Interactive initialization of the multilayer perceptron

Abstract

A new multilayer preceptor initialization method is proposed and compared experimentally with a traditional random initialization method. An operator maps training-set vectors into a two-variate space, inspects bi-variate training-set vectors and controls the complexity of the decision boundary. Simulations with sixteen real-world pattern classification tasks have shown that in small-scale pattern classification problems, often complex classification rules and non-linear decision boundaries are not necessary. However, in cases where non-linear decision boundaries are required, the proposed weight initialization method is useful.