On the choices of the parameters in general constrained learning algorithms

Abstract

This paper addresses the constrained learning algorithm (CLA) proposed by Perantonis et al, which is an efficient and fast back propagation (BP) algorithm formed by imposing the constraint condition, referred to as the a priori information, implicit in the issues into the conventional BP algorithm. It is found, through analyzing the CLA, that the choice of the values of the three learning parameters {δP, θp, η} in the algorithm is critical to successful application of the technique. Otherwise, the algorithm will not be able to converge within a limited time, or even diverge. This paper will discuss how to choose the three learning parameters based on an exhaustive understanding on the CLA. Finally, several computer simulation results show that our analyses and conclusions are completely correct. © Springer-Verlag 2003.