I started writing mine and got lost in it so thoroughly that I realized - about 8 years later - that it was a bit too ambitious.
The title made it simple Adaptive monitoring.
I worked in a small AI/ML company as an "SRE" (back then we haven't even heard of that acronym), so it seemed like a good fit. Just throw in some predictors, train it on the alerting data (when the oncall person just ignores it and the alert is not solved then it was a real false positive, right?) and so on. Also based on confidence scores decrease sampling rate for stuff that doesn't tend to break. Oh and also the local agent should pull the some pretrained predictors from the controller so it can locally switch to higher sampling if some patterns occur, this would catch those pesky hard to debug transient bugs/errors.
...
truth be told I got lost in chapter 1, because I had no idea what to actually write and implement from all this.
so no thesis, but years later I got a nice part of an ADHD diagnosis out of it!