東華大學圖書館 |

Machine Learning for Efficient & Robust Next-Generation Communication Systems /

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Machine Learning for Efficient & Robust Next-Generation Communication Systems // Noel Teku.
作者:	Teku, Noel,
面頁冊數:	1 electronic resource (161 pages)
附註:	Source: Dissertations Abstracts International, Volume: 86-12, Section: A.
Contained By:	Dissertations Abstracts International86-12A.
標題:	Electrical engineering. -
電子資源:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=32117618
ISBN:	9798286428809

Machine Learning for Efficient & Robust Next-Generation Communication Systems /
Teku, Noel,

Machine Learning for Efficient & Robust Next-Generation Communication Systems /Noel Teku. - 1 electronic resource (161 pages)

Source: Dissertations Abstracts International, Volume: 86-12, Section: A.

Machine learning (ML)-based techniques are increasingly being incorporated into next-generation wireless systems: both for improving fundamental building blocks (e.g., modulation classification, power allocation, channel decoding) as well as enabling new functionalities (e.g., AR/VR, autonomous vehicles). This dissertation makes the following contributions in these areas:As ML classifiers become integral to next-generation wireless systems, it is essential to ensure their predictions are delivered both reliably and with low delay-for instance, in applications like transmitting road condition assessments in vehicular networks or relaying critical health data from sensors to medical providers. In our first contribution, we analyze the fundamental information-theoretic tradeoffs between latency and end-to-end distortion when communicating the results of a classifier over a noisy communication system. We use techniques from finite blocklength channel capacity and show that lattice-based quantization of probability distributions leads to a significant reduction in latency compared to other baselines.In our second contribution, we present a new approach for using reinforcement learning (RL) to provide adaptive robustness to High Frequency (HF) channels. The HF band, which occupies the spectrum of 3 to 30 MHz, enables long-range communications by bouncing signals off the ionosphere with limited communication infrastructure. However, the turbulent nature of the channel, which causes frequent signal dropouts, has deterred the band from being used more heavily. To mitigate this challenge, we propose using RL to learn the optimal settings (e.g., tap length, step size, filter type, adaptive algorithm) of an adaptive equalizer and show that our techniques can provide better performance compared to adaptive equalizers with a fixed structure.In our third contribution, we devise an unsupervised learning-based framework to optimize cell-free networks (CFNs). CFNs deviate from the concept of having an access point (AP) be responsible for serving user equipment (UEs) within a fixed radius and instead deploy APs over a geographic region to collaboratively serve every UE. In doing so, CFNs increase the probability of coverage and achieve stronger diversity gains. To build on these improvements, we propose using an unsupervised neural network to learn how to split a UE's message across different APs in a manner that minimizes the total latency of the CFN. We show that our unsupervised technique is more effective in ensuring higher probabilities of lower latencies compared to decentralized baselines. Additionally, when noisy channel state information is assumed, our unsupervised technique is more robust in achieving a high likelihood of lower latencies compared to centralized baselines.In our final contribution, we investigate a complementary problem of ensuring privacy when aligning Large Language Models (LLMs). LLMs have been investigated for various applications, due to their broad knowledge base attained via pre-training on large corpora of data. However, it has been shown that LLMs can generate socially unacceptable responses. Alignment procedures have been proposed to train LLMs, using preference data collected from humans, to reinforce which types of responses are socially acceptable. While such methods are effective in regulating an LLM's responses, this type of training could be susceptible to leaking privacy-sensitive information of the human labelers. To mitigate this, we study the problem of LLM alignment with labeler privacy while maintaining the utility of the alignment process. To accomplish this, we present a novel privacy-preserving approach, namely PROPS (PROgressively Private Self-Alignment), a multi-stage algorithm capable of ensuring preference privacy without causing a significant drop in the utility of an LLM as it undergoes alignment.

