On the Performance Analysis of Resource Allocating Network Trained with Significant Patterns

摘要

Radial Basis Function (RBF) neural networks have been receiving wider attention among the neural network research community after the Back Propagation (BP) trained Multilayer Perceptron (MLP) Feed Forward Neural Networks. The main reasons are their simple structure and ease of training. Several variants of RBF learning algorithms are available for solving classification, function approximation and regression problems. In any neural network system, proper identification of architecture and their parameter initialization assumes greater significance to accomplish better performance. There have been a large number of proposals on automatic development of architectures during learning to ease out the frustration caused by trial and error based schemes to fix up the so called optimal architectures. However their focus has been confined to one objective, ie to reduce the number of RBF units. Among the several variants of constructive learning networks, Platt's Resource Allocating Network (RAN) is still one of the popular incremental learning networks that has been tried on several problem types and still research is underway to fine tune it to maximize its performance. RAN has several limitations as it is sensitive to outliers and builds larger networks resulting in memorization consequently leading poor generalization. The best part of this algorithm is its fast convergence. Hence there is substantial scope to improve its performance. In this paper, we propose two new ideas to improve the performance of RAN. In the first instance, we calculate the medoids from samples in respective pattern classes. Secondly we pick a small percentage of significant patterns based on the farthest neighbor concept. The medoids and the significant patterns are collectively used to train the RAN. The entire samples in the input space are used to test the generalization performance of the network. Benchmark datasets have been used to experiment the proposed techniques. Results reveal significant enhancement in the performance of the RAN compared to traditional RAN.