An Improved Policy Iteration Algorithm for Partially Observable MDPs