Search

The work presented in this thesis concerns the problem of dynamic vehicle routing. The motivation for this is the increasing demands on transportation services to deliver fast, efficient and reliable service. Systems are now needed for dispatching transportation requests that arrive dynamically throughout the scheduling horizon. Therefore the focus of this research is the dynamic pickup and delivery problem with time windows, where requests are not completely known in advance but become available during the scheduling horizon. All requests have to be satisfied by a given fleet of vehicles and each request has a pickup and delivery location, along with a time window at which services can take place. To solve the DPDPTW, our algorithm is embedded in a rolling horizon framework, thus allowing the problem to be viewed as a series of static sub-problems. This research begins by considering the static variant of the problem. Both heuristic and metaheuristic methods are applied and an analysis is performed across a range of well-known instances. Results competitive with the state of the art are obtained. For the dynamic problem, investigations are performed to identify how requests arriving dynamically should be incorporated into the solution. Varying degrees of urgency and proportions of dynamic requests have been examined. Further investigations look at improving the solutions over time and identifying appropriate improvement heuristics. Again competitive results are achieved across a range of instances from the literature. This continually increasing area of research covers many real-life problems such as a health courier service. Here, the problem consists of the pickup and delivery of mail, specimens and equipment between hospitals, GP surgeries and health centres. Final research applies our findings to a real-life example of this problem, both for static schedules and a real-time 24/7 service.

As one of the world's leading steel producers, Tata Steel generates and records substantial quantities of data on a daily basis, much of it linked to its manufacturing process. This research seeks to utilise the techniques developed under the general heading of Data Science to enable the company to derive maximum benefit from the data it records. In particular, the research centres on the data collected during the production of steel slabs from liquid iron at the company's Port Talbot operation. The key outcome of the research being to provide improved insight on the performance and maintenance of its continuous casters. At present, there are a number of existing organisational challenges, including poor communication and data sharing that inhibit best use of the collected data. Initial investigations focused on understanding existing data and identifying key data sources that could be exploited to explain events and activities surrounding caster performance. Initial findings revealed that, maintenance events data and real-time sensor data to be effective in identifying key performance indicators. The data sources considered contain rich information to explain both 'normal' and 'anomalous' casting behaviour patterns associated with an unexpected caster performance failure. A Data Science Methodology provided by Rollins (2016), includes business understanding, data understanding, data preparation and modelling, which has been employed to approach this problem to develop a failure prediction model. The knowledge gained from learning the pre-failure conditions from historical data has enabled a predictive model to be developed. This model has the ability enable Tata Steel to make informed decisions about when to take a caster out of production for maintenance. Ultimately, this research promotes better utilisation of current data to help optimise the scheduling of caster maintenance. The findings of this research present Tata Steel with the opportunity to minimise the occurrence of unnecessary caster maintenance; delivering productivity increase and cost reduction.

Missed appointments are estimated to cost the UK National Health Service (NHS) approximately £1 billion annually. Research that leads to a fuller understanding of the types of factors influencing spatial and temporal patterns of missed appointments is therefore timely. Current research is mainly focused on traditional statistical techniques that fail to capture the complex relationships in this type of data. This thesis uses machine learning approaches to investigate whether these factors are consistent across a range of medical specialties. A predictive model based on LightGBM was used to determine the risk-increasing and risk-mitigating factors for missing appointments, which were then used to assign a risk score to patients on an appointment-by-appointment basis for each specialty. Results show that the best predictors of missed appointments include the patient's age, appointment history, and the deprivation rank of their area of residence; however, the factors associated with missed appointments have also been shown to differ across medical specialties. This report shows that the use of machine learning techniques has real value in informing future targeted reminder mechanisms that can help reduce the burden on the NHS and improve patient care and well-being.

While the amount of unstructured text data continues to grow within the clinical domain, little modelling is carried out in comparison to other industries. My research goal in this thesis is to present machine learning models that can effectively discern the relationships within medical notes, tying symptoms and other elements to an associated medical speciality. There have been many studies in the clinical domain using natural language processing that have seen successes with document classification. However, the solutions proposed often rely on an external medical dictionary to annotate the data. My goal is the development of a classifier that shows that these relationships can be extracted from the original, unstructured text. Furthermore, the standard approach to documenting research in this area revolves around focusing on a single type of machine learning algorithm, be it the method of feature generation or the specific machine learning model chosen for the task. The results shown in this thesis address this issue by providing a comparative demonstration of multiple feature generation methods alongside a plethora of traditional machine learning and neural network-based models for classification. Lastly, existing research encounters issues with the procurement of suitable medical data, often defaulting to using datasets that have been curated for a specific task. This research instead uses real patient data from Digital Health and Care Wales (DHCW), selected randomly from cases between 2018 and 2019.The results produced in this thesis found that frequency-based feature generation performed substantially better than word embeddings when using a traditional machine learning model like logistic regression. However, using word embeddings with a neural network architecture yielded more comparable results. For the machine learning models themselves, the support vector machine (91%) and two transformer deep learning models (93%) produced the best results.