Training neural networks using modified differential evolution algorithm for classification problems

In recent years, progress in the field of artificial neural networks provides a very important tool for complex problems in pattern recognition, data mining and medical diagnosis. The training algorithms of neural networks play an important role for adjustment the network parameters. Different algorithms have been presented for training neural networks; the most common one is the use of gradient descent based algorithms such as back propagation algorithm. Getting trapped in local minima and possessing a very slow converging speed made the gradient based methods problematic. To resolve this many evolutionary algorithms have been adopted for the training of neural networks. In this paper, a modified differential evolution algorithm acronymed as 2sDE is employed as a new training algorithm for feedforward neural networks in order to resolve the problems of local optimization training algorithms such as trapping in local minima and the slow convergence. Effectiveness and efficiency of the proposed method are compared with other training algorithms on various classification problems.