NSF AI award $4M: an optimization institute for turning predictions into decisions (Georgia Tech, $20M estimated total)

1Optimization comes after prediction

What machine learning answers is "what is likely to happen." What an operation needs is "so what do we do," and those two have been handled by different technologies. The second belongs to optimization — the mathematics of finding the choice that best meets an objective under limited resources and constraints that must hold.

Routing deliveries, scheduling generation, planning production, allocating resources: all of them take that shape.

2What changes when AI and optimization are combined

Two motives connect them. One is that solving an optimization problem grows sharply more expensive as the problem grows, so machine learning is used to guess where good solutions lie and shorten the computation. The other is that when machine-learned predictions — demand, price, failure probability — become inputs to an optimization, the error in those predictions propagates into the quality of the decision.

The first is a question of speed; the second is a question of how error is handled when a prediction is handed to an optimizer. They are different in kind, and an institute is the chosen form because both span theory and application.

3Its place in the AI Research Institutes program

The record shows $4M obligated, a $20M estimated total, and a period from 2026-10-01 to 2031-09-30. Within the same AI Research Institutes program, the University of Washington's dynamic systems institute is recorded with an identical amount and period, and the same award date of 2026-08-14.

4What is likely, and what to do about it

Machine learning and optimisation answer different questions. This institute sits at the seam between them.

Two motives join them. One is to shorten optimisation by using machine learning to guess where good solutions lie. The other is to handle how error in a prediction propagates into the quality of a decision when that prediction becomes an input to the optimisation. The first is a problem of speed, the second of error, and they differ in kind. An institute is the form chosen because both span theory and application.