Design of neural estimators for multisensors: second order backpropagation, initialization and generalization

The static networks as multilayer perceptrons (MLP) are able to implement boolean logic functions, to partition the pattern space for classification problems and to approximate nonlinear functions. The present goal is to study their capabilities when used to effect a parametric estimation without an explicit model (supervised estimation). One can also consider varying parameters. The paper focuses on the combined use of second order backpropagation and performant initial values of the weights. The authors also study the effects of noise introduction into training sets in robustness and generalization faculties. These results are then applied to the extraction of parameters from real multisensor signals. They show a drastic reduction in training time, improved robustness against local minima and better generalization