FunSearch (2023): a milestone in AI-generated heuristics

Welcome back to OR-Path!

This is the first edition of GenAI + OR Radar: a place to follow research, company applications and developments connecting generative AI with Operations Research.

I’m starting with a retrospective. In 2023, FunSearch offered a compelling example of language models contributing to algorithm discovery. Understanding that contribution gives us a useful reference for the papers and tools that follow.

So let’s get straight to it.

Researchers at Google DeepMind and collaborating universities published FunSearch online in Nature on December 14, 2023. Its journal citation is dated 2024, which explains why both years appear in references.

The central question was intriguing: could a language model help discover useful new algorithms through a process that tests its ideas?

Imagine items arriving one at a time. Each must go into a container with limited capacity, and the goal is to use as few containers as possible. This is online bin packing.

FunSearch generated code for the rule deciding where each item should go. An automated evaluator tested candidate rules, and promising programs informed further attempts. Researchers supplied the problem specification, evaluation and initial program structure.

For a concrete illustration, watch the authors’ bin-packing animation: Best Fit is on the left and FunSearch on the right. Notice the unused space as the items arrive.

The paper reported new mathematical constructions for the cap set problem and bin-packing heuristics that outperformed first fit and best fit on the evaluated datasets.

The research line also continued. Evolution of Heuristics, published at ICML 2024, explored evolving both natural-language ideas and executable code, comparing its method with FunSearch.

For me, this is a concrete reason to revisit the paper: it provides an early reference point for discussing how generative models can participate in designing optimization methods.

These experiments support claims about the tested settings. They do not establish a universally better heuristic or document a production deployment of the bin-packing method. Human choices about the problem, evaluator and search structure remained central.

When reading subsequent work, I would keep the comparison conditions visible: which problems were tested, against which methods, and with what computational effort?

FunSearch gives this series an accessible starting point: a familiar optimization problem, a surprising way to explore algorithms and results we can examine. That combination is what I want to look for as GenAI + OR Radar follows the field.

Until next time 👋
Ricardo
OR-Path