※ 引述《hometoofar (家太遠了)》之銘言：
: I tried to do the primal with slack variables starting by using the 4 points
: to get 4 equations of w1, w2, b and the slack variable.
: I did some algebra to reduce the equations but I can't get anything that's
: really helpful. However I did notice something interesting..
: If I just add the 4 equations, it shows that sum of the 4 slack variable >= 4
: Then I just throw in w1=0, w2=0 and hope to find some b that works with the
: slack variable constraint. A value of b exists so that w1=0 and w2=0 and sum
: of slack variable = 4. I believe that it minimizes the equation.. but there
: are two things I am concering about...
: 1. This is not a good mathematical way... I can't just random pick some number
: and claim it to be the optimal.
: 2. What line does w=[0 0]T give? I know that this dataset is not linearly
: separatable unless I have a >1 slack variable for one of the o or x point,
: but I can at least draw such a line. The idea of w=[0 0]T doesn't make me
: feel right.
I think of solving the primal "without" slack variables.
If using slack variables, we can't match the dual.
(Of course, I first assume no advanced dual solution exists.)
Because the example is not separable, the primal is not solvable.
: Also I am stuck on dual too...
: I get the alpha equation, but I don't know how to find the values of alphas.
: I get the equation, I get the constraint that yTa = 0, but there are 4
: unknown alphas..
: Again, I try to make everything 0, simple, but I know it isn't good.
: (Like the page 23 example of lecture notes, set alpha1=0 meet the constraints,
: but it's not the right solution)
: I tried to find another set of alpha that aren't all 0 and minimize the
: equation, but that set gives me really big w...
For dual, I can compute alpha. I just say it needs a trick,too. = =+
Alpha is a zero column.
--
Certainly, my answer must not be correct ....
--
※ 發信站: 批踢踢實業坊(ptt.cc)
◆ From: 59.112.84.83