Skip to content Skip to navigation

Optimization OR & Statistics

CDT in Data Science and Artificial Intelligence

Reference Number: 2018-MAX-02

Abstract: The EPSRC CDT in Data Science accepted its last cohort in September 2018.  However, we now welcome applications to the Centre for Doctoral Training (CDT) in Data Science and Artificial Intelligence (DSAI). Students on this programme will integrate with the current EPSRC CDT in Data Science, although the programme is formally a separate one. Students with a very strong background in computer science, mathematics, physics, or engineering are particularly encouraged to apply.

Apply here.

Adaptive Monte Carlo Methods for Stochastic Differential Equations

Reference Number: 2018-HW-AMS-14

Abstract: Capturing rare events is crucial for accurate risk assessment and its successful management. An example in finance is computing the probability of a large, but rare, loss from a financial portfolio. Approximating expectations involving such rare events is difficult because, when using Monte Carlo, many of the generated samples do not contribute to the final outcome and the expensive samples are effectively wasted. Adaptive sampling methods resolve this issue by spending some minimal computational effort to determine if a sample is of interest and, only if it is, the sample accuracy is then improved by spending further computational effort.  Using concepts from stochastic analysis, probably theory and numerical analysis, this project will look at applying adaptive methods to compute outputs depending on stochastic differential equations rather than simple random variables.

Apply here.

Applied Probability and Electricity Networks

Reference Number: 2018-HW-AMS-06

Abstract: The decarbonisation of the electricity network is leading to substantial changes in both generation, e.g. renewable generators and increased electricity storage, and demand, e.g. electric vehicles and dynamic demand response. This is leading to the development of novel mathematical tools to address the challenges arising from these changes.  Many of these tools are related to ideas from probability and queueing theory including stability analysis, large deviations and decentralised algorithms.  In this PhD project the student will consider a number of models arising from power systems and explore the application of ideas and methods from queuing theory.

Apply here.

Astrophysical inference using the LIGO gravitational wave detectors

Abstract: In September 2015, the LIGO detectors observed gravitational waves from a binary black hole merger for the first time. LIGO has subsequently observed five further events. As LIGO makes further observations, the properties of the observed population will allow us to constrain the astrophysical properties and origin of the sources. This project will focus on developing techniques for such inference and applying them to the LIGO data set.

Apply here.

Developing data analysis techniques for low frequency gravitational wave detectors

Abstract: There are ongoing efforts to detect gravitational waves at nanohertz frequencies using the accurate timing of millisecond pulsars with pulsar timing arrays. ESA plans to launch a space based gravitational wave detector, LISA, observing in the millihertz band, and it has recently been realised that gravitational waves at microhertz frequencies can be detected by monitoring the positions of stars in the sky using astrometric satellites such as GAIA. Techniques for analysing the data from these various detectors and extracting the science are not fully developed and this project will focus on that. The project could concentrate on one of the three detectors or on methods that could be applied to all of them.

Apply here.


Reference Number: 2018-HW-AMS-18


Apply here.

High-Dimensional Extremes

Abstract: Extreme values of high-dimensional random vectors appear naturally in risk management. This thesis develops dimension reduction methods for multivariate extreme values, with a principal focus on high-dimensional extremes. A main goal is on tracking directions which account for a larger proportion of risk, and to use corresponding lower-dimensional representations to quantify risk in the raw high-dimensional random vector. Applications are envisioned in high- dimensional portfolios.

Apply here.

Managing uncertainty in government modelling

Abstract: There is considerable interest in government at present in management of uncertainty, in large part driven by the Aqua Book (guidance on producing quality analysis for government, see A range of projects would be possible, including statistical methodological work, use of modelling for decision support, developing ways of making uncertainty quantification methodology accessible to a wide analyst community, and communication of modelling results beyond the technical modelling community. There would be opportunities to collaborate with relevant government units, including through projects based at the Alan Turing Institute.

Mathematical and statistical modelling of energy systems

Abstract: I can supervise broadly across energy systems analysis, including in operational and planning optimization, uncertainty quantification in large scale energy system models, security of supply risk analysis, statistical modelling of renewable resource, and people/institutional/communication aspects of the use of modelling in public policy decisions. My research in these areas is strongly informed by industry and government links, and there is particular scope for linking basic research to important application in security of supply analysis and modelling of renewable resources (wind, solar etc) and their relationship with demand.

