automl / dac

Dynamic Algorithm Configuration

DAC

Code/Supplementary for the ECAI2020 paper
Dynamic Algorithm Configuration:
Foundation of a New Meta-Algorithmic Framework

If you used the presented benchmarks or framework in one of your research projects, please cite us:

@inproceedings{biedenkapp-ecai20,
  author    = {A. Biedenkapp and H. F. Bozkurt and T. Eimer and F. Hutter and M. Lindauer},
  title     = {Dynamic Algorithm Configuration: Foundation of a New Meta-Algorithmic Framework},
  booktitle = {Proceedings of the Twenty-fourth European Conference on Artificial Intelligence ({ECAI}'20)},
  year = {2020},
  month     = jun,
}

Installation using Anaconda

1. conda create -n dac python=3.7
2. conda activate dac
3. cat requirements.txt | xargs -L 1 -n 1 pip install
4. python setup.py install
5. (optional) conda install jupyter

Example calls

• Train 0.1-greedy tabular Q-learning on 1D-sigmoids with trinary actions for 10⁴ episodes:
  python dac/train/train_other.py --seed 0 -r 1 -n 10000 --epsilon_decay const -e 0.1 -l 1. --env 1D3M --out-dir 1D-trinary-action-Sigmoid
• Train DDQN on 1D-sigmoids with trinary actions for 10⁵/10⁴ steps/episodes:
  python dac/train/train_chainer_agent_on_toy.py --eval-n-runs 10 --eval-interval 10 --checkpoint_frequency 1000 --outdir 1D-trinary-action-Sigmoid/DQN --seed 0 --scenario 1D3M --steps 100000
• Run PS-SMAC on 1D-sigmoids with trinary actions for 10⁴ episodes:
  1. Find well performing parameter sequence:
     python dac/train/train_other.py --seed 0 -r 1 -n 10000 --epsilon_decay const -e 0.1 -l 1. --env 1D3M --out-dir 1D-trinary-action-Sigmoid --bo
  2. Validate the found sequence:
     python dac/train/train_other.py --seed 0 -r 1 -n 10000 --epsilon_decay const -e 0.1 -l 1. --env 1D3M --out-dir 1D-trinary-action-Sigmoid --bo --validate-bo 1D-trinary-action-Sigmoid/smac3*/run*

We provide a jupyter notebook with which you can inspect your own runs or the example results provided in example-results-1D-trinary-action-Sigmoid

License

This repository is licensed under the Apache License 2.0

Repository Structure

• cmds:
  Contains many more commands to run sigmoid experiments from the paper
• dac
  • envs:
    Contains all instance features and code for the presented white-box benchmarks
  • train:
    Contains code to train all the agents presented in the paper
• example-results-1D-trinary-action-Sigmoid
  • DQN:
    Result of the above example DQN call
  • smac3-output:
    Results of the call of PS-SMAC
  • tabular:
    Result of the above example tabular call
• Appendix.pdf
  The online appendix of the paper.
• example_plots.ipynb
  Jupyter-Notebook to plot the example results