Echo State Network With Probabilistic Regularization for Time Series Prediction
Xiufang Chen, Mei Liu, Shuai Li
Abstract
Recent decades have witnessed a trend that the echo state network (ESN) is widely utilized in field of time series prediction due to its powerful computational abilities. However, most of the existing research on ESN is conducted under the assumption that data is free of noise or polluted by the Gaussian noise, which lacks robustness or even fails to solve real-world tasks. This work handles this issue by proposing a probabilistic regularized ESN (PRESN) with robustness guaranteed. Specifically, we design a novel objective function for minimizing both the mean and variance of modeling error, and then a scheme is derived for getting output weights of the PRESN. Furthermore, generalization performance, robustness, and unbiased estimation abilities of the PRESN are revealed by theoretical analyses. Finally, experiments on a benchmark dataset and two real-world datasets are conducted to verify the performance of the proposed PRESN. The source code is publicly available at https://github.com/LongJin-lab/probabilistic-regularized-echo-state-network.