Search

Advances in machine learning techniques have been shown to bring benefit for analysing medical images. Whole-body scans contain multiple organs, which makes the manual annotation harder, hence the amount of data obtained is still limited. Meanwhile, machine learning needs a large amount of annotated data in order to perform well. Image segmentation is a crucial task in medical image analysis and beneficial for clinical diagnosis. One of the segmentation challenges is to infer the automatic segmentation quality between algorithms. Unfortunately, the manual annotations required to compute the segmentation accuracy are sometimes unavailable. In order to tackle this, we propose a new method to predict the segmentation evaluation on a per-case basis without annotated data. Segmenting scans acquired from different imaging centers than the one used for algorithm training, can lead to degraded segmentation results. This problem has brought the necessity of domain adaptation. We studied how to select useful samples from unlabeled target domain for supervised domain adaptation. This framework is shown to be less time-consuming with promising results to be useful in clinical practice. A novel dual-stream network is also introduced to learn useful features on unpaired images from different modalities. This network demonstrates its power in multi-modal learning to segment multiple abdominal organs on both modalities. Additionally, strategies to segment challenging small organs on whole-body scans are investigated. We show that using two-stage networks with weighting schemes is useful to solve class imbalance and multi-scale contextual information. Machine learning is not only helping the work of radiologists with faster automatic segmentation tools, but also leveraging the useful features for alleviating real clinical problems. We believe that our proposed methods could offer an important contribution to the field of medical image analysis; advancing our understanding on how machine-learning approaches could be applied in several problems in whole-body image analysis.

This thesis addresses two challenging acquisition problems for computer graphics: acquisition of layered reflectance, and acquiring common transparent objects (axially-symmetric containers) and liquids. The first problem is approached using a novel computational photography method for single shot separation of diffuse/specular reflectance as well as novel angular domain separation of layered reflectance. A polarized light field camera is presented as a key component for simultaneously capturing multiple states of polarization and this thesis demonstrates single-shot reflectance separation under various types of incident illumination. To tackle the second problem, a practical solution is proposed for high quality reconstruction of axially-symmetric transparent objects and estimation of the optical properties of various common liquids contained in such objects. Acquisition setup of the solution involves imaging such objects from a single viewpoint while illuminating them from directly behind with polarized patterns. A reconstruction step is then based on optimization of the object's geometry and its refractive index to minimize the difference between observed and simulated transmission and refraction of rays passing through the object. The same setup is also employed for developing a novel method for estimating optical properties of common liquids contained in natural concentrations in a known axially-symmetric transparent container.

The first part of this thesis discusses the application of artificial intelligence to stock price prediction and the building of another artificial intelligence model for stock price prediction during mergers and acquisitions. It is demonstrated that a combination of Long Short-Term Memory/Recurrent Neural Network + Convolutional Neural Network + Attention (LSTM/RNN+CNN+Attention) trained on daily and minute stock price and volume data are able to predict future stock prices and stock price changes three times better than traditional non-deep learning methods. Further research has been done on stock price prediction during Mergers and Acquisitions for technology unicorns, finding that when more information on M&A has been added into the input dataset, the accuracy can be improved by about 2%. It is the first time that a deep learning algorithm has been applied to modelling volume and price/price change, and to M&A circumstances. In the second part of this thesis, a Support Vector Machine (SVM) method is proposed, which enables adverting agencies to assess the effectiveness of their advertisement rapidly, quantitatively and without bias. This is crucial work for business/equity trading because when the product/service of a company sells well, then the corresponding company stock prices rise. The success of an advertising agency in their ability to produce compelling advertisement is therefore related to stock price prediction. This research explores models that outperform existing research; even the basic machine learning model improves the accuracy of a prediction to purchase from 70% up to 80%. The last part of this thesis incorporates a deep learning method to encrypt the data going between input nodes and the rest of the network. This method has the advantage of simultaneously preventing sensitive data being intercepted during the transmission from the network edge over insecure transports, and also anonymize incoming data in a way that could help satisfy privacy laws such as General Data Protection Regulation (GDPR). This research has provided a new data privacy and security structure that is more cost efficient than existed deep learning model done solely locally or remotely.

Continuous-time recurrent neural networks (CTRNNs) are both plausible models of cortical circuits and practically relevant for biologically inspired machine learning. To date, their dynamics are not well understood, especially when they are driven by external signals, and are hence typically treated as a "black box". This work builds on previous attempts to better understand the complex dynamics of CTRNNs. A series of computational experiments are presented whose aim is to reveal, through both visualisation and analysis, the dynamical transitions exhibited by these networks when driven by external periodic signals. On the applications front, the suitability of CTRNNs for episodic-like memory is explored, using the reservoir computing paradigm. This paradigm has emerged as a simple approach to overcoming important limitations of the gradient-based methods conventionally used to train recurrent neural networks. To explore their applicability to episodic-like memory, CTRNNs are trained to memorise and reconstruct sequences of high-dimensional images, representing episodic experiences. The various parameters affecting network performance are investigated, revealing a strong dependence on the dynamic regime the networks operate in. Finally, a simple computational experiment is presented using CTRNNs with modular, small-world connectivity. The aim is to visualise information broadcast in complex networks, as well as to provide computational support for the plausibility of the "connective core hypothesis", which posits a particular relationship between connectivity and information flow in the brain. The experiments and findings presented in this work offer a deeper and more intuitive understanding of the dynamics of continuous-time recurrent neural networks, and open up new avenues for future research in both their theory and application.

The development of microrobots is motivated by their potential to enable therapies and procedures that are currently not possible. Miniaturising conventional robot designs to micro-scales, however, is not feasible due to scaling laws and the difficulties in fabrication and assembly. MEMS (Microelectromechanical System) fabrication techniques are the gold standard for devices such as microgrippers, although they are rarely intended for in vivo applications due to constraints on their form-factor. A common embodiment of medical microrobots in the literature are untethered swimming- or crawling-type microrobots that are designed to navigate through the body or for use in microfluidic devices. Alternative tethered microsurgical tools, beyond MEMS, are underrepresented in the literature, while untethered microrobots have experienced a great progression over the last decade. One technology that has aided in many of the advances in microrobotics is two-photon polymerisation (2PP); a 3D printing technology with sub-micron resolution. 2PP is a highly versatile fabrication technique that has already proven itself to be adaptable to a wide range of applications including optics, metamaterials and microrobots. This thesis explores how 2PP could be used to create novel tethered microrobotic tools that are fabricated on fibres (both standard silica-based fibres and more advanced multimaterial fibres). Multichannel fibres (capable of delivering light, electric current, gas, liquids etc.) are a promising platform upon which to fabricate and interface with 2PP-printed end-effectors in order to incorporate both actuation and sensing capabilities. A fabrication protocol for 2PP-printed end-effectors on fibres was developed. Passive-compliant end-effectors were used for both manual and automatic manipulation of microobjects in a closed-chain configuration. Real-time force sensing was incorporated, first using visually-inferred information and then, improving upon this, using an integrated interferometric force-sensing element. A 3-fingered passive-compliant gripper, on the tip of an optical fibre, with this force sensor was fabricated, validated and used in manipulation tasks. Finally, two active-compliant pneumatically-actuated mechanisms on a hollow capillary fibre were demonstrated: a piston with continuous motion and a bistable gripper. The two approaches were characterised, and the bistable gripper was demonstrated to grasp and hold a microbject in its closed state (without holding power) before release.

Parallelism is everywhere, with co-processors such as Graphics Processing Units (GPUs) accelerating the performance of applications such as training deep-learning neural networks, climate forecasting, bitcoin mining, medical imaging, or data analytics on platforms ranging from desktop computers to cloud computing and high performance clusters to mobile phones. Code optimisations enable realising the available performance of such devices, and automating these optimisations enables performance portability of software between different architectures. In this thesis, we consider two code optimisations that can improve application performance by reducing the degree of hardware and software parallelism in a program execution: thread coarsening, which by merging threads reduces the number of threads launched, and artificial occupancy reduction, which limits the number of threads simultaneously processed by allocating superfluous resources. We show how occupancy prediction through re-compilation can enable the selection of near-optimal coarsening factors at compile-time, by which thread coarsening can be applied in a fully automated manner without requiring auto-tuning. We demonstrate that our approach can achieve a maximum speedup of 5.08x (1.30x average) across three different NVidia GPU architectures, two modes of coarsening, different problem sizes, and for code pre-optimised to different degrees. When trying to predict the likely effects of thread coarsening, it is important to consider the effects it might have on cache pressure. We describe how a fast static analysis based on partial symbolic execution can be implemented to identify cache line re-use in programs. We demonstrate how this heuristic approach can improve on the runtime and memory requirements of a more extensive re-use distance analysis by several orders of magnitude, causing it to be sufficiently light-weight for run-time execution. We show that the analysis is able to identify kernels that are likely to experience an increase in cache pressure after coarsening. We explore the interaction of thread coarsening and artificial occupancy reduction, which can have negative effects on cache pressure and processor workload, respectively. We show that the two optimisation techniques can cancel these out when applied in combination, and yield a performance improvement of 8% in some cases. We investigate whether the cache line re-use analysis can identify candidates for artificial occupancy reduction.

Medical imaging is a cornerstone of modern healthcare. The ability to acquire images from inside a patient has revolutionised the way doctors diagnose and treat diseases, with almost all clinical pipelines now involving imaging to some degree. The development of these imaging methods has led to the field of medical image computing, where a multitude of tools and techniques have been proposed to aid clinicians and researchers in interpreting and analysing these images. One such family of techniques involves generating synthetic medical images. Image synthesis techniques are wide and varied, ranging from basic phantoms, to disease atlases, to high resolution photo-realistic subject-specific images. Their applications are similarly diverse: for developing novel acquisition protocols, training and testing algorithms, visualising changes in disease, predicting particular image types from others, and improving image quality and resolution. This thesis examines the use of medical image synthesis with a particular focus on applications in neurodegenerative diseases. A method for synthesising subject-specific non-pathological images from pathological images is first proposed and used for the unsupervised brain lesion segmentation. We next show how generative adversarial networks (GANs) can be used to both analyse the structural changes seen in patients with Alzheimer's disease, and to add or remove these changes from patient images to produce a subject-specific prediction of disease progression. Finally we investigate how GAN-derived synthetic data can be used to increase the size of training datasets, and under what conditions this additional data can lead to an improvement across a variety of segmentation tasks. Within this context we explore two situations: where only a small amount of labelled data is available, and where a large amount of unlabelled data is also available. We show that the proposed methods can lead to significant improvements in segmentation results, especially when a small amount of labelled data is available.

