Bridging the Optimization Skills Gap With Generative AI

Despite a growing level of interest in mathematical optimization and decision intelligence, one thing that has been holding the field back is a significant skills gap—a topic I explored in greater depth in my last post.

Of course, building an optimization model is a specialized skill. In fact, it’s a set of skills: understanding a complex business problem, identifying an opportunity for improvement, building a model that captures that opportunity, and then assessing and refining that model based on the observed results.

Educational resources—including extensive examples, regular webinars, case studies and training events—can go a long way to bridge that skills gap and teach new learners how to get started with optimization on their own. However, such resources can only do so much regarding more advanced topics.

One relatively new technology that has the power to help close this gap is generative artificial intelligence (GenAI). Tools like ChatGPT have the potential to answer highly technical questions, help people build applications or even build the bulk of an application by itself. At Gurobi, we’ve recently begun to look deeper at where this technology can be applied in our field, and here’s what we found:

Using GenAI To Answer Technical Questions

The first question to consider is a fairly simple one: Can GenAI answer detailed technical questions from our domain?

We’ve seen a lot of evidence that it can provide answers to textbook problems, but those are often phrased so carefully that they are just begging to be answered correctly by a machine. To make things a bit more interesting, we chose a sampling of user questions from our community forum, where the questions aren’t usually stated quite as meticulously as in a textbook.

ChatGPT 3.5, released in November 2022, gave us underwhelming results. We found that it rarely produced accurate answers, and it stated the incorrect answers as confidently as it did the correct ones.

ChatGPT 4 was released not long after 3.5, but to our surprise, it produced correct answers to almost all of the user questions we fed it. The results were quite impressive and serve as an indication that this has great potential to be an invaluable new resource for those learning optimization.

If AI Can Do The Work For Us, Why Learn Optimization?

If ChatGPT can answer complex mathematical optimization questions, the obvious next question is, “To what extent will this new technology reduce or even eliminate the need to learn optimization?”

It’s always a bit dangerous to make predictions about something that is evolving as rapidly as GenAI, especially given the two very different experiences we had with versions that were released just a few months apart. With that caveat in mind, I’ll share my own recent experiences and also touch on more fundamental limits.

One seemingly significant opportunity to reduce the need for optimization modeling expertise is in programming assistants that make suggestions as you write code, one of the most prominent I’ve found being GitHub Copilot. These assistants can make the more mechanical aspects of software development less tedious, from automatically correcting syntax errors to reducing the need for repetition. (I’d be quite happy if I never have to implement bucket sort again.)

Nonetheless, I find that these tools aren’t a great help when it comes to more complex tasks. Copilot may reduce the amount of typing I have to do, but not the amount of thinking.

Enhancing (Not Replacing) Optimization Skills

The capabilities of GenAI are expanding every day, but one area where it seems unlikely to have a significant impact is the interface between software and the real world.

A big part of optimization modeling is understanding a business process, identifying an opportunity for improvement and making sure that the optimization model exploits this opportunity—not just when it is first deployed, but also over time. This requires communication (typically with multiple people), gathering the appropriate data (often from multiple, disparate sources) and, ultimately, experience. That’s going to be tough for an AI system to replicate.

We shouldn’t ignore the potential for future AI systems to translate descriptions of the problems that need to be solved into optimization models that can solve them. They can do this now with textbook problems, and their abilities in this area will only improve over time. We expect that these abilities will complement the skills of human experts, making them more productive while also helping them to continue refining their skills (and thereby narrowing the current skills gap).

While these tools won’t replace optimization software (since they solve very different problems), we expect they will help to make optimization tools more effective, because better models make it easier to find good solutions.

This article was originally published on Forbes.com.