Foundation Models and Learning for Operational Decisions

Queues, inventories, service networks, and platforms make decisions that repeat at scale, and they produce data as they run. Analytical models, data-driven methods, and machine learning each answer part of the question of how to make those decisions well. This direction asks how learning can take on more of it: learning from data, from feedback, and from simulation, and, as models and data grow, learning across many systems at once.

Ask about this direction

The questions

Learning from data. How can models learned from operational data, such as demand forecasts across many products, feed directly into decisions like where to ship and what to stock?

Learning from feedback. When a system is too complex, or changes too much, for a policy to be derived in closed form, how can decisions be learned by acting and observing, and stay reliable when the environment shifts?

Learning to simulate. How can generative models capture the randomness of real operational data well enough that simulators built on them can be trusted for planning and what-if questions?

Learning at scale. As models and training data grow, can one model serve a whole family of related systems rather than one system at a time? Many operational systems can be simulated faithfully, which makes training data far easier to come by than in many other fields.

Work so far

2018Working paper

Demand Prediction, Predictive Shipping, and Product Allocation for Large-Scale E-Commerce

Demand forecasting across thousands of products, driving shipping and stocking.

2023ICML

Robust Situational Reinforcement Learning in Face of Context Disturbances

Reinforcement learning that stays robust when its environment shifts.

2022NeurIPS

An Adaptive Deep RL Method for Non-Stationary Environments with Piecewise Stable Context

Reinforcement learning that adapts when its environment changes abruptly.

2025Operations Research

A Doubly Stochastic Simulator with Applications in Arrivals Modeling and Simulation

Generative simulation of arrival processes whose rate is itself random, with a consistency guarantee.

2022NeurIPS SyntheticData4ML Workshop

Mind Your Step: Continuous Conditional GANs with Generator Regularization

Generating data under conditions never seen in training.

A foundation model for operational systems is to be released soon.