Dynamic spectrum access and sharing through actor-critic deep reinforcement learning