Thesis

Robust decision-making under uncertainty : data-driven uncertainty and ambiguity set construction for robust optimisation and robust Markov decision processes

Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2026
Thesis identifier
  • T18226
Person Identifier (Local)
  • 202160648
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • In today’s dynamic and uncertain environments, Robust Decision-Making under Uncertainty is essential across diverse applications. However, conventional Robust Optimisation and Robust Markov Decision Processes suffer from overly conservative solutions. This is mainly due to simplistic or inaccurate uncertainty modelling. The heart of these fields is the proper construction of uncertainty and ambiguity sets that influence the outcome. In this thesis, we aim to explore these challenges by introducing novel, data-driven and machine learning approaches for constructing compact uncertainty sets. Through this, we seek to obtain significantly enhanced solution quality. Specifically, the thesis presents two methodological innovations. First, we study a novel uncertainty modelling approach for Robust Optimisation, utilizing a Bootstrapped Dirichlet Process Mixture Model to generate compact data-driven uncertainty sets that capture complex uncertainty distributions effectively. Second, we introduce an innovative framework for Robust Markov Decision Processes by integrating Bootstrapped Deep Q-Network ensembles with Bayesian ambiguity set construction. Consequently, we dynamically estimate both the centre and size of ambiguity sets from data in an online manner. Through extensive numerical experiments, the proposed methods demonstrate substantial improvements in solution quality, specifically in objective value for Robust Optimisation and worst-case return in Robust Markov Decision Processes, while outperforming conventional robust frameworks. We essentially addressed the fundamental challenge of overly conservative uncertainty modelling in robust decision-making by employing advanced statistical machine learning methods. Collectively, these contributions bridge predictive machine learning and prescriptive decision-making by translating predictive uncertainty into compact uncertainty sets that directly guide the Robust Optimisation and Robust Markov Decision Processes.
Advisor / supervisor
  • Liu, Bin
  • Akartunali, Kerem
Resource Type
DOI

Relations

Items