Source-linked AI summary
Sample-and-Score: Surrogate-Guided Coordinated Spatial Reuse in IEEE 802.11 Networks
Maksymilian Wojnar
TL;DR
Jointly optimizing simultaneous Wi-Fi transmissions is challenging because Co-SR has a combinatorial configuration space. This paper samples configurations and scores them with an offline-trained surrogate. On unseen topologies the surrogate recovers 52–77% of the perfect surrogate’s median rate, while a perfect surrogate reaches 95% of T-optimal throughput with 8192 candidates.
Problem
Co-SR configuration choices grow combinatorially, making real-time optimization challenging as networks scale.
Source: §I · p. 2Method
The method samples random Co-SR configurations, scores them with an offline-trained surrogate, and selects the best candidates.
Source: §I · p. 2, §III · p. 4Results
On unseen topologies with 2–16 APs, the surrogate recovers 52–77% of the perfect surrogate median rate; a perfect surrogate reaches 95% of T-optimal throughput with 8192 candidates.
Source: §IV · pp. 4–5Takeaways & Limitations
The surrogate exceeds DCF by 16% without online interaction and evaluates 1024 candidates over 30× faster than H-MAB convergence.
Source: §IV · p. 5, §V · p. 5The surrogate plateaus beyond 2048 candidates, with ranking many configurations identified as the main limitation.
Source: §V · p. 5
Abstract
from arXiv · showhide
Coordinated spatial reuse (Co-SR) in IEEE 802.11bn increases area throughput through simultaneous transmissions, yet the combinatorial configuration space makes real-time optimization challenging. We propose a sample-and-score approach: draw random Co-SR configurations, score each with a surrogate model, and select the best. The surrogate is trained offline and deployed without online fine-tuning or sequential exploration, yet performs close to multi-armed bandit and heuristic baselines that need multiple interaction steps or extensive measurements. On unseen topologies with 2-16 access points, it recovers 52-77% of a perfect surrogate's throughput. A perfect sample-and-score approach reaches 95% of the upper bound throughput with 8192 candidates.
I. INTRODUCTION
Co-SR can improve throughput by coordinating simultaneous transmissions, but jointly choosing power, served stations, and MCS is a delicate optimization problem. Existing approaches face scaling or adaptation challenges, motivating an offline surrogate that scores sampled configurations.
I. INTRODUCTION
- Co-SR jointly tunes transmission power, served stations, and MCS to balance simultaneous-transmission throughput gains against interference at receivers.The balance is particularly challenging in dense deployments with overlapping basic service sets.Source: §I · p. 2
- The combinatorial configuration space challenges real-time orchestration; heuristics become intractable as networks grow, while bandit convergence time increases with topology size.Deep reinforcement learning approaches use fixed catalogs and require action-space regeneration and retraining for unseen deployments.Source: §I · p. 2
- Prior methods commonly decouple or fix MCS selection, although MCS critically determines achievable rates.MCS negotiation is not formally part of the Co-SR specification.Source: §I · p. 2
- The proposed approach samples random Co-SR configurations, scores them with a surrogate, and selects the best ones for execution.The pipeline encodes configurations as graphs before scoring and selection.Source: §I · p. 2, Figure 1 · p. 2
- The surrogate is trained offline and replaces test-time simulation, requiring a single forward pass without the convergence period needed by online methods.The graph-structured representation is intended to generalize across topologies and scale to groups of up to 16 APs.Source: §I · p. 2
II. PROBLEM FORMULATION
A Co-SR configuration specifies the APs transmitting simultaneously and each active AP’s STA, power level, and MCS. The resulting configuration space grows rapidly with network size, making exhaustive search infeasible at transmission-opportunity timescales.
II. PROBLEM FORMULATION
- A Co-SR configuration specifies which APs transmit simultaneously and, for each active AP, its served STA, transmission power, and MCS index.Source: §II · p. 2
- The configuration space grows combinatorially with the number of APs and associated STAs.Each AP may also be inactive, adding a no-transmission option.Source: §II · pp. 2–3
- With 14 MCS indices, 4 power levels, and 4 STAs per AP, just 5 APs face over 1011 possible configurations, making exhaustive search infeasible at per-TXOP timescales.Source: §II · p. 3
- The evaluation uses a custom Monte Carlo MAPC simulator because no full-stack simulator such as ns-3 supports Co-SR.It models path loss, concurrent-transmission SINR, reception probability, fading, and signaling overhead under full-buffer downlink traffic.Source: §II · p. 3
III. PROPOSED METHOD
The method represents Co-SR configurations as graphs and uses an offline-trained surrogate to score sampled candidates. It relies on sparse, available RSSI links and can regenerate catalogs when network conditions change.
