Continuous versus discrete action spaces for deep reinforcement learning