Keynote Talk: Learning What to Optimize: ML Methods for Accessible Operations Research

In decision-making problems from applied operations research—transportation, logistics, pricing, and healthcare operations—there is a productive tension between model-free approaches common in ML and AI, like deep reinforcement learning, and model-based approaches from OR, like stochastic programming.

Model-free approaches offer generality, can be quickly scaled to new problems, and promise to improve performance without human intervention by adapting to data. Yet they are data-hungry, can make surprising errors, and lack interpretability. Model-based approaches offer interpretability and dependability, but require an expert to hand-craft and refine an appropriate model for each problem. Decision quality is often limited by model fidelity.

This talk explores how modern ML methods can make model-based optimization more accessible by automating the most laborious part of applied OR: working with domain experts to specify objective functions and constraints. We focus on problems where decision-makers cannot articulate their goals as a single mathematical equation—they know good solutions when they see them, but cannot write down the utility function. This is the reality in many operational settings: exam scheduling (what makes a schedule "fair"?), logistics planning (how to balance speed, cost and risk?), and online platform design (how to balance user experience, revenue, and profit?).

We present a line of work on preference learning for optimization. Starting with Bayesian Optimization with Preference Exploration (BOPE), we show how Bayesian methods can learn decision-maker utility functions through pairwise comparisons of candidate solutions. We then describe recent advances with LISTEN (LLM-based Iterative Selection with Trade-Off Evaluation from Natural-language), which leverages large language models as zero-shot preference oracles, enabling decision-makers to express goals directly in natural language rather than through repeated comparisons.

We demonstrate the impact of these methods through real-world applications: automated video dubbing for Instagram Reels at Meta, exam scheduling at Cornell, and logistics decision support for the US Marines. We conclude with a broader vision: how AI methods can automate the traditionally time-consuming work of engaging with decision makers in applied OR, making powerful model-based optimization accessible to non-experts and dramatically reducing the bottleneck of expert analyst time.