In this thesis, a number of problems are looked at predominately in the area of mobile agent communication protocols. Particularly, with the recent uptake of unstructured, large and dynamic networks there is a demand for cheap, ubiquitous and reliable protocols that will assist in the support of the network through tasks such as information dissemination, information search and retrieval, and network monitoring. One of the exciting new possible solutions to this is in the area of mobile agents where by a team of dedicated agents (physical or purely virtual) act independently or within a team on the network in the goal of solving simple, yet highly valuable, tasks. Independent agents may work within the bounds of their environment and without influence from their colleagues to solve simple or potentially complex tasks. Further, the agents may also act independently but in actuality work as a very complex team with simple underlying principles allowing for large scale networks and problems to be handled autonomously. In this work, two problems are investigated in detail, the Rendezvous Problem and Network Patrolling. In the rendezvous problem, the goal is to identify algorithms that will permit two agents to rendezvous within some known or unknown environment. This problem, while at first can feel trivial, it can be incredibly difficult when it is realised that all agents are expected to be identical at wake-up time, and that methods must be found that allow for the breaking of symmetry. This thesis provides an investigation into a number of models and environments and tries to provide optimal algorithms for allow rendezvous to occur. Secondly in the area of network patrolling, this work investigates the problem where there exists an environment, in which parts of it are viewed as being vital to network integrity and as such must be monitored. When there are less agents than vital regions the challenge of identifying traversal routes that minimise idle time becomes apparent. In this work an algorithm is presented that minimises this in the ring environment. Finally, this work also looks at other problems in distributed computing and provides exploratory foundation work that could provide alternative models for routing problems, distributed processing, community identification, and presents a number of open problems.

Medical image analysis and recognition is one of the most important tools in modern medicine. Different types of imaging technologies such as X-ray, ultrasonography, biopsy, computed tomography and optical coherence tomography have been widely used in clinical diagnosis for various kinds of diseases. However, in clinical applications, it is usually time consuming to examine an image manually. Moreover, there is always a subjective element related to the pathological examination of an image. This produces the potential risk of a doctor to make a wrong decision. Therefore, an automated technique will provide valuable assistance for physicians. By utilizing techniques from machine learning and image analysis, this thesis aims to construct reliable diagnostic models for medical image data so as to reduce the problems faced by medical experts in image examination. Through supervised learning of the image data, the diagnostic model can be constructed automatically. The process of image examination by human experts is very difficult to simulate, as the knowledge of medical experts is often fuzzy and not easy to be quantified. Therefore, the problem of automatic diagnosis based on images is usually converted to the problem of image classification. For the image classification tasks, using a single classifier is often hard to capture all aspects of image data distributions. Therefore, in this thesis, a classifier ensemble based on random subspace method is proposed to classify microscopic images. The multi-layer perceptrons are used as the base classifiers in the ensemble. Three types of feature extraction methods are selected for microscopic image description. The proposed method was evaluated on two microscopic image sets and showed promising results compared with the state-of-art results. In order to address the classification reliability in biomedical image classification problems, a novel cascade classification system is designed. Two random subspace based classifier ensembles are serially connected in the proposed system. In the first stage of the cascade system, an ensemble of support vector machines are used as the base classifiers. The second stage consists of a neural network classifier ensemble. Using the reject option, the images whose classification results cannot achieve the predefined rejection threshold at the current stage will be passed to the next stage for further consideration. The proposed cascade system was evaluated on a breast cancer biopsy image set and two UCI machine learning datasets, the experimental results showed that the proposed method can achieve high classification reliability and accuracy with small rejection rate. Many computer aided diagnosis systems face the problem of imbalance data. The datasets used for diagnosis are often imbalanced as the number of normal cases is usually larger than the number of the disease cases. Classifiers that generalize over the data are not the most appropriate choice in such an imbalanced situation. To tackle this problem, a novel one-class classifier ensemble is proposed. The Kernel Principle Components are selected as the base classifiers in the ensemble; the base classifiers are trained by different types of image features respectively and then combined using a product combining rule. The proposed one-class classifier ensemble is also embedded into the cascade scheme to improve classification reliability and accuracy. The proposed method was evaluated on two medical image sets. Favorable results were obtained comparing with the state-of-art results.

Finding easier and less resource-intensive ways of building knowledge resources is necessary to help broaden the coverage and use of semantic web technologies. Crowdsourcing presents a means through which knowledge can be efficiently acquired to build semantic resources. Crowds can be identified that represent communities whose knowledge could be used to build domain ontologies. This work presents a knowledge acquisition approach aimed at incorporating ontology engineering tasks into community crowd activity. The success of this approach is evaluated by the degree to which a crowd consensus is reached regarding the description of the target domain. Two experiments are described which test the effectiveness of the approach. The first experiment tests the approach by using a crowd that is aware of the knowledge acquisition task. In the second experiment, the crowd is unaware of the knowledge acquisition task and is motivated to contribute through the use of an interactive map. The results of these two experiments show that a similar consensus is reached from both experiments, suggesting that the approach offered provides a valid mechanism for incorporating knowledge acquisition tasks into routine crowd activity.

