A Two-Phase Genetic Local Search Algorithm for Feedforward Neural Network Training

In this work, a two-phase genetic local search algorithm is proposed to train the connection weights of the feedforward neural networks. Various evolutionary algorithms including evolution strategies, evolutionary programming, and genetic algorithms had been proposed to train the weights and/or architectures of neural networks. But, most of them did not have an effective crossover operator. In the proposed algorithm, an effective orthogonal array crossover operator was used. Two classes of architectures were adopted and the classification capability of these two neural network architectures trained by the proposed two-phase genetic local search algorithm was shown by applying them to the n-bit parity problem.