineering Department and ACCESS
Linnaeus Center, Royal Institute of Technology (KTH), Stockholm, Sweden. Corresponding author's address: Osquldas vag 10,
100-44 Stockholm, Sweden. E-mail: [email protected].
I. INTRODUCTION
Centralized approaches in multi-component systems require the collection of measurements
or other information to a central location (at possibly high communication and computational
cost), the computation of quantities of interest at this central location, and then the dissemination
of these quantities to (a subset of) the components. This approach is often inefficient or even
unrealizable (e.g., in ad-hoc networks that do not posses explicit routing mechanisms). Coopera-
tive distributed control/coordination algorithms and protocols have therefore received tremendous
attention, especially during the last decade. Several diverse research communities (e.g., biology,
physics, control, communication, and computer science) have made important contributions that
have resulted in many recent advances in so called consensus-based approaches (see, for example,
[2]) and in distributed computation of functions of geographically dispersed data, also known as
in-network computation (see, for example, [3] and references therein).
A distributed system or network consists of a set of components (nodes) that can share
information via connection links (edges), forming a directed interconnection topology (digraph).
In general, the objective of a consensus problem is to have all agents agree upon a certain (a
priori unknown) quantity of interest that is typically a function of some values that the nodes
initially posses. When the agents (asymptotically) reach agreement to the same value, we say
that the distributed system (asymptotically) reaches consensus. The problem of convergence of
discrete-time consensus algorithms was initially targeted by Tsitsikis et al. [4] and subsequently
by many other researchers (see, for example, [5] -- [13]). Typical applications include network
coordination problems involving self-organization, formation of patterns, parallel processing,
and distributed optimization, such as motion of mobile agents (e.g., coordination of unmanned
air vehicles, autonomous underwater vehicles, or satellites) and averaging of measurements in
wireless sensor networks.
Common challenges in consensus problems include the handling of node failures (e.g., due
to the draining of batteries in wireless sensor networks), transmission delays on the transfer of
data between agents, packet losses in wireless communication networks, and inaccurate sensor
measurements. As a result, agreement problems have been studied in networks of dynamical
agents, possibly with directed information flow, under delays and/or changing topologies. It is
worth pointing out that convergence of consensus algorithms can usually be established under
relatively weak requirements and that consensus protocols have been relatively successful in
addressing disturbances due to delays (e.g., [8], [9], [14]), packet drops (e.g., [15], [16]), and
changing interconnections (e.g., [6], [7]), or a combination of them (e.g., [5], [17], [18]). What
is different in this paper is that we devise a protocol that is able to overcome such limitations
while reaching consensus to the exact average of the values that the nodes initially posses. More
specifically, by utilizing the suggested protocol, each agent reaches the exact average of the initial
values of all the agents, even in the presence of (bounded) transmission delays and changes in
the interconnection topology. This is in sharp contrast with the above mentioned works in which
consensus is reached to a value that is typically a function of the disturbances involved (and
thus cannot be guaranteed a priori).
The average consensus problem studied in this paper aims to have the agents reach agreement
to the average of their initial values (see, for example, [19], [20]). It has been shown in [5] that,
under a fixed interconnection topology, average consensus can be achieved by performing a linear
iteration in a distributed fashion if the interconnection topology is both strongly connected and
balanced, while convex optimization [21] -- [23] requires update matrices that are doubly stochas-
tic. Even though various approaches have been proposed for forming balanced matrices (e.g.,
[24], [25]) and primitive doubly stochastic matrices (e.g., [11], [26]), which can subsequently
be used for reaching average consensus, most existing schemes are not applicable in digraphs
and/or fail in the presence of delays and changing interconnection topology. In particular, among
the limited existing algorithms that guarantee convergence to the exact average in a digraph (e.g.,
[27] -- [29]), few of them have addressed delays and topology changes, and it is unclear how/if
these techniques can be modified to overcome such disturbances.
The methodology developed in this paper is based on an algorithm suggested in [27] that
solves the average consensus problem in a static digraph using a linear iteration strategy in
which each node vj distributively sets the weights on its self-link and outgoing-links to be
(where D+
1+D+
is the out-degree of node vj). More generally, the set of weights needs to adhere to
j
1
j
the graph structure (i.e., be positive on each edge -- including self-edges -- and zero otherwise),
but is otherwise unrestricted as long as it forms a primitive column stochastic matrix P . More
generally, the set of weights needs to adhere to the graph structure (i.e., be positive on each edge
-- including self-edges -- and zero otherwise), but is otherwise unrestricted as long as it forms
a primitive column stochastic matrix P . Using the weights in matrix P , average consensus is
reached in [27] via ratio consensus, i.e., two linear iterations (with appropriately chosen initial
conditions) that run simultaneously so that the average can be obtained at each node by taking
the ratio of the two values it maintains for each of the two iterations. An equivalent approach
for gossiping algorithms was also proposed in [30], which is a generalization of the gossiping
algorithm proposed in [31]. The idea of ratio consensus can be traced back much earlier (see
the discussion on weak convergence at the "Bibliography and Discussion to §§3.1-3.2", pp. 98,
in [32]).
In this paper, we investigate the problem of discrete-time average consensus in a multi-
component system under a directed interconnection topology in the presence of bounded delays in
the communication links and changing interconnections (with communication links being added
or removed, as in a mobile network setting). First, we consider a fixed topology and we devise
a protocol, based on ratio consensus, where each node updates its information state by linearly
combining the available (possibly delayed) information received by its neighbors using constant
positive weights. We establish that, unlike other consensus approaches, this robustified version
of ratio consensus, henceforth called robustified ratio consensus, converges to the exact average
of the nodes initial values, despite the presence of arbitrary but bounded time-delays. Then,
we allow the communication links to change (at the same time we also allow communication
delays in the network) and enhance the proposed robustified ratio consensus algorithm so that
the algorithm is immune to arbitrary changing interconnection topology and delays. We show
that it is possible to asymptotically reach consensus to the exact average in a distributed fashion
for a network with changing communication links and delays, as long as the delays are bounded
and the unions of digraphs over consecutive time intervals form strongly connected digraphs
infinitely often.
The remainder of the paper is organised as follows. In Section II, the notation used throughout
the paper is provided, along with some background on graph theory that is needed for our
subsequent development. This section also outlines the algorithm proposed in this paper. In
Section III, we describe our model for communication link delays and our model for changing
interconnection topology in the multi-agent system. In Section IV we consider a fixed topology
and study the behavior of our algorithm in the presence of delays. In Section V we consider a
fixed set of nodes and allow changes in the communication links among them in order to study
the behavior of our algorithm in the presence of both interconnection topology changes and
delays. Finally, Section VI summarizes the results of the paper and draws directions for future
research.
A. Notation
II. NOTATION AND PRELIMINARIES
The sets of real, integer and natural numbers are denoted by R, Z and N, respectively; their
nonnegative counterparts are denoted by the subscript + (e.g., R+). Vectors are denoted by small
letters whereas matrices are denoted by capital letters. The transpose of matrix A is denoted by
AT . By 1 we denote the all-ones vector and by I we denote the identity matrix (of appropriate
dimensions). A matrix whose elements are nonnegative, called nonnegative matrix, is denoted by
A ≥ 0, and a matrix whose elements are positive, called positive matrix, is denoted by A > 0.
In multi-component systems with fixed communication links (edges), the exchange of infor-
mation between components (nodes) can be conveniently captured by a digraph G(V,E) of order
n (n ≥ 2), where V = {v1, v2, . . . , vn} is the set of nodes and E ⊆ V × V is the set of edges.
A directed edge from node vi to node vj is denoted by εji (cid:44) (vj, vi) ∈ E and represents a
communication link that allows node vj to receive information from node vi. A graph is said
to be undirected if and only if εji ∈ E implies εij ∈ E. In this paper, links are not required
to be bidirectional, i.e. we deal with digraphs; for this reason, we use the terms "graph" and
"digraph" interchangeably. Note that by convention and for notational purposes, we assume that
the given graph does not include any self-loops (i.e., εjj /∈ E for all vj ∈ V) although each node
vj obviously has a link (access) to its own information. A digraph is called strongly connected if
there exists a path from each vertex vi in the graph to each vertex vj (vj (cid:54)= vi). In other words,
for any vj, vi ∈ V, vj (cid:54)= vi, one can find a sequence of nodes vi = vl1, vl2, vl3, . . ., vlt = vj
(t ≥ 2) such that link (vlk+1, vlk) ∈ E for all k = 1, 2, . . . , t − 1.
j =(cid:12)(cid:12)N −
j
All nodes that can transmit information to node vj directly are said to be in-neighbors of node
j , is called the in-degree
j = {vi ∈ V εji ∈ E}. The cardinality of N −
(cid:12)(cid:12). The nodes that receive information from node vj belong
vj and belong to the set N −
of vj and is denoted by D−
to the set of out-neighbors of node vj, denoted by N +
N +
j , is called the out-degree of vj and is denoted by D+
In the algorithms we will consider, we will associate a positive weight pji for each (j, i) such
j = {vl ∈ V εlj ∈ E}. The cardinality of
j =(cid:12)(cid:12)N +
(cid:12)(cid:12).
j
that edge εji ∈ E ∪ {(vj, vj) vj ∈ V}. The nonnegative matrix P = [pji] ∈ Rn×n
as the entry at its jth row, ith column position) is a weighted adjacency matrix (also referred
(with pji
+
to as weight matrix) that has zero entries at locations that do not correspond to directed edges
(or self-edges) in the digraph. In other words, apart from the main diagonal, the zero/nonzero
structure of the adjacency matrix P matches exactly the given set of links in the digraph. We
use xj[k] ∈ R to denote the information state of node j at time step k. We first consider a
static network where the graph connectivity remains largely invariant (as it is usually the case
for distributed resources in applications, such as the power grid [27], [33]). At each time step k,
each node vj updates its information state to xj[k + 1] as a weighted linear combination of its
own value xj[k] and the available information received by its neighbors {xi[k] vi ∈ N −
j }. The
positive constant pji captures the weight of the information inflow from agent vi to agent vj.
In this work, since we deal with digraphs, we assume that each node vj chooses its self-weight
pjj and the weights plj on its out-going links vl ∈ N +
j . Hence, in its general form, each node
updates its information state xj[k + 1] according to
xj[k + 1] = pjjxj[k] +
pjixi[k] = pjjxj[k] +
xj←i[k], k = 0, 1, 2, . . .
(1)
(cid:88)
vi∈N −
j
(cid:88)
vi∈N −
j
where xj←i[k] (cid:44) pjixi[k], xi[k] ∈ R, is the value sent to node vj by node vi at time step k.
[Note that, since node vi chooses the weight pji, it is more convenient to sent xj←i[k] instead of
. . . xn[k])T and P = [pji] ∈
separately sending pji and xi[k].] If we let x[k] = (x1[k] x2[k]
Rn×n
+ , then (1) can be written in matrix form as
x[k + 1] = P x[k].
(2)
i xi[0]
n
Note that, with the exception of the diagonal entries, we have pji = 0, j (cid:54)= i, if and only if
(vj, vi) /∈ E. We say that the nodes asymptotically reach average consensus if limk→∞ xj[k] =
(cid:80)
for all vj ∈ V. The necessary and sufficient conditions for (2) to reach average consensus
are the following [20]: (a) P has a simple eigenvalue λi(P ) = 1 with left eigenvector 1T and
right eigenvector 1, and (b) all other eigenvalues of P (λj(P ), j (cid:54)= i) have magnitude less than
1 (λj(P ) < 1). If P ≥ 0 (as in our case), the necessary and sufficient condition is that P be a
primitive doubly stochastic matrix.
To capture dynamically changing topologies we will assume that we are given a fixed set of
components V = {v1, v2, . . . , vn} but the set of edges among them might change at various points
in time. This results in a sequence of digraphs of the form G[k] = (V,E[k]). Given a collection
of digraphs G[1], . . . ,G[m] (for some m ≥ 1) of the form G[k] = (V,E[k]), k = 1, 2, . . . , m,
the union digraph is defined as G1,2,...,m = (V,∪kE[k]). The collection of digraphs is said to be
jointly strongly connected, if its corresponding union graph G1,2,...,m forms a strongly connected
graph. A strongly connected graph certainly emerges if at least one of the graphs in the collection
is strongly connected, but it could also emerge even if none of the graphs forming the union is
strongly connected.
B. Ratio Consensus
In [27], the average consensus problem in a digraph is solved using ratio consensus. Each node
vj distributively sets positive weights on its self-link and out-going links so that the resulting
weight matrix P is primitive column stochastic, but not necessarily row stochastic. [Since the
graph is strongly connected, it will be sufficient for node vj to choose plj > 0 for vl ∈ N +
j ∪{vj}
j ∪{vj} plj = 1.] Average consensus is then reached by using
(zero otherwise) such that (cid:80)
vl∈N +
this weight matrix to run two linear iterations with appropriately chosen initial conditions and
by having each node take the ratio of the two values it maintains (one for each iteration). The
algorithm is stated below for a specific choice of weights, which assumes that each node knows
(this has the additional advantage of allowing
its out-degree and sets its link weights to
broadcasts, since the transmissions xl←j[k] (cid:44) pljxj[k] are identical for all vl ∈ N +
j ∪{vj}). Note,
however, that the algorithm works for any set of weights that adhere to the graph structure and
1+D+
1
j
form a primitive column stochastic weight matrix.
[27] Consider a strongly connected digraph G(V,E). Let yj[k] and zj[k] (for all
Lemma 1.
vj ∈ V and k = 0, 1, 2, . . .) be the result of the iterations
yj[k + 1] = pjjyj[k] +
zj[k + 1] = pjjzj[k] +
pjiyi[k] ,
pjizi[k] ,
(3)
(4)
(cid:88)
(cid:88)
j
vi∈N −
vi∈N −
j
where P = [pji] forms a primitive column stochastic matrix, and the initial conditions are
y0(2) . . . y0(V))T (cid:44) y0 and z[0] = 1. Then, the protocol asymptotically
y[0] = (y0(1)
converges to
where µj[k] =
yj[k]
zj[k]
.
lim
k→∞ µj[k] =
(cid:80)
v(cid:96)∈V y0((cid:96))
V
, ∀vj ∈ V ,
Note that the ratio consensus in [27] is actually a simpler version of more general algorithms
that have appeared under various names in the literature (e.g., weak ergodicity property in [32]
or the push-sum algorithm in [31]).
C. Products of SIA Matrices
A stochastic matrix P is called in [34] SIA (stochastic, indecomposable, and aperiodic) if the
limit Q = limk→∞ P k exists and has all of its columns identical. Specifically, Q = cP 1T for
some nonnegative vector cP . It can be shown that this definition of a SIA matrix is equivalent
to the standard definitions of indecomposability and aperiodicity for stochastic matrices.1 Let
A1, A2, . . . , Am be any square matrices of the same order. By a word (in the A's) of length
(cid:96) ∈ N we mean the product of (cid:96) A's (repetitions permitted). A stochastic, indecomposable, and
aperiodic (SIA) matrix is a column stochastic matrix B such that limk→∞ Bk exists and has all
columns the same, i.e., it is a rank one matrix of the form cB1T for some nonnegative column
vector cB. For the derivation of our results we make use of the theorem by Wolfowitz [34]
below.
Theorem 1.
[34] Let P = {P 1, P 2, . . . , P m} be a collection of column stochastic matrices
of the same order such that any word in the P 's is stochastic, indecomposable, and aperiodic
(SIA). For any > 0 there exists an integer ν() such that any word B = [bji] ∈ Rn×n
P 's) of length (cid:96) ≥ ν() satisfies δ(B) < , where δ(B) = maxj maxi1,i2 bj,i1 − bj,i2.
(in the
+
In words, Theorem 1 states that for large enough (cid:96), the product of (cid:96) matrices from the collection
P has all of its columns approximately the same. Note that the result does not mean that all
matrix products converge to a single matrix of the form c1T ; however, for large enough (cid:96), each
word B will take the form cB1T for some column vector cB.
III. MODELING DELAYS AND SWITCHING
A. Modeling Delays
We first focus on the average consensus problem in the presence of bounded delays when
the communication links among components are fixed and captured by an arbitrary strongly
connected digraph. More specifically, the transmission on the link from node vi to node vj at
time step k undergoes an a priori unknown delay τji[k], where τji[k] is an integer that satisfies
1A stochastic matrix P ∈ Rn×n is said to be decomposable if there exists a nonempty proper subset S ⊂ {1, 2, . . . , n} such
that pji = pij = 0 whenever vi ∈ S and vj /∈ S; also, P is indecomposable if it is not decomposable. A stochastic matrix
P is aperiodic if the Markov chain it describes is aperiodic, that is for every state i there exists ki such that for all k(cid:48) ≥ ki,
the probability of being at state i after k(cid:48) steps is greater than zero (for all k(cid:48)) or zero (for all k(cid:48)). Both indecomposability and
aperiodicity are properties that can be checked using the structure of the digraph that is induced by the zero/nonzero structure of
matrix P . Specifically, indecomposability follows from having a connected digraph with a single strongly connected component;
for an indecomposable matrix, aperiodicity is guaranteed as long as at least one component in the strongly connected component
has a self loop.
0 ≤ τji[k] ≤ ¯τji < ∞ (i.e., delays are bounded). The maximum delay is denoted by ¯τ =
max(vj ,vi)∈E ¯τji. We also assume that τjj[k] = 0, ∀vj ∈ V, at all time instances k (i.e., the own
value of a node is always available without delay).
Under this model, the information available to node vj at time step k (and which can be used to
update its value to xj[k + 1]) comprises of its own values and all values received by its neighbors
j ∪ {vj}}
by that time, i.e., it is a subset of the values in the set {xj←i[s] 0 ≤ s ≤ k, vi ∈ N −
(recall that, in the digraph setting we consider, node vi selects the weight of the link (vj, vi)
and thus sends to node vj the value xj←i[s] (cid:44) pjixi[s]). The protocol we will employ relies on
having each node vj update its information state at time step k to xj[k + 1] by combining (in
a linear fashion) its own value xj[k] and the possibly delayed information received at that time
step by its in-neighbors. In terms of the notation used above, this information is captured by
j ∪{vj}} . The exact way in which this information
{xj←i[s] 0 ≤ s ≤ k, s + τji[s] = k, vi ∈ N −
is used will be discussed later.
B. Modeling Switching
We consider a setting where the set of components in the multi-component system is fixed
to V = {v1, v2, . . . , vn}, but the (possibly directional) communication links between them are
allowed to change. A convenient way of capturing this is to assume that we have a sequence
of time-varying digraphs of the form G[k] = (V,E[k])). This means that at each time instant
k, each node vj has possibly different sets of in- and out-neighbors, denoted respectively by
j [k] and N +
N −
j [k]
and D+
j [k] = N +
j [k]. The in-degree and out-degree of node vj are defined as D−
j [k], respectively.
j [k] = N −
As in the case when there is no change in the interconnection topology, each node vj is in
charge of setting the weights plj[k], vl ∈ N +
j [k], on all links to its out-neighbors. Due to the
changing topology, the weight matrix will be time-varying and will be denoted by P [k]. What
is important is for P [k] to be column stochastic and to have positive weights at all links of the
graph including its diagonal elements. As in Lemma 1, nodes can easily set the weights on the
links to their out-neighbors to ensure column stochasticity as long as each node vj has knowledge
j [k] at each time step (in such case, each node vj ∈ V sets plj =
of its out-degree N +
for
vl ∈ N +
j [k] ∪ {vj}). There are various ways in which the out-degree information can become
available at each node (in undirected graphs this information is obviously available but even in
1
1+D+
j [k]
digraphs it can become available with simple protocols that we describe in more detail later).
Even if the out-degree information becomes available with some delay, the protocols we propose
can still be modified to reach consensus to the exact average of the nodes' initial values.
In our analysis of changing interconnection topology, we consider two cases.
(i) Switching without delays: When we have a varying interconnection topology and there exist
no delays in the communication links, each node vj updates its information state at time step
k to xj[k + 1] by combining its own state xj[k] and the available information received by its
neighbors {xj←i[k] vi ∈ N −
j [k]} (the latter information also includes the positive weights pji[k]
that capture the weight of the information inflow assigned by component vi to the link (vj, vi) at
time k). Here, we will consider two subcases: (a) each transmitting node knows its out-degree
as soon as the change takes place, and (b) each transmitting node knows its out-degree with
some delay.
(ii) Switching with delays: In this case, each node vj updates its information state at time step
k to xj[k + 1] by combining its own value xj[k] and the available (possibly delayed) information
{xj←i[s] 0 ≤ s ≤ k, s + τji[s] = k, vi ∈ N −
j [s]} (the latter information also includes the
positive weights pji[s], that component vi assigns to link (vj, vi) at time s. We consider again
the two cases (a) and (b) mentioned earlier where each node vj discovers its out-degree without
or with delay (the out-degree with delay will be made clear in the analysis), and also consider
a third case (c) in which a node vj discovers an established link with some delay.
IV. HANDLING DELAYS IN A DIGRAPH
links to its out-neighbors so that these weights are positive and satisfy(cid:80)
We first start with a static digraph, where each link transmission can undergo a bounded delay.
j } on
We assume that each node vj chooses its self weight pjj and the weights {plj vl ∈ N +
j ∪{vj} plj = 1 for
as in Lemma 1).
all vj ∈ V (a simple choice would be to set all of these weights to
vl∈N +
1
1+D+
j
We employ a protocol where each node updates its information state according to the following
relation:
xj[k + 1] = pjjxj[k] +
(cid:88)
vi∈N −
j
xj←i[k − r]Ik−r,ji[r] , k = 0, 1, 2, . . .
where xj←i[k − r] (cid:44) pjixi[k − r], xj[0] ∈ R is the initial state of node vj, and
r=0
¯τ(cid:88)
1,
Ik,ji(τ ) =
if τji[k] = τ,
0, otherwise.
(5)
(6)
In the absence of delays, we have τji[k] = 0 and the update relation (5) reduces to (1) with
constant weights. We will show that if (5) is employed in place of (1) and is used to run two
iterations as in Lemma 1, the resulting ratio consensus approach can still be used to calculate
the exact average, despite arbitrary but bounded delays at the communication links. Essentially,
we establish that (5) is a ratio consensus protocol tolerant to arbitrary but bounded delays.
Assumptions 1. For the analysis below we are given a digraph G(V,E) (that represents the
information exchange between agents in a multi-agent system). Each node vj ∈ V has an initial
value y0(j) and runs ratio consensus, i.e., two versions of the iteration in (5), one with initial
value y0(j) and one with initial value z0(j) = 1. We make the following assumptions:
and (vl, vj) ∈ E, and satisfy (cid:80)n
(A1) The digraph is strongly connected, and the (nonnegative) weights plj are positive for l = j
l=1 plj = 1 for all vj ∈ V (so that they form a column
stochastic matrix P ). For simplicity, we will assume that each node sets the weights on
the links to its out-neighbors to plj = 1
1+D+
j
for vl ∈ N +
j ∪ {vj} (zero otherwise).
(A2) There exists a finite ¯τ that uniformly bounds the delay terms; i.e. τji[k] ≤ ¯τ < ∞ for all
links (vj, vi) ∈ E for all time instants k. In addition, τjj[k] = 0 for all vj ∈ V and all k.
Note that Assumption (A1) is necessary for the successful operation of any distributed algo-
rithm seeking consensus. The particular choice of weights ensures that the weight matrix P is
primitive column stochastic. Assumption (A2) implies that no message is lost in the network and
every agent updates its value using values from its in-neighbors at least once every ¯τ consecutive
updates. The proof of the theorem below is developed in the remainder of this section.
Theorem 2. Consider a strongly connected digraph G(V,E). Let yj[k] and zj[k] (for all vj ∈ V
and k = 0, 1, 2, . . .) be the result of the iterations
yj[k + 1] = pjjyj[k] +
zj[k + 1] = pjjzj[k] +
yj←i[k − r]Ik−r,ji[r] ,
zj←i[k − r]Ik−r,ji[r] ,
(7)
(8)
(cid:88)
(cid:88)
vi∈N −
j
¯τ(cid:88)
¯τ(cid:88)
r=0
vi∈N −
j
r=0
under Assumptions (A1) and (A2). The initial conditions are y[0] = (y0(1) y0(2) . . . y0(V))T ≡
y0 and z[0] = 1, and Ik,ji is an indicator function that captures the bounded delay τji[k] on
link (vj, vi) at iteration k (as defined in (6), τji[k] ≤ ¯τ). Then, the solution to the average
, ∀vj ∈ V ,
consensus problem can be asymptotically obtained as lim
(cid:80)
k→∞ µj[k] =
v(cid:96)∈V y0((cid:96))
V
where µj[k] =
yj[k]
zj[k]
.
Notice that the two iterations in the above theorem are coupled via the delays (the indicator
functions Ik,ji are the same in both iterations). Our proof is based on an augmented representation
(digraph) that allows us to model the distributed system with bounded delays as described in
(5), which processes packets as soon as they arrive at the destination node. We then use this
augmented representation to establish that (for fixed communication topologies) the distributed
ratio consensus algorithm in (7) -- (8) will lead to asymptotic average consensus, regardless of the
nature and order of the delays, as long as they are bounded. Note that the nodes are not required
to know the delay of any packet or any upper bound of the delay; each node considers all the
packets that arrive at that time step, by including their value in the sum.
In the augmented digraph representation, we add extra, "virtual" nodes and use them to model
the delays. The maximum number of "virtual" nodes for each original node is bounded by
the maximum delay ¯τ. In particular, for each node vj ∈ V we introduce ¯τ "virtual" nodes
v(1)
holds the sum of the values that are
j
. At each time step k, virtual node v(τ )
, . . . , v(¯τ )
, v(2)
j
j
j
destined to arrive to node vj in τ steps. The augmented digraph has (¯τ + 1)V nodes and
(1 + 2¯τ )E edges. Before presenting the general case, we illustrate the construction of the
augmented digraph via an example.
Example 1. Consider the network of two agents exchanging information as shown in Figure 1.
Note that that the weights p11, p12, p21, and p22 are all strictly positive, and satisfy p11 + p21 = 1
and p22+p12 = 1; in the simple case presented in the introduction (and mentioned in Theorem 2),
we have p12 = p22 = 1/(1 +D+
1 ) = 1/2. Suppose the agents
experience delays that are bounded by 2 (¯τ = 2). Therefore, two extra "virtual" nodes will be
2 ) = 1/2 and p21 = p11 = 1/(1 +D+
added for each node (see Figure 2), depicting the states at which the delayed messages reside
before reaching their destination (refer to Figure 2).
Fig. 1. A simple example with two nodes when the links do not experience any delays.
(a) Graph representation of the network at a time instant
k = k1 when there exist no delays. As a result, each node
uses the value sent by its neighbor directly, plus the delayed
information (sent in previous time instances) that arrives at
time instant k = k1.
(b) Graph representation of the network at a time instant
k = k2 for which node v2 sends information to node v1
with delay τ12(k2) = 1, while node v1 sends information
to node v2 with delay τ21(k2) = 2.
Fig. 2.
A simple example with two nodes to illustrate the modeling of delays using the proposed augmented digraph. The
maximum allowable delay (¯τ) is 2. In Figure (a) there exist no delays between communication links at that time instant (k = k1),
whereas in Figure (b) both nodes experience delays. Active links are shown by boldface black lines in each case.
Taking x[k] = (x1[k] x2[k] x(1)
1 [k] x(1)
2 [k] x(2)
1 [k] x(2)
2 [k])T , the iteration in the augmented
v1v2p21p12p11p22v1v2p21p12p11p22v(2)2v(1)2v(1)1v(2)1p2111p21p1211p12v1v2p21p12p11p22v(2)2v(1)2v(1)1v(2)1p2111p21p1211p12 digraph can be written as x[k + 1] = P [k]x[k], where P [k] depends on the nature of delays. For
example, when there are no delays in the network, say at time instant k = k1, the network is
represented by Figure 2(a) and the weight matrix is given by P [k1] below. Similarly, if at time
instant k = k2 there is a delay of 2 at link (v2, v1) and a delay of 1 at link (v1, v2) (the network
is shown in Figure 2(b)), the matrix representation is given by P [k2] below.
p11
p12
1 0
0
p21
p22
0
1 0
0
0
0
0
0
0
0
0
0 0 1
0 0 0
0 0 0
0 0 0
0
0
0
1
0
0
.
p11
0
0
0
0
0
p21
p22
p12
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
P [k1] =
, P [k2] =
···
···
...
···
···
(cid:16)
In the general case, in a network of n = V nodes, we introduce ¯τ n nodes (for a total of
(¯τ + 1)n nodes) so that x[k + 1] = P [k]x[k], where
P [k] (cid:44)
P0[k]
P1[k]
...
P¯τ−1[k]
In×n
0
0
...
0
In×n
...
0
,
0
0
...
In×n
0
(9)
(cid:17)
, r = 1, 2, . . . ¯τ.
with x[k] = (cid:0)xT [k] x(1)[k] . . . x(¯τ )[k](cid:1)T and x(r)[k] =
P¯τ [k]
0
0
x(r)
1 [k] . . . x(r)
n [k]
Note that P0[k], P1[k], . . . , P¯τ [k] are appropriately defined nonnegative matrices that depend on
the link delays that are experienced by messages sent at time k. Specifically, Pr[k] is a matrix
associated only with the links of the graph for which the message was delayed by r steps at
time step k, and satisfies
Pr[k](j, i) =
P (j, i),
0,
if τji[k] = r, (j, i) ∈ E,
otherwise.
Note that, for each (j, i) ∈ E, only one of P0[k](j, i), P1[k](j, i), ..., P¯τ [k](j, i) is nonzero and
is equal to P (j, i). Thus, we also have
¯τ(cid:88)
r=0
P =
Pr[k] ,
k = 0, 1, 2, . . .
(10)
Matrix P [k] may take at most (¯τ + 1)E matrix values, where (¯τ + 1) is the total number of
states ("virtual" and original) for each node vj. Specifically, if there exists an edge (vj, vi) in the
original digraph, then that edge also exists in the augmented digraph along with edges (v(1)
, vi),
j ). However, among
the (¯τ + 1) entries of P [k] corresponding to the edge (vj, vi) only one of them could be nonzero
, vi), and also edges (vj, v(1)
, vi), . . ., (v(¯τ )
j ), . . ., (v(¯τ−1)
, v(2)
j
j ), (v(1)
j
(v(2)
j
, v(¯τ )
j
j
(and equal to pji); the others will be zero. In the sequel we do not require the matrix P [k] to
be known at each time step k; what we utilize is that P [k] will be a matrix from a finite set of
possible matrices P, which have certain useful properties.
