Hi! I am having a great time using both Tune and RLlib, big thanks to you for helping me in my Master’s!
I often find myself working on problems that require some parameters to be optimized on a logarithmic scale (like learning rate). If I noticed it correctly PB2 does not allow to define custom distributions, only an interval
(min, max). This is understandable as the Gaussian process is not just simply samples from the interval but selects according to its own calculation, however it is a simple transformation of scale that can be done manually, but can also be automatized in some way, like adding an extra element in the interval tuple:
(min, max, "log"), or allow converting from certain tune search space objects.