Hello NuPIC,

Here are some tips for swarming in no particular order:

1. Only swarm over a few thousand rows of data (3000 max). Tune this
with the "iterationCount" swarm setting.

2. Ensure you swarm over the same type of data you'll run the model
against. Using a sample of the exact same data is best.

3. To debug, start with a "small" swarm. It is really a DEBUG swarm.
It only runs one
model, basically testing that the swarm configuration being used is
valid. Do this by setting "swarmSize" to "small" in the swarm config.

4. Don't use large swarms unless you've gotten some promising results
from medium swarms. It is a mistake to jump right to
"swarmSize":"large" because it will take a REALLY long time for
potentially very small gain. If you are not getting any promising
results from a model created by a medium swarm, a large one probably
will not help.

5. To run faster, find out how many CPUs your computer has and use the
"maxWorkers" configuration to utilize them all. See
https://github.com/numenta/nupic/wiki/Running-Swarms for details.

6. Don't include too many columns of data. You might even just start
with the field you want to predict. Sometimes that's all HTM needs to
make decent predictions. Adding new columns of data will exponentially
increase your swarm time.

Regards,
---------
Matt Taylor
OS Community Flag-Bearer
Numenta

Reply via email to