Computational Rationality for Atari Games

This repository aims to provides agents that play Atari games with human-plausible active vision behavior.

Prerequisites

• Linux or MacOS system, Windows is NOT supported
• python (recommended version: 3.11)

Installation

All four installation steps are required

System dependencies

• The following system wide dependencies are needed. Exact names on non-debian distros may vary.
• cmake version 4 is not supported. Installl cmake version 3

sudo apt install cmake libx11-dev libglew-dev patchelf build-essential zlib1g-dev libglib2.0-0

Project Installation

git clone https://github.com/ciao-group/atari-cr
cd atari-cr
pip install -e .

ROMs

Install Atari ROMs by running

curl -L https://www.atarimania.com/roms/Atari-2600-VCS-ROM-Collection.zip -o x.zip && unzip x.zip
python -m atari_py.import_roms ROMS && rm -r ROMS 'HC ROMS' x.zip

or follow the steps described in https://github.com/openai/atari-py

Dataset

Download and decompress the Atari-HEAD dataset:

./data/Atari-HEAD/download.sh

Entry points

• Run python src/atari_cr/agents/dqn_atari_cr/main.py to train an agent. See python src/atari_cr/agents/dqn_atari_cr/main.py --help for viable arguments.
• Use python src/atari_cr/hyperparams.py to test multiple agents with different hyperparameters.
• Use ./run.sh to start hyperparameter tuning together with tensorboard.

Acknowledgement

We thank the authors of SugaRL from which this repository was forked.