Meta-learning-driven deep reinforcement learning for green resource management in 5G/6G networks

Future sixth-generation (6G) mobile networks must satisfy demanding service requirements while operating under increasingly stringent sustainability constraints and variable energy conditions. Static heuristics and conventional deep reinforcement learning (DRL) policies can adapt slowly to changes in traffic, radio conditions, and energy availability, leading to inefficient resource use andavoidable emissions. We propose a meta-reinforcement learning framework that combines Model-Agnostic Meta-Learning (MAML) with Proximal Policy Optimization (PPO) for energy-aware network orchestration. The method learns an initialization that can be adapted with a small number of gradient updates, supporting rapid response to changing operating conditions. We instantiate the framework in the Access and Mobility Management Function (AMF) of a virtualized core network, where joint CPU and memory allocation is formulated as a continuous-control problem with energy and carbon-intensity objectives.Evaluations against PPO, A2C, DQN, TD3, and SAC show improvements of up to 37% across energy efficiency, energy consumption, carbon footprint, and convergence-related metrics, while maintaining near-optimal service-level agreement compliance. These results indicate that meta-learned policy adaptation can support more sustainable resource management in dynamic 5G/6G corenetworks.