I am curious on how ML outputs can be used afterwards on a optimization setting.
In other words, your machine learning model gives you a function $f$ and you want to optimize afterwards a function $g$ that accounts for $f$.
I have mainly two questions on that :
Propagation of error : Your machine learning model is not 100% accurate. Let's say on average it has a 5% mean absolute percentage error, what does that mean for your optimization outputs ? How can one measure the impact of "error propagation".
Constraints on model to consider : In many settings, tree-based models produce the best results. That being said, the output is not necessarily a function that can be plugged easily in an optimization setting to my understanding. I'd be interested in understanding what kind of models would be appropriate if the overall goal is optimizing and not predicting.
Any research articles references would be highly appreciated :)