A search algorithm makes decisions of its own. It chooses which possibilities to examine, when to change direction, and how much effort to spend improving a solution. Designing those choices has traditionally involved mathematics, experiments, experience, and considerable invention.
Machine learning provides another way to explore the space of algorithmic ideas. It can learn which branching decisions work well, select among existing algorithms, discover useful configurations, and propose entirely new pieces of executable search logic. These are different levels of contribution, but together they suggest something intriguing: algorithms can help us design algorithms.
A search strategy has choices too
Imagine a local-search method improving a production schedule. It could swap two jobs, move one job later, or rearrange several operations together. Each kind of change defines a neighborhood, the set of schedules reachable by making those changes to the current schedule. The method must decide which neighborhood to explore and when to move on.
A human researcher may invent a rule based on the structure of the schedule. A learned policy might instead use observations from previous runs to predict which moves are promising. If the learned rule improves solution quality within the same total computation time, it has made the optimizer more effective. If evaluating the rule costs more than it saves, the apparent intelligence may add no practical benefit.
Each kind of change defines a neighborhood, the set of schedules reachable by making those changes to the current schedule.
The key is the entire search process, not the sophistication of an isolated prediction.
A continuum of algorithm development
Some systems learn configuration: settings such as neighborhood size or restart frequency. Others perform algorithm selection, choosing among alternative methods for a particular instance. Still others learn decisions inside a method, such as the next branch to explore or the next group of jobs to reconsider.
These activities lead naturally toward the generation of new procedures. A generative model can propose code implementing a heuristic, an algorithmic building block, or a different strategy for organizing search. The proposal becomes interesting when it is executed and evaluated, and when evidence from those evaluations helps guide the next attempt.
Calling configuration a lesser form of discovery misses the relationship between these levels. A particularly effective configuration can reveal a useful structural insight; a learned branching rule can become part of a novel algorithm; a generated program can combine familiar techniques in an unfamiliar way. What matters is what the resulting method actually does.
When candidate programs become research material
Systems such as FunSearch and AlphaEvolve have demonstrated that language models can propose executable ideas which are then tested by automated evaluators. FunSearch explored mathematical and algorithmic problems, including bin packing, while AlphaEvolve extended program-evolution techniques across a wider range of computational tasks. The important feature is the feedback loop: generate, evaluate, retain useful ideas, and generate again.
Testing such changes tells us whether the algorithm has captured something broad about the problem or merely adapted to one familiar setting.
This does not turn every generated program into a scientific discovery. A candidate may fail on larger instances, depend on peculiarities of the test set, or optimize a score that fails to represent the intended task. But such failures are part of how research identifies promising ideas. The most valuable results often come from examining why a method behaves differently, not just whether it tops one leaderboard.
What would make the discovery transferable?
Suppose a newly learned scheduling heuristic performs beautifully on factories with similar machines. Does it still help when machines have unique capabilities, setup times depend on the preceding product, or urgent work interrupts the plan? Testing such changes tells us whether the algorithm has captured something broad about the problem or merely adapted to one familiar setting.
This is where theoretical reasoning and empirical work meet. Mathematics can reveal properties of a method; benchmark experiments can show how it behaves; learning can help identify patterns too complex for one simple hand-written rule. They are complementary tools for discovering and understanding algorithms.
At 4colors, our interest is not limited to making established optimization software easier to configure. We develop algorithms, and learning-assisted methods offer additional ways to investigate how search can be structured, improved, and sometimes reimagined.
An optimization algorithm explores solutions. AI-assisted algorithm research explores possible ways of doing that exploration. The second search may be just as interesting as the first.