Computer graphics have evolved at a very fast pace over the last forty years. Most of the research in rendering has been focused on recreating visual effects that can be explained with geometric optics, such as reflections from diffuse and specular surfaces, refraction and volumetric scattering. Although geometric optics cover a wide range of effects related to light transport, some very impressive and colourful effects can only be explained and rendered with wave optics. This is the case of diffraction of light. Diffraction is a very common effect that causes dispersion of light i.e., the decomposition of white light into colourful patterns on a surface. It is caused by interferences between light waves when the geometry of a surface reaches a size below the coherence length of white light (around 65 micrometers). The most famous example of a diffractive surface is probably a Compact Disc on which the bits of information are stored along tracks that are small enough to diffract light. In this thesis, we present novel approaches to generate photorealistic computer renderings of diffraction of light from measurements of real-world surfaces. We present four practical measurement setups that employ commonly found hardware to acquire reflectance properties of both spatially-homogeneous diffractive surfaces and spatially-varying printed holographic surfaces. We also describe how such measurements can be employed in conjunction with a physically-based rendering model of diffraction to avoid Fourier optics simulations and therefore reduce the computational expense of diffraction rendering. Finally, we present techniques to render diffraction effects under arbitrary illumination at real-time framerates which is computationally very expensive with conventional techniques. These contributions constitute the first demonstration of realistic renderings of complex diffraction patterns observed in manufactured materials using practical measurement techniques at the interface of photography and optics. The algorithms presented in this thesis can be implemented in real-time applications such as video games and virtual reality experiences.

Continuous technological development in modern societies has increased the quality of life and average life-span of people. This imposes an extra burden on the current healthcare infrastructure, which also creates the opportunity for developing new, autonomous, assistive robots to help alleviate this extra workload. The research question explored the extent to which a prototypical robotic platform can be created and how it may be implemented in a hospital environment with the aim to assist the hospital staff with daily tasks, such as guiding patients and visitors, following patients to ensure safety, and making deliveries to and from rooms and workstations. In terms of major contributions, this thesis outlines five domains of the development of an actual robotic assistant prototype. Firstly, a comprehensive schematic design is presented in which mechanical, electrical, motor control and kinematics solutions have been examined in detail. Next, a new method has been proposed for assessing the intrinsic properties of different flooring-types using machine learning to classify mechanical vibrations. Thirdly, the technical challenge of enabling the robot to simultaneously map and localise itself in a dynamic environment has been addressed, whereby leg detection is introduced to ensure that, whilst mapping, the robot is able to distinguish between people and the background. The fourth contribution is geometric collision prediction into stabilised dynamic navigation methods, thus optimising the navigation ability to update real-time path planning in a dynamic environment. Lastly, the problem of detecting gaze at long distances has been addressed by means of a new eye-tracking hardware solution which combines infra-red eye tracking and depth sensing. The research serves both to provide a template for the development of comprehensive mobile assistive-robot solutions, and to address some of the inherent challenges currently present in introducing autonomous assistive robots in hospital environments.

Digital Education Games (DEGs) have been used to support children's learning in various domains. A number of existing studies on DEGs has focused on whether they could improve children's learning performance. However, only a few of them have attempted to address the critical question of how young children interact with DEGs. Bridging this gap is the main motivation for my research work that contributed to the applied body of knowledge in Human-Computer Interaction, Game-Based Learning, and Educational Technology. With the use of state-of-the-art eye-tracking technology, especially its applicability for mobile devices, a better understanding of young children's focused attention and interaction strategies when learning with DEGs can be obtained. Methodologically, my research work demonstrates how eye-tracking methods can be applied with young users. For the empirical studies of this research work, a dedicated DEG and its Cardboard version serving as control were designed with reference to the UK Early Years Foundation Stage Framework on numeracy. The DEG was developed based on the literature on game-based learning and young children's attention span. To validate the learning effect of both versions of the game, the between-subject experimental design was employed and a paper-based knowledge test was created. The research protocols and instruments were pilot tested and were improved, especially the eye-tracking calibration process. In analysing the eye-tracking data, fixation duration was used as proxy for focused attention, and gaze sequences were used to infer interaction strategies with a refined Gaze Sub-sequence Marking Scheme. Overall, the empirical findings provided valuable groundwork and implication for future work. Specifically, a more child-friendly eye-tracker hardware design is required to facilitate data collection with children playing DEGs. Practical game design for 5-year-olds needs to consider different factors, including game duration, game difficulty, voice-based instructions and interactive non-player-character.

Software is taking an increasingly prominent role in all of the areas relevant to social sustainability; e.g. by creating and enabling communities, facilitating connectedness between groups of people, enforcing standards, rules, and laws over individuals' and business' interactions. It enables the sharing of information and facilitate cooperation. Presently, there is no solution, framework, or method that supports software and requirements engineers in accounting for social sustainability requirements during software development. Most requirement engineers don't even understand what social sustainability is or what it may have to do with software. Thus, the main aim of this thesis is to deliver a method that de-mystifies the notion of social sustainability, and enables requirements engineers to integrate (otherwise implicit) social sustainability requirements into a software systems specification. This thesis tackles this issue by eliciting the structure of the social sustainability concern, as reported within in the current scientific literature (systematic literature review), and delivering a generic, repeatable approach through which each of the identified sub-concerns could be operationalised into reusable requirements. The proposed approach is instantiated for the equality concern, for which a value-based pattern is derived (qualitative analysis), which is operationalised through templates. An elicitation method is also derived using the pattern and usage guideline is proposed. Several targeted studies are designed to evaluate the developed pattern and method with both requirements engineering experts and general software users. The studies demonstrate the utility and (re-)usability of the method, pattern along with its template and requirements. This thesis reveals that the generic social sustainability requirements can be used (but contextualised) across all countries and cultures. It also declares that each social sustainability concern can be represented by a value pattern. For equality, this pattern is related to equal support of: stakeholders variability, goal achievement and access to services that will facilitate goal achievement.

Formal choreographic models yield a suitable frameworks to specify, analyse, and develop message-passing applications. In this application domain, orchestration has so far been preferred to choreography because the former offers a simpler approach to tame the complexity of developing distributed applications. On the one hand, orchestration yields a toolbox of handy and relatively simple solutions to practitioners. However, orchestration shows scalability limitations that choreographies seem able to tackle. Another key advantage of choreographies is the possibility of guaranteeing behavioural properties of applications "by construction". Correct components can be (semi)automatically derived from correct global specifications. In this context it is crucial to verify the correctness of global specifications. In this thesis, we adopt the global choreographies of Guanciale and Tuosto to specify global views while we abstract components as special automata. More precisely, we generalise the class of interface automata of de Alfaro and Henzinger to group interface automata and equip them with two internal products, instrumental to the verification of "sound" global specifications. In particular, these products allow us to characterise erroneous states that potentially spoil interactions. A main result of our approach is that the absence of those erroneous state is equivalent to well-formedness of global choreographies, that is a sufficient property guaranteeing well-behaviour of local components projected from the global specification.

Emotions elicited from interacting with technologies are fundamental components of user experience (UX). Two general approaches to measuring people's emotional experiences exist: memory-based and moment-based. Whereas memory-based approaches are susceptible to the peak-end effect, moment-based approaches appear to better reflect the actual experiences of emotions. In this thesis, it is proposed that assessment of emotions of think aloud verbalisations is a moment-based approach for measuring emotional experiences. To evaluate the effectiveness of this approach, two independent user studies were conducted, respectively with 46 and 35 participants, in the domain of online shopping. In both studies, three assessment methods were applied for measuring emotional experiences: Self-reports, Vocal expressions (manual analysis), and Vocal Expressions (automatic analysis). Study 1 adopted machine learning (ML) approaches to analyse participants' vocal expressions elicited during thinking aloud based on the discrete and dimensional models of emotions. While vocal analysis of verbal data was used as the momentbased approach, self-report questionnaires served as a means for the memory-based assessment. Results suggested that retrospective evaluations of emotions were significantly correlated with the most frequently elicited emotion (modal emotion) during the interaction. Study 2 expanded the extent of analysis of vocal expressions by taking a different set of ML techniques. Emotions were modelled by dimensional descriptors through binary pleasure (negative/positive), binary arousal (low/high) and ternary dominance (low/neutral/high). Results showed 66%, 70%, 41% recognition accuracy for Pleasure, Arousal and Dominance dimensions respectively. Moreover, elicited facial expressions were analysed to derive classification models to predict subjective self-reports. The main contribution of this thesis is the proposition and validation of an approach for automatic assessment of emotional experiences evoked during the think aloud protocol. Future work will investigate methods for improving the accuracy of automatic analysis, integration of multiple modalities (verbal and nonverbal) and different ML techniques in an actual automatic assessment tool.

We introduce a process calculus with a new action prefixing operator that allows to model locally controlled reversibility. Since the observation of covalent bonding in chemical reactions is the starting point of our work we call the process calculus the Calculus of Covalent Bonding (CCB). The calculus is based on CCSK, but adds an operator of the form (s; b), where s is a sequence of actions. Action b can only be executed once all actions in s are done and executing it requires to undo one action in s. By this we achieve control over when reverse actions happen without the need of a global control or a memory. The calculus also allows spontaneous undoing and sequences of forward and reverse actions. We give a formal definition of the calculus using Structural Operational Semantics rules. We also show properties of the calculus, in particular we show that it can model out-of-causal order reversibility. In order to demonstrate the use of our calculus we model the hydration of formaldehyde in water into methanediol. We use a system of four molecules, where various paths are possible to produce the final result. These path include a range of chemical situations. Our calculus can model most of them. We also model Base Excision Repair (BER), a bio-chemical process, on a higher level. This show that our calculus is also useful to model more general situations than the chemical reactions it was derived from.

The human brain hosts a colossal number of water molecules which are constantly moving due to Brownian motion. Their movement, random by nature, is restricted by the brain tissue walls. Magnetic Resonance Imaging (MRI) provides macroscopic measurements of the diffusion process in a non-invasive manner, i.e. diffusion MRI. Hidden in these measurements lies information about the underlying architecture. The ability to unravel tissue microstructure from the coarse-grained diffusion measurements is extremely valuable since this information is 2-3 orders of magnitude below typical MRI resolution. This makes diffusion MRI sensitive to pathological and developmental processes occurring at the mesoscopic scale, in the order of microns. Accessing this level of detail can lead to clinical biomarkers specific to early stages of neurodegenerative diseases or brain development. Computational models of biophysical tissue properties have been widely used in diffusion MRI research to elucidate the link between microstructural properties and MR signal formation. The potential increase in sensitivity and specificity in detecting brain microstructural changes is their major driving force. However, these models establish complex relationships between biophysical properties and the MR signal, making the inverse problem of recovering model parameters from noisy measurements ill-conditioned with conventional diffusion MRI acquisitions. This thesis explores ways to make diffusion MRI biophysical modelling more robust while maintaining time and hardware requirements that are feasible in clinical conditions. Firstly, we explore theoretically the benefits of incorporating functionally independent measurements, such as double diffusion encoding. Secondly, we propose an optimal experiment design framework that gives us, after exploring the whole multidimensional diffusion MRI measurement space, the acquisition that maximises accuracy and precision in the parameter estimation. Finally, we extract relevant information from histology images that can be used to feed or benchmark diffusion MRI models.