III. PROPOSED METHOD
- The graph includes AP–STA association and AP–AP interference links, while message passing infers residual cross-interference without collecting every AP–STA RSSI link.The retained measurements are available through beacon listening and acknowledgment frames; this approximation trades exactness for lower signaling overhead.Source: §III · p. 3
- The surrogate predicts candidate data rates from graph-encoded configurations using a graph network with transformer updates and a four-component Gaussian-mixture output.The mixture-density model is preferred to MSE because stochastic data rates are better captured by a mixture, and MSE overfits in the reported experiments.Source: §III · p. 3
- On a hold-out validation set, the surrogate achieves Spearman correlation ρ = 0.87 and R2 = 0.70.Source: §III · p. 3
- Training uses 50,000 examples from 1,000 sampled realizations across 11 scenario types with 2–10 APs and up to 50 STAs.Each realization contributes 40 fully random configurations and 10 using simulator-computed ideal MCS to increase high-rate-region coverage.Source: §III · pp. 3–4
- For inference, the method samples N configurations, scores their graph representations, and selects candidates using a chosen strategy such as top-k round-robin application.The resulting catalog is a scheduling aid for a controller’s policy, rather than a standalone scheduler.Source: §III · p. 4
- The method operates on a snapshot of network state; dynamic environments require periodic or adaptive catalog regeneration after throughput drops or association changes.Source: §III · p. 4
IV. PERFORMANCE EVALUATION
The evaluation tests candidate-count scaling in a residential scenario, generalization to unseen open-space topologies, and inference time against H-MAB. It also compares the surrogate with other baselines under the reported evaluation settings.
Evaluation settings
- The open-space evaluation uses 2–16 APs in a 75 m × 75 m area, while the residential scenario models a 2 × 2 grid of apartments with one AP and four STAs per apartment.Surrogate methods jointly sample MCS with other parameters; DCF and H-MAB use ideal MCS, and SR groups uses its own MCS heuristic.Source: §IV · p. 4
Generalization to unseen topologies
- 52–77%: on unseen 2–16-AP open-space topologies, the surrogate recovers this share of the perfect surrogate median rate, with consistently low variability.Recovery is 77% at 2 APs and 52–54% at 14–16 APs; the largest topologies are outside the 2–10-AP training range.Source: Figure 3 · p. 4, §IV · p. 5
Residential candidate-count scaling
- At N = 8192, the perfect surrogate reaches 973 Mbit/s (95% of T-optimal); the learned surrogate reaches 656 Mbit/s at N = 1024 and plateaus beyond N = 2048.The learned surrogate exceeds DCF by 16% without online interaction.Source: §IV · pp. 4–5
Inference time
- 121 ms at 4 APs and 381 ms at 16 APs: split-batch inference scores 1024 candidates in these times, while full-batch inference runs out of memory for N > 128.H-MAB takes 3780 ms to converge at 4 APs and fails at 16 APs; the paper reports the surrogate as over 30× faster with no online interaction.Source: Table II · p. 5, §IV · p. 5
Baseline comparison
- 16%: the surrogate exceeds DCF without online interaction, while H-MAB achieves higher average throughput but requires many steps and SR groups has runs spanning 473–1205 Mbit/s.SR groups also requires crosslink RSSI measurements for every AP–STA pair; the surrogate uses readily available measurements instead.Source: §IV · p. 5
V. CONCLUSION
The paper presents surrogate-guided sample-and-score optimization for Co-SR. Its perfect-surrogate results reach 95% of T-optimal throughput, while the current evaluation is limited to static topologies with full-buffer downlink traffic and no full-stack simulator.
Conclusion
- The approach draws random Co-SR configurations, scores them with a surrogate, and selects the best; the surrogate uses sparse, readily available RSSI links.The perfect surrogate reaches 95% of T-optimal throughput by increasing catalog coverage.Source: §V · p. 5
- The learned surrogate exceeds DCF without online interaction and evaluates 1024 candidates over 30× faster than H-MAB convergence.The paper identifies ranking many configurations as the main limitation and a direction for future work.Source: §V · p. 5
Limitations
- The evaluation uses no full-stack simulator and considers only static topologies with full-buffer downlink traffic.The paper names packet-level simulation and hardware-testbed validation as future steps.Source: §V · p. 5