Design of MLP using Evolutionary Strategy with Variable Length Chromosomes

This paper presents a novel approach in designing MLP neural networks by using evolutionary strategy with variable length chromosomes. In particular, unlike other similar approaches in which the maximum number of neurons must be determined beforehand, the proposed method can grow a network as large as possible with less computational cost. By redefining genetic operators such as mutation and crossover, the evolutionary approach can evolve chromosomes with different lengths; therefore, various networks with different number of neurons in hidden layer can be achieved. The empirical result shows that the evolutionary strategy proposed in this paper can be compared favorably to other alternative approaches for classification problems.