Sergey, there are a couple things I can think that might be the issue: 1. Are you looking at the predicted values or the probabilities associated with the predictions? 2. If your encoding has buckets too large then it may be predicted a bucket that includes 0 and some positive numbers. When this happens, it uses a exponential smoothing to pick the value to use for the predicted bucket. So you may change the encoder to make sure that 0, 1, 2, etc fall into different buckets.
On Mon, Jan 5, 2015 at 11:18 AM, Matthew Taylor <[email protected]> wrote: > Sergey, > > It is hard to help without knowing some more about your setup. For > example how is he feeding in the data? What are the scalar encoder > parameters? Has he swarmed on the data? How large is the training > data? > > Thanks, > --------- > Matt Taylor > OS Community Flag-Bearer > Numenta > > > On Thu, Jan 1, 2015 at 4:00 PM, Sergey Alexashenko > <[email protected]> wrote: > > Hey all, > > > > Running into a problem with TemporalMultiStep prediction, wondering if > > anyone has encountered anything similar. > > > > I am feeding NuPIC a stream of data that includes a numeric variable I am > > trying to predict. This variable follows a pretty obvious pattern, it is > 0 > > for 4 periods at a time and on the 5th period has positive numeric value. > > Here is a sample of the variable changing over time: > > 0,0,0,0,2,0,0,0,0,3,0,0,0,0,1... etc > > > > In CL params, I have NuPIC generate predictions for steps 1-5. > > > > However, instead of the prediction results I am expecting, e.g.: > > > > 0,0,0,0,4 or 0,3,0,0,0, depending where I am in the cycle, NuPIC makes > > predictions that look like this: > > > > 0.3,0.3,0.3,0.3,0.3 > > > > I was wondering if anyone has any idea what's going on and what I can do > to > > fix this issue? > > > > Appreciate the help, > > > > Sergey > > > > > >
