Hi Ricardo, Looks like others have answered your swarming question.
In your example, one big difference is in resolution. In 2) you need a smaller change in the numbers to cause a difference in the encoding. The number 1.5 would be encoded differently from 1 and 2 whereas with n=10 and w=3, at least two of the encodings would be identical. The best resolution really depends on the noisiness of your dataset, The other difference is in the size of w. You need w to be large enough so that SP columns will form a good solid representation of the number. If it is too small, then a small amount of noise can easily change the representation drastically. Our rule of thumb is to simply keep w fixed at 21, and just vary n to determine the resolution. (This is one of the things swarming does.) w=21 is large enough that small changes won't cause large changes to the SP output. --Subutai On Tue, Jan 20, 2015 at 4:00 AM, Ricardo Franco <[email protected]> wrote: > ok. I stopped using swarming because it takes an enternity to run in my > computer. But I'll give it a try again. I thought also in using some > statistical way to shorten the min/max range. > > My other question is about **n** and **w** values: > > If I understood correctly, the encoded value should have overlapping bits > to represent semantic similar values. Lets consider two different n/w > settings: > > 1) n = 10; w = 3. > > in such a way that it represents number 1 as: [1 1 1 0 0 0 0 0 0 0] > And number 2 as: [0 1 1 1 0 0 0 0 0 0] > > They (1 and 2) have 2 overlapping bits. > > 2) n = 30; w = 5 > > in such a way that it represents number 1 as: [1 1 1 1 1 0 0 0 0 ....] > And number 2 as: [0 0 0 1 1 1 1 1 0 0 0 0 0 .....] > > Again the numbers 1 and 2 have 2 overlapping bits. > > Comparing the two settings, should they give similar results? > > 2015-01-19 20:57 GMT-03:00 Subutai Ahmad <[email protected]>: > > >> You should set it to 15/50. You can fine tune the resolution of the >> encoder through swarming for best prediction accuracy. >> >> --Subutai >> >> On Sun, Jan 18, 2015 at 8:22 AM, Ricardo Franco <[email protected]> >> wrote: >> >>> Hello. Let me start with minValue and maxValue: >>> >>> Lets say I'm dealing with my city's temperature. >>> I know the range is about 15º ~ 50º. The temperature never goes below 15 >>> or above 50. I know this because I live here. >>> >>> Then I'll train my model using the last 4 months. In this period the >>> temperature ranged at 21º ~ 38º >>> >>> I'll make predictions for the next days/weeks. >>> >>> Which values should I set for minValue and maxValue? >>> >>> minValue = 15 >>> maxValue = 50 >>> >>> or >>> >>> minValue = 21 >>> maxValue = 38 >>> >>> -- >>> >>> Ricardo Franco Andrade >>> >>> *Web Developer* >>> >>> email: [email protected] >>> skype: ricardo.krieg >>> phone: +55 (86) 9569 8521 >>> linkedin: http://br.linkedin.com/in/ricardokrieg/ >>> github: https://github.com/ricardokrieg >>> >>> >>> >> > > > -- > > Ricardo Franco Andrade > > *Web Developer* > > email: [email protected] > skype: ricardo.krieg > phone: +55 (86) 9569 8521 > linkedin: http://br.linkedin.com/in/ricardokrieg/ > github: https://github.com/ricardokrieg > > >
