Distributed Coordinated Beamforming Based on Multi-Agent Reinforcement Learning in Multicell MISO Systems

Maximizing the sum-rate in multicell multiple input single output (MISO) systems is a non convex and NP-hard problem. Most existing algorithms trying to solve this problem are suboptimal with high computational cost and high system interaction overhead. In this paper, we propose a coordinated beamforming (CB) scheme based on multi-agent reinforcement learning (MARL) to maximize the sum-rate of the multicell MISO systems with limited information feedback and exchange. Specifically, the training of the proposed MARL network is guided by the actual sum-rate of the multiple cells, and the execution is performed totally locally by using the local channel quality information feedback. Simulation results show that compared to the existing distributed coordinated beamforming scheme, the proposed scheme achieves similar performance by using much reduced information overhead.