Before moving to UCL, I completed my PhD in Computer Science at the University of Sheffield. I conducted theoretical analyses of algorithm configurators, tools that optimise the settings of other algorithms. These tools are of particular interest today since they can be used to optimise the hyperparameters of machine learning models.
My research was the first to try to establish a rigorous grounding for these tools and address key questions such as whether a given configurator will ever be able to identify optimal parameter settings for an algrorithm, and, if so, how long will it require to do so. This work led us to the insight that the traditional way of measuring the performance of a configuration of an optimisation algorithm (by measuring the time it takes to reach the optimum of the problem) is problematic, and we proved that a faster approach is to look at the best solution achieved by the configuration within a given time budget.
We published these findings in two conference papers and extended the first in a paper published in Artificial Intelligence:
- George T. Hall, Pietro S. Oliveto, Dirk Sudholt. On the impact of the performance metric on efficient algorithm configuration. Artificial Intelligence (2022).
- George T. Hall. A Framework for the Runtime Analysis of Algorithm Configurators. Doctoral thesis (2021).
- George T. Hall, Pietro S. Oliveto, Dirk Sudholt. Fast perturbative algorithm configurators. International Conference on Parallel Problem Solving from Nature (2020).
- George T. Hall, Pietro S. Oliveto, Dirk Sudholt. Analysis of the performance of algorithm configurators for search heuristics with global mutation operators. Proceedings of the 2020 Genetic and Evolutionary Computation Conference (2020).
- George T. Hall, Pietro S. Oliveto, Dirk Sudholt. On the impact of the cutoff time on the performance of algorithm configurators. Proceedings of the Genetic and Evolutionary Computation Conference (2019).