Chaos-Based Reinforcement Learning When Introducing Refractoriness in Each Neuron

Abstract

Aiming for the emergence of “thinking”, we have proposed new reinforcement learning using a chaotic neural network. Then we have set up a hypothesis that the internal chaotic dynamics would grow up into “thinking” through learning. In our previous works, strong recurrent connection weights generate internal chaotic dynamics. On the other hand, chaotic dynamics are often generated by introducing refractoriness in each neuron. Refractoriness is the property that a firing neuron becomes insensitive for a while and observed in biological neurons. In this paper, in the chaos-based reinforcement learning, refractoriness is introduced in each neuron. It is shown that the network can learn a simple goal-reaching task through our new chaos-based reinforcement learning. It can learn with smaller recurrent connection weights than the case without refractoriness. By introducing refractoriness, the agent behavior becomes more exploratory and Lyapunov exponent becomes larger with the same recurrent weight range.