Search

Models of auditory salience aim to predict which sounds attract people's attention, and their proposed applications range from soundscape design to machine listening systems and object-based broadcasting. A few different types of models have been proposed, but one of the areas where most of them still fall short is spatial aspects of sound - they usually operate on mono signals and do not consider spatial auditory scenes. Part of the reason why this is the case might be that the relationship between auditory salience and position of sound is still not clear. In addition, methods used to measure auditory salience vary greatly, and authors in the field do not always use the same definition of salience. In Part I, this thesis aims to answer questions about the effect of spatial location of sound on auditory salience. This is done in four different experiments, which are based on previously published experimental methods but adapted to measure spatial effects. In general, the combined results of these experiments do not support the hypothesis that the spatial position of a sound alone influences how salient the sound is. However, they do show that unexpected changes in position might activate the deviance detection mechanism and therefore be salient. In addition, an experiment comparing three of the methods used reveals at least two dimensions of salience, which are measured by different methods to different extent. This emphasises the importance of carefully considering which experimental methods are used to measure auditory salience, and also providing a clear definition of what type of salience is of interest. Part II demonstrates how spatial position of sound can be incorporated into an auditory salience model. The results of experiments described in this thesis support the idea that the basis of auditory salience is the violation of expectations. The surprise caused by a sudden change in sound position can therefore be modelled by a Kalman-filter-based deviance detection model, which predicts experimental data discussed above with good accuracy. Finally, an example is given of how an application of such a model can improve the performance of a machine learning algorithm for acoustic event detection.

The amount of noise in web data is rapidly increasing, as is the number of users searching for information to suit their interest. The increase of web data has led to some critical issues, such as a high level of noise and irrelevant data. Given that the web is noisy, inconsistent and irrelevant by nature, finding useful information that defines interest of user has become a challenge. Existing research acknowledges that there is a need to propose machine learning tools capable of addressing problems with noise web data. Identifying and eliminating noise web data is critical to the web usage mining process. As the web evolves and more web data sources emerge, the level of noisiness also increases. Despite efforts by existing research to address noise in web data, a number of critical issues remain unresolved. For example, existing research work considers noise web data as irrelevant data that does not form part of the main content of a web page. Therefore, current machine learning tools focus on protecting the main content of a web page by eliminating noise/irrelevant data, such as advertisements, banners and external links etc. However, the main content of a web page can potentially be noise when user interests are considered. The position taken by the proposed research is based on the fact that noise web data can itself be useful when the interests of a user are considered prior to elimination. To justify this position, a Noise Web Data Learning (NWDL) approach which aims to learn noise web data prior to elimination is proposed. To the best of our knowledge, learning noise web data prior to elimination has not been addressed by current and relevant research works. The objective is to ensure that the interestingness of data on the web is defined by user interests over time. The proposed NWDL considers the following key aspects, 1) the significance of exit page in defining user interest level on web pages visited by a user. 2) The effect of the dynamic change of user interests towards the classification of web pages. Experiments conducted in this research shows that noise web data reduction process is user-centric, i.e., the dynamic changes of user interests influence the interestingness of web data. As a result, what is currently identified and eliminated as noise can be useful when user interests and their changes over time are considered. The findings from validation and evaluation of the proposed NWDL against existing tools shows that user interest over time significantly impacts the importance of data on the web. Given that current research work mainly identifies and eliminate noise web data without defining user interests, the process is not user-centric. A key contribution of the proposed research is to identify and learn noise web data taking into account user interests as they change over time prior to elimination. Ultimately the proposed NWDL contributes towards minimising loss of useful information otherwise considered as noise by the existing tool as well as reduce the level of noise data suggested to a user.

At present, the United Nations figures claim that the current world population would rise from 7.6 billion to 8.5 billion in 2030 and 9.7 billion in 2050. Therefore by 2050, 65 percent of the world's population would be living in urban mega-cities and each megacity would be accommodating around 10 million inhabitants. Such massive urbanization of growing population would be known as 21st century's 'Urban Age'. On the other side, by 2020 the growing population of elderly people above 65 years old would be increasing by 25 percent in EU countries and by 30 percent in other developing nations including Asia and North America. As a result, the growth of massive population and elderly inhabitants in urban cities would require an assisted living environment for independent and comfortable living experiences. As can be expected, a persuasive demand of assisted living environment would be vital to the humankind. The goal of an assisted living environment is to support the aging population and inhabitants to live independently in their own home and communities with the support of trained services and personal digital assistants. Therefore, the continuous growing demand of assisted living environment targets to improve the inhabitants comfort level and efficiency to do their ADL (Activity Daily Living) routine tasks by enabling the cooperation among various IoT smart objects and sensors which will understand the environmental surroundings and the inhabitant's contextual needs in a proactive manner. In this work, a Cognitive IoE (Internet of Everything) framework with ambient intelligence capability is proposed to observe the inhabitant activities with heterogeneous IoT network objects and sensors in a time series manner to perceive the inhabitant intentions and situations in the environment. The predictive regression model forecasts the inhabitant's activity patterns with regressive machine learning algorithms. The interconnected network objects (sensors and actuators) behave as agents to learn, think and adapt to contextual situations in the dynamic environment with no or minimum human intervention. Therefore, the first research challenge is to recognize the inhabitant's intentional-situation in the environment, and it is achieved by the Ambient Cognition Model(ACM). The ACM not only performs IoT data-fusion but also applies a statistical model for threshold and weight scheme to extract contextual information in a more systematic manner. The second research challenge of automating the predictive regression model to forecast the time series activity patterns of inhabitants is addressed within the Ambient-Expert Model(AEM). The hidden activity state patterns are identified, trained and tested with the supervised machine learning method of Hidden Markov Model, Recurrent-Neural Network, and Naive Bayes classifier. In addition, a recursive training mechanism of DATAWELL is integrated with the architecture to train(re-train) the model over new datasets and perform predictive analysis in a proactive manner. Furthermore, the unified framework CAiSH (Cognitive Ambient Intelligent Smart Home), built upon the integration of ACMand AEM architectures to a provide an intelligent IoT framework for the ambient intelligence smart home environment. The trained model uses maximum likelihood posterior probabilities to forecast the inhabitant's intentional activity states. The CAiSH works as a proactive digital assistant to the inhabitant provide a development platform for autonomous and enhanced assisted living services in the cognitive IoE environment. The research has been carried out on time-series data sets, deploying IoT lab to generate and collect time series data for the training and testing purpose and providing hands-on research experience on IoT prototype deployment. Overall, 5499 datasets of 30 SA (Spot-Activities) and 9 IA (Intention- Activities) data sets have been engaged for the training and evaluation. The result outputs are evaluated with MAE (Mean-Square Error), MAPE (Mean Absolute Percentage Error) and MAE (Mean Absolute Error) metrics for the prediction accuracy measures.

