[Paper] AWaRe-SAC: Proactive Slice Admission Control under Weather-Induced Capacity Uncertainty

Source: arXiv - 2601.05978v1

Overview

The paper introduces AWaRe‑SAC, a proactive slice‑admission‑control system designed for millimeter‑wave (mmWave) x‑haul networks that suffer capacity swings due to weather—especially rain. By forecasting link conditions with deep learning and coupling that prediction to a reinforcement‑learning (Q‑learning) controller, the authors demonstrate a way to keep service‑level agreements (SLAs) intact while boosting operator revenue.

Key Contributions

• Weather‑aware capacity predictor: A neural‑network model trained on real‑world mmWave link measurements that forecasts short‑term capacity drops caused by rain attenuation.
• Proactive admission controller (AWaRe‑SAC): A Q‑learning‑based algorithm that decides whether to admit, defer, or reject new network slices before the anticipated capacity degradation occurs.
• End‑to‑end evaluation on live deployment data: Experiments use traffic traces from a dense‑urban mmWave x‑haul testbed, incorporating realistic slice demand dynamics and rain‑fade models.
• Revenue‑focused objective: The framework jointly optimizes QoS compliance and long‑term operator revenue, showing 2–3× higher average earnings compared with reactive baselines.
• Scalable design: The solution runs in real time with modest computational overhead, making it suitable for integration into existing network‑orchestration platforms.

Methodology

1. Data collection & preprocessing – The authors gathered per‑minute capacity logs from a city‑wide mmWave backhaul network together with concurrent weather station data (rain rate, humidity).
2. Capacity forecasting – A lightweight LSTM (Long Short‑Term Memory) network predicts the next 5–10 minutes of link capacity, outputting a probability distribution over possible attenuation levels.
3. Reinforcement‑learning controller – The predicted capacity distribution feeds into a Q‑learning agent. The state includes current slice load, predicted capacity, and SLA slack; actions are admit, delay, or reject a slice request. The reward balances two terms: (i) penalty for QoS violations and (ii) revenue earned from admitted slices.
4. Training & deployment – The agent is trained offline using historical demand and weather traces, then deployed online where it continuously updates its Q‑values based on observed outcomes.
5. Benchmarking – The proactive approach is compared against (a) a reactive threshold‑based controller that reacts only after capacity drops, and (b) a static admission policy that ignores weather.

Results & Findings

• The LSTM predictor achieves a mean absolute error of 0.12 Gbps on 5‑minute ahead forecasts, sufficient to trigger timely admission decisions.
• Q‑learning converges after ~10 k training episodes, and the online policy remains stable even when rain intensity spikes abruptly.
• Revenue gains stem from the controller’s ability to pre‑emptively defer low‑priority slices during forecasted capacity dips, preserving high‑value traffic.

Practical Implications

• Network operators can embed AWaRe‑SAC into their orchestration stacks (e.g., ONAP, OSM) to automatically hedge against weather‑induced outages, reducing manual re‑configuration.
• Slice‑as‑a‑Service (SlaaS) platforms gain a data‑driven tool to price slices dynamically—high‑revenue slices can be prioritized when rain is expected, while lower‑margin slices are throttled or shifted.
• Edge‑cloud providers can use the forecast‑aware admission logic to decide where to place workloads (edge vs. central) based on anticipated backhaul capacity, improving end‑user latency.
• Hardware vendors may consider exposing real‑time weather metrics to the control plane, enabling tighter integration with predictive controllers like AWaRe‑SAC.
• Developer APIs can expose “admission‑confidence” scores, allowing applications to adapt (e.g., degrade video quality) before a QoS breach occurs.

Limitations & Future Work

• Geographic specificity: The model is trained on a single urban deployment; performance may vary in suburban or rural settings with different rain patterns.
• Prediction horizon: Forecasts beyond 10 minutes lose accuracy, limiting the controller’s look‑ahead capability for longer‑duration slices.
• Model complexity vs. edge constraints: While the LSTM is lightweight, ultra‑low‑power edge devices might still struggle; exploring even simpler predictors (e.g., ARIMA) is a possible avenue.
• Multi‑operator scenarios: The current framework assumes a single operator’s control over the entire x‑haul; extending to shared infrastructure with competing slice requests remains open.
• Broader weather factors: Future work could incorporate snow, fog, or temperature‑induced hardware variations to broaden the robustness of the admission controller.

Authors

• Dror Jacoby
• Yanzhi Li
• Shuyue Yu
• Nicola Di Cicco
• Hagit Messer
• Gil Zussman
• Igor Kadota

Paper Information

• arXiv ID: 2601.05978v1
• Categories: cs.NI, cs.LG
• Published: January 9, 2026
• PDF: Download PDF