Hi Alexander, The commercial product Grok uses a "pre-swarmed" model. The main business reason for doing this was to give users results very quickly on cheap servers.
>From an algorithm standpoint this is not too bad because the inputs to Grok are very restricted: timestamp + single numerical metric value. We've found that for Grok-like data we can usually get by with a single set of parameters. Basically we swarmed on a bunch of different datasets and tried to find one parameter set that worked well across most of the datasets. This is the essentially the same set of parameters that are in hotgym_anomaly today. For single numerical value + timestamp I think it is a very good starting point. For multiple fields or extremely different looking datasets or for multi-step prediction I would still recommend swarming first to get the best performance. --Subutai On Mon, Apr 21, 2014 at 9:46 AM, Alexander van Dijk <[email protected]>wrote: > Hi Subutai, > > From discussions with the team here, I have been wondering for a while > whether the commercial Grok implementation uses swarming. Ritchie has been > using the swarming code to try to downselect parameters from our aviation > safety dataset using swarming (not always successful from what i > understand), but from what I've read and seen sofar about Grok, it appears > it is doing a 1-parameter-to-time model building, and builds individual > models for each of the parameters coming in from the AWS instances. > > I was wondering whether my understanding is correct, and if so, what the > reasoning is behind going down this route, e.g. is it to go down to the > lowest level of model building first to learn the simplest functioning, and > build up from there lateron using swarming or some other method? > > Thanks, > Alexander > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org >
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
