Atari¶
We use ale_py==0.7.5 as the codebase.
See https://github.com/mgbellemare/Arcade-Learning-Environment/tree/v0.7.5
Env Wrappers¶
Currently it includes these wrappers: random-noops / fire-reset /
episodic-life / frame-skip / action-repeat / image-resize /
reward-clip. The wrapper execution order is the same as
OpenAI Baselines.
Options¶
task_id (str): see available tasks below;num_envs (int): how many environments you would like to create;batch_size (int): the expected batch size for return result, default tonum_envs;num_threads (int): the maximum thread number for executing the actualenv.step, default tobatch_size;seed (int): the environment seed, default to42;max_episode_steps (int): the maximum number of steps for one episode, default to27000, which corresponds to 108000 frames or roughly 30 minutes of game-play (Hessel et al. 2018, Table 3) because of the 4 skipped frames;img_height (int): the desired observation image height, default to84;img_width (int): the desired observation image width, default to84;stack_num (int): the number of frames to stack for a single observation, default to4;gray_scale (bool): whether to use gray scale env wrapper, default toTrue;frame_skip (int): the number of frames to execute one repeated action, only the last frame would be kept, default to4;noop_max (int): the maximum number of no-op action being executed when calling a singleenv.reset, default to30;episodic_life (bool): make end-of-life == end-of-episode, but only reset on true game over. It helps the value estimation. Default toFalse;zero_discount_on_life_loss (bool): when the agent losses a life, thediscountin dm_env.TimeStep is set to 0. This option doesn’t affect gym’s behavior (since there is nodiscountfield in gym’s API). Default toFalse;reward_clip (bool): whether to change the reward tosign(reward), default toFalse;repeat_action_probability (float): the action repeat probability in ALE configuration, default to 0 (no action repeat to perform deterministic result);use_inter_area_resize (bool): whether to usecv::INTER_AREAfor image resize, default toTrue.use_fire_reset (bool): whether to usefire-resetwrapper, default toTrue.full_action_space (bool): whether to use full action space of ALE of 18 actions, default toFalse.
Observation Space¶
The observation image size should be (stack_num, img_height, img_width),
(4, 84, 84) by default. For a single frame, it has been gray-scaled and
resized inside the c++ code.
Action Space¶
Each Atari games has its own discrete action space.
Available Tasks¶
Note: Our Atari environments ALE settings follow gym’s *NoFrameSkip-v4
(with openai/baselines wrapper) instead of *-v5 by default, see the
related discussions at
Issue #14.
Adventure-v5AirRaid-v5Alien-v5Amidar-v5Assault-v5Asterix-v5Asteroids-v5Atlantis-v5Atlantis2-v5Backgammon-v5BankHeist-v5BasicMath-v5BattleZone-v5BeamRider-v5Berzerk-v5Blackjack-v5Bowling-v5Boxing-v5Breakout-v5Carnival-v5Casino-v5Centipede-v5ChopperCommand-v5CrazyClimber-v5Crossbow-v5Darkchambers-v5Defender-v5DemonAttack-v5DonkeyKong-v5DoubleDunk-v5Earthworld-v5ElevatorAction-v5Enduro-v5Entombed-v5Et-v5FishingDerby-v5FlagCapture-v5Freeway-v5Frogger-v5Frostbite-v5Galaxian-v5Gopher-v5Gravitar-v5Hangman-v5HauntedHouse-v5Hero-v5HumanCannonball-v5IceHockey-v5Jamesbond-v5JourneyEscape-v5Kaboom-v5Kangaroo-v5KeystoneKapers-v5KingKong-v5Klax-v5Koolaid-v5Krull-v5KungFuMaster-v5LaserGates-v5LostLuggage-v5MarioBros-v5MiniatureGolf-v5MontezumaRevenge-v5MrDo-v5MsPacman-v5NameThisGame-v5Othello-v5Pacman-v5Phoenix-v5Pitfall-v5Pitfall2-v5Pong-v5Pooyan-v5PrivateEye-v5Qbert-v5Riverraid-v5RoadRunner-v5Robotank-v5Seaquest-v5SirLancelot-v5Skiing-v5Solaris-v5SpaceInvaders-v5SpaceWar-v5StarGunner-v5Superman-v5Surround-v5Tennis-v5Tetris-v5TicTacToe3d-v5TimePilot-v5Trondead-v5Turmoil-v5Tutankham-v5UpNDown-v5Venture-v5VideoCheckers-v5VideoChess-v5VideoCube-v5VideoPinball-v5WizardOfWor-v5WordZapper-v5YarsRevenge-v5Zaxxon-v5