Very often data relevant to one search is not located at a single site, it may be widely-distributed and in many different forms. Similarly there may be a number of algorithms that may be applied to a single Knowledge Discovery in Databases (KDD) task with no obvious “best” algorithm. There is a clear advantage to be gained from a software organisation that can locate, evaluate, consolidate and mine data from diverse sources and/or apply a diverse number of algorithms. Multi-agent systems (MAS) often deal with complex applications that require distributed problem solving. Since MAS are often distributed and agents have proactive and reactive features, combining Data Mining (DM) with MAS for Data Mining (DM) intensive applications is therefore appealing. This thesis discusses a number of research issues concerned with the viability of Multi-Agent systems for Data Mining (MADM). The problem addressed by this thesis is that of investigating the usefulness of MAS in the context of DM. This thesis also examines the issues affecting the design and implementation of a generic and extendible agent-based data mining framework. The principal research issues associated with MADM are those of experience and resource sharing, flexibility and extendibility, and protection of privacy and intellectual property rights. To investigate and evaluate proposed solutions to MADM issues, an Extendible Multi-Agent Data mining System (EMADS) was developed. This framework promotes the ideas of high availability and high performance without compromising data or DM algorithm integrity. The proposed framework provides a highly flexible and extendible data-mining platform. The resulting system allows users to build collaborative DM approaches. The proposed framework has been applied to a number of DM scenarios. Experimental tests on real data have confirmed its effectiveness.

Argument systems are computational models that enable an artificial intelligent agent to reason via argumentation. Basically, the computations in argument systems can be viewed as search problems. In general, for a wide range of such problems existing algorithms lack five important features. Firstly, there is no comprehensive study that shows which algorithm among existing others is the most efficient in solving a particular problem. Secondly, there is no work that establishes the use of cost-effective heuristics leading to more efficient algorithms. Thirdly, mechanisms for pruning the search space are understudied, and hence, further pruning techniques might be neglected. Fourthly, diverse decision problems, for extended models of argument systems, are left without dedicated algorithms fine-tuned to the specific requirements of the respective extended model. Fifthly, some existing algorithms are presented in a high level that leaves some aspects of the computations unspecified, and therefore, implementations are rendered open to different interpretations. The work presented in this thesis tries to address all these concerns. Concisely, the presented work is centered around a widely studied view of what computationally defines an argument system. According to this view, an argument system is a pair: a set of abstract arguments and a binary relation that captures the conflicting arguments. Then, to resolve an instance of argument systems the acceptable arguments must be decided according to a set of criteria that collectively define the argumentation semantics. For different motivations there are various argumentation semantics. Equally, several proposals in the literature present extended models that stretch the basic two components of an argument system usually by incorporating more elements and/or broadening the nature of the existing components. This work designs algorithms that solve decision problems in the basic form of argument systems as well as in some other extended models. Likewise, new algorithms are developed that deal with different argumentation semantics. We evaluate our algorithms against existing algorithms experimentally where sufficient indications highlight that the new algorithms are superior with respect to their running time.

This thesis proposes the novel concept of Vertex Unique Labelled Subgraph (VULS) mining with respect to the field of graph-based knowledge discovery (or graph mining). The objective of the research is to investigate the benefits that the concept of VULS can offer in the context of vertex classification. A VULS is a subgraph with a particular structure and edge labelling that has a unique vertex labelling associated with it within a given (set of) host graph(s). VULS can describe highly discriminative and significant local geometries each with a particular associated vertex label pattern. This knowledge can then be used to predict vertex labels in 'unseen' graphs (graphs with edge labels, but without vertex labels). Thus this research is directed at identifying (mining) VULS, of various forms, that 'best' serve to both capture effectively graph information, while at the same time allowing for the generation of effective vertex label predictors (classifiers). To this end, four VULS classifiers are proposed, directed at mining four different kinds of VULS: (i) complete, (ii) minimal, (iii) frequent and (iv) minimal frequent. The thesis describes and discusses each of these in detail including, in each case, the theoretical definition and algorithms with respect to VULS identification and prediction. A full evaluation of each of the VULS categories is also presented. VULS has wide applicability in areas where the domain of interest can be represented in the form of some sort of a graph. The evaluation was primarily directed at predicting a form of deformation, known as springback, that occurs in the Asymmetric Incremental Sheet Forming (AISF) manufacturing process. For the evaluation two flat-topped, square-based, pyramid shapes were used. Each pyramid had been manufactured twice using Steel and twice using Titanium. The utilisation of VULS was also explored by applying the VULS concept to the field of satellite image interpretation. Satellite data describing two villages located in a rural part of the Ethiopian hinterland were used for this purpose. In each case the ground surface was represented in a similar manner to the way that AISF sheet metal surfaces were represented, with the $z$ dimension describing the grey scale value. The idea here was to predict vertex labels describing ground type. As will become apparent, from the work presented in this thesis, the VULS concept is well suited to the task of 3D surface classification with respect to AISF and satellite imagery. The thesis demonstrates that the use of frequent VULS (rather than the other forms of VULS considered) produces more efficient results in the AISF sheet metal forming application domain, whilst the use of minimal VULS provided promising results in the context of the satellite image interpretation domain. The reported evaluation also indicates that a sound foundation has been established for future work on more general VULS based vertex classification.

