Writing software is a difficult and expensive task. Its automation is hence very valuable. Search algorithms have been successfully used to tackle many software engineering problems. Unfortunately, for some problems the traditional techniques have been of only limited scope, and search algorithms have not been used yet. We hence propose a novel framework that is based on a co-evolution of programs and test cases to tackle these difficult problems. This framework can be used to tackle software engineering tasks such as Automatic Refinement, Fault Correction and Improving Non-functional Criteria. These tasks are very difficult, and their automation in literature has been limited. To get a better understanding of how search algorithms work, there is the need of a theoretical foundation. That would help to get better insight of search based software engineering. We provide first theoretical analyses for search based software testing, which is one of the main components of our co-evolutionary framework. This thesis gives the important contribution of presenting a novel framework, and we then study its application to three difficult software engineering problems. In this thesis we also give the important contribution of defining a first theoretical foundation.

The Object Constraint Language (OCL) is a declarative language and is used to make the Unified Modeling Language (UML) models well-defined through defining a set of constraints. However, the syntactic complexity of OCL makes the writing of OCL code difficult. A natural language based interface can be useful in making the process of writing OCL expressions easy and simple. However, the translation of natural language (NL) text to object constraint language (OCL) code is a challenging task on account of the informal nature of natural languages as various syntactic and semantic ambiguities make the process of NL translation to formal languages more complex. However, in our approach the usage of SBVR not only provides natural languages a formal abstract syntax representation but it is also close to OCL syntax. In this thesis, a framework is presented to facilitate the users of the UML tools so that they can write invariants and pre/post conditions in English. The results of the case studies manifest that a natural language based approach to generate OCL constraints can not only help in significantly improving usability of OCL but also outperforms the most closely related techniques in terms of effectiveness and effort required in generating OCL.

The rigid structure of medical hierarchies within UK hospitals can become the source of dissatisfaction and conflict for medical personnel, the repercussions of which can be disastrous for patients and staff. The research reported herein presents the results of an investigation into the use of Virtual Reality (VR) simulation and conventional story-boarded techniques to empower medical personnel to challenge decisions they feel are inappropriate. Prototype applications were crafted from a selection of transcribed ‘challenge events’ acquired from an opportunistic sample of clinical staff. Data obtained from an initial investigation were used to establish attitudes toward challenging and evaluate the findings of the literature to generate research questions and objectives. Medical personnel who engaged with both media as part of an experimental phase assessed their viability as potential training resources to help foster the ability to challenge. Analysis of this experiment suggested that both techniques are viable tools in the delivery of decision-making training and could potentially deliver impact into other applications within healthcare. To increase the realism of the training material, the technologies should be presented in a format appropriate for those with limited ‘gaming’ experience and allow a credible level of interaction with the environment and characters.

Today's distributed programs are often written using either explicit message passing or Remote Procedure Calls (RPCs) that are not natively integrated in the language. It is difficult to establish the correctness of programs written this way compared to programs written for a single computer. We propose a generalisation of RPCs that are natively integrated in a functional programming language meaning that they have support for higher-order calls across node boundaries. Our focus is on how such languages can be compiled correctly and efficiently. We present four different solutions. Two of them are based on interaction semantics --- the Geometry of Interaction and game semantics --- and two are extensions of conventional abstract machines --- the Krivine machine and the SECD machine. To target as general distributed systems as possible our solutions support RPCs without sending code. We prove the correctness of the abstract machines with respect to their single-node execution, and show their viability for use for compilation by implementing prototype compilers based on them. The conventionally based machines are shown to enable efficient programs. Our intention is that these abstract machines can form the foundation for future programming languages that use the idea of higher-order RPCs.

Dependency parsing is one of the important natural language processing tasks that assigns syntactic trees to texts. Due to the wider availability of dependency corpora and improved parsing and machine learning techniques, parsing accuracies of supervised learning-based systems have been significantly improved. However, due to the nature of supervised learning, those parsing systems highly rely on the manually annotated training corpora. They work reasonably good on the in-domain data but the performance drops significantly when tested on out-of-domain texts. To bridge the performance gap between in-domain and out-of-domain, this thesis investigates three semi-supervised techniques for out-of-domain dependency parsing, namely co-training, self-training and dependency language models. Our approaches use easily obtainable unlabelled data to improve out-of-domain parsing accuracies without the need of expensive corpora annotation. The evaluations on several English domains and multi-lingual data show quite good improvements on parsing accuracy. Overall this work conducted a survey of semi-supervised methods for out-of-domain dependency parsing, where I extended and compared a number of important semi-supervised methods in a unified framework. The comparison between those techniques shows that self-training works equally well as co-training on out-of-domain parsing, while dependency language models can improve both in- and out-of-domain accuracies.

