Parameter estimation using decomposed algorithms with fast convergence rates

This paper addresses the use of decomposed algorithms for numerical computation of parameter estimates, where the estimation problem is solved in stages. At each stage, an estimate of a subset of parameters is computed by minimizing a cost function with the remaining parameters fixed. In this paper, we study the convergence speed performance of these algorithms, and present a method based on the epsilon decomposition algorithm of Siljak to identify parameter subsets that lead to decomposed algorithms with fast convergence properties. The ideas and results presented are applied to speed and parameter estimation for induction machines. We present several improvements over heuristic decompositions used in the past for this application, and provide experimental verification.