A learning framework for optimized control of wireless links

Abstract

The proliferation of wireless communication systems has garnered significant attention, driven by the exponential growth of interconnected nodes and the emergence of applications with diverse Quality-of-Service (QoS) requirements. These supported use cases exhibit a wide spectrum of demanding QoS requirements. In particular, fifth-generation (5G) communication systems have been architected to concurrently support three different use cases: Ultra-Reliable Low-Latency Communications (URLLC), massive Machine-Type Communications (mMTC), and enhanced Mobile BroadBand (eMBB), each with its unique application-specific requirements, all operating within shared network resources. Satisfying these requirements mandates operating the network in an optimized fashion, a formidable challenge given the inherent dynamism of channel conditions. At the heart of enhancing the operational efficiency of modern communication systems lies the optimization of transmission parameters (e.g., modulation and coding schemes, guard intervals, and more). However, the dynamic adaptation of transmission parameters is contingent upon channel knowledge, which is acquired through a feedback process. This feedback mechanism is sensitive to the estimation and compensation of carrier frequency offsets. Hence, this thesis focuses on three pivotal components within the communication pipeline: Carrier Frequency Offset (CFO) estimation, Channel State Information (CSI) feedback compression, and link adaptation. The intrinsic complexities associated with analytically modeling these challenges, coupled with the accessibility of abundant datasets and the extraordinary efficacy demonstrated by artificial intelligence (AI) and machine learning (ML) algorithms, have catalyzed the integration of AI and ML methodologies in addressing these issues. Moreover, the incorporation of AI and ML techniques holds the promise of significantly reducing execution times by circumventing the conventional iterative algorithms traditionally employed in such endeavors. This dissertation proposes several novel ML-based solutions for the aforementioned three challenging problems within a comprehensive framework. The proposed solutions improve the accuracy, reliability, and efficiency of communication systems. This dissertation is structured into three parts, each dedicated to addressing one of the three core problems investigated. Precisely, PART 1 is focused on tackling the CFO estimation problem, PART 2 delves into the CSI feedback compression problem, and PART 3 is devoted to the Link Adaptation problem. Our contributions to PART 1 (the CFO estimation) can be summarized as follows: - Introducing ensemble learning, specifically the Gradient Boosting Machine (GBM) algorithm, within the CFO estimation problem. Adopting the GBM enhances the generalization capabilities of our proposed solution while aligning it with the inherent resource constraints encountered at the User Equipment (UE). - Presenting the BiModule CFO Estimation (BMCE) technique, which represents an innovative approach to combining the predictions generated by the estimation module with the outcomes derived from an auxiliary module designed to model the temporal correlation among CFO values. Combining the direct CFO estimation derived from synchronization preambles with CFO forecasting results in a notable enhancement of prediction accuracy, yielding a 16% improvement over sole reliance on direct estimation methods. Our contributions to PART 2 (the CSI feedback compression) can be summarized as follows: - Introducing the Variational Autoencoders (VAE) to tackle the problem of noisy feedback channels. It has been empirically demonstrated that VAEs outperform conventional point estimation autoencoders in terms of reconstruction accuracy, providing a more effective means of managing the impact of noisy channels. - Proposing a customized version of the VAE loss in order to further optimize the performance of VAE in the context of feedback problems. This customized loss function is precisely designed to align with the specific requirements and characteristics of feedbackrelated challenges (noisy feedback channels), thereby contributing to more accurate and meaningful reconstructions. - Proposing an alternative solution rooted in learning theory to address the recognized limitations of autoencoder-based approaches. This novel approach not only addresses the limitations of conventional autoencoders but also outperforms both traditional and learning-based solutions by a substantial margin in terms of reconstruction accuracy. Finally, in PART 3 (link adaptation), our contributions can be summarized as follows: - Proposing a novel modeling for the link adaptation problem as a multilabel multiclass classification that provides a new framework for tackling this intricate issue. - Proposing a customized loss function to train the classification models while increasing the system reliability (by minimizing the false positive errors). - Proposing a comprehensive subsampling criterion for training link adaptation models instead of the random sampling adopted in the literature. The utilization of this novel subsampling criterion for training models on limited datasets yields significant improvements, with performance enhancements of up to 50% observed in specific scenarios. - Developing a sophisticated neural architecture for solving the joint compression-adaptation problem. This architecture is accompanied by a customized loss function and training procedure, collectively representing a comprehensive solution framework for efficiently managing the complexities of joint compression and adaptation tasks.

Date	16 Oct 2023
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Mohamed Cheriet (Supervisor) & Kim Khoa Nguyen (Co-supervisor)