​Embedding methods for searching latent representations of the data are very important tools for unsupervised and supervised machine learning as well as information visualisation. Over the years, such methods have continually progressed towards the ability to capture and analyse the structure and latent characteristics of larger and more complex data. In this thesis, we examine the problem of developing efficient and reliable embedding methods for revealing, understanding, and exploiting the different aspects of the relational data. We split our work into three pieces, where each deals with a different relational data structure. In the first part, we are handling with the weighted bipartite relational structure. Based on the relational measurements between two groups of heterogeneous objects, our goal is to generate low dimensional representations of these two different types of objects in a unified common space. We propose a novel method that models the embedding of each object type symmetrically to the other type, subject to flexible scale constraints and weighting parameters. The embedding generation relies on an efficient optimisation despatched using matrix decomposition. And we have also proposed a simple way of measuring the conformity between the original object relations and the ones re-estimated from the embeddings, in order to achieve model selection by identifying the optimal model parameters with a simple search procedure. We show that our proposed method achieves consistently better or on-par results on multiple synthetic datasets and real world ones from the text mining domain when compared with existing embedding generation approaches. In the second part of this thesis, we focus on the multi-relational data, where objects are interlinked by various relation types. Embedding approaches are very popular in this field, they typically encode objects and relation types with hidden representations and use the operations between them to compute the positive scalars corresponding to the linkages' likelihood score. In this work, we aim at further improving the existing embedding techniques by taking into account the multiple facets of the different patterns and behaviours of each relation type. To the best of our knowledge, this is the first latent representation model which considers relational representations to be dependent on the objects they relate in this field. The multi-modality of the relation type over different objects is effectively formulated as a projection matrix over the space spanned by the object vectors. Two large benchmark knowledge bases are used to evaluate the performance with respect to the link prediction task. And a new test data partition scheme is proposed to offer a better understanding of the behaviour of a link prediction model. In the last part of this thesis, a much more complex relational structure is considered. In particular, we aim at developing novel embedding methods for jointly modelling the linkage structure and objects' attributes. Traditionally, link prediction task is carried out on either the linkage structure or the objects' attributes, which does not aware of their semantic connections and is insufficient for handling the complex link prediction task. Thus, our goal in this work is to build a reliable model that can fuse both sources of information to improve the link prediction problem. The key idea of our approach is to encode both the linkage validities and the nodes neighbourhood information into embedding-based conditional probabilities. Another important aspect of our proposed algorithm is that we utilise a margin-based contrastive training process for encoding the linkage structure, which relies on a more appropriate assumption and dramatically reduces the number of training links. In the experiments, our proposed method indeed improves the link prediction performance on three citation/hyperlink datasets, when compared with those methods relying on only the nodes' attributes or the linkage structure, and it also achieves much better performances compared with the state-of-arts.

Nowadays, powerful and easy to use editing software which is available to almost everyone allows forgers to create convincing digital forgeries. As multimedia applications require a certain level of trust in the integrity and authenticity of the data become more common, there is an increasing need to restore some of the lost trustworthiness of digital media. In multimedia forensics, Digital Signature and Digital Watermarking have long been commonly used in video authentication, but these methods have proven to have shortcomings. The main drawback of these techniques is that information must generally be inserted at the time of video capture or before video broadcasting. Both techniques require two stages are: at the sender side and then at the receiver side, which in some real world applications is not feasible. For the problem of source type identification, digital fingerprints are usually extracted and then compared with a dataset of possible fingerprints to determine the acquisition devices. Photo-Response Non-Uniformity (PRNU), which is caused by the different sensitivity of pixels to the light, has proven to be a distinctive link between the camera and its images/videos. With this in mind, this thesis proposes several new digital forensic techniques to detect evidence of manipulations in digital video content based on blind techniques (Chapter 4 and Chapter 5) where there is no need for per-embedded watermarks or per-generated digital signature. These methods showed potential to be reliable techniques in digital video authentication based on the local video information. For the problem of determining the source of digital evidence, this thesis proposes a G-PRNU method (in Chapter 3) that overcomes the accuracy obtained in PRNU method in the problem of digital videos source type identification and it is less computationally expensive. Each proposed method was tested on a dataset of videos and detailed experimental results are presented.

