A SUPERIOR TRAINING STRATEGY FOR THREE-LAYER FEEDFORWARD ARTIFICIAL NEURAL NETWORKS

Abstract

A new algorithm is proposed for the identification of three-layer feedforward artificial neural networks. The algorithm, entitled LLSSIM, partitions the weight space into two major groups: the input- hidden and hidden -output weights. The input- hidden weights are trained using a multi -start SIMPLEX algorithm and the hidden -output weights are identified using a conditional linear- least- square estimation approach. Architectural design is accomplished by progressive addition of nodes to the hidden layer. The LLSSIM approach provides globally superior weight estimates with fewer function evaluations than the conventional back propagation (BPA) and adaptive back propagation (ABPA) strategies. Monte -carlo testing on the XOR problem, two function approximation problems, and a rainfall- runoff modeling problem show LLSSIM to be more effective, efficient and stable than BPA and ABPA.