Atari#

Atari environments are simulated via the Arcade Learning Environment (ALE) [1].

Action Space#

The action space a subset of the following discrete set of legal actions:

Num

Action

0

NOOP

1

FIRE

2

UP

3

RIGHT

4

LEFT

5

DOWN

6

UPRIGHT

7

UPLEFT

8

DOWNRIGHT

9

DOWNLEFT

10

UPFIRE

11

RIGHTFIRE

12

LEFTFIRE

13

DOWNFIRE

14

UPRIGHTFIRE

15

UPLEFTFIRE

16

DOWNRIGHTFIRE

17

DOWNLEFTFIRE

If you use v0 or v4 and the environment is initialized via make, the action space will usually be much smaller since most legal actions don’t have any effect. Thus, the enumeration of the actions will differ. The action space can be expanded to the full legal space by passing the keyword argument full_action_space=True to make.

The reduced action space of an Atari environment may depend on the flavor of the game. You can specify the flavor by providing the arguments difficulty and mode when constructing the environment. This documentation only provides details on the action spaces of default flavors.

Observation Space#

The observation issued by an Atari environment may be:

  • the RGB image that is displayed to a human player,

  • a grayscale version of that image or

  • the state of the 128 Bytes of RAM of the console.

Rewards#

The exact reward dynamics depend on the environment and are usually documented in the game’s manual. You can find these manuals on AtariAge.

Stochasticity#

It was pointed out in [1] that Atari games are entirely deterministic. Thus, agents could achieve state of the art performance by simply memorizing an optimal sequence of actions while completely ignoring observations from the environment. To avoid this, ALE implements sticky actions: Instead of always simulating the action passed to the environment, there is a small probability that the previously executed action is used instead.

On top of this, Gym implements stochastic frame skipping: In each environment step, the action is repeated for a random number of frames. This behavior may be altered by setting the keyword argument frameskip to either a positive integer or a tuple of two positive integers. If frameskip is an integer, frame skipping is deterministic, and in each step the action is repeated frameskip many times. Otherwise, if frameskip is a tuple, the number of skipped frames is chosen uniformly at random between frameskip[0] (inclusive) and frameskip[1] (exclusive) in each environment step.

Flavors#

Some games allow the user to set a difficulty level and a game mode. Different modes/difficulties may have different game dynamics and (if a reduced action space is used) different action spaces. We follow the convention of [2] and refer to the combination of difficulty level and game mode as a flavor of a game. The following table shows the available modes and difficulty levels for different Atari games:

Environment

Valid Modes

Valid Difficulties

Default Mode

Adventure

[0, 1, 2]

[0, ..., 3]

0

AirRaid

[1, ..., 8]

[0]

1

Alien

[0, ..., 3]

[0, ..., 3]

0

Amidar

[0]

[0, 3]

0

Assault

[0]

[0]

0

Asterix

[0]

[0]

0

Asteroids

[0, ..., 31, 128]

[0, 3]

0

Atlantis

[0, ..., 3]

[0]

0

BankHeist

[0, 4, 8, 12, 16, 20, 24, 28]

[0, ..., 3]

0

BattleZone

[1, 2, 3]

[0]

1

BeamRider

[0]

[0, 1]

0

Berzerk

[1, 2, 3, 4, 5, 6, 7, 8, 9, 16, 17, 18]

[0]

1

Bowling

[0, 2, 4]

[0, 1]

0

Boxing

[0]

[0, ..., 3]

0

Breakout

[0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44]

[0, 1]

0

Carnival

[0]

[0]

0

Centipede

[22, 86]

[0]

22

ChopperCommand

[0, 2]

[0, 1]

0

CrazyClimber

[0, ..., 3]

[0, 1]

0

Defender

[1, ..., 9, 16]

[0, 1]

1

DemonAttack

[1, 3, 5, 7]

[0, 1]

1

DoubleDunk

[0, ..., 15]

[0]

0

ElevatorAction

[0]

[0]

0

Enduro

[0]

[0]

0

FishingDerby

[0]

[0, ..., 3]

0

Freeway

[0, ..., 7]

[0, 1]

0

Frostbite

[0, 2]

[0]

0

Gopher

[0, 2]

[0, 1]

0

Gravitar

[0, ..., 4]

[0]

0

Hero

[0, ..., 4]

[0]

0

IceHockey

[0, 2]

[0, ..., 3]

0

Jamesbond

[0, 1]

[0]

0

JourneyEscape

[0]

[0, 1]

0

Kangaroo

[0, 1]

[0]

0

Krull

[0]

[0]

0

KungFuMaster

[0]

[0]

0

MontezumaRevenge

[0]

[0]

0

MsPacman

[0, ..., 3]

[0]

