Online Learning Based Successive Interference Cancellation for Fair Coexistence of Heterogeneous Systems Over the Unlicensed Spectrum

Abstract

Due to the scarcity of the licensed spectrum and the continuously increasing demand forwireless capacity in communication and networking systems, heterogeneous wireless technologies have begun operating in unlicensed frequency bands. However, existing collision avoidance-based MAC protocols for unlicensed bands, including Listen-Before-Talk (LBT) and Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA), suffer from low spectrum utilization efficiency. Achieving harmonious and effective coexistence among heterogeneous networks presents several critical challenges. First, while enabling concurrent transmissions—combined with the use of interference cancellation (IC) techniques—can significantly improve spectrum efficiency, IC operates at the Physical layer, whereas managing concurrent transmissions and optimizing channel access strategies must be handled at the MAC layer. This is particularly difficult in heterogeneous networks, where devices typically cannot directly coordinate with one another. Second, although offline training can be used to estimate channel and link quality for each set of concurrent transmissions, it incurs substantial overhead and delay, and is poorly suited to dynamic wireless environment characterized by fading, user mobility, or frequent topology changes (e.g., links continuously joining or leaving the network). Third, resource allocation and scheduling in shared spectrum scenarios often face strict latency constraints dictated by Ultra-Reliable Low-Latency Communication (URLLC) applications in 5G systems. Designing online learning algorithms that are both efficient and capable of meeting these stringent timing requirements remains a significant challenge. Finally, conventional IC techniques are often inadequate in heterogeneous network coexistence scenarios, where concurrent transmissions are uncoordinated and unpredictable. These limitations hinder the ability of these techniques to simultaneously meet high reliability and low latency requirements. In this dissertation, we aim to address the aforementioned challenges and enhance theperformance (spectrum efficiency, throughput, reliability, and latency) of heterogeneous network coexistence. First, we propose allowing concurrent transmissions of heterogeneous links to take place (instead of avoiding them) by adopting IC techniques that alleviate interference within the same or across different wireless technologies. The concurrent transmissions are achieved via adjusting energy detection (Clear Channel Assessment) threshold, which remains compatible with existing Listen-Before-Talk (LBT) and CSMA/CA protocols. To demonstrate the feasibility and benefits of cross-technology IC, we implement a prototype successive interference cancellation (SIC) receiver for LTE/Wi-Fi coexistence on USRP devices. Second, we propose an IC-aware MAC protocol that enables concurrent transmissions and optimizes the channel access strategy at the MAC layer, so as to mitigate the interference of coexisting technologies and improve spectrum efficiency and fairness. For the MAC protocol design, we introduce a training phase to discover the topology and learn the channel and link quality information of each concurrent transmission set without direct message exchanges between the heterogeneous technologies. In the transmission phase, all links make optimized channel access decisions in a distributed manner. Because the training incurs significant overhead when there is a large number of links, we propose multi-armed bandit (MAB) based online learning algorithms to improve the overall throughput and link-level fairness. Traditional MAB approaches are insufficient for our problem, as the reward of an individual arm (i.e., a link) depends on the specific meta-arm (i.e., the set of concurrently transmitting links, or CTS) selected by the learning agent. Therefore, we introduce a new formulation, called fair probabilistic MAB (FP-MAB). To solve the FP-MAB problem, we propose two algorithms: Fair Probabilistic Explore-Then-Commit (FP-ETC) and Fair Probabilistic Optimism in the Face of Uncertainty (FP-OFU). Both algorithms probabilistically play an arm over time to meet the fairness requirements. Furthermore, we enhance the efficiency of the FP-MAB algorithms to support delay-sensitive resource allocation and scheduling applications in spectrum sharing scenarios. To this end, we propose the Efficient Fair Probabilistic Multi-Armed Bandit (EFP-MAB) algorithm, which significantly reduces the computational complexity. Unlike FP-OFU, which constructs elliptical confidence intervals, EFP-MAB adopts rectangular confidence intervals for each meta-arm. Finally, we investigate real-world implementation of SIC for heterogeneous network coexistence. Specifically, we introduce DL-SIC, a deep learning-based framework that significantly advances the capabilities of traditional SIC receiver in heterogeneous shared-spectrum environment. Unlike traditional SIC receivers that rely on received signal strength and often suffer from high decoding latency or high bit error rate (BER), DL-SIC leverages deep neural networks to accurately identify protocol types and determine optimal decoding orders in real time, even under unknown or dynamic power conditions.