More complex neural nets

[rll]

More complex neural nets

[dementrock]

Hi @TheBirdie, thanks for your suggestions! We will update the interface of the policy so that it can optionally take a customized network, which should be closer to the interface we have for categorical policy, as shown here: https://github.com/rllab/rllab/blob/master/rllab/policies/categorical_mlp_policy.py

For splitting the inputs, there is a method on space instances called unflatten and a batched version called unflatten_n, which could be found e.g. here:

• https://github.com/rllab/rllab/blob/master/rllab/spaces/box.py#L57
• https://github.com/rllab/rllab/blob/master/rllab/spaces/product.py#L45

Are these similar to what you are looking for?

[dementrock]

I just pushed a commit with the aforementioned updates to the policy interface:

bc1b506

[TheBirdie]

Awesome! It looks exactly like what I was looking for.
Thank you very much for the work/support @dementrock!