Identification of a weighted combination of multivariable local linear state-space systems from input and output data

Discusses a method for the determination of a weighted combination of local linear state-space systems from input and output data. The method is iterative and each iteration consists of two steps. The first step is to determine the weighting functions given the local models. This problem is solved by using an extended Kalman smoother. The second step is to identify the local models given the weights. For this step we optimize a cost function that represents a tradeoff between local and global learning. For this optimization we use a gradient search method in combination with an appropriate projection in the parameter space to deal with similarity transformations