Learning-Based Computation-Aware Access-Point Clustering and Joint Radio–Compute Allocation in Dynamic Cell-Free MEC Networks

User-centric cell-free massive MIMO is a key 6G technology to meet strict latency and energy-efficiency demands
of compute-intensive applications. In practical O-RAN deployments, the central unit and distributed unit compute capacity
fluctuates with network-slicing dynamics. Relying on fixed compute resources and access-point (AP) clustering based solely on
radio conditions leads to radio–compute mismatches, which in turn create bottlenecks that degrade both latency and energy
efficiency. To address this drawback, we propose a communication and computation-aware AP-clustering strategy with joint
resource allocation, leveraging a multi-agent reinforcement learning framework. The learned policies reduce average device energy consumption by 45% versus a fixed-compute RL benchmark and by 70% versus an offloading-first FPC heuristic across 10–50 Mbit/s, while maintaining 98–100% deadline satisfaction.