Control of Chaotic Neural Networks for Optimisation Problems Using Reinforcement Learning

摘要

We attempt to control chaotic neural networks for optimization problems from a periodic state to non-periodic states that include solutions. In this report, a chaotic neural network to solve a five-city travelling salesman problem (TSP) is controlled with the reinforcement learning. In the reinforcement learning for the control, the reward for a control action is given by evaluating the energy decreasing and the firing rate of the network. Even with such a simple reinforcement learning, the control makes the network to show the states that include the solu-tions of the TSP, however it fails to obtain the optimal solution with the above simple criterion for the reinforcement learning.