Issue
Primarily for debugging etc. It would be convenient to define a new "machine learning tool" type that just evaluates a given function. We will not directly address the dimension reduction handling here, so the user would have to know about this. But the idea is that one could e.g. define the actual forward map "G" in simple settings and pass this in.
Possible solution
- define a new tool called "UserDefinedMLT" which is built with a function,
- optimize does nothing and predict just calls the function and uses the result to produce a mean and covariance