Influence of the Chaotic Property on Reinforcement Learning Using a Chaotic Neural Network

摘要

Aiming for the emergence of higher complicated dynamic function such as "thinking", our group has set up a hypothesis that internal chaotic dynamics in an agent's chaotic neural network grows from "exploration" to "thinking" through reinforcement learning, and proposed a new learning method for that. However, even after learning in a simple obstacle avoidance task, the agent sometimes moved irregularly and collided with the obstacle. By reducing the scale of the recurrent connection weights, which is expected to have a deep relation to the chaotic property, the problem was reduced. Then in this paper, the learning performance depending on the recurrent weight scale is observed. The scale has an appropriate value as can be seen in FORCE learning in reservoir computing.