The sum rate maximization beamforming problem for a multi-cell multi-user multiple-input single-output interference channel(MISO-IC)system is considered.Conventionally,the centralized and distributed beamforming solut...The sum rate maximization beamforming problem for a multi-cell multi-user multiple-input single-output interference channel(MISO-IC)system is considered.Conventionally,the centralized and distributed beamforming solutions to the MISO-IC system have high computational complexity and bear a heavy burden of channel state information exchange between base stations(BSs),which becomes even much worse in a large-scale antenna system.To address this,we propose a distributed deep reinforcement learning(DRL)based approach with lim⁃ited information exchange.Specifically,the original beamforming problem is decomposed of the problems of beam direction design and power allocation and the costs of information exchange between BSs are significantly reduced.In particular,each BS is provided with an inde⁃pendent deep deterministic policy gradient network that can learn to choose the beam direction scheme and simultaneously allocate power to users.Simulation results illustrate that the proposed DRL-based approach has comparable sum rate performance with much less information exchange over the conventional distributed beamforming solutions.展开更多
基金supported by the joint research project with ZTE Corporation under Grant No.HC-CN-2020120002.
文摘The sum rate maximization beamforming problem for a multi-cell multi-user multiple-input single-output interference channel(MISO-IC)system is considered.Conventionally,the centralized and distributed beamforming solutions to the MISO-IC system have high computational complexity and bear a heavy burden of channel state information exchange between base stations(BSs),which becomes even much worse in a large-scale antenna system.To address this,we propose a distributed deep reinforcement learning(DRL)based approach with lim⁃ited information exchange.Specifically,the original beamforming problem is decomposed of the problems of beam direction design and power allocation and the costs of information exchange between BSs are significantly reduced.In particular,each BS is provided with an inde⁃pendent deep deterministic policy gradient network that can learn to choose the beam direction scheme and simultaneously allocate power to users.Simulation results illustrate that the proposed DRL-based approach has comparable sum rate performance with much less information exchange over the conventional distributed beamforming solutions.
