自然勾配を用いたポートハミルトン系のための強化学習の高速化

抄録

In this paper, we accelerated a reinforcement learning algorithm for port-Hamiltonian systems using a natural gradient method. The proposed algorithm consists of an actor-critic structure wherein the actor generates a control input according to a policy and learns the policy using a temporal difference (TD) error, and the critic computes the TD error and learns a state-value function. Furthermore, the reinforcement learning algorithm for port-Hamiltonian systems has two types of the policy parameters which the proposed algorithm learns using the natural gradient method. Additionally, the proposed method was applied to the problem of swing-up control for an inverted pendulum through numerical simulation. The simulation result showed that the learning speed of the proposed method was higher than that of the existing method.