A New Distributed Reinforcement Learning Approach for Multiagent Cooperation Using Team-mate Modeling and Joint Action Generalization

This paper focuses on the issue of distributed reinforcement learning (RL) for decision-making in cooperative multi-agent systems. Although this problem has been a topic of interest to many researchers, results obtained from these works aren’t sufficient and several difficulties have not yet been resolved, such as, the curse of dimensionality and the multi-agent coordination problem. These issues are aggravated exponentially as the number of agents or states increases, resulting in large memory requirement, slowness in learning speed, coordination failure and even no system convergence. As a solution, a new distributed RL algorithm, called the ThMLA-JAG method, is proposed here. Its main idea is to decompose the coor- dination of all agents into several two-agent coordination and to use a team-mate model for managing other agents’ experiences. Validation tests on a pursuit game show that the proposed method overcomes the aforementioned limitations and is a good alternative to RL methods when dealing with cooperative learning in mics environments while avoiding collisions with obstacles and other learners.