Learning of Correlation in Autonomous Decentralized Multi-Agent with Individual Objectives
Takumi Aotani, Taisuke Kobayashi, Kenji Sugimoto · The Proceedings of JSME annual Conference on Robotics and Mechatronics (Robomec) · 2018
Multi-agent reinforcement learning (MARL) is a framework to make multiple agents (e.g., robots) in the same environment learn their policies simultaneously using reinforcement learning. In the conventional MARL, although decentralization is essential for feasible learning, rewards for the agents have been given from a centralized system (named as top-down MARL). To achieve the completely distributed autonomous systems, we tackle a new paradigm named bottom-up MARL, where the agents get respective rewards. The bottom-up MARL requires to share the respective rewards for creating orderly group behaviors, and therefore, methods to do so were investigated through simulations. We found that the orderly group behaviors could be created by considering the relationship between the agents.