Neural network training and simulation using a multidimensional optimization system

Abstract

A new approach is presented to neural network simulation and training that is based on the use of general purpose optimization software. This approach requires that the training problem should be formulated as the minimization of a cost function of the network weights. This cost function is a user written code called by the optimization system, which in turn provides the user with a variety of minimization procedures that can be combined via user programmable minimization strategies. Experimental results concerning several learning paradigms indicate that the approach is very convenient and effective and leads to the discovery of efficient training strategies.