An efficient differential evolution algorithm with approximate fitness functions using neural networks

Abstract

We develop an efficient differential evolution (DE) with neural networks-based approximating technique for computationally expensive problems, called DE-ANN hereinafter. We employ multilayer feedforward ANN to approximate the original problems for reducing the numbers of costly problems in DE. We also implement a fast training algorithm whose data samples use the population of DE. In the evolution process of DE, we combine the individual-based and generation-based methods for approximate model control. We compared the proposed algorithm with the conventional DE on three benchmark test functions. The experimental results showed that DE-ANN had capacity to be employed to deal with the computationally demanding real-world problems. © 2010 Springer-Verlag.