Landslides exhibit complicated and destructive behaviour with disastrous consequences. A landslide event is preceded by slope failure, which is governed by laws of soil mechanics. Once initiated, landslides often propagate downslope at rapid rates, exhibiting fluid-like behaviour. Landslide propagation is therefore frequently characterised by laws of fluid dynamics. Numerical models are vital for an improved understanding of these catastrophic events. Existing models are incapable of adequately simulating the dynamics of both initiation and propagation. Smoothed Particle Hydrodynamics (SPH) is a meshless method that is able to capture large displacements and rapid velocities, and it has been frequently applied to simulate landslide propagation. However, SPH is susceptible to numerical instabilities, that are particularly detrimental with regards to simulations of soil. These must be eliminated for SPH to be an ideal tool for general landslide modelling. The majority of approaches at removing the numerical instabilities from SPH are not universal -- some are applicable for small displacements problems only, while others require the tuning of artificial model parameters. In this research, a novel numerical model is developed capable of accurately simulating landslide behaviour -- including initiation and propagation. The numerical model -- Stress-Particle SPH -- removes instabilities in a way that does not require artificial parameter tuning. The method is an extension of SPH, and involves calculating velocities and stresses on two separate sets of particles -- nodes and stress-points. Previous literature suggests that the addition of stress-points have the potential to effectively stabilise SPH. Despite this, their implementation within SPH is relatively unexplored, and stress-points have only been applied to a limited range of problems. In this research, Stress-Particle SPH is extended for applicability to landslide behaviour, allowing the numerically stable simulation of high displacement problems with Stress-Particle SPH for the first time. The developments presented in this research offer the potential for SPH to tackle a broad range of problems beyond its current capabilities.

Quantified Boolean formulas (QBF), which form the canonical PSPACE-complete decision problem, are a decidable fragment of first-order logic. Any problem that can be solved within a polynomial-size space can be encoded succinctly as a QBF, including many concrete problems in computer science from domains such as verification, synthesis and planning. Automated solvers for QBF are now reaching the point of industrial applicability. In this thesis, we focus on dependency awareness, a dedicated solving paradigm for QBF. We show that dependency schemes can be envisaged in terms of dependency quantified Boolean formulas (DQBF), exposing strong connections between these two previously disparate entities. By introducing new lower-bound techniques for QBF proof systems, we study the relative strengths of models of dependency-aware solving, including the proposal of new, stronger models. Proof Complexity: Using the strategy extraction paradigm, we introduce new lower-bound techniques that apply to resolution-based QBF proof systems. In particular, we use the technique to prove exponential lower bounds for a new family of QBFs called the equality formulas. Our technique also affords considerably simpler, more intuitive proofs of some existing QBF proof-size lower bounds. Models of Solving: We apply our lower bound techniques to show new separations for QBF proof systems parametrised by dependency schemes. We also propose new models of dynamic dependency-aware solving and prove that they are exponentially stronger than the existing static models. Finally, we introduce Merge Resolution, a proof system modelling CDCL-style solving for DQBF, which is the first of its kind.

The Holy Quran is the essential resource for Islamic sciences and Arabic language. Therefore, numerous Quranic search applications have been built to facilitate the retrieval of knowledge from the Quran. This thesis presents a novel Arabic Quran semantic search model. First, this thesis evaluated existing search tools constructed for the Holy Quran, against 13 criteria depending on: search features, output features, the precision of the retrieved verses, recall database size, and types of database contents. Then, the study reviewed the existing Quran ontologies and compared them against 11 criteria. Some deficits have been found in all these ontologies. Additionally, a single Quranic ontology does not cover most of the knowledge in the Quran. Therefore, I developed a new Arabic-English Quran ontology from ten datasets related to the Quran such as: Quran chapter and verse names, Quran word meanings, and Quran topics. The main aim of developing a Quranic ontology is to facilitate the retrieval of knowledge from the Quran. Additionally, the Quran ontology will enrich the raw Arabic and English Quran text with Islamic semantic tags. Furthermore, I developed the first Annotated Corpus of Quran Questions and Answers in Arabic. This corpus has 2200 pairs of question and answer collected from trusted Islamic sources. Each pair of question and answer is labelled with 5 tags. Examples of tags are: question type: either factoid or descriptive, topic of question-based on the Quran ontology, and question class. Finally, the thesis explains a new semantic search model for the Arabic Quran based on my Quran ontology. This model aims at overcoming limitations in the existing Quran search applications. This search tool employs both Information Retrieval techniques and semantic search technologies. The performance of this search model is evaluated by using The Annotated Corpus of Arabic Quran Questions and Answers.

Healthcare processes are complex and may vary considerably among the same cohort of patients. Process mining techniques play a significant role in automating the construction of healthcare models using a system's event log. An event log is a data type that records any event that occurs within the process. It is a basic element of any information system and has three main components: process instance id, event and time when an event has occurred. Using ordinary techniques of process mining in healthcare produces `spaghetti-like' models which are difficult to understand and thus have little value. Previously published studies have highlighted the importance of event abstraction which is considered as a central tool for reducing complexity and improving efficiency. Although studies have successfully improved the understandability of process models, they have generally relied on involvement from a domain expert. Untangling these `spaghetti-like' models with the help of domain experts can be expensive and time-consuming. Machine learning techniques such as Hidden Markov Model (HMM) has been used for modelling sequential data for a long time. State transition modelling has also been explored by process mining research and is advocated for sequence clustering purposes where a model is trained over a group of sequences and then used to evaluate if a process instance is more likely to be generated from this model or not. However, state transition models can also be utilised for detecting hidden processes which can be used subsequently for process abstraction. In this thesis, we aim to address healthcare process complexity using unsupervised abstraction. We adopt an unsupervised method for detecting hidden processes using HMM and the Viterbi algorithm. The method in this research includes eight stages; event logs extraction, preprocessing, learning, decoding, optimisation, selection, visualisation and lastly model evaluation. One of the main contributions of this research is the design of two different types of process model optimisation which are strict and soft optimisations. Models that are selected by the proposed optimisation address the limitations of other standard metrics that can be used for model selection in HMM such as Bayesian Information criteria (BIC). Two different real healthcare data sources are used in this research namely the Medical Information Mart for Intensive Care (MIMIC-III) from Boston, USA and the Patients Pathway Manager (PPM) from Leeds, UK. Models are trained using the MIMIC-III medical event log and then tested using the PPM dataset to be evaluated later by a domain expert. Three breast cancer case studies that range in complexity are extracted. The results of our method have significantly improved model complexity and provided a conceptually valid abstraction for several care patterns. Promising results are demonstrated in the improvement of the precision and fitness of the abstracted models. The abstracted models can then be used as a middle step for bringing structure to unstructured processes which helps in finding cohorts of patients based on similar healthcare processes. The healthcare processes of a cohort of patients can then be modelled using any process mining tool where their process similarity could not be captured in the complex models.

This thesis focuses on proposing a novel heat management solution which could be applied in the next generation aircraft. The need for novel solutions in this sector arises primarily from the severe installation constrains and the increased use of electronics which in turn creates the need for more efficient cooling. The research is accomplished by focusing on numerical modelling of heat exchanger performance using computational fluid dynamics. In Chapters 1-3 an overview of the topics relating to the project is provided. It was particularly focused on current heat exchangers used in industry and the recent advancements within the aircraft thermal management systems driving this project. Various additive layer manufacturing techniques are compared and the most suitable one for heat exchangers is identified. In addition, current analytical/empirical design methodology of heat exchangers is compared to existing numerical modelling work. In Chapter 4 a two dimensional sinusoidal heat exchanger channel geometry was taken from a published numerical study. It was used to assess the impact of various numerical modelling assumptions at the highest computational resolution feasible. It was found that for the flows occurring in the transitional Reynolds number regime a particular care needs to be taken in numerical solution setup to predict flow and heat transfer accurately which is often not demonstrated in the literature. In Chapter 5 a plate-fin heat exchanger with serrated corrugation was experimentally and numerically evaluated allowing to build a robust numerical modelling framework for heat exchangers. During the process a novel approach to model heat exchanger corrugation was proposed. It uses a slice of a heat exchanger core and models both hot and cold fluid streams separated by a solid. It enables simulating the cross-flow heat transfer effects directly and eliminates the need for analytical/empirical models still popular within industry. The data from this novel corrugation model was then used in the heat exchanger unit model and produced a better agreement with the experimental data than normally obtained in industry. In Chapter 6 the research focused on applying additive layer manufacturing to heat exchangers. A series of novel heat transfer ideas were designed and manufactured of titanium using selective layer melting. The most promising inter-layer heat exchanger corrugation was incorporated in a novel proof of concept heat exchanger design with manifold headers. It was then evaluated numerically and compared to a more conventional pin-fin heat exchanger. Overall, the novel heat exchanger design led to increased heat transfer with no penalty in flow resistance compared to the pin-fin heat exchanger. This novel design, whilst is a proof of concept heat exchanger, is a significant step in industry and opens the way for the next generation more efficient heat transfer solutions.

Mathematical optimization is a common tool for numerous real-world optimization problems. However, in some application domains there is a scope for improvement of currently used optimization techniques. For example, this is typically the case for applications that contain features which are difficult to model, and applications of interdisciplinary nature where no strong optimization knowledge is available. The goal of this thesis is to demonstrate how to overcome these challenges by considering five problems from two application domains. The first domain that we address is scheduling in Cloud computing systems, in which we investigate three selected problems. First, we study scheduling problems where jobs are required to start immediately when they are submitted to the system. This requirement is ubiquitous in Cloud computing but has not yet been addressed in mathematical scheduling. Our main contributions are (a) providing the formal model, (b) the development of exact and efficient solution algorithms, and (c) proofs of correctness of the algorithms. Second, we investigate the problem of energy-aware scheduling in Cloud data centers. The objective is to assign computing tasks to machines such that the energy required to operate the data center, i.e., the energy required to operate computing devices plus the energy required to cool computing devices, is minimized. Our main contributions are (a) the mathematical model, and (b) the development of efficient heuristics. Third, we address the problem of evaluating scheduling algorithms in a realistic environment. To this end we develop an approach that supports mathematicians to evaluate scheduling algorithms through simulation with realistic instances. Our main contributions are the development of (a) a formal model, and (b) efficient heuristics. The second application domain considered is powerline routing. We are given two points on a geographic area and respective terrain characteristics. The objective is to find a ``good'' route (which depends on the terrain), connecting both points along which a powerline should be built. Within this application domain, we study two selected problems. First, we study a geometric shortest path problem, an abstract and simplified version of the powerline routing problem. We introduce the concept of the k-neighborhood and contribute various analytical results. Second, we investigate the actual powerline routing problem. To this end, we develop algorithms that are built upon the theoretical insights obtained in the previous study. Our main contributions are (a) the development of exact algorithms and efficient heuristics, and (b) a comprehensive evaluation through two real-world case studies. Some parts of the research presented in this thesis have been published in refereed publications [119], [110], [109].

