Block-successive approximation for a discounted Markov decision model