Deep reinforcement learning-based resource joint optimization for millimeter-wave massive MIMO systems

Aiming at the problem of low throughput and energy efficiency caused by limited wireless resources, huge power consumption, and mutual constraints between energy efficiency and system capacity in millimeter-wave large-scale multiple-input multiple-output systems, a resource co-optimization method based on deep reinforcement learning was proposed. The method was adopted in a three-stage strategy, firstly, an RF beamformer was constructed to reduce the hardware cost and total power consumption through a small number of RF chains; secondly, a baseband precoder was designed using the effective channel state information; and finally, a two-tier deep reinforcement learning architecture was designed and applied to realize dynamic discrete bandwidth and continuous power resource allocation. Experimental results show that the proposed joint optimization method significantly improves the throughput and energy efficiency of the system compared with the single-stage all-digital precoding and hybrid precoding equal resource allocation methods and the particle swarm optimization-based resource allocation algorithm.