š This shouldnāt be a competition, ML + OR, 7 ways AI will reshape OR
Local Optimum: short, imperfect-yet-useful ideas - Edition #15
Welcome to a new edition of Local Optimum: a short, imperfect-yet-useful collection of ideas related to optimization, decision-making, and applied Operations Research.
Letās dive in! šŖ
1) āļø This shouldnāt be a competition
The recent FICO post showing Xpress edging out DeepMindās AlphaEvolve on several open maths benchmarks is impressive in its own right.
It also reminds us weāre not really watching a head-to-head ābattleā, but two different paradigms at work.
AlphaEvolve is an end-to-end AI agent that evolves code with almost no human guidance, while Xpress is a purpose-built global-optimization engine that shines once an analyst has translated the real-world question into a mathematical model.
In practice, these approaches are complementary: let the solver deliver certifiably optimal answers when you can formalize the problem, and let AI heuristics explore the unknown when you canāt, or when you want to minimize human time in the loop.
Seeing both push the frontier is good news for everyone in Operations Research.
2) š ML + OR
The marriage between Machine Learning and Mathematical Optimization is here to stay.
There are multiple proofs of that, and one of them is this one about AI4OPT, where Junyang Cai, a PhD student. He says:
I develop data-driven strategies that guide the solverās internal decision-making. By learning from historical problem data and solver behavior, our machine learning models can predict effective branching strategies and other key actions. This leads to a more informed and efficient search process, reducing overall solution times.
Read the full email by clicking in the image (and consider subscribing to AI4OPT newsletter!):
3) š® 7 ways AI will reshape OR
Iāve been thinking a lot about how fast things are moving lately. Agentic AI, LLMs writing code, solvers inside chatbots... Itās tempting to believe that everything old is about to become obsolete.
But hereās the thing: some ideas age better than others.
In February, I published my predictions for how AI will transform Operations Research. Looking back, none of them feel outdated. In fact, the world seems to be moving right toward them.
So if you missed it, hereās your chance to catch up:
Deep Learning to the rescue of Linear Programming
Almost two months ago, I introduced you to how to integrate Deep Learning into OR pipelines.
One thing I said back then was:
But if I like one approach, it's using Graph Neural Networks for optimization problems. They have a lot of potential because we can model many optimization problems as graphs.
(spoiler alert: in a few weeks, we'll explore that option here in Feasible)
Now the time has come, and Iām pretty happy to say itās a collaborative article!
Weāll cover:
š Understanding the power grid and its optimization problems
šøļø The challenge of Linear Programming and how Graph Neural Networks help
ā” Performance gains in real-world problems
If you want to understand how to use GNNs to warmstart Linear Programming problems, this will be useful. See you Monday!
And thatās it for today!
If youāre finding this newsletter valuable, consider doing any of these:
1) š Subscribe to the full version: if you arenāt already, consider becoming a paid subscriber. Youāll get access to the full archive, a private chat group, and 30% off new products.
2) š¤š» Collaborate with Feasible. Iām always looking for great products and services that I can recommend to subscribers. Also, if you want to write an article with me, Iām open to that! If you are interested in reaching an audience of Operations Research Engineers, you may want to do that here.
3) š¤ Share the newsletter with a friend, and earn rewards in compensation. Youāre just one referral away from getting The Modern OR Engineer Playbook: Mindset, methods, and metrics to deliver Optimization that matters.
If you have any comments or feedback, just respond to this email!
Have a nice day ahead āļø
Borja.
