Efficient Reinforcement Learning from Partial Observability