An python implementation for GLEET: "Auto-configuring Exploration-Exploitation Tradeoff in Evolutionary Computation via Deep Reinforcement Learning", which is accepted by GECCO 2024.
pip install -r requirements.txt
- To train GLEET
python run.py --train --problem=pro_id --dim=problem_dimension --backbone=backbone_algorithm [--run_name=run_name_to_specify --resume=resume_path]
where --train switch the running mode to training. --problem is the function id for training, which range from 1 to 10 in CEC2021 dataset. --problem_dimension is the dimension of running function, you can specify it as 10 or 30. --backbone specify the backbone algorithm for GLEET, where you can specify it to be PSO or DMSPSO. --run_name specify the running name for this run for better identification. --resume specify the loading path for GLEET if you want to resume training base on any former model.
An example can be:
python run.py --train --problem=1 --dim=10 --backbone=PSO --run_name=train_problem1_10dim
which means we start training GLEET with PSO as backbone algorithm in 10 dimensional 1st problem (Bent cigar to be specific) from scratch.
In addition, the training output model will be saved in ./outputs, and you can load from it if needed. The training logging output will be saved in ./logs.
- To rollout GLEET or the backbone algorithm
python run.py --test --problem=pro_id --dim=problem_dimension --backbone=backbone_algorithm [--load_path=path to load for GLEET]
where --test switch the running mode to testing. The meanings of problem, dim, backbone are the same as before. --load_path specify the loading path of some former trained model. If the load_path is given, the backbone algorithm will be running under the control of the loaded agent. If the load_path isn't given, the backbone algorithm will be running under the default configuration without any control.
An example can be:
python run.py --test --problem=1 --dim=10 --backbone=DMSPSO --load_path=outputs/func1_10.pt
which means we firstly load a pre_trained model from load_path and use it to control DMSPSO in 10 dimensional 1st problem (Bent cigar to be specific).
The PDF version of the paper is available here. If you find our GLEET useful, please cite it in your publications or projects.
@inproceedings{ma2024auto,
title={Auto-configuring exploration-exploitation tradeoff in evolutionary computation via deep reinforcement learning},
author={Ma, Zeyuan and Chen, Jiacheng and Guo, Hongshu and Ma, Yining and Gong, Yue-Jiao},
booktitle={Proceedings of the Genetic and Evolutionary Computation Conference},
pages={1497--1505},
year={2024}
}