At IEEE International Conference on Systems, Man, and Cybernetics (SMC2018) held in Miyazaki, Japan, my student presented following topic.
“Bottom-up Multi-agent Reinforcement Learning for Selective Cooperation”
In this research, to make multi-agent reinforcement learning (MARL) truly distributed autonomous system, bottom-up MARL is raised as a new problem setting. In this problem, the necessary condition to acquire orderly group behavior is derived mathematically, and as a practical solution, we proposed a prediction of other agents’ rewards and their precisions for reward shaping. Furthermore, the precisions are reduced by an exploration bonus according to the amount of policy update. As a result, we succeeded in learning cooperative tasks on a dynamic robot simulator.