Robust Symbolic Regression for Dynamical System Identification

Authors: Ramzi Dakhmouche, Ivan Lunati, Hossein Gorji

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The numerical experiments demonstrate the competitive performance of SDFL in comparison to the state-of-the-art. We illustrate the performance of the proposed scheme on the prototypical problem of Kuramoto networks and a standard benchmark of single-cell RNA sequence trajectory data.
Researcher Affiliation Academia Ramzi Dakhmouche EMAIL Institute of Mathematics, EPFL Laboratory for Computational Engineering, EMPA Ivan Lunati EMAIL Laboratory for Computational Engineering, EMPA Hossein Gorji EMAIL Laboratory for Computational Engineering, EMPA
Pseudocode Yes Algorithm 1 Symbolic Distribution Flow Learner Algorithm 2 Symbolic Distribution Flow Learner [extended description]
Open Source Code Yes Additionally, to foster reproducibility, a Python implementation of SDFL has been made public at https://github.com/Ramzisofo/SDFL.
Open Datasets Yes Then, we conduct an evaluation on a real-world dataset of embryoid stem cell trajectories (Moon et al., 2019).
Dataset Splits No The paper mentions 'unseen data' and 'Training Sample Size (TSS) per snapshot' with specific sample sizes for evaluation (e.g., m=50), but does not provide specific details on how the overall datasets (Kuramoto, single-cell RNA-seq) were split into training, validation, or test sets using percentages, counts, or specific methodologies.
Hardware Specification Yes For a fair comparison, all the reported running times are obtained on an Intel(R) Core(TM) i7-7500U CPU.
Software Dependencies No The paper mentions a "Python implementation of SDFL" and "publicly available implementations of JKOnet and Trajectory Net" but does not specify any version numbers for Python or any specific libraries/solvers used.
Experiment Setup Yes For the implementation of SDFL, we set the building operations consisting of {+, , , , cos, sin, exp} with a maximum of L = 20 operations per expression, with a number of episodes of 500 to 1000. For the recovery of the Kuramoto system, we use 15 snapshots with time-stamps ti = 2i for 1 i 15, and we set K = 1/3. ... For JKOnet, we use a small regularization parameter ε = 0.001 to make its target closer to the Wasserstein distance.