RDF is widely used in the Semantic Web for representing ontology data. Many real world RDF collections are large and contain complex graph relationships that represent knowledge in a particular domain. Such large RDF collections evolve as a consequence of their representation of the changing world. Evolution in Semantic Web content produces difference files (deltas) that track changes between ontology versions. These changes may represent ontology modifications or simply changes in application data. An ontology is typically expressed in a combination of OWL, RDFS and RDF knowledge representation languages. A data repository that represents an ontology may be large and may be duplicated over the Internet, often in the form of a relational data store. Although this data may be distributed over the Internet, it needs to be managed and updated in the face of such evolutionary changes. In view of the size of typical collections, it is important to derive efficient ways of propagating updates to distributed datastores. The deltas can be used to reduce the storage and bandwidth overhead involved in disseminating ontology updates. Minimising the delta size can be achieved by reasoning over the underlying knowledge base. OWL 2 is a development of the OWL 1 standard that incorporate new features to aid application construction. Among the sub languages of OWL 2, OWL 2 RL/RDF provides an enriched rule set that extends the semantic capability of the OWL environment. This additional semantic content can be exploited in change detection approaches that strive to minimise the alterations to be made when ontologies are updated. The presence of blank nodes (i.e. nodes that are neither a URI nor a literal) in RDF collections provides a further challenge to ontology change detection. This is a consequence of the practical problems they introduce when comparing data structures before and after an update. The contribution of this thesis is a detailed analysis of the performance of RDF change detection techniques. In addition, the work proposes a new approach to maintaining the consistency of RDF by using knowledge embedded in the structure to generate efficient update transactions. The evaluation of this approach indicates that it reduces the overall update size, at the cost of increasing the processing time needed to generate the transactions. In the light of OWL 2 RL/RDF, this thesis examines the potential for reducing the delta size by pruning the application of unnecessary rules from the reasoning process and using an approach to delta generation that produces a small number of updates. It also assesses the impact of alternative approaches to handling blank nodes during the change detection process in ontology structures. The results indicate that pruning the rule set is a potentially expensive process but has the benefit of reducing the joins over relational data stores when carrying out the subsequent inferencing.

This thesis is concerned with the fact that people routinely appropriate interactivetechnology. Much of the work in this project was conducted at The Public, aninteractive art gallery in West Bromwich. Examples of appropriation that arepresented range from interactive art, the game Minecraft™, to mundane objectsencountered in daily life. Research Questions posed in this study are: • What are the dynamics of appropriation? • What is the relationship of appropriation to affordance? • How do individuals experience appropriation? Appropriation is the mechanism by which we make objects in the world relevant and personal. This PhD has revealed three dimensions of appropriation namely: • Control: both in terms of ownership and virtuosity. • Ensoulment: the mechanism through which we ascribe personal significance to artefacts. • Affordance: the experiential relationship to artefacts concerned with action on and with them. Appropriation is revealed as a mechanism through which people understand potential action with technology. A traditional view is that people learn how to use a system and once its canonical use is established new uses or appropriations are discovered. What is revealed in this study is that appropriation is bound to our perception of action with technology, commonly explained through the concept of affordance. Appropriation is revealed as the initial act in human encounters with technology.

The distribution of digital multimedia, namely, audio, video, documents, images and Web pages is commonplace across today's Internet. The successful distribution of such multimedia, in particular video can be achieved using a number of proven architectures such as Internet Protocol Television (IPTV) and Over-The-Top (OTT) services. In order to maximise the scope and reach of this multimedia a need to combine aspects of the two architectures has arisen due to the rapid uptake of multimedia steaming to the plethora of Internet enabled devices that both architectures encompass. Content Delivery Networks (CDNs) have been proposed as an effective means to facilitate this unification in order to distribute multimedia in an efficient manner that enhances end-users' Web experience by replicating or copying content to edge of network locations in proximity to the end-user. However, CDNs often face resource over-provisioning, performance degradation and Service Level Agreement (SLA) violations, thus incurring high operational costs, hardware under-utilisation and a limited scope and scale of their services. The emergence of Cloud computing as a commercial reality has created an opportunity whereby Internet Service Providers (ISPs) can leverage their Cloud resources to disseminate multimedia. However, Cloud resource provisioning techniques can still result in over-provisioning and under-utilisation. To move beyond these shortcomings this thesis sets out to establish the basis for developing advanced and efficient techniques to enable the utilisation of Cloud-based resources in a highly scalable, and cost-effective manner that reduces over-provisioning and under-utilisation while minimising latency and therefore maintaining the QoS/QoE expected by end-users for streaming multimedia.

