GP + Gradient Descent for Better Symbolic Regression Performance

Python implementation for embedding learnable parameters in GP programs and combining the standard GP evolutionary approach with a gradient-based refinement of the individuals employing the Adam optimizer. We propose two different algorithms that differ in how these parameters are shared in the expression operators.

link to the conference paper: https://drive.google.com/file/d/1AFN1ubYF5um8AVzI8l6RzPgzMufe16vZ/view?usp=sharing link to the poster: https://drive.google.com/file/d/17h8WIzMIrItwWuC3YuoZ6W8pcDYJ16if/view?usp=sharing

contacts: gloria.pietropolli@phd.units.it

Instructions

Code runs with Python 3.8 on Ubuntu 20.04.

To install the required libraries, enter the following command:

pip install -r requirements.txt 

To run the code, enter the following command:

python main.py --alg --dataset --save_dir --hyperparams_file --comp_budget --e_in_evo --e_after_evo --lr

where the inputs arguments stand for:

• --alg is the algorithm considered (that can be: gp, gpgda, gpgdc, opgda, opgdc)
• --dataset is the dataset selected for the training
• --save_dir is the directory where results are saved
• --hyperparams_file is the file containing hyperparameters for the training
• --comp_budget is the total computational budget for the training
• --e_in_evo is the gradient-based optimization performed during the evolution
• --e_in_evo is the gradient-based optimization performed after the evolution
• --lr is the learning rate that governs the gradient-based optimizer algorithm

that will return fitness results for the 150 runs performed and save them in the save_dir folder.

The code to reproduce all the box-plot of the paper (for all of the datasets) is contained in the folder analysis, it is sufficient to run:

python3 boxplot.py