Averaging vs. Discounting in Dynamic Programming: a Counterexample

Abstract

We consider countable state, finite action dynamic programming problems with bounded rewards. Under Blackwell's optimality criterion, a policy is optimal if it maximizes the expected discounted total return for all values of the discount factor sufficiently close to 1. We give an example where a policy meets that optimality criterion, but is not optimal with respect to Derman's average cost criterion. We also give conditions under which this pathology cannot occur.