Dr. Philippe Capdepuy

Department of Computer Science
University of Hertfordshire
Hatfield AL10 9AB
United Kingdom

Phone:	+44/1707/28-5118
Email:	p.capdepuy@herts.ac.uk

Research
Current Position
The Perception-Action Loop
PhD Work
Publications
Miscellaneous

Research

I obtained my PhD in computer science from the University of Hertfordshire in the Adaptive Systems Research Group in July 2010. My principal supervisor was Dr. Daniel Polani and my second supervisor was Pr. Chrystopher L. Nehaniv. I successfully defended my thesis entitled "Informational Principles of Perception-Action Loops and Collective Behaviours" on the 14th of July 2010. My examiners were Dr Hugo Touchette and Dr Neil Davey.
My main research subject is the perception-action loop and its information theoretic properties. I study how information theoretic quantities such as empowerment, the amount of perceivable control an agent has, can be used as a drive to shape the behaviour of an agent. I also study these aspects in the context of multi-agent systems, i.e. I try to identify the conditions in which multiple agents behaving as a group have more abilities than the sum of their own behaviour. In this context I also try to adress issues concerning self-organization in agent societies from an information theory perspective. The main assumption is that information quantities are universal and not semantically biased, but that they allow us to analyze the semantics that an agent can build through its interactions with a given environment.

Information theory and the perception-action loop
Multi-agent systems
Self-organization and complex systems
Origins of life and cognition
Artificial chemistries

Current Position

I am now working in the industry, doing applied research in robotics at Generation Robots (under the brand HumaRobotics). I'm in charge of developing autonomous behaviours for the Nao robot from Aldebaran Robotics. We have successfully managed to make him play Connect Four against a human opponent (follow the link for extra videos and information). Here are some technical points of interest:

Everything runs directly on the robot, no extra computers or sensors are used.
We use custom vision algorithms that allow fast detection and tracking of objects in control loops (around 100ms).
The controller for putting chips in the board is learned from sensorimotor data.
Learning is fast enough to be performed on-the-fly during the task.
We are working on a fuzzy-logic based behaviour description language that incorporates emotional dynamics.

The Perception-Action Loop

Using the formalism of causal Bayesian graphs, it is possible to model the perception-action loop of an embodied agent. In this model, the (S)ensors, (A)ctuators and the (R)est of the environment are represented as random variables at different points in time. The relationships between these variables are modelled as conditional probability distributions. A typical perception-action loop for a memoryless reactive agent is then represented as such:

The red arrow on the figure represents what is called the actuation channel or the sensorimotor channel of the agent. The causal Bayesian graph formalism allows us to express various information-theoretic quantities. One of them, referred to as empowerment, is defined as the capacity of the sensorimotor channel. More precisely, 1-step empowerment is defined as:

Empowerment can be understood as the amount of perceivable control an agent as onto its future sensors. Because this control is mediated through the environment, it is bounded from above by the amount of control the agent has on the state of the environment. However the latter quantity is not directly accessible to the agent.

PhD Work

My PhD work focuses on collective behaviour in multi-agent systems. More precisely we use information theoretic quantities such as empowerment as a drive toward collective behaviour. We carried out different simulation experiments whose results reinforce our hypothesis: collective behaviour can be the result of maximizing potential information transfer of the agents' perception-action loops. Here is a video of the kind of behaviours obtained when using these principles:

Publications

Journal Papers:

(2010) Capdepuy P., Nehaniv C.L. and Polani D.: Perception-Action Loops of Multiple Agents: Informational Aspects and the Impact of Coordination, (submitted).

Conference papers:

(2008) Capdepuy P., Nehaniv C.L. and Polani D.: Adaptation of the Perception-Action Loop Using Active Channel Sampling, Proceedings of the 2008 NASA/ESA Conference on Adaptive Hardware and Systems. (Preprint,Bibtex)

(2007) Polani D., Capdepuy P., and Nehaniv C.L.: Maximal, Enforced, and Potential Entropy Production: towards an Understanding of Mechanisms for the Generation of Complexity, Proceedings of the 2007 International Conference on Complex Systems.

