III: Medium: Towards Data-Efficient Decision-Focused Learning: A Modularized Pre-Training and Fine-Tuning Approach

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

Many real-life decision-making tasks involve working with uncertainty, where decisions must be made to minimize costs affected by underlying factors that influence uncertainty. For example, in healthcare, allocating resources for diseases require anticipating future cases in various regions; in finance, funds rely on predicting future stock return rates to continually adjust their portfolios and maximize expected returns; and in wildlife conservation, patrol decisions to prevent poaching are based on predictions of poachers' behavior. To address underlying uncertainty issues, this project develops new methods for tailoring predictive models to capture the underlying uncertainty factors and integrate them into decision-making problems. This research will develop a general decision-focused learning framework, which will be applied in public health, environmental sustainability, and urban planning. Furthermore, this research will foster a diverse group of doctoral and undergraduate students at Georgia Tech who collaborate across disciplines, and create a graduate-level course focused on machine learning for decision-making at Georgia Tech. Part of the uncertainty issues incurrent decision-focused learning methods is based on data scarcity issues, due to challenges in task-specific training protocols, model misspecification, and data distribution shifts during deployment. This project addresses the pressing need for data-efficient decision-focused learning that can effectively solve stochastic decision-making problems (i.e., problems with high levels of uncertainty) with limited data, while addressing the key challenges of model misspecification and quickly adapting to changing distributions in the deployment phase. It will create a pre-training and fine-tuning learning paradigm that modularizes the key components in decision-focused learning for stochastic optimization. The specific objectives include: (1) pretraining key components to obtain general task-agnostic base modules that can be reused in different tasks; (2) fine-tuning the pretrained modules for specific decision-making tasks rapidly using task-specific data; and (3) developing active data collection techniques using model uncertainty to improve model performance and adapt the model to changing distributions. By exploiting broader data across multiple tasks during pre-training and requiring only a small amount of task-specific data for fine-tuning, this framework creates flexible predictive models that reduce model misspecification, improve data efficiency, and mitigate distribution shifts during deployment to produce better decisions. These research thrusts compose a general decision-focused learning framework. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.