Multi-agent reinforcement learning for control systems: Challenges and proposals

Abstract

Multi-agent Reinforcement Learning (MARL) methods offer a promising alternative to traditional analytical approaches for the design of control systems. We review the most important MARL algorithms from a control perspective focusing on on-line and model-free methods. We review some of sophisticated developments in the state-of-theart of single-agent Reinforcement Learning which may be transferred to MARL, listing the most important remaining challenges. We also propose some ideas for future research aiming to overcome some of these challenges.