A homotopy approach to improving PEM identification of ARMAX models (Q5947642)

The paper presents a homotopy approach to improving the prediction error method (PEM) identification of ARMAX model. PEM estimates of ARMAX model parameters are determined as the global minimum of the criterion function, which is however not always unimodal because of the MA noise model part. Because an optimization-based PEM identification algorithm has a potential risk to be stuck at a local minimum, that circumstance results in a poorly identified model. A homotopy continuation method is introduced to solve this problem. The utilized idea is to start the estimation with the criterion function for PEM identification of the ARX model, which is gradually deformed into the actual one for PEM identification of the ARMAX model as the algorithm iterates. By building the deformation into the usual recursive procedure for the ARMAX identification, and introducing a scheme to control the solution continuously staying in the global minima of the deformed criterion functions, the homotopy-based PEM identification algorithm is implemented. It is implemented in such a way that it has very good convergence performance, with only little increase in computational load compared to the usual PEM algorithm. In the simulations, the homotopy-based prediction method (HBPEM) algorithm is compared with the bias-remedy least mean square equation error algorithm (BRLEA) that is said to have better convergence performance than the so-called ``composite regressor algorithm'' (CRA). It is found that BRLEA provides a compromise between the convergence and bias aspects. It is shown that the ``remedy parameter'' should be chosen very carefully in order to have a stable estimation with small biases. It is declared that an attempt was made to introduce a homotopy continuation method into BRLEA, but no improvement has been found. It is found that in comparison to the proposed HBPEM, the BRLEA has lower convergence rate, larger biases in the parameter estimates and is more sensitive to initial values. However, since it does not make any assumption on noise structure, it seems to have better robustness against the unmodeled dynamics. Simulation results show that as far as ARMAX identification is considered, the proposed HBPEM algorithm performs much better than the BRLEA.