An Adaptive Memetic Algorithm with Rank-Based Mutation for Artificial Neural Network Architecture Optimization

Abstract

Designing a well-generalized architecture for artificial neural networks (ANNs) is an important task. This paper presents an adaptive memetic algorithm with a rank-based mutation, denoted as AMARM, to design ANN architectures. The proposed algorithm introduces an adaptive multi-local search mechanism to simultaneously fine-tune the number of hidden neurons and connection weights. The adaptation of the multi-local search mechanism is achieved by identifying effective local searches based on their search characteristics. Such an algorithm is distinguishable from previous evolutionary algorithm-based methods that incorporate one single local search for evolving ANN architectures. Furthermore, a rank-based mutation strategy is devised for avoiding premature convergence during evolution. The performance of the proposed algorithm has been evaluated on a number of benchmark problems and compared with related work. The results show that the AMARM can be used to design compact ANN architectures with good generalization capability, outperforming related work.