Incorporating expert advice into reinforcement learning using constructive neural networks

This paper presents and investigates a novel approach to using expert advice to speed up the learning performance of an agent operating within a rein- forcement learning framework. This is accomplished through the use of a constructive neural network based on radial basis functions. It is demonstrated that incorporating advice from a human teacher can substantially improve the perform- ance of a reinforcement learning agent, and that the constructive algorithm pro- posed is particularly effective at aiding the early performance of the agent, whilst reducing the amount of feedback required from the teacher. The use of construc- tive networks within a reinforcement learning context is a relatively new area of research in itself, and so this paper also provides a review of the previous work in this area, as a guide for future researchers.