Dynamic Changes of Population Size in Training of Artificial Neural Networks

In this paper an adaptive differential evolution algorithm with dynamic changes of population size is presented. In proposed algorithm an adaptive selection of control parameters of the algorithm are introduced. Due to these parameters selection, the algorithm gives better results than differential evolution algorithm without this modification. Also, in presented algorithm dynamic changes of population size are introduced. This modification try to overcome limitations connected with premature convergence of the algorithm. Due to dynamic changes of population size, the algorithm can easier get out from local minimum. The proposed algorithm is used to train artificial neural networks. Results obtained are compared with those obtained using: adaptive differential evolution algorithm without dynamic changes of population size, method based on evolutionary algorithm, error back-propagation algorithm, and Levenberg-Marquardt algorithm.