Performance Enhancement of Reservoir Computing with Multi-Clustered Structure

Abstract

Reservoir computing (RC) is an attractive model of recurrent neural networks. In RC framework, a sparse interconnected hidden layer called reservoir, is the only training-required component that the simple linear regression method is capable to train it. This advantage makes the RC easier to implement and apply to engineering application than other RNNs. Recently, many studies have been focused on optimizing the performance of RC via learning algorithms or topological structures, which have been demonstrated critical to RC. In this paper, we propose a novel reservoir network with multi-clustered structure which is organized according to a developmental time windows algorithm. Experimental results showed that the model can significantly reduce the computational errors and standard deviation of the errors on the numeric prediction benchmark of Mackey-Glass chaotic time series. Moreover, we found that there exist the optimal values of the cluster size and time window size, where RC reaches the best performance in terms of computational error. And interestingly, in this case, the reservoir network exhibits the largest Shannon information entropy of neuronal activity patterns, indicating that information can be effectively transmitted in neural networks with complex dynamics. Our results showed that the multi-clustered generating method effectively promote the computational capability of reservoir computing.