The best way to get good predictions is to swarm over data and use the best model returned by the swarm for that predicted field. That model is not the one that is used in river-runner. I am using a generic set of model params that work well for most scalar input fields. These params are not optimized for prediction, so the predictions are not good.
--------- Matt Taylor OS Community Flag-Bearer Numenta On Sat, Oct 17, 2015 at 8:05 AM, Marek Otahal <[email protected]> wrote: > Great timing Matt! > > I was just planing to run a bunch of NAB & River data through HTM for this > weekend :) I guess https://github.com/nupic-community/river-runner is > what I'm looking for, right? > > PS: I don't understand the note about predictions, afaik anomaly detection > is based on predictions, so either the models are not optimized at all, or > are optimized for predictions (and anomalies as well). > PPS: This leads me to "Do you have optimized params for specific streams > of River?" If not, are you interested in some? Where to publish..? > > Cheers, > > On Sat, Oct 17, 2015 at 4:37 PM, Matthew Taylor <[email protected]> wrote: > >> Hello NuPIC, >> >> Getting data out of River View is easier than ever with >> https://github.com/nupic-community/riverpy. I have also updated >> https://github.com/nupic-community/river-runner to use riverpy as a data >> client, providing another example. >> >> This new client is useful because I have recently made changes in River >> View that disallow the retrieval of very large amounts of data in one >> request. Riverpy provides a data cursor with a way to navigate through >> large streams of data easily. >> >> Regards, >> --------- >> Matt Taylor >> OS Community Flag-Bearer >> Numenta >> > > > > -- > Marek Otahal :o) >
