Improved bound on the worst case complexity of Policy Iteration

Solving Markov Decision Processes (MDPs) is a recurrent task in engineering. Even though it is known that solutions for minimizing the infinite horizon expected reward can be found in polynomial time using Linear Programming techniques, iterative methods like the Policy Iteration algorithm (PI) remain usually the most efficient in practice. This method is guaranteed to converge in a finite number of steps. Unfortunately, it is known that it may require an exponential number of steps in the size of the problem to converge. On the other hand, many open questions remain considering the actual worst case complexity. In this work, we provide the first improvement over the fifteen years old upper bound from Mansour & Singh (1999) by showing that PI requires at most k k−1 · kn n + o ( k n n ) iterations to converge, where n is the number of states of the MDP and k is the maximum number of actions per state. Perhaps more importantly, we also show that this bound is optimal for an important relaxation of the problem.