🤖 machine learning

FM4WiFi: Flow Matching for Multi-AP Coordination in Dense Deployments of Beyond Wi-Fi 8 Networks

FM4WiFi is a generative machine learning framework that leverages flow matching and autoencoders to enable scalable, single-step network-wide coordinated spatial reuse for beyond Wi-Fi 8 dense deployments, overcoming the limitations of existing pair-wise coordination and slow convergence methods.

Original authors: Maksymilian Wojnar, Krzysztof Rusek, Katarzyna Kosek-Szott, Szymon Szott

Published 2026-08-06

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Maksymilian Wojnar, Krzysztof Rusek, Katarzyna Kosek-Szott, Szymon Szott

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the internet as a giant, bustling city where Wi-Fi signals are the cars driving on invisible roads. For years, these cars have been driving somewhat chaotically: every time a car wants to move, it checks if the road is clear, and if it is, it zooms ahead. If two cars try to enter the same intersection at the same time, they crash, and both have to wait. This "random access" method works okay in a quiet neighborhood, but in a crowded city like a stadium or a busy office building, it leads to gridlock. The roads are full, but no one is moving fast.

To fix this, engineers are building a new traffic system called "Multi-AP Coordination." Instead of every car deciding for itself, a central traffic controller tells groups of cars when to drive and how fast to go, so they can all move at the same time without crashing. This is the promise of the upcoming Wi-Fi 8 standard. However, there's a catch: figuring out the perfect schedule for hundreds of cars in a massive city is a math problem so huge that even the fastest supercomputers get stuck trying to solve it. If the controller takes too long to think, the cars sit idle, and the traffic jam gets worse. This is the puzzle a new paper, FM4WiFi, tries to solve.

The Problem: The Traffic Jam of the Future

The authors of this paper, researchers from AGH University of Krakow, are looking at a specific type of traffic control called Coordinated Spatial Reuse (Co-SR). In simple terms, this is when multiple Wi-Fi towers (Access Points) agree to transmit data at the same time, but at lower power levels, so they don't drown each other out. It's like a group of friends whispering in a library; they can all talk at once because they are quiet enough not to disturb the neighbors.

The problem is that in the current Wi-Fi 8 rules, this whispering is only allowed between two towers at a time. The researchers argue that for the future of Wi-Fi (beyond Wi-Fi 8), we need to coordinate many towers at once. But here's the nightmare: the number of possible ways to arrange these whispers is astronomical. If you have 30 towers, the number of combinations is so big that traditional computers can't calculate the best schedule in time. Existing methods either take too long to figure it out, require too much communication between the towers, or make unrealistic guesses about how the signals behave.

The Solution: An AI That "Dreams" the Best Schedule

Enter FM4WiFi. The researchers propose a new way to solve this traffic jam using a type of Artificial Intelligence called Generative AI. Instead of trying to calculate every single possibility (which is like checking every single route on a map to find the shortest one), the AI learns to "dream up" good solutions instantly.

Think of it like a master chef who has tasted thousands of perfect dishes. When you ask for a meal, the chef doesn't measure out every single grain of salt and pepper from scratch. Instead, they instantly create a delicious dish based on their deep understanding of what "good" tastes like. FM4WiFi does the same for Wi-Fi. It has been trained offline on a massive dataset of different network "topologies" (different arrangements of towers and devices). It has learned the "latent" (hidden) patterns of what a good Wi-Fi schedule looks like.

The system works in three main steps, like a high-tech assembly line:

The Encoder (The Observer): It looks at the current network state—who is connected, how strong the signals are, and where the interference is—and compresses this complex information into a simple, compact summary.
The Flow Matching Model (The Dreamer): This is the heart of the system. Using a technique called "Flow Matching," it takes that summary and instantly generates a list of potential schedules. It doesn't just guess; it flows from a state of "random noise" to a "perfect schedule" in a single, smooth step. It can generate dozens of different valid schedules in a fraction of a second.
The Surrogate Predictor (The Judge): Since the AI can't test these schedules on the real network (that would be too slow), it uses a "digital twin" or a fast calculator to predict how well each generated schedule would work. It picks the best ones and sends them to the network.

What They Found: Speed and Scale

The researchers tested their system in two ways: through detailed computer simulations and on a real-world testbed using actual laptops and Raspberry Pis.

The results were striking. In simulations, FM4WiFi could handle networks with 30 or more Access Points and produce a working schedule in less than one second. In contrast, the best existing methods (like complex math solvers or other AI approaches) either crashed, took minutes or hours to solve, or simply couldn't handle networks larger than a few towers.

