Took me awhile to get back to this, but I have some news at least. :) I looked at your example code, but was a bit confused, so I modified an existing code sample I have to do predictions on your "5s and 6s" data set. See:
https://github.com/numenta/nupic.workshop/tree/fives-and-sixes/part-1-scalar-input And the resulting predictions match perfectly: https://plot.ly/~rhyolight/301/just-some-data/ In particular, see the model params I used: https://github.com/numenta/nupic.workshop/blob/fives-and-sixes/part-1-scalar-input/model_params/model_params_fives_sixes.json And also this bit identifying the RDSE "resolution" based on the min/max might be what was missing from the previous example I gave you: https://github.com/numenta/nupic.workshop/blob/fives-and-sixes/part-1-scalar-input/run_prediction.py#L36-L41 I hope that helps? --------- Matt Taylor OS Community Flag-Bearer Numenta On Thu, Apr 28, 2016 at 7:41 AM, Alexandre Vivmond <[email protected]> wrote: > I appreciate that you're going the extra mile here in helping me out. > I'll try to keep it short then, I've run 2 swarms, > -- The first setup -- > Swarm size: medium > Input data size: 20000 lines > "last_record": 3000 > "maxValue": 6.0 > "minValue": 5.0 > Once the swarm had run its course, I ran the OPF with the swarm's > generated model_params.py file. > The output file showed that HTM struggles to learn the pattern > 5,5,5,5,5,5,5,5,5,5,6,5,5,... predicting the 6 seemingly randomly. > > -- The second setup -- > Same as above, except I followed your previous advice about using > a RandomDistributedScalarEncoder instead of the regular ScalarEncoder. > Again the output file showed pretty much the same thing as for the > previous setup > > If you want to double check for yourself, I provided all the files in the > attachment that you would need to test it yourself. > All I want really, is to be sure that my setup is not wrong, and that > Nupic's results really show that the above mentioned pattern truly is hard > for HTM to learn. >
