Presented at RSJ2017

At RSJ2017 in SAITAMA, I presented a following content.

“Actor-Critic Reinforcement Learning for Acquiring Global Optimum”

In this presentation, I solved the problem of the conventional actor-critic algorithm in reinforcement learning, namely it would fall into local optimum. Specifically, I employed student-t distribution as a stochastic policy, instead of normal distribution, because it has efficient exploration ability and conservative learning ability.