I've got a journal paper accepted in IEEE Compt. Int. Letters

That’s right! We published a journal paper for you, fine people working on or interested in Symbolic Regression. 🚀

Our paper, “Exhaustive Canonical Skeletons Generation: A Formal Dataset for Decomposable Symbolic Regression,” coauthored with Dr Aniko Ekart, has just been accepted for publication at IEEE Computational Intelligence Letters, a new journal of the IEEE Computational Intelligence Society.

Most dataset generation for training Transformer models for Neural SR relies on random sampling, which creates significant structural redundancy and incomplete functional coverage. To address this, we created ExCaUS: an exhaustive dataset of canonical “skeleton” expressions generated via a deterministic, constrained context-free grammar. These canonical skeletons subsume more specific functional forms, eliminating redundant expressions and avoiding repeated computational efforts. We also paired it with a numerical data sampling strategy that avoids numerical instabilities without needing artificial protected operators. Finally, to put SR algorithms to the test, we’re introducing ExCAUS-Bench, a rigorous 150-problem benchmark where current state-of-the-art methods still significantly struggle!

Check out the links below to learn more:

📄 Read the (extended) paper: https://arxiv.org/pdf/2610.20108

🛠️ ExCaUS dataset: https://github.com/ACAIRA-Aston/SRSkeletons

📊 ExCAUS-Bench benchmark: https://github.com/ACAIRA-Aston/UnivariateSRBenchmark