MetaMTO: Learn where, what and how to transfer knowledge in Evolutionary Multitasking through Deep Reinforcement Learning

Citing MetaMTO

MetaMTO is accepted by IEEE TEVC (IF=15.9, SCI Q1 Top). The PDF version of the paper is available here. If you find our work useful, please cite it in your publications or projects.

@article{zhan2026learning,
  title={Learning where, what and how to transfer: A multi-role reinforcement learning approach for evolutionary multitasking},
  author={Zhan, Jiajun and Ma, Zeyuan and Gong, Yue-Jiao and Tan, Kay Chen},
  journal={IEEE Transactions on Evolutionary Computation},
  year={2026},
  publisher={IEEE}
}

Requirements

Create the conda environment with python 3.9.23 and torch 2.7.1, then install packages:

conda create -n MetaMTO python=3.9.23
conda activate MetaMTO
pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Train

To train the model, run:

python main.py

For more adjustable settings, please refer to config.py for details.

Recording results: Log files will be saved to ./log, the file structure is as follow:

log
|--run_name
   |--logging files
   |--...

The saved checkpoints will be saved to ./saved_models, the file structure is as follow:

saved_models
|--run_name
   |--episode_0
   |--episode_1
   |--...

Rollout

Modify load_name (The run_name of the trained model) in config.py.

Then run:

python test.py

to rollout the trained models.