Diversification of energy sources and transport decarbonisation are growing concerns of modern societies. Alternative fuels play an important role in addressing these challenges. For the spark ignition (SI) engine, the propensity of the fuel and fuel blends to auto-ignite is a critical characteristic that limits engine efficiency, which can be assessed by the ignition delays (τi). Severity of knock is also dependent upon the duration of heat release rate - the excitation time (τe). In this thesis, detailed evaluations of τi and τe are employed to study the tendency of methane to detonate in comparison with other fuels, employing the detonation peninsula on the ξ/ɛ diagram. The ξ parameter is the ratio of acoustic to auto-ignitive velocity, whereas ε is the ratio of the acoustic wave resistance time in a hot spot to the τe. It is shown that stoichiometric methane/air exhibits very good anti-knock properties in comparison with other fuels under turbocharged engine running conditions. The changes in the auto-ignition behaviour caused by the progressive addition of n-butanol (at 10%, 20%, 40% and 85% vol n-butanol) to gasoline (RON 95, MON 86.6) and its toluene reference fuel (TRF) are studied computationally and experimentally in a rapid compression machine (RCM) under stoichiometric condition at 2 MPa and at 678-916 K. At low temperatures, n-butanol acts as an octane enhancer, reducing low temperature heat release and increasing ignition delays, with marginal additional effects for blends above 40%. This is supported by the results from ξ /ɛ diagram, where higher n-butanol blends lie further away from the developing detonation region. A brute-force sensitivity analysis of the surrogate model suggests that the main reaction inhibiting ignition at low temperatures is H abstraction from the α-site of n-butanol, even for the 10% blend. At higher temperatures, the behaviour reverses as the chain branching routes from H abstraction by OH from the γ-site of n-butanol and from the α-site by HO2 become more dominant, promoting ignition. For the lower blends, the largest discrepancies between simulations and experiments are found in the negative temperature coefficient (NTC) region, where a larger number of reactions contribute to the uncertainty in predicting τi. For the higher blends, the largest discrepancies occur at low temperatures, indicating that uncertainties within the low temperature n-butanol chemistry need to be resolved. Regarding τe, the addition of n-butanol to the TRF blends has a negligible effect. Furthermore, τe, is not influenced by NTC chemistry.

Cloud Computing has transformed the way in which enterprises and individuals are utilising the Information Technology (IT) by offering on-demand services such as applications, platforms and infrastructures for their customers with reasonable prices based on their usage (e.g., pay-as-you-go model). However, the wide adoption of Cloud Computing and the growing number of Cloud customers have increased the overall operational costs for Cloud providers, especially with the increasing cost of energy consumed to operate Cloud services. Consequently, Cloud providers consider energy consumption as one of the most important cost factors to be maintained within their infrastructures. In order to achieve energy efficiency and reduce the operational costs for Cloud services, reactive and proactive management mechanisms can be used to efficiently manage Cloud resources and reduce energy-related costs while maintaining service performance requirements. However, these mechanisms need to be supported with performance and energy awareness not only at the physical machine (PM) level but also at virtual machine (VM) level in order to make enhanced cost decisions. Moreover, estimating the future cost of Cloud services can help the cloud service providers offer suitable services that meet their customers' requirements. This thesis introduces a Cloud system architecture along with a novel Cost Modeller component that aims to enable the awareness of energy consumption, performance variation and cost in a Cloud environment. To fulfil this aim, an energy-based cost model is firstly developed to attribute the PM's energy consumption to VMs and measures the actual resource usage, power consumption and the total cost for each VM. An energy-based cost prediction framework is then introduced to predict workload, power consumption and estimate the total cost of the VMs during service operation based on historical workload data. Finally, a performance and energy-based cost prediction framework is introduced to combine VMs consolidation and resource provisioning in order to design cost-effective strategies while taking into consideration the trade-off among cost, energy efficiency and performance variation of Cloud services. The evaluation of the proposed research on a Cloud testbed shows that the proposed energy-based cost model is capable of fairly attributing the PMs energy consumption to heterogeneous VMs, thus enabling cost and energy awareness at the VM level. Compared with actual results obtained in the Cloud testbed, the predicted results show that the proposed energy-based cost prediction framework is capable of predicting workload, power consumption and estimating the total cost for heterogeneous VMs based on historical workload patterns. Additionally, the results have shown that the proposed performance and energy-based cost prediction framework is capable to estimate the total cost of heterogeneous VMs by considering their resource usage and power consumption, while maintaining the expected level of service performance. The application of the proposed research provides the awareness of energy consumption, performance variation and cost at the virtual level in Cloud environments, which contributes to overcoming the challenge of identifying the most cost-effective strategies for Cloud services. The outcomes of this research can be used and incorporated in reactive and proactive management mechanisms to make enhanced cost decisions supported by performance and energy awareness in order to efficiently manage Cloud resources. This has the potential to contribute to a reduction in energy consumption, and therefore lowering the total cost for Cloud providers while maintaining the service performance.

Identifying lexical semantic relations in the text has been a long-standing dream of artificial intelligence and the target of many researchers' attention over the past years. This thesis addresses the problem of identifying antonymy relations, such as(hot/cold) in an automatic method. This work presents three key points in capturing antonymy word pairs: extracting word pairs examples from a textual corpus, representing antonymy in a pair vector space model, and using a machine learning classifier to predict the antonymy relation. Researchers have found that discriminating antonymy from synonymy is a non-trivial task. Both relations show similar semantic distributions as they are found in similar contexts. This issue affects many similarity-based applications by displaying opposite words instead of synonyms. Moreover, both traditional and modern vector space models such as Bag-of-Words and Word Embeddings models show poor discrimination between antonymy and synonymy words. Therefore, this work proposed antonymy pair vector representation based on symmetric classified patterns extracted from a corpus. Besides, we are motivated by extracting novel antonymy and opposites relations between word pairs. This research aims to capture and predict antonymy pairs generated by a textual corpus to make computers able to understand and capture opposition relation in the text. Our research proposes the Antonymy classifier which combines two approaches: the pattern-based approach and a machine learning classifier. We use the pattern-based approach to extract word pairs and patterns. We also propose using distant supervision learning to label the extracted pairs automatically. Distant supervision uses an external knowledge base (the Open Multilingual WordNet) to generate positive and negative antonymy instances. It also extracts every sentence from a corpus which shows both canonical antonymy pairs such as(national/international)and non-canonical antonymy or opposites pairs such as (internal/international) that might provide statistical evidence for an antonymy relation. In addition, this work presents a pattern classifier model which automatically extracts and classifies antonymy patterns by computing the average co-occurrence association between positive(antonymy) and negative (non-antonymy)instances in the training set. A part of these patterns such as (between X and Y, both X and Y, from X to Y ) were found in related linguistic studies on manual patterns extraction and analysis. We also found some novel textual patterns that are highly associated with antonymy pairs such as (however X or Y, what is X and what is Y) and more. This work also shows experiments in extracting and predicting antonymy on the English BNC and SkELL corpora and the Arabic ArTenTen corpus. The overall outcomes showed a positive prediction improvement in distinguishing antonym pairs compared to previous attempts. Also, we presented new antonymy pairs that are not found in the English and Arabic WordNet. The antonymy classifier model uses a machine learning algorithm to extract and classify novel adjectival and noun antonymous pairs such as (verbal/visual), (input/output), (life/death) and(material/spiritual). Therefore, the work presented in this research is a promising method for better extraction and classification of antonymy pairs and patterns in a corpus.

In this thesis we initiate the study of the proof complexity of modal resolution systems. To our knowledge there is no previous work on the proof complexity of such systems. This is in sharp contrast to the situation for propositional logic where resolution is the most studied proof system, in part due to its close links with satisfiability solving. We focus primarily on the proof complexity of two recently proposed modal resolution systems of Nalon, Hustadt and Dixon, one of which forms the basis of an existing modal theorem prover. We begin by showing that not only are these two proof systems equivalent in terms of their proof complexity, they are also equivalent to a number of natural refinements. We further compare the proof complexity of these systems with an older, more complicated modal resolution system of Enjalbert and Farinas del Cerro, showing that this older system p-simulates the more streamlined calculi. We then investigate lower bound techniques for modal resolution. Here we see that whilst some propositional lower bound techniques (i.e. feasible interpolation) can be lifted to the modal setting with only minor modifications, other propositional techniques (i.e. size-width) fail completely. We further develop a new lower bound technique for modal resolution using Prover-Delayer games. This technique can be used to establish "genuine" modal lower bounds (i.e lower bounds on the number of modal inferences) for the size of tree-like modal resolution proofs. We apply this technique to a new family of modal formulas, called the modal pigeonhole principle to demonstrate that these formulas require exponential size modal resolution proofs. Finally we compare the proof complexity of tree-like modal resolution systems with that of modal Frege systems, using our modal pigeonhole principle to obtain a "genuinely" modal separation between them.

Temporal planning has become an increasingly popular area of research in recent years. Much of this is because the modelling of many real world problems requires the consideration of time and how actions interact and depend on one another. A key phenomenon that modelling with time allows us to accommodate when planning is the notion of concurrency, where actions occur simultaneously. Required Concurrency is where certain actions must occur in parallel due to one or more dependencies between actions. Dealing with concurrency in temporal planning requires careful reasoning when making choices about which actions to use during search. Inference can be used to assist in problem solving when there is required concurrency because certain actions must occur at the same time. The research presented in this thesis explores how problems of required concurrency can be solved using more inference and less search. Specifically, the work focuses on pair-wise cases of required concurrency. Novel techniques are presented for detecting required concurrency. In addition, new infer and search algorithms capable of exploiting inference from these detections and reducing search are developed. The algorithms for pattern detection, the combined infer and search strategies, along with the inference engine that powers them are implemented as an extension of POPF. We present an evaluation of the practical benefit from this new planner that we call POPI. Additionally, we present a formal analysis provably showing which unique types of required concurrency exist between pairs of durative actions.

