Learning desirable actions in two-player two-action games

Reinforcement learning is widely used to let an autonomous agent learn actions in an environment, and recently, it is used in a multi-agent context in which several agents share an environment. Most of multi-agent reinforcement learning algorithms aim to converge to a Nash equilibrium of game theory, but it does not necessarily mean a desirable result On the other hand, there are several methods aiming to depart from unfavorable Nash equilibria, but they use other agents' information for learning and the condition whether or not they work has not yet been analyzed and discussed in detail. In this paper, we first see the sufficient conditions of symmetric two-player two-action games that show whether or not reinforcement learning agents learn to bring the desirable result After that, we construct a new method that does not need any other agents' information for learning.