An Incremental Self-Learning Algorithm with Robustness against Impulsive Noise

Abstract

In this paper, we propose a self-learning algorithm with the incremental learning capability, in which correntropy is selected as the cost function. Based on the correntropy concept of information theoretic learning (ITL), the incremental self-learning algorithm is presented for nonlinear system identification and the real world prediction problems against impulsive noises. Correntropy is a measure of information content that represents a similarity measure between two arbitrary random variables and has the advantages of rejecting outliers or impulsive noises. The proposed algorithm combines the neural network structure and fuzzy model, wherein the network hidden node (i.e. rule base) starts with empty and grows online based on the incremental learning method. Moreover, the maximum correntropy criterion (MCC) is applied to update the weight parameters for self-learning during the incremental process. The incremental process under two criteria is validated in detail through a nonlinear identification problem under impulsive environment. The performance evaluation of the proposed algorithm is also carried out on a nonlinear system identification and Time Series Data Library under both noise-free and impulsive noise conditions. The simulation results demonstrate that the proposed algorithm has better identification and prediction accuracy with the least number of rules and training time, and it also owns robustness for handling superior impulsive noise when compared to other algorithms.