Sym-Q

Inspired by how human experts refine and adapt expressions, we introduce Symbolic Qnetwork (Sym-Q), a novel reinforcement learningbased model that redefines symbolic regression as a sequential decision-making task. Sym-Q leverages supervised demonstrations and refines expressions based on reward signals indicating the quality of fitting precision. Its distinctive ability to manage the complexity of expression trees and perform precise step-wise updates significantly enhances flexibility and efficiency. Our results demonstrate that Sym-Q excels not only in recovering underlying mathematical structures but also uniquely learns to efficiently refine the output expression based on reward signals, thereby discovering underlying expressions. Sym-Q paves the way for more intuitive and impactful discoveries in physical science, marking a substantial advancement in the field of symbolic regression. This repository contains the official Pytorch implementation for the paper "Interactive Symbolic Regression through Offline Reinforcement Learning: A Co-Design Framework" and "Sym-Q: Adaptive Symbolic Regression via Sequential Decision-Making".

Workspace Setup

To create the symbolic regression workspace with the necessary repositories, follow these steps:

mkdir symbolic_ws
cd symbolic_ws
git clone https://github.com/SymposiumOrganization/NeuralSymbolicRegressionThatScales.git
git clone https://github.com/AILWQ/Joint_Supervised_Learning_for_SR.git
git clone https://github.com/Wenqi-Z/Sym-Q.git
cp -R NeuralSymbolicRegressionThatScales/src/nesymres Joint_Supervised_Learning_for_SR/src
mkdir -p Joint_Supervised_Learning_for_SR/Dataset/2_var/5000000/Train
mkdir -p Joint_Supervised_Learning_for_SR/Dataset/2_var/5000000/SSDNC

After running the above commands, your workspace should have the following structure:

symbolic_ws
│
└───NeuralSymbolicRegressionThatScales
│
└───Joint_Supervised_Learning_for_SR
│ 
└───Sym-Q

Environment Setup

Create and activate a Conda environment for Sym-Q:

conda create -n SymQ python=3.9
conda activate SymQ
pip install -e NeuralSymbolicRegressionThatScales/src/
pip install -r Sym-Q/requirements.txt
cd Joint_Supervised_Learning_for_SR
export PYTHONPATH="${PYTHONPATH}:$(pwd)"

Dataset Creation

Generate the dataset for training:

cd Joint_Supervised_Learning_for_SR/data_creation
python add_points_to_json.py
python gen_SSDNC_benchmark.py

Quick Start

Train the Sym-Q model:

cd Sym-Q
python train.py

This generates a folder /model/GENERATION_TIME with the model weights and evaluation data. Generate beamsearch results for evaluation:

python beam_search.py --weights_path PATH_TO_WEIGHTS --target ssdnc_r2

V2

To enhance flexibility and scalability, we have modularized the pipeline, allowing the model to utilize various encoders and handle inputs with different dimensions. For more detailed information, please refer to Sym-Q/symq_v2.

Acknowledgment

Our implementation is mainly refers to the following codebases. We gratefully thank the authors for their wonderful works.

SymbolicGPT, DSO, NeSymReS, T-JSL

Citing this work

If you found our work useful and used code, please use the following citation:

@article{tian2025co,
  title={Interactive Symbolic Regression through Offline Reinforcement Learning: A Co-Design Framework},
  author={Tian, Yuan and Zhou, Wenqi and Viscione, Michele and Dong, Hao and Kammer, David S and Fink, Olga},
  journal={arXiv preprint arXiv:2502.02917},
  year={2025}
}

@article{tian2024sym,
  title={Sym-Q: Adaptive Symbolic Regression via Sequential Decision-Making},
  author={Tian, Yuan and Zhou, Wenqi and Dong, Hao and Kammer, David S and Fink, Olga},
  journal={arXiv preprint arXiv:2402.05306},
  year={2024}
}