The asymptotic behavior of undiscounted value iteration in Markov decision problems