Proposition 1. Let P = {P 1, P 2, . . . , P (¯τ +1)E} be the set of all possible P [k] as defined in (9).
Then, for integer (cid:96), (cid:96) ≥ ¯τ + 1, any (cid:96)-length word B = P [k + (cid:96)]P [k + (cid:96) − 1] . . . P [k + 1] is SIA.
Moreover, for (cid:96) ≥ n(¯τ + 1), the first n rows of matrix B will be positive with minimum entry
greater or equal to cmin ≡(cid:16) 1D+
max
(cid:17)n(¯τ +1)
, where D+
max = maxvj∈V D+
j .
Proof of Proposition 1:
In order to prove that B = P [k + (cid:96)] . . . P [k + 2]P [k + 1] is SIA,
we have to show that it is (i) column stochastic, (ii) indecomposable, and (iii) aperiodic.
(i) Column Stochasticity: This is easy to see as it is equivalent to proving that the product of
two or more column stochastic matrices of the same order is also a column stochastic matrix
(the result follows easily by induction and is standard).
(ii) Indecomposability: We argue indecomposability for (cid:96) ≥ ¯τ + 1 (the result also holds for any
0 ≤ (cid:96) < ¯τ + 1 but we do not discuss the proof here due to space limitations). Write matrix B
in block form as
B =
B0,0
B1,0
...
B0,1
B1,1
...
B0,2
B1,2
...
B¯τ−1,0 B¯τ−1,1 B¯τ−1,2
B¯τ ,0
B¯τ ,1
B¯τ ,2
,
B0,¯τ
B1,¯τ
...
···
···
...
··· B¯τ−1,¯τ
···
B¯τ ,¯τ
where all blocks are nonnegative matrices of size n× n. We will argue that (i) the zero/nonzero
structure of B0,0 corresponds to a graph that is strongly connected, and (ii) each of B0,0, B0,1,
B0,2, ..., B0,¯τ has strictly positive entries on its diagonal. These two facts establish that the graph
that corresponds to the zero/nonzero structure of the overall matrix B has the following property:
(i) any pair of non-virtual nodes (i.e., the top n nodes) can be connected via a directed path
(that can actually involve only non-virtual nodes); (ii) all other (virtual) nodes have an outgoing
link to at least one of the non-virtual nodes. Therefore, the set of non-virtual nodes is part of a
strongly connected component; this component could potentially involve other (virtual) nodes in
the graph, but no other strongly connected component exists. Thus, matrix B is indecomposable.
For fact (i), we need to explain why B0,0 corresponds to a graph of n nodes that is strongly
connected. It is not hard to see that one can write
B0,0 = (Π(cid:96)
l=¯τ +2P0[k +l])P¯τ [k +1]+E0,0
l=2P0[k +l])P0[k +1]+(Π(cid:96)
l=3P0[k +l])P1[k +1]+...+(Π(cid:96)
l=l1
where Πl2
A[l] ≡ A[l2]A[l2 − 1]...A[l1] (Πl2
A[l] ≡ I for l2 = l1 − 1 and zero otherwise)
and E0,0 is a nonnegative matrix (that can be expressed as the sum of various products of
the nonnegative2 blocks composing the P matrices). Since the diagonal elements in matrix
l=l1
Π(cid:96)
P0[k + l] (for l = 1, 2, ..., (cid:96) are strictly positive, we know that the diagonals of each product
l=mP0[k + l], m = 2, 3, ..., ¯τ + 2, will be strictly positive and thus the elements of each term
l=mP0[k + l])Pm−2[k + 1] will be positive at the locations where Pm−2[k] is positive. Thus,
from the expression for B0,0 above, the zero/nonzero structure of B0,0 corresponds to a graph of
(Π(cid:96)
2The fact that the blocks are nonnegative is important because it means that nonzero entries created by some products cannot
be cancelled by nonzero entries of other products.
n nodes that includes all the edges in(cid:80)¯τ
r=0 Pr[k + 1] = P (recall (10)); thus, all edges in the
original graph are included and, since the original graph is strongly connected, B0,0 corresponds
to a graph that is strongly connected.
For fact (ii), we need to explain why each B0,r, r = 0, 1, ..., ¯τ, has strictly positive diagonal
entries. For r = 0, this follows for the discussion above. For r = 1, 2, ..., ¯τ, we can also write
B0,r = (Π(cid:96)
l=r+2P0[k + l])P0[k + 1 + r] + E0,r ,
where E0,r is again a nonnegative matrix that can be expressed as the sum of various products
of nonnegative matrices. Since the diagonal elements in matrix P0[k + l] (for l = 1, 2, ..., (cid:96)) are
strictly positive, we know that the diagonals of each B0,r will be strictly positive.
(iii) Aperiodicity: Since the graph corresponding to B is indecomposable, aperiodicity is easily
established due to the fact that the diagonal entries that correspond to the original (non-virtual)
nodes in the strongly connected component are nonzero (it is sufficient for at least one of them
to be nonzero).
B will be positive with minimum entry greater or equal to cmin ≡(cid:16) 1D+
To prove the second part of the proposition (i.e., for (cid:96) ≥ n(¯τ + 1), the first n rows of matrix
), notice that for
(cid:17)n(¯τ +1)
max
(cid:96) = n(¯τ + 1) we can write B as
B = BinBin−1...Bi2
Bi1 ,
(cid:125)
(cid:124)
(cid:123)(cid:122)
B(cid:48)
where each Bim is the product of ¯τ + 1 consecutive P , i.e.,
Bim = P [k + m(¯τ + 1)] . . . P [k + (m − 1)(¯τ + 1) + 2]P [k + (m − 1)(¯τ + 1) + 1] .
From the discussion on indecomposability, we know that each Bim has blocks B(im)
B(im)
0,¯τ
0,1 , ...,
corresponds to a graph that includes the
such that the zero/nonzero structure of B(im)
0,0
0,0 , B(im)
original strongly connected graph of size n and has a positive diagonal. Thus, the product of
n − 1 such blocks will result in a strictly positive diagonal block for matrix B(cid:48). An additional
multiplication by Bi1 on the right, will ensure that each of the top ¯τ + 1 blocks of matrix B will
be strictly positive. Since B involves the product of n(¯τ + 1) nonnegative matrices P (whose
minimum nonzero entry is
1D+
max
), the minimum entry in B will be greater or equal to cmin.
For (cid:96) = n(¯τ + 1) + 1, we have a matrix product of the form BP [k + 1], where B is the
product of n(¯τ + 1) matrices P (thus, its top n rows are strictly positive with minimum entry
cmin). Since P [k +1] is a column stochastic matrix, we can easily conclude that matrix BP [k +1]
will also have its top n rows positive with minimum entry cmin. The claim in the second part
of the proposition (that, for (cid:96) ≥ n(¯τ + 1), the first n rows of matrix B will be positive with
minimum entry greater or equal to cmin), then follows easily by induction.
Proof of Theorem 2:
If we use the augmented graph representation with initial conditions
¯y[0] = [yT
0 0 0 . . . 0]T and ¯z[0] = [1T 0 0 . . . 0]T , we can write
¯y[k] = P [k] . . . P [2]P [1]
¯z[k] = P [k] . . . P [2]P [1]
(cid:125)
(cid:125)
¯y[0] ,
¯z[0] .
(cid:124)
(cid:124)
(cid:123)(cid:122)
(cid:123)(cid:122)
Bk
Bk
By Proposition 1 and Wolfowitz's Theorem, we know that for any > 0, the resulting word Bk
satisfies (for k ≥ ν()) Bk = cBk
1T + Ek, where cBk is an appropriate nonnegative vector, and
Ek is an error matrix with entries with absolute value smaller than /2 (i.e., Ek(j, i) < /2 for
all j, i). Taking k to also satisfy k ≥ n(¯τ + 1), it follows from Proposition 1 that each of the
first n entries of cBk (i.e., the entries that correspond to non-virtual nodes) will be greater than
cmin. Without loss of generality, we take < 2cmin in the remainder of this discussion.
With the above notation at hand, we have for vj ∈ V
µj[k] (cid:44) ¯yj[k]
¯zj[k]
=
Bk(j, :)¯y[0]
Bk(j, :)¯z[0]
=
cBk(j)(1T + eT
cBk(j)(1T + eT
k )¯y[0]
k )¯z[0]
=
(1T + eT
(1T + eT
k )¯y[0]
k )¯z[0]
,
where cBk(j) is the jth element of vector cBk, and eT
and satisfies emax(k) ≡ maxi{ek(i)} < /2.
k = Ek(j, :) is the jth row of matrix Ek
Since ¯z[0] = 1 ≥ 0 (elementwise), the denominator of the above expression can be bounded
n(1 − emax(k)) ≤ (1T + eT
k )¯z[0] ≤ n(1 + emax(k)) .
Similarly, assuming that (cid:80)
(cid:80)
l ¯yl[0] =(cid:80)
l yl[0] > 0 (when (cid:80)
l ¯yl[0] =(cid:80)
l yl[0] < 0 or (cid:80)
l ¯yl[0] =
l yl[0] = 0 we can apply a similar analysis), we can bound the numerator of the above
expression as
where Σy =(cid:80)
Σy − emax(k)Σy ≤ (1T + eT
l ¯yl[0] =(cid:80)
l yl[0] and Σy =(cid:80)
l ¯yl[0] =(cid:80)
k )¯y[0] ≤ Σy + emax(k)Σy ,
l yl[0]. Putting the above inequalities
together, we obtain
Σy − emax(k)Σy
n(1 + emax(k))
≤ µj[k] (cid:44) ¯yj[k]
¯zj[k]
≤ Σy + emax(k)Σy
n(1 − emax(k))
,
(cid:80)
which can be relaxed (after some algebraic manipulations) to µ∗ − Mk ≤ µj[k] ≤ µ∗ + Mk,
where µ∗ =
. By Wolfowitz theorem,
we can take k as large as necessary to make Mk arbitrarily small (by ensuring that and thus
is the exact average and Mk = µ∗ (Σy+Σy)emax(k)
Σy(1−emax(k))
l yl[0]
n
emax(k)) is as small as desired).
Example 2. Consider the directed network on the left of Figure 3 where each node vj chooses
its self-weight and the weights of its outgoing links to be (1 + D+
j )−1 so that the weight matrix
P is primitive column stochastic as shown on the right of the figure. Each node vj updates its
information state xj[k] using equation (3), so that the information state for the whole network is
given by x[k +1] = P x[k]. We first use the update formula (1) with y[0] = (−1 2 3 4 2)T ≡ y0
and no delays (¯τ = 0). Since the update matrix is column stochastic, iteration (3) for this network
converges, but not necessarily to the average (as shown in Figure 4(a) the nodes do not even
reach consensus). As suggested in [27], by simultaneously running two iterations y[k] and z[k]
(using the weights in matrix P ) with initial conditions y[0] = y0 and z[0] = 1, respectively, then
average consensus is asymptotically reached for the ratio yj[k]/zj[k] (see Figure 4(b)).
We now consider delays by taking the maximum allowable delay to be ¯τ = 5. At each link
at each time instant, the delay is an integer in {0, 1, 2, ..., 5} (in our simulations each possible
delay is chosen with probability 1/6). If we run our update formula (as in (5)) for the network
in Figure 3 with weights P and x[0] = y0, the algorithm does not converge (see Figure 5(a)).
However, if we run ratio consensus in (7) -- (8) with initial conditions y[0] = y0 and z[0] = 1
P =
0
0
1/3
1/3 1/3
0
1/3 1/3 1/2
0
0
0
1/3 1/2
0
0
1/2
0
0
0
1/3
1/2 1/3
0
1/3
Fig. 3. A simple digraph of five nodes when the links do not experience any delays.
(a) The iteration of (1) for the network in Figure 3 does
not lead to average consensus (not even consensus) for the
digraph, since P is not a doubly stochastic matrix.
(b) By running ratio consensus in (7) -- (8) with appropriate
initial conditions, average consensus is reached (no delays
are introduced yet, i.e., ¯τ = 0).
Fig. 4.
Iteration (3) converges but does not reach consensus (left). By simultaneously running two iterations y[k] and z[k] (using
the weights in matrix P ) with initial conditions y[0] = y0 and z[0] = 1, respectively, then average consensus is asymptotically
reached for the ratio yj[k]/zj[k] (right).
respectively, then average consensus is asymptotically reached for the ratio yj[k]/zj[k] (Figure
5(b)). This demonstrates the validity of our theoretical analysis, both in the sense that each of
the individual iterations does not convergence and also in the sense that the ratios converge to
the average of the initial values.
It is obvious from the simulations that the convergence speed of the algorithm depends on the
delays (e.g., longer delays should result in slower convergence of the algorithm). Our discussion
did not characterize the worst-case combination of delays but, nevertheless, the final average
value is not affected by the particular realization of delays.
v1v2v3v4v51/31/31/21/21/31/31/31/31/31/21/21/31/302468101214161820−1−0.500.511.522.533.54Values at each node vs Number of iterationsNumber of iterationsValue at each node02468101214161820−1−0.500.511.522.533.54Ratio at each node vs Number of iterationsNumber of iterationsRatio at each node (a) The update rule (5) does not converge for the digraph,
due to the presence of delays.
(b) By running ratio consensus in (7) -- (8) with appropriate
initial conditions, we can reach average consensus even in
the presence of delays.
Fig. 5. The update formula in (5) for the network on the left of Figure 3 with weights P and x[0] = y0, does not converge
(left); however, if we run ratio consensus in (7) -- (8) with initial conditions y[0] = y0 and z[0] = 1 respectively, then average
consensus is asymptotically reached for the ratio yj[k]/zj[k] (right).
To gain additional insight into the problem, we also consider the convergence of node 1 under
(i) different upper bounds in delays (see left of Figure 6 where delays are equally likely as
before), and (ii) varying network size (see right of Figure 6 where random geometric graphs
of different sizes are used and ¯τ = 5 with delays being equally likely as before). It is obvious
from the simulations that the convergence speed of the algorithm depends on the delays (e.g.,
longer delays result in slower convergence). Nevertheless, for fixed ¯τ it appears that the size of
the network has no effect on the convergence time (at least for geometric graphs).
Fig. 6. Convergence of the ratio at node 1 for different upper bound ¯τ on delays (left) and different network sizes (right).
05101520253035404550−1−0.500.511.522.533.54Values at each node vs Number of iterationsNumber of iterationsValue at each node05101520253035404550−1−0.500.511.522.533.54Ratio at each node vs Number of iterationsNumber of iterationsRatio at each node0102030405060708090100−101234(cid:53)(cid:68)(cid:87)(cid:76)(cid:82)s at node 1 vs Number of iterationsNumber of iterationsRatios at node 1 τ = 0τ = 5τ = 10τ = 15τ = 200510152025303540−2−1.5−1−0.500.511.5(cid:53)(cid:68)(cid:87)(cid:76)(cid:82)s at node 1 vs Number of iterationsNumber of iterationsRatio at node 1 n=10n=20n=30n=50n=100 Remark 1. In [14], [35], the following update formula is suggested
(cid:88)
vi∈N −
j
xj[k + 1] = p(cid:48)
jjxj[k] +
jixi[k − dji[k]] ,
p(cid:48)
k = 0, 1, 2, . . .
(11)
ji form a doubly stochastic weight matrix P (cid:48) = [p(cid:48)
where x[0] = y0, the weights p(cid:48)
ji] and dji[k]
is chosen so that node vj uses in its update the most recently seen value from node vi (i.e.,
dji[k] = minτji[k−t]=t,0≤t≤¯τ{t}). Since the weight matrix P (cid:48) = [p(cid:48)
is doubly stochastic,
we know that in the absence of delays the iteration in (11) would reach asymptotic average
ji] ∈ Rn×n
+
consensus. The iteration also reaches consensus in the presence of delays (regardless of the
delays introduced, as long as they are bounded [35]), but not necessarily to the exact average
of the initial values. The value the nodes converge to depends on the specific delays that are
introduced during the execution of the iteration.
V. HANDLING CHANGING INTERCONNECTIONS
In this section, we extend the previous setting to include time-varying communication links
(in addition to bounded delays on each link). We assume that we have a time-varying digraph,
in which the set of nodes is fixed but the communication links can change, i.e., at time step
k the interconnections between components in the multi-component system are captured by a
digraph G[k] = (V,E[k]). For the analysis below, we let ¯G = {G1,G2, . . . ,Gν}, ν ≤ 2n2−n, be
the set of all possible digraphs3 defined for a given set of nodes V. We start our analysis by
considering the simplest case where we have changing interconnection topology without delays
and assuming that each node knows its (instantaneous) out-degree at that particular time instant.
We start with the assumptions below (some of these assumptions are relaxed later on).
Assumptions 2. For the analysis below, the interconnection topology is described by a sequence
of digraphs G[1], G[2], . . ., G[k], . . ., of the form G[k] = (V,E[k]).
3Each of the n nodes may be connected (out-going link) with up to (n − 1) other nodes. As a result, we have n(n − 1)
possible links, each of which can be either present or not. Hence, we have 2n(n−1) possible graph combinations. Of course,
depending on the underlying application, some of these interconnection topologies may be unrealizable.
(B1) At each time instant k, each node vj knows the number of nodes receiving its message
(i.e., the number of its out-neighbors D+
j [k]).
(B2) There exist no delays in the delivery of messages.
(B3) We can find an infinite sequence of times t0, t1, . . . , tm, ..., where t0 = 0, 0 < tm+1 − tm ≤
(cid:96) < ∞, with the property that for any m ∈ Z+ the union of graphs G[tm],G[tm +
1], . . . ,G[tm+1 − 1] is strongly connected.
Remark 2. Assumption (B1) requires that the transmitting node knows the number of nodes
receiving its messages at each time instant. In an undirected graph setting, this is not too difficult;
in a digraph setting, this is not as straightforward but there are many ways in which knowledge
of the out-degree might be possible. For example, there can be an acknowledgement signal (ACK)
via a distress signal (special tone in a control slot or some separate control channel) sent at
higher power than normal so that it is received by transmitters in its vicinity [36]. Knowledge
of the out-degree is also possible if the nodes periodically perform checks to determine the
number of their out-neighbors (e.g., by periodically transmitting the distress signals mentioned
above). As we discuss later, at the cost of little additional complexity, the nodes can also handle
situations where they learn their out-degree with some delay. Assumption (B2) is made to keep
things simple and is relaxed later. Assumption (B3) stems from the fact that we require that there
exists paths between any pair of nodes infinitely often.
In its general form, each node updates its information state according to the following relation:
(cid:88)
vi∈N −
j [k]
xj[k + 1] = pjj[k]xj[k] +
xj←i[k] ,
k = 0, 1, . . .
(12)
where xj←i[k] (cid:44) pji[k]xi[k] is the information sent from node vi to node vj at time step k, and
xj[0] ∈ R is the initial state of node vj. Since the out-degree is known at the transmitting node,
j [k] ∪ {vj} (this
each node vj can easily set the (positive) weights plj[k] =
l=1 plj[k] = 1 for all vj ∈ V but more generally, as in the case of a fixed
topology, each node only needs to ensure that the weights on its out-going links are positive
choice satisfies (cid:80)n
for vl ∈ N +
1
1+D+
j [k]
and sum to unity). Note that unspecified weights in P [k] are set to zero and correspond to pairs
of nodes (vl, vj) that are not connected at time step k, i.e., plj[k] = 0, for all (vl, vj) /∈ E[k],
l (cid:54)= j. If we let x[k] = (x1[k] x2[k]
then (12) can be
. . . xn[0])T ≡ xT
0 .
written in matrix form as x[k + 1] = P [k]x[k], where x[0] = (x1[0] x2[0]
Note that, with the specific choice of matrix P [k] (based on the out-degree of each node as
. . . xn[k])T and P [k] = [pji[k]] ∈ Rn×n
+
described above), the matrices P [k] are column stochastic and have strictly positive elements on
their diagonal. This fact will be important in our proof later on which utilizes Theorem 1 on a
particular set of matrices.
Remark 3. Throughout the operation of the algorithm, communication links can be initiated or
terminated by either (a) the receiving node, or (b) the transmitting node. Possible communication
protocols to perform these tasks are described briefly below.
(a) When node vl wants to receive messages from node vj (e.g., because it is in the neighborhood
of vj), it can send a distress signal to pass this request to vj (alternatively, node vl can send
the message to node vj using some path in the digraph or using some sort of flooding scheme).
When node vj receives the request from vl, it sends an acknowledgement packet (directly to node
vl) and the communication link is initiated. In practice, this might not necessarily require node
vj to transmit a separate package to node vl (e.g., in a wireless broadcast setting) or to transmit
at a higher power (e.g., if vl is already in its range); however, it does imply that node vj will
adjust its self-weight and the weights plj, vl ∈ N +
j , on the links to its out-neighbors in order to
ensure that column stochasticity is preserved. If, on the other hand, node vl wants to terminate
the communication link, it sends (or broadcasts if there exists a single communication channel
and the message cannot be specifically directed to node vj) a distress signal destined for node vj
(alternatively, it can use a flooding-like strategy via the paths in the digraph); as soon as node
vl receives an acknowledgement from node vj along with the latest message with values for the
last update, then the link can be terminated. If node vl does not receive the acknowledgement
message from node vj the link remains active.
(b) Note that if the transmitting node vj wants to terminate a communication link to node vl, it
is enough to simply initiate such a request to node vl (since a direct link is available).
Lemma 2. Consider a sequence of graphs of the form G[k] = (V,E[k])), k = 0, 1, 2, ..., such that
there exists an infinite sequence of time instants t0, t1, . . . , tm, . . ., where t0 = 0, 0 < tm+1−tm ≤
(cid:96) < ∞, m ∈ Z+, with the property that for any m ∈ Z+ the union of graphs G[tm],G[tm +
1], . . . ,G[tm+1 − 1] is strongly connected. Let yj[k], ∀vj ∈ V, be the result of iteration (12) with
j [k] ∪ {vj} (zeros otherwise) and initial conditions y[0] = y0, and
plj[k] =
j [k] ∪ {vj}
let zj[k], ∀vj ∈ V, be the result of iteration (12) with plj[k] =
(zeros otherwise) and with initial condition z[0] = 1. Then, the solution to the average consensus
for vl ∈ N +
for vl ∈ N +
1
1+D+
j [k]
1
1+D+
j [k]
(cid:80)
problem in the presence of dynamically changing topologies can be obtained as lim
k→∞ µj[k] =
v(cid:96)∈V y0((cid:96))
V
, ∀vj ∈ V , where µj[k] =
yj[k]
zj[k]
.
Proof: Let P tm+1−tm
(cid:44) P [tm+1 − 1]P [tm+1 − 2] . . . P [tm]. Since the union of graphs from
time instant tm until tm+1− 1, i.e., the set of graphs G[tm],G[tm + 1], . . . ,G[tm+1− 1], is strongly
connected and each matrix involved in the product has strictly positive elements on the diagonals,
matrix P tm+1−tm is SIA for m ∈ Z+. Furthermore, products of matrices of the form P tm+1−tm are
SIA (since each such matrix has positive elements on its diagonal, the product of such matrices
will have a positive entry at its (j, i) position if at least one of the matrices has a positive
element at its (j, i) position; hence, the zero/nonzero structure of the product will correspond
to strongly connected graph). Hence, according to Theorem 1, for any > 0, there exist a
finite integer ν() ∈ N, such that a finite word W given by the product of a collection of ν
stochastic matrices of the form P tm+1−tm has all of its columns approximately the same, i.e.,
1T + E, where cWν is a nonnegative column vector
P tk+ν−tk+ν−1 . . . P tk+2−tk+1P tk+1−tk = cWν
and matrix E has entries that are bounded in absolute value by /2. From this point onwards,
the proof continues as in the proof of Theorem 2.
A. Changing interconnection topology with communication delays
Assumptions 3. In the presence of delays, we make the following extra assumption:
(C1) There exists a finite ¯τ that uniformly bounds the delay terms, i.e. τji[k] ≤ τ ji ≤ τ; this is
the same as in assumption (A2).
In this case, each node updates its information state according to the following iteration:
xj[k + 1] = pjj[k]xj[k] +
xj←i[k − r]Ik−r,ji[r] ,
k = 0, 1, 2, . . .
(13)
¯τ(cid:88)
(cid:88)
r=0
vi∈N −
j [k−r]
where xj←i[k − r] (cid:44) pji[k − r]xi[k − r] is the value sent from node vi to node vj at time step
k − r that occurs delay r, xj[0] ∈ R is the initial value of node vj, and the values pji[k] ≥ 0
depend on the topology of the graph at time k.
To handle delays in a network of n = V nodes, we introduce ¯τ n nodes (for a total of (¯τ +1)n
nodes) so that we can write
where (as before)
with
x[k + 1] = P [k]x[k] ,
P [k] (cid:44)
P0[k]
P1[k]
...
P¯τ−1[k]
P¯τ [k]
In×n
0
0
...
0
0
In×n
...
0
0
···
···
...
···
···
x[k] =(cid:0)xT [k] x(1)[k] . . . x(¯τ )[k](cid:1)T
(cid:16)
(cid:17)
x(r)[k] =
x(r)
1 [k] . . . x(r)
n [k]
,
r = 1, 2, . . . ¯τ .
(14)
(15)
,
0
0
...
In×n
0
,
As before, P0[k], P1[k], . . . , P¯τ [k] are appropriately defined nonnegative matrices, such that
¯τ(cid:88)
P [k] =
Pr[k],
i.e., the sum P [k] of all the nonnegative matrices Pr[k], r ∈ {0, 1, 2, . . . , ¯τ}, gives the weights of
the zero-delay interconnection topology at time instant k. The difference from the case when only
r=0
delays are present in the network is that the interconnection topology is dynamically changing
and the weights at each time instant might differ (mathematically, this means that the left matrix
in the above equation changes with k). The proposed protocol is able to asymptotically reach
average consensus, as stated in Lemma 3 below. The proof is similar to the proof for delays
with no changes in the interconnection topology and is omitted.
Lemma 3. Consider a sequence of graphs of the form G[k] = (V,E[k])), k = 0, 1, 2, ..., such that
there exists an infinite sequence of time instants t0, t1, . . . , tm, . . ., where t0 = 0, 0 < tm+1−tm ≤
(cid:96) < ∞, m ∈ Z+, with the property that for any m ∈ Z+ the union of graphs G[tm],G[tm +
1], . . . ,G[tm+1 − 1] is strongly connected. Let yj[k] for all vj ∈ V be the result of iteration (13)
j [k] ∪ {vj} (zeros otherwise) and initial conditions y[0] = y0,
with plj[k] =
j [k]∪{vj}
and let zj[k], ∀vj ∈ V, be the result of iteration (12) with plj[k] =
(zeros otherwise) and with initial condition z[0] = 1. The indicator function Ik,ji captures
the bounded delay τji[k] on link (vj, vi) at iteration k (as defined in (6), τji[k] ≤ ¯τ). Then,
the solution to the average consensus problem can be asymptotically obtained as lim
k→∞ µj[k] =
for vl ∈ N +
1
1+D+
j [k]
for vl ∈ N +
1
1+D+
j [k]
(cid:80)
vl∈V y0(l)
V
, ∀vj ∈ V , where µj[k] = yj [k]
zj [k].
We now discuss the case in which a node, say node vj, receives an indication (e.g., an
j [k− 1], no longer receives
acknowledgement message) that one of its out-neighbors, say vl ∈ N +
its transmissions. In other words, node vl /∈ N +
j [k] but node vj finds out about it with some
bounded delay that we denote by Tlj[k]. Such bounded delays could arise from communication
protocols in a variety of ways, e.g., when using periodic acknowledgement signals like the
distress signals discussed in Remark 3. Another way for node vj to discover that its out-degree
has changed is by using acknowledgement signals that arrive at node vj through paths in the
network that connect the out-neighbors of node vj to node vj (note that a direct link between
an out-neighbor of node vj and node vj may not necessarily exist in a digraph).
We use Tlj[k] ≤ T < ∞ to denote the time it takes for node vj to realize that node vl is no
longer in the set N +
j [k]. The problem that node vj has to address at time k + Tlj[k] when it
realizes that, at time steps k, k + 1, . . . , k + Tlj[k− 1], it erroneously assumed an out-degree D+
je
that (supposing no other changes) was greater than the true out-degree D+
j , is that the weighted
message from node vj that was not eventually conveyed to the out-neighbor vl (because the link
was terminated) needs to be accounted for. The most straightforward way to reconcile this is
to add these values back to node vj. This can be done easily as long as node vj keeps track
of the messages it has recently transmitted -- within the last T steps -- to its out-neighbors.
Note, however, that node vj has to track these messages for each of its perceived out-neighbors,
because if more than one out-neighbors terminate their links with vj (possibly at different time
steps), then node vj must know what needs to be added back to its own value for each such
former out-neighbor. One way to think about this in terms of the augmented digraph is that node
vj adds, for each of its perceived out-neighbors, T "virtual" nodes that loop back to itself. These
virtual nodes essentially keep track of the values that have been sent to each out-neighbor in the
last T steps. The following example discusses this issue in more detail.
Consider, for example, node vj with two out-neighbors (D+
j = 2) shown in Figure 7(a).
Suppose that the maximum delay required for an acknowledgement signal from any out-neighbor
of node vj is 2 (i.e., T = 2). Then, the model for the part of the network consisting of node
vj and its out-neighbors vl1, vl2 is as shown in Figure 7(a). Suppose now that out-neighbor vl2
terminates the link vl2 ← vj and node vj receives an ACK with delay 2. This means that node
vj erroneously considered an out-degree of 2 (instead of 1) for the last 2 updates. In this model,
the message is passed through the two extra "virtual" nodes (added in a self-loop), allowing us
to loop the weighted message back to node vj (see Figure 7(b)). Therefore, node vj is able to
recover the lost values (sent to node vl2 that is no longer an out-neighbor). It is worth pointing
out at this point that, unlike the previous case of fixed interconnections with bounded delays,
the virtual nodes are no longer simply a question of modeling; in fact, in this example, we
have to ensure that node vj essentially implements the functionality of the virtual buffers (by
remembering the messages that it has sent to its out-neighbors.
Example 3. Consider the simple network with 3 nodes in Figure 8 and the following scenario:
at instant k1 the network of the three nodes has no delays or interconnection topology changes;
(a) A node vj and its two out-neighbors. The maximum
delay required for an acknowledgement signal from the out-
neighbors of node vj is 2.