(2007) Capdepuy P., Nehaniv C.L. and Polani D.: Maximization of Potential Information Flow as a Universal Utility for Collective Behaviour, Proceedings of the First IEEE Symposium on Artificial Life, p. 207-213. (Preprint,Bibtex)

(2007) Capdepuy P., Nehaniv C.L. and Polani D.: Constructing the Basic Umwelt of Artificial Agents: An Information-Theoretic Approach, Proceedings of the Ninth European Conference on Artificial Life, LNCS/LNAI 4648, p. 375-383, Springer. (Preprint,Poster,Bibtex)

(2007) Capdepuy P., Nehaniv C.L. and Polani D.: Grounding Action-Selection in Event-Based Anticipation, Proceedings of the Ninth European Conference on Artificial Life, LNCS/LNAI 4648, p. 253-262, Springer. (Preprint,Poster,Bibtex)

(2007) Capdepuy P., Nehaniv C.L. and Polani D.: Construction of an Internal Predictive Model by Event Anticipation, Anticipatory Behavior in Adaptive Learning Systems - From Brains to Individual and Social Behavior. (Preprint,Bibtex)

(2006) Capdepuy P., Nehaniv C.L. and Polani D.: An Information Theory Based Anticipation Architecture, Ninth International Conference on the Simulation of Adaptive Behavior 2006, Third Workshop on Anticipatory Behavior in Adaptive Learning Systems. (Preprint,Bibtex)

(2005) Capdepuy P. and Kaltenbach M.: Autonomous Systems: Computational Autopoiesis through Artificial Chemistries, European Conference on Artificial Life 2005, Workshop on Artificial Chemistries. (Preprint,Bibtex)

Miscellaneous

Some people

Tom Anthony A PhD student involved in related topics
Arnaud Blanchard My former deskmate, involved in affective computing, robotics...
Jim Crutchfield A researcher involved in computational mechanics, evolution of complexity...
Nicolas Despres A good old friend doing research in programming stuff (I just can't tell more, have a look)
Francis Heylighen A researcher involved in evolution of complexity and evolutionary cybernetic approaches.
Tim Hutton A researcher involved in artificial chemistries and protocells modelling.
Alexander Klyubin Recently got his PhD. Much of my research work is a continuation or is related to his.
Barry McMullin A researcher much involved in artificial life and artificial chemistries.
Chrystopher L. Nehaniv My second supervisor, involved in constructive biology, theoretical computer science and other topics.
John von Neumann A brilliant mathematician that, among other things, pionnered computer science and artificial self-replication.
Daniel Polani My main PhD supervisor, involved in information theory applied to the perception-action loop, sensors evolution and other topics.
Ilya Prigogine A great researcher (Nobel prize) who was involved in evolution of complexity, non-equilibrium thermodynamics, self-organization and many other topics...

Some concepts

Artificial Chemistries A good review of the different kind of artificial chemistries.
Autopoiesis A good introduction to Maturana and Varela's theory of autopoiesis.
Cybernetics Website of Principia Cybernetica Project, a good introduction to cybernetics in various contexts, with many links to related subjects.
Dissipative structures A good description of what is this phenomenon.
Information Theory The wikipedia description of information theory. Good introduction with links to related subjects.

Some good french websites

Automates Intelligents News on robotics, artificial intelligence, sciences... Very good quality.
Cognitive Sciences Wiki A french wiki dedicated to cognitive science and related stuff initiated by Arnaud Blanchard. Feel free to contribute.
Futura Sciences Articles, news and forums about science in general...
RISC An institution affiliated with the CNRS (french national research institute) that relays and centralizes information on cognitive sciences.
Vie Artificielle Articles, news and forums about artificial life, robotics, artificial intelligence...