English

ISBN: 9798286428809Subjects--Topical Terms:

649834
Electrical engineering.
Subjects--Index Terms:

Autonomous vehicles

Machine Learning for Efficient & Robust Next-Generation Communication Systems /
LDR:05331nmm a22004333i 4500 001 2435885
005 20251029060413.5
006 m o d
007 cr|nu||||||||
008 261011s2025 miu||||||m |||||||eng d
020 $a 9798286428809
035 $a (MiAaPQD)AAI32117618
035 $a AAI32117618
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Teku, Noel, $e author. $3 3826359
245 1 0 $a Machine Learning for Efficient & Robust Next-Generation Communication Systems / $c Noel Teku.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2025
300 $a 1 electronic resource (161 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 86-12, Section: A.
500 $a Advisors: Tandon, Ravi Committee members: Lazos, Loukas; Djordjevic, Ivan.
502 $b Ph.D. $c The University of Arizona $d 2025.
520 $a Machine learning (ML)-based techniques are increasingly being incorporated into next-generation wireless systems: both for improving fundamental building blocks (e.g., modulation classification, power allocation, channel decoding) as well as enabling new functionalities (e.g., AR/VR, autonomous vehicles). This dissertation makes the following contributions in these areas:As ML classifiers become integral to next-generation wireless systems, it is essential to ensure their predictions are delivered both reliably and with low delay-for instance, in applications like transmitting road condition assessments in vehicular networks or relaying critical health data from sensors to medical providers. In our first contribution, we analyze the fundamental information-theoretic tradeoffs between latency and end-to-end distortion when communicating the results of a classifier over a noisy communication system. We use techniques from finite blocklength channel capacity and show that lattice-based quantization of probability distributions leads to a significant reduction in latency compared to other baselines.In our second contribution, we present a new approach for using reinforcement learning (RL) to provide adaptive robustness to High Frequency (HF) channels. The HF band, which occupies the spectrum of 3 to 30 MHz, enables long-range communications by bouncing signals off the ionosphere with limited communication infrastructure. However, the turbulent nature of the channel, which causes frequent signal dropouts, has deterred the band from being used more heavily. To mitigate this challenge, we propose using RL to learn the optimal settings (e.g., tap length, step size, filter type, adaptive algorithm) of an adaptive equalizer and show that our techniques can provide better performance compared to adaptive equalizers with a fixed structure.In our third contribution, we devise an unsupervised learning-based framework to optimize cell-free networks (CFNs). CFNs deviate from the concept of having an access point (AP) be responsible for serving user equipment (UEs) within a fixed radius and instead deploy APs over a geographic region to collaboratively serve every UE. In doing so, CFNs increase the probability of coverage and achieve stronger diversity gains. To build on these improvements, we propose using an unsupervised neural network to learn how to split a UE's message across different APs in a manner that minimizes the total latency of the CFN. We show that our unsupervised technique is more effective in ensuring higher probabilities of lower latencies compared to decentralized baselines. Additionally, when noisy channel state information is assumed, our unsupervised technique is more robust in achieving a high likelihood of lower latencies compared to centralized baselines.In our final contribution, we investigate a complementary problem of ensuring privacy when aligning Large Language Models (LLMs). LLMs have been investigated for various applications, due to their broad knowledge base attained via pre-training on large corpora of data. However, it has been shown that LLMs can generate socially unacceptable responses. Alignment procedures have been proposed to train LLMs, using preference data collected from humans, to reinforce which types of responses are socially acceptable. While such methods are effective in regulating an LLM's responses, this type of training could be susceptible to leaking privacy-sensitive information of the human labelers. To mitigate this, we study the problem of LLM alignment with labeler privacy while maintaining the utility of the alignment process. To accomplish this, we present a novel privacy-preserving approach, namely PROPS (PROgressively Private Self-Alignment), a multi-stage algorithm capable of ensuring preference privacy without causing a significant drop in the utility of an LLM as it undergoes alignment.
546 $a English
590 $a School code: 0009
650 4 $a Electrical engineering. $3 649834
650 4 $a Computer engineering. $3 621879
650 4 $a Technical communication. $3 3172863
653 $a Autonomous vehicles
653 $a Machine learning
653 $a Wireless systems
653 $a High Frequency
653 $a Cell-free networks
690 $a 0544
690 $a 0643
690 $a 0464
710 2 $a The University of Arizona. $b Electrical & Computer Engineering. $e degree granting institution. $3 3815673
720 1 $a Tandon, Ravi $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 86-12A.
790 $a 0009
791 $a Ph.D.
792 $a 2025
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=32117618