(b) A model in which node vj directs the weighted messages
of the links that no longer exist as delayed information to
itself. In this example, node vl2 terminated the link and the
ACK arrived at node vj with maximum delay 2.
Fig. 7. The weighted messages from node vj that were not conveyed to the out-neighbor vl2 (because the link was terminated)
are added back to the value of node vj.
at time instant k2, link v3 ← v1 terminates, and an acknowledgement is sent to node v1 via
node v2, from which there exists a link to node v1. In addition, any message from v2 to v1 can
be delayed by at most 1 iteration. For simplicity, we assume that the delays in all other links
Fig. 8. A model in which link v3 ← v1 terminates, and an acknowledgement is send to node v1 via node v2, from which
there exists a direct link to node v1. In addition, the message from v2 to v1 is delayed by 1.
are zero and the connection with link v2 cannot be lost; hence, no additional loops need to be
inserted. Taking x[k] = (x1[k] x2[k] x3[k] x(1)
3 [k])T , the matrix representations
1 [k] x(1)
3 [k] x(2)
vl1vl2vjpl1jpl2jpjjv(1)l1v(2)l11pl1jpl1j1v(1)l2v(2)l21pl2jpl2j1vl1vl2vjpl1jpl2jpjjv(1)l1v(2)l11pl1jpl1j1v(1)l2v(2)l21pl2jpl2j1v2v3v1p21p23p31p12p11p22p33v(1)3v(2)31p31p311v(1)11p12 at time instances k1 and k2 are captured by
P [k1] =
p11
p12
0
1 1
p21
p22
p23
0 0
p31
0
0
0
0
0
0
0
p33
0 0
0
0
0
0 0
0 0
0 0
0
0
0
0
1
0
, P [k2] =
p11
0
0
p21
p22
p23
0
0
0
p31
0
p33
p12
0
0
0
0
0
1
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
1
0
.
In the general case, in a network of n = V nodes, we introduce max(τ n, T n) nodes (for a
total of (max(τ n, T n) + n) nodes) and we proceed as in (14).
Remark 4. There are also cases in which the transmitting node vj may not have knowledge
of its out-degree at time instant k. Such situations can also be handled if, at each time instant
k, node vj (i) knows the number of nodes with which it has established a communication link
in the past and have not been officially terminated yet, and (ii) is able to multicast a table
of values to each of these out-neighbors. One way to do this is to employ the communication
protocol proposed in [37] where, at each time instant, each node vj broadcasts its own state
(as updated via the iterations in equation (12)), as well as the sum of all the values, called the
total mass in [37], that have been broadcasted to each neighboring node vl ∈ N +
j so far. If, for
any reason, some messages are lost (dropped) or the communication link disappears for some
time-period, the total mass will enable the receiving out-neighbor to retrieve the information of
the lost messages, with some time-delay. Thus, even though the communication links may not be
reliable and can even change, the problem boils down to dealing with delayed information (as
in Section IV). Note, however, that each node vj needs to keep track of its own current state, the
total mass transmitted to each neighboring node vl ∈ N +
j [k] (the total mass can be different for
each node vl due to, for example, newly established communication links), and the total mass
received from each neighboring node vi ∈ N −
j [k] that transmits information to node vj. Since
different information might need to be transmitted to each node vl at each time instant k, node
vj is required to broadcast a table of values with entries for each receiving node.
Example 4. We illustrate how the algorithm operates via a small network of six nodes. Each
j [k])−1 (such that the sum of all weights plj[k], vl ∈ N +
node vj chooses its self-weight and the weight of its outgoing links at each time instant k to
be (1 + D+
j [k] ∪ {vj}, assigned by each
node vj to links to its out-neighbors at time step k is equal to 1). First, suppose the nodes
experience only changes in interconnection topology but no delays. When each node updates
its information state xj[k] using equation (12), the information state for the whole network is
given by x[k + 1] = P [k]x[k], where P [k] depends on the links present at time instant k. For
example, at time instants k = k1 and k = k2, whose interconnection topologies are captured by
the graphs in Figures 9(a) and 9(b), matrices P [k1] and P [k2], respectively, are given by
P [k1] =
1/4
1/2
0
1/4 1/2 1/4
1/4
1/4
0
0
0
0
0
0
0
0
0
1/2
0
0
0
0
0
0
1/3
1/4
1/4
1/2
0
0
1/2
1/3
1/4
1/2
0
1/3
, P [k2] =
.
1/3
1/3
0
1/3
1/3
1/5
1/3
0
1/5
0
0
0
1/3
1/5
1/2
1/2
0
0
0
0
0
0
0
0
0
1/5
0
1/2
1/2
1/5
1/2
0
1/2
0
0
0
We use twice the update formula (12) with initial conditions y[0] = (−1 1 2 3 4 3)T and
z[0] = 1T respectively. In our simulations, we generate at each iteration a new random graph
with six nodes that includes a directed link (vj, vi) from node vi to node vj (vi, vj ∈ V , vi (cid:54)= vj)
with some probability p independently between different links. Note that once the graph is chosen
at iteration k, the update matrix P [k] will be column stochastic. A realization of the ratios at
each node is shown in Figure 10(a); in this case, the average is 2. When delays are present with
maximum delay ¯τ = 5 we use the update formula (13) with the same initial conditions and we
observe that the system again converges to the exact average (as shown in Figure 10(b)), but
with a slower convergence.
Remark 5. In many settings, it might be more natural for the communication protocol to allow
the receiver to set the weights of the incoming values, since it is easier for each receiving node to
know from which (and how many) nodes it has received a message. Indeed, consensus in multi-
(a) Connections and weights at instant k1.
(b) Connections and weights at instant k2.
Fig. 9. A network of six nodes, where each node vj chooses its self-weight and the weight of the links to its out-neighbors
to be (1 + D+
j [k])−1.
(a) Ratios at each node for changing interconnection topol-
ogy and no delays.
(b) Ratios at each node for changing interconnection topol-
ogy with delays.
Fig. 10. We use twice the update formula (12) with initial conditions y[0] = (−1 1 2 3 4 3)T and z[0] = 1T respectively,
and plot the ratio yj[k]/zj[k] for each node vj under changing interconnection topology but no delays (left). When delays are
present with maximum delay ¯τ = 5 we use the update formula (13) with the same initial conditions and observe that the ratios
again converge to the average, but with a slower convergence (right).
agent systems in digraphs in the presence of changing interconnection topology and time-varying
delays have been studied in [17], which provided sufficient conditions for the multi-agent system
to reach consensus when using a protocol that relies on a single iteration and uses weights that
v1v2v3v4v5v61/41/41/21/21/31/21/41/21/41/41/41/41/21/21/31/3v1v2v3v4v5v61/31/51/21/21/21/31/31/41/31/31/51/51/51/51/21/21/2051015202530−1−0.500.511.522.533.54Values at each node vs Number of iterationsNumber of iterationsValue at each node051015202530−1−0.500.511.522.533.54Values at each node vs Number of iterationsNumber of iterationsValue at each node form a row stochastic matrix (compared to the two iterations and weights that form column
stochastic matrices proposed in this paper). Event though [17] reaches consensus, it does not
necessarily reach consensus to the exact average of the initial values of the nodes; in fact, the
value to which this approach (and others that rely on a single iteration and weights that form
row stochastic matrices) converge depends on the delay magnitude and profile.
Technically, our approach relies on weak convergence of the backward product of column
stochastic matrices whereas [17] and others rely on strong convergence of the backward product
of row stochastic matrices (which is equivalent to a forward product of column stochastic
matrices4). This means that in approaches that depend on row stochastic matrices the product of
matrices converges to a rank one matrix (which would have to be equal to 1
n
11T if convergence
to the average of the initial values is desirable); on the contrary, in the proposed approach
we do not have convergence of the product but we have instead weak convergence to rank one
matrices. This means that for a large number of iterations, each matrix product gets closer to
a rank one matrix but this rank one matrix is not necessarily the same for each iteration step.
However, by running two iterations and focusing on the ratio of the two iteration values, we
exploit weak convergence and are able to obtain the exact average.
VI. CONCLUSIONS
In this paper, we studied distributed strategies for a discrete-time networked system to reach
asymptotic average consensus in the presence of time-delays and dynamically changing topolo-
gies. By assuming that nodes in the multi-agent system have knowledge of their out-degree (i.e.,
the number of nodes to which they send information to) and by modeling the time-delays using
an augmented graphical model, we have shown that our proposed discrete-time strategy reaches
asymptotic average consensus in a distributed fashion, in the presence of dynamically changing
interconnection topology for whatever the realization of delays, as long as they are bounded
and the union graph of the graph topologies over consecutive time intervals forms a strongly
4Note that weak and strong convergence are equivalent for forward products of column stochastic matrices (see Theorem 4.17
in the book of Seneta "Non-negative Matrices and Markov Chains").
connected graph infinitely often.
REFERENCES
[1] C. N. Hadjicostis and T. Charalambous, "Asynchronous coordination of distributed energy resources for the provisioning
of ancillary services," in Proceedings of 47th Annual Allerton Conference on Communication, Control, and Computing,
Sep. 2011, pp. 1500 -- 1507.
[2] R. Olfati-Saber, J. Fax, and R. Murray, "Consensus and cooperation in networked multi-agent systems," Proceedings of
the IEEE, vol. 95, no. 1, pp. 215 -- 233, Jan. 2007.
[3] A. Giridhar and P. R. Kumar, "Toward a theory of in-network computation in wireless sensor networks," IEEE
Communications Magazine, vol. 44, no. 4, pp. 98 -- 107, April 2006.
[4] J. Tsitsiklis, D. Bertsekas, and M. Athans, "Distributed asynchronous deterministic and stochastic gradient optimization
algorithms," IEEE Transactions on Automatic Control, vol. 31, no. 9, pp. 803 -- 812, Sep. 1986.
[5] R. Olfati-Saber and R. Murray, "Consensus problems in networks of agents with switching topology and time-delays,"
IEEE Transactions on Automatic Control, vol. 49, no. 9, pp. 1520 -- 1533, Sep. 2004.
[6] W. Ren and R. Beard, "Consensus seeking in multiagent systems under dynamically changing interaction topologies,"
IEEE Transactions on Automatic Control, vol. 50, no. 5, pp. 655 -- 661, May 2005.
[7] L. Moreau, "Stability of multiagent systems with time-dependent communication links," IEEE Transactions on Automatic
Control, vol. 50, no. 2, pp. 169 -- 182, Feb. 2005.
[8] D. Angeli and P. Bliman, "Stability of leaderless discrete-time multi-agent systems," Mathematics of Control, Signals, and
Systems, vol. 18, no. 4, pp. 293 -- 322, July 2006.
[9] P. Bliman and G. Ferrari-Trecate, "Average consensus problems in networks of agents with delayed communications,"
Automatica, vol. 44, no. 8, pp. 1985 -- 1995, Aug. 2008.
[10] M. Franceschelli, A. Giua, and C. Seatzu, "A gossip-based algorithm for discrete consensus over heterogeneous networks,"
IEEE Transactions on Automatic Control,, vol. 55, no. 5, pp. 1244 -- 1249, May 2010.
[11] A. D. Dom´ınguez-Garc´ıa and C. N. Hadjicostis, "Distributed strategies for average consensus in directed graphs," IEEE
Transactions on Automatic Control, 2012 (to appear).
[12] K. Cai and H. Ishii, "Average consensus on general digraphs," in Proceedings of the 50th IEEE Conference on Decision
and Control and European Control Conference, Dec. 2011, pp. 1956 -- 1961.
[13] D. Varagnolo, G. Pillonetto, and L. Schenato, "Distributed parametric and nonparametric regression with on-line
performance bounds computation," Automatica, vol. 48, no. 10, pp. 2468 -- 2481, October 2012.
[14] L. Fang and P. Antsaklis, "Information consensus of asynchronous discrete-time multi-agent systems," in Proceedings of
American Control Conference, June 2005, pp. 1883 -- 1888.
[15] S. Patterson, B. Bamieh, and A. El Abbadi, "Distributed average consensus with stochastic communication failures," in
Proceedings of the 46th IEEE Conference on Decision and Control, Dec. 2007, pp. 4215 -- 4220.
[16] F. Fagnani and S. Zampieri, "Average consensus with packet drop communication," SIAM J. Control Optim., vol. 48, no. 1,
pp. 102 -- 133, Feb. 2009.
[17] F. Xiao and L. Wang, "State consensus for multi-agent systems with switching topologies and time-varying delays,"
International Journal of Control, vol. 79, no. 10, pp. 1277 -- 1284, Oct. 2006.
[18] A. Nedic, A. Ozdaglar, and P. Parrilo, "Constrained consensus and optimization in multi-agent networks," IEEE Transactions
on Automatic Control, vol. 55, no. 4, pp. 922 -- 938, April 2010.
[19] A. Jadbabaie, J. Lin, and A. Morse, "Coordination of groups of mobile autonomous agents using nearest neighbor rules,"
IEEE Transactions on Automatic Control, vol. 48, no. 6, pp. 988 -- 1001, June 2003.
[20] L. Xiao and S. Boyd, "Fast linear iterations for distributed averaging," Systems and Control Letters, vol. 53, no. 1, pp.
65 -- 78, Sep. 2004.
[21] B. Johansson, M. Rabi, and M. Johansson, "A randomized incremental subgradient method for distributed optimization in
networked systems," SIAM J. on Optimization, vol. 20, no. 3, pp. 1157 -- 1170, Aug. 2009.
[22] F. Zanella, D. Varagnolo, A. Cenedese, G. Pillonetto, and L. Schenato, "Newton-Raphson consensus for distributed convex
optimization," in Proceedings of the 50th IEEE Conference on Decision and Control and European Control Conference,
Dec. 2011, pp. 5917 -- 5922.
[23] -- -- , "Asynchronous Newton-Raphson consensus for distributed convex optimization," in Proceedings of 3rd IFAC
Workshop on Distributed Estimation and Control in Networked Systems (NecSys'12), Sep. 2012, pp. 133 -- 138.
[24] B. Gharesifard and J. Cort´es, "Distributed strategies for making a digraph weight-balanced," in Proceedings of the 47th
Annual Allerton Conference on Communication, Control, and Computing, Sep. 2009, pp. 771 -- 777.
[25] C. N. Hadjicostis and A. Rikos, "Distributed strategies for balancing a weighted digraph," in Proceedings of
the
Mediterranean Conference on Control and Automation, July 2012, pp. 1141 -- 1146.
[26] B. Gharesifard and J. Cort´es, "When does a digraph admit a doubly stochastic adjacency matrix?" in Proceedings of the
American Control Conference, July 2010, pp. 2440 -- 2445.
[27] A. D. Dom´ınguez-Garc´ıa and C. N. Hadjicostis, "Coordination and control of distributed energy resources for provision
of ancillary services," in Proceedings of the First IEEE International Conference on Smart Grid Communications, Oct.
2010, pp. 537 -- 542.
[28] M. Franceschelli, A. Giua, and C. Seatzu, "Distributed averaging in sensor networks based on broadcast gossip algorithms,"
IEEE Sensors Journal, vol. 11, no. 3, pp. 808 -- 817, March 2011.
[29] K. Cai and H. Ishii, "Average consensus on general strongly connected digraphs," Automatica, vol. 48, no. 11, pp. 2750 --
2761, Nov. 2012.
[30] F. B´en´ezit, V. Blondel, P. Thiran, J. Tsitsiklis, and M. Vetterli, "Weighted gossip: Distributed averaging using non-doubly
stochastic matrices," in Proceedings of the IEEE International Symposium on Information Theory, June 2010, pp. 1753 --
1757.
[31] D. Kempe, A. Dobra, and J. Gehrke, "Gossip-based computation of aggregate information," in Proceedings of the 44th
Annual IEEE Symposium on Foundations of Computer Science, ser. FOCS '03, 2003, pp. 482 -- 491.
[32] E. Seneta, Non-negative Matrices and Markov Chains, 2nd ed. New York, NY: Springer, 2006.
[33] S. M. Amin and B. Wollenberg, "Toward a smart grid: Power delivery for the 21st century," IEEE Power and Energy
Magazine, vol. 3, no. 5, pp. 34 -- 41, Sept.-Oct. 2005.
[34] J. Wolfowitz, "Products of indecomposable, aperiodic, stochastic matrices," Proceedings of the American Mathematical
Society, vol. 15, no. 5, pp. 733 -- 737, October 1963.
[35] B. Lubachevsky and D. Mitra, "A chaotic asynchronous algorithm for computing the fixed point of a nonnegative matrix
of unit spectral radius," Journal of the ACM, vol. 33, no. 1, pp. 130 -- 150, Jan. 1986.
[36] N. Bambos, S. C. Chen, and G. J. Pottie, "Channel access algorithms with active link protection for wireless communication
networks with power control," IEEE/ACM Transactions on Networking, vol. 8, no. 5, pp. 583 -- 597, May 2000.
[37] N. H. Vaidya, C. N. Hadjicostis, and A. D. Dom´ınguez-Garc´ıa, "Robust average consensus over packet dropping links:
Analysis via coefficients of ergodicity," in Proceedings of 51st IEEE Conference on Decision and Control, Dec. 2012, pp.
2761 -- 2766.
|
1603.00939 | 2 | 1603 | 2016-07-29T18:42:59 | Routing Autonomous Vehicles in Congested Transportation Networks: Structural Properties and Coordination Algorithms | [
"cs.MA"
] | This paper considers the problem of routing and rebalancing a shared fleet of autonomous (i.e., self-driving) vehicles providing on-demand mobility within a capacitated transportation network, where congestion might disrupt throughput. We model the problem within a network flow framework and show that under relatively mild assumptions the rebalancing vehicles, if properly coordinated, do not lead to an increase in congestion (in stark contrast to common belief). From an algorithmic standpoint, such theoretical insight suggests that the problem of routing customers and rebalancing vehicles can be decoupled, which leads to a computationally-efficient routing and rebalancing algorithm for the autonomous vehicles. Numerical experiments and case studies corroborate our theoretical insights and show that the proposed algorithm outperforms state-of-the-art point-to-point methods by avoiding excess congestion on the road. Collectively, this paper provides a rigorous approach to the problem of congestion-aware, system-wide coordination of autonomously driving vehicles, and to the characterization of the sustainability of such robotic systems. | cs.MA | cs | Routing Autonomous Vehicles in Congested Transportation
Networks: Structural Properties and Coordination Algorithms
Rick Zhang,* Federico Rossi,* and Marco Pavone
the same). Indeed, it has been argued that the presence of
many rebalancing vehicles may contribute to an increase in
congestion [6], [7]. These statements, however, do not take
into account that in an AMoD system the operator has control
over the actions (destination and routes) of the vehicles, and
may route vehicles intelligently to avoid increasing congestion
or perhaps even decrease it.
6
1
0
2
l
u
J
9
2
]
A
M
.
s
c
[
2
v
9
3
9
0
0
.
3
0
6
1
:
v
i
X
r
a
Abstract-This paper considers the problem of routing and
rebalancing a shared fleet of autonomous (i.e., self-driving)
vehicles providing on-demand mobility within a capacitated trans-
portation network, where congestion might disrupt throughput.
We model the problem within a network flow framework and
show that under relatively mild assumptions the rebalancing
vehicles,
if properly coordinated, do not lead to an increase
in congestion (in stark contrast to common belief). From an
algorithmic standpoint, such theoretical insight suggests that the
problem of routing customers and rebalancing vehicles can be
decoupled, which leads to a computationally-efficient routing and
rebalancing algorithm for the autonomous vehicles. Numerical
experiments and case studies corroborate our theoretical insights
and show that the proposed algorithm outperforms state-of-the-
art point-to-point methods by avoiding excess congestion on the
road. Collectively, this paper provides a rigorous approach to
the problem of congestion-aware, system-wide coordination of
autonomously driving vehicles, and to the characterization of
the sustainability of such robotic systems.
I. INTRODUCTION
Autonomous (i.e., robotic, self-driving) vehicles are rapidly
becoming a reality and hold great promise for increasing
safety and enhancing mobility for those unable or unwilling
to drive [1], [2]. A particularly attractive operational paradigm
involves coordinating a fleet of autonomous vehicles to pro-
vide on-demand service to customers, also called autonomous
mobility-on-demand (AMoD). An AMoD system may reduce
the cost of travel [3] as well as provide additional sustainability
benefits such as increased overall vehicle utilization, reduced
demand for urban parking infrastructure, and reduced pollution
(with electric vehicles) [1]. The key benefits of AMoD are
realized through vehicle sharing, where each vehicle, after
servicing a customer, drives itself to the location of the
next customer or rebalances itself throughout
the city in
anticipation of future customer demand [4].
In terms of traffic congestion, however, there has been no
consensus on whether autonomous vehicles in general, and
AMoD systems in particular, will ultimately be beneficial or
detrimental. It has been argued that by having faster reaction
times, autonomous vehicles may be able to drive faster and
follow other vehicles at closer distances without compromising
safety, thereby effectively increasing the capacity of a road
and reducing congestion. They may also be able to interact
with traffic lights to reduce full stops at intersections [5].
On the downside, the process of vehicle rebalancing (empty
vehicle trips) increases the total number of vehicles on the
road (assuming the number of vehicles with customers stays
*These authors contributed equally to this work.
This research was supported by National Science Foundation under CA-
REER Award CMMI-1454737 and by the Dr. Cleve B. Moler Stanford
Graduate Fellowship. Rick Zhang, Federico Rossi, and Marco Pavone are with
the Department of Aeronautics & Astronautics, Stanford University, Stanford,
CA 94305 {rickz, frossi2, pavone}@stanford.edu
Accordingly, the goal of this paper is twofold. First, on an
engineering level, we aim to devise routing and rebalancing
algorithms for an autonomous vehicle fleet that seek to mini-
mize congestion. Second, on a socio-economic level, we aim to
rigorously address the concern that autonomous cars may lead
to increased congestion and thus disrupt currently congested
transportation infrastructures.
Literature review: In this paper, we investigate the problem
of controlling an AMoD system within a road network in
the presence of congestion effects. Previous work on AMoD
systems have primarily concentrated on the rebalancing prob-
lem [4], [3], whereby one strives to allocate empty vehicles
throughout a city while minimizing fuel costs or customer
wait times. The rebalancing problem has been studied in [4]
using a fluidic model and in [8] using a queueing network
model. An alternative formulation is the one-to-one pickup and
delivery problem [9], where a fleet of vehicles service pickup
and delivery requests within a given region. Combinatorial
asymptotically optimal algorithms for pickup and delivery
problems were presented in [10], [11], and generalized to
road networks in [12]. Almost all current approaches assume
point-to-point travel between origins and destinations (no road
network), and even routing problems on road networks (e.g.
[12]) do not take into account vehicle-to-vehicle interactions
that would cause congestion and reduce system throughput.
On the other hand, traffic congestion has been studied in
economics and transportation for nearly a century. The first
congestion models [13], [14], [15] sought to formalize the rela-
tionship between vehicle speed, density, and flow. Since then,
approaches to modeling congestion have included empirical
[16], simulation-based [17], [18], [19], queueing-theoretical
[20], and optimization [21], [22]. While there have been many
high fidelity congestion models that can accurately predict
traffic patterns, the primary goal of congestion modeling has
been the analysis of traffic behavior. Efforts to control traffic
have been limited to the control of intersections [23], [24]
and freeway on-ramps [25] because human drivers behave
non-cooperatively. The problem of cooperative, system-wide
routing (a key benefit of AMoD systems) is similar to the
dynamic traffic assignment problem (DTA) [22] and to [26],
[27] in the case of online routing. The key difference is that
these approaches only optimize routes for passenger vehicles
while we seek to optimize the routes of both passenger
vehicles and empty rebalancing vehicles.
Statement of contributions: The contribution of this paper is
threefold. First, we model an AMoD system within a network
flow framework, whereby customer-carrying and empty rebal-
ancing vehicles are represented as flows over a capacitated
road network (in such model, when the flow of vehicles along
a road reaches a critical capacity value, congestion effects
occur). Within this model, we provide a cut condition for
the road graph that needs to be satisfied for congestion-free
customer and rebalancing flows to exist. Most importantly,
under the assumption of a symmetric road network, we inves-
tigate an existential result that leads to two key conclusions:
(1) rebalancing does not
increase congestion, and (2) for
certain cost functions, the problems of finding customer and
rebalancing flows can be decoupled. Second, leveraging the
theoretical insights, we propose a computationally-efficient
algorithm for congestion-aware routing and rebalancing of
an AMoD system that is broadly applicable to time-varying,
possibly asymmetric road networks. Third, through numerical
studies on real-world traffic data, we validate our assumptions
and show that the proposed real-time routing and rebalancing
algorithm outperforms state-of-the-art point-to-point rebalanc-
ing algorithms in terms of lower customer wait
times by
avoiding excess congestion on the road.
Organization: The remainder of this paper is organized as
follows: in Section II we present a network flow model of an
AMoD system on a capacitated road network and formulate
the routing and rebalancing problem. In Section III we present
key structural properties of the model including fundamental
limitations of performance and conditions for the existence of
feasible (in particular, congestion-free) solutions. The insights
from Section III are used to develop a practical real-time
routing and rebalancing algorithm in Section IV. Numerical
studies and simulation results are presented in Section V, and
in Section VI we draw conclusions and discuss directions for
future work.
II. MODEL DESCRIPTION AND PROBLEM FORMULATION
In this section we formulate a network flow model for an
AMoD system operating over a capacitated road network. The
model allows us to derive key structural insights into the vehi-
cle routing and rebalancing problem, and motivates the design
of real-time, congestion-aware algorithms for coordinating the
robotic vehicles. We start in Section II-A with a discussion
of our congestion model; then, in Section II-B we provide a
detailed description of the overall AMoD system model.
A. Congestion Model
We use a simplified congestion model consistent with clas-
sical traffic flow theory [13]. In classical traffic flow theory,
at low vehicle densities on a road link, vehicles travel at the
free flow speed of the road (imposed by the speed limit). This
is referred to as the free flow phase of traffic. In this phase,
the free flow speed is approximately constant [28]. The flow,
or flow rate, is the number of vehicles passing through the
link per unit time, and is given by the product of the speed
and density of vehicles. When the flow of vehicles reaches
an empirically observed critical value, the flow reaches its
maximum. Beyond the critical flow rate, vehicle speeds are
dramatically reduced and the flow decreases, signaling the
beginning of traffic congestion. The maximum stationary flow
rate is called the capacity of the road link in the literature.
In our approach, road capacities are modeled as constraints
on the flow of vehicles. In this way, the model captures the
behavior of vehicles up to the onset of congestion.
is not
This simplified congestion model is adequate for our pur-
poses because the goal
to analyze the behavior of
vehicles in congested networks, but to control vehicles in
order to avoid the onset of congestion. We also do not
explicitly model delays at intersections, spillback behavior due
to congestion, or bottleneck behavior due to the reduction of
the number of lanes on a road link. An extension to our model
that accommodates (limited) congestion on links is presented
in Section V-A.
B. Network Flow Model of AMoD system
We consider a road network modeled as a directed graph
G = (V,E), where V denotes the node set and E ⊆ V × V
denotes the edge set. Figure 1 shows one such network. The
nodes v in V represent intersections and locations for trip
origins/destinations, and the edges (u, v) in E represent road
links. As discussed in Section II-A, congestion is modeled
by imposing capacity constraints on the road links: each
constraint represents the capacity of the road upon the onset
of congestion. Specifically, for each road link (u, v) ∈ E, we
denote by c(u, v) : E (cid:55)→ N>0 the capacity of that link. When
the flow rate on a road link is less than the capacity of the
link, all vehicles are assumed to travel at the free flow speed,
or the speed limit of the link. For each road link (u, v) ∈ E,
we denote by t(u, v) : E (cid:55)→ R≥0 the corresponding free flow
time required to traverse road link (u, v). Conversely, when the
flow rate on a road link is larger than the capacity of the link,
the traversal time is assumed equal to ∞ (we reiterate that our
focus in this section is on avoiding the onset of congestion).
We assume that the road network is capacity-symmetric
(or symmetric for short): for any cut1 (S, ¯S) of G(V,E), the
overall capacity of the edges connecting nodes in S to nodes
in ¯S equals the overall capacity of the edges connecting nodes
in ¯S to nodes in S, that is
(cid:88)
(cid:88)
c(u, v) =
c(v, u)
(u,v)∈E: u∈S, v∈ ¯S
(v,u)∈E: u∈S, v∈ ¯S
It is easy to verify that a network is capacity-symmetric if
and only if the overall capacity entering each node equals the
capacity exiting each node., i.e.
c(u, v) =
(cid:88)
(cid:88)
c(v, w)
u∈V:(u,v)∈E
w∈V:(v,w)∈E
If all edges have symmetrical capacity, i.e., for all (u, v) ∈ E,
c(u, v) = c(v, u), then the network is capacity-symmetric. The
converse statement, however, is not true in general.
Transportation requests are described by the tuple (s, t, λ),
where s ∈ V is the origin of the requests, t ∈ V is the
destination, and λ ∈ R>0 is the rate of requests, in customers
per unit
time. Transportation requests are assumed to be
stationary and deterministic, i.e., the rate of requests does not
change with time and is a deterministic quantity. The set of
1For any subset of nodes S ⊆ V, we define a cut (S, ¯S) ⊆ E as the set
of edges whose origin lies in S and whose destination lies in ¯S = {V \ S}.
Formally, (S, ¯S) := {(u, v) ∈ E : u ∈ S, v ∈ ¯S}.
with transportation requests, i.e. to ensure that every outbound
customer flow is matched by an inbound flow of rebalancing
vehicles and vice versa. Finally, constraint (6) enforces the
capacity constraint on each link (function 1x denotes the
indicator function of the Boolean variable x = {true, false},
that
is 1x equals one if x is true, and equals zero if x
is false). Note that the CRRP is a linear program and, in
particular, a special instance of the fractional multi-commodity
flow problem [29].
We denote a customer flow {fm(u, v)}(u,v),m that satisfies
Equations (2), (3), (4) and (6) as a feasible customer flow.
For a given set of feasible customer flows {fm(u, v)}(u,v),m,
we denote a flow {fR(u, v)}(u,v) that satisfies Equation (5)
and such that the combined flows {fm(u, v), fR(u, v)}(u,v),m
satisfy Equation (6) as a feasible rebalancing flow. We remark
that a rebalancing flow that is feasible with respect to a set of
customer flows may be infeasible for a different collection of
customer flows.
