Investigate why swarms pick up constant fields

[rhyolight]

Investigate constant field contribution in swarms. Why are they selected, and what can be done for them not to be selected.

Check whether 0.2% improvement is sufficient to include a field.

This issue was plucked from the backlog of Grok issues after the swarming process was open-sourced. It included some propriety data sets from previous customers that could not be included publicly. To reproduce this but, the assignee will likely need to test different data sets by swarming over them and find one where swarming selects a constant field.

[rhyolight]

@subutai This seemed like a pretty big deal before we OSed NuPIC. Do you agree that it is still important?

[breznak]

Can you explain what is meant by a 'constant field'?
Eg. class SP(var1=1, const2=2)
or
x,sinx,const
0,0,1
0.1,0.05,1
?

Also

Check whether 0.2% improvement is sufficient to include a field.
could this be part of swarming too? eg.
considerableSwarmingImprovement=swarmFloat(0.1, 0.4) ?

On Wed, Aug 20, 2014 at 11:16 PM, Matthew Taylor notifications@github.com
wrote:

@subutai https://github.com/subutai This seemed like a pretty big deal
before we OSed NuPIC. Do you agree that it is still important?

—
Reply to this email directly or view it on GitHub
#323 (comment).

Marek Otahal :o)

[rhyolight]

A constant field meaning the value for that column of input data does not change over time.

[subutai]

@rhyolight It's a corner case. Because there is some randomness, sometimes a constant field ends up adding positive contributions. This is particularly true if we don't have much data or if the algorithm is not doing that well anyway.

@breznak I think that is actually a parameter to swarming. Swarming wouldn't include a field unless it improved the error by at least that amount. There is some way to specify this - not sure how.