Quote:
Originally Posted by dobrokot
If I take output of linear regression, I usually have a little (15 of 100) points which are misclassified. But it does not help PLA, because first addition Xi to W destabilize "almost good" value of W. I can even invert sign of L.R. output, it doesn't affect number of iterations (despite number of misclassified items on first iteration changes from 15 to 9599, I verified)
But if I use (L.R. output)*N, it does help for PLA, greatly reduces number of iterations.
Is it supposed, that usage of linear regression output as W should greatly change number of iterations? If yes, does it means I should to search for other bugs in my code?
Output of L.R. should be used as is, or should be multiplied to make it "stronger" ?

I think for this question you should set N=10 and just use output of linear regression. It should reduce the number if iterations of PLA. I tried it for N = 100 and agree with you that linear regression doesn't help improve iterations.