Adapting Similarity Measures to Agent Types in Opponent Modelling.pdfVIP

Adapting Similarity Measures to Agent Types in Opponent Modelling Timo Steffens ¨ Institute of Cognitive Science Osnabruck ¨ 49078 Osnabruck, Germany tsteffen@uos.de Abstract This paper is part of ongoing research that aims at generating similarity-measures from incomplete domain- This paper proposes a case-based reasoning approach knowledge [7]. For the RoboCup application we propose to for opponent modelling in multi-agent systems that exploit s adapt similarity measures to classification goals. A general the advantages of lazy learning: The same observations are method is introduced which uses imperfect problem solving used for different classification goals such as action predic- knowledge for the classification goals. The role of the agent tion of various agent types. While the observations remain is used as the entry point into the knowledge network, and the same, it is the similarity measure that is adapted to the backward chaining reveals the relevant attributes for pre- type or role of the agents. Adaptation is done by integrat- dicting the agent’s action. ing incomplete problem solving knowledge into the similar- The next section gives an overview of the evaluation- ity measure automatical