Spatially periodic policies

[rmojgani]

Hi @wadaniel ,

I am thinking of a property of the policy that you might have thought about before, so decided to share my thoughts with you.

I have a one action per agent policy implemented, but then realized probably my policy should also be periodic, as my spatial domain is doubly periodic.

I think this should appear in my upsampling function:

RLonKorali/experiments/flowControl_turb_code/_model/turb.py

Line 472 in 39d2113

def upsample(self, action):

this is what I mean:
below is an image of contour plot of forcing, but as you can see the values on left and right of the domain are not continuous, but I was expecting them to be (since my domain is periodic)

Let me know what you think

[wadaniel]

Hi,
yes thats a good observation and I did this aswell. you must take care of it in the upsampling procedure (with upsampling, you mean interpolation on the finer grid, right?)

[rmojgani]

Hi,
(with upsampling, you mean interpolation on the finer grid, right?)
yes.

I am thinking of passing through a periodic map, e.g., sin(x) and sin(y), any better ideas ?

[wadaniel]

Hm I thought this could be handled much easier and just linearly interpolate the most right and the most left stencils on the coarse grid to introduce periodicity in x?

[rmojgani]

This is the update :
https://colab.research.google.com/drive/1JBs71VmJHwWQ89q3kAlLpyVON6CFVrUZ?usp=sharing