What Makes a Good Co-Evolutionary Learning Environment?

There is growing evidence to suggest that the success of a co-evolutionary learning system may depend critically on the nature of the environment in which the learner is placed, and on certain attributes of the task domain, rather than the details of the particular learning algorithm employed. We discuss how a learning system can be modeled as a meta-level game between abstract entities which we call performer, innltrator and evaluator. Learning can sometimes fail due to collusive suboptimal equilibria in this meta-game of learning. But some domains have special attributes which seem to prevent such collusions and thereby facilitate co-evolutionary advancement. A better understanding of these issues may help to improve the design of co-evolutionary learning systems in the future.