Fair AI-STA for Legacy Wi-Fi: Enhancing Sensing and Power Management with Deep Q-Learning

Peng Yi, Wenchi Cheng, Zhongjie Ju, Jingqing Wang, Jiahui Pan, Yuehui Ouyang, Wei ZhangarXiv 2024 · DOI · arXiv

Abstract

With the increasing complexity of Wi-Fi networks and the iterative evolution of 802.11 protocols, the Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) protocol faces significant challenges in achieving fair channel access and efficient resource allocation between legacy and modern Wi-Fi devices. To address these challenges, we propose an AI-driven Station (AI-STA) equipped with a Deep Q-Learning (DQN) module that dynamically adjusts its receive sensitivity threshold and transmit power. The AI-STA algorithm aims to maximize fairness in resource allocation while ensuring diverse Quality of Service (QoS) requirements are met. The performance of the AI-STA is evaluated through discrete event simulations in a Wi-Fi network, demonstrating that it outperforms traditional stations in fairness and QoS metrics. Although the AI-STA does not exhibit exceptionally superior performance, it holds significant potential for meeting QoS and fairness requirements with the inclusion of additional MAC parameters. The proposed AI-driven Sensitivity and Power algorithm offers a robust framework for optimizing sensitivity and power control in AI-STA devices within legacy Wi-Fi networks.

Introduction

In modern Wi-Fi environments, the "listen-before-talk" mechanism of the Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) protocol is increasingly strained. As Wi-Fi standards evolve from 802.11n to 802.11ax and beyond, newer devices often possess capabilities that can inadvertently crowd out older "legacy" hardware. High-density environments create a difficult trade-off between aggressive sensing—where a device ignores distant interference to transmit more frequently—and high transmit power, which can overcome noise but creates significant interference for neighboring Overlapping Basic Service Sets (OBSS).

Standard protocols are often too rigid to balance these factors, leading to uneven channel contention. This paper introduces an AI-driven Station (AI-STA) that utilizes Deep Q-Learning to dynamically tune its sensing thresholds and transmit power, ensuring it meets its own Quality of Service (QoS) needs without unfairly depriving legacy devices of channel access.

Methodology

The proposed AI-STA manages three primary operational parameters: the Carrier Sensing Threshold (CST), the Receiver Sensitivity Threshold (RST), and the Transmit Power ($P_s$). By adjusting these "knobs," the station can adapt its behavior to the current network state.

System model of the AI-STA interacting with legacy devices and OBSS.

A core innovation of this approach is the "civilized" reward function. Rather than simply maximizing its own throughput, the AI-STA seeks to maximize an Independent Fairness Metric based on Jain’s Fairness Coefficient. Since a single station lacks global network statistics, the AI-STA "eavesdrops" on Request to Send/Clear to Send (RTS/CTS) control packets to build a duration table of channel occupancy by other devices. The DQN agent is rewarded for fairness and penalized if its own throughput, latency, jitter, or packet loss rates fall below predefined QoS thresholds.

Experimental Setup

The performance of the AI-STA was evaluated using discrete event simulations in ns-3, integrated with ns3-ai. The simulation environment modeled a typical home Wi-Fi scenario (IEEE 802.11ax) featuring 5 Access Points and 14 Stations. To ensure realistic results, the traffic profile was reconstructed from real-time mobile gaming data, representing demanding, low-latency requirements.

Results and Discussion

The AI-STA was compared against a standard 802.11ax baseline and a traditional Dynamic Sensitivity Control (DSC) algorithm. Key findings include:

  • Throughput and Fairness: The AI-STA consistently outperformed both the baseline and DSC, reaching a throughput of nearly 4.8 Mbps (compared to the baseline's 4.5 Mbps) while maintaining a significantly higher fairness score.
  • Aggression Control: The Carrier Sensing Threshold (CST) served as the primary indicator of algorithmic aggression. As shown in the results, the AI-STA learned to be less aggressive than the DSC algorithm, staying closer to the baseline to preserve network-wide fairness.
Comparison of Carrier Sensing Threshold (CST) behavior over time for DQN, DSC, and baseline algorithms.
  • Convergence: The DQN agent successfully stabilized its CST, RST, and transmit power levels within approximately 20-30 seconds of simulated time, demonstrating effective learning of the environment.
  • Trade-offs: While throughput was improved and smoothed, the AI-STA exhibited slightly higher average delay and jitter compared to the baseline. This is attributed to the station waiting for optimal moments to transmit to maintain fairness.

Conclusion

The AI-driven Sensitivity and Power algorithm provides a robust framework for optimizing device behavior in legacy Wi-Fi networks. While the current model focuses on sensing and power, future work will explore integrating additional MAC layer parameters, such as contention window size and backoff time, to further optimize latency and jitter. Additionally, investigating the interactions between multiple AI-STAs remains a critical area for ensuring cooperative rather than competitive reinforcement learning in dense deployments.

Summary updated

Abstract

With the increasing complexity of Wi-Fi networks and the iterative evolution of 802.11 protocols, the Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) protocol faces significant challenges in achieving fair channel access and efficient resource allocation between legacy and modern Wi-Fi devices. To address these challenges, we propose an AI-driven Station (AI-STA) equipped with a Deep Q-Learning (DQN) module that dynamically adjusts its receive sensitivity threshold and transmit power. The AI-STA algorithm aims to maximize fairness in resource allocation while ensuring diverse Quality of Service (QoS) requirements are met. The performance of the AI-STA is evaluated through discrete event simulations in a Wi-Fi network, demonstrating that it outperforms traditional stations in fairness and QoS metrics. Although the AI-STA does not exhibit exceptionally superior performance, it holds significant potential for meeting QoS and fairness requirements with the inclusion of additional MAC parameters. The proposed AI-driven Sensitivity and Power algorithm offers a robust framework for optimizing sensitivity and power control in AI-STA devices within legacy Wi-Fi networks.
BibTeX
@misc{yi2024fair,
  author = {Peng Yi and Wenchi Cheng and Zhongjie Ju and Jingqing Wang and Jiahui Pan and Yuehui Ouyang and Wei Zhang},
  title = {{Fair AI-STA for Legacy Wi-Fi: Enhancing Sensing and Power Management with Deep Q-Learning}},
  year = {2024},
  eprint = {2412.10874},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2412.10874},
}
Figure