Adaptive Federated Learning for 6G: A Multi-Agent Architecture for 6G Edge Intelligence

In a highly dense ultra-6G environment, synchronising decisions across all three layers of communications, computations, and controls is no longer theoretical but practical; edge servers, mobile devices, and cloud servers will have to operate in real time. Furthermore, they will have to adapt to changing link conditions and consume less power over those links, and also have to process non-IID data sets. Unfortunately, none of today’s learning-based orchestration approaches is robust in this environment. This work is an introduction to a hierarchical trust-aware orchestration framework, which is capable of working reliably in the presence of real 6G conditions. Multi-Agent Architecture for 6G Edge Intelligence (MAAFL-6G), the proposed system is comprised of a combination of agents (device level, edge level and cloud level) that utilise Hierarchical Multi-Agent eXtended Proximal Policy Optimisation based Federated Learning (H-MAXPPO) policy to make decisions on the appropriate scale in time. Reliability increases to about 99.8%, and more importantly, it stays there instead of wobbling or collapsing mid-training. The latency, which starts above two seconds in early rounds, settles comfortably around 0.48−0.55 seconds. Energy consumption remains low and steady, roughly 0.0175-0.0195 J per task, whereas centralised baselines burn significantly more. Even the subtler metrics, such as cross-layer stability and the speed at which trust scores settle, point in the same direction: the hierarchical approach is simply more resilient to the messy, multi-timescale nature of 6G environments than either centralised or purely local learning.