I am using Mathematica 11.2 with the CUDA paclet 11.2.22 on Windows 10 64-bit.
I have Microsoft Visual Studio 2015 Community.
Although CUDACCompliers gives an empty list, (with only VS 2015 installed) CUDAFunctionLoad works fine in Mathematica.
But I have just installed Visual Studio 2017 Community to program in C++ with the just released nVIDIA CUDA Toolkit 9.0.
The CUDA paclet 11.2 is still based on nVIDIA CUDA Toolkit 8.0 which does not support VS 2017.
CUDACCompliers is still an empty list, although VS 2015 and VS 2017 are both installed.
But CUDAFunctionLoad selects VS 2017 to compile, which fails for the reason I just mentioned.
How can I overrule Mathematica's choice for VS 2017? It should select VS 2015 for now...
I'm not one to answer, but check the directory (in your Home directory) that Mathematica uses to store your prefereces and also .m in Mathematica's program directory. (.m is the "simple text" version of .nb, you can say)
The choice very well may be in one of those .m, and if it is: you can edit it.
An alternate strategy (other than waiting for WR to fix it, which i assume you already requested by support email), would be to trick it: in your program folders or "program links" make 2015 to run where 2017 is found, tricking it to run 2015 instead of 2017. You'd then need a different "program link" if you were starting 2017 for some other purpose.