Improving Evolutionary Learning of Cooperative Behavior by Including Accountability of Strategy Components

摘要

We present an improvement to evolutionary learning of cooperative behavior which incorporates some accountability measure for strategy components into the evolutionary learning process. Our evolutionary approach is based on evolving sets of prototypical situation-action pairs (strategies) that together, with the nearest-neighbor rule, represent the decision making of our agents. The basic idea of our improvement is to collect data for each pair showing the results of its applications. We then choose those pairs in the parent strategies that had positive results for the construction of new sets of pairs for our strategies. Our experiments within the OLEMAS system show that the incorporation of accountability results in substantial improvements of both on-and off-line learning when compared to the basic evolutionary approach. In nearly all experiments, either the agent teams required less learning time or found better strategies. In many cases both were observed.