Learning and generalization of noisy mappings using a modified PROBART neural network

Abstract

Incremental function approximation using the PROBART neural network offers many advantages over conventional feedforward networks. These include dynamic node allocation based on the complexity of the function approximation task, guaranteed convergence, and the ability to handle noise in the training data. However, the PROBART network does not generalize very well to untrained data. In this paper, a modified PROBART is proposed to overcome this deficiency. This modification replaces the winner-take-all mode of prediction of the PROBART with a distributed mode of prediction. This distributed mode enables several neurons to cooperate during prediction and, thus, provides better generalization capabilities even in noisy conditions. Computer simulations are conducted to evaluate the performance of the modified PROBART neural network using three benchmark nonlinear function approximation tasks. The prediction accuracy of the modified PROBART network compares favorably to the PROBART, fuzzy ARTMAP, and ART-EMAP networks for all these tasks.