Systematic choice of initial points in local search: Extensions and application to Neural Networks

Abstract

Recently the authors presented (Inform. Process. Lett. 30 (1989) 67-72) a method for systematically initializing local search procedures which outperforms the commonly employed random initialization methods. In applying this technique to Neural Networks problems, the authors were led to two extensions of this result. The first extension shows that the resulting superior performance is retained when the underlying search algorithm is probabilistic, and the second extension allows application to continuous spaces. These extensions came about because of the recently demonstrated effectiveness (N. Baba, Neural Networks 2 (1989) 367-373) of random search methods in solving Neural Network problems which are formulated in continuous spaces. © 1991.