'Optimization under Uncertainty via History Matching and Statistical Emulation', a seminar by Dr Hailiang Du (Durham University)

Optimization under Uncertainty via History Matching and Statistical Emulation

Abstract

There have been increasing demands for solving optimization problems arising from complex physical systems. Conventional optimization approaches typically provide a deterministic solution for decision support based on computer simulators, but they are often unable to account for multiple sources of uncertainty. For complex high dimensional systems, further simplifications are commonly introduced to make optimization tractable, which can result in solutions that are suboptimal or misleading. Even when an optimization problem is well resolved, relying on a single optimal solution offers limited value for operational and long term planning. In this work, we introduce a general optimization framework for complex systems, treating optimization as a history matching problem addressed using statistical emulation and uncertainty quantification. Emulation constructs fast statistical approximations to expensive simulator based objective function evaluations, enabling efficient identification of candidate solutions. Uncertainty quantification is then used to characterise multiple sources of uncertainty associated with each candidate solution, implemented through Bayes linear analysis. A practical wind farm planning case study demonstrates that the proposed framework can overcome key limitations of conventional optimization approaches and provide a diverse set of near optimal solutions for more robust decision support.

About the speaker

Dr. Hailiang Du is an Associate Professor in the Department of Mathematical Sciences and Co-Director of the Institute of Hazard, Risk and Resilience (IHRR) at Durham University. He is also a Senior Visiting Fellow at the London School of Economics’ Data Science Institute and Global School of Sustainability. His research focuses on uncertainty quantification, probabilistic forecasting, machine learning, and nonlinear dynamical systems, with applications to weather and climate risk, energy systems, and decision-making under uncertainty.