Treinamento de redes perceptron utilizando janela dinâmica

Abstract

In this work we discuss neural networks and the bias-variance dilemma. We propose the Window method to be inserted into supervisioned neural training with noise data. The method has an intrinsic caracteristic of regularization, because it tries to eliminate noise while the network is beeing trained, reducing its in uence of the adjustment of network weights. We implement and analize the method in adaptive logic networks (ALN) and at multilayer perceptrons (MLP). Finally, we test the network in aplications as function aproximation, adaptive lters and time series prediction.