Techniques for analysis of evolved prisoner's dilemma strategies with fingerprints

It is easy to generate strategies for games such as the iterated prisoner's dilemma using evolutionary computation, but much harder to analyze those strategies. Fingerprints are a functional signatures of game playing agents that capture essential features of an agent's strategy while ignoring implementation details. Using functional fingerprints, it is practical to cluster agents and to rapidly identify common agent types in spite of the representational obfuscation often generated by evolutionary training techniques. In this paper, a set of 1080 agents from 30 evolved populations are subjected to analysis using fingerprint based techniques. Filtration is used to remove first well known and then later common strategies. A novel clustering technique, multi-clustering, is then used to cluster the remaining strategies. Filtration and multiclustering, used together, smooth the analysis of evolved agents. Agents playing known strategies are quickly identified and removed from the agent pool, unknown types are clustered into plausible groupings. A previously unsuspected tendency of evolution to prefer finite state strategies composed of a single communicating class is documented.