Hi, Dino, Yes, ML can help deal with unpredictable link issues *if* there are some underlying statistics at work. However, it’s generally more useful to track such links as faulty and replace them than to use AI to “adapt” to their failure patterns.
I looked at ML techniques for predicting connection parameters (e.g., AI meets TCB-sharing) years ago, but it turned out to not find anything we didn’t know going in (diurnal patterns, association based on routing prefix, etc.). Joe — Dr. Joe Touch, temporal epistemologist www.strayalpha.com > On Feb 25, 2022, at 3:31 PM, Dino Farinacci <farina...@gmail.com> wrote: > > >> >> >>> On Feb 25, 2022, at 3:07 PM, Dino Farinacci <farina...@gmail.com> wrote: >>> >>> >>>> >>>> We use all three in the Internet (longest prefix, ARP/LISP, and RIP/OSPF, >>>> respectively). >>> >>> But we haven't used ML. Wonder what people think about that? >> >> Machine learning? > > Yes. > >> As in the failed DARPA Intelligent Nets effort? > > I don't know. > >> If so, ARP is learning - just the simplest kind. Without name structure >> that maps to network topology, there’s nothing else to learn. > > Yes, probably better than constand retraining ML models. > >> Or did you mean something else? >> >> Joe > > Use probabability to know which next-hops to use in a FIB entry. Start with a > static topology baseline and then prune paths based on probablility of > failure (typical links going up and down, nodes not reliable, etc). > > This is very high-level. And this is NOT SDN. All predications are done local > to a router. Just like my phone and car do it for other types of > applications/use-cases. > > Dino > >
_______________________________________________ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
