GESR: A Geometric Evolution Model for Symbolic Regression

Abstract | Setup | Run the model | Dependencies | Examples |

This repository is the official implementation of "GESR"

Abstract

In this work, we propose a Geometric Evolution Symbolic Regression(GESR) algorithm. Three key modules are presented in GESR to enhance the approximation: (1) a new semantic gradient concept, proposed from the observation of inaccurate approximation results within semantic backpropagation, to assist the exploration and improve the accuracy of semantics approximation; (2) a new geometric semantic search operator, tailored for efficiently approximating the target formula directly in the sparse topological space, to obtain more accurate and interpretable solutions under strict program size constraints; (3) the Levenberg-Marquardt algorithm with L1 regularization, used for the adjustment of expression structures and the optimization of global subtree weights to assist the proposed geometric semantic search operator.

Setup

Prerequisites

• Supported Operation Systems: Linux
• CUDA driver version >11.8

Install dependencies

Using conda and the environment.yml file:

• Run conda env create -n GESR -f environment.yml.

Run the model

To launch a model training, a command without any additional arguments can be used, which will launch a model training for livermore datasets:

python ./autorun.py

You can also use additional arguments to train the specified datasets or specify the save path, like:

python ./autorun.py -train_dataset ./Dataset/easy/train/ -test_dataset ./Dataset/easy/test/ -save_file ./result/Feynman_easy_result.txt -seed 1111

Dependencies

• python3
• numpy
• pycuda

Examples

When you run the commands before in a terminal, you will get the following prints if it works successfully:

License

The majority of this repository is released under the MIT license as found in the LICENSE file.