An Improved Small-World Topology for Optimizing the Performance of Echo State Network

By optimizing the topology of echo state network (ESN) to improve network performance, SW-ESN (an ESN with Small-Worldness) is proposed in this paper. The dynamic neuron pool has small-world property, and the learning performance of the small-world topology as a reservoir of ESN is studied. Then, establish the relationship function between the distance of the network nodes and its connection weight. The inversion operator is used to invert the weight with a certain probability to ensure the reasonable distribution of the positive and negative of the weight. Study the parameter adjustment of SW-ESN, and the impact of small-world topology on the echo state network via small-worldness, memory capacity, and nonlinear time series prediction. Through experiments, it is well proved the superiority of ESN with small-world topology in the aspect of fitting ability and robustness.