Intelligent Multi-Model Minimum Variance Controllers

摘要

In this paper a class of intelligent Multi Model Partitioning Minimum Variance (MMMV) controllers for input/output models is presented. The method is based on the derivation of the adaptive model conditional Minimum Variance (MV) control. Thus, the problem is reduced to selecting the true model among a set of "candidate" models, using the well known multi-model partitioning theory of Lainiotis, for general (not necessarily Gaussian) data pdf's. The cumulative control is the average of the model conditional minimum variance controls, weighted by the respective a-posteriori probability that the particular model is the true model. Then tow different realizations are performed. First we use a bank of conventional Kalman filters. Second we implement a bank of neural networks, using as learning algorithm some localized approaches of the Extended Kalman filter. Since the underlying system's structure is usually unknown, different order models are used in the banks. This approach has been proved, via simulations, very effective. Indeed, the method is 100% successful in selecting the correct order model and identifying the model's parameters that are necessary for the controller's design. Finally, other interesting features of the resulting controllers are discussed.