Diverse PSRO

Diverse PSRO

Diverse PSRO is a variation of the Policy Space Response Oracle algorithm which promotes training a behaviourally diverse set of policies by using the theory of determinantal point processes (DPPs). This approach allows to train less exploitable more diverse strategies as well as bringing a new geometrically interpretable way of measuring population diversity.

How to run Diverse PSRO

The code on this repository can be run by cloning the repository

git clone https://github.com/diversepsro/diverse_psro

Creating a new Anaconda environment

conda env create -f environment.yml
conda activate diverse_psro

You can now run Random Games of Skill by executing

python3 random_games_skill.py

You can now run Real World Meta-Games by executing

python3 spinning_tops_dpp.py

You can now run Non-transitive Mixture Model by executing

python3 non_mixture_model.py

Performance of Diverse PSRO

Diverse PSRO is evaluated in three different settings, each of them using a different version of diverse oracle.

Game	Oracle
Random Games of Skill	Diverse BR
Real World Meta-Games	Diverse BR
Non-transitive mixture model	Diverse gradient ascent

Random Games of Skill

Real World Meta-Games

Non-transitive mixture model