Collaborative Research: Operations-Driven Machine Learning

点击翻译按钮获取中文摘要

This award will contribute to the Nation's prosperity and welfare by capitalizing on the increased availability and accessibility of data to improve operational decision making. Operational decisions are ubiquitous in all aspects of the commercial economy, and even incremental improvements in operations can have major impacts in the competitiveness of such sectors as transportation, logistics, healthcare delivery, supply chain management. Similarly, public sector service operations involve decision making to wisely invest limited public resources. The ongoing data revolution has created great opportunities for leveraging large scale data to improve operational decision making. This award will support research in new techniques to make effective use of these data in the management of operations. This project provides a broadly applicable framework for addressing operational decisions and will result in improved performance and efficiency in practice. The project will involve outreach engagements with diverse organizations, including a nonprofit foster care agency. Current operational decision-making often involve two significant challenges: prediction and optimization. These tasks are usually addressed sequentially: key parameters are first predicted using modern statistical machine learning tools, and then planning decisions are made using these predictions within a complex optimization model. This project advances a new, broadly applicable framework, called Smart "Predict, then Optimize" (SPO), that effectively addresses the prediction and optimization challenges in tandem. In this new framework, operational performance is measured by the true objective value of the solutions generated from the predicted parameters. This project investigates the statistical and computational properties of novel loss functions in the SPO framework, including convex surrogates as well as non-convex formulations. The project will also develop new algorithms for training machine learning models, such as linear models, logistic models, and decision trees, using the new loss functions, and will extend the SPO framework to handle regularization, robustness, different data primitives, and dynamic data collection with exploration-exploitation tradeoffs.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.