For a given set of optimal flows {f∗
m(u, v)}(u,v),m and
{f∗
R(u, v)}(u,v), the minimum number of vehicles needed to
implement them is given by
(cid:88)
(cid:88)
m∈M
(u,v)∈E
(cid:16)
Vmin =
t(u, v)
∗
m(u, v) + f
∗
R(u, v)
f
(cid:17) .
This follows from a similar analysis done in [4] for point-to-
point networks. Hence, the cost function (1) is aligned with
the desire of minimizing the number of vehicles needed to
operate an AMoD system.
D. Discussion
A few comments are in order. First, we assume that trans-
portation requests are time invariant. This assumption is valid
when transportation requests change slowly with respect to
the average duration of a customer's trip, which is often
the case in dense urban environments [30]. Additionally, in
Section IV we will present algorithmic tools that allow one
to extend the insights gained from the time-invariant case
to the time-varying counterpart. Second, the assumption of
single-occupancy for the vehicles models most of the existing
(human) one-way vehicle sharing systems (where the driver
is considered "part" of the vehicle), and chiefly disallows
the provision of ride-sharing or carpooling service (this is an
aspect left for future research). Third, as also discussed in
Section II-A, our congestion model is simpler and less accurate
than typical congestion models used in the transportation
community. However, our model lends itself to efficient real-
time optimization and thus it is well-suited to the control of
fleets of autonomous vehicles. Existing high-fidelity conges-
tion models should be regarded as complementary and could
be used offline to identify the congestion thresholds used in our
model. Fourth, while we have defined the CRRP in terms of
fractional flows, an integer-valued counterpart can be defined
and (approximately) solved to find optimal routes for each
individual customer and vehicle. Algorithmic aspects will be
investigated in depth in Section IV, with the goal of devising
practical, real-time routing and rebalancing algorithms. Fifth,
trip requests are assumed to be known. In practice,
trip
requests can be reserved in advance, estimated from historical
data, or estimated in real time. Finally, the assumption of
Fig. 1.
A road network modeling Lower Manhattan and the Financial
District. Nodes (denoted by small black dots) model intersections; select
nodes, denoted by colored circular and square markers, model passenger trips'
origins and destinations. Different trip requests are denoted by different colors.
Roads are modeled as edges; line thickness is proportional to road capacity.
transportation requests is denoted by M = {(sm, tm, λm)}m,
and its cardinality is denoted by M.
Single-occupancy vehicles travel within the network while
servicing the transportation requests. We denote fm(u, v) :
E (cid:55)→ R≥0, m = {1, . . . , M}, as the customer flow for
requests m on edge (u, v), i.e., the amount of flow from origin
sm to destination tm that uses link (u, v). We also denote
fR(u, v) : E (cid:55)→ R≥0 as the rebalancing flow on edge (u, v),
i.e., the amount of rebalancing flow traversing edge (u, v)
needed to realign the vehicles with the asymmetric distribution
of transportation requests.
C. The Routing Problem
The goal is to compute flows for the autonomous vehicles
that (i) transfer customers to their desired destinations in
minimum time (customer-carrying trips) and (ii) rebalance
vehicles throughout the network to realign the vehicle fleet
with transportation demand (customer-empty trips). Specifi-
cally, the Congestion-free Routing and Rebalancing Problem
(CRRP) is formally defined as follows. Given a capacitated,
symmetric network G(V,E), a set of transportation requests
M = {(sm, tm, λm)}m, and a weight factor ρ > 0, solve
(cid:88)
(u,v)∈E
fm(·,·),fR(·,·)
u∈V
m∈M
minimize
(cid:88)
subject to (cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
u∈V
u∈V
u∈V
=
t(u, v)fm(u, v)+ρ
t(u, v)fR(u, v)
fm(u, sm) + λm =
(1)
fm(sm, w) ∀m ∈ M (2)
fm(u, tm) = λm +
fm(tm, w) ∀m ∈ M (3)
(cid:88)
(cid:88)
w∈V
w∈V
(cid:88)
(cid:88)
w∈V
m∈M
(cid:88)
m∈M
fm(u, v) =
fm(v, w)
∀m ∈ M, v ∈ V \ {sm, tm}
fR(u, v) +
1v=tm λm
w∈V
fR(u, v) +
fR(v, w) +
(cid:88)
m∈M
1v=sm λm ∀v ∈ V
fm(u, v) ≤ c(u, v) ∀(u, v) ∈ E
(4)
(5)
(6)
(cid:88)
(u,v)∈E
The cost function (1) is a weighted sum (with weight
ρ) of the overall duration of all passenger trips and the
duration of rebalancing trips. Constraints (2), (3) and (4)
enforce continuity of each trip (i.e., flow conservation) across
nodes. Constraint (5) ensures that vehicles are rebalanced
throughout the road network to re-align vehicle distribution
capacity-symmetric road networks indeed appears reasonable
for a number of major U.S. metropolitan areas (note that
this assumption is much less restrictive than assuming every
individual road is capacity-symmetric). In the Supplementary
Material, by using OpenStreetMap data [31], we provide a
rigorous characterization in terms of capacity symmetry of
the road networks of New York City, Chicago, Los Angeles
and other major U.S. cities. The results consistently show that
urban road networks are usually symmetric to a very high
degree. Additionally, several of our theoretical and algorithmic
results extend to the case where this assumption is lifted, as
it will be highlighted throughout the paper.
III. STRUCTURAL PROPERTIES OF THE
NETWORK FLOW MODEL
In this section we provide two key structural results for
the network flow model presented in Section II-B. First,
we provide a cut condition that needs to be satisfied for
feasible customer and rebalancing flows to exist. In other
words, this condition provides a fundamental limitation of
performance for congestion-free AMoD service in a given road
network. Second, we investigate an existential result (our main
theoretical result) that is germane to two key conclusions: (1)
rebalancing does not increase congestion in symmetric road
networks, and (2) for certain cost functions, the problems of
finding customer and rebalancing flows can be decoupled – an
insight that will be heavily exploited in subsequent sections.
A. Fundamental Limitations
(S, ¯S),
the overall flow exiting a cut
Fout(S, ¯S) :=(cid:80)
m∈M(cid:80)
i.e., Cout(S, ¯S) = (cid:80)
We start with a few definitions. For a given set of
feasible customer flows {fm(u, v)}(u,v),m, we denote by
Fout(S, ¯S)
i.e.,
u∈S,v∈ ¯S fm(u, v). Similarly, we de-
note by Cout(S, ¯S) the capacity of the network exiting S,
u∈S,v∈ ¯S c(u, v). Analogously, Fin(S, ¯S)
denotes the overall flow entering S from ¯S, i.e., Fin(S, ¯S) :=
Fout( ¯S,S), and Cin(S, ¯S) denotes the capacity entering S
from ¯S,
that
the arguments leading to the main result of this subsection
(Theorem 3.4) do not require the assumption of capacity
symmetry; hence, Theorem 3.4 holds for asymmetric road
networks as well.
i.e., Cin(S, ¯S) := Cout( ¯S,S). We highlight
The next technical lemma (whose proof is provided in the
Supplementary Material) shows that the net flow leaving set
S equals the difference between the flow originating from the
origins sm in S and the flow exiting through the destinations
tm in S, that is,
Lemma 3.1 (Net flow across a cut): Consider a set of fea-
sible customer flows {fm(u, v)}(u,v),m. Then, for every cut
(S, ¯S), the net flow leaving set S satisfies
Fout(S, ¯S) − Fin(S, ¯S) =
1sm∈S λm − (cid:88)
(cid:88)
1tm∈S λm.
m∈M
m∈M
We now state two additional lemmas (whose proofs are
given in the Supplementary Material) providing, respectively,
lower and upper bounds for the outflows Fout(S, ¯S).
Lemma 3.2 (Lower bound for outflow): Consider a set of
feasible customer flows {fm(u, v)}(u,v),m. Then, for any cut
(S, ¯S), the overall flow Fout(S, ¯S) exiting cut (S, ¯S) is lower
bounded according to(cid:88)
m∈M
1sm∈S,tm∈ ¯S λm ≤ Fout(S, ¯S).
Lemma 3.3 (Upper bound for outflow): Assume there ex-
feasible customer and rebalancing flows
ists a set of
{fm(u, v), fR(u, v)}(u,v),m. Then, for every cut (S, ¯S),
1) Fout(S, ¯S) ≤ Cout(S, ¯S), and
2) Fout(S, ¯S) ≤ Cin(S, ¯S).
We are now in a position to present a structural (i.e., flow-
independent) necessary condition for the existence of feasible
customer and rebalancing flows.
Theorem 3.4 (Necessary condition for feasible flows): A
necessary condition for the existence of a set of feasible
customer and rebalancing flows {fm(u, v), fR(u, v)}(u,v),m,
is that, for every cut (S, ¯S),
m∈M 1sm∈S,tm∈ ¯S λm ≤ Cout(S, ¯S), and
m∈M 1sm∈S,tm∈ ¯S λm ≤ Cin(S, ¯S).
Proof: The theorem is a trivial consequence of Lemmas
1) (cid:80)
2) (cid:80)
3.2 and 3.3.
Theorem 3.4 essentially provides a structural fundamental
limitation of performance for a given road network: if the cut
conditions in Theorem 3.4 are not met, then there is no hope
of finding congestion-free customer and rebalancing flows.
We reiterate that Theorem 3.4 holds for both symmetric and
asymmetric networks (for a symmetric network, claim 2) in
Lemma 3.3 and condition 2) in Theorem 3.4 are redundant).
B. Existence of Congestion-Free Flows
In this section we address the following question: assuming
there exists a feasible customer flow, is it always possible to
find a feasible rebalancing flow? As we will see, the answer
to this question is affirmative and has both conceptual and
algorithmic implications.
Theorem 3.5 (Feasible rebalancing): Assume there exists a
set of feasible customer flows {fm(u, v)}(u,v),m. Then, it is
always possible to find a set of feasible rebalancing flows
{fR(u, v)}(u,v).
Proof: We prove the theorem for the special case where
no node v ∈ V is associated with both an origin and a
destination for the transportation requests in M. This is
without loss of generality, as the general case where a node v
has both an origin and a destination assigned can be reduced to
this special case, by associating with node v a "shadow" node
so that (i) all destinations are assigned to the shadow node
and (ii) node v and its shadow node are mutually connected
via an infinite-capacity, zero-travel-time edge.
We start the proof by defining the concepts of partial rebal-
ancing flows and defective origins and destinations. Specifi-
cally, a partial rebalancing flow, denoted as { fR(u, v)}(u,v),
is a set of mappings from E to R≥0 obeying the following
properties:
1) It satisfies constraint (5) at every node that is not an origin
nor a destination, that is ∀ v ∈ {V \{{sm}m ∪{tm}m}},
fR(u, v) =
fR(v, w).
(cid:88)
w∈V
(cid:88)
u∈V
2) It violates constraint (5) in the "≤ direction" at every
node that is an origin, that is ∀ v ∈ V such that ∃m ∈
M : v = sm,(cid:88)
fR(u, v) ≤ (cid:88)
(cid:88)
w∈V
fR(u, v) +
fR(v, w) +
(cid:88)
1v=tm λm ≥ (cid:88)
m∈M
m∈M
w∈V
u∈V
v = tm,(cid:88)
u∈V
3) It violates constraint (5) in the "≥ direction" at every node
that is a destination, that is ∀ v ∈ V such that ∃m ∈ M :
1v=sm λm.
fR(v, w).
4) The combined customer and partial rebalancing flows
{fm(u, v), fR(u, v)}(u,v),m satisfy Equation (6) for every
edge (u, v) ∈ E.
Note that the trivial zero flow, that is fR(u, v) = 0 for all
(u, v) ∈ E, is a partial rebalancing flow (in other words, the set
of partial rebalancing flows in not empty). Clearly a feasible
rebalancing flow is also a partial rebalancing flow, but the
opposite is not necessarily true.
For a given partial rebalancing flow, we denote an origin
node, that is a node v ∈ V such that v = sm for some
m = 1, . . . , M, as a defective origin if Equation (5) is not
satisfied at v = sm (in other words, the strict inequality <
holds). Analogously, we denote a destination node, that is a
node v ∈ V such that v = tm for some m = 1, . . . , M,
as a defective destination if Equation (5) is not satisfied at
v = tm (in other words, the strict inequality > holds). The
next lemma (whose proof is provided in the Supplementary
Material) links the concepts of partial rebalancing flows and
defective origins/destinations.
Lemma 3.6 (Co-existence of defective origins/destinations):
For every partial rebalancing flow that
is not a feasible
rebalancing flow, there exists at least one node u ∈ V that is
a defective origin, and one node v ∈ V that is a defective
destination.
For a given set of customer flows {fm(u, v)}(u,v),m and
partial rebalancing flows { fR(u, v)}(u,v), we call an edge
(u, v) ∈ E saturated if Equation (6) holds with equality
for that edge. We call a path saturated if at least one of
the edges along the path is saturated. We now prove the
existence of a special partial rebalancing flow where defective
destinations and defective origins are separated by a graph
cut formed exclusively by saturated edges (this result, and its
consequences, are illustrated in Figure 2).
Lemma 3.7: Assume there exists a set of feasible customer
flows {fm(u, v)}(u,v),m, but there does not exist a set of
feasible rebalancing flows {fR(u, v)}(u,v). Then, there exists
a partial rebalancing flow { fR(u, v)}(u,v) that induces a graph
cut (S, ¯S) with the following properties: (i) all defective
destinations are in S, (ii) all defective origins are in ¯S, and
(iii) all edges in (S, ¯S) are saturated.
We are now in a position to prove Theorem 3.5. The proof
is by contradiction. Assume that a set of feasible rebalancing
flows {fR(u, v)}(u,v) does not exist. Then Lemma 3.7 shows
that there exists a partial rebalancing flow { fR(u, v)}(u,v)
and a cut (S, ¯S) such that all defective destinations under
{ fR(u, v)}(u,v) belong to S and all defective origins belong
to ¯S. Let us denote the sum of all partial rebalancing flows
in (S, ¯S)−Fout(S, ¯S)− F reb
Combining the above equations, one obtains
Fin(S, ¯S)+ F reb
out (S, ¯S), we follow a procedure
in (S, ¯S) − F reb
To compute F reb
similar to the one used in Lemma 3.1. Summing Equation (5)
(cid:88)
over all nodes in S, one obtains,
out (S, ¯S) ≤ Cin(S, ¯S)−Cout(S, ¯S).
(cid:34)(cid:88)
(cid:88)
v∈S
u∈V
m∈M
1v=tm λm
(cid:35)
fR(v, w) +
1v=sm λm
.
fR(u, v) +
(cid:34)(cid:88)
w∈V
(cid:88)
v∈S
>
(cid:35)
(cid:88)
m∈M
Fig. 2. A graphical representation of Lemma 3.7. If there exists a set of
feasible customer flows but there does not exist a set of feasible rebalancing
flows, one can find a partial rebalancing flow where all the defective origins,
represented as blue circles, are separated from all the defective destinations,
represented as blue squares, by a cut of saturated edges (shown in red). Note
that not all saturated edges necessarily belong to the cut. In the proof of
Theorem 3.5 we show that the capacity of such a cut (S, ¯S) is asymmetric,
i.e., Cout < Cin – a contradiction that leads to the claim of Theorem 3.5.
across cut (S, ¯S) as
out (S, ¯S) :=
F reb
(cid:88)
fR(u, v),
u∈S,v∈ ¯S
in (S, ¯S) := F reb
out ( ¯S,S). Since all
and, analogously, define F reb
edges in the cut (S, ¯S) are saturated under { fR(u, v)}(u,v),
one has, due to Equation (6), the equality
Cout(S, ¯S) = Fout(S, ¯S) + F reb
Fin(S, ¯S) + F reb
out (S, ¯S).
in (S, ¯S) ≤ Cin(S, ¯S).
Additionally, again due to Equation (6), one has the inequality
m∈M
1tm∈S λm.
out (S, ¯S) >
(cid:88)
(cid:88)
The strict inequality is due to the fact that for a partial
rebalancing flow that is not feasible there exists at least one
defective destination (Lemma 3.6), which, by construction,
must belong to S. Simplifying those flows fR(u, v) for which
both u and v are in S (as such flows appear on both sides of
the above inequality), one obtains
in (S, ¯S) − F reb
F reb
Also, by Lemma 3.1,
Fout(S, ¯S) − Fin(S, ¯S) =
m∈M
Collecting all the results so far, we conclude that
in (S, ¯S) − Fout(S, ¯S) − F reb
1sm∈S λm − (cid:88)
1sm∈S λm − (cid:88)
0 < Fin(S, ¯S) + F reb
= Cin(S, ¯S) − Cout(S, ¯S).
Hence, we reached the conclusion that Cin(S, ¯S) −
Cout(S, ¯S) > 0, or, in other words, the capacity of graph
G(V,E) across cut (S, ¯S) is not symmetric. This contradicts
the assumption that graph G(V,E) is capacity-symmetric, and
the claim follows.
out (S, ¯S)
1tm∈S λm.
m∈M
m∈M
The importance of Theorem 3.5 is twofold. First, perhaps
surprisingly, it shows that for symmetric road networks it is
always possible to rebalance the autonomous vehicles without
increasing congestion – in other words, the rebalancing of
autonomous vehicles in a symmetric road network does not
lead to an increase in congestion. Second, from an algorithmic
standpoint, if the cost function in the CRRP only depends
on the customer flows (that is, ρ = 0 and the goal is to
minimize the customers' travel times), then the CRRP problem
can be decoupled and the customers and rebalancing flows can
be solved separately without loss of optimality. This insight
will be instrumental in Section IV to the design of real-time
algorithms for routing and rebalancing.
We conclude this section by noticing that the CRRP, from a
computational standpoint, can be reduced to an instance of the
Minimum-Cost Multi-Commodity Flow problem (Min-MCF),
a classic problem in network flow theory [29]. The problem
can be efficiently solved either via linear programming (the
size of the linear program is E(M + 1)), or via spe-
cialized combinatorial algorithms [32], [33], [34]. However,
the solution to the CRRP provides static fractional flows,
which are not directly implementable for the operation of
actual AMoD systems. Practical algorithms (inspired by the
theoretical CRRP model) are presented in the next section.
IV. REAL-TIME CONGESTION-AWARE
ROUTING AND REBALANCING
A natural approach to routing and rebalancing would be
to periodically resolve the CRRP within a receding-horizon,
batch-processing scheme (a common scheme for the control
of transportation networks [35], [4], [36]). This approach,
however, is not directly implementable as the solution to the
CRRP provides fractional flows (as opposed to routes for the
individual vehicles). This shortcoming can be addressed by
considering an integral version of the CRRP (dubbed integral
CRRP), whereby the flows are integer-valued and can be thus
easily translated into routes for the individual vehicles, e.g.
through a flow decomposition algorithm [37]. The integral
CRRP, however, is an instance of the integral Minimum-Cost
Multi-Commodity Flow problem, which is known to be NP-
hard [38], [39]. Naıve rounding techniques are inapplicable:
rounding a solution for the (non-integral) CRRP does not yield,
in general, feasible integral flows, and hence feasible routes.
For example, continuity of vehicles and customers can not
be guaranteed, and vehicles may appear and disappear along
a route. In general, to the best of our knowledge, there are
no polynomial-time approximation schemes for the integral
Minimum-Cost Multi-Commodity Flow problem.
On the positive side, the integral CRRP admits a decoupling
result akin to Theorem 3.5: given a set of feasible, integral
customer flows, one can always find a set of feasible, integral
rebalancing flows. (In fact, the proof of Theorem 3.5 does not
exploit anywhere the property that the flows are fractional, and
thus the proof extends virtually unchanged to the case where
the flows are integer-valued). Our approach is to leverage
this insight (and more in general the theoretical results from
Section III) to design a heuristic, yet efficient approximation
to the integral CRRP that (i) scales to large-scale systems,
and (ii) is general, in the sense that can be broadly applied to
time-varying, asymmetric networks.
Specifically, we consider as objective the minimization of
times, which, from Section III and
the generalization of
the customers' travel
the aforementioned discussion about
Theorem 3.5 to integral flows, suggests that customer routing
can be decoupled from vehicle rebalancing (strictly speaking,
this statement is only valid for static and symmetric networks –
its generalization beyond these assumptions will be addressed
numerically in Section V). Accordingly, to emulate the real-
world operation of an AMoD system, we divide a given
city into geographic regions (also referred to as "stations"
in some formulations) [4], [8], and each arriving customer
is assigned the closest vehicle within that region (vehicle
imbalance across regions is handled separately by the vehicle
rebalancing algorithm, discussed below). We apply a greedy,
yet computationally-efficient and congestion-aware approach
for customer routing where customers are routed to their
destinations using the shortest-time path as computed by an A∗
algorithm [40]. The travel time along each edge is computed
using a heuristic delay function that is related to the current
volume of traffic on each edge. In this work, for each edge
(u, v) ∈ E we use the simple Bureau of Public Roads (BPR)
delay model [41]
(cid:32)
(cid:18) f (u, v)
c(u, v)
(cid:19)β(cid:33)
,
We assume the residual capacity cR(u, v) of an edge (u, v),
defined as the difference between its overall capacity c(u, v)
and the current number of vehicles along that edge, is known
and remains approximately constant over the rebalancing time
horizon. In case the overall rebalancing problem is not feasible
(i.e. it is not possible to move all excess vehicles to regions
that have a deficit of vehicles while satisfying the congestion
constraints), we define slack variables with cost C that allow
the optimizer to select a subset of vehicles and rebalancing
where f (u, v) :=(cid:80)M
td(u, v) := t(u, v)
1 + α
m=1 fm(u, v) + fR(u, v) is the total flow
on edge (u, v), and α and β are usually set to 0.15 and 4
respectively. Note that customer routing is event-based, i.e, a
routing choice is made as soon as a customer arrives.
i
(t) := vi(t) +(cid:80)
Separately from customer routing, vehicle rebalancing from
one region to another is performed every thor > 0 time units as
a batch process (unlike customer routing, which is an event-
based process). Denote by vi(t) the number of vehicles in
region i at time t, and by vji(t) the number of vehicles
traveling from region j to i that will arrive in the next tvicinity
time units. Let vown
j vji(t) be the number of
vehicles currently "owned" by region i (i.e., in the vicinity of
such region). Denote by ve
i (t) the number of excess vehicles
in region i, or the number of vehicles left after servicing the
customers waiting within region i. From its definition, ve
i (t) is
(t)− ci(t), where ci(t) is the number of
given by ve
customers within region i. Finally, denote by vd
i (t) the desired
number of vehicles within region i. For example, for an even
distribution of excess vehicles, vd
i (t)/N, where
i ve
N is the number regions. Note that the vd
i (t)'s are rounded
so they take on integer values. The set of origin regions (i.e.,
regions that should send out vehicles), SR, and destination
regions (i.e., regions that should receive vehicles), TR, for the
rebalancing vehicles are then determined by comparing ve
i (t)
and vd
i (t) ∝ (cid:80)
i (t) = vown
i (t), specifically,
i
if ve
if ve
i (t) > vd
i (t) < vd
i (t),
i (t),
region i ∈ SR
region i ∈ TR.
routes of maximum cardinality such that each link does not
become congested. The slack variables are denoted as dsi for
each i ∈ SR, and dtj for each j ∈ TR.
time units, the rebalancing vehicle routes are
computed by solving the following integer linear program
(cid:88)
(cid:88)
t(u, v) fR(u, v) +
Cdsi +
Cdti
fR(u, v) + 1v∈SR (ve
i∈SR
v(t) − vd
i∈TR
v (t) − dsv)
fR(v, w) + 1v∈TR (vd
v (t) − ve
v(t) − dtv),
fR(·,·),{dsi},{dtj}
Every thor
minimize
(cid:88)
subject to (cid:88)
(cid:88)
(u,v)∈E
u∈V
=
w∈V
for all v ∈ V
fR(u, v) ≤ cR(u, v),
fR(u, v) ∈ N,
dsi, dtj ∈ N,
for all (u, v) ∈ E
for all (u, v) ∈ E
for all i ∈ SR, j ∈ TR
The set of (integral) rebalancing flows {fR(u, v)}(u,v) is then
decomposed into a set of rebalancing paths via a flow decom-
position algorithm [37]. Each rebalancing path connects one
origin region with one destination region: thus, rebalancing
paths represent the set of routes that excess vehicles should
follow to rebalance to regions with a deficit of vehicles.
The rebalancing optimization problem is an instance of
the Minimum Cost Flow problem. If all edge capacities are
integral,
the linear relaxation of the Minimum Cost Flow
problem enjoys a totally unimodular constraint matrix [29].
Hence, the linear relaxation will necessarily have an integer
optimal solution, which will be a fortiori an optimal solution
to the original Minimum Cost Flow problem. It follows that
an integer-valued solution to the rebalancing optimization
problem can be computed efficiently, namely in polynomial
time, e.g., via linear programming. Several efficient combina-
torial algorithms [29] are also available, whose computational
performance is typically significantly better.
The favorable computational properties of the routing and
rebalancing algorithm presented in this section enable appli-
cation to large-scale systems, as described next.
V. NUMERICAL EXPERIMENTS
In this section, we characterize the effect of rebalancing
on congestion in asymmetric network and explore the perfor-
mance of the algorithm presented in Section IV on real-world
road topologies with real customer demands.
A. Characterization of Congestion due to Rebalancing in
Asymmetric Networks
The theoretical results in Section III are proven for capacity-
symmetric networks, which are in general a reasonable model
for typical urban road networks (we refer the reader to the
Supplementary Material for a detailed analysis of capacity
symmetry for major U.S. cities). Nevertheless, it is of interest
to characterize the applicability of our theoretical results
(chiefly, the existential result in Theorem 3.5) to road networks
that significantly violate the capacity-symmetry property. In
other words, we study to what degree rebalancing might lead
to an increase in congestion if the network is asymmetric.
To this purpose, we compute solutions to the CRRP for
road networks with varying degrees of capacity asymmetry
and we compare corresponding travel times to those obtained
by computing optimal routes in the absence of rebalancing (as
it would be the case, e.g., if the vehicles were privately owned).
We focus on the road network portrayed in Figure 3(a),
which captures all major streets and avenues in Manhattan.
Transportation requests are based on actual taxi rides in New
York City on March 1, 2012 from 6 to 8 p.m. (courtesy of the
New York Taxi and Limousine Commission). We randomly
selected about one third of the trips that occurred in that time
frame (roughly 17,000 trips) and we adjusted the capacities
of the roads such that the flows induced by these trips would
approach the threshold of congestion. The roads considered all
have similar speed limits and comparable number of lanes and
thus we assign to each edge in the network the same capacity,
specifically, one vehicle every 23.6 seconds. This capacity is
consistent with the observations that (i) the customer flow is
only 30% of the real one (so road capacity is reduced accord-
ingly) and (ii) taxis only contribute to a fraction of the overall
traffic in Manhattan. Nevertheless, we stress that the capacity
was selected specifically to ensure that the flow induced by the
trips would approach the threshold of congestion before any
asymmetry is induced. To investigate the effects of network
asymmetry, we introduce an artificial capacity asymmetry
into the baseline Manhattan road network by progressively
reducing the capacity of all northbound avenues.
In order to gain a quantitative understanding of the effect
of rebalancing on congestion and travel times, we introduce
slack variables δC(u, v), associated with a cost cc(u, v), to
each congestion constraint (6). The cost cc(u, v) is selected
so that the optimization algorithm will select a congestion-
free solution whenever one is available. Once a solution is
found, the actual travel time on each (possibly congested) link
is computed with the heuristic BPR delay model [41] presented
in Section IV. This approach maintains feasibility even in the
congested traffic regime, and hence allows us to assess the
impact of rebalancing on congestion in asymmetric networks.
Figure 3(b) summarizes the results of our simulations. In the
baseline case, no artificial capacity asymmetry is introduced,
i.e., the fractional capacity reduction of northbound avenues
is equal to 0%. In this case, the customer routing problem
with no rebalancing (essentially, the CRRP problem with the
rebalancing flows constrained to be equal to zero) admits
a congestion-free solution. On the other hand,
the CRRP
requires a (very small) relaxation of the congestion constraints.
Overall, the difference between the travel times in the two
cases is very small and approximately equal to 2.12%, in line
with the fact that New York City's road graph has largely
symmetric capacity, as discussed in Section II and shown in
the Supplementary Material. Interestingly, even with a massive
50% reduction in northbound capacity,
times when
rebalancing vehicles are present are within 4.12% of those
obtained assuming no rebalancing is performed. Collectively,
these results show that
in Theorem
3.5, proven under the assumption of a symmetric network,
appears to extend (even though approximately) to asymmetric
networks. In particular, it appears that vehicle rebalancing does
not lead to an appreciable increase in congestion under very
general conditions.
the existential result
travel
We conclude this section by noticing that for a 40% reduc-
TABLE I
RESULTS OF THE REAL-TIME SIMULATIONS
Congestion-aware
Performance metric
# of trips completed
mean wait time (all trips)
mean travel time (completed trips)
mean service time (completed trips)
% with wait time > 5 minutes
mean # of rebalancing vehicles
49,585
163.57 s
265.13 s
286.96 s
5.4%
204
Baseline
42,219
406.03 s
275.19 s
324 s
20%
1489
Fig. 3.
Left: Manhattan road network. One-way roads are represented as
dashed lines. Centers of rebalancing regions are represented in red. Right:
Customer travel times with and without rebalancing for different levels of
network asymmetry.
tion in capacity, the travel times with vehicle rebalancing dip
slightly lower than those without. This effect is due to our use
of the BPR link delay model: while in our theoretical model
the time required to traverse a link is constant so long as a link
is uncongested, the link delay in the BPR model varies by as
much as 15% between free-flow and the onset of congestion.
B. Congestion-Aware Real-time Rebalancing
In this section we evaluate the performance of the real-
time routing and rebalancing algorithm presented in Section
IV against a baseline approach that does not explicitly take
congestion into account. We simulate 7, 000 vehicles providing
service to actual taxi requests on March 1, 2012, for two hours
between 6 and 8 p.m., using the same Manhattan road network
as in the previous section (see Figure 3(a)). Taxi requests
are clustered into 88 regions corresponding to a subset of
nodes in the road network. Road capacities are reduced to
account for exogenous vehicles on the roads to the point that
congestion occurs along some routes during the simulation.
The free flow speed of the vehicles is set to 25 mph (11 m/s)
and approximately 55,000 trip requests (from the taxi data set
discussed before) are simulated using a time step of 6 seconds.
