Visualize estimated distributions of the TPE

[Marvin Kastner]

Dear all,

Has anybody tried to visualize the estimated continuous or discrete distributions of the TPE?

For educational purposes it could be great to actually peek inside the working algorithm and show people what is happening under the hood.

My first guess is that one needs to inject some own code into the `build_posterior` function at https://github.com/hyperopt/hyperopt/blob/master/hyperopt/tpe.py#L650 which will dump the required information that can be used to draw the graphs in a second step. Alternatively, one could heavily sample from the two distributions (in the code called a and b, in the paper l and g) and use that data as a basis for the visualization. Maybe somebody has an even better idea? Something I missed in the code?

I am looking forward to ideas and inspirations of how to approach this - I will share my result in case it works of course. Thanks a lot in advance!

Best wishes

Marvin