Light-weight Reinforcement Learning with Function Approximation for Real-life Control Tasks