This thesis is to investigate the stability analysis problem of fuzzy-model-based (FMB) control systems and its application to continuum manipulator. The stability analysis of T-S FMB control system is conducted on the basis of Lyapunov stability theory relaxed by membership-function-dependent (MFD) approach and extended to some practical control problems, such as control input saturation and guaranteed cost of system indexes, where numerical examples are raised to verify the effectiveness of each proposed fuzzy control method. An example of continuum manipulator is developed to show the process and advantage of applying fuzzy logic and fuzzy-model- based control method to a real complex practical system. The main works and contributions of the thesis are summarized in the following three parts: 1) The first part of work is presented in Chapter 3. It aims to consider the output feedback tracking control with control input saturation problem under the framework of Takagi-Sugeno (T-S) FMB control. A fuzzy controller is employed to close the feedback loop, which aims to drive the system states to follow those of a stable reference model subject to H∞ performance. To enhance the fuzzy controller design flexibility, the number of rules and premise membership functions are not necessarily required to be the same as those of fuzzy model. To address the control input saturation problem, linear sectors are created by local upper and lower bounds to include the possible control area such that the nonlinear saturation problem can be tackled. Then, the membership-function-dependent (MFD) technique is used to embed the information of membership functions to the stability conditions in the form of linear matrix inequalities (LMIs). A numerical example is given to demonstrate the effectiveness of proposed method dealing with different levels of control input saturation problems. 2) The second part of the thesis work is presented in Chapter 4. It considers the guaranteed cost stability analysis in the T-S FMB control system. A weighted quadratic cost function is considered as the cost index to measure the performance of the closed-loop system in terms of the system states, system outputs and control signals. The stability of the FMB control system is investigated by the Lyapunov stability theory subject to the minimization of cost index for performance realization. An MFD approach using the piece-wise linear membership 3 functions (PLMFs) technique is employed to include the information of membership functions into the stability analysis. MFD stability conditions in terms of LMIs are obtained to determine the system stability and feedback gains with the consideration of the system performance measured by the cost function. An numerical example is raised to demonstrate the effectiveness and merits of proposed method. 3) The third part of the work focuses on applying fuzzy logic and FMB control to a practical example of nonlinear systems. Based on the curve geometry of a continuum manipulator, the kinetic and potential energies can be calculated as the integration of the energies regarding to each slice perpendicular to the backbone of the manipulator. By applying Euler-Lagrangian equation of motion, we can obtain the dynamic model of continuum manipulator with the capabilities of bending and contractile. Two traditional nonlinear control methods, namely inverse dynamic control and sliding mode control, are implemented to drive the system states to follow the desired trajectory of a stable reference system. In order to improve the tracking performance and solve the chattering problem particularly in sliding mode control, a fuzzy sliding mode controller is proposed by applying fuzzy logic theory, which varies the value of feedback gains adaptively to the manipulator configurations and successfully attenuates the chattering problem in the traditional sliding mode control method. In order to further apply the FMB control on the practical example of nonlinear systems, the dynamic model of continuum manipulator needs to be transformed to a polynomial fuzzy model, where the fuzzy model of two-link rigid body manipulator is developed preparatively as a relatively simple practical nonlinear system. Then the development polynomial fuzzy model of continuum manipulator is proceeded with necessary simplifications.

Diabetes is a chronic disease that affects more than 300 million people worldwide. It is caused by abnormal glucose levels in the body. These levels are typically measured by piercing the fingers to collect blood, which is then placed on an electrochemical strip that, when inserted into a device it reads the glucose content in the blood. Nowadays, there is no commercially-established system that allows the monitoring of glucose levels without drawing blood, despite the need of millions of people that check their glucose levels on a daily basis. This thesis presents the progress towards developing a novel non-invasive glucose monitoring system that uses wireless transmission of millimetre waves (around 60 GHz) to detect glucose changes in the blood. The thesis also proposes an antireflection coating metamaterial film to enhance the diagnostic accuracy of the non-invasive glucose system. The work presented in this thesis combines theory, simulations and experiments to prove the possibility to detect glucose non-invasively. As a proof of concept, the idea is tested on controlled aqueous glucose solutions showing a nearly linear dependence between glucose levels and transmitted power. This idea is also tested in in-vivo pre-clinical experiments with a porcine subject and ten human subjects. For the in-vivo experimental campaigns, the device shows the ability to detect blood glucose spikes after the glucose is injected into the bloodstream. The last part of the thesis proposes the idea of using thin, micro-structured, metal-dielectric films in the propagation path of the signal to improve the sensitivity of the non-invasive glucose detection method. These films are antireflection coating metamaterials that match the impedance of the material that is placed behind them, thereby reducing the reflections and maximizing the transmission of the signal. Different metamaterial films are designed to operate in contact with acrylic, skin and skin-blood layers. Their performance is assessed by theory, simulations and experiments including in-vivo testing, demonstrating the antireflection effect.

The goal of expected 5th Generation (5G) wireless networks is to bring ultra high data rates, network throughput and service quality to mobile users in near future. Regarding to these expectations, the present virtualized and softwarized network architecture such as Cloud-Radio Access Network (C-RAN) and Software Defined Networks (SDN) is responsible for performing multiple key tasks such as allocating radio resources efficiently, providing Network Function Virtualization (NFV) and intelligent inter-network control/coordination, and flexibly sharing network resources among different network tenants and etc. Although these mentioned techniques and approaches already achieved certain progresses but also face various challenges because of the realistic engineering constraints and complicated network scenarios. To this end, this thesis contributes a series of 5G wireless networks optimization frameworks and efficient algorithms tackling different network problems in both theoretical and practical ways. The specific works of the thesis include the following proposals: a deep research of control plane in future SDN architecture, which is capable to provide intelligent control functions. Furthermore, based on such SDN based architecture, a control plane signal optimization framework is defined and solved, which optimally reduces the potential handover signals and balances control load between multiple control planes in the SDN enabled networks. On the other hand, by introducing network virtualization technologies, this thesis further proposes an optimization framework to flexibly share radio resources between network operators by using both intra-tenant and inter-tenant management entity. In addition, downlink signal transmission power control is also introduced to the framework to engage further resource reuse among tenants and achieve certain degrees improvement in energy efficiency during network operations. Due to the high complexity of the proposed optimization problem, various algorithms are proposed to solve this inter-tenant resource sharing problem efficiently, and numerical results demonstrate a satisfying improvements in various network performance. Finally, a service rate aware resource sharing optimization framework is proposed to bring the previous inter-tenant sharing framework to a more realistic network architecturing level by considering fronthaul limitation and potential multi-cell coordinated resource transmission. Since the high complexity of this optimization problem, novel algorithm is also designed and numerical result demonstrates satisfying gains in a very complicated 5G networks environment.

Genomics, owing to its immediate applications in medicine, forensics, evolu-tionary and molecular biology etc., has witnessed a dramatic advancement in the technology of acquiring and generating data. Consequently, the botleneck of the information-extraction pipeline has shifted from data-acquisition to the computational capacity for storing and processing prodigious amounts of data. Uncertainty and identification of regularity in data are two key aspects contributing to the complexity of the task of mining knowledge and insights from genomic data. One form of macro-level uncertainty arises in sequential data when a single representation is used for a multitude of strings which are by and large similar. For example, in human genomics, the reference genome has been represented as a single sequence so far. Now, with the availability of a vast collection of human genomes, the so called reference cohorts seem more sensible in order to avoid the reference-bias presented by a single genomic sequence. Different representations have recently been explored in an attempt to organise human genomic sequences in reference cohorts. Each such representation has its own challenges. Moreover, in genomic sequences, local regularity (a term encapsulating various forms of repetitions) is often flanked by regions of interest - genes, for example - which are, in comparison, not regular. In other words, the regularity of a local segment of genomic data is indicative of it being a potential biologically-important region. One of the multiple possible ways to express this notion of local regularity of strings can be in terms of unbordered factors of a string. A border of a string - one of the central properties characterising the regularity associated with repetitions - is its (possibly empty) proper factor occurring both as a prefix and as a suffix. This dissertation presents an assortment of efficient novel algorithms - based on string algorithms and data-structures - to solve three problems that find direct or indirect applications in genomic data analysis. Specifically, the presented algorithms handle the uncertainty arising in the representation of an ensemble of sequences as well as characterise the regularity present in a sequence in terms of unbordered factors. Firstly, we present an optimal algorithm - in terms of both time and space -improving the state-of-the-art, to identify Superbubbles (a special type of self-contained subgraphs, each with a single source and a single sink) in de Bruijn sequence graphs for genome assembly. Identifying these motifs in a reference graph is crucial for overcoming the lack of a coordinate system in the graphical representation of a reference cohort. Secondly, we introduce another representation for sequential data with macrolevel uncertainty, called Elastic-degenerate strings. The motivation is to condense a set of genomes (with variations) as a reference cohort. An elastic-degenerate string is a string in which an elastic-degenerate symbol can occur at one or more positions; each such symbol corresponds to a set of two or more variable-length strings. We not only formalise the concept of elastic-degenerate strings but also present a practically efficient algorithm to solve the pattern matching problem in a given elastic-degenerate text. Lastly, we provide a quasilinear time algorithm to compute the Longest Unbordered Factor Array of a string w for general alphabets. This array specifies the length of the maximal unbordered factor (the longest factor which does not have a border) starting at each position of w. This is a major improvement on the running time of the currently best worst-case algorithm working in O(n1.5) time for integer alphabets, where n is the length of w. Although this problem is rooted in theory, the data-structures proposed in this algorithm can be used to characterise the regularity of a sequence; this has possible applications in genomics.

With the advanced telecommunication technologies, the patients can receive a good quality of medical care. In the past decades, many telecommunication technologies have been proposed and used to process patients' electronic medical records (EMR). For the most of current EMR systems, patients only can get access to some simple medical files, for example, the personal details, description of health condition from physicians, chart records between patients and physicians, and payment bill, etc. However, with more detailed medical records, physicians could provide more effective medical services, which is not only saving the vital time for patients but also medical expenses. In this research, we proposed the systems that patients and physicians can get access to the EMR including medical images and videos with low transmission delay at any time. The primary objective of this research is to come up with new allocation systems and cache patient's EMR in different location of the patient's daily life. The patients and physicians can access EMR timely when needed. The proposed systems could decrease the transmission delay and provide more effective and efficient medical care for patients. Three systems have been proposed based on femtocaching, dynamic vision sensor, edge computing, game theory and deep Q-learning in this research. For the first system, we allocate patient's medical records (text/words, images, and videos) to the femtocachings which are set up close to the nearest hospitals of the home area, workplace, family home, friends' home and other places according to the patient's social life. Femtocaching has been used to choose and cache the best medical files. Dynamic vision sensor has been proposed to decrease the size of medical video files due to the fact that the medical video files account for a large part of EMR. Dynamic vision sensor takes video according to the reflection of light and how fast the subjects move. It is the first time that femtocaching technology and dynamic vision sensor have been proposed in the health care area. Different from the previous system, for the second system, an auction-based and non-cooperative game theory algorithm based on edge computing has been proposed for patients to share their edge devices to others instead of themselves only. Edge computing has been used to process medical records instead of cloud computing. There has been a large volume of research of cloud computing that has been conducted in the health care area in the last decade. However, there is an increase in the amount of medical data correlating with the rapidly increasing number of internet of things (IoT) devices. Cloud computing, especially, is not efficient enough for those health devices that require very short response times, this could cause intolerable network latency. Edge computing could cache and process computing tasks at the edge of the network without transferring to the cloud network and protect the privacy of patients. A competitive game mechanism is proposed by combining with the patient's health condition and telecommunication channel quality on the edge. The system would compare the priority of patients and allocate medical files based on the results of the game theory algorithm. In our third system, we proposed a new sharing algorithm based on deep Q-learning to allocate and cache the medical records of multiple patients. The new system provides more fair and humanized medical care services comparing with the last system. In our last system, all the medical parameters in the game theory allocation algorithm have the same weights, which would cause the situation that the patient has higher illness severity but lower priority. However, in this new system, the patients with a higher level of illness severity have higher priority and lower waiting time to be served. The system would not stop seeking the caching path for patients until the best Q value is found. At last, according to the overall picture of the research conducted, the main conclusions together with some directions for the future works are presented.

