A Multi-agent Reinforcement Learning with Weighted Experience Sharing

Abstract

Reinforcement Learning, also sometimes called learning by rewards and punishments is the problem faced by an agent that must learn behavior through trial-and-error interactions with a dynamic environment [1]. With repeated trials however, it is expected that the agent learns to perfect its behavior overtime. In this paper we simulate the reinforcement learning process of a mobile agent on a grid space and examine the situation in which multiple reinforcement learning agents can be used to speed up the learning process by sharing their Q-values. We propose a sharing method which takes into consideration the weight of the experience acquired by each agent on the occasion of visiting a state and taking an action.