Hi Alex,

My team and I are close to finishing the semester's work, we have one last
task to finish before our project is demo-able. We are currently running
two weeks of sample data (time-stamped apache web logs) through a python
file which manipulates the data until it is in usable form and feeds it
into your default NuPIC model. When the code reaches the first troublesome
period within the two weeks (we have been told to expect anomalies since
the company experienced a service outage) the NuPIC model correctly
identifies the response times as anomalous.

While that is a great relief (to know the detection is working properly) we
encounter our problem after the data has been parsed and the next
troublesome section of timestamped data is reached. Roughly 48 hours after
the first outage, hundreds of data points later, we expect to detect our
second anomalous period but the NuPIC model does not identify anything as
out of the ordinary. The response times spike largely (from an average of
200 to an average of 5000 for almost three hours) but do not trigger the
model's detection.

We are working on testing the dataset now, by scaling it down and giving
the model a truncated data set where the first outage does not exist, to
see if the model can correctly identify the second outage. Our questions
might be difficult to answer, I can provide supplementary code base and the
data set we are using.

Do any possible issues come to mind? We do not turn off the "learning"
aspect of the model and are wondering if the first anomaly (although it was
hundreds of data points prior to the second outage) caused NuPIC's pattern
recognition to justify the second outage. Should we look to save the NuPIC
model externally (while it has seen nothing but normal timestamped points)
and each time an anomaly is discovered we revert to the original NuPIC
model, in hopes of keeping the model "clean"?

Thank you for any help,

Arthur Harris

On Sat, Apr 16, 2016 at 7:53 PM, Alex Lavin <[email protected]> wrote:

> Hi Arthur,
> Looks like your email slipped through the cracks, sorry. What’s the
> current status of your project?
>
> Best,
> Alex
>
> --
> Alexander Lavin
> Software Engineer
> Numenta
>

Reply via email to