A Neighbor Generation Mechanism Optimizing Neural Networks

Abstract

This paper proposes the utilization of new neighbor generation method in conjunction with search techniques. The proposed mechanism works by adding a random number between –n and +n to the connection weights, where n is the weight value of each respective connection. This value may be multiplied by an adjustable ratio. The present paper shows the results of experiments with three optimization algorithms: simulated annealing, tabu search and hybrid system for the optimization of MLP network architectures and weights. In the context of solving the odor recognition problem in an artificial nose, the proposed mechanism has proven very efficient in finding minimal network architectures with a better generalization performance than the hybrid system mechanism used.