The growing need for computer aided processing of knowledge has led to an increasing interest in description logics (DLs), which are applied to encode knowledge in order to make it explicit and accessible to logical reasoning. DLs and in particular the family around the DL ALC have therefore been thoroughly investigated w.r.t. their complexity theory and proof theory. The question arises which expressiveness these logics actually have. The expressiveness of a logic can be inferred by a model theoretic characterisation. On concept level, these DLs are akin to modal logics whose model theoretic properties have been investigated. Yet the model theoretic investigation of the DLs with their TBoxes, which are an original part of DLs usually not considered in context of modal logics, have remained unstudied. This thesis studies the model theoretic properties of ALC, ALCI, ALCQ, as well as ALCO, ALCQO, ALCQIO and EL. It presents model theoretic properties, which characterise these logics as fragments of the first order logic (FO). The characterisations are not only carried out on concept level and on concept level extended by the universal role, but focus in particular on TBoxes. The properties used to characterise the logics are `natural' notions w.r.t. the logic under investigation: On the concept-level, each of the logics is characterised by an adapted form of bisimulation and simulation, respectively. TBoxes of ALC, ALCI and ALCQ are characterised as fragments of FO which are invariant under global bisimulation and disjoint unions. The logics ALCO, ALCQO and ALCQIO, which incorporate individuals, are characterised w.r.t. to the class K of all interpretations which interpret individuals as singleton sets. The characterisations for TBoxes of ALCO and ALCQO both require, additionally to being invariant under the appropriate notion of global bisimulation and an adapted version of disjoint unions, that an FO-sentence is, under certain circumstances, preserved under forward generated subinterpretations. FO-sentences equivalent to ALCQIO-TBoxes, are - due to ALCQIO's inverse roles - characterised similarly to ALCO and ALCQO but have as third additional requirement that they are preserved under generated subinterpretations. EL as sub-boolean DL is characterised on concept level as the FO-fragment which is preserved under simulation and preserved under direct products. Equally valid is the characterisation by being preserved under simulation and having minimal models. For EL-TBoxes, a global version of simulation was not sufficient but FO-sentences of EL-TBoxes are invariant under global equi-simulation, disjoint unions and direct products. For each of these description logics, the characteristic concepts are explicated and the characterisation is accompanied by an investigation under which notion of saturation the logic in hand enjoys the Hennessy-and-Milner-Property. As application of the results we determine the minimal globally bisimilar companion w.r.t. ALCQO-bisimulation and introduce the L1-to-L2-rewritability problem for TBoxes, where L1 and L2 are (description) logics. The latter is the problem to decide whether or not an L1-TBox can be equivalently expressed as L2-TBox. We give algorithms which decide ALCI-to-ALC-rewritability and ALC-to-EL-rewritability.

This thesis describes research undertaken in the field of web data mining. More specifically this research is directed at investigating solutions to the Website Boundary Detection (WBD) problem. WBD is the problem of identifying the collection of all web pages that are part of a single website, which is an open problem. Potential solutions to WBD can be beneficial with respect to tasks such as archiving web content and the automated construction of web directories. A pre-requisite to any WBD approach is that of a definition of a website. This thesis commences with a discussion of previous definitions of a website, and subsequently proposes a definition of a website which is used with respect to the WBD solution approaches presented later in this thesis. The WBD problem may be addressed in either the static or the dynamic context. Both are considered in this thesis. Static approaches require all web page data to be available a priori in order to make a decision on what pages are within a website boundary. While dynamic approaches make decisions on portions of the web data, and incrementally build a representation of the pages within a website boundary. There are three main approaches to the WBD problem presented in this thesis; the first two are static approaches, and the final one is a dynamic approach. The first static approach presented in this thesis concentrates on the types of features that can be used to represent web pages. This approach presents a practical solution to the WBD problem by applying clustering algorithms to various combinations of features. Further analysis investigates the ``best'' combination of features to be used in terms of WBD performance. The second static approach investigates graph partitioning techniques based on the structural properties of the web graph in order to produce WBD solutions. Two variations of the approach are considered, a hierarchical graph partitioning technique, and a method based on minimum cuts of flow networks. The final approach for the evaluation of WBD solutions presented in this research considers the dynamic context. The proposed dynamic approach uses both structural properties and various feature representations of web pages in order to incrementally build a website boundary as the pages of the web graph are traversed. The evaluation of the approaches presented in this thesis was conducted using web graphs from four academic departments hosted by the University of Liverpool. Both the static and dynamic approaches produce appropriate WBD solutions, however. The reported evaluation suggests that the dynamic approach to resolving the WBD problem offers additional benefits over a static approach due to the lower resource cost of gathering and processing typically smaller amounts of web data.

