Variable structure control based on-line learning design for continuous time multilayer networks

The purpose of this paper is to introduce variable structure-based-on-line learning algorithms for continuous time two layer and three layer perceptron networks with non-linear and linear activation functions. The computer implementation of the proposed algorithms result in a temporal learning capabilities of a neural network with dynamically adjusted weights, and zero convergence of the learning error in a finite time. The performance of the considered networks is tested in terms of solving a tracking problem of a sine signal.