The simulated speed of the vehicles on each link depend on
the number of vehicles in the link, and is calculated using the
BPR model. Other delay factors such as traffic signals, turning
times, and pedestrian blocking are not simulated.
Three simulations are performed, namely (i) assuming every
customer has access to a private vehicle with no rebalanc-
ing, (ii) using the congestion-aware routing and rebalancing
algorithm presented in Section IV, and (iii) using a baseline
rebalancing algorithm. The baseline approach is derived from
the real-time rebalancing algorithm presented in [8], which is a
point-to-point algorithm that computes rebalancing origins and
destinations without considering the underlying road network.
In the baseline approach, customer routes are computed in
the same way as in Section IV. For rebalancing, the origins
and destinations are first solved using the algorithm provided
in [8], then the routes are computed using the A∗ algorithm
much like the customer routes. In simulations (ii) and (iii),
rebalancing is performed every 2 minutes.
Table I presents a summary of the performance results
for simulations (ii) and (iii). Note that the service time is
the total
time a customer spends in the system (waiting
+ traveling). Only data from simulations (ii) and (iii) are
presented in Table I because the only applicable performance
metric in simulation (i) is the mean travel time which was
264.69 s. Comparing our algorithm with (i), we notice that
the additional rebalancing vehicles have no significant impact
on the travel time. Comparing our algorithm with (iii), we
notice that the congestion-aware algorithm outperforms the
baseline algorithm in every metric: low congestion allows the
vehicles to service customers faster, resulting in a reduction in
wait times as well as travel times. The baseline algorithm will
send rebalancing vehicles to stations with a deficit of vehicles
regardless of the level of congestion in the road network. This
results in many more empty vehicles dispatched to rebalance
the system (see Table I), which causes heavy congestion in the
network2. Our congestion-aware algorithm drastically reduces
this effect, resulting in very few congested road links.
VI. CONCLUSIONS AND FUTURE WORK
In this paper we presented a network flow model of an
autonomous mobility-on-demand system on a capacitated road
network. We formulated the routing and rebalancing problem
and showed that on symmetric road networks, it is always
possible to route rebalancing vehicles in a coordinated way
that does not
increase traffic congestion. Using a model
road network of Manhattan, we showed that rebalancing
did not increase congestion even for moderate degrees of
network asymmetry. We leveraged the theoretical
insights
to develop a computationally efficient real-time congestion-
aware routing and rebalancing algorithm and demonstrated its
performance over state-of-the-art point-to-point rebalancing al-
gorithms through simulation. This highlighted the importance
of congestion awareness in the design and implementation of
control strategies for a fleet of self-driving vehicles.
This work opens the field to many future avenues of
research. First, note that the solution to the integral CRRP
can directly be used as a practical routing algorithm. For
large scale systems, high-quality approximate solutions for the
integral CRRP may be obtained using randomized algorithms
[42], [43]. Second, from a modeling perspective, we would like
to study the inclusion of stochastic information (e.g., demand
prediction, travel time uncertainty) for the routing and rebal-
ancing problem, as well as a richer set of performance metrics
and constraints (e.g., time windows to pick up customers).
Third, it is worthwhile to study how our results give intuition
into business models for autonomous urban mobility (e.g. fleet
sizes). Fourth, it is of interest to explore other approaches
that may reduce congestion, including ride-sharing, demand
staggering, and integration with public transit to create an
intermodal transportation network. Fifth, we would like to
explore decentralized architectures for cooperative routing
and rebalancing. Finally, we would like to demonstrate the
real-world performance of the algorithms using high fidelity
2See the Media Extension, available at https://youtu.be/7OivaJi6CHU
0102030405060Capacity reduction on northbound avenues [%]400060008000100001200014000Overall travel time [s]Routing and rebalancingCustomers only (no rebalancing) 2014.
microscopic traffic simulators and by implementing them on
real fleets of self-driving vehicles.
ACKNOWLEDGEMENT
The authors would like to thank Zachary Sunberg for his
analysis on the road network symmetry of U.S. cities.
REFERENCES
[1] W. J. Mitchell, C. E. Borroni Bird, and L. D. Burns, Reinventing the
Automobile: Personal Urban Mobility for the 21st Century. Cambridge,
MA: The MIT Press, 2010.
[2] Google, "Just Press Go: Designing a Self-Driving Vehicle." Tech. Rep.,
[3] K. Spieser, K. Treleaven, R. Zhang, E. Frazzoli, D. Morton, and
M. Pavone, "Toward a Systematic Approach to the Design and Eval-
uation of Automated Mobility-On-Demand Systems: A Case Study in
Singapore," in Lecture Notes in Mobility.
Springer, Jun. 2014, pp.
229–245.
[4] M. Pavone, S. L. Smith, E. Frazzoli, and D. Rus, "Robotic Load
Balancing for Mobility-On-Demand Systems," International Journal of
Robotics Research, vol. 31, no. 7, pp. 839–854, Jun. 2012.
[5] J. P´erez, F. Seco, V. Milan´es, A. Jim´enez, J. C. D´ıaz, and T. De Pedro,
"An RFID-based Intelligent Vehicle Speed Controller Using Active
Traffic Signals," Sensors, vol. 10, no. 6, pp. 5872–5887, 2010.
[6] B. Templeton, "Traffic Congestion & Capacity," 2015, available at http:
//www.templetons.com/brad/robocars/congestion.html.
at
2016,
[7] M. Barnard, "Autonomous Cars Likely to Increase Congestion,"
http://cleantechnica.com/2016/01/17/
Jan.
autonomous-cars-likely-increase-congestion.
available
[8] R. Zhang and M. Pavone, "A Queueing Network Approach to the
Analysis and Control of Mobility-On-Demand Systems," in American
Control Conference, Chicago, IL, Jul. 2015, pp. 4702–4709.
[9] G. Berbeglia, J.-F. Cordeau, and G. Laporte, "Dynamic pickup and
delivery problems," European Journal of Operational Research, vol.
202, no. 1, pp. 8–15, 2010.
[10] K. Treleaven, M. Pavone, and E. Frazzoli, "An Asymptotically Optimal
Algorithm for Pickup and Delivery Problems," in Proc. IEEE Conf. on
Decision and Control, Orlando, FL, Dec. 2011, pp. 584–590.
[11] --, "Asymptotically Optimal Algorithms for One-to-One Pickup and
Delivery Problems With Applications to Transportation Systems," IEEE
Transactions on Automatic Control, vol. 58, no. 9, pp. 2261–2276, Sep.
2013.
[12] --, "Models and Efficient Algorithms for Pickup and Delivery
Problems on Roadmaps," in Proc. IEEE Conf. on Decision and Control,
Maui, HI, Dec. 2012, pp. 5691–5698.
[13] J. G. Wardrop, "Some Theoretical Aspects of Road Traffic Research,"
Thomas
in ICE Proceedings: Engineering Divisions, vol. 1, no. 3.
Telford, 1952, pp. 325–362.
[14] M. J. Lighthill and G. B. Whitham, "On kinematic waves. I. Flood
movement in long rivers," in Proceedings of the Royal Society of London
A: Mathematical, Physical and Engineering Sciences, vol. 229, no. 1178.
The Royal Society, 1955, pp. 281–316.
[15] C. F. Daganzo, "The cell transmission model: A dynamic representation
of highway traffic consistent with the hydrodynamic theory," Transporta-
tion Research Part B: Methodological, vol. 28, no. 4, pp. 269–287, 1994.
[16] B. S. Kerner, Introduction to modern traffic flow theory and control: the
long road to three-phase traffic theory. Springer Science & Business
Media, 2009.
[17] M. Treiber, A. Hennecke, and D. Helbing, "Microscopic simulation of
Springer, 2000,
congested traffic," in Traffic and Granular Flow 99.
pp. 365–376.
[18] Q. Yang and H. N. Koutsopoulos, "A microscopic traffic simulator
for evaluation of dynamic traffic management systems," Transportation
Research Part C: Emerging Technologies, vol. 4, no. 3, pp. 113–129,
1996.
[19] M. Balmer, M. Rieser, K. Meister, D. Charypar, N. Lefebvre, K. Nagel,
and K. Axhausen, "MATSim-T: Architecture and simulation times,"
Multi-agent systems for traffic and transportation engineering, pp. 57–
78, 2009.
[20] C. Osorio and M. Bierlaire, "An analytic finite capacity queueing
network model capturing the propagation of congestion and blocking,"
European Journal of Operational Research, vol. 196, no. 3, pp. 996–
1007, 2009.
[21] S. Peeta and H. S. Mahmassani, "System optimal and user equilibrium
time-dependent traffic assignment in congested networks," Annals of
Operations Research, vol. 60, no. 1, pp. 81–113, 1995.
[22] B. N. Janson, "Dynamic traffic assignment for urban road networks,"
Transportation Research Part B: Methodological, vol. 25, no. 2, pp.
143–161, 1991.
[23] T. Le, P. Kov´acs, N. Walton, H. L. Vu, L. L. H. Andrew, and S. S. P.
Hoogendoorn, "Decentralized signal control for urban road networks,"
Transportation Research Part C: Emerging Technologies, 2015.
[24] N. Xiao, E. Frazzoli, Y. Luo, Y. Li, Y. Wang, and D. Wang, "Throughput
optimality of extended back-pressure traffic signal control algorithm," in
Control and Automation (MED), 2015 23th Mediterranean Conference
on.
IEEE, 2015, pp. 1059–1064.
[25] M. Papageorgiou, H. Hadj Salem, and J.-M. Blosseville, "ALINEA:
A local feedback control law for on-ramp metering," Transportation
Research Record, no. 1320, 1991.
[26] D. Wilkie, J. P. van den Berg, M. C. Lin, and D. Manocha, "Self-Aware
Traffic Route Planning," in AAAI, 2011.
[27] D. Wilkie, C. Baykal, and M. C. Lin, "Participatory Route Planning,"
in Proceedings of the 22nd ACM SIGSPATIAL International Conference
on Advances in Geographic Information Systems. ACM, 2014, pp.
213–222.
[28] B. S. Kerner, "Traffic Congestion, Modeling Approaches to," in
Encyclopedia of Complexity and Systems Science, R. A. Meyers, Ed.
Springer New York, 2009, pp. 9302–9355.
[29] Ravindra K. Ahuja, Thomas L. Magnanti, and James B. Orlin, Network
Flows: Theory, Algorithms and Applications. Upper Saddle River, New
Jersey 07458: Prentice Hall, 1993.
[30] H. Neuburger, "The economics of heavily congested roads," Transporta-
tion Research, vol. 5, no. 4, pp. 283 – 293, 1971.
[31] M. Haklay and P. Weber, "OpenStreetMap: User-Generated Street
Maps," Pervasive Computing, IEEE, vol. 7, no. 4, pp. 12–18, Oct 2008.
[32] Andrew V. Goldberg, Eva Tardos, and Robert E. Tarjan, "Network Flow
Algorithms," Cornell University Operations Research and Industrial
Engineering, Tech. Rep., 1989.
[33] T. Leighton, F. Makedon, S. Plotkin, C. Stein,
´E. Tardos, and
S. Tragoudas, "Fast approximation algorithms for multicommodity flow
problems," Journal of Computer and System Sciences, vol. 50, no. 2,
pp. 228–243, 1995.
[34] A. V. Goldberg, J. D. Oldham, S. Plotkin, and C. Stein, "An Implemen-
tation of a Combinatorial Approximation Algorithm for Minimum-Cost
Multicommodity Flow," in Integer Programming and Combinatorial Op-
timization, ser. Lecture Notes in Computer Science, R. Bixby, E. Boyd,
and R. Ros-Mercado, Eds. Springer Berlin Heidelberg, 1998, vol. 1412,
pp. 338–352.
[35] K. T. Seow, N. H. Dang, and D. H. Lee, "A collaborative multiagent
taxi-dispatch system," IEEE Transactions on Automation Sciences and
Engineering, vol. 7, no. 3, pp. 607–616, 2010.
[36] R. Zhang, F. Rossi, and M. Pavone, "Model Predictive Control of
Autonomous Mobility-on-Demand Systems (Extended version)," Sep.
2015, available at http://arxiv.org/abs/1509.03985.
[37] L. R. Ford and D. R. Fulkerson, Flows in Networks.
Princeton
University Press, 1962.
[38] R. M. Karp, "On the computational complexity of combinatorial prob-
lems," in Networks, Networks (USA), (Proceedings of the Symposium on
Large-Scale Networks, Evanston, IL, USA, 18-19 April 1974.), vol. 5,
no. 1, Jan. 1975, pp. 45–68.
[39] S. Even, A. Itai, and A. Shamir, "On the Complexity of Timetable and
Multicommodity Flow Problems," SIAM Journal on Computing, vol. 5,
no. 4, pp. 691–703, 1976.
[40] P. Hart, N. Nilsson, and B. Raphael, "A Formal Basis for the Heuristic
Determination of Minimum Cost Paths," Systems Science and Cybernet-
ics, IEEE Transactions on, vol. 4, no. 2, pp. 100–107, July 1968.
[41] Bureau of Public Roads, "Traffic Assignment Manual," U.S. Department
of Commerce, Urban Planning Division, Washington, D.C (1964), Tech.
Rep., 1964.
[42] P. Raghavan and C. D. Tompson, "Randomized Rounding: A Technique
for Provably Good Algorithms and Algorithmic Proofs," Combinatorica,
vol. 7, no. 4, pp. 365–374, 1987.
[43] A. Srinivasan, "A survey of the role of multicommodity flow and
randomization in network design and routing," American Mathematical
Society, Series in Discrete Mathematics and Theoretical Computer
Science, vol. 43, pp. 271–302, 1999.
SUPPLEMENTARY MATERIAL:
PROOFS OF TECHNICAL RESULTS
=
Proof of Lemma 3.1: We compute the sum over all
customer flows m ∈ M and over all nodes v ∈ V of the
node balance equation for flow m at node v (Equation (3) if
node v is the source of m, Equation (4) if node v is the sink
of m, or Equation (2) otherwise). We obtain
(cid:33)
(cid:88)
(cid:88)
v∈S
m∈M
(cid:32)(cid:88)
(cid:88)
(cid:88)
u∈V
(cid:32)(cid:88)
=
(cid:33)
fm(v, w) + 1v=tmλm
.
v∈S
m∈M
w∈V
For any edge (u, v) such that u, v ∈ S, the customer flow
fm(u, v) appears on both sides of the equation. Thus the
equation above simplifies to
fm(u, v) + 1v=sm λm
(cid:88)
(cid:88)
m∈M
v∈S
u∈ ¯S
(cid:88)
(cid:88)
(cid:88)
m∈M
m∈M
fm(u, v) + 1v=smλm
(cid:88)
(cid:88)
fm(v, w) + 1v=tmλm
,
which leads to the claim of the lemma
Fin(S, ¯S) +
v∈S
1sm∈S λm = Fout(S, ¯S) +
w∈ ¯S
(cid:88)
m∈M
1tm∈S λm.
Proof of Lemma 3.2: Adding Equations (2), (3) and (4)
over all nodes in S and over all flows whose origin is in S
and whose destination is in ¯S, one obtains
(cid:88)
(cid:32)(cid:88)
(cid:88)
u∈V
(cid:33)
=
(cid:33)
fm(v, w)
.
fm(u, v) + 1v=sm λm
(cid:32)(cid:88)
w∈V
(cid:88)
v∈S
m:sm∈S,tm∈ ¯S
Flows fm(u, v) such that both u and v are in S appear on
both sides of the equation. Simplifying, one obtains
m:sm∈S,tm∈ ¯S
v∈S
(cid:88)
(cid:88)
(cid:88)
m:sm∈S,tm∈ ¯S
m:sm∈S,tm∈ ¯S
fm(u, v)
v∈S,u∈ ¯S
λm =
v∈S,w∈ ¯S
(cid:88)
fm(v, w) − (cid:88)
(cid:88)
(cid:88)
≥ (cid:88)
(cid:88)
v∈S,w∈ ¯S
m∈M
m:sm∈S,tm∈ ¯S
v∈S,w∈ ¯S
fm(v, w).
The first term on the right-hand side represents a lower bound
for Fout(S, ¯S), since
Fout(S, ¯S) =
fm(v, w)
Furthermore, the second term on the right-hand side is upper-
bounded by zero. The lemma follows.
Proof of Lemma 3.3: The first condition follows trivially
from equation (6). As for the second condition, consider a cut
(S, ¯S). Analogously as for the definitions of Fin(S, ¯S) and
Fout(S, ¯S), let F reb
out (S, ¯S) denote, respectively,
the overall rebalancing flow entering (exiting) cut (S, ¯S).
in (S, ¯S) and F reb
(cid:88)
1sm∈S λm − (cid:88)
Summing equation (5) over all nodes in S, one easily obtains
in (S, ¯S) − F reb
F reb
1tm∈S λm.
Combining the above equation with Lemma 3.1, one obtains
out (S, ¯S) =
m∈M
m∈M
in (S, ¯S) − F reb
F reb
out (S, ¯S) = Fout(S, ¯S) − Fin(S, ¯S),
in other words, rebalancing flows should make up the dif-
ference between the customer inflows and outflows across cut
(S, ¯S). Accordingly, the total inflow of vehicles across (S, ¯S),
in (S, ¯S), satisfies the inequality
F tot
in (S, ¯S) : = Fin(S, ¯S) + F reb
F tot
= Fin(S, ¯S) + F reb
≥ Fout(S, ¯S).
customer
the
in (S, ¯S)
out (S, ¯S) + Fout(S, ¯S) − Fin(S, ¯S)
Since
flows
{fm(u, v), fR(u, v)}(u,v),m are feasible,
then, by equation
in (S, ¯S) ≤ Cin(S, ¯S). Collecting the results, one obtains
(6), F tot
the second condition.
Proof of Lemma 3.6: By contradiction. Since the flow
{ fR(u, v)}(u,v) is not a feasible rebalancing flow, there exists
at least one defective origin or a defective destination. Assume
that there exists at least one defective destination, say a node
tj where Equation (5) is violated:
1tj =tm
fR(u, tj) +
fR(tj, w),
rebalancing
(cid:88)
(cid:88)
(cid:88)
λm >
and
u∈V
m∈M
w∈V
Now, assume that there does not exist any defective origin. By
summing Equation (5) over all nodes v ∈ V and simplifying
all flows fR(u, v) (as they appear on both sides of the resulting
equation), one obtains
(cid:88)
(cid:88)
that is (cid:80)
m∈M λm > (cid:80)
m∈M
v∈V
(cid:88)
(cid:88)
v∈V
m∈M
m∈M λm, which is a contradiction.
Noticing that the symmetric case where we assume that there
exists at least one defective destination leads to an analogous
contradiction, the lemma follows.
Proof of Lemma 3.7: The proof is constructive and
constructs the desired partial rebalancing flow by starting with
the trivial zero flow fR(u, v) = 0 for all (u, v) ∈ E. Let
Vor, def := {s1, . . . , sVor, def} and Vdest, def := {t1, . . . , tVdest, def}
be the set of defective origins and destinations, respectively,
under such flow. Then, the zero flow is iteratively updated
according to the following procedure:
1) Look for a path between a node in Vdest, def and a node
in Vor, def that is not saturated (note that for rebalancing
flows, paths go from destinations to origins). If no such
path exists, quit. Otherwise, go to Step 2.
2) Add the same amount of flow on all edges along the path
until either (i) one of the edges becomes saturated or (ii)
constraint (5) is fulfilled either at the defective origin or
at the defective destination. Note that the resulting flow
remains a partial rebalancing flow.
3) Update sets Vor, def and Vdest, def for the new partial rebal-
ancing flow and go to Step 1.
The algorithm terminates. To show this, we prove the
invariant that if a node is no longer defective for the updated
partial rebalancing flow (in other words, Step 2 ends due to
condition (ii)), it will not become defective at a later stage.
Consider a defective destination node v that becomes non-
1v=tm λm >
1v=smλm,
the speed limits, and the number of lanes are obtained from
OpenStreetMap data [31].
D(S, ¯S) as
For a cut (S, ¯S), we define its fractional capacity disparity
D(S, ¯S) := 2
(cid:12)(cid:12)Cout(S, ¯S) − Cin(S, ¯S)(cid:12)(cid:12)
Cout(S, ¯S) + Cin(S, ¯S)
.
Table VI shows the average (over 1000 samples) fractional
capacity disparity for several US urban centers. As expected,
the road networks for such cities appear to posses a very high
degree of capacity-symmetry, which validates the symmetry
assumption made in Section III.
AVERAGE FRACTIONAL CAPACITY DISPARITY FOR SEVERAL MAJOR
URBAN CENTERS IN THE UNITED STATES.
TABLE II
Urban center
Chicago, IL
New York, NY
Colorado Springs, CO
Los Angeles, CA
Mobile, AL
Portland, OR
Avg. frac. capacity disparity
1.2972 ·10−4
1.6556 ·10−4
3.1772 ·10−4
0.9233 ·10−4
1.9368 ·10−4
1.0769 ·10−4
Std. dev.
1.003 · 10−4
1.304 · 10−4
2.308 · 10−4
0.676 · 10−4
1.452 · 10−4
0.778 · 10−4
defective under the updated partial rebalancing flow (the proof
for defective origins is analogous). Then, at the subsequent
stage it cannot be considered as a destination in Step 1 (as it
is no longer in set Vdest, def). If a path that does not contain
v is selected, then v stays non-defective. Otherwise, if a path
that contains v is selected, then, after Step 2, both the inbound
flow (that is the flow into v) and the outbound flow (that is
the flow out of v) will be increased by the same quantity, and
the node will stay non-defective. An induction on the stages
then proves the claim. As the number of paths is finite, and
sets Vor, def and Vdest, def cannot have any nodes added, the
algorithm terminates after a finite number of stages.
The output of the algorithm (denoted, with a slight abuse
of notation, as { fR(u, v)}(u,v)) is a partial rebalancing flow
that is not feasible (as, by assumption, there does not exist a
set of feasible rebalancing flows). Therefore, by Lemma 3.6,
such partial rebalancing flow has at least one defective origin
and at least one defective destination. Let us define Ens := E \
{(u, v) : (u, v) is saturated} as the collection of non-saturated
edges under the flows {fm(u, v)}(u,v),m and { fR(u, v)}(u,v).
For any defective destination and any defective origin, all paths
connecting them contain at least one saturated edge (due to
the exit condition in Step 1). Therefore, the graph Gns(V,Ens)
has two properties: (i) it is disconnected (that is, it is not
possible to find a direct path between every pair of nodes
in V by using edges in Ens), and (ii) a defective origin and a
defective destination can not be in the same strongly connected
component (hence, graph Gns(V,Ens) can be partitioned into
at least two strongly connected components).
We now find the cut (S, ¯S) as follows. If a strongly
connected component of Gns contains defective destinations,
we assign its nodes to set S. If a strongly connected component
contains defective origins, we assign its nodes to set ¯S. If
a strongly connected component contains neither defective
origins nor destinations, we assign its nodes to S (one could
also assign its nodes to ¯S, but such choice is immaterial for
our purposes). By construction, (S, ¯S) is a cut, and its edges
are all saturated. Furthermore, set S only contains destination
nodes, and set ¯S only contains origin nodes, which concludes
the proof.
SUPPLEMENTARY MATERIAL: CAPACITY SYMMETRY
WITHIN URBAN CENTERS IN THE US
The existential result in Section III, Theorem 3.5, relies on
the assumption that the road network is capacity-symmetric,
i.e., for every cut (S, ¯S), Cout(S, ¯S) = Cin(S, ¯S). One may
wonder whether this assumption is (approximately) met in
practice. From an intuitive standpoint, one might argue that
transportation networks within urban centers are indeed de-
signed to be capacity symmetric, so as to avoid accumulation
of traffic flow in some directions. We corroborate this intuition
by computing the imbalance between the outbound capacity
(i.e., Cout) and the inbound capacity (i.e., Cin) for 1000
randomly-selected cuts within several urban centers in the
United States. For each edge (u, v) ∈ E, we approximate
its capacity as proportional to the product of the speed limit
vmax(u, v) on that edge and the number of lanes L(u, v), that
is, c(u, v) ∝ vmax(u, v) · L(u, v). The road graph G(V,E),
|
1711.08225 | 1 | 1711 | 2017-11-22T10:56:31 | Micro and Macro Pedestrian Dynamics in Counterflow: the Impact of Social Groups | [
"cs.MA",
"cs.CY"
] | Although it is widely recognised that the presence of groups influences microscopic and aggregated pedestrian dynamics, a precise characterisation of the phenomenon still calls for evidences and insights. The present paper describes micro and macro level original analyses on data characterising pedestrian behaviour in presence of counter-flows and grouping, in particular dyads, acquired through controlled experiments. Results suggest that the presence of dyads and their tendency to walk in a line-abreast formation influences the formation of lanes and, in turn, aggregated observables, such as overall specific flow. | cs.MA | cs |
Micro and Macro Pedestrian Dynamics in
Counterflow: the Impact of Social Groups
Luca Crociani, Andrea Gorrini, Claudio Feliciani, Giuseppe Vizzari, Katsuhiro
Nishinari, Stefania Bandini
Abstract Although it is widely recognised that the presence of groups influences
microscopic and aggregated pedestrian dynamics, a precise characterisation of the
phenomenon still calls for evidences and insights. The present paper describes mi-
cro and macro level original analyses on data characterising pedestrian behaviour in
presence of counter-flows and grouping, in particular dyads, acquired through con-
trolled experiments. Results suggest that the presence of dyads and their tendency
to walk in a line-abreast formation influences the formation of lanes and, in turn,
aggregated observables, such as overall specific flow.
1 Introduction
Research on pedestrian dynamics has systematically analysed the influence of group
behaviour only in the last years (see, e.g., [9, 2, 8]): although observations and exper-
iments agree on some aggregated and microscopic effects of the presence of groups
(e.g. group members walk slower than individuals), there is still need for additional
insights, for instance on the spatial patterns assumed by groups in their movement
and in general on the interaction among different factors influencing overall pedes-
trian dynamics (e.g. do obstacles still make egress from a room smoother in presence
of groups?). Models incorporating mechanisms reproducing the cohesion of group
members, in fact, are just partially able to reproduce overall phenomena related to
Luca Crociani( ) · Andrea Gorrini · Giuseppe Vizzari · Stefania Bandini
CSAI research center, University of Milano-Bicocca, Milan, Italy.
e-mail: {name.surname}@disco.unimib.it
Claudio Feliciani · Katsuhiro Nishinari · Stefania Bandini
The University of Tokyo, Tokyo, Japan.
e-mail: [email protected]
Katsuhiro Nishinari
e-mail: [email protected]
1
2
Crociani, L., Gorrini, A., Feliciani, C., Vizzari, G., Nishinari, K., Bandini, S.
Fig. 1 A screenshot from the video of the experiments and a schematic representation of the set-
ting.
the presence of groups in the simulated population of pedestrians (see, e.g., [1] in
which groups preserve their cohesion and they move slower than individuals) and
they would benefit from additional insights on how members manage their move-
ments balancing (for instance) goal orientation, tendency to stay close to other mem-
bers, opportunities offered by the presence of lanes.
In this framework, the present work discusses results of experiments carried
out to investigate the potentially combined impact of counter-flow situations [13]
and grouping [11]. Experiment 1 [5] tested the impact of four different configu-
rations of counter-flow in a corridor setting (from uni-directional to fully balanced
bi-directional flow). Experiment 2 [7] replicated the same procedures and about half
of participants were paired to compose dyads (the simplest and most frequent type
of group), asking them to walk close to their companion. In the following, both the
experimental procedures and the achieved results will be presented in details.
2 Description of Experiments
The two experiments have been performed on June 13, 2015 at the Research Center
for Advanced Science and Technology of The University of Tokyo (Tokyo, JAPAN).
Experiments have been executed in a corridor-like setting composed as in Fig. 1.
The central area of 10× 3 m2 was recorded for the tracking of participants and it
was surrounded by two start areas of 12× 3 m2 and two buffer zones of 2 m length
that allowed participants to reach a stable speed at the measurement area. Partici-
pants were asked to wear coloured caps so that trajectories could be automatically
recorded with the software PeTrack[3].
Each experiment was composed of four procedures, to which 54 male students
participated. To achieve a more consistent dataset, every procedure was iterated four
times. The aim of the whole investigation has been to test the following hypotheses:
(Hp1) the increase of flow ratio negatively impacts the speed of pedestrians; (Hp2)
the cohesion of dyad members affects their speed; (Hp3) the cohesion of dyads leads
to a lower pedestrian flow at a macroscopic level.
Micro and Macro Pedestrian Dynamics in Counterflow
3
Fig. 2 Experiments and procedure tested in this investigation.
With flow ratio we denote the rate between the minor flow and the total flow
in bidirectional scenarios. Flow ratio was managed as independent variable among
four experimental procedures, as graphically exemplified in Fig. 2.
At the beginning of each iteration and according to the tested flow ratio, pedes-
trians were placed in the marked positions of the start areas. Fig. 2 exemplifies the
arrangement in all experimental procedures. In case of Exp. 2, roughly 44% of the
participants (24 out of the 54 total) was configured as dyads. These were formed
by coupling two random members and asking them to possibly walk close to the
other companion during the iteration. As shown in Fig. 2, dyads could be initially
disposed either in line abreast or river-like pattern, except in Proc. 2 where only
the latter was possible for dyads belonging to the minor flow. In the other cases the
choice was purely random.
3 Data Analysis
Individual data analysis on each of the two experiments has been already described
in [5] and [7]. As expected, both facing a counter-flow and being part of a group
was found to influence the walking speed of pedestrians. In addition, there is a
difference between the behavior found in balanced and un-balanced configurations
of counter-flow in terms of lanes' formation and amount of lateral motion required to
avoid conflicts with participants from the opposite direction [5, 6]. In this paper we
will compare the results of the two experiments, with the aim of verifying whether
their spatial patterns and different speeds affect the dynamics at a more macroscopic
level.
4
Crociani, L., Gorrini, A., Feliciani, C., Vizzari, G., Nishinari, K., Bandini, S.
(a)
(b)
Fig. 3 Distribution of pedestrian speeds (a) and local densities (b) among procedures of Exp. 2.
Black dots indicate the mean, red lines the median and the box size defines the standard deviation.
