A resource limited immune approach for evolving architecture and weights of multilayer neural network

Abstract

A resource limited immune approach (RLIA) was developed to evolve architecture and initial connection weights of multilayer neural networks. Then, with Back-Propagation (BP) algorithm, the appropriate connection weights can be found. The RLIA-BP classifier, which is derived from hybrid algorithm mentioned above, is demonstrated on SPOT multi-spectral image data, vowel data and Iris data effectively. The simulation results demonstrate that RLIA-BP classifier possesses better performance comparing with Bayes maximum-likelihood classifier, k-nearest neighbor classifier (k-NN), BP neural network (BP-MLP) classifier and Resource limited artificial immune classifier (AIRS) in pattern classification. © 2010 Springer-Verlag.