Opponent Modelling and Strategy Evolution in the Iterated Prisoner’s Dilemma

Learning and evolution are two adaptive processes in the natural world that have been modelled in the study of artificial intelligence in computer science. In both biology and in artificial intelligence, learning and evolution are complementary processes. The nature of the interactions between learning and evolution has been the subject of much research in scientific disciplines. Evolution of a species can promote learning in individuals and learning in individuals can guide evolution of the species. In artificial intelligence, both learning and evolution can be applied to achieve similar goals. But which is more effective for a particular problem? What if, rather than being complementary, learning and evolution were in competition? Which is more important, being smart or having good breeding? In this paper the game of the Iterated Prisoner’s Dilemma is used as a context for investigating the answers to these questions. Evolutionary techniques have been widely applied to find good strategies for the game. The evolutionary approach to finding strategies is continued here and is measured against a learning approach. Can an adaptive learning agent out-perform evolved strategies? How does evolution respond to the presence of an adaptive agent? The experiments show that opponent modelling can be very effective against evolved populations of deterministic strategies. When stochastic strategies are considered, evolution produces strategies that are able to exploit the learning agent. However, the learning agent is able to recognise that being exploited is better than retaliation in this zero-sum game and is able to achieve near-optimal performance against these non-deterministic players.