Adherence Improves Cooperation in Sequential Social Dilemmas

Abstract

Social dilemmas have guided research on mutual cooperation for decades, especially the two-person social dilemma. Most famously, Tit-for-Tat performs very well in tournaments of the Prisoner’s Dilemma. Nevertheless, they treat the options to cooperate or defect only as an atomic action, which cannot satisfy the complexity of the real world. In recent research, these options to cooperate or defect were temporally extended. Here, we propose a novel adherence-based multi-agent reinforcement learning algorithm for achieving cooperation and coordination by rewarding agents who adhere to other agents. The evaluation of adherence is based on counterfactual reasoning. During training, each agent observes the changes in the actions of other agents by replacing its current action, thereby calculating the degree of adherence of other agents to its behavior. Using adherence as an intrinsic reward enables agents to consider the collective, thus promoting cooperation. In addition, the adherence rewards of all agents are calculated in a decentralized way. We experiment in sequential social dilemma environments, and the results demonstrate the potential for the algorithm to enhance cooperation and coordination and significantly increase the scores of the deep RL agents.