3.1 Microscopic Analysis on Dyads
A comparison of speeds among dyads and individuals in Exp. 2 shows that dyads
are slower in procedures characterised by a counter-flow situation. In presence of a
uni-directional flow essentially exempt from collisions (Proc. 1), on the other hand,
the difference is rather small (see Fig. 3(a)). This suggests that the bi-directional
flow affects more the spatial pattern of the dyad members, that more frequently
switch from the desired line-abreast pattern to a river-like one. Moreover, it is also
observed that group members have perceived a sensibly higher density during the
procedures with counter-flow: Fig. 3(b) shows the distributions of local densities for
all procedures of Exp. 2, calculated using the Voronoi method [10] (density values
used here are instantaneous and collected from the time the first participant enters
the measurement area to the time the last one leaves it). While average density is
almost equal in Proc. 1, the difference becomes already noticeable in the second
one. Later on we will show that this is due to the fact that dyad members tend to
walk close to each other compared to the other individual pedestrians, that instead
take more frequently the opportunities offered by unoccupied gaps in front of them.
A first analysis on the distributions of relative positions of dyads, with respect
to their centroid, has shown a decrease of distance between them with the increase
of counter-flow conditions [7]. Moreover, it is observed that conflicts arising from
the bi-directional flow frequently lead dyad members to assume a river-like pattern,
which is barely visible in the first procedure. As with the analysis on speed and
local density distribution, no significant difference arises between Proc. 3 and 4 of
the second experiment. The relation between density, speed and relative positions
is shown in Fig. 4(a) and (b). While a dependency between density and angular
Micro and Macro Pedestrian Dynamics in Counterflow
5
(a)
(b)
(c)
(d)
Fig. 4 (a – b) Relative position of dyads according to their centroid. Positions are rotated so that
the movement direction is up. Colors indicate the information of local density (a) or instantaneous
speed (b). (c – d) Relations between density, distance and speed of dyads.
arrangement (i.e. spatial pattern) of dyads was not found to be significant, it is ap-
parent how points of high densities are mostly close to the center (the few outliers
are probably due to a transient stretched river-like pattern) and they describe a pat-
tern with an elliptical shape, whose long side is associated to the walking direction.
The same regularity is also visible with the data about the speed: points associated
to higher speeds are located in the outer part of the dataset, while close to the center
speeds are lower, about 0.6 m/s.
The elliptical appearance of diagrams in Fig. 4(a) and (b) is not surprising and it
reflects the physics of pedestrian movement already considered with former works
on the modelling side (e.g. [4]). According to these data, it is possible to define a
6
Crociani, L., Gorrini, A., Feliciani, C., Vizzari, G., Nishinari, K., Bandini, S.
distance metric that applies distortion on the y-axis and helps to analyse the relation
between density, speed and elliptical distance:
(cid:114)
f (x,y) =
x2 +
(cid:16) y
(cid:17)2
2
The outcome of this analysis is shown in Fig. 4 (c) and (d). It is fair to state that
the defined elliptical distance between dyad members acts as a mediator between
the fundamental characteristics of the dynamics (more logically the local density
leads group members to walk closer and not vice-versa). On the other hand, this
analysis suggests that there is a positive effect of the density on the cohesion of
dyad members, that consequently affects their instantaneous speed. However, the
relation between walking speed and elliptical distance is less clear and points in
Fig. 4 (d) appear to be in a rather large area which is difficult to describe using a
linear function. Considering these observations, we can say that models of dyads
should be able to reproduce a growing trend between elliptical distance and speed.
An additional analysis carried out on microscopic data about the instantaneous
position of pedestrians is focused on evaluating the position of the other nearby
pedestrians: this kind of analysis, shown in Fig. 5, highlights a different kind of be-
haviour between members of dyads and individuals with regard to the lane formation
phenomenon. We focus in particular on Proc. 4 since it is the most interesting one
in terms of macroscopic results. It must be said that lane formation is a rather fuzzy
concept and several methods are proposed in the literature as attempt for its quantifi-
cation: [5], for example, analyse the rotation of the pedestrian directions to achieve
a numerical value describing the stability of lanes. We also do not try to provide
a definition of lane, but the data describing the proxemic behaviour of individuals
and dyads, respectively shown in Fig. 5(a) and (b), show that, on one side, there is
a clear following behaviour for the individuals, where the most frequent position of
neighbouring pedestrians is in a spot about 1 m ahead. On the other hand, members
of dyads mostly try to keep a line-abreast pattern: the most frequent positions for
neighbours are in fact on the side instead that ahead the considered pedestrian. In
other words, lanes composed of dyad members tended to be wider and this led to a
less efficient utilisation of the space available on average.
3.2 Effects of dyads at a macroscopic scale
Previous results highlighted the effects of density and counter-flow situations on the
behaviour of dyads at a very detailed scale whereas we present here the aggregated
effect of these microscopic observations on the overall pedestrian flow at different
levels of density.
The presence of groups in Proc. 1 did not bring to significant differences, since
only a simple free-flow situation emerged from it. In counter-flow situations, in-
stead, differences become apparent and the most interesting result is represented
by the scenario with a perfectly balanced counter-flow, whose data are reported in
Micro and Macro Pedestrian Dynamics in Counterflow
7
(a)
(b)
Fig. 5 Distribution of relative positions of neighbour pedestrians, according to the position of each
individual (a) or dyad member (b).
Fig. 6 Comparison of fundamental diagrams in the form density–speed (left) and density–flow
(right) of Exp. 1 and 2 - Proc. 4.
Fig. 6. The diagram shows very little difference at low densities, but starting from
0.5 peds/m2 the specific pedestrian flow observed in the experiment with dyads
grows at a slower rate, compared to the Exp. 1. The range of observed densities
does not reach a critical density for any of the experiments, but the trend of the di-
agrams supports the conjecture that the situation of Exp. 2 would lead to a lower
maximum flow.
8
4 Conclusions and Future Works
Crociani, L., Gorrini, A., Feliciani, C., Vizzari, G., Nishinari, K., Bandini, S.
The paper has presented original results of analyses of pedestrian dynamics achieved
through an experimental observation aimed at characterising the influence of dyads,
both at micro and macroscopic level. Micro-level results underline that different
counterflow situations affect local density, and that groups walk slower compared to
singletons, depending also on their spatial patterns at variable density situations. The
introduction of dyads in the pedestrian demand leads to a higher level of measurable
density in analogous initial conditions and a more chaotic macro-level dynamics
characterized by fragmented lanes, inducing a lower observed specific flow.
Future works are aimed, on one hand, to transfer the achieved results to the mod-
elling activities in presence of groups (preliminary results are discussed in another
paper in this volume [12]), but additional observations and experiments would be
needed to further investigate whether previously observed aggregated phenomena
are still observed in the presence of groups.
References
1. Bandini, S., Crociani, L., Gorrini, A., Vizzari, G.: An agent-based model of pedestrian dy-
namics considering groups: A real world case study. In: 17th International IEEE Conference
on Intelligent Transportation Systems (ITSC), pp. 572–577 (2014)
2. Bandini, S., Gorrini, A., Vizzari, G.: Towards an integrated approach to crowd analysis and
crowd synthesis: A case study and first results. Pattern Recognition Letters 44, 16–29 (2014)
3. Boltes, M., Seyfried, A.: Collecting pedestrian trajectories. Neurocomputing 100, 127–133
(2013)
4. Chraibi, M., Seyfried, A., Schadschneider, A.: Generalized centrifugal-force model for pedes-
trian dynamics. Phys. Rev. E 82(4), 46,111 (2010)
5. Feliciani, C., Nishinari, K.: Empirical analysis of the lane formation process in bidirectional
pedestrian flow. Physical Review E 94(3), 032,304 (2016)
6. Feliciani, C., Nishinari, K.: Estimation of capacity in bidirectional pedestrian flow by detecting
congestion and phase transitions (2017). Submitted, May 2017
7. Gorrini, A., Crociani, L., Feliciani, C., Zhao, P., Nishinari, K., Bandini, S.: Social groups
arXiv preprint
and pedestrian crowds: Experiment on dyads in a counter flow scenario.
arXiv:1610.08325 (2016)
8. von Kruchten, C., Schadschneider, A.: Empirical study on social groups in pedestrian evacu-
ation dynamics. Physica A: Statistical Mechanics and its Applications 475, 129–141 (2017)
9. Moussaıd, M., Perozo, N., Garnier, S., Helbing, D., Theraulaz, G.: The walking behaviour of
pedestrian social groups and its impact on crowd dynamics. PLoS ONE 5(4), e10,047 (2010)
10. Tordeux, A., Zhang, J., Steffen, B., Seyfried, A.: Quantitative comparison of estimations for
the density within pedestrian streams. Journal of statistical mechanics: theory and experiment
2015(6), P06,030 (2015)
11. Zanlungo, F., Ikeda, T., Kanda, T.: Potential for the dynamics of pedestrians in a socially
interacting group. Physical Review E 89(1), 012,811 (2014)
12. Zeng, Y., Crociani, L., Gorrini, A., Vizzari, G., Song, W.: Investigating the effect of social
groups in uni-directional pedestrian flow. In: in this volume
13. Zhang, J., Klingsch, W., Schadschneider, A., Seyfried, A.: Ordering in bidirectional pedestrian
flows and its influence on the fundamental diagram. Journal of Statistical Mechanics (02), 9
(2012)
|
1801.08159 | 1 | 1801 | 2018-01-24T19:38:00 | Adversarial Classification on Social Networks | [
"cs.MA",
"cs.SI"
] | The spread of unwanted or malicious content through social media has become a major challenge. Traditional examples of this include social network spam, but an important new concern is the propagation of fake news through social media. A common approach for mitigating this problem is by using standard statistical classification to distinguish malicious (e.g., fake news) instances from benign (e.g., actual news stories). However, such an approach ignores the fact that malicious instances propagate through the network, which is consequential both in quantifying consequences (e.g., fake news diffusing through the network), and capturing detection redundancy (bad content can be detected at different nodes). An additional concern is evasion attacks, whereby the generators of malicious instances modify the nature of these to escape detection. We model this problem as a Stackelberg game between the defender who is choosing parameters of the detection model, and an attacker, who is choosing both the node at which to initiate malicious spread, and the nature of malicious entities. We develop a novel bi-level programming approach for this problem, as well as a novel solution approach based on implicit function gradients, and experimentally demonstrate the advantage of our approach over alternatives which ignore network structure. | cs.MA | cs | Adversarial Classification on Social Networks
Sixie Yu
Yevgeniy Vorobeychik
Scott Alfeld
Electrical Engineering and Computer
Science, Vanderbilt University
Electrical Engineering and Computer
Science, Vanderbilt University
Computer Science, Amherst College
Amherst, MA
[email protected]
Nashville, TN
[email protected]
Nashville, TN
yevgeniy.vorobeychik@vanderbilt.
edu
8
1
0
2
n
a
J
4
2
]
A
M
.
s
c
[
1
v
9
5
1
8
0
.
1
0
8
1
:
v
i
X
r
a
ABSTRACT
The spread of unwanted or malicious content through social me-
dia has become a major challenge. Traditional examples of this
include social network spam, but an important new concern is the
propagation of fake news through social media. A common ap-
proach for mitigating this problem is by using standard statistical
classification to distinguish malicious (e.g., fake news) instances
from benign (e.g., actual news stories). However, such an approach
ignores the fact that malicious instances propagate through the
network, which is consequential both in quantifying consequences
(e.g., fake news diffusing through the network), and capturing de-
tection redundancy (bad content can be detected at different nodes).
An additional concern is evasion attacks, whereby the generators of
malicious instances modify the nature of these to escape detection.
We model this problem as a Stackelberg game between the defender
who is choosing parameters of the detection model, and an attacker,
who is choosing both the node at which to initiate malicious spread,
and the nature of malicious entities. We develop a novel bi-level
programming approach for this problem, as well as a novel solution
approach based on implicit function gradients, and experimentally
demonstrate the advantage of our approach over alternatives which
ignore network structure.
1 INTRODUCTION
Consider a large online social network, such as Facebook or Twitter.
It enables unprecedented levels of social interaction in the digital
space, as well as sharing of valuable information among individuals.
It is also a treasure trove of potentially vulnerable individuals to
exploit for unscrupulous parties who wish to gain an economic,
social, or political advantage. In a perfect world, the social network
is an enabler, allowing diffusion of valuable information. We can
think of this "benign" information as originating stochastically
from some node, and subsequently propagating over the network
to its neighbors (e.g., through retweeting a news story), then their
neighbors, and so on. But just as the network is a conduit for val-
ueable information, so it is for "malicious" content. However, such
undesirable content can be targeted: first, by selecting an influential
starting point on the network (akin to influence maximization), and
second, by tuning the content for maximal impact. For example, an
adversary may craft the headline of a fake news story to capture
the most attention. Consider the illustration in Figure 1, where an
Proc. of the 17th International Conference on Autonomous Agents and Multiagent Systems
(AAMAS 2018), M. Dastani, G. Sukthankar, E. Andre, S. Koenig (eds.), July 2018, Stockholm,
Sweden
© 2018 Authors of the paper
https://doi.org/doi
attacker crafts a fake news story and shares it with Adam. This
story is then shared by Adam with his friends, and so on.
Figure 1: An example of the propagation of malicious con-
tents.
These are not abstract concerns. Recently, widespread malicious
content (e.g., fake news, antisocial posts) in online social networks
has become a major concern. For example, considering that over
50% adults in the U.S. regard social media as their primary sources
for news [17], the negative impact of fake news can be substantial.
According to Allcott et al. [1] over 37 million news stories that are
later proved fake were shared on Facebook in the last three months
of 2016 U.S. presidential election. In addition to fake news, anti-
social posts in online communities negatively affect other users
and damage community dynamics [9], while social network spam
and phish can defraud users and spread malicious software [11].
The managers of online social networks are not powerless against
these threats, and can deploy detection methods, such as statistical
classifiers, to identify and stop the spread of malicious content.
However, such traditional mitigations have not as yet proved ad-
equate. We focus on two of the reasons for this: first, adversaries
can tune content to avoid being detected, and second, traditional
learning approaches do not account for network structure. The im-
plication of network structure mediating both spread and detection
has in turn two consequences: first, we have to account for impact
of detection errors in terms of benign or malicious content subse-
quently propgatating through the network, and second, the fact that
we can potentially detect malicious content at multiple nodes on
the network creates a degree of redundancy. Consequently, while
traditional detection methods use training data to learn a single
"global" classifier of malicious and benign content, we show that
specializing such learning to network structure, and using different
classifiers at different nodes can dramatically improve performance.
To address the problem of malicious content detection on social
networks, we propose two significant modeling innovations. First,
we explicitly model the diffusion process of content over networks
AAMAS'18, July 2018, Stockholm, Sweden
Sixie Yu, Yevgeniy Vorobeychik, and Scott Alfeld
as a function of content (or, rather, features thereof). This is a gener-
alization of typical network influence models which abstract away
the nature of information being shared. It is also a crucial general-
ization in our setting, as it allows us to directly model the balancing
act by the attacker between increasing social influence and avoiding
detection. Second, we consider the problem of designing a collection
of heterogeneous statistical detectors which explicitly account for net-
work structure and diffusion at the level of individual nodes, rather
than merely training data of past benign and malicious instances.
We formalize the overall problem faced as a Stackelberg game be-
tween a defender (manager of the social network) who deploys a
collection of heterogeneous detectors, and an attacker who opti-
mally chooses both the starting node for malicious content, and the
content itself. This results in a complex bi-level optimization prob-
lem, and we introduce a novel technical approach for solving it, first
considering a naive model in which the defender knows the node
being attacked, which allows us to develop a projected gradient
descent approach for solving this restricted problem, and subse-
quently utilizing this to devise a heuristic algorithm for tackling the
original problem. We show that our approach offers a dramatic im-
provement over both traditional homogeneous statistical detection
and a common adversarial classification approach.
Related Work. A number of prior efforts have considered lim-
iting adversarial influence on social networks. Most of these pit
two influence maximization players against one another, with both
choosing a subset of nodes to maximize the spread of their own
influence (blocking the influence of the other). For example, Cerenet
et al. [7] consider the problem of blocking a "bad" campaign using
a "good" campaign that spreads and thereby neutralizes the "bad"
influence. Similarly, Tsai et al. [26] study a zero-sum game between
two parties with competing interests in a networked environment,
with each party choosing a subset of nodes for initial influence.
Vorobeychik et al. [28] considered an influence blocking game in
which the defender chooses from a small set of security configu-
rations for each node, while the attacker chooses an initial set of
nodes to start a malicious cascade. The main differences between
this prior work and ours is that (a) our diffusion process depends on
the malicious content in addition to network topology, (b) detection
at each node is explicitly accomplished using machine learning
techniques, rather than an abstract small set of configurations, and
(c) we consider an attacker who, in addition to choosing the starting
point of a malicious cascade, chooses the content in part to evade
the machine learning-based detectors. The issue of using hetero-
geneous (personalized) filters was previously studied by Laszka
et al. [19], but this work did not consider network structure or
adversarial evasion.
Our paper is also related to prior research in single-agent in-
fluence maximization and adversarial learning. Kempe et al. [18]
initiated the study of influence maximization, where the goal is to
select a set of nodes to maximize the total subset of network affected
for discrete-time diffusion processes. Rodriguez et al. [16] and Du
et al. [13–15] considered the continuous-time diffusion process to
model information diffusion; we extend this model. Prior adver-
sarial machine learning work, in turn, focuses on the design of a
single detector (classifier) that is robust to evasion attacks [6, 12, 20].
However, this work does not consider malicious content spreading
over a social network.
2 MODEL
We are interested in protecting a set of agents on a social network
from malicious content originating from an external source, while
allowing regular (benign) content to diffuse. The social network is
represented by a graph G = (V , E), where V is the set of vertices
(agents) and E is the set of edges. An edge between a pair of nodes
represents communication or influence between them. For example,
an edge from i to j may mean that j can see and repost a video
or a news article shared by i. For simplicity, we assume that the
network is undirected; generalization is direct.
We suppose that each message (benign or malicious) originates
from a node on the network (which may differ for different mes-
sages) and then propagates to others. We utilize a finite set of
instances diffusing over the network (of both malicious and benign
content) as a training dataset D. Each instance, malicious or benign,
is represented by a feature vector x ∈ Rn where n is the dimension
of the feature space. The dataset D is partitioned into D+ and D−,
where D+ corresponds to malicious and D− to benign instances.
To analyze the diffusion of benign and malicious content on
social networks in the presence of an adversary, we develop formal
models of (a) the diffusion process, (b) the defender who aims to
prevent the spread of malicious content while allowing benign
content diffuse, (c) the attacker who attempts to maximize the
influence of a malicious message, and (d) the game between the
attacker and defender. We present these next.
2.1 Continuous-Time Diffusion
Given an undirected network with a known topology, we use a
continuous-time diffusion process to model the propagation of con-
tent (malicious or benign) through the social network, extending
Rodriguez et al. [16]. In our model, diffusion will depend not merely
on the network structure, but also on the nature of the item propa-
gating through the network, which we quantify by a feature vector
x as above.
Suppose that the diffusion process for a single message originates
at a node s. First, x is transmitted from s to its direct neighbors. The
time taken by a propagation through an edge e is sampled from
a distribution over time, fe(t; we , x), which is a function of the
edge itself and the entity x, and parametrized by we. The affected
(influenced) neighbors of s then propagate x to their neighbors, and
so on. We assume that an affected agent remains affected through
the diffusion process.
Given a sample of propagation times over all edges, the time ti
taken to affect an agent i is the length of the shortest path between s
and i, where the weights of edges are propagation times associated
with these edges. The continuous-time diffusion process is supplied
with a time window T , which is used to simulate time-sensitive
natures of propagation, for example, people are generally concerned
about a news for several months but not for years. An agent is
affected if and only if its shortest path to s is less than or equal
to T . The diffusion process terminates when the path from s to
every unaffected agent is above T . We define the influence σ(s, x)
Adversarial Classification on Social Networks
AAMAS'18, July 2018, Stockholm, Sweden
of an instance x initially affecting a network node s as the expected
number of affected agents over a fixed time window T .
Figure 2: Rayleigh distributions with different 1/γ 2.
We assume that the distributions associated with edges are
Rayleigh distributions (illustrated in Figure 2), which have den-
γ 2 e−t 2/(2γ 2), where t ≥ 0 and γ is the scale
sity function f (t; γ) = t
parameter.1 The Rayleigh distribution is commonly used in epi-
demiology and survival analysis [29] and has been recently applied
to model information diffusion in social networks [14, 16]. In order
to account for heterogeneity among mutual interactions of agents,
and to let the influence of a process depend on the content being
diffused, we parameterize the Rayleigh distribution of each edge
by letting 1/γ 2 = wT x, where w is sampled from the uniform dis-
tribution over [0, 1]. This parameterization results in the following
density function for an arbitrary edge:
T x)e
(1)
We denote by W = {we∀e ∈ E} the joint parametrization of all
edges.
Throughout, we assume that the parameters W are given, and
known to both the defender and attacker. A number of other re-
search studies explore how to learn these parameter vectors from
data [14, 15].
fe(t; we , x) = t(we
2 t 2(we
− 1
T x)
.
2.2 Defender Model
To protect the network against the spread of malicious content, the
network manager-henceforth, the defender-can deploy statisti-
cal detection, which considers a particular instance (e.g., a tweet
with an embedded link) and decides whether or not it is safe. The
traditional way of deploying such a system is to take the dataset
D of labeled malicious and benign examples, train a classifier, and
use it to detect new malicious content. However, this approach
entirely ignores the fact that a social network mediates the spread
of both malicious and benign entities. Moreover, both the nature
(as captured in the feature vector) and the origin of malicious in-
stances are deliberate decisions by the adversary aiming to maximize
impact (and harm, from the defender's perspective). Our key inno-
vations are (a) to allow heterogeneous parametrization of classifiers
1It is straightforward to allow for alternative distributions, such as Weibull.
deployed at different nodes, and (b) to explicitly consider both diffu-
sion and adversarial manipulation during learning. In combination,
this enables us to significantly boost detection effectiveness in social
network settings.
Let Θ = {θ1, θ2, · · · , θV } be a vector of parameters of detection
models deployed on the network where each θi ∈ Θ represents
the model used for content shared by node i.2 We now extend
our definition of expected influence to be a function of detector
parameters, denoting it by σ(i, Θ, x), since any content x (malicious
or benign) starting at node i which is classified as malicious at a
node j (not necessarily the same as i) will be blocked from spreading
any further.
We define the defender's utility as
σ(i, Θ, x) − (1 − α)
Ud = α
x ∈D−
i∈V
σ(s, Θ, z(x)),
(2)
x ∈D +
where s is the starting node targeted by the adversary, which is
subsequently modified by the same adversary into z(x) (in an at-
tempt to bypass detection) when the original content used by the
adversary is x. The first part of the utility represents the influence
of benign content that the defender wishes to maximize, while the
second part denotes the influence of malicious content that the
defender aims to limit, with α trading off the relative importance
of these two considerations. Observe that we assume that benign
content originates uniformly over the set of nodes, while malicious
origin is selected by the adversary. The defender's action space is
the set of all possible parameters Θ of the detectors deployed at all
network nodes. Note that, as is typical in machine learning, we are
using the finite labeled dataset D as a proxy for expected utility
with respect to malicious and benign content generated from the
same distribution as the data.
2.3 Attacker Model
The attacker's decision is twofold: (1) find a node s ∈ V to start
diffusion; and (2) transform malicious content from x (its original,
or ideal, form) into another feature vector z(x) with the aim of
avoiding detection. The first decision is reminiscent of the influ-
ence maximization problem[18]. The second decision is commonly
known as the evasion attack on classifiers [20, 23]. In our case, the
adversary attempts to balance three considerations: (a) impact, me-
diated by the diffusion of malicious content, (b) evasion, or avoiding
being detected (a critical consideration for impact as well), and (c) a
cost of modifying original "ideal" content into another form, which
corresponds to the associated reduced effectiveness of the trans-
formed content, or effort involved in the transformation. We impose
this last consideration as a hard constraint that z(x)− xp ≤ ϵ for
an exogenously specified ϵ, where ∥ · ∥p is the lp norm.
Consider the collection of detectors with parameters Θ deployed
on the network. We say that a malicious instance is detected at
a node i if 1[θi(x) = 1] = 1, where 1(·) is the 0-1 indicator func-
tion. The optimization problem of the attacker corresponding to an
2Below, we focus on θi corresponding to detection thresholds as an illustration;
generalization is direct.
AAMAS'18, July 2018, Stockholm, Sweden
Sixie Yu, Yevgeniy Vorobeychik, and Scott Alfeld
original malicious instance x ∈ D+ is then:
max
i,z
s.t
σ(i, Θ, z)
z − xp ≤ ϵ
1[θj(z) = 1] = 0,∀j ∈ V
(3)
where the first constraint is the attacker's budget limit, while the
second constraint requires that the attack instance z remains unde-
tected. If Problem (3) does not have a feasible solution, the attacker
sends the original malicious instance without any modification.
Consequently, the pair (s, z(x)) in the defender's utility function
above are the solutions to Problem (3).
2.4 Stackelberg Game Formulation
We formally model the interaction between the defender and the
attacker as a Stackelberg game in which the defender is the leader
(choosing parameters of node-level detectors) and the attacker the
follower (choosing a node to start malicious diffusion, as well as the
content thereof). We assume that the attacker knows Θ, as well as
all relevant parameters (such as W) before constructing its attack.
The equilibrium of this game is the joint choice of (Θ, s(Θ), z(x; Θ)),
where s(Θ) and z(x; Θ) solve Problem (3), thereby maximizing the
attacker's utility, and Θ maximizes the defender's utility given s and
z. More precisely, we aim to find a Strong Stackelberg Equilibrium
(SSE), where the attacker breaks ties in the defender's favor.
We propose finding solutions to this Stackelberg game using the
following optimization problem:
σ(i, Θ, x) − (1 − α)
σ(s, Θ, z(x))
max
Θ
s.t . :
α
x ∈D−
∀x ∈ D
∀x ∈ D
∀x ∈ D
i
+ :
+ :
+ :
x ∈D +
σ(j, Θ, z)
(s, z(x)) ∈ arg max
z(x) − xp ≤ ϵ
1[θk(x) = 1] = 0,∀k ∈ V
j,z
(4)
problem by leveraging Karush-Kuhn-Tucker (KKT) [5] conditions.
The assumption that the defender knows the node being attacked
allows us to solve the resulting single-level optimization problem
using projected gradient descent.
3.1 Collapsing the Bilevel Problem
A continuous-time diffusion process proceeds in a breadth-first-
search fashion. It starts from an agent i trying to influence each of
its neighbors. Then its neighbors try to influence their neighbors,
and so on. Notice that once an agent becomes affected, it is no
longer affected by others. The main consequence of this propagation
process is that it results in a propagation tree rooted at i, with
its structure intimately connected to the starting node i. This is
where we leverage the assumption that the defender knows the
starting node of the attack: in this case, the tree structure can be
pre-computed, and fixed for the optimization problem.
We divide the agents traversed by the tree into several layers in
terms of their distances to the source, where each layer is indexed
by l. Since the structure of the tree depends on i, l is a function of
i, l(i). An example of the influence propagation tree is depicted in
Figure 3, where the first layer consists of {j, k, · · · , д}. The number
next to each edge represents the weight sampled from the associated
distribution.
We define a matrix Al ∈ RNl×n where Nl is the number of
agents in layer l and n is the feature dimension of x. Each row of Al
corresponds to the parametrization vector w of an edge in layer l
(an edge is in layer l if one of its endpoint is in layer l − 1 while the
other is in layer l; the source is always in layer zero). For example, in
Figure 3, A1 = [wT
iд]. The product of Al x is a vector
in RNl , where each element corresponds to the parameter 1/γ 2 of
an edge in layer l.
ik; · · · ; wT
ij; wT
This is a hierarchical optimization problem, where the upper-level
optimization corresponds to maximizing the defender's utility. The
constraints of the upper-level optimization are called the lower-
level optimization, which is the attacker's optimization problem.
The optimization problem (4) is generally intractable for several
reasons. First, Problem (4) is a bilevel optimization problem [10],
which is hard even when the upper- and lower-level problems are
both linear [10]. The second difficulty lies in maximizing σ(i, Θ, x)
(the attacker's problem), as the objective function does not have
an explicit closed-form expression. In what follows, we develop a
principled approach to address these technical challenges.
3 SOLUTION APPROACH
We start by making a strong assumption that the defender knows
the node being attacked. This will allow us to make considerable
progress in transforming the problem into a significantly more
tractable form. Subsequently, we relax this assumption, develop-
ing an effective heuristic algorithm for computing the SSE of the
original problem.
First, we utilize the tree structure of a continuous-time diffusion
process to convert (4) into a tractable bilevel optimization. We then
collapse the bilevel optimization into a single-level optimization
Figure 3: A example continuous-time diffusion process.
Recall that a sample of random variables from Rayleigh distri-
butions associated with edges corresponds to a sample of weights
associated with these edges. With a fixed time window T , small
edge weights result in wider diffusion of the content over the social
network. For example, in Figure 3 if the number next to each edge
represents a sample of weights, then with T = 1 the propagation
starting from i can only reach agents j and k. However, if we assume
that in another sample ti, j , ti,k , ti,д all become 0.1, then with the
same time window the propagation can reach every agent in the net-
work. Consequently, the attacker's goal is to increase 1/γ 2 = wT
e x
for each edge e. This suggests that in order to increase 1/γ 2 the
attacker can modify the malicious instance x such that the inner
products between x and the parameter vectors we of edges are
Adversarial Classification on Social Networks
AAMAS'18, July 2018, Stockholm, Sweden
noting again that l and Nl depend on j, the starting node of the
diffusion process. From the expression (8), the defender tries to find
Θ that minimizes the impact of false positives while maximizing
the impact of true negatives. This is because if each benign instance
x ∈ D− is correctly identified (false-positive rates are zero and
true-negative rates are one), the summation at the second line of
expression (8) will attain its maximum possible value.
In addition to facilitating the propagation of benign contents,
the defender wants to limit the propagation of malicious contents,
which is embodied in 2 . The equations in 2 are similar to those
in 1 , except that the summation is over malicious contents D+,
and 2 is accounting for the false negatives. In this case, cl,s is a
function of z(x), the adversarial feature vector which transforms x
into another, z.
We now re-formulate the problem (4) as a new bilevel optimiza-
tion problem (9). The upper-level problem corresponds to the de-
fender's strategy (6), and the lower-level problem to the attacker's
optimization problem (5). Here, s is again the node chosen by the
attacker.
