Learning to cooperate in a social dilemma: a satisficing approach to bargaining

摘要: Learning in many multi-agent settings is inherently repeated play. This calls into question the naive application of single play Nash equilibria learning and suggests, instead, give-and-take principles bargaining. We modify analyze a satisficing algorithm based on (Karandikar et al., 1998) that compatible with bargaining perspective. form relaxation search converges to equilibrium without knowledge game payoffs or other agents' actions. then develop an M action, N player social dilemma encodes key elements Prisoner's Dilemma. instructive because it characterizes dilemmas more than two agents choices. show how several different algorithms behave this dilemma, demonstrate converges, high probability, Pareto efficient solution self against selfish agents. Finally, we present theoretical results characterize behavior algorithm.