Maximum a Posteriori Based Regularization Parameter Selection

Paper in proceeding, 2011

The l(1) norm regularized least square technique has been proposed as an efficient method to calculate sparse solutions. However, the choice of the regularization parameter is still an unsolved problem, especially when the number of nonzero elements is unknown. In this paper we first design different ML estimators by interpreting the l(1) norm regularization as a MAP estimator with a Laplacian model for data. We also utilize the MDL criterion to decide on the regularization parameter. The performance of these new methods are evaluated in the context of estimating the Directions Of Arrival (DOA) for the simulated data and compared. The simulations show that the performance of the different forms of the MAP estimator are approximately equal in the one snapshot case, where MDL may not work. But for the multiple snapshot case both methods can be used.