At 32th Symposium for Distributed Autonomous Systems in Tokyo, I presented a following content.
“Adaptive Eligibility Traces towards Online Deep Reinforcement Learning”
In this presentation, I proposed a new eligibility traces for deep reinforcement learning, which has an adaptive decaying factor according to output divergences between the past and latest parameters. This proposed method could outperformed the previous methods (i.e., the cases with/without the standard eligibility traces) at several robotic benchmark tasks in simulation.
In addition, the following paper was presented by corresponding student.
“Multi-agent Reinforcement Learning Considering Interests by Reward Prediction with Multivariate Distribution”