Specifically, in a simulation of a dense office building with 16 towers, FM4WiFi found a schedule that delivered high data speeds in just 0.62 seconds. The traditional math-based methods took nearly 30 seconds to find a slightly worse solution, and one of the other AI methods (H-MAB) never finished calculating at all.

Even more impressively, when they tested this on a real physical testbed with six Wi-Fi towers and eight devices, the system worked "zero-shot." This means they didn't retrain the AI for this specific room; they just turned it on, and it immediately figured out how to coordinate the real hardware better than the standard "round-robin" (take turns) method or just letting everyone transmit at once.

The Limits and the Future

The paper is careful to note that this isn't magic. The system relies on a central controller connected by wires to all the towers, and it needs a powerful computer (a GPU) to do the fast calculations. Also, because it's a generative model, it doesn't guarantee a perfect solution every single time; it finds a very good solution very quickly. The researchers also point out that while it handles the "who talks to whom" and "how loud" parts brilliantly, it currently assumes the "what language to speak" (the data rate or MCS) is handled perfectly by a separate oracle, though they show their system can learn to do this too.

The authors conclude that by shifting from "searching" for a solution to "generating" one, we can finally unlock the potential of coordinated Wi-Fi in dense, crowded environments. It's a shift from trying to solve a puzzle by checking every piece to simply knowing what the finished picture looks like and painting it instantly. This approach suggests that for the chaotic, crowded networks of the future, the best way to manage traffic might not be to calculate every rule, but to learn the rhythm of the flow and dance to it.

Technical Summary: FM4WiFi: Flow Matching for Multi-AP Coordination in Dense Deployments of Beyond Wi-Fi 8 Networks

1. Problem Statement

The evolution of Wi-Fi networks is shifting from competitive, random channel access to tightly coordinated operation across Access Points (APs), a transition embodied in the upcoming Wi-Fi 8 standard's Multi-AP Coordination (MAPC). However, the current MAPC specification restricts cooperation to pairs of APs, fundamentally limiting the performance gains achievable in dense deployments (e.g., stadiums, industrial warehouses).

The specific challenge addressed is Coordinated Spatial Reuse (Co-SR), where multiple APs transmit concurrently at reduced power to mitigate interference. Effective Co-SR requires the joint selection and configuration of:

AP Groups: Which APs transmit simultaneously.
Station Selection: Which stations (STAs) these APs serve.
Wi-Fi Parameters: Modulation and Coding Scheme (MCS) indices and transmit power levels.

Existing solutions for Co-SR scheduling fail to scale to network-wide deployments (30+ APs) due to:

Signaling Overhead: Heavy reliance on global RSSI/SINR reporting.
Computational Complexity: Solver-based methods and Reinforcement Learning (RL) approaches suffer from exponential growth in computation time or convergence steps as network size increases.
Rigidity: Traditional RL methods often rely on fixed, pre-computed action spaces that cannot adapt to unseen topologies without retraining.
Idealized Assumptions: Many approaches decouple MCS selection or assume perfect MCS values, failing to capture the interplay between interference management and achievable data rates.

2. Methodology: The FM4WiFi Pipeline

The authors propose FM4WiFi, a generative machine learning pipeline designed to produce high-quality Co-SR configurations in a single inference step. The system operates offline for training and online for inference, consisting of three core components:

A. Graph Representation and Autoencoder

The Wi-Fi network is modeled as an undirected graph $G=(V, E)$ , where vertices represent APs and STAs, and edges represent transmission links and interference relationships.

Features: Edges carry a 7-attribute feature vector including RSSI, link type, activity status, selection status, MCS index, transmit power, and success probability.
GNN Autoencoder: A Graph Neural Network (GNN) based on the Message Passing Neural Network (MPNN) framework learns compact latent representations ( $z_e \in \mathbb{R}^{d_z}$ ) of the network state. It uses a Variational Autoencoder (VAE) architecture with a novel mixture categorical loss to handle heterogeneous attributes (continuous, categorical, and "N/A" values) effectively. This allows the model to generalize across graphs of varying sizes and topologies.

B. Flow Matching (FM) Generative Model

Instead of searching a fixed action space, FM4WiFi treats Co-SR configuration as a conditional generation task.