The objective of this work is object retrieval in large scale image datasets, where the object is specified by an image query and retrieval should be immediate at run time. Such a system has a wide variety of applications including object or location recognition, video search, near duplicate detection and 3D reconstruction. The task is very challenging because of large variations in the imaged object appearance due to changes in lighting conditions, scale and viewpoint, as well as partial occlusions. A starting point of established systems which tackle the same task is detection of viewpoint invariant features, which are then quantized into visual words and efficient retrieval is performed using an inverted index. We make the following three improvements to the standard framework: (i) a new method to compare SIFT descriptors (RootSIFT) which yields superior performance without increasing processing or storage requirements; (ii) a novel discriminative method for query expansion; (iii) a new feature augmentation method. Scaling up to searching millions of images involves either distributing storage and computation across many computers, or employing very compact image representations on a single computer combined with memory-efficient approximate nearest neighbour search (ANN). We take the latter approach and improve VLAD, a popular compact image descriptor, using: (i) a new normalization method to alleviate the burstiness effect; (ii) vocabulary adaptation to reduce influence of using a bad visual vocabulary; (iii) extraction of multiple VLADs for retrieval and localization of small objects. We also propose a method, SCT, for extremely low bit-rate compression of descriptor sets in order to reduce the memory footprint of ANN. The problem of finding images of an object in an unannotated image corpus starting from a textual query is also considered. Our approach is to first obtain multiple images of the queried object using textual Google image search, and then use these images to visually query the target database. We show that issuing multiple queries significantly improves recall and enables the system to find quite challenging occurrences of the queried object. Current retrieval techniques work only for objects which have a light coating of texture, while failing completely for smooth (fairly textureless) objects best described by shape. We present a scalable approach to smooth object retrieval and illustrate it on sculptures. A smooth object is represented by its imaged shape using a set of quantized semi-local boundary descriptors (a bag-of-boundaries); the representation is suited to the standard visual word based object retrieval. Furthermore, we describe a method for automatically determining the title and sculptor of an imaged sculpture using the proposed smooth object retrieval system.

The financial reporting world has recently faced a number of changes due to the impact of the Internet. Today, the revolution in business communication is accelerating and more data is being shared by a large number of participant users, aside from the company’s internal management, including: clients, business partners, financial market analysts, investors and government regulators. These changes have led to the development of eXtensible Business Reporting Language (XBRL), which is an opensource Internet-based financial reporting language. XBRL is an extension of eXtensible Markup Language (XML) that provides machinereadable tags for each individual data element in each financial statement. XBRL is likely to be used as a platform that offers universal standards for defining business information. XBRL can ease the preparation, analysis, and exchange of business information along each part of financial reporting supply chain and across companies around the world. It can also increase the efficiency for all related users of business data. This study has analysed the accuracy of XBRL outputs by conducting a literature review and by checking the accuracy of the real company XBRL filing submissions that are published publicly. This study has found that there were many errors in these public XBRL documents that were caused either through a few basic common errors or from mistakes in related financial information. Therefore, this study has aimed to discover any possible causes of errors in XBRL filings. It has also aimed to find a way to detect those errors. Consequently, this study conducted a semantic checking system that aimed to detect XBRL errors and so enhance the accuracy of financial statements. To develop the semantic checking system, the results of an error finding analysis were combined, filtered, and classified into each category of errors, including the integration of accounting, business, and technology knowledge to fulfil the system. A process flow for the semantic checking system was created to help understand both the method and the rule. The rules were then set up to determine the different aspect of errors, which had a different method to manage and reduce errors. The semantic checking system was created in terms of the information specification of the XBRL filings. The system was designed to be more practical for the users by presenting the relationship between the real data and accounting practice. Moreover, a prototype was produced and the case study method was applied to prove the development of the system. This step was able to ensure the accuracy of the semantic checking system. Finally, this semantic checking system has been shown to improve the accuracy of XBRL filings. It also emphasises the importance of employing XBRL preparers who are aware of all of the possible issues that may arise while preparing an XBRL-based filing submission.

Semantic Web data sources such as DBpedia are a rich resource of structured representations of knowledge about geographical features and provide potential data for computing the results of Question Answering System queries that require geo-spatial computations. Retrieval from these resources of all content that is relevant to a particular spatial query of, for example, containment, proximity or crossing is not always straightforward as the geometry is usually confined to point representations and there is considerable inconsistency in the way in which geographical features are referenced to locations. In DBpedia, some geographical feature instances have point coordinates, others have qualitative properties that provide explicit or implicit spatial relationships between named places, and some have neither of these. This thesis demonstrates that structured geo-spatial query, a form of question answering, on DBpedia can be performed with a hybrid query method that exploits quantitative and qualitative spatial properties in combination with a high quality reference geo-dataset that can help to support a full range of geo-spatial query operators such as proximity, containment and crossing as well as vague directional queries such as Find airports north of London?. A quantitative model based on the spatial directional relations in DBpedia has been used to assist in query processing. Evaluation experiments confirm the benefits of combining qualitative and quantitative methods for containment queries and of employing high-quality spatial data, as opposed to DBpedia points, as reference objects for proximity queries, particularly for linear features. The high quality geo-data also enabled answering questions impossible to answer with SemanticWeb resources alone, such as finding geographic features within some distance from a region boundary. The contributions were validated by a prototype geo-spatial query system that combined qualitative and quantitative processing and included ranking answers for directional queries based on models derived from DBpedia contributed data.

The primary objective of this thesis is to leverage on the framework of H∞ loop-shaping control to formulate efficient and powerful optimization algorithms in LMI framework for the synthesis of performance loop-shaping weights. The H∞ loop-shaping design procedure is an efficient controller synthesis technique that combines classical loop-shaping concepts with H∞ synthesis. This procedure establishes a good tradeoff between robust stability and robust performance of a closed-loop system in a systematic manner. However, the selection of pre- and/or post-compensators, a crucial step in the design procedure, is nontrivial as factors such as the right half plane poles/zeros of the nominal plant, roll-off rate around the crossover frequency, strength of cross-coupling in multi-input multi-output systems, expected bandwidth, etc. must be adequately considered.Firstly, a frequency-dependent weight optimization framework is formulated in state-space form in order to remove the dependency on frequency while retaining the objective of maximizing the robust stability margin of a closed-loop system. This formulation facilitates the synthesis of low-order controllers, which is desirable from an implementation perspective.A weight optimization framework that incorporates smoothness constraints in order to prevent the cancellation of important modes of the system, for example, lightly damped poles/zeros of flexible structures, is subsequently formulated. The proposed formulation is intuitive from a design perspective as the smoothness constraints are expressed as gradient constraints on a log-log scale in dB/decade, consistent with the notation used in Bode plot for single-input single-output systems and singular value plots for multi-input multi-output systems.Thereafter, an optimization framework that maximizes the robust performance of a closed-loop system is presented. The philosophy in this framework is in line with practical design objectives that give the best achievable robust performance on a particular problem once a level of robust stability margin is demanded.Lastly, a novel unmanned vehicle is proposed. The vehicle uses a full six-degree-of-freedom tri-rotor actuation, capable of fully decoupled thrust and torque vectoring in all the 3D space. This vehicle can act as an unmanned ground vehicle or unmanned aerial vehicle, but the objective herein is restricted to the upright stability of the vehicle while operating on the ground as this is a precursor to rolling motion. The full nonlinear model of the vehicle is derived and linearized for subsequent controller synthesis, and this is thereafter validated by means of numerical simulations.

