Projection Network for more than 1 action with differing action spaces

[sidney-tio]

Following the discussion from #37, I developed a MultiCategoricalProjectionNetwork that splits logits and creates the respective Categorical distribution. I tried to adhere to the same pattern as far as I could; It can be found here: https://gist.github.com/sidney-tio/66abada949f1b629dd9ee28777d402d5

If the team would like, I could raise a PR based on the gist I developed. From what I see, these are the to-dos to make it PR-worthy:

• [x] add tests
• [x] add more detailed docstrings
• [x] add masks

[sguada]

Have you tried instead having a nested action space? In which each action can have different number of actions?

[sidney-tio]

No, I don't think I have tried that. I assume you are referring to something like a gym.spaces.Dict type of nested structure where we could specify {'action1': 4, 'action2': 3}? Could you elaborate further?

[sguada]

Yeah you can use gym.spaces.Dict or directly nested ArraysSpecs to define the actions. Then each one can have their own Categorical distribution and sampling will sample all of them.

[sidney-tio]

my current workflow was to generate a spec from a gym.spaces.MultiDiscrete instance before creating the network.

I can see why something like a nested action space would be useful. I also just tried from a gym.spaces.Dict; would need loop through the iterable before generating the respective Categorical distributions.

i'll add a function to check for iterable and extract the relevant information. let me make the changes and, if its okay, I will raise a draft PR

[sidney-tio]

hello, not sure if it was missed, but the PR for this issue is up. Could I request for a review please?