Two Controller Networks Automatically Constructed Through System Linearisations and Learning

This study aims at comparing two linear controller networks as well as two methods to automaticly construct their architecture. The general idea of a controller network is to use a number of linear local controllers valid for diierent operating regions of a non-linear system. The two controller networks studied here are the \Clustered Controller Network" (CCN) and the \Model-Controller Network" (MCN). They diier by the method used for the selection of the controllers at each instant. In the CCN, the controllers are selected according to a spatial clustering of the operating space whereas in the MCN the selection of the controllers depends of the performance of the model associated to each local controller. The two diier-ent methods to construct the architecture of these controller networks are the \multiple oo-equilibrium system linearisations" and the \learning control through incre-mental network construction". It is shown that these network construction methods make the two controller networks general and systematic non-linear controller design approaches. However, the selection method applied by the MCN is preferable for control purposes since it is directly related to the controller capability unlike the method implemented by the CCN. In other hand, the exibility of the controller selection applied by the MCN makes accurate local control learning diicult to achieve. A mixture of this two methods of controller selection should remove these problems.