Cooperation without exploitation between self-interested agents

Abstract

We study how two self-interested agents that play a sequence of randomly generated normal form games, each game played once, can achieve cooperation without being exploited. The agent learns if the opponent is willing to cooperate by tracking the attitude of its opponent, which tells how much the opponent values its own payoff relative to the agent's payoff. We present experimental results obtained against different types of non-stationary opponents. The results show that a small number of games is sufficient to achieve cooperation. © 2013 Springer-Verlag.