env.step(action) not executing correct action

[kautenja]

env.step(action) not executing correct action

[Kautenja]

Ah, my thought is that you're running the environment without refining the action space (which is required as of 3.0.0 due to a new emulator in the back-end). Assuming you're looking for the standard action space similar to what is in 2.3.1 (not exactly the same), you'd need to refine your template to include a BinarySpaceToDiscreteSpaceEnv wrapper from the nes_py.wrappers package and pass in the SIMPLE_MOVEMENT action list from gym_super_mario_bros.actions:

from nes_py.wrappers import BinarySpaceToDiscreteSpaceEnv
import gym_super_mario_bros
from gym_super_mario_bros.actions import SIMPLE_MOVEMENT
env = gym_super_mario_bros.make('SuperMarioBros-v0')
env = BinarySpaceToDiscreteSpaceEnv(env, SIMPLE_MOVEMENT)

done = True
for step in range(5000):
    if done:
        state = env.reset()
    state, reward, done, info = env.step(env.action_space.sample())
    env.render()

env.close()

This is because the nes_py emulator allows the full range of 256 discrete actions on an NES joypad by default, and thus requires tuning to more realistic action spaces.

gym_super_mario_bros.actions defines three action lists (though you're welcome to create your own):

Name	Movement
RIGHT_ONLY	the agent can only run/walk/jump right
SIMPLE_MOVEMENT	the agent can do everything from: right only, walk left, and jump in place
COMPLEX_MOVEMENT	the agent can do everything from: simple movement, run/walk/jump left, and press down (enter pipes).

Hopefully this addresses your issue!

NOTE 3.0.0 has a serious bug as noted in #49, proceed at your own risk with this version. Might be best to wait until #49 is resolved - in 3.0.1 - to move up from 2.3.1. Hopefully that bug will be addressed sometime this week / week-end.

[griase94]

Allright good to know, thank you for your fast reply!
You should probably document this, since it took me several hours to find out that it was an issue which only happened in the latest Version ;)

[Kautenja]

Sorry for the confusion! I had meant to fix that bug much earlier, but got locked up with some work projects. However, I believe I have resolved the bug as of versions 3.0.3. Feel free to upgrade now, as >v3 uses a new back-end that is much much faster and should cut down training time considerably. Closing this issue as I believe it is resolved.