Optimal weight adaptation for multilayer neural networks

A new approach for training multilayer neural networks is proposed based on the scheme of optimization of a multistage decision process. The optimal weights are computed on a layer-by-layer basis starting from the output layer. At each layer, a new representation of an error function expressed in terms of the weights and the dynamic desired summation inputs to each neuron is presented, and minimization of the error function yields the optimum weights. Simulation results for XOR and parity checker problems are also provided.<<ETX>>