Supervised learning for feed-forward neural networks: a new minimax approach for fast convergence

An approach to the problem of the learning process for feedforward neural networks, based on an optimization point of view, is proposed. The developed algorithm is a minimax method based on a configuration of the quasi-Newton and steepest-descent methods. The optimum point is reached by minimizing the maximum of the error functions of the network without requiring any tuning of internal parameters. The algorithm is tested on several widespread benchmarks and shows superior convergence properties when compared with other algorithms available in the literature. Significant experimental results are included