Deep Reinforcement Learning for Channel Selection and Power Allocation in D2D Communications

Device-to-device (D2D) communication is regarded as a key technical component of the fifth-generation (5G), D2D communication usually reuses spectrum resources with cellular users (CUs). To mitigate interference to cellular links and improve spectrum efficiency, this paper investigates a sum-rate maximization problem in the underlay of D2D communication. Particularly, a joint channel selection and power allocation framework based on multi-agent deep reinforcement learning is proposed, named Double Deep Q-Network (DDQN). It can adeptly select the channel and allocate power in a dynamic environment. The proposed scheme only requires local information and some outdated nonlocal information, which reduces signaling overheads significantly. Simulation results show that the proposed scheme can improve the D2D sum rate and ensure quality-of-service (QoS) of CUs compared with other benchmarks.