Thesis Title:
Balancing Empiricism and Conservatism: A Scalable Approach for Data-Driven Decision Making under Uncertainty
Thesis Abstract:
In operations management, organizations need to schedule their production and investment to ensure efficiency and profit. These decisions rely on essential information such as customer demands, raw material prices, and exchange rates observed over time. However, access is limited because such information is uncertain or opaque. Hence, scientifically and quantitatively making decisions based on incomplete or historical information has garnered interest in both academia and industry.
There are two philosophies in decision-making under uncertainty: empiricism and conservatism. Empiricists advocate fully adapting decisions to observed data—empirical risk minimization (ERM). This approach can ignore sampling noise and distributional shifts. Hence, ERM can suffer significant performance issues because decisions fitted to historical data may perform poorly on future data. Conversely, conservatism argues for caution and risk hedging, ensuring decisions perform reasonably well even in worst-case scenarios. This tension motivates methods that balance fidelity to data with robustness.
Both empiricism and conservatism have evident drawbacks. Combining them is natural but quantifying this combination while ensuring scalability is subtle. In this project, we propose a scalable approach based on interpolated distributionally robust optimization (DRO). DRO is a popular framework for data-driven decision-making under uncertainty but often incurs high computational cost and limited tractability. We propose a relaxed version of DRO by interpolating between two extreme solutions, making it scalable to large-scale decision-making problems. Our interpolation tunes single parameter to trade off empirical fit and robustness. After providing performance guarantees, we will apply this approach to address widely concerned challenges such as healthcare resource allocation.
Primary Supervisor:
Prof Wolfram Wiesemann and Prof Ryan Cory-Wright