Hi Arthur, I’m happy to help. What granularity is the raw data? How do you combine the data over 15 seconds into one value? If you're taking the mean over a lot of events it's possible you're washing out spikes in the data. Maybe taking the max over each 15 second period will give your model better info. We (specifically Yuwei) have been working on an algorithm to automatically find the best data aggregation. It’s still experimental code in our research repo [1], but should be straightforward for you to try out.
It looks like you're using anomaly likelihood correctly in NuPICModel.py, but the learning_amount part is odd; looks like it's never used, and redundant if-statements. In AnomalyTry.py I see you're using the log likelihood (good). Have you tried different values than 0.5 for the likelihood_cutoff? Also you can explicitly set the learning period by initializing the AnomalyLikelihood instance with claLearningPeriod [2]. [1] https://github.com/numenta/nupic.research/blob/master/projects/wavelet_dataAggregation/param_finder_runner.py [2] https://github.com/numenta/nupic/blob/master/src/nupic/algorithms/anomaly_likelihood.py#L84 Best, Alex -- Alexander Lavin Software Engineer Numenta
