ReLAx: Deep Reinforcement Learning Based Resource Allocation for Next-G RANs

Next-Generation Radio Access Network (Next-G RAN) will leverage a novel architecture that accelerates the tran-sition from inflexible networks to agile and adaptable networks. In this paper, we introduce a novel Deep Reinforcement Learning-based Resource Allocation (ReLAx) framework to deal with the joint optimization of UE association and power allocation in Next-G RAN systems. The ReLAx problem has been formulated to maximize the network Energy Efficiency (EE) under the constraints of Quality of Service (QoS), fronthaul link, functional split configuration, and transmit power budget. The optimization problem is cast as non-convex and NP-complete. A multitask Deep Deterministic Policy Gradient (DDPG) method is proposed to solve the complexity, in which two actors are trained to gen-erate UE association and power allocation, respectively. To speed up the training process and reduce computational resources, we introduce soft multitask learning as a constraint during training so that one model would not drift too far away from the other one. Our real-time experiments on a fully containerized Next-G RAN testbed show the effect of functional splits on CPU utilization and system latency. In addition, simulation results show that the proposed resource allocation solution outperforms competing traditional algorithms, such as standard DDPG and Weighted Minimum Mean Square Error (WMMSE).