Optimisation-compatible learning for structured decision-making
File(s)
Author(s)
Triantafyllou, Niki
Type
Thesis or dissertation
Abstract
Structured decision-making problems in process systems engineering often combine discrete decisions, continuous variables, operational constraints, and expensive simulator evaluations. Mathematical optimisation provides a principled framework for these problems, but large mixed-integer models and simulation-based optimisation tasks can become computationally prohibitive. Machine learning can exploit recurring structure across related instances, but standard predictive models do not guarantee feasibility or solution quality.
This thesis develops optimisation-compatible learning methods in which learned components are designed to support optimisation rather than replace it. The methods address two problem classes, namely mixed-integer optimisation and expensive black-box process optimisation.
For mixed-integer optimisation, the thesis presents a learning-enhanced dimensionality reduction framework in which feed-forward and convolutional neural networks identify complicating binary variables that define reduced optimisation models solved by exact branch-and-cut solvers. It then introduces an energy-based graph neural network framework that learns energy functions over combinatorial configurations and embeds them within mixed-integer inference problems, allowing hard constraints to be enforced at inference time.
For black-box process optimisation, the thesis develops reduced-space Bayesian optimisation methods that use global sensitivity analysis to identify influential process variables and optimise in physically interpretable reduced spaces. This is extended to multi-fidelity Bayesian optimisation, where low-fidelity approximations and high-fidelity simulator evaluations are combined under limited computational budgets.
Together, these contributions show how learning can improve the scalability and data efficiency of structured optimisation while preserving feasibility, constraint representation, and decision quality.
This thesis develops optimisation-compatible learning methods in which learned components are designed to support optimisation rather than replace it. The methods address two problem classes, namely mixed-integer optimisation and expensive black-box process optimisation.
For mixed-integer optimisation, the thesis presents a learning-enhanced dimensionality reduction framework in which feed-forward and convolutional neural networks identify complicating binary variables that define reduced optimisation models solved by exact branch-and-cut solvers. It then introduces an energy-based graph neural network framework that learns energy functions over combinatorial configurations and embeds them within mixed-integer inference problems, allowing hard constraints to be enforced at inference time.
For black-box process optimisation, the thesis develops reduced-space Bayesian optimisation methods that use global sensitivity analysis to identify influential process variables and optimise in physically interpretable reduced spaces. This is extended to multi-fidelity Bayesian optimisation, where low-fidelity approximations and high-fidelity simulator evaluations are combined under limited computational budgets.
Together, these contributions show how learning can improve the scalability and data efficiency of structured optimisation while preserving feasibility, constraint representation, and decision quality.
Version
Open Access
Date Issued
2026-06-18
Date Awarded
2026-09-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Papathanasiou, Maria
Sponsor
Imperial College London
Publisher Department
Department of Chemical Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)