Weights and structure determination of pruning-while-growing type for 3-input power-activation feed-forward neuronet

Abstract

In this paper, a new type of 3-input power-activation feed-forward neuronet (3IPFN) is constructed and investigated. For the 3IPFN, a novel weights-and-structure-determination (WASD) algorithm is presented to solve data approximation and prediction problems. With the weights-direct-determination (WDD) method exploited, the WASD algorithm can obtain the optimal weights of the 3IPFN between hidden layer and output layer directly. Moreover, the WASD algorithm determines the optimal structure (i.e., the optimal number of hidden-layer neurons) of the 3IPFN adaptively by growing and pruning hidden-layer neurons during the training process. Numerical results of illustrative examples highlight the efficacy of the 3IPFN equipped with the so-called WASD algorithm.