Learning and Cooperation in Sequential Games

The predictions of classical game theory for one-shot and finitely repeated play of many 2x2 simultaneous games do not correspond to human behavior observed in laboratory experiments. The promising results of learning models in tracking human behavior coupled with the growing electronic market and the number of e-commerce applications has resulted in an increased interest in studying the behavior of adaptive artificial agents in different economic games. We model agents with a reinforcement learning algorithm and analyze cooperative behavior in a sequential prisoner's dilemma game. Our results demonstrate the ability of artificial agents to learn cooperative behavior even in sequential games where defection is the subgame perfect Nash equilibrium. We attribute the reciprocal-like behavior to the structural flow of information, which reduces the risks of exploitation faced by the second-mover. Additionally, we analyze the impact of the second-mover's temptation payoff and pay-off risks on the rate of cooperative behavior.