At 24th Robotics Symposia held in Toyama, my student presented following topic.
“Cooperative-Competitive Learning by Online Classification of Interests between Multiple Agents”
The purpose of this study is generalization of bottom-up multi-agent reinforcement learning to make it possible to learn even if interests between agents are not explicitly given. We proposed the ways to find the interests as the correlations of rewards between agents in online manner, and to utilize it for reward shaping. Simulations showed that this generalization enables learning even in environments that contain agents with competition relationships, which cannot be handled in the previous work.