T2 - HALF DAY ==================================================================== Learning of Behaviour in Agents and Multi-Agent Systems: Dimitar Kazakov, Daniel Kudenko =================================================================================== ABSTRACT OF TUTORIAL Learning Agents is an emerging multi-disciplinary area encompassing Computer Science, Software Engineering, Biology, as well as Cognitive and Social Sciences. Learning is a highly important component of intelligence. We believe that a tutorial on (machine) learning of agent behaviour will be very useful for all participants interested in the development of intelligent agent technologies. The primary objectives of the proposed tutorial are to (1) present a novel synthesis combining distinct lines of AI work, (2) motivate and explain an AI topic of emerging importance and (3) introduce expert non-specialists to an AI sub-area. The sub-area of Learning Agents is situated at the intersection between Machine Learning and Agents. The tutorial will cover the necessary minimum of background in both areas to concentrate on the methodology of this emerging AI domain, and demonstrate its ideas on well-chosen focus topics. The latter two items will represent more than half of the material covered. The chosen balance of material should make the tutorial of sufficient interest and novelty to experts in either ML or Agents, while keeping it self- contained and open to other attendants with general AI background. The tutorial will also help introduce novices to major topics of Artificial Intelligence, since it will assume no prior knowledge of and will introduce the participants to the basic concepts of Machine Learning and Agents, while keeping as its primary goal the discussion of the area at their intersection. Further, a number of seminal contributions will be discussed when the two larger areas are introduced, hence the tutorial will serve to a certain extent as a survey of a mature area of AI research and/or practice. After taking the tutorial, we expect students to: - have a basic knowledge of fundamental machine learning concepts; - understand the problems involved in applying ML techniques to situated agents; - appreciate the added complexity and issues involved when moving from a single learning agent to a multi-agent system; - be able to apply the new knowledge in his/her application domain of choice. ====================================================================== INTENDED AUDIENCE Graduate students of AI and up. ========================================================================== BACKGROUND KNOWLEDGE REQUIRED We will not assume any prior knowledge of Machine Learning or Agents. ====================================================================================== DETAILED OUTLINE The tutorial will be split into two parts, with a short break between them. Two important distinctions will be made throughout the tutorial. Firstly, one will distinguish between the evolution of behaviour through natural selection in the course of several generations of agents, and individual learning, which improves an agent's performance during its lifespan. It will be shown how and under which circumstances does the evolution of inherited behaviour select in favour of agents that have an innate ability to learn. Secondly, single agent learning using no interaction between agents will be compared to multi-agent learning, where such interaction is employed with benefit. PART I (duration: 2 hours): Evolution of Behaviour and Single Agent Learning Introduction ·Introduction to the general research area of machine learning which has historically focused on disembodied agents. ·Discussion of the need for learning of behaviour in agents. ·Evolution vs. individual learning of behaviour. Evolution of Behaviour ·Introduction to the basic concepts of evolution through natural selection. ·The Baldwin effect or the pros and cons of the ability to learn. ·Neo-Darwinism, selfish genes, inclusive fitness and the evolution of behaviour. ·Focus topic: evolution of co-ordination through kinship-driven altruism. Machine Learning (ML) ·Fundamental ML concepts (such as learning biases and version space). ·Taxonomy and examples of learning approaches. ·The move from disembodied to situated ML. Integrating ML into the Agent Architecture:Single-Agent Learning ·The relevance of above ML concepts to the design of learning agents. ·Agent-related issues in ML (such as timeliness of learning and time complexity of the theories learned). PART II (duration: 2 hours): Learning in Multi-Agent Systems Principles of Multi-agent Learning ·Moving from single-agent to multi-agent systems: multiple single-agent learning vs. social multi-agent learning ·Co-evolution Specialisation among Agents ·Learning to distribute pre-defined team roles ·Learning of team roles from scratch Learning and Communication ·Sharing sensory inputs ·Sharing of experience ·Merging learning results Distributed Learning and Data Mining Focus Topic: Learning of Coordination ·Principles of coordination ·Learning of coordination in single-stage games ·Learning of coordination for Markov Decision Processes ·Assessing the risk of mis-coordination ·Evolution of coordination Future Directions and Challenges ========================================================================== BIOGRAPHIES OF PRESENTERS Dimitir Kazakov Department of Computer Science, University of York, Heslington, York, YO10 5DD, UK Email: lastname@cs.york.ac.uk, WWW: http://www-users.cs.york.ac.uk/~lastname/ Dr. Dimitar Kazakov is a lecturer in Machine Learning at the Department of Computer Science at the University of York. His main research interests are at the intersection of Machine Learning, Natural Language, and Agents. Dr Kazakov has several publications in the areas of Machine Learning, Natural Language Processing and Learning Agents, in issues such as the journals of Machine Learning and AI Communications. His current interests are in the field of Evolution of Collaboration, Communication and Language in Multi-Agent Systems. He was a co-chair of the second AAMAS symposium (London, 2002) and is the chair of the forthcoming AAMAS-3 symposium (Aberystwyth, 2003). Dr Kazakov is an associate co-ordinator of the ALAD SIG of the European Network of excellence AgentLink. Daniel Dudenko Department of Computer Science, University of York, Heslington, York, YO10 5DD, UK Email: lastname@cs.york.ac.uk, WWW: http://www-users.cs.york.ac.uk/~lastname/ Dr. Daniel Kudenko is a lecturer at the Department of Computer Science at the University of York. His research interests are in the areas of learning in multi-agent systems, specifically reinforcement learning of coordination and distributed learning, adaptive information agents, and negotiation simulation. Dr. Kudenko has several publications in major AI conferences and journals in the area of Machine Learning and Agents. He has chaired the AISB'01 Symposium on Adaptive Agents and Multi-Agent Systems (AAMAS), co-chaired AAMAS-2 (London, 2002) and will co-chair the upcoming AAMAS-3 (Aberystwyth, 2003). He also served on the programme committees of ICML 2000 and CIA 2002. Dr Kudenko is the coordinator of the ALAD SIG of the European Network of excellence AgentLink.