Scaling Up Reinforcement Learning for Traffic Smoothing