Adaptive Random Search Approach to Identification of Neural Network Model

Abstract

Identifying a neural network model is equivalent to multidimensional, nonlinear optimization. This paper presents a modified adaptive random search scheme for the optimization. The idea is to introduce a sophisticated probability density function (PDF) into a usual random search scheme for generating search vector. The new PDF provides two parameters that are used respectively to control local search range and search direction based on the past success-failure information so as to improve the searching efficiency. Computer simulations show that the new adaptive random search algorithm is a good alternative for the case where it is difficult to apply the well-known backpropagation algorithm.