Improved deep reinforcement learning for car-following decision-making