Propagate log of Categorical distribution parameters?

[bvdmitri]

@bartvanerp

We now propagate vectors with elements in [0, 1], but what if we would instead propagate the log of these elements for better robustness? I think this might also get rid of having to clamp the variables.

[bvdmitri]

@bartvanerp https://github.com/cjdoris/LogarithmicNumbers.jl

[bartvanerp]

Seems interesting, thanks for sharing!

[bartvanerp]

Following up on https://github.com/biaslab/RxInfer.jl/issues/94, we need to reduce the number of occurences of tiny / huge as much as possible. Getting rid of the clamping for Categorical distributions will be a first important step.