Experiments in Learning Prototypical Situations for Variants of the Pursuit Game

We present an approach to learning cooperative behavior of agents. Our approach is based on classifying situations with the help of the nearest-neighbor rule. In this context, learning amounts to evolving a set of good prototypical situations. With each prototypical situation an action is associated that should be executed in that situation. A set of prototypical situation/action pairs together with the nearest-neighbor rule represent the behavior of an agent. We demonstrate the utility of our approach in the light of variants of the well-known pursuit game. To this end, we present a classification of variants of the pursuit game, and we report on the results of our approach obtained for variants regarding several aspects of the classification. A first implementation of our approach that utilizes a genetic algorithm to conduct the search for a set of suitable prototypical situation/action pairs was able to handle many different variants.