This new algorithm pairs well with another tool we just released bsky.app/profile/foli...

This paper will be presented at #GECCO2025 but you can already read the pre-print at arxiv.org/abs/2501.17859

The source-code and binaries are available at github.com/folivetti/sr... together with additional info.

#GECCO2025 #SymbolicRegression #GeneticProgramming

This tool is compatible with a lot of popular SR algorithms such as 𝐎𝐩𝐞𝐫𝐨𝐧, 𝐏𝐲𝐒𝐑, 𝐁𝐢𝐧𝐠𝐨, 𝐅𝐄𝐀𝐓, 𝐈𝐓𝐄𝐀, 𝐓𝐈𝐑, 𝐆𝐏-𝐆𝐎𝐌𝐄𝐀, and many more!

It is especially compatible with the new algorithm eggp (bsky.app/profile/foli...)

You can also search for the top building blocks by frequency and average fitness. This can help unlock new discoveries from the knowledge built during the search.

A little demo of what it can do:

asciinema.org/a/713509

Once you load the expressions, you can explore them by requesting the top-N expressions associated with different filters: by size, by number of parameters or ... 𝐛𝐲 𝐩𝐚𝐭𝐭𝐞𝐫𝐧 𝐦𝐚𝐭𝐜𝐡𝐢𝐧𝐠 parts of the expressions!

With this tool, you can load a set of symbolic models obtained from any combination of multiple runs from the same or different algorithms, or with different hyper-parameters, using only the returned Pareto front or the entire search history.

This work will be presented at #GECCO2025 but we already got an arXiv version available at arxiv.org/abs/2501.17848 and the source-code at github.com/folivetti/sr...

All the experiments can be replicated with github.com/folivetti/eg...

This improves the results to such an extent that a traditional GP (as in tinyGP) can compete with the popular Operon and PySR algorithms.

The current implementation comes with many customization options including exporting the generated e-graph for further exploration.

For example, if in the first generation the search visits the expression (x+x) it will avoid generating this expression again in future generations and most of their equivalents, such as 2x, 3x/x, 3x - x, etc.

Tell me about it :-D We're currently fighting to split the course into CS and non-CS version. But I'm already happy that we don't teach this course in Java anymore :-P

The main challenge here is to make them understand how a function works, what are variables, and why the order of instructions matter :-) not the easiest of the courses to teach...

Here we have the course "Information Processing" (which I'm teaching right now) that is basically programming 101 for a mix of students (CS and non-CS). The main goal here is to teach the very basics so each student is prepared to learn their respective niche language when time comes.

I feel the same, and not only as a reviewer, as an author as well, especially when we get good reviewers :)

is it really useful in practice to have explicit feature transformation in these templates? (e.g. f(x1^2,x1*x2)). It seems to me that the user is doing the SR job :-P Wouldn't a syntax like `p1 * x1 + f1 - sin(f2)` be enough and cleaner?