This thesis investigates the use of synthetic talking heads, with lip, tongue and face movements synchronized with synthesized or natural speech, in technology-enhanced learning. This work applies talking heads in a speech tutoring application for teaching English as a second language. Previous studies have shown that speech perception is aided by visual information, but more research is needed to determine the effectiveness of visualization of articulators in pronunciation training. This thesis explores whether or not visual speech technology can give an improvement in learning pronunciation. This thesis investigates techniques for audiovisual speech synthesis, using both viseme-based and data-driven approaches to implement multiple talking heads. Intelligibility studies found the audiovisual heads to be more intelligible than audio alone, and the data-driven head was found to be more intelligible than the viseme-driven implementation. The talking heads are applied in a pronunciation-training application, which is evaluated by second-language learners to investigate the benefit of visual speech in technology-enhanced learning. User trials explored the efficacy of the software in demonstrating the /b/–/p/ contrast in English. The results indicate that learners showed an improvement in listening and pronunciation after using the software, while the benefit of visualization compared to auditory training alone varied between individuals. User evaluations found that the talking heads were perceived to be helpful in learning pronunciation, and the positive feedback on the tutoring system suggests that the use of talking heads in technology-enhanced learning could be useful in addition to traditional methods.

Extraction, representation, organisation and application of metadata about audio recordings are in the concern of semantic audio analysis. Our broad interpretation, aligned with recent developments in the field, includes methodological aspects of semantic audio, such as those related to information management, knowledge representation and applications of the extracted information. In particular, we look at how Semantic Web technologies may be used to enhance information management practices in two audio related areas: music informatics and music production. In the first area, we are concerned with music information retrieval (MIR) and related research. We examine how structured data may be used to support reproducibility and provenance of extracted information, and aim to support multi-modality and context adaptation in the analysis. In creative music production, our goals can be summarised as follows: O↵-the-shelf sound editors do not hold appropriately structured information about the edited material, thus human-computer interaction is inefficient. We believe that recent developments in sound analysis and music understanding are capable of bringing about significant improvements in the music production workflow. Providing visual cues related to music structure can serve as an example of intelligent, context-dependent functionality. The central contributions of this work are a Semantic Web ontology for describing recording studios, including a model of technological artefacts used in music production, methodologies for collecting data about music production workflows and describing the work of audio engineers which facilitates capturing their contribution to music production, and finally a framework for creating Web-based applications for automated audio analysis. This has applications demonstrating how Semantic Web technologies and ontologies can facilitate interoperability between music research tools, and the creation of semantic audio software, for instance, for music recommendation, temperament estimation or multi-modal music tutoring.

In this work. we propose a unified framework called Kalman filter based Radial Basis Functions (KF-RBF) for online functional prediction based on the Radial Basis Functions and the Kalman Filter. The data are nonstationary spatio-ternporal observations irregularly sampled in the spatial domain. We shall assume that a Functional Auto-Regressive (FAR) model is generating the system dynamics. Therefore. to account for the spatial variation. a Radial Basis Function (RBF) network is fitted to the spatial data at every time step. To capture the temporal variation, the regression surfaces arc allowed to change with time. This is achieved by proposing a linear state space model for the RBF weight vectors to evolve temporally. With a fixed functional basis in expressing all regressions. the FAR model call then he re-formulated as a Vector Auto-Regressive (VAR) model embedded in a Kalman Filter. Therefore functional predictions. normally taken place in the Hilbert space. can now be easily implemented 011 a computer. The advantages of our approach are as follows. First it is computationally simple: using the KF. we can obtain the posterior and predictive distributions in closed form. This allows for quick implementation of the model. and provides for full probabilistic inference for the forecasts. Second, the model requires no restrictive assumptions such as stationarity. isotropy or separability of the space/time correlation functions. Third. the method applies to non-lattice data. in which the number and location of sensors can change over time. This framework proposed is further extended by generalizing the real-valued. scalar weights in the functional autoregressive model to operators ill the Reproducing Kernel Hilbert Space (RKHS). This essentially implies that a larger. more intricate class of functions can be represented by this functional autoregressive approach. In other words. the unknown function is expressed as a sum of transformed functions mapped from the past functions in the RKHS. This bigger class of functions can potentially yield a better candidate that is "closer". in the norm sense. to the unknown function. In our research. the KF is used despite the system and observational noise covariance are both unknown. These uncertainties may significantly impact the filter performance. resulting in sub- optimality or divergence. A multiple-model strategy is proposed in view of this. This is motivated by the Interactive Multiple Model (IMM) algorithm in which a collection of filters with different noise characteristics is run in parallel. This strategy avoids the problems associated with the estimation of the noise covariance matrices. Furthermore. it also allows future measurements to be predicted without the assumption of time stationarity of the disturbance terms.

