Adam Lerer and Alexander Peysakhovich
In social dilemmas individuals face a temptation to increase their payoffs in the short run at a cost to the long run total welfare. Much is known about how cooperation can be stabilized in the simplest of such settings: repeated Prisoner’s Dilemma games. However, there is relatively little work on generalizing these insights to more complex situations. We start to fill this gap by showing how to use modern reinforcement learning methods to generalize a highly successful Prisoner’s Dilemma strategy: tit-for-tat. We construct artificial agents that act in ways that are simple to understand, nice (begin by cooperating), provokable (try to avoid being exploited), and forgiving (following a bad turn try to return to mutual cooperation). We show both theoretically and experimentally that generalized tit-for-tat agents can maintain cooperation in more complex environments. In contrast, we show that employing purely reactive training techniques can lead to agents whose behavior results in socially inefficient outcomes.
The paper is here.