Deep Reinforcement Learning Based Subchannel Selection and Power Allocation in Wireless Networks with Imperfect CSI.

Resource management is important for wireless networks. However, most model-driven resource allocation algorithms are limited by computational complexity and suboptimal spectral efficiency. This paper focuses on learning-based resource management to improve spectral efficiency in multicellular networks with imperfect channel state information (CSI). Considering imperfect CSI, the joint subchannel selection and power allocation problem can be formulated as a probability-constrained non-convex optimization problem. By means of parameter transformation, the non-convex optimization problem with probabilistic constraints is first transformed into a non-probabilistic optimization problem. Then, to solve this problem, a dual-module network based on dueling deep Q-network and deep deterministic policy gradient algorithm is proposed to maximize spectral efficiency. Simulation results show that the proposed dual-module network outperforms model-driven optimization algorithms such as fractional programming and existing deep reinforcement learning algorithms in terms of spectral efficiency.