Information Based Selection of Neural Networks Training Data for S.I. Engine Mapping

The paper deals with the application of two techniques for the selection of the training data set used for the identification of Neural Network black-box engine models; the research starts from previous studies on Sequential Experimental Design for regression based engine models. The implemented methodologies rely on the Active Learning approach (i.e. active selection of training data) and are oriented to drive the experiments for the Neural Network training. The methods allow to select the most significant examples leading to an improvement of model generalization with respect to a heuristic choice of the training data. The data selection is performed making use of two different formulation, originally proposed by MacKay and Cohn, based on the Shannon's Statistic Entropy and on the Mean Error Variance respectively. These techniques have been applied to assist the training of artificial Neural Networks for the estimation of engine torque and exhaust emissions of an S.I. engine, to be embedded into a powertrain dynamic model for the optimal design of engine control strategies (O.D.E.C.S.), now in use at Magneti Marelli.