Reinforcement Learning with Efficent Active Feature Acquisition