QoS-Aware State-Augmented Learnable Framework for 5G NR-U/Wi-Fi Coexistence: Impact of Parameter Selection and Enhanced Collision Resolution

Unlicensed spectrum supports diverse traffic with stringent Quality-of-Service (QoS) requirements. In NR-U/Wi-Fi coexistence,the values of MAC parameters critically influence delay, collision behavior, and airtime fairness and efficiency. In this paper, we investigate the impact of (i) cost scaling and violation modeling, (ii) choice of MAC parameters, and (iii) an enhanced collision resolution scheme for the Listen-Before-Talk (LBT) mechanism on the performance of a state-augmented constrained reinforcement learning controller for NR-U/Wi-Fi coexistence. Coexistence control is formulated as a constrained Markov decision process with an explicit delay constraint for high-priority traffic and fairness as the optimization goal. Our simulation results show three key findings: (1) signed, threshold-invariant cost scaling with temporal smoothing stabilizes learning and strengthens long-term constraint adherence; (2) use of the contention window parameter for control provides smoother adaptation and better delay compliance than other MAC parameters; and (3) enhanced LBT significantly reduces collisions and improves airtime efficiency. These findings provide practical insights for achieving robust, QoS-aware coexistence control.

💡 Research Summary

This paper presents a comprehensive analysis of a learnable framework for ensuring Quality-of-Service (QoS) in unlicensed spectrum bands where 5G New Radio in Unlicensed spectrum (NR-U) and Wi-Fi networks coexist. The core challenge lies in managing decentralized medium access, where MAC-layer parameters critically influence delay, collision probability, and airtime fairness. The study builds upon the QoS-aware State-Augmented Learnable (QaSAL) framework, which formulates coexistence control as a Constrained Markov Decision Process (CMDP). The primary goal is to maximize airtime fairness (using Jain’s Fairness Index) while strictly adhering to a delay constraint for high-priority traffic (PC1).

The main contribution of this work is a detailed investigation into three critical design choices that determine the practical performance of such a learning-based controller:

1. Impact of Cost Scaling and Violation Modeling: The paper identifies that raw constraint violation signals are noisy and lead to unstable learning. To address this, the authors propose a “signed, threshold-invariant scaling” technique. This method normalizes the violation signal relative to its threshold, applies a hyperbolic tangent (tanh) function for smooth scaling, and uses exponential moving averaging for temporal smoothing. Crucially, the agent’s cost function only observes the negative (violating) component to apply efficient penalties, while the dual variable update uses the full signed signal to dynamically adjust constraint pressure. This approach stabilizes training and strengthens long-term constraint adherence without complex reward engineering.

2. Choice of MAC Control Parameter: The research compares the effectiveness of controlling three key MAC parameters: Contention Window (CW), Arbitration Inter-Frame Spacing Number (AIFSN), and Maximum Channel Occupancy Time (MCOT). Simulation results demonstrate that using the CW as the control action yields superior performance. Controlling CW, which directly sets the backoff range and thus channel access aggressiveness, provides smoother adaptation and better compliance with the PC1 delay constraint compared to adjusting AIFSN (which controls deferment time) or MCOT (which controls transmission duration). This finding offers a practical guideline for parameter selection in real-world implementations.

3. Effect of an Enhanced Collision Resolution Scheme: The paper analyzes the integration of a Collision-Resolution LBT (CR-LBT) mechanism for NR-U. Traditional LBT transmits a Reservation Signal (RS) after backoff until the next slot boundary, which can lead to simultaneous-start collisions. CR-LBT replaces this long RS with a series of short “collision-resolution slots,” allowing a gNB to perform quick sensing and defer if it detects another transmission starting. This enhanced physical-layer mechanism is shown to significantly reduce the number of collisions and consequently improve overall airtime efficiency, creating a more favorable environment for the learning controller to meet its QoS targets.

The experimental setup involves a discrete-event simulator modeling saturated traffic conditions. The results validate that the proposed enhancements—proper cost scaling, selecting CW as the control knob, and employing CR-LBT—collectively enable the QaSAL framework to achieve robust, QoS-aware control. The paper concludes that successful coexistence management requires co-optimization across multiple layers: the learning algorithm design, the selection of the right control parameters, and the underlying physical/MAC-layer mechanisms for collision avoidance. These insights provide a valuable roadmap for designing adaptive and reliable controllers for next-generation unlicensed spectrum sharing.