(1 − α)
1[θs(x) = 0]
1[θj(x) = 0]
x ∈D +
− α
l
min
Θ
j
x ∈D−
s.t : ∀x ∈ D
kl cT
l,s Al z(x)
kl cT
l, j Al x
l
l
+ : z(x) ← arg max
kl 1T Al z
(9)
z
(6)
s.t . ∀x ∈ D
∀x ∈ D
∀x ∈ D
+ : z(x) − xp ≤ ϵ
+ : 1[θk(z(x)) = 1] = 0,∀k ∈ V
+ : z(x) ⪰ 0,
where the last constraint ensures that wT z(x) ≥ 0 for all attacks
z(x).
The final step, inspired by [24, 25], is to convert (9) into a single-
level optimization problem via the KKT [5] conditions of the lower-
level problem. With appropriate norm constraints (e.g., l2 norm)
and a convex relaxation of the indicator functions (i.e., convex
surrogates of the indicator functions), the lower-level problem of
(9) is convex. A convex optimization problem can be equivalently
represented by its KKT conditions[8]. The single-level optimization
problem then becomes:
(7)
(cid:18)
Fd
min
s.t . ∀x :
Θ
−
∂z
large. Consequently, we can formulate the attacker's optimization
problem with respect to malicious content z for a given original
feature vector x as
max
z
s.t .
l
kl 1T Al z
z − xp ≤ ϵ
1[θk(z) = 1] = 0,∀k ∈ V .
(5)
The attacker aims to make 1/γ 2 for each edge as large as possible,
which is captured by the objective function 1T Al z, where 1 ∈ RNl
is a vector with all elements equal to one. Intuitively, this means the
attacker is trying to maximize on average the parameter 1/γ 2 of
every edge at layer l. Here, [k1, k2, · · · , kl] is a vector of decreasing
coefficients that provides more flexibility to modeling the attacker's
behavior: they are used to re-weight the importance of each layer.
For example, setting k1 = e0, k2 = e−1, · · · , kl = e−l models the
attacker who tries to make malicious instances spread wider at the
earlier layers of the diffusion.
We now use similar ideas to convert the upper-level optimization
problem of (4) into a more tractable form. Suppose that the node
being attacked is s (and known to the defender). Then the defender
wants the detection model at j to accurately identify both malicious
and benign contents. This is achieved by the two indicator func-
tions inside 1 and 2 in the reformulated objective function of the
defender (6):
max
Θ
α
j
kl cT
x ∈D−
1[θj(x) = 0]
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
− (1 − α)
1[θs(z(x)) = 0]
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
l,s Al z(x)
l, j Al x
x ∈D +
kl cT
1
l
l
2
Notice that this expression includes a vector cl, j ∈ RNl that does
not appear in (5). cl, j is a function of Θ and x, for a given node j
which triggers diffusion (which we omit below for clarity):
1[θl1(x) = 0]
1[θl2(x) = 0]
...
(x) = 0].
1[θlNl
cl, j =
(x) = 0](cid:105)
Slightly abusing notation, we let li , i ∈ [1, 2, · · · , Nl] denote the
ith agent in layer l. The term kl cT
l, j Al x in 1 can be expanded as
follows:
kl cT
l, j Al x
(cid:104)
(cid:18)
= kl
= kl
1[θl1(x) = 0], . . . , 1[θlNl
1[θl1(x) = 0]wT
l1x + · · · + 1[θlNl
x
wT
l1
...
wT
lNl
x
(x) = 0]wT
lNl
(8)
(cid:19)
x
,
(cid:19)
= 0
(10)
l,s Al z + λд(z, x) + µT h(z, Θ) − ηT z
l
kl cT
λд(z, x) = 0, λ ≥ 0
д(z, x) ≤ 0
η ⊙ (−z) = 0, η ⪰ 0
h(z, Θ) = 0
where Fd is the objective function of Problem (9), and λ, µ, η are
vectors of lagrangian multipliers. д(z, x) = z − xp − ϵ ≤ 0 is the
attacker's budget constraint. h(x, Θ) is the set of equality constraints
Sixie Yu, Yevgeniy Vorobeychik, and Scott Alfeld
implicitly define z(x) in terms of Θ:
l kl cT
l,s Al z + λд(z, x) + µT h(z, Θ) − ηT z
(14)
f(Θ, z, λ, µ, η) =
(cid:18)
−
∂z
(cid:19)
Θ and the attacked malicious instance z satisfy f(Θ, z, λ, µ, η) = 0.
The Implicit Function Theorem[31] states that if f(Θ, z, λ, µ, η) is
continuous and differentiable and the Jacobian matrix
λд(z, x)
µT h(z, Θ)
η ⊙ (−z)
(cid:20) ∂f
∂z
∂f
∂λ
∂f
∂µ
∂f
∂η
(cid:21)
AAMAS'18, July 2018, Stockholm, Sweden
1[θj(z) = 1] = 0,∀j ∈ V . η ⊙ (−z) is the Hadamard (elementwise)
product between η and (−z) .
3.2 Projected Gradient Descent
In this section we demonstrate how to solve the single-level opti-
mization obtained above by projected gradient descent. The key
technical challenge is that we don't have an explicit representation
of the gradients with respect to the defender's decision Θ, as these
are indirectly related via the optimal solution to the attacker's opti-
mization problem. We derive these gradients based on the implicit
function of the defender's utility with respect to Θ.
We begin by outlining the overall iterative projected gradient
descent procedure. In iteration t we update the parameters of de-
tection models by taking a projected gradient step:
(t) − βt∇Θ Fd
(t +1) = ProjAd
Θ
Θ
(11)
where Ad is the feasible domain of Θ and βt is the learning rate.
With Θ(t +1) we solve for z(t +1), which is the optimal attack for a
fixed Θ(t +1). ∇Θ Fd is the gradient of the upper-level problem.
Expanding ∇Θ Fd using the chain rule and still using s as the
initially attacked node, we obtain
(cid:19)
(cid:12)(cid:12)Θ=Θ(t)
(cid:18)
∇Θ Fd = (1 − α) 1 − α 2
1 =
x ∈D +
2 =
x ∈D−
∂Θ
(cid:21)
l
l kl cT
∂Θ
(cid:20) ∂1[θs(z(x)) = 0]
1[θs(z(x)) = 0] ∂[
l,s Al z(x)+
kl cT
l,s Al z(x)]
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
(cid:20) ∂1[θj(x) = 0]
(a)
1[θj(x) = 0] ∂[
(cid:21)
kl cT
l, j Al x]
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
l
l kl cT
∂Θ
l, j Al x+
∂Θ
j
(b)
In
In both 1 and 2 we note that ∂1[θj(x)=0]
is dependent on the
specific detection models. We will give a concrete example of their
derivation in Section 3.5.
l,s Al z(x) there are two terms that are functions of Θ:
cl,s and z(x). Consequently, (a) can be expanded as:
l kl cT
∂Θ
(cid:20) ∂z(x)
(cid:21)T
∂Θl
(cid:21)
Al z(x) +
AT
l cl,s
.
(13)
(a) =
kl
l
(cid:20) ∂cl,s
∂Θl
(12)
kl
Al x
.
(16)
has full rank, there is a unique implicit function I(Θ) = (z, λ, µ, η).
Moreover, the derivative of ∂I
= −(cid:104) ∂f
∂Θ is:
∂λ ∂f
∂z ∂f
∂µ ∂f
∂η
(cid:105)−1(cid:18) ∂f
(cid:19)
∂Θ
∂I
∂Θ
.
(15)
∂z is the Jacobian matrix of f(Θ, z, λ, µ, η) with respect to z, and
∂f
∈ Rn×N is the first n rows of ∂I
so on. ∂z
can be
∂Θ
column-wise indexed by the nodes at layer l.
(b) can be similarly expanded as we had done for (a), except that
the attacker does not modify benign content, so that x ∈ D− is no
longer a function of Θ:
∂Θ, where ∂z
∂Θl
(b) =
l
j
(cid:20) ∂cl, j
∂Θl
(cid:21)
The full projected gradient descent approach is given by Algo-
rithm 1.
Algorithm 1 Find Defense Strategy
1: Input: agent j
2: Initialize: Θ(0), λ, µ, η, β0
3: for t = 1 · · · k do
Θ(t +1) = ProjAd
4:
5: end for
6: return Θ(k +1)
(cid:18)
Θ(t) − βt∇Θ Fd
(cid:19)
(cid:12)(cid:12)Θ=Θ(t)
Note that only the detection models of those agents at layer l have
contribution to cl,s. Thus, ∂cl,s
is a Jacobian matrix with dimension
Nl × Nl , where Nl is the number of agents at layer l and Θl denotes
∂Θl
the detection models of those Nl agents. Since cl,s is also dependent
on the specific detection models of agents, we defer its derivation
to Section 3.5.
is a n × Nl Jacobian matrix and is the main difficulty
because we do not have an explicit function of the attacker's optimal
decision z(x) with respect to Θl . Fortunately, the constraints in (10)
∂z(x)
∂Θl
3.3 Optimal Attack
So far, we had assumed that the network node being attacked is
fixed. However, the ultimate goal is to allow the attacker to choose
both the node s, and the modification of the malicious content z. We
begin our generalization by first allowing the attacker to optimize
these jointly.
The full attacker algorithm which results is described in Algo-
rithm 2.
Adversarial Classification on Social Networks
AAMAS'18, July 2018, Stockholm, Sweden
Algorithm 2 Optimal Attack Strategy
1: Input: Θ, x
2: Initialize: ret = []
3: for i = 1 · · · V do
x(i) ← Solve (5)
4:
Ua(i) ← Optimal objective value of (5)
5:
6:
7: end for
8: z, s ← OptimalTuple(ret)
9: return z, s
(cid:0)i, z(i, x), Ua(i)(cid:1) appended to ret
Recall that the tree structure of a propagation is dependent on
the agent being attacked, which makes the objective function of (5)
a function of the agent being attacked. Thus, for a given fixed Θ, the
attacker iterates through each agent i and solves the problem (5),
assuming the propagation starts from i, resulting in associated util-
ity Ua(i) and an attacked instance z(i, x). Then i, z(i, x), and Ua(i)
are appended into a list of a 3-tuples (the sixth step in Algorithm
2). When the iteration completes the attacker picks the optimal
3-tuple in terms of utility (eighth step in Algorithm 2, where the
function OptimalTuple(ret) finds the optimal 3-tuple from the list
ret). The node s and the corresponding attack instance z in this
optimal 3-tuple become the optimal attack.
3.4 SSE Heuristic Algorithm
Now we take the final step, relaxing the assumption that the at-
tacker chooses a fixed node to attack which is known to the de-
fender prior to choosing Θ. Our main observation is that fixing
s in the defender's algorithm above allows us to find a collection
of heterogeneous detector parameters Θ, and we can evaluate the
actual utility of the associated defense (i.e., if the attacker optimally
chooses both s and z in response) by using Algorithm 2. We use
this insight to devise a simple heuristic: iterate over all potential
nodes s that can be attacked, compute the associated defense Θ(s)
(using the optimistic definition of defender's utility in which s is
assumed fixed), then find the actual optimal attack in response for
each x ∈ D+. Finally, choose the Θ(s) which has the best actual
defender utility.
This heuristic algorithm is described in Algorithm 3. The fifth
Algorithm 3 Optimal Defense Strategy
1: Input: G = (V , E),W, D
2: for j = 1 · · · V do
3:
4:
5:
6: end for
7: j ← arg maxj Ud(j)
8: return Θj
Θj ← Apply Algorithm 1
∀x ∈ D+ : (s, z(x)) ← Apply Algorithm 2
Ud(j) ← DefenderUtility(Θj ,(s, z(x)))
step in the algorithm includes the function DefenderUtility, which
evaluates the defender's utility Ud(j). Note that the input argument
s of this function is used to determine the tree structure of the
propagation started from s.
∂Θ
and ∂cl, j
∂Θl
Recall that Algorithm 1 solves (10), which depends on the specific
detection model to compute the relevant gradients. Therefore, in
what follows, we present a concrete example of how to solve (10)
where detection models are logistic regressions. Specifically, we
illustrate how to derive the two terms, ∂1[θj(z)=0]
that
depend on particular details of the detection model.
3.5 Illustration: Logistic Regression
We consider the logistic regression model used for detection at
individual nodes to illustrate the ideas developed above. For a node
i, its detection model has two components: the logistic regression
1+e−ϕT x , where ϕ is the weight vector of the logistic regression
and x the instance propagated to i, and a detection threshold θi
(which is the parameter the defender will optimize). An instance is
≤ θi. Thus (slightly abusing notation
classified as benign if
as before), θi(x) (cid:44) 0 (x is classified as malicious) if
≥ θi.
With the specific forms of the detection models we can derive
∂1[θj(x)=0]
(omitting the node index s or j for clarity). A
technical challenge is that the indicator function 1(·) is not con-
tinuous or differentiable, which means that it's difficult to char-
acterize its derivative with respect to Θ. However, observe that
and ∂cl
∂Θl
1+e−ϕT x
1+e−ϕT x
∂Θ
1
1
1
1
1−θj
1+e−ϕT x
function for 1[·]. Then ∂1[θj(x)=0]
the jth element equal to
for logistic regression θj(x) = 0(cid:0)
log(cid:0) θj
(cid:1) ≥ ϕT x. Therefore we use log(cid:0) θj
(cid:1) − ϕT x
(cid:1) − ϕT x
log(cid:0) θl1
log(cid:0) θl2
1−θl1
1−θl2
...
∂Θ
θj−θ 2
cl =
1−θj
1
j
≤ θj
(cid:1) is equivalent to
(cid:1)−ϕT x as a surrogate
(17)
and ∂cl
∂Θl
becomes a Nl × Nl diagonal matrix:
is a N -dimension vector with
. The cl vector then becomes:
1
θl1−θ 2
l1
. . .
∂cl
∂Θl
=
1
θNl −θ 2
Nl
(18)
and ∂1[θj(x)=0]
∂Θ
With equations (12)-(16), ∂cl
, we can now calculate
∂Θl
∇Θ Fd. Since the thresholds θi ∈ [0, 1], the defender's action space
is [0, 1]N . When updating Θ by (11) we therefore project it back to
[0, 1]N in each iteration.
4 EXPERIMENTS
In this section we experimentally evaluate our proposed approach.
We used the Spam dataset [22] from UCI machine learning reposi-
tory as the training dataset for the logistic regression model. The
Spam dataset contains 4601 emails, where each email is represented
by a 57-dimension feature vector. We divided the dataset into three
disjoint subsets: D′ used to learn the logistic regression (tuning
the weight vectors with thresholds setting to 0.5) as well as other
models to which we compare, Dtrain used in Algorithm 3 to find
the optimal defense strategy, and Dtest to test the performance of
AAMAS'18, July 2018, Stockholm, Sweden
Sixie Yu, Yevgeniy Vorobeychik, and Scott Alfeld
the defense strategy. The sizes of D′, Dtrain, and Dtest are 3681, 460,
and 460, respectively. They are all randomly sampled from D.
Our experiments were conducted on two synthetic networks
with 64 nodes: Barabasi-Albert preferential attachment networks
(BA) [2] and Watts-Strogatz networks (Small-World) [30]. BA is
characterized by its power-law degree distribution, where connec-
tivity is heavily skewed towards high-degree nodes. The power-law
degree distribution, P(k) ∼ k−r , gives the probability that a ran-
domly selected node has k neighbors. The degree distributions of
many real-world social networks have previously been shown to
be reasonably approximated by the power-law distribution with
r ∈ [2.1, 2.4] [3]. Our experiments for BA were conducted across
two sets of parameters: r = 2.1 and r = 2.3.
The Small-World topology is well-known for balancing short-
est path distance between pairs of nodes and local clustering in a
way as to qualitatively resemble real networks [27]. In our experi-
ments we consider two kinds of Small-World networks. The first
has average length of shortest path equal to 5.9 and local clustering
coefficient equal to 0.144. In this case the local clustering coeffi-
cient is close to what had been observed in large-scale Facebook
friendship networks [27]. The second one has average shortest path
length of 5 and local clustering coefficient of 0.08, where the local
clustering coefficient is close to that for the electric power grid of
the western United States [30].
Our node-level detectors use logistic regression, with our algo-
rithm producing the threshold for these. The trade-off parameter
α was set to 0.5 and the time window T was set to 1. We applied
standard pre-processing techniques to transform each feature to
lie between zero and one. The attacker's budget is measured by
squared l2 norm and the budget limit ϵ is varied from 0.001 to 0.01.
We compare our strategy with three others based on traditional
approaches: Baseline, Re-training, and Personalized-single-threshold;
we describe these next.
Baseline: This is the typical approach which simply learns a
logistic regression on training data, sets all thresholds to 0.5, and
deploys this model at all nodes.
Re-training: The idea of re-training, common in adversarial
classification, is to iteratively augment the original training data
with attacked instances, re-training the logistic regression each
time, until convergence [4, 21]. The logistic regressions deployed
at the nodes are homogeneous, with all thresholds being 0.5.
Personalized-single-threshold: This strategy is only allowed
to tune a single agent's threshold. It has access to Dtrain that in-
cludes unattacked emails. The strategy iterates throught each node
i and finds its optimal threshold. The optimality is measured by the
defender's utility as defined in (2), where the expected influence of
an instance is estimated by simulating 1000 propagations started
from i. Then the strategy picks the node with largest utility and
sets its optimal threshold.
As stated earlier, network topologies and parameter vectors as-
sociated with edges are assumed to be known by both the defender
and the attacker. The attacker has full knowledge about the defense
strategy, including the weight vectors of logistic regressions as well
as their thresholds. As in the definition of Stackelberg game, the
evaluation procedure lets the defender first choose its strategy Θ∗,
and then the attacker computes its best response, which chooses
the initial node for the attack s and transformations of malicious
content z aimed at evading the classifier. Finally the defender's util-
ity is calculated by (2), where the expected influence is estimated by
simulating 1000 propagations originating from s for each malicious
instance z.
The experimental results for BA (r = 2.1) and Small-World (aver-
age length of shortest path=5.9 and local clustering coefficient=0.144)
are shown in Figure 4, and the experimental results for BA (r = 2.3)
and Small-World (average length of shortest path=5 and local clus-
tering coefficient=0.08) are shown in Figure 5.
Figure 4: The performance of each defense strategy. Each
bar is averaged over 10 random topologies. Left: BA. Right:
Small-world)
Figure 5: The performance of each defense strategy. Each
bar is averaged over 10 random topologies. Left: BA. Right:
Small-world)
As we can observe from the experiments, our algorithm out-
performs all of the alternatives in nearly every instance; the sole
exception is when the attacker budget is 0.001, which effectively
eliminates the adversarial component from learning. For larger bud-
gets, our algorithm remarkably robust even as other algorithms per-
form quite poorly, so that when ϵ = 0.01, there is a rather dramatic
gap between our approach and all alternatives. Not surprisingly,
the most dramatic differences can be observed in the BA topology:
with the large variance in the degree distribution of different nodes,
our heterogeneous detection is particularly valuable in this setting.
In contrast, the degradation of the other methods on Small-World
topologies is not quite as dramatic, although the improvement of-
fered by the proposed approach is still quite pronounced. Among
the alternatives, it is also revealing that personalizing thresholds re-
sults in second-best performance: again, takng account of network
topology is crucial; somewhat surprisingly, it often outperforms
re-training, which explicitly accounts for adversarial evasion, but
not network topology.
Adversarial Classification on Social Networks
AAMAS'18, July 2018, Stockholm, Sweden
5 CONCLUSION
We address the problem of adversarial detection of malicious con-
tent spreading through social networks. Traditional approaches use
with a homogeneous detector or a personalized filtering approach.
Both ignore (and thus fail to exploit knowledge of) the network
topology, and most filtering approaches in prior literature ignore
the presence of an adversary. We present a combination of modeling
and algorithmic advances to systematically address this problem.
On the modeling side, we extend diffusion modeling to allow for de-
pendence on the content propagating through the network, model
the attacker as choosing both the malicious content, and initial
target on the social network, and allow the defender to choose het-
erogeneous detectors over the network to block malicious content
while allowing benign diffusion. On the algorithmic side, we solve
the resulting Stackelberg game by first representing it as a bilevel
program, then collapsing this program into a single-level program
by exploiting the problem structure and applying KKT conditions,
and finally deriving a projected gradient descent algorithm using
explicit and implicit gradient information. Our experiments show
that our approach dramatically outperforms, homogeneous classifi-
cation, adversarial learning, and heterogeneous but non-adversarial
alternatives.
6 ACKNOWLEDGEMENTS
This work was supported, in part by the National Science Founda-
tion (CNS-1640624, IIS-1649972, and IIS-1526860), Office of Naval
Research (N00014-15-1-2621) and Army Research Office (W911NF-
16-1-0069).
[16] M. Gomez-Rodriguez and B. Schölkopf. Influence maximization in continuous
time diffusion networks. In Proceedings of the 29th International Coference on
International Conference on Machine Learning, pages 579–586. Omnipress, 2012.
[17] J. Holcomb, J. Gottfried, and A. Mitchell. News use across social media platforms.
Pew Research Journalism Project, 2013.
[18] D. Kempe, J. Kleinberg, and É. Tardos. Maximizing the spread of influence through
a social network. In Proceedings of the ninth ACM SIGKDD, pages 137–146. ACM,
2003.
[19] A. Laszka, Y. Vorobeychik, and X. Koutsoukos. Optimal personalized filtering
against spear-phishing attacks. In AAAI Conference on Artificial Intelligence, 2015.
[20] B. Li and Y. Vorobeychik. Feature cross-substitution in adversarial classification.
In Advances in neural information processing systems, pages 2087–2095, 2014.
[21] B. Li, Y. Vorobeychik, and X. Chen. A general retraining framework for scalable
adversarial classification. arXiv preprint arXiv:1604.02606, 2016.
[22] M. Lichman. UCI machine learning repository, 2013.
[23] D. Lowd and C. Meek. Adversarial learning. In Proceedings of the eleventh ACM
SIGKDD, pages 641–647. ACM, 2005.
[24] S. Mei and X. Zhu. The security of latent dirichlet allocation.
In Artificial
Intelligence and Statistics, pages 681–689, 2015.
[25] S. Mei and X. Zhu. Using machine teaching to identify optimal training-set
attacks on machine learners. In AAAI Conference on Artificial Intelligence, pages
2871–2877, 2015.
[26] J. Tsai, T. H. Nguyen, and M. Tambe. Security games for controlling contagion.
In AAAI Conference on Artificial Intelligence, 2012.
[27] J. Ugander, B. Karrer, L. Backstrom, and C. Marlow. The anatomy of the facebook
social graph. arXiv preprint arXiv:1111.4503, 2011.
[28] Y. Vorobeychik and J. Letchford. Securing interdependent assets. Autonomous
Agents and Multi-Agent Systems, 29(2):305–333, 2015.
[29] J. Wallinga and P. Teunis. Different epidemic curves for severe acute respiratory
syndrome reveal similar impacts of control measures. American Journal of
Epidemiology, 160(6):509–516, 2004.
[30] D. J. Watts and S. H. Strogatz. Collective dynamics of small-world networks.
Nature, 393(6684):440–442, 1998.
[31] V. A. Zorich and R. Cooke. Mathematical analysis i. 2004.
REFERENCES
[1] H. Allcott and M. Gentzkow. Social media and fake news in the 2016 election.
Technical report, National Bureau of Economic Research, 2017.
[2] A.-L. Barabási and R. Albert. Emergence of scaling in random networks. Science,
286(5439):509–512, 1999.
[3] A.-L. Barabâsi, H. Jeong, Z. Néda, E. Ravasz, A. Schubert, and T. Vicsek. Evolution
of the social network of scientific collaborations. Physica A: Statistical mechanics
and its applications, 311(3):590–614, 2002.
[4] M. Barreno, B. Nelson, R. Sears, A. D. Joseph, and J. D. Tygar. Can machine
learning be secure? In Proceedings of the 2006 ACM Symposium on Information,
computer and communications security, pages 16–25. ACM, 2006.
[5] S. Boyd and L. Vandenberghe. Convex optimization. Cambridge university press,
2004.
[6] M. Brückner and T. Scheffer. Stackelberg games for adversarial prediction prob-
lems. In Proceedings of the 17th ACM SIGKDD, pages 547–555. ACM, 2011.
[7] C. Budak, D. Agrawal, and A. El Abbadi. Limiting the spread of misinformation
in social networks. In Proceedings of the 20th international conference on World
wide web, pages 665–674. ACM, 2011.
[8] C. J. Burges. A tutorial on support vector machines for pattern recognition. Data
mining and knowledge discovery, 2(2):121–167, 1998.
[9] J. Cheng, C. Danescu-Niculescu-Mizil, and J. Leskovec. Antisocial behavior in
online discussion communities. In International Conference on Weblogs and Social
Media, pages 61–70, 2015.
[10] B. Colson, P. Marcotte, and G. Savard. An overview of bilevel optimization.
Annals of operations research, 153(1):235–256, 2007.
[11] G. V. Cormack et al. Email spam filtering: A systematic review. Foundations and
Trends® in Information Retrieval, 1(4):335–455, 2008.
[12] N. Dalvi, P. Domingos, S. Sanghai, D. Verma, et al. Adversarial classification. In
Proceedings of the tenth ACM SIGKDD, pages 99–108. ACM, 2004.
[13] N. Du, L. Song, M. G. Rodriguez, and H. Zha. Scalable influence estimation in
continuous-time diffusion networks. In Advances in neural information processing
systems, pages 3147–3155, 2013.
[14] N. Du, L. Song, H. Woo, and H. Zha. Uncover topic-sensitive information diffusion
networks. In Artificial Intelligence and Statistics, pages 229–237, 2013.
[15] N. Du, L. Song, M. Yuan, and A. J. Smola. Learning networks of heterogeneous
influence. In Advances in Neural Information Processing Systems, pages 2780–2788,
2012.
|
1001.4419 | 1 | 1001 | 2010-01-25T13:10:54 | A Framework to Manage the Complex Organisation of Collaborating: Its Application to Autonomous Systems | [
"cs.MA"
] | In this paper we present an analysis of the complexities of large group collaboration and its application to develop detailed requirements for collaboration schema for Autonomous Systems (AS). These requirements flow from our development of a framework for collaboration that provides a basis for designing, supporting and managing complex collaborative systems that can be applied and tested in various real world settings. We present the concepts of "collaborative flow" and "working as one" as descriptive expressions of what good collaborative teamwork can be in such scenarios. The paper considers the application of the framework within different scenarios and discuses the utility of the framework in modelling and supporting collaboration in complex organisational structures. | cs.MA | cs | A Framework to Manage the Complex Organization of
Collaborating: Its Application to Autonomous Systems
Peter Johnson
Rachid Hourizi
Neil Carrigan
Nick Forbes
Human Computer Systems Group, Department of Computer Science
School of Computer Science and Engineering, University of Bath, Bath, UK
[email protected]
In this paper we present an analysis of the complexities of large group collaboration and its applica-
tion to develop detailed requirements for collaboration schema for Autonomous Systems (AS). These
requirements flow from our development of a framework for collaboration that provides a basis for
designing, supporting and managing complex collaborative systems that can be applied and tested in
various real world settings. We present the concepts of "collaborative flow" and "working as one"
as descriptive expressions of what good collaborative teamwork can be in such scenarios. The paper
considers the application of the framework within different scenarios and discuses the utility of the
framework in modelling and supporting collaboration in complex organisational structures.
1
Introduction
This paper addresses important conceptual issues concerning collaboration in groups and sets out a de-
scription of the nature of the complexities of large group collaboration. It provides a basis for thinking
about the structural aspects of collaboration in Virtual Organisations from both a technical and social
perspective. It describes succinctly how the findings may be applied to autonomous systems, based upon
our past and current research on collaboration [8, 10] and autonomous systems [9]. Why should we be
interested in both group and collaborative working for complex human and autonomous systems (AS)?
Group working is rather obvious in that many activities will require more resources, capability and effort
that a single AS cannot provide. The second is less obvious, why should we be interested in collaborative
working? The nature of working in teams and managing groups is often harder than we realize, the costs
of working together can at times outstrip the benefits. The question is what does it mean to work to-
gether? Simply bringing a collection of people or agents or machines together does not achieve working
together. The difference between a group that works together and one that does not shows its effects in
many ways including the quality and efficiency of the work, the ease of working in the group, the ease of
managing the group and the level of confidence one can have in that unit. A group which works together
well achieves a high-level of flow to its work enabling a state of operation in which the individuals and the
group as a whole are fully immersed in what they are doing by a feeling of energized focus, full involve-
ment, and success in the process of the activity [9]. To have a group (of people or AS) work together
requires more than just enabling them to communicate and coordinate well, it requires collaboration.
Collaboration requires both good communication and good coordination; it also requires that they work
as one with shared goals, shared understanding, with a common ground[5, 14]. It does not necessarily
require a leader and can occur in different group structures [4]. Collaborative groups can achieve flow
and produce greater quality of product more efficiently [17], with easier group working and with easier
management. Moreover collaborative groups can achieve flow and produce greater levels of creativity
[3, 2]. The aim of this paper is to understand how large groups can function as collaborative groups and
how this needs to be adapted to be applied to particular settings, in this first instance AS. Achieving this
Jeremy W. Bryans and John S. Fitzgerald (Eds.):
Formal Aspects of Virtual Organisations 2009 (FAVO2009)
EPTCS 16, 2010, pp. 51 -- 63, doi:10.4204/EPTCS.16.5
c(cid:13) Peter Johnson, Rachid Hourizi, Neil Carrigan, Nick Forbes
This work is licensed under the
Creative Commons Attribution License.
52
TheOrganisationofCollaboration
aim has required us to develop a conceptual understanding of the nature of large group collaboration, of
the ways in which it is achieved and the benefits that accrue when that achievement takes place. In apply-
ing this to AS requires us to identify the particular characteristics of large group collaboration involving
autonomous systems, the problems and challenges that can arise whilst it is taking place and the ways
in which those problems and challenges can be overcome. With respect to its application to deploying
multiple AS the potential benefits are well documented [9, 13]. These benefits include the availability of
a greater range of resources than are possessed by any single AS (e.g. the ability to search an area that
includes both aerial and underwater threats), the enhancement of mission completion (e.g. the speed of
completion, parsimony of resource utilization and reliability of outcome) and the achievement of mis-
sion objectives that lie beyond the scope of any single system (e.g. the simultaneous screening of many
thousands of people inside a crowded street or public space). To take full advantage of these benefits we
must develop systems that can do more than co-exist. More specifically, we must design systems that
can coordinate their roles, objectives, data, resources and activities in such a way as to achieve smooth,
low cost work with minimal disruptions and conflicts. We describe this smooth, efficient multi-actor
activity as "collaboration" and draw upon previous work [11] to understand its optimal operation as one
involving multiple actors "working as one" and achieving collaborative "flow" [15]. This smooth effi-
cient collaboration is difficult to achieve, even when group sizes are small (i.e. in groups of five or less,
co-located actors, pursuing clearly defined, well-understood tasks) and the goals, actions, understandings
and impact of other actors are easy to identify. Achieving collaborative flow in situations that require the
involvement of larger sized and/or multiple-groups of actors is yet more difficult. Each of those actors
may have different capabilities, pursue multiple goals and be involved in many different activities. In
this context, the designers, managers and participants of large-group collaborations cannot rely upon the
existence of shared or common understanding, such as that which exists within smaller groups [4, 12].
