Optimum training design for neural network in synthesis of robust model predictive control

摘要

The paper deals with determining the neural network model uncertainty for the purpose of robust controller design. The approach presented in the paper is based on the application of optimum experimental design for the choice of sequences providing the most informative data during the training of neural network. As a criterion quantifying the quality of training process a measure operating on the Fisher Information Matrix related to the estimates of network parameters is used. Then, it is possible to analyze the variance of the predicted network response and estimate how possible variations of parameter values influence the changes observed in the predicted model output. This allows to construct an appropriate cost function for the control system taking into account the model uncertainty and incorporate it into model predictive control scheme.