Learning and Optimizing Large Scale Multi-Agent System

        --A Mean Field Game Approach with Safe Reinforcement Learning

Abstract: Most complicated and coordinated tasks performed by the Large Scale Multi-Agent System (LS-MAS) require huge information exchange between the members of the team. This restricts the maximum population of LS-MAS due to the notorious "Curse of Dimensionality" and induces an interplay between the total number of agents and complexity of system optimization. The development of theory and algorithms that can break the "Curse of Dimensionality" while addressing the interplay, is, therefore, necessary to facilitate the optimal design of LS-MAS. In this talk, we will first introduce an emerging game theory, i.e. Mean-Field Game, to reformulate LS-MAS optimization problem without causing “Curse of Dimensionality” even while the number of agents is continuously increasing. Then, a novel reinforcement learning algorithm has been derived to obtain optimal control for LS-MAS by solving a coupled HJB-FPK equation from Mean-Field Game. Furthermore, to strengthen the practicality of derived learning technique, a safe reinforcement learning structure has been developed which cannot only learn the optimal control for LS-MAS but also strengthen the resiliency of LS-MAS in practical. Eventually, we will demonstrate the effectiveness of proposed work through numerical simulation results.