A Decision-Mechanism for Reactive and Coordinated Agents

In this thesis we present preliminary results in the development of LIBRA (LInk oping Behavior Representation for Agents). LIBRA is a rule-based system for specifying the behavior of automated agents that combine reactivity to an uncertain and rapidly changing environment and coordination. A central requirement is that the behavior speci cation should be made by users who are not computer and AI specialists. Two application domains are considered: air-combat simulation and a simulated soccer domain from the perspective of the RoboCup competition. The behavior of an agent is speci ed in terms of prioritized production rules organized in a decision tree. Coordinated behaviors are encoded in the decision trees of the individual agents. The agents initiate tactics depending on the situation and recognize the tactics that the other team members have initiated in order to take their part in it. What links individual behavior descriptions together are explicit communication and common means to describe the situation the agents nd themselves in. Acknowledgments First of all, I would like to express my thanks to my supervisor, Anders Torne, for having given me the possibility of working in an interesting and fruitful project and for support and guidance. I would also like to thank my second supervisor, Dimiter Driankov, for the useful comments on this thesis and support. I am specially grateful to Erik Sandewall for having given me the opportunity to be part of a stimulating research environment as the one in IDA, and for his guidance and support during all this period. I would like to thank Patrick Lambrix for useful comments and discussions about this work and Ivan Rankin for proof-reading the thesis and helping to improve the English. My partner in work and life, Lars Karlsson, has given a great contribution to this work, both co-authoring several of the papers that form the basis of this thesis and keeping me in the best mood while I was writing them. This work would have not been possible without the collaboration of Saab. I would like to thank the personnel in Saab: Jennie Andersson, Johan Gorsjo, and Gunnar Frisk for support and discussion. In particular I would like to thank Jennie Andersson for her patience and support in solving the technical problems during the interface between our system and the Saab system and for the discussions and comments about this work. I am also indebted to Milind Tambe, University of Southern California, and Gary Eiserman, Decision-Science Application Inc, for the help in understanding two of the systems presented in this thesis: TacAir-Soar and Brawler. I am really grateful for their patience in answering all my questions. Finally, I would like to thank my colleagues in RTSLAB and RKLLAB and the rest of IDA for providing a stimulating research environment. Special thank goes to the administrative and technical sta ; particularly Anne Eskilsson, Lillemor Wallgren, Lise-Lotte Svensson, Leif Finmo and Gunilla F ak.