Atari Environments

Basketball Pong

Boxing

Combat: Plane

Combat: Tank

Double Dunk

Entombed: Competitive

Entombed: Cooperative

Flag Capture

Foozpong

Ice Hockey

Joust

Mario Bros

Maze Craze: Blockade

Othello

Pong

Quadrapong

Space Invaders

Space War

Surround

Tennis

Video Checkers

Volleyball Pong

Warlords

Wizard of Wor

Overview

The Atari environments are based off the Arcade Learning Environment. This environment was instrumental in the development of modern reinforcement learning, and so we hope that our multi-agent version of it will be useful in the development of multi-agent reinforcement learning.

Games Overview

Most games have two players, with the exception of Warlords and a couple of Pong variations which have four players.

Environment Details

The ALE has been studied extensively and a few notable problems have been identified:

Preprocessing

We encourage the use of the supersuit library for preprocessing. The unique dependencies for this set of environments can be installed via:

pip install supersuit

Here is some example usage for the Atari preprocessing:

from supersuit import resize, frame_skip, frame_stack, sticky_actions
from pettingzoo.atari import space_invaders_v0

env = space_invaders_v0.env()

# repeat_action_probability is set to 0.25 to introduce non-determinism to the system
env = sticky_actions(env, repeat_action_probability=0.25)

# downscale observation for faster processing
env = resize(env, (84, 84))

# allow agent to see everything on the screen despite Atari's flickering screen problem
env = frame_stack(env, 4)

# skip frames for faster processing and less control
# to be compatable with gym, use frame_skip(env, (2,5))
env = frame_skip(env, 4)

Common Parameters

All the Atari environments have the following environment parameters:

<atari_game>.env(obs_type='rgb_image', full_action_space=True, max_frames=100000)

obs_type: There are three possible values for this parameter:

full_action_space: the effective action space of the Atari games is often smaller than the full space of 18 moves. Setting this to False shrinks the available action space to that smaller space.

max_frames: the number of frames (the number of steps that each agent can take) until game terminates.

Citation

Multiplayer games within the Arcade Learning Environment were introduced in:

@article{terry2020arcade,
  Title = {Multiplayer Support for the Arcade Learning Environment},
  Author = {Terry, Justin K and Black, Benjamin},
  journal={arXiv preprint arXiv:2009.09341},
  year={2020}
}

The Arcade Learning Environment was originally introduced in:

@Article{bellemare13arcade,
  author = { {Bellemare}, M.~G. and {Naddaf}, Y. and {Veness}, J. and {Bowling}, M.},
  title = {The Arcade Learning Environment: An Evaluation Platform for General Agents},
  journal = {Journal of Artificial Intelligence Research},
  year = "2013",
  month = "jun",
  volume = "47",
  pages = "253--279",
}

Various to the Arcade Learning Environment were introduced in:

@article{machado2018revisiting,
  title={Revisiting the arcade learning environment: Evaluation protocols and open problems for general agents},
  author={Machado, Marlos C and Bellemare, Marc G and Talvitie, Erik and Veness, Joel and Hausknecht, Matthew and Bowling, Michael},
  journal={Journal of Artificial Intelligence Research},
  volume={61},
  pages={523--562},
  year={2018}
}