Self-creating and adaptive learning of RBF networks: merging soft-competition clustering algorithm with network growth technique

摘要

Proposes a hybrid learning algorithm of RBF neural networks. The number of hidden neurons is decided by a network growth technique. A membership function is introduced into training center vectors of Gaussian functions. The reciprocal of the fuzzy factor, which increases during iteration, is considered as the temperature in simulated annealing. This algorithm can not only effectively overcome initial weight sensitivity problems and the dead-node problem of the c-means clustering algorithm, but also dynamically determines the hidden neurons. Experimental results show that the algorithm proposed in the paper is effective.