关键词: 密集型网络, 模型驱动, 资源分配, 深度强化学习, 交替方向乘子法

Abstract: In order to optimize resource allocation of dense network under limited channel state information, a model-driven learning framework combined with alternating direction method of multipliers, as well as deep reinforcement learning algorithm, is proposed. This framework differs from data-driven ones, which enables one-to-one modeling of specific problems. The steps on how to model resource allocation include: alternately optimizing base station selection, power, and subcarrier allocation with alternating direction method of multipliers; using deep reinforcement learning algorithm to optimize weights, solve target functions and improve performance of the system; using effective channel state information instead of redundant information to reduce overhead on communication; adding constraints on users’ quality of service requirements to maximize cell spectral efficiency while ensuring user experience, which can maximize the spectral efficiency of the cell while ensuring users’ experience. The simulation results show that the model-driven learning framework can converge in a small number of iterations.

Key words: dense network, model-driven, resource allocation, deep reinforcement learning, alternating direction multiplier method