Optimizing roundabout management via deep reinforcement learning with safety and comfort constraints