Co-evolution of agent strategies in N-player dilemmas

Abstract

Understanding how cooperative behaviour emerges within a population of individuals has been the focus of a great deal of research in the multi-agent systems community. In this paper, we examine the effectiveness of two different learning mechanisms -- an evolutionary-based technique and a social imitation technique -- in promoting and maintaining cooperation in the spatial N-player Iterated Prisoner's Dilemma (NIPD) game. Comprehensive Monte Carlo simulation experiments show that both mechanisms are able to evolve high levels of cooperation in the NIPD despite the diminished impact of direct reciprocation. However, the performance of evolutionary learning is significantly better than social learning, especially for larger population sizes. Our conclusion implies that when designing autonomous agents situated in complex environments, the use of evolutionary-based adaptation mechanisms will help realising efficient collective actions.