Performance analysis of a MLP weight initialization algorithm

The determination of the initial weights is an important issue in multilayer perceptron design. Recently, we have proposed a new approach to weight initialization based on discriminant analysis techniques. In this paper, the performances of multilayer perceptrons (MLPs) initialized by non-parametric discriminant analysis are compared to those of randomly initialized MLPs using several synthetic and real-world benchmark learning tasks. Simulation results confirm that the proposed scheme yields a better initial state, as compared to randomly initialized MLPs. This leads to an improvement in the generalization performance and a reduction in training time, especially for complex or ill-posed problems.