Multilayer perceptron training with inaccurate derivative information
Jouni Lampinen, Arto Selonen · 2002
Presents an algorithm for using possibly inaccurate knowledge of model derivatives as a part of the training data for a multilayer perceptron network (MLP). In many practical process control problems there are many well-known rules about the effect of control variables to the target variables. With the presented algorithm the basically data driven neural networks model can be trained to comply with these a priori rules, making the models more correct and decreasing the amount of required training data. Since the training of the rules is based on statistical error minimization the rules may be numerically inaccurate or contradictory. This makes the collection and maintenance of the rule bases much less expensive than in rule based expert systems. Currently the authors are incorporating the derivative based training into a commercial neural network process control tool.