I am trying to reproduce the results obtained by Dahm et al. in the paper Learning Light Transport the Reinforced Way.
This method takes advantage of the similarity between the Bellman equation (Q-Learning) and the rendering equation.
Here's the overview of the code authors inserted in the first version of their paper:
What is, in the first function,
idxPrev? Intuition suggests it is the index of the previous cell.
ray.o is probably the origin.
Moreover, in the same function, I understand that
getLastAttenuation() returns the color of
idxCurr. How is getLastAttenuation calculated?
The action space is discretized, apparently, dividing the hemisphere built on the hitting point in different stratums. According to what I understand, the Q-Learning algorithm finds the best action (so the best stratum) for each cell in the space, and then a ray is scattered randomly in that stratum. But a stratum covers 360 degrees, so I don't see how this is helpful to scattered a ray towards the light.
In the function
sampleScatteringDIrFromQtable(), the PDF is calculated based on the
ps (not clear) multiplied by the number of patches. So, the action is related to circular stratums of the hemisphere, or the number of patches?
I divide the hemispheres over the hitting points in 24 equally-sized patches. Then I scatter rays, for now, not based on a uniform distribution on the hemisphere, but through the center of these 24 equally-sized patched. In this case, what would the PDF be?