Thus we want to use some methods to deal with the situation of overfitting to noise.

Reminding us of the approach of pocket, instinction called us.(We can not shatter points, but we could find someone better.)

However, this is not a QP problem anymore if we simply add them togather.

Thus we need a factor that define the degreee of error to make this issue solvable, or linear.

 

 

With the added parameter, we can represent the new function of soft-margin SVM. There for when making some mathematic pruning, 

We would find something familar:

 

To Simplify, is to be alive~~

Finaly, we got what we want:

Thus we can find suitible α and when it comes to b :

For any free SV we found, they could be used to compute the factor b.

Now we are ready to analyze the marks and distribution of points in free SV.

We can clearly see the role of ζn, along with the changing of α, which is the violation amout for ons specific point.

Applying the former stratigies, we should try the simplest model first. Cross-Vaildation is also useful.

non-SV is with no contribution to Eloocv, thus we get a classic bound, which indicates the relationship between #SV and Eloocv.

This is a useful tool to check the model, if there are too many SV in a model relatively speaking, we may delete this possiblity.