Seminars & Events

CSE Community Seminar | September 12, 2025

Presenter Swati Gupta, Associate Professor, MIT Sloan School of Management, MIT

Navigating Underspecified Optimization Problems using Portfolios for Human-AI Augmented Decision-Making

Abstract

Optimization has long furnished us with precise solutions when objectives and problem parameters are clearly defined. Yet with growing human–AI interactive systems and decision-makers facing real-world challenges—from healthcare to disaster management to infrastructure planning—the stated problem is often ambiguous (e.g., human to agentic AI: “prioritize allocation of resources to those with higher need”). In these regimes, optimization should not aim for a single answer, resorting to a black-box formulation, but rather serve as the foundation for constructing meaningful portfolios of solutions—structured menus of diverse, provably high-quality options that represent the space of potential problem formulations. This is a key step towards democratizing access to optimization and AI.

I will present recent work on approximation algorithms for combinatorial optimization problems, such as facility location (EC 2023) and scheduling (SODA 2025), and demonstrate how these ideas can be extended to multi-stakeholder reinforcement learning (ICML 2025). Each of these results constructs a portfolio to provably approximate a class of generalized p-means or ordered norms. Portfolios provide compact yet representative sets of solutions that capture competing objectives (e.g., efficiency v/s fairness) and trade-offs. Looking forward, I will discuss ongoing work on bringing these ideas to core optimization algorithms like online mirror descent. I will discuss recent results on constructing a portfolio of mirror maps, where learning across maps yields polynomial (in dimension) improvements in regret compared to classical Euclidean and entropic settings.

Together, these results illustrate how optimization can expand beyond single-solution paradigms to provide the theoretical scaffolding for AI systems—creating menus of possibilities that support robust, adaptive, and participatory decision-making. This talk is based on multiple joint works with Mohit Singh, Milind Tambe, Jai Moondra, Cheol Woo Kim, Shresth Verma, Madeleine Pollack, and Lingkai Wong.