Evolutionary dynamic optimisation (EDO), or the study of applying evolutionary algorithms to dynamic optimisation problems (DOPs) is the focus of this thesis. Based on two comprehensive literature reviews on existing academic EDO research and real-world DOPs, this thesis for the first time identifies some important gaps in current academic research where some common types of problems and problem characteristics have not been covered. In an attempt to close some of these gaps, the thesis makes the following contributions: First, the thesis helps to characterise DOPs better by providing a new definition framework, two new sets of benchmark problems (for certain classes of continuous DOPs) and several new sets of performance measures (for certain classes of continuous DOPs). Second, the thesis studies continuous dynamic constrained optimisation problems (DCOPs), an important and common class of DOPs that have not been studied in EDO research. Contributions include developing novel optimisation approaches (with superior results to existing methods), analysing representative characteristics of DCOPs, identifying the strengths/weaknesses of existing methods and suggesting requirements for an algorithm to solve DCOPs effectively. Third, the thesis studies dynamic time-linkage optimisation problems (DTPs), another important and common class of DOPs that have not been well-studied in EDO research. Contributions include developing a new optimisation approach (with better results than existing methods in certain classes of DTPs), analysing the characteristics of DTPs and the strengths and weaknesses of existing EDO methods in solving certain classes of DTPs.

Creativity is a useful attribute for people to have. It allows them to solve unfamiliar problems, introduce novelty to established domains, and to understand and assimilate new information and situations - all things we would like computers to be able to do too. However, these creative attributes do not exist in isolation: they occur in a context in which people tend to solve problems routinely where possible rather than consider non-standard ideas. These more mundane attributes might also be useful for problem solving computers, for the same reasons they are useful for us. However, they are often ignored in attempts to implement systems capable of producing remarkable outputs. We explore how the study of both human and computational creativity can inform an approach to help computers to display useful, complete problem-solving behaviour similar to our own: that is, robust, exible and, where possible and appropriate, surprising. We describe a knowledge-based model that incorporates a genetic algorithm with some characteristics of our own approach to knowledge reuse. The model is driven by direct interactions with problem scenarios. Descriptions of the role or appearance of key themes and concepts in literature in functioning problem-solving systems is lacking; we suggest that they appear as artefacts of the operation of our model. We demonstrate that it is capable of solving routine problems flexibly and effectively. We also demonstrate that it can solve problems that would be effectively impossible for a genetic algorithm operating without the benefit of knowledge-driven biasing. Artefacts of the behaviour of the model could, in certain scenarios, lead to the appearance of non-routine or surprising solutions.

Information leakage occurs when a system exposes its secret information to an unauthorised entity. Information flow analysis is concerned with tracking flows of information through systems to determine whether they process information securely or leak information. We present a novel information flow model that permits an arbitrary amount of secret and publicly-observable information to occur at any point and in any order in a system. This is an improvement over previous models, which generally assume that systems process a single piece of secret information present before execution and produce a single piece of publicly-observable information upon termination. Our model precisely quantifies the information leakage from secret to publicly-observable values at user-defined points - hence, a "point-to-point" model - using the information-theoretic measures of mutual information and min-entropy leakage; it is ideal for analysing systems of low to moderate complexity. We also present a relaxed version of our information flow model that estimates, rather than computes, the measures of mutual information and min-entropy leakage via sampling of a system. We use statistical techniques to bound the accuracy of the estimates this model provides. We demonstrate how our relaxed model is more suitable for analysing complex systems by implementing it in a quantitative information flow analysis tool for Java programs.

The thesis investigates the problem of identifying active modules of bipartite metabolic network. We devise a method of motif projection, and the extraction of clusters from active modules based on the concentration of active-motifs in the network. Our results reveal the existence of hierarchical structure. We model regulation of metabolism as an interaction between a metabolic network and a gene regulatory network in the form of interconnected network. We devise two module detection algorithms for interconnected network to evaluate the molecular changes of activity that are associated with cellular responses. The first module detection algorithm is formulated based on information map of random walks that is capable of inferring modules based on topological and activity of nodes. The proposed algorithm has faster execution time and produces comparably close performance as previous work. The second algorithm takes into account of strong regulatory activities in the gene regulatory layer to support the active regions in the metabolic layer. The integration of gene information allows the formation of large modules with better recall. In conclusion, our findings indicate the importance of no longer modelling complex biological systems as a single network, but to view them as flow of information of multiple molecular spaces.

The rail industry has focused on the improvement of maintenance through the effective use of online condition monitoring of rolling stock and rail infrastructure in order to reduce the occurrence of unexpected catastrophic failures and disruption that arises from them to an absolute minimum. The basic components comprising a railway wheelset are the wheels, axle and axle bearings. Detection of wheelset faults in a timely manner increases efficiency as it helps minimise maintenance costs and increase availability. The main aim of this project has been the development of a novel integrated online acoustic emission (AE) and vibration testing technique for the detection of wheel and axle bearing defects as early as possible and well before they result in catastrophic failure and subsequently derailment. The approach employed within this research study has been based on the combined use of accelerometers and high-frequency acoustic emission sensors mounted on the rail or axle box using magnetic hold-downs. Within the framework of this project several experiments have been carried out under laboratory conditions, as well as in the field at the Long Marston Test Track and in Cropredy on the Chiltern Railway line to London.