Multimodel Control Design Using Unsupervised Classifiers

Multimodel approaches derive a smooth control law from the blending of local controllers using the concept of validities and domain overlapping. In this paper, it is demonstrated that unsupervised classification algorithms can be of a great help to design such parameters as the number of the models and their respective clusters, which will be performed using a respectively Rival Penalized Competitive Learning (RPCL) and simple or fuzzy K-means algorithms. The classical multimodel approach follows by deriving parametric model identification using the classification results for models orders and then parameters estimation. The determination of the global system control parameters results from a fusion of models control parameters. The case of a second order nonlinear system is studied to illustrate the efficiency of the proposed approach, and it is shown that this approach is much simpler that other multimodel control design methods which generally require a huge number of neighboring models.