Soft materials have opened new possibilities in robotics: the use of inherently adaptable mechanical structures allows soft robots to negotiate with uncertain and unstructured tasks. Nevertheless, their sizable elastic deformations pose a limitation to their modelling and control. Moreover, the maximum achievable stiffness is typically too low to provide performances comparable to those of their rigid-linked counterparts. The research presented in this thesis aims at overcoming these limitations and bridging the gap between traditional and soft robotics by proposing a novel design paradigm for a Variable Stiffness System (VSS) based on soft materials with the potential to be employed in a wide range of application areas. The proposed design takes inspiration from the antagonistic stiffening mechanism of muscles in nature, in which the balancing of two opposing contractile muscle forces allows to achieve infinite stable configurations. A combination of flexible inflatable membranes and flexible, yet inextensible fabric sleeves is used to enable fine stiffness tuning, by mean of pneumatic actuation. The conjoint use of tendon-driven actuation is proposed to enable not only stiffness controllability, but also shape-shifting and shape-locking capabilities. The use of fabric allows for significantly higher pressures to be used, thus, larger forces can be exerted on the environment, still making use of soft materials. This has led to contributions within the areas of industrial collaborative robots, where the concept of stiffness-controllable robotic link has been explored to enhance safety in Human-Robot Interaction (HRI); in surgical robotics where the use of the proposed antagonistic actuation mechanism has been investigated to improve the dexterity of laparoscopic tools; and in rehabilitation robotics, where the same mechanism is employed to improve the ergonomics of state-of-the-art exoskeletons for hand-rehabilitation. This thesis shows how this novel design concept can be applied to effectively improve or replace a wide range of state-of-the-art robotic systems.

Reflective DDoS (DRDoS) attacks are an almost 20-years-old phenomenon but despite all this time, there are still no effective countermeasures. The main reason for this is that all the current solutions operate reactively. Attackers first discover a DRDoS vulnerability in a protocol, abuse it to create attacks, and finally security experts attempt to fix the particular problem. However, by the time the vulnerability is mitigated, the attackers simply find a new DRDoS vulnerability to exploit, and the cycle continues. In this thesis, we propose a novel approach to limit DRDoS attacks based on the hypothesis that it is possible to identify DRDoS protocol vulnerabilities systematically and proactively. We have developed a systematic analysis to achieve this and when we applied it to five well-known network protocols used in DRDoS attacks (CharGen, QotD, DNS, SNMP, and NTP) we discovered that only 'the tip of the iceberg' is revealed by the attackers: there are many more ways in which DRDoS attacks can be created using these protocols. Altogether we have identified 131 different DRDoS attack scenarios, 76 of which are not reported in the literature. Moreover, we have tested 85 of them empirically. Current statistics for DRDoS attacks, our systematic analysis, outcomes from analysing protocols and experimental validation are set out in this thesis together with a strong argument as to why this novel systematic and proactive approach is required to effectively break the cycle and mitigate DRDoS attacks in the long term.

This thesis keeps a close eye on the cache allocation problems in future mobile networks. Mobility-aware cache allocation methods are tailored to the Information/ Content Centric Networking, Software Define Networking as well as Network Function Virtualization paradigms which are considered as the enablers towards flexible 5G networks. The methods developed were depicted as various optimization frameworks that were based on classic theories including Game Theory and Optimization Theory. The formulized optimization frameworks were solved via various methods i.e. Integer Linear Programming, Convex Optimization and dual methods. Heuristic methods (e.g., Genetic Algorithm) were utilized for reducing the computational difficulty of the methods developed. The attainable system performance of the caching allocation methods developed were tested in a wide range of simulations. The methods developed have achieved excellent performance especially in the public transportation scenarios. The research conducted in the thesis has not put much attention on the two sub-problems: what content to be cached and when to update the cache profiles. The research of these two sub-problems remains open and it can be a strong supportive complement to the methods developed in this thesis. The thesis has contributed to the new network paradigms by enhancing their ability of handling high mobility applications. In summary, the research conducted in the thesis should provide theoretic guidance for the caching deployment in future mobile networks.

In this thesis, we study problems on a type of uncertain sequences called weighted sequences. In weighted sequences, for every position of the sequence and every letter of the alphabet a probability of occurrence of this letter at this position is specified. This type of sequences are commonly used to represent imprecise or uncertain data, for example in molecular biology, where they are known as PositionWeight Matrices. Normally in weighted sequences problem a cumulative weight threshold 1/z ∈ (0,1] is given and we only consider the substrings with cumulative probability at least 1/z, which we call it valid. Two main problems on weighted sequences are addressed in this thesis: pattern matching and weighted indexing. An application of weighted indexing is also presented to solve palindromic factorization in weighted sequences. • In pattern matching problem we are given a pattern of length m, a text of length n > m and a cumulative weight threshold 1/z. Our aim is to find all valid occurrences of the pattern in the text. We provide two types of average-case algorithms: one is to solve the problem with searching time o(n) and the other one is to solve the problem with searching time O(nzlogm/m). Both algorithms require a specific weighted ratio z/m . An average-case algorithm for approximate pattern matching on weighted sequences is also provided in the thesis. • In weighted indexing problem we are aiming to build an index containing all the valid substrings of a given weighted sequences of length n. We introduce a type of property string of length nz called z-estimation which contains all valid substrings, and provide algorithms to construct weighted suffix tree and weighted suffix array in time and space O(nz) based on z-estimation. • In palindromic factorization problem we are given a weighted sequence of length n and a threshold 1/z. We find all the maximum palindrome in the sequence and give a maximal-palindromic factorization of the given sequence. Our algorithm is based on weighted suffix tree and solve the problem with time and space complexity O(nz). Experimental results are given using synthetic or real data with DNA alphabet (A, C, G, T) for all the algorithms to prove our complexity theorem or to present the time and space performance.

Over the course of the last few decades, Internet has evolved to form a major technological component in our society. Accompanying this exponential growth are many challenges embracing both users expectations and ambitions of network providers. Internet Protocol (IP) is continuously being extended from the native Internet towards ubiquitous packet delivery models for the next generation of cellular networks. The two families of networks with their means of access, technologies and architectures are converging. The mutable access network structures, deployment practicalities and global traffic increase, necessitate advancements of packet delivery models. Evolution of the cellular network architectures introduces the flat-IP structure and novel topological arrangements in the backhaul. A flexible routing within an efficient Traffic Engineering (TE) is needed for novel random all-IP access network topologies with heterogeneous wireless accesses. These requirements converge with the conventional IP intra-domain routing concepts. Multi-Plane Routing (MPR) is a TE approach comprised of offline and online features tailored for all-IP access networks. Diversity is leveraged upon via multiple paths between traffic sources and destinations rendered by slicing the physical topology into multiple logical routing planes. Performance is enhanced utilising the entire topology while minimising the MPR-incurred protocol overhead. In this thesis, MPR extensions reflecting the evolution of access networks and emerging traffic types are studied. Minimum Set Cover (MSC) mathematical approach for random graphs representations of IP access network is applied accommodating novel features in offline TE: capacity planning/projections in conjunction with path correlations. Initially, a MPR utilization of the whole topology is validated via extensive simulations for diverse models of IP packet delivery improving upon key performance criteria and supporting meshing. The study proceeds to broader cases of random topologies of different sizes and sparseness where MSC-approach proves as superior by managing such complex routing resources. Additionally, Quality of Service (QoS) is tested for many traffic types including the Tactile traffic with its stringent requirements.

Human action recognition is one of the important research areas in computer vision. Manipulation action recognition which contains complex human-object interactions is a challenging problem in this area. Especially for the manipulation actions in the kitchen scenario, occlusions among objects and human body parts and the transformations of objects increase the difficulty for action recognition. Previous methods on manipulation action recognition often rely on high-level representation approaches such as object detection and human body part detection, which contain expensive work of object and human annotations. Meanwhile, the object transformation information has not been considered comprehensively. This thesis proposes a method for manipulation action recognition by generating and mining superpixel groups in the videos. Manual annotations of objects and human body parts are not required in the method. We develop a new mid-level representation method called superpixel groups to capture object parts, human body parts and object transformation information in manipulation actions. A hierarchical structure can be built based on the superpixel groups. A participant-based mining algorithm is introduced in this thesis. The proposed mining approach combines discriminativity and representativity characteristics to retrieve discriminative patterns for each action, which is more effective than mining methods that only use discriminativity characteristic. We evaluate the proposed method on two challenging manipulation action datasets, achieving state-of-the-art result on the 10-class classification problem in the 50 Salads dataset in terms of frame-wise accuracy. Our method also obtains comparable result with the methods using additional object detection and human skin detection in the "Actions for Cooking Eggs" dataset contest.

Ideation is described as the process of generating useful ideas to help reach a specific goal or outcome. It is one of the task types with the longest standing roles in experimental studies of creativity. Despite being a well studied concept, many aspects of the cognitive processes underpinning ideation are poorly understood. A main model of thinking that is referred to when looking at ideation studies is the concept of Search for Ideas in Associative Memory (SIAM), which suggests that ideas are generated in semantically related clusters, due to the way we synthesise these ideas from memory. The SIAM model of thinking is likened to that of information foraging theories in this thesis, which forms the basis for our studies. In this thesis we investigate the time-course pattern of idea generation in an ideation task and ask whether we can find any influences on this. In particular, we look to classic stopping heuristics to try to explain the time-course of ideas across a continuous single question ideation task and the giving-up strategies used in a multiple sequential question ideation task. Throughout the work presented in this thesis, we ran into methodological problems with inter-rater disagreement between coders rating the sets of ideas on commonly used metrics such as novelty and value, as well as classifying ideas in semantic categories. The second area we investigate in this thesis is whether we can support ideation in such a way that judgement of ideas becomes more reliable. We ran three ideation studies in order to develop theory in support of applications to assist in the ideation process. In our first study, we looked at continuous ideation across a single ideation task. Key findings from this study showed a weak verification of semantic clustering of ideas generated, however, we were able to replicate main effects others had found using this method. In the second study, we looked at ideation across multiple questions. Our key finding directly mirrors that found in other (non-ideation) studies on discretionary task switching: people combine rate of return and sub-goal completion as strategies for abandoning a task. In the third study, we addressed the issues of low inter-rater agreement by performing simple instructional manipulations for ideators. Key findings show a that the use of thematic roles have a positive effect on judged value of ideas. Novelty judgement agreement was also positively affected, although this was seen across conditions. Although a lot more research is required in order to fully understand the cognitive processes in ideation, we hope that this work shows that by detailed experimental analysis of ideation we might learn some possible interface interventions through which ideators might be supported.