Mechanism: A Flow Matching model learns a continuous-time velocity field in the latent space of the autoencoder. It is trained to map a noise sample to a target latent configuration conditioned on recent network observations (context).
Training: The model minimizes the difference between the predicted velocity and the target velocity along a linear interpolation path between noise and data.
Inference: Configurations are generated by solving an Ordinary Differential Equation (ODE) using Euler's method, starting from noise and evolving toward a valid configuration in very few steps (e.g., 6 steps). This allows for the parallel generation of hundreds of candidate configurations.

C. Surrogate Data Rate Predictor

To evaluate candidates without access to a live network or a computationally expensive digital twin, a surrogate model is employed.

Function: This model acts as a learned digital twin, predicting the expected aggregate effective data rate for a given configuration.
Architecture: It mirrors the FM backbone and outputs a Gaussian Mixture Model (MDN) to capture the distribution of possible data rates.
Selection: Candidates generated by the FM model are scored by the surrogate, and the top- $k$ configurations are selected for execution.

D. Inference Workflow

Context Encoding: The central controller encodes recent network observations (RSSI, outcomes) into a latent context.
Candidate Generation: The FM model generates $N_{cand}$ (e.g., 128) candidate configurations in parallel.
Evaluation: The surrogate model scores all candidates.
Decoding & Execution: The top- $k$ (e.g., 8) candidates are decoded into explicit scheduling decisions (AP-STA pairs, MCS, power) and applied to the network.

3. Key Contributions

First Generative Pipeline for Co-SR: Introduces a framework that learns the distribution of high-quality configurations offline and generates them online, moving beyond search-based or fixed-action-space RL approaches.
Joint Optimization: Uniquely optimizes AP grouping, station selection, MCS, and transmit power simultaneously within a single model, eliminating the need for idealized MCS assumptions found in prior work.
Scalability and Speed: Demonstrates the ability to generate schedules in under 1 second for networks with 30+ APs, a scale where existing analytical and RL baselines become intractable.
Dynamic Robustness: Capable of immediate adaptation to topology changes, station joins, and mobility via event-triggered regeneration without retraining, contrasting with fixed-action-space RL methods.
Reproducibility and Testbed: Provides open-source code for the pipeline and dataset generation, alongside the first open-source experimental testbed on commercial-off-the-shelf (COTS) devices for evaluating Co-SR.

4. Evaluation Results

The authors evaluated FM4WiFi across extensive simulations and a real-world hardware testbed.

Scalability: In residential scenarios (up to 5x6 AP grids), FM4WiFi scales to 30+ APs with sub-second inference. In contrast, H-MAB (a state-of-the-art probing baseline) fails to converge within time budgets beyond 16 APs, and analytical solvers (T-Optimal, F-Optimal) become computationally prohibitive (taking minutes to hours).
Performance:
- Static Scenarios: FM4WiFi with its own MCS selection matches or exceeds H-MAB at medium-to-large scales and stays within 7% of the theoretical upper bound (F-Optimal) at large scales.
- Dynamic Scenarios: In mobility tests (stations moving, reassociating), FM4WiFi maintains stable throughput (680–760 Mbit/s in enterprise scenarios) by regenerating configurations upon topology changes. H-MAB fails to converge in these dynamic environments, remaining significantly lower.
- Open-Space: FM4WiFi matches H-MAB performance in small open-space networks (2–5 APs) when both use oracle MCS, confirming the generative model's ability to learn the underlying solution distribution.
Experimental Validation: On a 6-AP COTS testbed, FM4WiFi operated zero-shot on an unseen topology, outperforming simple round-robin and "all-at-once" baselines. While H-MAB eventually achieved higher throughput, it required a convergence period, whereas FM4WiFi provided immediate high performance.

5. Significance and Claims

The paper claims that FM4WiFi addresses the fundamental barrier of scalability in coordinated Wi-Fi. By reframing Co-SR scheduling from a combinatorial search problem to a conditional generation problem, the method decouples inference time from network size.

The authors position this work as a "learn-once, deploy-anywhere" paradigm. Unlike methods that require per-deployment learning or struggle with unseen topologies, FM4WiFi leverages offline training on diverse scenarios to generalize to new, complex deployments instantly. The significance lies in enabling network-wide coordination for Beyond Wi-Fi 8 systems, where the sheer density of APs renders traditional coordination mechanisms ineffective.

The authors modestly acknowledge limitations: the system requires a central controller with wired backhaul and hardware acceleration; it is a stochastic method without deterministic QoS guarantees; and the surrogate predictor introduces approximation errors. However, they assert that the approach offers a promising path forward for MAC-layer problems where action spaces are too large to enumerate and environments are too diverse for per-deployment learning.