DRL-Based Green Resource Provisioning for 5G and Beyond Networks

Networks play a crucial role in our daily lives by efficiently delivering multimedia services. However, network operators face the challenge of ensuring efficient resource provisioning while balancing profit maximization and environmental objectives. Deep reinforcement learning (DRL) has emerged as an approach for resource allocation, leveraging observed data to make near-optimal decisions. However, the dynamic and complex nature of large-scale environments, such as wireless networks, poses challenges in designing appropriate rewards for DRL agents. This study investigates the feasibility of parallelization approaches to create network environments that facilitate efficient and generalizable learning for DRL algorithms. We propose two service provisioning solutions: DRL-MCTS for wireless networks and DRL-VNF for wired networks. These solutions prioritize green network objectives, including minimizing power consumption for nodes and links and selecting low carbon-emitting links. We compare our proposals to baseline methods, including a greedy algorithm, WMMSE, and the optimal solution. Through extensive experimental evaluations, our methods demonstrate significant improvements in reducing the network's environmental footprint while meeting Quality of Service requirements. In specific scenarios, our proposed solutions outperform the baseline approaches by up to 55%.