Reinforcement learning based on extreme learning machine

Abstract

Extreme learning machine not only has the best generalization performance but also has simple structure and convenient calculation. In this paper, its merits are used for reinforcement learning. The use of extreme learning machine on Q function approximation can improve the speed of reinforcement learning. As the number of hidden layer nodes is equal to that of samples, the larger sample size will seriously affect the learning speed. To solve this problem, a rolling time-window mechanism is introduced to the algorithm, which can reduce the size of the sample space to a certain extent. Finally, our algorithm is compared with a reinforcement learning based on a traditional BP neural network using a boat problem. Simulation results show that the proposed algorithm is faster and more effective. © 2012 Springer-Verlag.