An Adaptive Memetic Algorithm for Designing Artificial Neural Networks

This paper presents an adaptive memetic algorithm (AMA) to evolve ANN architectures. In the AMA, multi-local searches are introduced and adaptively employed to simultaneously fine-tune the number of hidden neurons and connection weights of ANN architectures. The adaptation strategy is based on the characteristics of different local searches and their effectiveness during the evolutionary process. Such an algorithm is distinguishable from most previous evolutionary algorithms, which incorporate one single local search for evolving the ANN architectures. The performance of the AMA has been evaluated on three benchmark problems and compared with other related methods. The results show that the proposed AMA can obtain a satisfactory ANN architecture, outperforming related methods.