Multiple clusters echo state network for chaotic time series prediction

Abstract

The chaotic time series prediction problem is considered. A novel type of cortex-like neural network model, i.e. multi-clusters echo state network model （MCESN）, regulated by a group of five growth-factors, is proposed. It is shown that characters of MCESN’ topology can be effectively determined by the growth-factors group; and that it is the MCESN possessing both small-world and scale-free properties of complex network that corresponds to the better prediction performance. In addition, Monte Carlo simulation experiments show that MCESN not only can be trained by easy algorithm, but also can achieve higher accuracy and less standard deviation prediction results than classical echo state networks.