Skip to content

Latest commit

 

History

History
65 lines (41 loc) · 1.64 KB

Tips_Atari_Env.md

File metadata and controls

65 lines (41 loc) · 1.64 KB
Raw

OpenAI Gym

  • Link: https://gym.openai.com/

  • Original Paper: Bellemare, Marc G., et al. "The arcade learning environment: An evaluation platform for general agents." Journal of Artificial Intelligence Research 47 (2013): 253-279.

  • Num_Games: As of 4/4/2019, 797 Games are available

from gym import envs

all_envs = envs.registry.all()
env_ids = [env_spec.id for env_spec in all_envs]
print("Currently {} Games are available".format(len(env_ids)))

Naming rule of Environments

In Atari ALE, OpenAI follows some rules below

  • {}-v0
  • {}-v4
  • {}Deterministic-v0
  • {}Deterministic-v4
  • {}NoFrameskip-v0
  • {}NoFrameskip-v4
  • {}-ram-v0
  • {}-ram-v4
  • {}-ramDeterministic-v0
  • {}-ramDeterministic-v4
  • {}-ramNoFrameskip-v0
  • {}-ramNoFrameskip-v4

Details

v0 vs v4??

  • v0: repeat_action_probability = 0.25
    • Gym repeats the same action at the previous time step with the probability of 0.25
  • v0: repeat_action_probability = 0.0
    • Gym repeats the same action at the previous time step with the probability of 0.0

ram or not??

  • ram: Gym observes the RAM of the ALE environment and returns as an obsevation
  • others: Frame of the ALE environment is an observation

Simply it is better **not ** to use ram option

Deterministic/NoFrameskip or not

  • Deterministic: Gym always repeat the action four times
  • NoFrameskip: Gym does not repeat the action
  • others: Gym randomly repeat the action between 2 to 4 times

Tips

  • If you want to manually repeat the same action, then DO NOT use NoFrameskip
  • If you want to just focus on the algorithm without handling the repetition of actions, then use Deterministic