Model selection for regularized least-squares algorithm in learning theory