An incremental algorithm for learning radial basis function networks

This paper presents and evaluates an algorithm for incrementally constructing radial basis function networks, a class of neural networks which looks more suitable for adaptive control applications than the more popular backpropagation networks. The algorithm has been inspired by the CART algorithm developed by Breiman for generation regression trees. The algorithm proved to work well on a number of tests and exhibits performances comparable to the one step learning. An evaluation on the standard case study of the Mackey-Glass temporal series is reported.