Learning-Rate Adjusting Q-Learning for Prisoner's Dilemma Games

摘要

Many multiagent Q-learning algorithms have been proposed to date, and most of them aim to converge to a Nash equilibrium, which is not desirable in games like the Prisoner's Dilemma (PD). In the previous paper, the author proposed the utility-based Q-learning for PD, which used utilities as rewards in order to maintain mutual cooperation once it had occurred. However, since the agent's action depends on the relation of Q-values the agent has, the mutual cooperation can be maintained by adjusting the learning rate of Q-learning. Thus, in this paper, we deal with the learning rate directly and introduce a new Q-learning method called the learning-rate adjusting Q-learning, or LRA-Q.