An optimal design method for multilayer feedforward networks

Today, there exist many examples of artificial neural network (ANN) technology implementations. By far the most successful of these have been with multilayer feedforward networks, primarily utilizing the backpropagation (BPN) paradigm. These networks are universal classifiers and as such are able to address various engineering problems. However the designing and building of these networks is not well defined. In fact, there may not exist a practical step-by-step method of design which can be broadly applied since theoretically there are an infinite number of configurations which would have to be tested to identify the optimal design. Practically, if certain network parameters are bounded over a reasonable range it is possible to design an optimal network within these guidelines. In this paper, a BPN network is designed by applying this method. The results suggest the method is efficient, reliable and probably yields the absolute optimal network. The nonlinear systems modeling and simulation problem, power flow analysis, is undertaken with the BPN network being compared with the classical Newton-Raphson method.