FedCritic-MIMO: Communication-Efficient Serverless Federated Critic Learning for Massive-MIMO Resource Control in Open and Disaggregated 6G RANs

In 6G radio access networks (RANs), controllers are increasingly disaggregated and independently deployed. This creates a significant challenge: how can these controllers coordinate their resource allocation—such as user scheduling, power control, and beamforming—to manage interference and meet quality-of-service (QoS) requirements without relying on a centralized trainer or a common parameter server?

The authors propose FedCritic-MIMO, a serverless multi-agent reinforcement learning (MARL) framework. Unlike traditional approaches that require centralized training or the sharing of full models, FedCritic-MIMO allows each base station (BS) to maintain its own local actor and personalized critic components. Collaboration is limited to a shared critic subnetwork, which is exchanged peer-to-peer over an interference-aware graph. To ensure communication efficiency, the framework employs three key mechanisms:

1. Utility-aware triggering: Updates are only sent when they are deemed significant based on critic innovation, queue urgency, and interference intensity.
2. Adaptive layer-wise top-k sparsification: Only the most important model parameters are transmitted, with error feedback to maintain accuracy.
3. Interference-aware fusion: The system weights updates from neighbors based on the strength of their physical interference coupling.

Simulations in strongly interference-coupled, reuse-1 massive-MIMO OFDMA environments demonstrate that FedCritic-MIMO outperforms existing heuristic, independent-learning, and centralized-training baselines. It achieves the highest held-out network throughput and improves both user-rate distribution and mean SINR. Furthermore, the framework successfully satisfies long-term QoS constraints while reducing critic-communication overhead by approximately 76% compared to uncompressed distributed exchange. The authors also provide theoretical guarantees for the convergence of this compressed, peer-to-peer critic recursion.