This paper proposes a hyper-parameter estimation algorithm for the regularized least squares problem in the empirical Bayesian approach arising from FIR model identification using OBFs (orthonormal basis functions)-based kernels. The algorithm consists of two steps by dividing the decision variables into two groups and alternately minimizing with respect to each group. It is shown that DC (difference of convex functions) programming is effectively applicable in the algorithm because the search space is shown to be bounded. The paper includes a couple of numerical simulations to show the efficiency of the method.