Hello! I finally started testing my implementation on real data. I took a classic dataset, which reflects the amount of energy consumed by houses in a city.
On the graphs below, you can see the training stage at the very beginning, the identification of anomalies, but are there too many false “anomalies”?
Encoder parametrs:
Day of week: 100 bits, 21 active, periodic;
Time of day: 100 bits, 21 active, periodic;
Energy: 200 bits, 31 active, range is 0…70, not periodic.
SP - im use my own SP, his always provide 2% of active columns (2% is 50, total columns 2500). Maybe problem in my SP, now I can’t say for sure =(
TM - slightly different from vanila HTM TM, but I don’t think it matters. LEARNING_TRESHOLD is 5 and ACTIVATION_THRESHOLD is 10.
These values seem low to me, which could be causing too many predictions. It totally depends tho on columnCount and newSynapseCount. Would you share the whole parameters file?
Here’s the default used in the NuPIC hotgym example for comparison: