The NN approach sounds interesting, but there are no technical details and it might be hype-based marketing. Is anyone familiar with solid published work on applying ML/AI to optimize network control (or, as done here, optimize application parameters depending on network conditions).
Funnily enough, 10 years ago as a required entry step before being able to do a thesis I scribbled something[0] about distributed components with probes (some automatic, some with user feedback exactly as suggested here) and knobs driven by some fuzzy logic feedback loop, to be e.g applied to codecs and buffer size in an A/V streaming scenario† (possibly bidirectional). Concluded by saying you could get better, self-improving results by using genetic algorithms and/or AI such as NN. Was basically laughed at during defense and dismissed as being way too sci-fi. Got sick of such constant condescending attitudes and dumped research, pursuing my career as a software engineer. Had I pursued I could have been the author on such a paper you're looking for! Pretty sure I could have patented the shit out of this anyway but that's not how I roll regarding software patents. Now I lost the LaTeX source and Java code as well as the resulting PDF, and can't even download the thing as a memento since it's behind some stupid paywall.
† turned out MS launched instant play adaptative 1080p streaming some years later on Xbox 360, which was exactly that.
[0] Supervision d'une Application à Base de Composants afin de Respecter des Contraintes Temporelles - https://hal.inria.fr/inria-00000787