Experiments on Regularizing MLP Models with Background Knowledge

. In this contribution we present results of using possibly inaccurate knowledge of model derivatives as part of the training data for a multilayer perceptron network (MLP). Even simple constraints offer significant improvements and the resulting models give better prediction performance than traditional data driven MLP models. 1 Introduction An increasingly important application of neural networks is in modeling nonlinear plants for simulation and control purposes. Often the training must be done with measurements from normal operating situations. Then, depending on the operating statistics during the data collection, many important features of the process behavior may be lacking from the data. An important type of knowledge from any process to be controlled is the effect of control variables to the output variables. Typically the process experts can tell the direction and coarse magnitude of the change in product variables as function of the control variables, even though they could..