Evaluation of Gaussian Linear Model Order Selection Approaches

Model order selection approaches are usually evaluated in simulations by comparing the resulting model orders to the true model order. In this paper, the mean Kullback-Leibler divergence (MKD) between the selected model and the true model is proposed as an objective measure for evaluating different model order selection approaches in simulations. For Gaussian linear model order selection problems the Kullback-Leibler divergence are reduced to simple forms and the MKD can be easily computed. Simulation results show that the MKD is a reasonable measure to evaluate different Gaussian linear model order selection approaches, in terms of signal processing.