Skip to content

CyberAgentAILab/adaptively-perturbed-md

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Adaptively Perturbed Mirror Descent for Learning in Games

tl;dr

This paper proposes a novel variant of Mirror Descent that achieves last-iterate convergence.

Installation

pip install -r requirements.txt

Run Experiments

In order to investigate the performance of APMD in Three-Player Biased Rock-Paper-Scissors with full feedback, execute the following command:

# D_{psi}=KL G=KL
$ python main.py feedback=full n_trials=10 T=100000 game=three_biased_rps algorithm=APMD algorithm.learning_rate=0.1 algorithm.perturbation_strength=0.1 algorithm.random_init=True algorithm.regularizer=entropy algorithm.perturbation_divergence=kl algorithm.update_slingshot_freq=100
# D_{psi}=KL G=Reverse KL
$ python main.py feedback=full n_trials=10 T=100000 game=three_biased_rps algorithm=APMD algorithm.learning_rate=0.1 algorithm.perturbation_strength=0.1 algorithm.random_init=True algorithm.regularizer=entropy algorithm.perturbation_divergence=reverse_kl  algorithm.update_slingshot_freq=100
# D_{psi}=Squared L2 G=Squared L2
$ python main.py feedback=full n_trials=10 T=100000 game=three_biased_rps algorithm=APMD algorithm.learning_rate=0.1 algorithm.perturbation_strength=1.0 algorithm.random_init=True algorithm.regularizer=l2 algorithm.perturbation_divergence=l2 algorithm.update_slingshot_freq=20

To evaluate APMD via an experiment in Three-Player Biased Rock-Paper-Scissors with noisy feedback, execute the following command:

# D_{psi}=KL G=KL
$ python main.py feedback=noisy n_trials=10 T=100000 game=three_biased_rps algorithm=APMD algorithm.learning_rate=0.01 algorithm.perturbation_strength=0.1 algorithm.random_init=True algorithm.regularizer=entropy algorithm.perturbation_divergence=kl algorithm.update_slingshot_freq=1000
# D_{psi}=KL G=Reverse KL
$ python main.py feedback=noisy n_trials=10 T=100000 game=three_biased_rps algorithm=APMD algorithm.learning_rate=0.01 algorithm.perturbation_strength=0.1 algorithm.random_init=True algorithm.regularizer=entropy algorithm.perturbation_divergence=reverse_kl algorithm.update_slingshot_freq=1000
# D_{psi}=Squared L2 G=Squared L2
$ python main.py feedback=noisy n_trials=10 T=100000 game=three_biased_rps algorithm=APMD algorithm.learning_rate=0.01 algorithm.perturbation_strength=1.0 algorithm.random_init=True algorithm.regularizer=l2 algorithm.perturbation_divergence=l2 algorithm.update_slingshot_freq=200

Reference

Kenshi Abe, Kaito Ariu, Mitsuki Sakamoto, and Atsushi Iwasaki. Adaptively perturbed mirror descent for learning in games. In ICML, 2024

Bibtex:

@inproceedings{
abe2024adaptively,
  title={Adaptively Perturbed Mirror Descent for Learning in Games},
  author={Abe, Kenshi and Ariu, Kaito and Sakamoto, Mitsuki and Iwasaki, Atsushi},
  booktitle={ICML},
  year={2024}
}

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages