State size/Observation space

[rtang23]

Is there any command that will allow me to grab the state size or the observation space of the environment?

[Christian Kauten]

for the observation space, environments define an observation_space property (a proprietary gym object for spaces). To get the raw shape of the numpy array state use observation_space.shape. You can also sample from the space using observation_space.sample(). Heres a quick test script

from gym_super_mario_bros import make


env = make('SuperMarioBros-v0')
print(env.observation_space)
print(env.observation_space.shape)
print(env.observation_space.sample())

By state size, I'm assuming you mean the number of bytes required for a given state? There is no direct command but the calculation is as follows:

f = 256 * 240 * 3 = 184,320 # the number of bytes per frame (tensor)
r = 4  # the number of bytes for a reward (float)
d = 1 # the number of bytes for a done flag (boolean)
s = f + r + d = 184,325 bytes

I made a script to get the empirical state size from the

import sys
import numpy as np
from gym_super_mario_bros import make


env = make('SuperMarioBros-v0')
env.reset()
state, reward, done, info = env.step(0)


size = (
    np.prod(state.shape) * state.itemsize +
    sys.getsizeof(reward) +
    sys.getsizeof(done)
)


print('{}B per state'.format(size))

I get 184368B per state when I run the above snippet. It appears that both booleans and floats are actually allocated 24 bytes.

[rtang23]

You're a goddam hero!