Structure study and automatic search of neural networks for real-valued function approximation

Summary form only given. We studied the problem of using feedforward neural networks (FNNs) to approximate unknown real-valued functions in an industrial application-the mappings between automobile engine control variables and performance parameters and gained heuristic knowledge about network structures from a vast number of experiments. Incorporating such heuristic knowledge, we developed a program for automatic search of optimal FNNs based on evolutionary computation techniques, called PASS (Program for Automatic Structure Search). PASS has been successfully applied to the engine mapping problem and has shown the promise to be a general and efficient tool for automatic determination of FNNs for real-valued function approximation.