So far, the controllers using the switching scheme of multiple models have been proposed for controlling the plants whose parameter values change occasionally and largely owing to wide varieties of environments involving different operating conditions, uncertainty, and so forth. Although the stability of the controllers has been analyzed, the efficiency of the models has not been examined in detail, where the present learning algorithm of competitive associative net called CAN2 is formalized and utilized to minimize the energy defined as the sum of identification errors of switched models so that the models can be efficiently utilized when the identification errors are large. In order for the CAN2 to learn and switch multiple models for reproducing the forward dynamics of the plant, two kinds of CAN2, called CAN2-1 and CAN2-2, are introduced. The former basically is for learning to achieve a piecewise linear approximation of nonlinear plant dynamics, and the latter basically is for learning to achieve adaptive linear approximation of time-varying plant dynamics. The CAN2-1 and CAN2-2 are embodied into the conventional model-based predictive controllers, which are called APC1 (adaptive predictive controller 1) and APC2, respectively. Computer simulation clarifies the behavior of the APC1 and APC2 and shows that the APC1 is useful for nonconvex nonlinear plants while the APC2 is for time-varying plants.
View full abstract