Matching points between images is difficult because of the many possible variations between images of the same scene or 3D feature. This includes different viewpoints, changing lighting conditions, occlusion and noise. In this work, I demonstrate that the use of a local descriptor which is individual to each interest-point improves matching performance over using one globally. I propose two different approaches. The first is to use a pooling operation, based on geometric blur, which is individual to each interest-point. This is achieved by estimating how each interest-point will appear in other images through generating synthetic warps. The second is to learn an optimal combination of base-descriptors for each interest-point so as to obtain an optimal descriptor for each interest-point. This is achieved through supervised learning based on having multiple instances of the same interest-point. Another difficult problem is detecting people in images of artwork, because of the huge variation in the ways people are depicted. This included the media used (e.g. pencil, paint and sculpture) and the range of poses and projections, including Cubism at the extreme. I demonstrate that state-of-the-art CNN based methods yield improved performance if the models are fine tuned on artwork, however their performance on photos is subsequently reduced. This shows that these approaches cannot simultaneously generalise over and perform well on both photos and artwork. The underlying theme of this dissertation is the proposition that these and other algorithms lack generalisation because their invariance is at too low a level, lacking high level semantic information. For further work, I suggest the incorporation of high level semantic information into interest-point matching and better modelling of the structure of people which is shown to be essential for detecting people across different depictive styles.

The quadratic advantage in query-complexity that Grover's algorithm offers for the unstructured search problem is well-known by the cryptographic community and the security estimates for many cryptosystems are based upon quantum resource estimates for Grover's algorithm. In essence, Grover's algorithm consists of the repeated application of a quantum oracle, unique to the instance of the search problem, and some less costly circuitry. This quantum oracle is essentially a reversible boolean circuit, implemented via quantum gates. This thesis examines the optimisation of quantum search routines in the noise-free quantum circuit model of computation, with the goal of negating a portion of the cost associated with the necessity to implement quantum oracles in Grover's algorithm. The importance of such gains is illustrated by considering the approximate circuit-size of Grover's algorithm for the single-target search problem, $\frac{\pi}{4} \cdot 2^{\frac{n}{2}} \cdot \text{poly}(n)$ where $\text{poly}(n)$ is the cost of implementing the quantum oracle. If $n = 100$ and $\text{poly}(n) =n^3$, then the approximate quantum circuit-size is $2^{70}$ compared to the lower bound implied by query-complexity of $2^{50}$ and this overhead of $2^{20}$ has an associated real-world cost in terms of the number of quantum gates which must be implemented. Reduction of this overhead can therefore result in gains for any potential implementation. Optimisation of the reversible circuitry which implements the quantum oracle can achieve this and this thesis demonstrates that this can also be achieved by other methods for certain problems by using modifications of Grover's algorithm. As a motivating example throughout this thesis, we apply these techniques to lower quantum resource estimates for cryptanalysis of the {\it Multivariate Quadratic problem} over $\mathbb{F}_2$ and later in cryptanalysis of the {\it Advanced Encryption Standard} (AES).

Stop motion animation is a popular creative medium with applications across film, games, education and health industries. Traditionally this type of animation has been known as a two dimensional (2D) art form, where the artist iteratively deforms a character and then takes photographs of it. The artist's task can be overwhelming as he has to reshape a character into hundreds of poses to obtain just a few seconds of animation. Moreover, features that took a lot of effort to create remain unseen due to the preferred 2D method of visualization. The current project was a collaboration between Fat Pebble Games Studio and the Centre for Digital Entertainment (CDE) from the University of Bath. The aim was to create a novel pipeline for reconstructing and enhancing stop motion animation from three dimensional (3D) character scans, obtained from multi-view images. Deformation, non-rigid registration and interpolation techniques were used to fulfil this aim. These procedures were aided by information about the material a character is made from and the character's structure. The underling inquiry of the project was to see whether reverse engineering the artist's plasticine modelling process can result in physically plausible deformations between scans and more accurate non-rigid registrations. Message: A specialized pipeline for animation reconstruction and enhancement of handmade, plasticine characters can be created by combining deformation, non-rigid registration and interpolation. These techniques can be adapted to imitate the physical properties of plasticine, by including information about the material and structure of the original models.

This work introduces a variant of Gundersen, Heijltjes, and Parigot's atomic lambda calculus. This typeable calculus with explicit sharing extends the Curry-Howard interpretation of open deduction already established by the mentioned. We extend the intuitionisitc proof system with a switch inference rule, which corresponds to an end-of-scope operator i.e. the switch rule can only be applied to subterms of an abstraction that do not have the binding variable occuring freely inside. Combining this notion of scope with the previous calculus, where duplication of terms is performed atomically, results in the ability to duplicate the spine of an abstraction, where the spine is the direct path from the binder to the bound variable. Spine duplication has been witnessed previously with strategies involving graphs and labels, but this work uses environments and remains in a typed setting. We prove that this calculus preserves strong normalisation with respect to the lambda calculus, satisfies the confluence property, and can duplicate the spine on any abstraction.

The 3D Web emerged as an agglomeration of technologies that brought the third dimension to the World Wide Web. Its forms spanned from being systems with limited 3D capabilities to complete and complex Web-Based Virtual Worlds. The advent of the 3D Web provided great opportunities to museums by giving them an innovative medium to disseminate collections' information and associated interpretations in the form of digital artefacts, and virtual reconstructions thus leading to a new revolutionary way in cultural heritage curation, preservation and dissemination thereby reaching a wider audience. This audience consumes 3D Web material on a myriad of devices (mobile devices, tablets and personal computers) and network regimes (WiFi, 4G, 3G, etc.). Choreographing and presenting 3D Web components across all these heterogeneous platforms and network regimes present a significant challenge yet to overcome. The challenge is to achieve a good user Quality of Experience (QoE) across all these platforms. This means that different levels of fidelity of media may be appropriate. Therefore, servers hosting those media types need to adapt to the capabilities of a wide range of networks and devices. To achieve this, the research contributes the design and implementation of Hannibal, an adaptive QoS & QoE-aware engine that allows Web-Based Virtual Museums to deliver the best possible user experience across those platforms. In order to ensure effective adaptivity of 3D content, this research furthers the understanding of the 3D web in terms of Quality of Service (QoS) through empirical investigations studying how 3D Web components perform and what are their bottlenecks and in terms of QoE studying the subjective perception of fidelity of 3D Digital Heritage artefacts. Results of these experiments lead to the design and implementation of Hannibal.

This research considered a novel approach to network security by combining a new networking architecture based on the Identifier-Locator Network Protocol (ILNP) and the existing Domain Name System (DNS). Specifically, the investigations considered the mitigation of network-level and transport-level based Denial of Service (DoS) attacks. The solutions presented for DoS are applicable to secure servers that are visible externally from an enterprise network. DoS was chosen as an area of concern because in recent years DoS has become the most common and hard to defend against attacks. The novelty of this approach was to consider the way the DNS and ILNP can work together, transparently to the application, within an enterprise scenario. This was achieved by the introduction of a new application-level access control function - the Capability Management System (CMS) - which applies configuration at the application level (DNS data) and network level (ILNP namespaces). CMS provides dynamic, ephemeral identity and location information to clients and servers, in order to effectively partition legitimate traffic from attack traffic. This was achieved without modifying existing network components such as switches and routers and making standard use of existing functions, such as access control lists, and DNS servers, all within a single trust domain that is under the control of the enterprise. The prime objectives of this research were: • to defend against DoS attacks with the use of naming and DNS within an enterprise scenario. • to increase the attacker's effort in launching a successful DoS attack. • to reduce the visibility of vulnerabilities that can be discovered by an attacker by active probing approaches. • to practically demonstrate the effectiveness of ILNP and DNS working together to provide a solution for DoS mitigation. The solution methodology is based on the use of network and transport level capabilities, dynamic changes to DNS data, and a Moving Target Defence (MTD) paradigm. There are three solutions presented which use ILNP namespaces. These solutions are referred to as identifier-based, locator-based, and combined identifier-locator based solutions, respectively. ILNP-based node identity values were used to provide transport-level per-client server capabilities, thereby providing per-client isolation of traffic. ILNP locator values were used to allow a provision of network-level traffic separation for externally accessible enterprise services. Then, the identifier and locator solutions were combined, showing the possibility of protecting the services, with per-client traffic control and topological traffic path separation. All solutions were site-based solutions and did not require any modification in the core/external network, or the active cooperation of an ISP, therefore limiting the trust domain to the enterprise itself. Experiments were conducted to evaluate all the solutions on a test-bed consisting of off-the-shelf hardware, open-source software, an implementation of the CMS written in C, all running on Linux. The discussion includes considering the efficacy of the solutions, comparisons with existing methods, the performance of each solution, and critical analysis highlighting future improvements that could be made.

Managing data is one of the main challenges in distributed systems and computer science in general. Data is created, shared, and managed across heterogeneous distributed systems of users, services, applications, and devices without a clear and comprehensive data model. This technological fragmentation and lack of a common data model result in a poor understanding of what data is, how it evolves over time, how it should be managed in a distributed system, and how it should be protected and shared. From a user perspective, for example, backing up data over multiple devices is a hard and error-prone process, or synchronising data with a cloud storage service can result in conflicts and unpredictable behaviours. This thesis identifies three challenges in data management: (1) how to extend the current data abstractions so that content, for example, is accessible irrespective of its location, versionable, and easy to distribute; (2) how to enable transparent data storage relative to locations, users, applications, and services; and (3) how to allow data owners to protect data against malicious users and automatically control content over a distributed system. These challenges are studied in detail in relation to the current state of the art and addressed throughout the rest of the thesis. The artefact of this work is the Sea of Stuff (SOS), a generic data model of immutable self-describing location-independent entities that allow the construction of a distributed system where data is accessible and organised irrespective of its location, easy to protect, and can be automatically managed according to a set of user-defined rules. The evaluation of this thesis demonstrates the viability of the SOS model for managing data in a distributed system and using user-defined rules to automatically manage data across multiple nodes.

Understanding the brain and creating adaptable artificial intelligence systems are both at the forefront of science and engineering. Brains have the capacity to adapt in response to novel stimuli and efficiently encode vast amounts of information for later use through learning manifested as neural plasticity. While there are a range of known plasticity mechanisms, this work focuses on brain plasticity involving neurons preferentially connecting and disconnecting with each other in a process called structural synaptic plasticity. This process is known to occur in biology, however it is difficult to study experimentally due to the long time scales involved, therefore this work explores it through simulation. Structural plasticity involves complex changes to the morphology of neural cells with the end goal of altering the connectivity of neural circuits. Morphological changes are ignored to preserve computational tractability and the size of a synapse is equated with its efficacy; this is taken to be a measure of its stability. A model of structural plasticity is implemented on the SpiNNaker neuromorphic computation platform. It operates in real time and in conjunction with spike-timing dependent plasticity, and generates higher quality topographic maps compared to experiments not involving synaptic rewiring. These networks further benefit from the rewiring mechanism creating inhibitory connections that stabilise the spiking activity of the network. Further, the same model and architecture are used in the context of handwritten digit classification, and they show that the structural plasticity model is sufficient for neurons to learn the statistics of the input in an unsupervised fashion. Finally, neurons perform unsupervised motion classification through self-organisation. Neural responses become highly tuned and resemble those observed experimentally in Visual Cortex and Superior Colliculus: a highly asymmetric response evoked by motion. For this behaviour to arise, the networks are simulated for over 5 hours, modelling the development of these circuits from the point when no synapses exist until neurons reach their full synaptic capacity. SpiNNaker's real time operation and massive parallelism ensure simulation execution is tractable, and a sufficient number can be performed to obtain statistically significant results. The work described in this thesis aims to show that structural plasticity is an important and powerful learning mechanism which can now be incorporated seamlessly in Spiking Neural Network simulations executed on SpiNNaker.

