Electrical and Electronic Engineering - Theses
Permanent URI for this collection
Search Results
1 - 3 of 3
-
ItemDistributed Failure-Tolerant Anomaly Detection in Cognitive Radio Networks( 2023-04)The communications landscape has seen exciting developments through the emergence of small, low-cost, wireless devices. Developments in these devices have led to unprecedented connectivity and distributed computational resources—ready to support new applications. Such applications provide new benefits to end users (through cognitive radio and Internet-of-Things, IoT, to name a few), as well as new attack vectors for malicious users—with a higher number of exposed devices and communications. In this work, we investigate the use of these new wireless networking devices to make wireless communication and networking more secure by analysing wireless activity throughout a network and training anomaly detection models to identify any unusual behaviour. Using their flexible communications, onboard computation, and ability to record wireless network data, we explore state-of-the-art methods to learn patterns in network behaviour using distributed sensing and computational resources. These methods span classical and modern anomaly detection approaches, each with its own benefits and drawbacks in terms of performance, resource usage, and reliability. Throughout this work, the tradeoff between these benefits and drawbacks is outlined and new collaborative anomaly detection methods are proposed. The methods and tools in this thesis have been analysed in various network environments, to strengthen present and future wireless networks.
-
ItemInformation-theoretic Analysis For Machine Learning and Transfer Learning: Bounds and Applications( 2023-03)Traditional machine learning is characterized by the assumptions that the training data and target data are drawn from the same distributions. However, in practice, obtaining these data may be expensive and difficult. Transfer learning, or domain adaptation, is concerned with machine learning problems in which training and testing data come from possibly different distributions. The domain adaptation problems are widely investigated and used to improve the predictive results for one certain domain by transferring useful information from another (possibly) related domain where it is easy and cheap to obtain the data. Therefore, developing high-performance transfer learning techniques is necessary. One may ask how do we guarantee that the transfer learning is useful and efficient? In this thesis, we investigate the learning performance of the transfer learning algorithms from an information-theoretic perspective, where one broad line of work considers the learning setting where in the training phase we only have access to labelled data from the source distribution mu, possibly with some additional unlabelled or labelled data from the target distribution mu' that we are interested in the testing phase. A popular approach in this context is to formulate a measure of discrepancy between the distributions mu and mu' and to give test error bounds in terms of this discrepancy. In this sense, we are particularly interested in the generalization error, which is defined as the difference between the empirical training loss and the population loss under mu' for a given algorithm, and this quantity indicates if the output hypothesis of the algorithm has been overfitted (or underfitted). This quantity can be viewed as the distribution (over both data and algorithm) divergences between the training and testing phases. From this perspective, the information-theoretic approach will benefit from different perspectives. In this thesis, we first give a review of information-theoretic analysis for generalization error in traditional machine learning problems with identical training and testing data distributions. We then propose a fast generalization framework that enhances learning performance by identifying the key conditions and improving the learning rate, where the improvement shifts the typical information-theoretic bounds from sublinear convergence to linear convergence. Next, we extend this analysis to transfer learning under various learning settings, viewed from different perspectives. Initially, we use the variational representation of KL divergence to derive upper bounds for general transfer learning algorithms under the batch learning setting. These data-algorithm-dependent bounds offer valuable insights into the impact of domain divergence on generalization ability. We then extend the batch learning setting to the online learning setting, viewed from a Bayesian perspective, and consider transfer learning under the supervised learning setting. We view prediction from a causal perspective using the proposed potential outcome framework and derive corresponding excess risks under different distribution shifting scenarios. These bounds are useful in orienting general transfer learning problems and identifying whether transfer learning is practical. To demonstrate the practical applications of our theoretical results, we propose bound-based algorithms and show their versatility in real-world problems.
-
ItemRisk Management Frameworks and Methodologies for Modern and Resilient Power Systems Planning Using Machine Learning Techniques( 2020)Renewable energy technologies, customer behaviour, and new regulations are key factors contributing to a change in power generation paradigm that is becoming increasingly decentralized and embedded in the distribution network. The new paradigm, together with strong opportunities, is bringing challenges for power networks that must be adequately anticipated and planned to maintain the security and reliability of the power supply. This research addresses two key challenges for developed power networks and one challenge for developing networks located in countries vulnerable to extreme weather events. For developed power networks, this research formulated risk assessment models based on Artificial Intelligence techniques that enable power system planners to analyse vast numbers of scenarios and assess the impact of voltage excursions and reverse power flows as a result of elevated penetration of distributed energy resources. The novelty of the work is derived from the scalability of the proposed models and its end-to-end approach that includes financial modelling of the impacts. For the developing network, this research developed one risk-based methodology to assess resilience to extreme weather events that is linked to power system planning. The novelty of the proposed methodology is derived from the problem formulation that explicitly considers both the technical power system resilience and the social community energy resilience in quantifiable terms that are linked to power system planning via an optimization problem.