Reinforcement Learning Using Approximate Belief States