The Thing That We Tried Didn't Work Very Well : Deictic Representation in Reinforcement Learning