0

NameThisGame

[8, 24, 40]

[0, 1]

8

Phoenix

[0]

[0]

0

Pitfall

[0]

[0]

0

Pong

[0, 1]

[0, ..., 3]

0

Pooyan

[10, 30, 50, 70]

[0]

10

PrivateEye

[0, ..., 4]

[0, ..., 3]

0

Qbert

[0]

[0, 1]

0

Riverraid

[0]

[0, 1]

0

RoadRunner

[0]

[0]

0

Robotank

[0]

[0]

0

Seaquest

[0]

[0, 1]

0

Skiing

[0]

[0]

0

Solaris

[0]

[0]

0

SpaceInvaders

[0, ..., 15]

[0, 1]

0

StarGunner

[0, ..., 3]

[0]

0

Tennis

[0, 2]

[0, ..., 3]

0

TimePilot

[0]

[0, 1, 2]

0

Tutankham

[0, 4, 8, 12]

[0]

0

UpNDown

[0]

[0, ..., 3]

0

Venture

[0]

[0, ..., 3]

0

VideoPinball

[0, 2]

[0, 1]

0

WizardOfWor

[0]

[0, 1]

0

YarsRevenge

[0, 32, 64, 96]

[0, 1]

0

Zaxxon

[0, 8, 16, 24]

[0]

0

Common Arguments#

When initializing Atari environments via gym.make, you may pass some additional arguments. These work for any Atari environment. However, legal values for mode and difficulty depend on the environment.

mode: int. Game mode, see [2]. Legal values depend on the environment and are listed in the table above.

difficulty: int. Difficulty of the game, see [2]. Legal values depend on the environment and are listed in the table above. Together with mode, this determines the “flavor” of the game.

obs_type: str. This argument determines what observations are returned by the environment:

  • “ram”: The 128 Bytes of RAM are returned

  • “rgb”: An RGB rendering of the game is returned

  • “grayscale”: A grayscale rendering is returned

frameskip: int or a tuple of two ints. This argument controls stochastic frame skipping, as described in the section on stochasticity.

repeat_action_probability: float. The probability that an action sticks, as described in the section on stochasticity.

full_action_space: bool. If set to True, the action space consists of all legal actions on the console. Otherwise, the action space will be reduced to a subset.

render_mode: str. Specifies the rendering mode:

  • “human”: We’ll interactively display the screen and enable game sounds. This will lock emulation to the ROMs specified FPS

  • “rgb_array”: we’ll return the rgb key in step metadata with the current environment RGB frame.

It is highly recommended to specify render_mode during construction instead of calling env.render(). This will guarantee proper scaling, audio support, and proper framerates

Version History and Naming Schemes#

All Atari games are available in three versions. They differ in the default settings of the arguments above. The differences are listed in the following table:

Version

frameskip=

repeat_action_probability=

full_action_space=

v0

(2, 5,)

0.25

False

v4

(2, 5,)

0.0

False

v5

5

0.25

True

Version v5 follows the best practices outlined in [2]. Thus, it is recommended to transition to v5 and customize the environment using the arguments above, if necessary.

For each Atari game, several different configurations are registered in OpenAI Gym. The naming schemes are analgous for v0 and v4. Let us take a look at all variations of Amidar-v0 that are registered with OpenAI gym:

Name

obs_type=

frameskip=

repeat_action_probability=

full_action_space=

Amidar-v0

"rgb"

(2, 5,)

0.25

False

AmidarDeterministic-v0

"rgb"

4

0.0

False

AmidarNoframeskip-v0

"rgb"

1

0.25

False

Amidar-ram-v0

"ram"

(2, 5,)

0.25

False

Amidar-ramDeterministic-v0

"ram"

4

0.0

False

Amidar-ramNoframeskip-v0

"ram"

1

0.25

False

Things change in v5: The suffixes “Deterministic” and “NoFrameskip” are no longer available. Instead, you must specify the environment configuration via arguments passed to gym.make. Moreover, the v5 environments are in the “ALE” namespace. The suffix “-ram” is still available. Thus, we get the following table:

Name

obs_type=

frameskip=

repeat_action_probability=

full_action_space=

ALE/Amidar-v5

"rgb"

5

0.25

True

ALE/Amidar-ram-v5

"ram"

5

0.25

True

References#

[1] MG Bellemare, Y Naddaf, J Veness, and M Bowling.
“The arcade learning environment: An evaluation platform for general agents.”
Journal of Artificial Intelligence Research (2012).

[2] Machado et al.
“Revisiting the Arcade Learning Environment: Evaluation Protocols and Open Problems for General Agents”
Journal of Artificial Intelligence Research (2018)
URL: https://jair.org/index.php/jair/article/view/11182