Apply here.

PDE-constrained optimization in scientific processes

Reference Number:

Abstract: A vast number of important and challenging applications in mathematics and engineering are governed by inverse problems. One crucial class of these problems, which has significant applicability to real-world processes, including those of fluid flow, chemical and biological mechanisms, medical imaging, and others, is that of PDE-constrained optimization. However, whereas such problems can typically be written in a precise form, generating accurate numerical solutions on the discrete level is a highly non-trivial task, due to the dimension and complexity of the matrix systems involved. In order to tackle practical problems, it is essential to devise strategies for storing and working with systems of huge dimensions, which result from fine discretizations of the PDEs in space and time variables. In this project, "all-at-once" solvers coupled with appropriate preconditioning techniques will be derived for these systems, in such a way that one may achieve fast and robust convergence in theory and in practice. This project is related to the EPSRC Fellowship.

Apply here.

Scalable inference for mixtures of experts

Abstract: Mixtures of experts are a popular class of models in statistics and machine learning, with applications in a variety of fields. They provide highly flexible regression and overcome parametric and stationary assumptions of standard regression models by partitioning the input space into regions where such assumptions need only locally within region. Infinite mixtures of experts allow the data to determine the number of local regions present, without the need to fix this number in advance. This project will develop scalable inference for infinite mixtures of experts, combining scalable experts in high-dimensional input spaces with efficient inference methods for large sample sizes.

Statistical analysis of writing style in literature and social media (with application to cybersecurity)

Abstract: Suppose that the manuscript of a newly discovered play purported to be written by Shakespeare is found in an antique shop.  Professors of English Literature begin to debate whether the author was indeed Shakespeare, or an imposter. To what extent can statistical analysis help answer this question? in fact, the use of statistical methods for tasks such as this is quite common. The technical name for this is "authorship attribution" - creating quantitative models to describe the writing styles of authors and using these to assess whether they are the true authors of written texts of interests. Until recently, this was a niche field restricted to solving academic literary disputes. However the rise of the internet and social media has made such tools directly relevant to cybersecurity. Suppose that an anonymous post is made on an extremist Internet forum. Can analysis of the writing style be used to identify (or exclude) potential authors as part of a deanonymisation analysis? Or suppose that a person's Twitter account is hijacked, and a hacker starts posting new content on it. Can statistical analysis identify the change in writing style when it occurs? This project will examine traditional methods for authorship attribution where there is only a small number of candidate authors (Shakespeare, etc) and extend them to the modern internet era where there may be millions of possible authors, and only small text fragments associated with each. Potential tools include statistical classification, clustering, and Bayesian hierarchal modelling.

Apply here.

Statistical mechanics for reaction-diffusion systems

Reference Number: ACM

Abstract: Many systems in chemistry, biology and engineering can be modelled by reaction-diffusion equations in which the populations of a number of different species move and interact or react, leading to the interchange of population or mass between the species. When the populations are large and/or reside in a complex environment (such as particles suspended in a turbulent flow), a powerful approach is to use techniques from statistical mechanics, which describes the 'average' behaviour of such systems. Dynamical density functional theory is one such approach that has met with great success over the past decade or so. This project will extend existing models, which generally describe only the dynamics, to include the reaction terms. The topics covered can be tailored to the interest of the student, covering both rigorous analysis and numerics. Techniques include statistical mechanics, stochastic dynamics, mathematical modelling, homogenisation theory of PDEs, and computational methods such as pseudo-spectral methods and finite elements. This project also has strong links to the work of members of the School of Engineering. 

Apply here.

Statistical methods for earthquake forecasting

Abstract: Statistical analysis of earthquake recurrence and its input into seismic hazard assessments is of interest to both the insurance and disaster management industries. This is typically based on probability models for the expected time between earthquakes in a particular region. Building these statistical models is complicated by the limited number of events in historical records. Many current approaches resolve this by combining earthquake histories of multiple regions into a single catalogue. However, more sophisticated methods for combining information across multiple regions and coping with limited data in a more principled and flexible manner are needed. This project will focus on Bayesian hierarchical modelling to allow natural pooling of information across related regions, along with machine learning/artificial intelligence approaches to gain insight into earthquake recurrence and the variability in seismic cycles. (No previous experience in these methods is required).

Apply here.

