A constructive spiking neural network for reinforcement learning in autonomous control

Abstract

This paper presents a method that draws upon reinforcement learning to perform autonomous learning through the automatic construction of a spiking artificial neural network. Constructive neural networks have been applied previously to state and action-value function approximation but have encountered problems of excessive growth of the network, difficulty generalising across a range of problems and a lack of clarity in the operation of resultant networks. The results presented here demonstrate that rapid learning of the control of an inverted pendulum can be achieved with automatic construction of an efficient spiking neural network with internal reward-value associations. This provides a new approach to reinforcement learning and automatic neural network construction for autonomous learning.