The Flux switching motor is a new class of reluctance machine that has demonstrated potential as a possible replacement for brushed-dc motor in many applications. However the design and optimisation of the motor and its drive system are rather complicated and not much past knowledge and guidelines are available to aid the engineer(s) in the design of the machine.;The development of flexible and versatile design optimisation software to facilitate the design and optimisation of FS motor and drive is presented. The design optimisation software incorporates a genetic algorithm optimisation tool and dynamic simulation model with third party finite element analysis software.;The developed genetic algorithm optimisation program integrated with finite element analysis software provides the engineer with the necessary optimisation tools capable of interfacing with the FEA software. This has allowed many FSM lamination designs to be created without any requirement of user feedback once the program is initialised. In addition the application of the developed design tool can also be extended to other electromagnetic devices.;A dynamic simulation model of the FS motor drive system has been developed. The model can either be used as a standalone program or be integrated into the optimisation software. The dynamic simulation model consisted of a simple time-stepping electrical equivalent circuit coupled with a switch control algorithm, a winding optimisation model and an iron loss model. When interfaced with the FEA software it can support rapid estimation of the motor dynamic performance. The developed optimisation software has been used to design and optimise FS motors and the results have demonstrated the potential of genetic algorithms in design optimisation of the machine.

This thesis investigates stereo vision based techniques for supporting work with teleoperated mobile robots. In particular, the tasks of navigation and control of the robot in unknown environments as well as control of the manipulator are considered. Support is required with map overviews and images from novel viewpoints. Challenges are to create maps and novel images in real time, exclusively from image sequences of a calibrated stereo camera that is allowed to move arbitrarily in three dimensions. Furthermore, no assumptions are made about the environment. These requirements have led to research in the areas of stereo vision, camera motion estimation, mapping and novel view synthesis. Several contributions have been made in these areas. Firstly, the behaviour of stereo correlation is analysed and a new real time stereo algorithm derived, which has reduced matching errors compared to traditional algorithms. Secondly, a new robust real time camera motion estimation method is described, which exclusively uses stereo images and permits arbitrary camera motion. Thirdly, a new method for creating two-dimensional maps from images of a stereo camera under arbitrary three-dimensional motion is presented. Finally, novel view synthesis is performed in disparity space of source stereo images using a new method. All the proposed techniques have been integrated into the Immediate Reality Scanner (IRIS) system. The system performs incremental mapping, immediate virtual walkthroughs and dynamic novel views exclusively from a stereo camera under arbitrary three-dimensional motion. Mapping and immediate virtual walkthroughs can be performed in real time, concurrently to scanning the environment. The thesis describes camera calibration, stereo processing, camera motion estimation, map building, novel view synthesis and the integration into IRIS, including evaluations in all areas. It is concluded that IRIS does not only fulfil the requirements of the anticipated application on teleoperated mobile robots, but could also be used in a wide range of robotics and non-robotics applications. It is believed that IRIS is the first system that allows incremental mapping and immediate virtual walkthroughs in real time concurrently to scanning an unconstrained environment with an arbitrarily moving stereo camera

The emergence of platform-independent, mobile code technologies has created big opportunities for Internet-based applications. Mobile agents are being utilized to perform a variety of tasks from personalized computing to business-critical transactions. Unfortunately, these advances were not matched by correspondent research into the reliability of these new technologies. This work has been undertaken to investigate the faulttolerance of this new paradigm. Agent programs' mobility and autonomy of execution has introduced a new class of failures different to that of traditional distributed systems. Therefore, fault tolerance is one of the main problems that must be resolved to improve the adoption of an agents' paradigm. The investigation of mobile agents reliability in this thesis resulted in the development of REMA (REliable Mobile Agents), which guarantees the reliable execution, migration, and communication of mobile agents in the presence of faults that might affect the agents hosts or their communication network. We introduced an algorithm for the transparent detection of faults that might affect agent execution, migration, and communication. A decentralized structure was used to divide the agent dynamic distributed system into network-partitioning proof spaces. Lightweight messaging was adopted as the basic error detection engine, which together with the loosely coupled detection managers provided an efficient, low overhead detection mechanism for agent-based distributed processing. The problem of taking checkpoint of agent execution is hampered by the lack of the accessibility of the underlying structure of the JVM. Thus, an alternative solution has been achieved through the REMA Checkpoint and Recovery (REMA-CR) package. REMA-CR provides the developer with powerful classes and methods that allow for capturing the critical data of agents' execution. The developed recovery protocol offers a communication-pairs, independent checkpointing strategy at a low-cost, that covers all possible faults that might invalidate reliable agent execution, migration and communication and maintains the exactly once execution property. The results and the performance of REMA confirmed our objectives of providing a fault tolerant wrapper for agents and their applications with acceptable overhead cost.