In smaller groups, each actor is often able to follow the goals, activity, tasks, resources and capabilities
of each other actor (1:1 understanding) [14]. In large groups, by contrast this 1:1 understanding is less
prevalent (i.e. an actor may understand the goals, activities, tasks, resources and capabilities of some
but not all other actors).
In this context actors within large group collaborations may require strate-
gies/mechanisms that allow them to develop and use more abstract or group-level understanding of each
others goals, actions tasks etc [4], in addition to the more detailed understanding of these attributes that
they may have of some subset of actors in the group. (Note: even in a large group there may still be
some 1:1 understanding but it will not be developed between every actor in the group and every other
actor). Larger-scale collaborations are, as a result, both qualitatively and quantitatively different from the
small-group collaboration, in the sense that the possibility for a variations and individual differences in
the goals, actions and understandings that exist within the group may not be understood easily or in great
detail by the group or its members. Consequently, large-scale collaborations require both the group(s)
and its (their) members to manage the understandings of and contributions to the multiple goals and
activities. Moreover, in highly dynamic situations in which the goals, resources, group members etc are
likely to be changeable or emergent there is even greater complexity to the collaboration structures and
processes.
1.1 Collaboration: the application to AS
Our research (as part of the SEAS DTC- Systems Engineering for Autonomous Systems Defence Tech-
nology Centre) has identified the capabilities required of AS that enable participation in these large-scale
collaborations, considers the costs of deploying AS without such capabilities, tests the benefits of deploy-
ments involving collaboratively capable AS and demonstrates the effects of such deployment in authentic
PeterJohnson,RachidHourizi,NeilCarrigan,NickForbes
53
scenarios of use. The research addresses four questions crucial for effective AS collaboration:
1. What mechanisms or strategies for coping do groups of AS need to achieve/maintain collaboration?
2. When AS have to collaborate with other AS within these large groups, what are the coping strate-
gies/mechanisms that they require for this?
3. What coping mechanisms/strategies do AS have to initiate a request for collaboration within large
groups?
4. What efforts are required by People to work with AS that either have or do not have those capabil-
ities (i.e. what are the savings to the task and collaboration costs imposed upon those humans)?
1.2 Benefits of Collaborative AS
The understanding gained by addressing these four questions has enabled the identification of the func-
tionality required of next-generation autonomous systems, capable of managing their own contribution
to wider system goals and has the potential to deliver the one human to multiple platform vision. This
will provide for:
• increased "Flow" in the work undertaken by large groups of AS,
• improved coordination/reduced coordination problems within those groups, fewer interruptions,
• fewer/less severe communication problems,
• easier and more efficient management of large groups of AS,
• improved quality of performance and product.
The research contributes to an understanding of the additional "coping strategies" that AS might adopt
in response to the large numbers of actors, goals, activities, understandings and potential conflicts that
exist within large scale collaborations and the capabilities that those AS need in order to implement those
strategies. It also considers two factors in producing an initial requirement specification: the complexity
of the tasks and the complexity of the collaborations. The former includes the nature of the goal
relationships and the latter the nature of the collaborative relationships. Both factors are relevant to
understanding and managing collaboration in virtual organisation. In this paper we report the capabilities
and understandings that AS will require if they are to implement these coping strategies. These strategies
include mechanisms that address:
1. Task Structure issues: The large number of and variability in the goals, activities and state-
descriptions inherent in large group collaboration must be identified, negotiated, monitored and
judged. One possible solution is to have more abstract representations of each one that provide
less variability (though loosing all aspects of that variability may not be desirable).
2. Group Structure issues: The large and possible varied structure of the group(s) themselves. The
issues of group structure and how that affects group awareness, group communication and group
coordination. For example, one possible solution is appropriate division of the collaborative activ-
ity into subsections that can then be managed by addressing each subdivision only via a specified
middle-manager or gate-keeper (of course, the consequence of this can be lack of awareness in the
group and the gate keeper becomes the potential bottleneck).
54
TheOrganisationofCollaboration
2 Large Collaboration Capability Requirements
Previous research identified the capabilities needed to take part in small-group collaborations [5, 14, 13,
11] as well as the characteristics of large-groups [9, 4, 16] that mean those capabilities are insufficient
for the participation in and management of large collaborations. Previous work does not identify the
collaborative structures and processes needed to work as one or achieve a smooth effortless flow when
working in large groups. To address this, we have developed a conceptual framework to allow us to
develop the detailed structures and processes needed for large collaborations in highly dynamic situations
of high emergence. We present an overview of this in this paper for a more detailed expose see [9].
2.1 A Framework of Large-Group Collaboration
In considering the collaboration requirements, we focused upon the need for a collaborative group to
be able to manage emergent properties and dynamic changes to: 1.
the organization (wider group)
within which a collaboration takes place, 2. the internal (sub-) groups that undertake tasks within that
collaboration, 3. the tasks themselves, and 4. the resources required to undertake those tasks. As we
have already briefly mentioned collaboration and conflict are inseparable, in that potential and actual
conflicting situations arise within collaborations. Hence designing collaborations without conflict is
impossible, instead we recognise that the collaborative structure has to have conflict mechanisms. We
have identified three aspects of conflict: a) the avoidance of conflict before it occurs, b) the identification
of conflict that cannot be avoided, c) the resolution of the conflict identified in b, and, in the execution of
those three components the need for communication and coordination of the factors identified above. The
framework from which particular collaborative capability requirements were identified is summarized in
Figure 1, below:
2.1.1 The Need for Small Group Collaboration Mechanisms
The starting point for the large group collaboration framework is the vast amount of previous research
carried out on small group collaborations (such as [5, 14, 13, 11, 12]). Rather than offer this as an alter-
native to those it should be seen as building upon them. Thus the mechanisms required for large group
collaboration are additional to the requirements for small group collaboration moreover, small groups
can exist within large groups as well independently from them. Consequently, both sets of mechanisms
are needed and need to be satisfied to engage in teamwork that has "collaborative flow" and the ability
to "work as one". The capabilities we have identified that are required of small group collaboration are
reported elsewhere [13, 11, 12] and will not be repeated here.
2.1.2 The Need for Large Group Collaboration Mechanisms
As stated earlier, the large numbers of actors, goals, actions and resources that make up a large group
collaboration place additional requirements for collaboration mechanisms to enable the achievement of
collaborative flow in the context of the highly dynamic and emergent aspects of the required teamwork.
Figure 2. Below situates the problem space for large groups collaboration (adapted from [7]). Our
focus is in P4, where there are problems of broad extent coupling diverse complex subsystems. In P1,
where we have problems of limited scope and limited complexity, simple coordination mechanisms will
suffice. In P2, where we have problems of a limited extent but with high complexity, mechanisms for
self-coordinating groups will suffice. In P3, where we find problems of low complexity but high extent
and diversity, structured collaboration mechanism are needed and will suffice.
PeterJohnson,RachidHourizi,NeilCarrigan,NickForbes
55
Figure 1: Framework for large group collaborations.
Hence, those involved in large collaborations will need further capabilities and mechanisms in addi-
tion to those identified for small group collaboration. More specifically for example, they will require
capabilities that will allow them to dynamically re-distribute resources, dynamically share goals within
groups of actors etc where the extent and diversity of these are too large to allow 1:1 monitoring of
each others action, sub-goal and outcome. Hence in large group collaborations the importance of further
mechanisms and capability to address P4 in figure 2.
2.2 The Framework
The Framework presented in Figure 1. describes a number of structures and process relationships that
come into play and influence large-group collaborations. There are a number of structures (on the left-
hand side of the framework and coloured green) covering task, organisation, group and resources, where
within each there are many alternative structural relationships. For example, there are many different
types of task structures and, and within a given task structure there will be different types of relation-
ships between the elements of the task. For example, within a task structure there may be hierarchical
goal structuring and a network structure for the various procedures used. Similarly, there are many dif-
ferent types of organisation structures and within a given organisational structure there are different types
of relations. Hence each of the task, organisation, group and resource, structure cells of the framework
represent components that are themselves complex. Moreover, they are dynamically changed by both
internal and external factors and interact with each other. Hence the different task group, organisation
and resource structures interact with each other to deliver a collaborative mechanism. For further details
of the various task, group, organisation and resource structures see [9]. The central cells of the frame-
56
TheOrganisationofCollaboration
Figure 2: The problem space for large group collaborations.
work (coloured orange in figure 1) capture the different coordination and communication structures and
processes that may exist within a large-group collaboration. The coordination structures and processes
operate across and within each task, group organisation and resource structures. For example processes
and structures for coordinating resources relative to their use in tasks by groups between organisations
are detailed here. Similarly, communication structures and processes operate to ensure that appropri-
ate information, understanding and awareness is achieved both about and across the tasks, organisations
groups and resources. Moreover, particular communication structures and processes may take different
forms across the collaboration. For example the communication processes within one sub-group may
be strictly hierarchical, while in another it may be possible for anyone to communicate with anyone and
everyone directly. (As above, for further detail see [9]). The right hand-side of the framework (coloured
pink in figure 1.) addresses the processes and structures that are required in a collaboration to avoid,
identify and resolve conflict. This is explicit in the framework because of the importance of conflict
to collaboration. The three cells collectively capture the processes and structures managing and offset-
ting conflict as it arises and before it arises that enable resource conflicts, task conflicts, group conflicts,
and organisational conflicts to be overcome. Moreover, the relationships between the cells on the green
(left-hand side) and the pink cells (right-hand side) of figure 1. are "piped through the coordination and
communication processes and structures. Hence, the subject matter of the coordination and communi-
cation processes relate to the resolution, identification and avoidance of conflict relating to task, group,
organisational and resource properties. The need for these may arise as a result of external factors forc-
ing change or from internal factors requiring adaption and change in one part resulting from change in
another part of the collaborative system. For the purposes of this paper we describe the whole system
PeterJohnson,RachidHourizi,NeilCarrigan,NickForbes
57
driven from the perspective of the right-hand (Conflict) side of figure 1, however we could equally well
describe the system starting from the left-hand (task, group, organisation, resource) side.
2.3 Avoid Conflict
Conflict is an integral part of collaboration [1]; it is the management and reduction of conflict that leads
to successful collaborations. Hence to understand collaboration one must consider conflict avoidance,
identification and resolution. The first area described is the avoidance of conflict between the members
of a large group and their understandings of the ways in which the task at hand would be achieved.
Furthermore we also consider the roles, goals and actions to be adopted during the achievement of the
task and the allocation of available resources in the course of that adoption. This section describes that
first part of the framework.
2.3.1 Task Structures - Goals
There are many different models of task structuring (indeed we have ourselves contributed to these, see
for example, [6]) however, while they may use different terms and have different intended uses they
share a number of properties in common. Of concern here is the properties of collaboration structures
and process required to achieve tasks, rather than the modelling or analysis of tasks themselves. The
collaborating group must distribute the groups work amongst its various actors and sub-groups. This
distribution requires a set of capabilities of group members if it is to be achieved without external inter-
vention. Those capabilities include the identification of local goals that will, when correctly scheduled
and completed, lead to the achievement of the groups, high-level goals, the mapping of those sub-goals
to the achievement of high-level goals (i.e. an understanding of the relationship between the achieve-
ment of local goals within smaller sub-groups and the progress of the wider group towards its shared,
higher level goals), the allocation of those sub-goals to appropriate actors and/or sub-groups and, where
appropriate, the re-negotiation of that allocation with those actors and sub-groups.
2.3.2 Task Structure -- Actions
The actors and sub-groups must have and/or negotiate an understanding of the actions needed to achieve
their local sub-goals. In some situations, this negotiation will require only an agreement that sub-goals
will be achieved, in others that they will be achieved to a specified schedule (in line with the dependencies
that exist between sub-groups and their local objectives). In other situations, the trust held by managers
and leaders of the group will be so low that individual sub-groups maybe required to provide a detailed
description of the way that sub-goals will be achieved, rather than being left to make their own way
towards their own objectives, and hence gives rise to the need for further communication and coordination
in the collaboration.
2.3.3 Organisational Structure
The organisational structure refers to the pattern of relationships that exist both within a group(s) and
in the organisation(s) in which the group(s) exists. For example a hierarchical group may exist within
a coalition of services (as we find when military allies are created or when different care services work
together). In homogeneous groups (i.e. groups with a single structure and in which each participant has
an identical understanding of the group structure), certain conflicts would not arise i.e. the structure of
the group would, by definition, be known by all group members. With heterogeneous groups participants
58
TheOrganisationofCollaboration
would need to avoid conflict between competing understandings of the group structure to be adopted
(e.g. what roles were needed and their definition, the process of allocating roles, the responsibilities of
those roles etc). Participation in either "homogenous" and "heterogeneous" groups would require a num-
ber of key properties (i.e. collaboration requirements) to address external and internally driven dynamic
changes: Structuring (and restructuring) of the group for appropriate resources (and resource changes),
structuring (and restructuring) of the group to enable effective co-ordination and communications, struc-
turing (and restructuring) to enable task (and changes to task) goals, structuring (and restructuring) of
group to meet external organisational needs (and organisation changes), - in all cases to avoid/mitigate
conflicts arising in the collaboration.
2.3.4 Group Structure - Roles
Both hetero- and homogenous groups must agree the specific roles to be played by individual actors in
the course of a collaborative activity. In a flat structure (i.e. one, in which each actor is able to interact
with and influence the work of each other actor), these collaborative roles will differ very little from each
other (i.e. the information, requests and responses passed from one actor to another will follow similar
patterns, though the specific part of the collaborative task undertaken may vary from actor to actor).
In more structured groups, some actors will be asked to perform management roles (i.e.
to direct
the activity of other actors and / or to channel the information between the wider group and their sub-
ordinates), some will be asked to subordinate themselves to such managers and some to adopt both
management and subordinate roles (i.e. to become middle-managers). The specific roles adopted within
a particular collaboration will influence and be influenced by both the capabilities of the actors participat-
ing in a particular collaborative group and on the organizational structure of that group. Some groups of
actors may, for example collaborate without the need for a central manager, some groups may have one
or more managers imposed upon them by a higher authority and some may select managers by following
a pre-determined algorithm such as voting amongst themselves. If correctly managed, the structuring of
a large group through such an allocation of roles will allow large groups to cope with the impossibility of
monitoring every actor, action and objective, whilst ensuring that the group goals are achieved, available
resources are utilized appropriately, collaborative flow is maintained and the group can adapt to dynamic
changes such as the loss of an individual actor the loss of a resource or the alteration of a high level goal
by an external authority.
2.3.5 Group Structure - Actors
In either context, (i.e., homogeneous or heterogeneous organisational structures), the actors responsible
for assembling a large, collaborative group of their peers would need to be able to identify the resources
and capabilities needed for the completion of the task at hand (e.g. if the task involves the construction
of a wall the need for a certain number of bricks, a quantity of mortar and abilities to both lay bricks
and mix mortar). In more sophisticated cases, this capability may extend to the identification of different
combinations of resources and capabilities, any one of which could be used to satisfy the requirements
of the task (e.g.
in the case of the wall-building example, either the bricks, mortar and construction
capabilities, or an ability to transport an pre-fabricated wall from a storage location to the construction
site).
In such cases, the capability requirements demanded of any actor involved in group-assembly
would then include the identification of appropriate resource/capability combinations, the mapping of
specific combinations into collaborative task completion outcomes plus the identification of resources
and capabilities available by potential group members.
PeterJohnson,RachidHourizi,NeilCarrigan,NickForbes
59
2.3.6 Resources
The resources needed for the completion of tasks may be drawn from either local or global sources. In
some cases, specific actors or sub-groups will hold resources. For example, an information resource
may be available only to the sub-group, organisation, or location in which it is held.
In other cases
central resource stores may be appropriate information may be made available to all members of the
collaborative group, regardless of their small (sub) group membership. Similarly, a single fuel dump
may be managed by a single sub-group and allocated to other group members in case of need.
2.4 Identify Conflict
Despite the best attempts of the participants involved in a specific collaboration dynamic changes (e.g.
in group composition, group structure, overall or local goals, prevailing activity and available resources)
will inevitably present the opportunity for unanticipated conflicts to arise.
In order to resolve these
unanticipated conflicts, actors will need to both maintain awareness of those different factors (from group
composition to resource availability) and, as the next section makes clear, manage or resolve conflicts
arising within and between them.
2.4.1 Maintain Awareness.
In collaborating groups members must maintain awareness of the complex components of the prevailing
collaboration if they are to first identify and subsequently address the potential for conflict.
Organisational and Group Structure & Processes They must, for example, maintain awareness
of the organisational structure or structures that govern the prevailing collaboration. For example, in
homogenous groups of AS that awareness can be relatively easily maintained, since each of the AS
involved will, by definition have an identical understanding of group structure (i.e. they will all have an
identical understanding of the hierarchical, holarchical or other group structure under which the group
is operating). It should be noted that such common understanding cannot be assumed in heterogeneous
groups. Even within homogeneous groups actors involved in large group collaborations must maintain
some awareness of the composition of the group within which they are working. This does not mean that
they must maintain understanding of every individual actor and group within the wider collaboration in
which they are involved. They must be able to identify local managers, subordinates and contacts, must
understand how communication can be achieved with each one and must also be aware of changes in
role allocation that causes a new actors to be their manager, subordinate or contact:
Task Structure & Processes In any dynamic environment, a truly collaborative actor must also
maintain awareness of the groups goals and sub-goals. Once again, no comprehensive awareness of all
such goals either can or must be maintained by a single actor. This limitation notwithstanding each
actor in a large-group collaboration must maintain awareness of their own objectives and, depending on
their role, those of managers, subordinates and/or contacts in other sub-groups. This understanding is
important if actors are to coordinate their activity with others and, ultimately ensure that this activity
contributes to the groups shared goals. Each actor must also maintain awareness of their own actions
and their effects on the goals of the other group members with which they are in contact (the managers,
subordinates and contacts identified above).
In large group collaborations, the wider effects of each
action may not be understood by every member of the wider group but an effective group structure will
ensure that an appropriate understanding will be available to those in key roles and, as a consequence,
that collaborative flow is maintained as the group progresses towards its shared goals.
60
TheOrganisationofCollaboration
Resources. Finally, actors must maintain awareness of the resources that they need to complete
their activities. They must follow the extent to which they hold sufficient reserves of each resource (e.g.
knowledge, fuel or payload) themselves, the amount of each resource that can be obtained within the
sub-group to which they belong and, in case of need, the nature and amount of the resources that can be
obtained from a central store.
2.4.2 Manage Dynamic Changes
To exemplify the interactions and interconnectedness of the different components of the framework we
consider how when a change occurs to one part it influences and affects everything else in the framework.
Though maintaining awareness of a groups initial configuration and of the changes to that collaboration
are central to an actors participation in collaborative work, it should be noted that this awareness is
not, in itself, sufficient for collaborative flow to be maintained in many collaborations. In addition to
understanding a developing large-group collaboration, actor(s) must also be able to adapt to (manage)
dynamic changes within that collaboration. They must, for example, be able to manage changes to the
composition of the group i.e. to deal with situations in which loss of functionality, changing priorities
or instruction from a higher authority cause actor to either join or leave a group involved in collaborative
activity. In the homogenous, structured, large groups those joiners and leavers may have no immediate
impact upon a particular actor (because the joining or leaving actor affects only a remote subgroup),
may change the composition and therefore activity of the local sub-group or may lead to the replacement
of an actors immediate superior, subordinate or contact within other sub-groups. This in turn may lead
to dynamic changes to the task structure and processes in the light of a new role being adopted by
the actor. Those changes to group composition (or indeed other changes e.g. the loss of an important
resource, a change in the environment within which the group is acting or fresh instructions from a remote
authority) may in turn cause dynamic changes to local goals, high-level goals or both. In response to
changes in group goals or composition, the actions of an individual actor may also need to be dynamically
altered. The adoption of a new group structure, role or local (sub) goal will each cause a collaborative
actor to reconsider their activity, the schedule to which that activity must adhere or both together. Finally,
the resources needed by and available to an actor working within a large group are likely to change in
the course of collaborative activity of any complexity. A lost communication channel can lead to the
loss of information resources (i.e.
those resources which were supplied by other actors), a blocked
physical pathway may lead to lost fuel supplies and the loss of a superior or contact actors will prevent
communication beyond the immediate sub-group within which an actor is operating. Adaptation to such
losses may only be possible if actors posses the capabilities to: a) dynamically change goals, group
structure and activity in light of changes to resource needs and b) dynamically change goals, group
structure and activity in light of changes to resource availability.
2.5 Address & Resolve Conflict
When conflicts are identified in the course of a large-scale collaboration actors will also need the capa-
bility to address and resolve them. In perhaps the least disruptive case, the actor(s) identifying a conflict
may also be able to resolve it. This may require the revision or change of individual tasks (which in-turn
may impact upon the wider task of the group). It may also require an alternative or additional resource
usage (e.g. the allocation of more time to a transport task in exchange for a lower fuel usage). To the
extent that these changes are made they must be done so with communication and coordination to the
appropriate other members of the collaboration to ensure that awareness and the potential for further
PeterJohnson,RachidHourizi,NeilCarrigan,NickForbes
61
conflict is minimised. Hence we see a further instance of the complex interactions that occur with the
collaboration framework
2.6 Communication and Coordination Structures and Processes
In describing the framework here we have left largely implicit the detail of the communication and coor-
dination structures and processes needed. Briefly, the coordination structures and processes characterise
the dependencies and the means for ensuring the required states between those dependencies are main-
tained. These exist both in the individual cells of the framework and between the cells of the framework.
For example within the task structures there will be coordination processes needed to ensure that the com-
pletion of tasks are coordinated within the group. Furthermore, between the resource and task structures
there will be coordination process to ensure the availability of resources in timely manner. Similarly,
the communication processes and structures are the rules governing the routes and the form of commu-
nication within the collaboration. These may be universal (i.e. one set of rules applies to all) or may
be diverse (i.e. different rules apply to different parts) and static or dynamic (i.e. they may be allowed
to change over-time and/or events or not). These communication rules will again have implications and
influences upon the group, task, organisation and resource structures and upon how conflict is avoided,
identified and resolved in collaboration. (For a fuller description see [9]).
3 APPLICATION TO AUTONOMOUS SYSTEMS and Beyond
We have briefly described the collaboration framework from the perspective of the "Identify conflict",
"Avoid conflict" and Resolve Conflict sections of Figure 1 and considered some of the capabilities that
collaborative groups of people and/or autonomous systems will need. Future work will develop simu-
lations of autonomous systems that implement those capabilities and will then use those simulations as
the basis for testing the validity and practicality of those candidate requirements. The research will also
apply the collaboration framework to heterogeneous groups of AS to consider:
• different communication pathways between AS in different parts of the wider collaboration,
• different understandings of the division of labor between different parts of the group,
• different allocations of resource both to individual AS and the sub-groups to which they belong,
'item different coping strategies in the case that conflicts arise, such as those described above.
3.1 Further Issues
In applying the framework to AS we recognise there maybe collaboration problems arising that cannot
be resolved by the AS encountering the problem. This may occur because AS have not been designed to
play a full collaborative part in an activity. Alternatively, AS may incur damage, resources run out and
group members lose contact with the wider group. In both cases this may affect individual and group
ability to;
• avoid group structure, group composition, role, goal, activity and resource conflicts,
• identify group structure, group composition, role, goal, activity and resource conflicts and
• resolve group structure, group composition, role, goal, activity and resource conflicts.
Moreover, this may lead to a situation:
62
TheOrganisationofCollaboration
• where a partial or complete inability to avoid conflicts will result in an increased need to identify
and resolve conflicts, and
• a partial or complete inability to identify conflicts will obscure any remaining capability an AS
holds to resolve those same conflicts.
In such situations, we may use the framework to identify a fallback solutions or "work-around"
that can be used to improve large group AS functionality (where a full collaborative capability is not
available), and to identify the collaboration costs and limitations associated with each work-around.
3.2 Relevance to Virtual Organisations
The framework we have presented here is of a conceptual nature, largely of relevance to those interested
in understanding and/or developing collaborative organisations. It provides a basis for developing de-
tailed models of the interactions that go on in collaboration and for modelling the structures and processes
in a collaborative organisation. An important area is the development of mechanisms and technologies
to support such collaborative organisations that can (and must) be developed following this framework.
Research we are engaged in beyond our AS work includes work with health organisations and with lo-
cal authorities where we are helping them to engage in collaborative decision-making, and collaborative
service provision. In many cases these collaborations are creating new virtual organisations (VO) that
come together to deliver and develop services, and which involve people from many different individual
organisations and groups. To support these applications we need to develop mechanisms and technolo-
gies that address: 1. the development of a VO, 2. the functioning of a VO, and 3. the assessment of the
collaboration in a VO. We envisage the development of collaborative VOs using the framework to also
require tools and languages to allow the proposed structures to be expressed and reasoned about as an
aid to design. To support the functioning of the VO we envisage that environments and tools that support
easy and efficient sharing of information, formation of policy, decision-making and communication and
coordination will be needed. While to support assessment of collaboration we will need to develop met-
rics of such aspects as the amount of communication or ease of communication, the amount of consensus,
and sharing that exists within the organisation and the ability to avoid, identify and resolve conflict. All
of these require language and software technologies that can carry with them a change of culture that
allows organisations to work collaboratively to meet the demands of complex dynamic situations.
4 Conclusions
In conclusion, therefore, this paper extends our understanding of collaborative structures and processes.
It situates large group collaboration within the broader context of social emergence [16]. It has led to a set
of requirements for future generations of autonomous systems capable of participating in collaborative
activity. We also identify areas in which future research must extend the work presented here. More
specifically, those extensions must include:
• Investigation of AS reaction to "individually-unsolvable" role, goal, action and resource conflict,
• Testing of both the validity and utility of the requirements identified here,
• Application to human collaborations in service delivery.
Development of technologies to support dynamic collaboration in large-scale groups. These extensions
are the subject of our further research.
PeterJohnson,RachidHourizi,NeilCarrigan,NickForbes
63
Acknowledgements
Some of the work reported in this paper was funded by the Systems Engineering for Autonomous Sys-
tems (SEAS) Defence Technology Centre established by the UK Ministry of Defence.
References
[1] A. L. Cohen, D. Cash & M. J. Muller: Designing to support adversarial collaboration.
In: The
2000 ACM Conference on Computer Supported Cooperative Work (CSCW'00). pp. 31 -- 39. DOI=
http://doi.acm.org/10.1145/358916.358948.
[2] T. Coughlan & P. Johnson (2006): Interaction in creative tasks. In: R. Grinter, T. Rodden, P. Aoki, E. Cutrell,
R. Jeffries & G. Olson, editors: TheSIGCHIConferenceonHumanFactorsinComputingSystems(CHI).
Montr´eal, Qu´ebec, Canada, pp. 531 -- 540. DOI= http://doi.acm.org/10.1145/1124772.1124854.
[3] M. Csikszentmihaly (1991): Flow: The Psychology of Optimal Experience. Harper Collins, New York.
[4] E. Hutchins: The technology of team navigation. In: J. Galegher, R. E. Kraut & C. Egido, editors: Intellectual
Teamwork: Social andTechnologicalFoundationsofCooperativeWork. L. Erlbaum Associates, Hillsdale,
NJ, pp. 191 -- 220.
[5] H. Johnson & J. Hyde (2003): Towards modeling individual and collaborative construction of jigsaws
using task knowledge structures (TKS). Trans. Comput.-Hum. Interact 10(4), pp. 339 -- 287. DOI=
http://doi.acm.org/10.1145/966930.966934.
[6] H. Johnson & P. Johnson (1991): Task Knowledge Structures: Psychological Basis and Integration into
System Design. ActaPsychologica, pp. 3 -- 26.
[7] N. L. Johnson: Science of C.I. Resources for Change.
In: Mark Tovey, editor: Collective Intelligence:
Creatingaprosperousworldatpeace.
[8] P. Johnson (2004):
Interactions, collaborations and breakdowns.
on Task Models and Diagrams (TAMODIA,
http://doi.acm.org/10.1145/1045446.1045448.
In: The 3rd Annual Conference
86. ACM, Prague, Czech Republic, pp. 1 -- 3. DOI=
[9] P. Johnson & R. Hourizi: Collaborative Mechanism to Manage Large Groups of Autonomous Systems. Tech-
nical Report MP018-1, Systems Engineering for Autonomous Systems Defence Technology Centre.
[10] P. Johnson, J. May & H. Johnson (2003): Introduction to multiple and collaborative tasks. ACM Trans.
Comp.-Hum.Interact.10(4), pp. 277 -- 280. DOI= http://doi.acm.org/10.1145/966930.966931.
[11] C. Middup, F. Nemetz & P. Johnson: MP003-8. Technical Report, Systems Engineering for Autonomous
Systems Defence Technology Centre.
[12] C. Middup, F. Nemetz & P. Johnson: MP003-9. Technical Report, Systems Engineering for Autonomous
Systems Defence Technology Centre.
[13] C. Middup, F. Nemetz, A. Nossier & P. Johnson: MP003-7. Technical Report, Systems Engineering for
Autonomous Systems Defence Technology Centre.
[14] D. Pinelle & C. Gutwin (2001): Group Task Analysis for Groupware Usability Evaluations. In: 10thIEEE
internationalWorkshopsonEnablingTechnologies:infrastructureForCollaborativeEnterprises(WETICE).
IEEE Computer Society, pp. 102 -- 107.
[15] R. K. Sawyer (2003): Group Creativity: Music, Theatre, Collaboration. Lawrence Erlbaum.
[16] R. K. Sawyer (2005): Social emergence: Societies as complex systems. Cambridge University Press.
[17] R. K. Sawyer (2007): Group genius: The creative power of collaboration. BasicBooks, New York.
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.