Structure study of feedforward neural networks for approximation of highly nonlinear real-valued functions

We used feedforward neural networks (NNs) to approximate highly nonlinear real-valued functions for an industrial application-the mappings between automobile engine control variables and performance parameters. Back-propagation (BP) was applied for training the networks. Our experiments showed that with the same input and output layers, the same transfer function in the hidden layer(s), and the same total number of hidden nodes, four-layered networks with more nodes in the first hidden layer than in the second hidden layer out-performed the three-layered network (i.e., the one with a single hidden layer) in accuracy and training efficiency. Such fact held under different sample functions used, different initial conditions, different training periods, and different total numbers of hidden nodes. It seems a valuable heuristic for guiding automatic processes for structure optimization of feedforward NNs