Re: [问题] max violating pair

楼主: Evenroll (怎样才能充满活力?)   2005-11-09 09:48:51
※ 引述《Evenroll (研二不是应该很清闲吗?)》之铭言:
: I have a question about max violating pair.
: since
: I_up = {t| alpha_t < C, y_t = 1 or alpha_t > 0, y_t = -1 }
: and
: I_low = {t| alpha_t < C, y_t = -1 or alpha_t > 0, y_t = 1 }
: then
: at begining, alpha is set to zero as a initial value
: then all i don't belong to I_up or I_low, since no alpha_i > 0.
: should I just randomly pick up 2 index i & j ?
: if I just randomly choose i & j at the first iteration,
: I still encounter this problem at the second iteration,
: since only alph_i and alpha_j are changed,
: all other i are still zero and they don't belong to I_up or I_low...
: so...
: I think that in the definition of I_up and I_low,
: maybe we should rewrite the inequality to be alpha_t >= 0 or alpha_t <= C
: is that right?
also,
I found that if there is an alpha_i that 0 < alpha_i < C
then the index i will belong to I_up and I_low at the same time.
I also find that in the example in ch 6.3,
even if alpha = [1,1,0,0]' at the first iteration
the example still could find I_up = {2, 4}, I_low = {1, 3}
I couldn't understand how he find I_up and I_low.
Is there anything I misunderstand in the algorithm?

Links booklink

Contact Us: admin [ a t ] ucptt.com