Abstract: The exponential growth of wireless devices and stringent reliability requirements of emerging applications demand fundamental improvements in distributed channel access mechanisms for unlicensed bands. Current Wi-Fi systems, which rely on binary exponential backoff (BEB), suffer from suboptimal collision resolution in dense deployments and persistent fairness challenges due to inherent randomness. This paper introduces a multi-agent reinforcement learning framework that integrates artificial intelligence (AI) optimization with legacy device coexistence. We first develop a dynamic backoff selection mechanism that adapts to real-time channel conditions through access deferral events while maintaining full compatibility with conventional CSMA/CA operations. Second, we introduce a fairness quantification metric aligned with enhanced distributed channel access (EDCA) principles to ensure equitable medium access opportunities. Finally, we propose a centralized training decentralized execution (CTDE) architecture incorporating neighborhood activity patterns as observational inputs, optimized via constrained multi-agent proximal policy optimization (MAPPO) to jointly minimize collisions and guarantee fairness. Experimental results demonstrate that our solution significantly reduces collision probability compared to conventional BEB while preserving backward compatibility with commercial Wi-Fi devices. The proposed fairness metric effectively eliminates starvation risks in heterogeneous scenarios.
Introduction
Modern Wi-Fi relies on a mechanism called Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA). When a device wants to transmit, it selects a random "backoff" number and waits. If it detects another device transmitting, it pauses its countdown—a process known as Binary Exponential Backoff (BEB). However, BEB is fundamentally "blind" to real-time congestion patterns, leading to high collision rates in dense environments like stadiums or apartment buildings.
This paper introduces FC-MAPPO, a multi-agent reinforcement learning framework that optimizes Wi-Fi backoff timers. By making decisions during channel "deferral events" rather than at the microsecond level, the system achieves a 6% reduction in collisions and a 10% throughput boost while remaining fully backward-compatible with legacy Wi-Fi devices.
The FC-MAPPO Framework
The researchers propose Fairness-Constrained Multi-Agent Proximal Policy Optimization (FC-MAPPO) to solve two primary challenges: the extreme speed of Wi-Fi operations and the need for fair coexistence with legacy hardware.
Defer Access-Based Backoff
Wi-Fi slots are only 9 microseconds long, which is too fast for most AI models to process in real-time. To overcome this, the AI re-evaluates its strategy only after "access deferral events"—periods when the channel is busy because a neighbor is transmitting. This reinterprets the waiting period as a "virtual slot," operating on a millisecond timescale that is computationally feasible for modern hardware.
Fairness as a Mathematical Constraint
To prevent AI agents from "cheating" by always picking the smallest backoff to hog the channel, the authors defined a Backoff Ratio ($B/CW$). Under ideal random conditions, this ratio should average 0.5. The AI is constrained via Reward-Constrained Policy Optimization (RCPO) to maintain an average ratio $\ge 0.5$, ensuring it does not starve legacy devices of airtime.
Architecture and Training
The model utilizes a Centralized Training, Decentralized Execution (CTDE) architecture. Agents are trained together in an ns-3 based simulator where they can observe the global network state. Once deployed, each Wi-Fi station makes autonomous decisions based only on local observations, such as the delay since the last successful transmission.
The underlying architecture uses a permutation-invariant transformer. This allows the AI to observe an arbitrary number of neighbors and their activity patterns without needing to identify specific devices, enhancing scalability in dynamic environments.
Key Results and Discussion
The study tested the framework in saturated uplink traffic scenarios with 4 to 10 stations. The results highlight significant improvements over standard 802.11ax (Wi-Fi 6) protocols:
- Collision Reduction: In dense 10-station scenarios, standard Wi-Fi hit a collision rate of nearly 40%, while FC-MAPPO kept it closer to 34%.
- Throughput Gains: RL-enhanced stations saw a 10% throughput increase. Interestingly, the system is "altruistic"; by reducing overall channel noise, even legacy devices saw performance improvements.
- Fairness Convergence: Training data confirms that the agents successfully learn to maximize rewards while adhering to the fairness constraint.
As shown in the training curves, the reward per transmission stabilizes as the agents learn optimal strategies. Simultaneously, the backoff ratio converges toward the 0.5 target, validating that the AI provides equitable access rather than simply out-competing other devices.
Conclusion
FC-MAPPO demonstrates that AI can fundamentally improve distributed channel access without sacrificing backward compatibility. By aligning AI decision-making with natural channel deferral events, the framework bypasses hardware latency limitations. Future work will focus on adapting these strategies for bursty, non-saturated traffic patterns and larger-scale deployments.
Summary updated