This thesis describes research into automatic knowledge extraction (AKE) from text. In particular, the automatic knowledge extraction is to produce or help to produce knowledge in the format of an existing hypermedia knowledge-based system: HyperTutor. The whole AKE process is divided into two main stages: concept extraction and relation extraction. Automatic thematic concept (keyword) extraction (TCE) is described in detail. Two approaches for TCE are presented and evaluated. One of them is a machine learning approach based on artificial neural networks (ANNs). New measures of evaluation of this novel approach are introduced, based on the concept of generalisation. These include natural generalisation (NG) and pure generalisation (PG). Measures commonly used in knowledge extraction research, i.e. recall and precision, are applied in their normal binary form, but analogue versions are developed to assess the performance of the ANN-based approach. A comparison with chance (CWC) measure is also applied to the results. A stemming analysis method has also been attempted at sense-level and word-level. The results show that thematic concepts can be automatically extracted from text using an ANN plus a lexical semantic resource. The ANN alone produces best result for non-keywords and overall. Word level stemming analysis alone is the best for identifying keywords, while sense level analysis provides the most balanced results between keywords and non-keywords. The baseline comparison for the ANN method shows that the ANN method adds value to the external lexicon. The CWC measure shows that both the ANN and stemming methods work much better than chance. Domain portability of the keyword extraction techniques developed is addressed. Although the ANN itself in not transferable, the ANN method is transferable with consistent performance between domains. The stemming analysis approach also transfers well between domains, although not as well as the ANN method. An attempt to answer the question of how the ANN learns the problem of thematic concept extraction is also presented, based on the analysis of the weights in the trained ANN. Analysis is carried out on three different aspects: relation level analysis tries to find out if some kind of relations are more important than others, or if they are equally important to the ANN; path level analysis aims to identify what kinds of paths are more likely to lead to a noun being classified as a keyword; and analysis of category information helps to explain how category affects keyword recognition. This analysis has confirmed the hypothesis of close relationships that are distinct and characteristic of the seed word-KW relationship. An important type of relation, named verb-noun relation, is targeted in the attempt of relation extraction. This is novel. Parse tree based and tagger based approaches have been investigated. The parse tree based method produces high precision and low recall. The main reason for the low recall is parser failure. This reflects the current limitation of the parser techniques is not advanced enough to process texts in the real world. The tagger-based approach produces high recall and low precision. The style of writing may have great impact on the results of relation extraction. The experiments have shown that both of the approaches perform better for one of the documents compared with the other. This thesis also proposes issues for future work in AKE research.

Handwritiiig recognition has been the subject of intensive research for many years. However, despite the best effort of many researchers, the problem of handwriting recognition is far from being solved. The greatest difficulty in cursive script recognition is due to the large variation of shapes that can result from the different writing styles. A common way to address this problem is to accommodate the variability in the feature set. However, such systems are limited in the range of writing styles that they can successfully deal with. An alternative approach has been to minimise the variability within the handwriting itself. Techniques such as normalisation, slant correction, restricting the number of objects to be recognised (i.e. numerical character, name of city) etc. have been shown to be partially effective. However further work remains to be done in order to cope with the variation problem. Here it is hypothesised that a pre-classification of writer style would provide an effective means of managing style variation and hence achieve better recognition results. The main aim of this thesis is to investigate alternative ways of addressing problems brought about by the variability of human handwriting; in particular those problems related to the recognition of off-line cursive handwriting. Style has been further broken down into case and quality of handwriting. Case classification of handwriting is proposed as a means of limiting the size of the template database used for word recognition. The quality of handwriting has been defined in terms of its legibility. It is proposed that this approach would lead to determining the legibility of an unknown sample prior to recognition. So as to select a recogniser that is suited to the quality of handwriting of the unknown sample. Two non-parametric classification techniques are applied to features extracted from the word image contours in order to compare their effectiveness in classifying words into upper, lower and mixed cases and further into legible, illegible and middle (between legible and illegible) classes. In the first method, a Multiple Discriminant Analysis (MDA) is used to transform the space of the extracted feature (36 dimensions) into an optimal discriminate space for a nearest mean based classifier. In the second method, a Probabilistic Neural Network (PNN) based on Bayes strategy and non-parametric estimation of probability density function is used. The experimental results show that PNN gives superior classification results when compared to MDA for both types of style classification. A number of experiments have been carried out using unseen data to determine the effectiveness of the above techniques. For a two-class word case classification problem the PNN approach yields 100% (lower/upper), 88%(upper/mixed) and 81%(lower/mixed) correct classification. For three-class word case classification the rate of correct classification is 73%. The same approach when applied to legible, illegible and middle style classification handwriting provides 86.5% (legible/illegible), 75.5% (legible/middle) and 90.5% (middle/illegible) correct classification for two classes. For three-class legibility classification the rate of correct classification is 67.33%. Style variation remains an open subject for further research. Word case and legibility are demonstrated to provide positive steps towards a more tangible definition of style. This research has demonstrated that a holistic classification technique is effective in dealing with the concept of style in a quantifiable manner. The experimental results indicate that further word level features are needed to further improve classification. This together with additional style categories would lead to more effective means of managing variability.

An adaptive test is one in which the number of test items and the order in which the items are presented are computed during the delivery of the test so as to obtain an accurate estimate of a student's knowledge, with a minimum number of test items. This thesis is concerned with the design and development of computerised adaptive tests for use within educational settings. Just as, in the same setting, intelligent tutoring systems are designed to emulate human tutors, adaptive testing systems can be designed to mimic effective informal examiners. The thesis focuses on the role of adaptive testing in student modelling, and demonstrates the practicality of constructing such tests using expert emulation. The thesis makes the case that, for small scale adaptive tests, a construction process based on the knowledge acquisition technique of expert systems is practical and economical. Several experiments in knowledge acquisition for the construction of an adaptive test are described, in particular, experiments to elicit information for the domain knowledge, the student model and the problem progression strategy. It shows how a description of a particular problem domain may be captured using traditional techniques that are supported by software development in the constraint logic extension to Prolog. It also discusses knowledge acquisition techniques for determining the sequence in which questions should be asked. A student modelling architecture called SKATE is presented. This incorporates an adaptive testing strategy called XP, which was elicted from a human expert. The strategy, XP, is evaluated using simulations of students. This approach to evaluation facilitates comparisons between approaches to testing and is potentially useful in tuning adaptive tests.

