Evolution guided GFlowNets (EGFN)

This repository is based on Distributional-GFlowNets.

PyTorch implementation for our paper

Evolution guided generative flow networks.

Transaction of Machine Learning Research

Zarif Ikram, Ling Pan, Dianbo Liu

We train our in two ways. Using evolutionary algorithms, we evolve a population of agent parameters to learn the optimal parameter population 🤖🤖🤖 that maximize the reward signal. While evolution in the evolution step, the generated samples get stored in the prioritized replay buffer (PRB). Using the offline samples from the replay buffer and online samples from the current policy, we train a GFlowNets agent 🤖 using gradient descent.

Setup

First, clone our repository.

cd ./E-GFN

Install Anaconda environment in case not available. Run the following command next.

./setup.zsh

conda activate e-gfn

Hypergrid task

First, navigate to the directory.

cd ./grid

Examples to run the grid experiments

python run_hydra.py ndim=5 horizon=20 method=db_egfn n_train_steps=2500 replay_sample_size=16 seed=$seed R0=0.00001

Ablation studies

All ablations are in the scripts directory. To run long time horizon ablation, run

./scripts/sparsity.sh

Other ablations include-

./scripts/long_time_horizon.sh

./scripts/generalizability.sh

./scripts/ablation_population_size.sh

./scripts/gamma.sh

./scripts/ablation_elite_population.sh

./scripts/buffer_size.sh

To change the configurations for the experiment, simply change the configs/main.yaml file.

Antibody sequence task

First, navigate to the directory.

cd ./POAS

Examples to run the experiments

scripts/comparison.sh

hu4D5 CDR H3 design task

First, navigate to the directory.

cd ./HER

Examples to run the experiments

scripts/comparison.sh

Molecule task

First, navigate to the directory.

cd ./mols

Next change the proxy_path and bpath variable in the gflownet.py file.

Running the experiments

To run the exeriment, run the following command.

python gflownet.py obj=fm_egfn sample_prob=0.2

To change the configurations for the experiment, simply change the configs/main_gfn.yaml file.

Citations

If you use the code for your research, please cite:

@article{
	ikram2025evolution,
	title={Evolution guided generative flow networks},
	author={Ikram, Zarif and Pan, Ling and Liu, Dianbo},
	journal={Transactions on Machine Learning Research},
	year={2025},
	url={https://openreview.net/forum?id=UgZIR6TF5N},
}