Writing an Environment about rllab HOT 6 CLOSED

Hi @cynthiatliu, sorry but I don't totally understand your setup. Do you mean that the action is either a discrete choice (1 out of N), or a two-dimensional continuous number?

You will probably need to implement a new action type, similar to how they are defined right now:

• https://github.com/rllab/rllab/blob/master/rllab/spaces/discrete.py
• https://github.com/rllab/rllab/blob/master/rllab/spaces/box.py
• https://github.com/rllab/rllab/blob/master/rllab/spaces/product.py

Your scenario is more like a union type rather than a product type. Since you need to be able to convert to/from a vector representation (see flatten and unflatten), one possible scheme is that you use the first entry in the vector to denote whether the action is a discrete choice or a 2D vector. The subsequent entries encode the actual choice.

from rllab.

Yes, that's what I mean. And thanks! I'll work on it.

from rllab.

Hi,

For this action space, how should the method "new_tensor_variables" be overwritten? Thanks,

from rllab.

Depending on your representation. If it's either a discrete or a 1D vector, you can use the same code as the Box type: https://github.com/rllab/rllab/blob/master/rllab/spaces/box.py#L71

from rllab.

It's either a discrete or a 2D vector.

Additionally, how should the actor distinguish between the types of actions? Part of the actor's action is to choose what type of action, as you know--does that create an extra dimension?

from rllab.

Solved the problems 👍 Thanks!

from rllab.