Since the agent paradigm emerged, agent researchers have faced the challenge of building open societies in which heterogeneous and independently designed entities can work towards similar or different ends. Open societies involve agents that do not necessarily share the same interests, that do not know and might not trust each other, but that can work together and help each other. One of the key omissions in the computational representation of open societies relates to the need for norms in multi-agent systems, that help to cope with the heterogeneity, the autonomy and the diversity of interests among their members. This also requires agents that can reason about norms because their participation in a society, rather than predefined, must be voluntary. So, these agents must understand why norms should be adopted and complied with, and why the authority and the power of agents in a society must be respected. This thesis addresses both the introduction of norms in systems of autonomous agents, and the modelling of agents that can reason about norms. The thesis makes three main contributions. First, it develops a framework of normative concepts that enables agents to reason about norms and the society in which they participate. Second, it provides the means for agents to identify situations of power and to use these powers both for the satisfaction of their goals and to understand why the goals of other agents must be satisfied. This is required since agents in an open society must interact with other agents which are also autonomous, and power represents a means to influence them. Third, this thesis provides models for agents that adopt and comply with norms not as an end, but as the result of a deliberation process in which if their goals and motivations are taken into account. This enables agents to voluntarily decide whether participating in a society is important for the achievement of their goals.

This thesis explores the behaviour-oriented nature of intelligence and presents the definition and use of Fuzzy Behavioural Maps (FBMs) as a flexible development framework for providing complex autonomous agent behaviour. This thesis provides a proof-of-concept for simple FBMs, including some experimental results in Mobile Robotics and Fuzzy Logic Control. This practical work shows the design of a collision avoidance behaviour (of a mobile robot) using a simple FBM and, the implementation of this using a Fuzzy Logic Controller (FLC). The FBM incorporates three causally related sensorimotor activities (moving around, perceiving obstacles and, varying speed). This Collision Avoidance FBM is designed (in more detail) using fuzzy relations (between levels of perception, motion and variation of speed) in the form of fuzzy control rules. The FLC stores and manipulates these fuzzy control (FBM) rules using fuzzy inference mechanisms and other related implementation parameters (fuzzy sets and fuzzy logic operators). The resulting FBM-FLC architecture controls the behaviour patterns of the agent. Its fuzzy inference mechanisms determine the level of activation of each FBM node while driving appropriate control actions over the creature's motors. The thesis validates (demonstrates the general fitness of) this control architecture through various pilot tests (computer simulations). This practical work also serves to emphasise some benefits in the use of FLC techniques to implement FBMs (e.g. flexibility of the fuzzy aggregation methods and fuzzy granularity).More generally, the thesis presents and validates a FBM Framework to develop more complex autonomous agent behaviour. This framework represents a top-down approach to derive the BB models using generic FBMs, levels of abstraction and refinement stages. Its major scope is to capture and model behavioural dynamics at different levels of abstraction (through different levels of refinement). Most obviously, the framework maps some required behaviours into connection structures of behaviour-producing modules that are causally related. But the main idea is following as many refinement stages as required to complete the development process. These refinement stages help to identify lower design parameters (i.e. control actions) rather than linguistic variables, fuzzy sets or, fuzzy inference mechanisms. They facilitate the definition of the behaviours selected from first levels of abstraction. Further, the thesis proposes taking the FBM Framework into the implementation levels that are required to build BB control architecture and provides and application case study. This describes how to develop a complex, non-hierarchical, multi-agent behaviour system using the refinement capabilities of the FBM Framework. Finally, the thesis introduces some more general ideas about the use of this framework to cope with some, current complexity issues around the behaviour-oriented nature of intelligence.

Advances in digital technology in the last decade have motivated the development of very efficient and high quality speech compression algorithms. While in the early low bit rate coding systems, the main target was the production of intelligible speech at low bit rates, expansion of new applications such as mobile satellite systems increased the demand for reducing the transmission bandwidth and achieving higher speech quality. This resulted in the development of efficient parametric models for speech production system. These models were the basis of powerful speech compression algorithms such as CELP, MBE, MELP and WI. The performance of a speech coder not only depends on the speech production model employed but also on the accurate estimation of speech parameters. Periodicity, also known as pitch, is one of the speech parameters that greatly affect the synthesised speech quality. Thus, the subject of pitch determination has attracted much research in the area of low bit rate coding. In these studies it is assumed that for a short segment of speech, called frame, the pitch is fixed or smoothly evolving. The pitch estimation algorithms generally fail to determine irregular variations, which can occur at onset and offset speech segments. In order to overcome this problem, a novel preprocessing method, which detects irregular pitch variations and modifies the speech signal such as to improve the accuracy of the pitch estimation, is proposed. This method results in more regular speech while maintaining perceptual speech quality. The perceptual quality of the synthesised speech may also be improved using postfiltering techniques. Conventional postfiltering methods generally consider the enhancement of the whole speech spectrum. This may result in the broadening of the first formant, which leads to the increase of quantisation noise for this formant. A new postfiltering technique, which is based on factorising the linear prediction synthesis filter, is proposed. This provides more control over the formant bandwidth and attenuation of spectral speech valleys. Key words: Pitch smoothing, speech pre-processor, postfiltering.