Compatibility with MultiDiscrete spaces as observation_space

[hill-a]

Compatibility with MultiDiscrete spaces as observation_space

[araffin]

Hello,
Thanks for reporting the issue.
Could you please follow the issue template and provide a minimal yet complete code to reproduce the error?

[Guilherme-B]

Hi @araffin ,

Updated the main topic to your request. Sorry.

[araffin]

Hello,
Thanks for the update.
I suspect your issue is related to this one: openai/gym#1191
What is your gym version?

[Guilherme-B]

Hi again!

My gym version is: 0.10.5

I tried printing the shape upon env creation and on dummy_vec_env, just one line before the place where the error is thrown, and they output the same shape, but inline with the bug you identified:

Original shape: ((2, 2),)
Env: ((2, 2),)

Hsa the new version fixed it?

[araffin]

Are you sure you are using 0.10.5? Did you install it with pip?
On my computer, the following code raises an error (with this gym version):

spaces.MultiDiscrete([[0, 12], [0, 250]])

AssertionError: nvec should be a 1d array (or list) of ints

What you want to do is:

spaces.MultiDiscrete([12, 250])

no ? (12 and 250 discrete actions)

[Guilherme-B]

That's strange, yes I installed it via Pip and it is definitely 0.10.5.
I never got an error, it is how the Hands‑On Intelligent Agents with OpenAI Gym book defines the space is to be used, but, I have changed it to your recommendation (although I'm working with an observation space, not an action space but I take it its the same), and the problem persists, exactly as before.

[araffin]

I never got an error

I created a colab notebook so you can reproduce the error (with gym 0.10.5): https://colab.research.google.com/drive/1KDWriJMxBv8rtqjvGEu6Qz9bzZvtKt-B

With the latest version of gym (0.10.8), there is no error but the shape is changed (and that causes the error in the vec env): https://colab.research.google.com/drive/1oDOxpECm5mU-joHy_lnLwPGBQkgTaHCh

I have changed it to your recommendation

If you don't have any error with your original, then my recommendation won't work anyway.

[Guilherme-B]

Ah I see.

Should I then use an older gym version before the shape is changed?

[araffin]

Yes,
I would recommend using 0.10.5 (with my recommendation for declaring the space).
I am waiting for OpenAI answer. If breaking things (i.e. no backward compatibility) is what they wanted, we will have to patch stable-baselines, otherwise, we will wait for a fix from gym.

[Guilherme-B]

You are right.

All fixed up, thank you for your help. Hopefully they will fix their end soon enough.
Now to get SubprocEnv working, which seems to not work. I'll run some more testing and open a new issue if needed.

Once again, thank you.