In pattern recognition, better classification or regression results usually depend on highly discriminative features (also known as attributes) of datasets. Machine learning plays a significant role in the performance improvement for classification and regression. Different from the conventional machine learning approaches which train all features in one batch by some predictive algorithms like neural networks and genetic algorithms, Incremental Attribute Learning (IAL) is a novel supervised machine learning approach which gradually trains one or more features step by step. Such a strategy enables features with greater discrimination abilities to be trained in an earlier step, and avoids interference among relevant features. Previous studies have confirmed that IAL is able to generate accurate results with lower error rates. If features with different discrimination abilities are sorted in different training order, the final results may be strongly influenced. Therefore, the way to sequentially sort features with some orderings and simultaneously reduce the pattern recognition error rates based on IAL inevitably becomes an important issue in this study. Compared with the applicable yet time-consuming contribution-based feature ordering methods which were derived in previous studies, more efficient feature ordering approaches for IAL are presented to tackle classification problems in this study. In the first approach, feature orderings are calculated by statistical correlations between input and output. The second approach is based on mutual information, which employs minimal-redundancy-maximal- relevance criterion (mRMR), a well-known feature selection method, for feature ordering. The third method is improved by Fisher's Linear Discriminant (FLD). Firstly, Single Discriminability (SD) of features is presented based on FLD, which can cope with both univariate and multivariate output classification problems. Secondly, a new feature ordering metric called Accumulative Discriminability (AD) is developed based on SD. This metric is designed for IAL classification with dynamic feature dimensions. It computes the multidimensional feature discrimination ability in each step for all imported features including those imported in previous steps during the IAL training. AD can be treated as a metric for accumulative effect, while SD only measures the one-dimensional feature discrimination ability in each step. Experimental results show that all these three approaches can exhibit better performance than the conventional one-batch training method. Furthermore, the results of AD are the best of the three, because AD is much fitter for the properties of IAL, where feature number in IAL is increasing. Moreover, studies on the combination use of feature ordering and selection in IAL is also presented in this thesis. As a pre-process of machine learning for pattern recognition, sometimes feature orderings are inevitably employed together with feature selection. Experimental results show that at times these integrated approaches can obtain a better performance than non-integrated approaches yet sometimes not. Additionally, feature ordering approaches for solving regression problems are also demonstrated in this study. Experimental results show that a proper feature ordering is also one of the key elements to enhance the accuracy of the results obtained.

Limit order book markets are a rich research area, not only because these markets generated huge amounts of data (at an exceedingly high rate), but also because the fine level of detail that their data enables one to explore market microstructure in unprecedented ways. Due to the large quantity and rich details of the data in such market, one has to leverage the power of computers to perform both the analysis and modeling work. This calls for both new algorithms and infrastructure to perform the computing tasks effectively and efficiently. Motivated by the questions and challenges I see there, I started my research first from a engineering perspective and then moved to a quantitative perspective. My aim was to find my way through this newly emerging area and develop a systematic approach to seek, study and solve the potential questions in it. I will graph and explain my findings and results in this thesis, hoping that they will help and inspire further research work. To discipline and guide myself with a clear goal in the long journey exploring the world of limit order book markets, I focus on liquidity modeling. I try to predict trading volume from a daily scale to intra-day distributions, with the aim to design trading algorithms to reduce transaction costs and market impact. Within a microstructure context, I try to model the self-exciting nature of trading events with both a stochastic process approach and a statistical approach. Prediction methods are proposed to help trading algorithms to react to big trade events in real time. I use two different modelling approaches. One is based on stochastic processes that have nice mathematical properties, while the other one is driven by statistics extracted directly from the data. I try to examine them in a unified and scientific way so that it is easy to compare the strengthes and weaknesses of each of them. Empirical findings are given to support the rationale behind all of the proposed algorithms.