cavalab/srbench

What it solves

SRBench addresses the lack of standardized, reproducible, and strong benchmarks in the field of symbolic regression (SR). It aims to bridge the gap between the genetic programming (GP) and machine learning (ML) communities by providing a unified framework to compare modern SR methods against strong baselines using large, diverse datasets.

How it works

The project provides a living benchmark that evaluates multiple symbolic regression methods under a unified experimental setup. To ensure consistency and reproducibility, each method is run from a Docker container with hyperparameter tuning and multiple independent runs per dataset. To maintain a unified interface, contributing methods must follow a scikit-learn compatible API.

Who it’s for

Researchers and developers in the ML and GP communities who are developing new symbolic regression algorithms or seeking reliable baselines for comparison in their research papers.

Highlights

• Extensive Method Roster: Evaluates up to 25 different symbolic regression methods (in the 2025 edition).
• Diverse Datasets: Uses datasets from the Penn Machine Learning Benchmarks (PMLB) and first-principles regression problems.
• Reproducible Setup: Employs Docker containers for environment isolation and consistent execution.
• Open Source: All experiment code and raw results are publicly available for the community.