A memetic imperialist competitive algorithm with chaotic maps for multi-layer neural network training

The performance of artificial neural networks (ANNs) is largely dependent on the success of the training process. Gradient descent-based methods are the most widely used training algorithms but have drawbacks such as ending up in local minima. One approach to overcome this is to use population-based algorithms such as the imperialist competitive algorithm (ICA) which is inspired by the imperialist competition between countries. In this paper, we present a new memetic approach for neural network training to improve the efficacy of ANNs. Our proposed approach – Memetic Imperialist Competitive Algorithm with Chaotic Maps (MICA-CM) – is based on a memetic ICA and chaotic maps, which are responsible for exploration of the search space, while back-propagation is used for an effective local search on the best solution obtained by ICA. Experiment results confirm our proposed algorithm to be highly competitive compared to other recently reported methods.