Hello NuPIC,

The wiki page on anomaly scores mentions that some cases call for the use
of a moving average on the raw anomaly score[1]. However, it does not
mention examples in which this is used. I would like to know more about
applications using a moving average of the anomaly score.

I'm experimenting with custom generated datasets. In the datasets that I
added randomized noise to, the predictions are impressive, but sometimes
seem 'too perfect'. For example, the prediction will be a little off, but
it will follow the spikes caused by noisiness impeccably.

By contrast, a dataset without noise and a seemingly clearcut pattern will
result in a gradual breakdown of the prediction, where it will start
switching values in consecutive steps, or spiking to an extreme prediction
in the middle of following the pattern accurately.

With regards,

Casper Rooker
[email protected]

[1]:
https://github.com/numenta/nupic/wiki/Anomaly-Detection-and-Anomaly-Scores#results

Reply via email to