Towards improving Bayesian optimisation for the physical sciences via long-term planning
File(s)
Author(s)
Folch, Jose Pablo
Type
Thesis
Abstract
Design of experiments is a vital area of research in all sciences. Initially an arduous and expensive task due to the need to manually do experiments, there has been massive opportunity to advance scientific research through the use of automation. Bayesian optimisation has emerged as the leading tool for expensive black-box maximisation, in particular with regards to selecting the configuration of expensive-to-train machine learning models.
In the physical sciences we face a much more complicated task, where the dynamics of experimentation are subject to strict physical constraints, time-delays, and varying accuracy of measurements. In order to unleash the potential of Bayesian optimisation for scientific discovery it is vital to adapt the algorithms to all the constraints faced in real-world scenarios. In this thesis we will tackle three of those problems through planning.
We look at multi-fidelity and asynchronous batch optimisation. These refer to the setting where we can use cheap measurements to approximate expensive experiments which may take a long time to run. The desynchronisation between experiments and the potential to parallelise them results in the asynchronous setting, where we must choose new experiments while taking into account other yet-unfinished experiments. We show there is a close relationship between both areas and propose an algorithm for combining them efficiently.
Later, we look at two related physical problems. Traditionally Bayesian optimisation can freely query any experiment in the search space. However, in the physical sciences we often cannot make large changes between experiments as it may be costly or break the system dynamics. By planning full trajectories through the search space we are able to efficiently optimise the underlying functions, and extend Bayesian optimisation to a whole range of physical problems. We showcase connections with previous Bayesian optimisation approaches, real-world applications, and experimentally validate the benefits of the proposed methods.
In the physical sciences we face a much more complicated task, where the dynamics of experimentation are subject to strict physical constraints, time-delays, and varying accuracy of measurements. In order to unleash the potential of Bayesian optimisation for scientific discovery it is vital to adapt the algorithms to all the constraints faced in real-world scenarios. In this thesis we will tackle three of those problems through planning.
We look at multi-fidelity and asynchronous batch optimisation. These refer to the setting where we can use cheap measurements to approximate expensive experiments which may take a long time to run. The desynchronisation between experiments and the potential to parallelise them results in the asynchronous setting, where we must choose new experiments while taking into account other yet-unfinished experiments. We show there is a close relationship between both areas and propose an algorithm for combining them efficiently.
Later, we look at two related physical problems. Traditionally Bayesian optimisation can freely query any experiment in the search space. However, in the physical sciences we often cannot make large changes between experiments as it may be costly or break the system dynamics. By planning full trajectories through the search space we are able to efficiently optimise the underlying functions, and extend Bayesian optimisation to a whole range of physical problems. We showcase connections with previous Bayesian optimisation approaches, real-world applications, and experimentally validate the benefits of the proposed methods.
Version
Open Access
Date Issued
2024-10-03
Date Awarded
01/02/2025
License URL
Advisor
Misener, Ruth
van der Wilk, Mark
Sponsor
BASF
Engineering and Physical Sciences Research Council
Grant Number
EP/S023151/1
Publisher Department
Department of Mathematics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)