Skip to content

chandar-lab/adaptive-hanabi

Repository files navigation

Adaptive Hanabi

Introduction

This repo contains the implementation of the benchmark described in our paper Towards Few-shot Coordination: Revisiting Ad-hoc Teamplay Challenge In the Game of Hanabi.

The codebase is mostly based on off-belief-learning repo.

hanabi-learning-environment is a modified version of the original HLE from Deepmind.

Environment Setup

Please refer to the setup instruction oon off-belief-learning repo.

Run the Code

To pre-train hanabi agents,

cd pyhanabi
sh scripts/iql.sh

To finetune an agent to a pre-traind agent, use the following script:

cd pyhanabi
sh scripts/adaptation.sh

Note that, before running the script, --load_model and --coop_agents should be specified that shows the path to directory of the learner and cooperative partners checkpoints.

Download Models

To download the trained models used in the paper, go to models folder and run

sh download_pool.sh

About

No description, website, or topics provided.

Resources

License

Code of conduct

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published