General Profile:

I have been at the university of Bristol for 4 years studying an integrated masters in Engineering Mathematics. The degree offers a multidisciplinary tool-kit to solve modern problems with mathematical models and computer simulations. I decided to specialise my last 2 years of study on topics related to machine-learning and spent a summer working for the university researching finite-element optimisation. Outside of university, I worked for a data-science company to investigate air-quality for large construction sites.

A key chapter of my decision to become a researcher in AI involves my third year project where I worked in a team to create an end-to-end medical symptom-diagnosis chatbot. This project opened my eyes to the many ways AI research can contribute towards improving aspects of society. I decided to specialise my 4^th year’s thesis on classification problems in machine learning. My thesis explored the relationship between classification performance and characteristics (and the size) of data distributions and produced evidence to suggest ‘more data isn’t necessarily good’.

I am very excited to be part of the IAI CDT; I love maths and computer programming and look forward to contributing to towards building responsible AI systems. Outside of academia, I love to do all kinds of sports (rugby, juijutsu, frisbee) and play the guitar. My next goal is to learn how to salsa...

Research Project Summary:

Time-series forecasting, a discipline within data analysis, holds a pivotal role in academia and practical domains due to its capacity to predict future trends based on historical data patterns. The exploration of time-series forecasting is crucial for various domains, ranging from: healthcare, transport networks, geographical systems, and financial markets.

Most machine learning methods for time-series forecasting are trained to predict one-step-ahead. However, this is a limiting framework for tasks where it is important to predict a system multiple steps ahead in time. The problem in multi-step-ahead forecasting lies in: is it better to take one big jump, or multiple smaller jumps?

Given how general this problem is, advances in multi-step-ahead forecasting have the potential to save lives, such as for early prediction of neurodegenerative diseases, and reduce financial risks for transport systems sustaining national economy.

As an example, consider the case of predicting ten steps ahead in some system, the forecaster can try and predict ten steps in one big (direct) step, or one step at a time with small (recursive) steps. The trouble is that, learning big steps is harder than learning small steps, but recursively predicting with small steps leads to accumulating error. There have been multiple accounts of this problem, and there is empirical evidence for both strategies succeeding over the other.

Given this uncertainty on how to address multi-step error accumulation, this thesis presents the following research hypotheses:

1. The instance-wise ranking-distribution of strategies determines the severity of the multi- step problem.

2. Dynamic model selection methods can address the multi-step problem by reducing the ranking error and the overall error, in comparison to traditional strategies.

3. Dynamic multi-step-ahead-forecasting strategies are explainable to better understand how characteristics of the time-series relate to strategy-optimality.

This project falls within the EPSRC A.I. and robotics themes as well as healthcare technologies, energy, and engineering. Advancements in multi-step-ahead forecasting will allow for safer planning for autonomous agents in robotics, better prediction of neurodegenerative conditions in healthcare, more efficient load balancing in energy distribution, and more accurate models for transport networks in engineering