MULTI-AGENT REINFORCEMENT LEARNING: FROM ORGANIC COMPUTING TO GAME THEORY

Dr. K. Vimala

Abstract: Complex systems consisting of multiple agents that interact both with each other as well as their environment can often be found in both nature and technical applications. This paper gives an overview of important Multi-Agent Reinforcement Learning (MARL) concepts, challenges and current research directions. It shortly introduces traditional reinforcement learning and then shows how MARL problems can be modelled as stochastic games. Here, the type of problem and the system configuration can lead to different algorithms and training goals. Key challenges such as the curse of dimensionality, choosing the right learning goal and the coordination problem are outlined. Especially, aspects of MARL that have previously been considered from a critical point of view are discussed with regards to if and how the current research has addressed these criticism or shifted their focus. The wide range of possible MARL applications is hinted at by examples from recent research. Further, MARL is assessed from an Organic Computing point of view where it takes a central role in the context of self-learning and self-adapting systems.  
Keywords: Reinforcement learning, Multi-Agent, Organic Computing, Game Theory