The use of Web standards, compact publication guidelines, and open data initiatives have motivated many public and private organisations to publish data on the Web, giving rise to a global data space. Consuming data from heterogeneous data sources published on the Web requires integration at scale. The pay-as-you-go approach to data integration (PAYG) addresses integration at scale, relying on automatic techniques to provide candidate integrations. The high reliance on automatic techniques gives rise to uncertainty. Uncertainty may arise and propagate to all the tasks of the life cycle of a PAYG approach whose effect may be manifested in the quality of an automatically generated integration. Quantifying the uncertainty on the outcomes of a bootstrapped integration is a crucial task that can help in understanding the decisions made by the automatic algorithms, aiming to reduce such uncertainty that ultimately can improve the quality of an integration. In this thesis, we address the issue of quantifying the uncertainty that arises during the bootstrapping phase of PAYG in the context of Dataspaces. In particular, two approaches are proposed: (i) an approach to quantify the uncertainty in mapping generation using internal evidence; (ii) an approach to quantify the uncertainty on the quality of an entire integration using user feedback in a pay-as-you-go manner. More specifically, this thesis makes the following contributions: (i) a principled methodology to derive degrees of belief on mappings that builds on Bayesian infer- ence to assimilate evidence in the form of fitness scores associated to mappings during mapping generation; (ii) a novel methodology to quantify the uncertainty on the quality of an entire integration by assimilating user feedback on tuple results; (iii) an experi- mental evaluation of the proposed techniques on a real-world integration scenario. The experimental evaluation of the contributed techniques presented in this dis- sertation provides empirical evidence of their cost-effectiveness, when applied in syn- thetic and real-world scenarios, in quantifying the quality of a pay-as-you-go data in- tegration.

Parkinson's Disease is a neurodegenerative condition with no cure and a wide variety of idiosyncratic motor and non-motor symptoms that impact people's quality of life. In infrequent and short clinical consultations, health care professionals use validated clinical scales that are time-consuming and rely on clinical expertise or patients' perception. Moreover, symptoms can fluctuate within or between days due to the natural progression of the disease or medication and thus, a short session only provides a snapshot and not a longitudinal image of the disease. In practice, it is difficult for clinicians to tailor treatments and medication to each person's symptoms. Therefore, a granular, continuous, and objective methodology to track symptom fluctuations would improve patients' quality of life and make health services more efficient. The complexity of Parkinson's makes it a challenging but suitable study case to investigate new personalised approaches supported by technology and tailored to each patient's condition. This work aims to explore the use of behavioural inferences extracted from smartphone data to track the fluctuations in Parkinson's Disease symptoms. We recruited 7 participants with early Parkinson's (Hoehn & Yahr scale ≤ 3), instrumented their personal Android and iOS mobiles, and collected a dataset with up to 22 data sources, 24/7, during 222 to 380 days and within the technological, ethical, and UX limitations of a real-world deployment. As part of this dataset, we assessed their symptom severity using four clinical scales and seven cognitive tests every six weeks and patients self-reported the severity of their three main symptoms on a daily basis. Through an agile iterative design process, we generated knowledge that supports ground truth collection for longitudinal Parkinson's tracking. We identified five design implications for future Parkinson's self-reporting tools. We published PaperStream, an open-source tool to create and encode questionnaires or surveys automatically. Also, we created a paper diary that blends digital and analogue data collection which participants used to gather daily self-reported symptom severity scores with average answer compliance of 96.39% (SD=0.05) among 7 participants monitored between 222 and 380 days. We created an adaptative method to track weekly self-reported fluctuations in pain, gait and fatigue, the first one of its kind for Parkinson's Precision Medicine. Our method adapts to an individual's condition (personalised) and tracks symptoms while participants carry on with their regular lives (naturalistic), over time (longitudinal) and without people's input (unobtrusive). Instead of relying on the structure of the collected smartphone data to find patterns, our method starts with a generalised set of predictions based on mobility and activity recognition features that are overfitted to each patient's symptoms. We showed that this method tracks fluctuations better than chance with an agreement measured by Cohen's kappa between 0.19 and 0.53 finding a different subset of predictions for each participant. Also, exploring how this method can be deployed in the real world, we show that splitting our dataset into personalisation/testing sets is not a suitable strategy to identify a relevant subset of smartphone features to track symptoms on unseen data. Our results should be considered under the assumptions we made when we designed our method as well as the three primary sources of error in our methodology: the validity of the smartphone features, the confounders in participants behaviour, and the drawbacks of self-reporting. Nonetheless, this work is evidence that the unobtrusive and personalised monitoring methodology we proposed has the potential to track fluctuations in Parkinson's symptoms with the most impact on patients' daily life. Thus, we encourage researchers to tackle questions that remain open: how can we refine the mobility and activity recognition smartphone features we used? What other behavioural areas and smartphone features can be exploited to monitor other motor and non-motor symptoms? What other ground truth collection methods can minimise the problems related to analogue selfreported data? What is the clinical and self-management utility of the weekly fluctuations we tracked with our personalised models? What parameters and models of our method can be improved? Moreover, in the future, how can we include people with Parkinson's in the development and evaluation of their personalised models?

A classifier making predictions in noisy environments with limited data will inevitably make some mistakes. However, some types of error are more harmful than others. For example, the cost of misidentifying an explosive as a benign electronic device is far greater than the cost of misidentifying a laptop as an explosive. In such cases it is of paramount significance that the relative cost of different types of errors is taken into account when constructing and assessing algorithms for making predictions under uncertainty. This thesis will develop a distribution dependent theory of learning with asymmetric costs. For a comprehensive coverage we consider four different learning scenarios: (1) Supervised learning with a fixed cost matrix, (2) Weakly supervised cost-sensitive learning with unknown costs, (3) Sequential cost-sensitive learning with bandit feedback and (4) Neyman Pearson classification for qualitatively asymmetric costs. In each of these scenarios we seek to identify the key properties of the data distribution which influence the difficulty of the learning problem. These include the distribution's dimensionality, smoothness and the level of noise. We shall pay particular attention to minimax rates which give the minimum risk attainable by a classification algorithm, uniformly, over a given family of distributions. We begin our analysis with a supervised setting in which the costs of the possible prediction errors are given by a known cost matrix. In this setting we give a natural generalisation of Tysbakov's margin condition which quantifies the level of noise in the data in relation to the optimal decision boundaries for the cost matrix. We then identify the minimax rates with an exponent which depends solely upon the dimensionality of the distribution, the smoothness of the regression function and the level of noise. Significantly, these rates hold for all compact sub-manifolds and for any reasonable cost matrix. We then show that the minimax rates are actually attained by a simple cost-sensitive k-nearest neighbour method. Departing from the fully supervised setting we turn to a class of weakly supervised cost-sensitive learning problems with stochastic and feature dependent costs. We show that the minimax rates have the same exponent as their supervised counter parts and once again, these rates are attainable by simple k-nearest neighbour methods. We then move to an online scenario in which the learner sequentially makes a prediction and then incurs the cost of their prediction, without ever observing either the true class label or the costs that would have been incurred if an alternative prediction had been made. This problem is equivalent to the problem of multi-armed bandits with covariates. We present the conceptually simple k-NN UCB algorithm for this problem, which combines the k-NN method from supervised learning with the UCB algorithm for multi-armed bandits. We establish a minimax optimal regret bound for this algorithm (up to logarithmic terms) which demonstrates that the method is capable of adapting automatically to the unknown intrinsic dimensionality of the distribution. An alternative approach to learning with asymmetric costs is the Neyman-Pearson paradigm. In this paradigm the learner is given a set of samples from both a null and a positive class. The learner's goal is to construct a classifier which minimises the false negative rate whilst satisfying a user-specified upper bound on the false positive rate. This approach to classification is appropriate in applications such as medical diagnosis when the different types of errors are of a fundamentally different category. We present a non-parametric approach to Neyman-Pearson learning which side-steps the necessity of density estimation and applies when the data is distributed on an unknown manifold within a high-dimensional ambient feature space. We prove a high probability distribution-dependent performance bound. Finally, we present the first minimax lower bounds in the Neyman-Pearson setting. Overall this thesis presents a clear picture of the way in which the properties of the distribution affect the optimal learning rates for high-dimensional classification problems with asymmetric costs. Moreover, we demonstrate that these optimal rates are often attained by non-parametric techniques which are both conceptually simple and straightforward to implement.

Managers in complex organisations often have to make quick decisions on whether new information sharing developments are worth undertaking or not. Such decisions are hard to make, especially at an enterprise level. Both costs and risks are regularly underestimated. Existing approaches to managing risk and estimating cost are principally focused on creating detailed predictions based on substantial models of the planned development. They aim to support project managers throughout the development process itself, rather than giving a low-cost indicator for use in early-stage decision making. Our objective is to help managers and stakeholders of large, complex organisations, such as the National Health Service (NHS) in the UK, make better informed decisions on points of cost and risk of new software systems that will reuse or extend their existing information infrastructure, before any implementation is undertaken. We analysed 18 case studies describing recent software developments introduced by providers of health care services, looking for common points of high cost and risk. From the case studies analysis, we found that the movement of data within and between organisations was a key indicator of high cost and risk. Data movement can be hindered by numerous technical barriers, but also by other challenges arising from social aspects of an organisation. Hence, we devised a catalogue of socio-technical data movement anti-patterns that under certain conditions can introduce high cost and risk to the organisation. In this thesis, we propose a new method aiming to identify places of high cost and risk when existing data needs to move to a new development. The method is low-cost and combines both technical and social aspects, but relies only on information that is likely to be already known to key stakeholders, or will be cheap to acquire. The method is based on the data journey model, a new lightweight technique that captures movements of data within or between organisations. The data journey model describes an abstraction of large, complex eco-systems focusing on the high-level journeys data take through networks of people and systems. To assess the effectiveness of our method and the accuracy of our predictions, we applied the method in real world settings in the NHS domain. We worked with clinicians to model the movements of data in five NHS studies from different Foundation Trusts across the UK. The results of the evaluation showed that our method was able to cheaply and quickly identify most of the points of high cost/risk that the hospital staff had identified, along with several other possible directions that the staff did not identify for themselves, but agreed could be promising. Finally, the results of the evaluation showed that the data journey modelling method can be completed in less than a couple of hours (including training). Also, the simplicity of our modelling technique can empower domain experts with no particular modelling expertise to quickly identify opportunities for cost savings in new developments, as well as existing ones.