Stochastic differential equations, sampling and big data

Reference Number: ACM

Abstract: For details on the range of potential topics in this area please contact Konstantinos Zygalakis.

Apply here.

Stochastic hybrid modelling of chemical systems

Reference Number: ACM

Abstract: It is well known that stochasticity can play a fundamental role in various biochemical processes, such as cell regulatory networks and enzyme cascades. Isothermal, well-mixed systems can be adequately modelled by Markov processes and, for such systems, methods such as Gillespies algorithm are typically employed. While such schemes are easy to implement and are exact, the computational cost of simulating such systems can become prohibitive as the frequency of the reaction events increases. This has motivated numerous coarse grained schemes, where the fast reactions are approximated either using Langevin dynamics or deterministically. While such approaches provide a good approximation for systems where all reactants are present in large concentrations, the approximation breaks down when the fast chemical species exist in small concentrations, giving rise to signicant errors in the simulation. This project will be concerned with developing further and analysing a new hybrid approach (a stochastic dierential equation with jumps) capable of dealing with more general systems.

Apply here.

Stochastic optimisation algorithms for data science

Reference Number: 2018-HW-AMS-19

Abstract: In the information age, many scientific and technological applications (ranging from healthcare to energy management, risk assertion, or astronomy) rely on the acquisition and analysis of huge data volumes. These data need to be transformed into interpretable information, for subsequent decision making processes. The related methodologies fall into the remit of what is now called data science. Handling the data volumes efficiently is a significant challenge, and powerful mathematical tools are required.
In this PhD project, the student will focus on the class of stochastic optimisation algorithms. These iterative methods are particularly suitable to handle big datasets (e.g. they allow to split the data into small variables, selected randomly). The objective will be two folds: (i) develop new such algorithms with convergence guaranties, and (ii) analyse the behaviour of the developed methods through simulations. A prime example will be in astronomical imaging where images and data can both be of the order of terabit.

Apply here.

Tests of fundamental physics with gravitational waves

Abstract: Gravitational waves are generated in a highly dynamical regime of gravity that has not been well constrained observationally. Observations with LIGO and future detectors such as LISA will provide a unique way to probe the properties of gravitational theory in this regime. This project will develop techniques for producing systematic and comprehensive constraints on deviations form general relativity in future observations.

Apply here.

Very large scale optimization with Interior Point Methods

Abstract: This PhD project will develop theory and implementation of interior point methods (IPMs) for huge scale constrained optimization problems. IPMs are particularly well-suited to solving very large problems and indeed as demonstrated by excessive computational experience, they stay beyond any competition. They are proved to solve problems of dimension N in no more that the square root of N iterations. However, in practice, they never need more than log(N) iterations. This project will aim at closing the gap between the best to-date known worst-case complexity result and the observed excellent behaviour of IPMs. 

Apply here.

Volatility-Models for Multivariate Extremes

Abstract: Statistical models for multivariate extremes are one of the most commonly used strategies to approach problems where it is necessary to understand the association of variables during extreme scenarios (say joint extreme losses in a portfolio). This thesis develops the angular volatility as a natural modeling tool for settings where the structure of dependence between extreme values may be changing over time. The target is on developing measures which can track the dynamics governing extremal dependence of losses in a portfolio over time.

Apply here.

• Model diagnostics for infectious epidemics

Reference Number: 2018-HW-AMS-17

Abstract: Stochastic modelling of communicable disease outbreaks is challenging due to inter-dependence in the involved transmission dynamics and imperfect observation of infection-related events. Estimation in such models is now well established, but research on model assessment and comparison is still under progress. This project will build on recently developed tools for epidemic model diagnostics to investigate the use of Bayesian methodology related to latent residuals, in cases where the epidemic outbreak is: (a) under-reported, (b) at early stages of its course, and/or (c) under the impact of intervention measures.

Apply here.

• Predictive performance modelling for financial investment strategies

Reference Number: 2018-HW-AMS-16

Abstract: The development of multi-asset quantitative strategies is crucial in the investment industry and requires the analysis of vast amounts of data to gain insight about performance. This project will investigate the use of statistical predictive models and performance measures for risk evaluation in financial investment strategies, aiming to provide increased levels of predictive robustness in modelling under different scenarios. Methodology related to back-testing and statistical machine learning will be used, and model uncertainty will be accounted for under a Bayesian approach.

Apply here.