C parameter does not affect accuracy (LibSVM) I'm working on LibSVM and I'm using linear kernel. But i have a problem with C parameter, it does not affect accuracy during test set! Is it possible?
I send a picture with the distributions of the two classes and the separator line. This line did not change when I changed C. I have no idea why! :-( 
 A: The decision boundary seems to be in a good place. What kind of changes do you expect to see by modifying $C$? It seems to me that this hyperplane already has close to the lowest possible misclassification error, e.g. increasing $C$ would no longer have an effect.
You can probably see an effect if you make $C$ much smaller, to minimize the support values instead of misclassification error. This is probably not what you want though (lower accuracy).
Using robust eyeballing™ on your plot, I suspect the following:


*

*Gaussian 0 has mean $[0,0]$,

*Gaussian 1 has mean $[1.5, 1.5]$,

*The slope of the decision boundary is about $-1$.


If my eyeballing skills did not fail me, this decision boundary is optimal: perpendicular to the line segment connecting the means of Gaussian 0 and 1 (which has slope 1) and it seems to intersect it around $[0.75, 0.75]$. If for a given value of $C$ you find the optimal decision boundary, increasing $C$ will not affect the boundary anymore.
A: The decision boundary seems already optimal to me, and I assume you would like to observe the effect of parameter changes on the decision boundary. If that is the case, first make sure you scale the data in advance. Then I would suggest you implement the grid search on C values on a wide range ($10^{-5},...,10^5)$), and observe the training accuracy. If C increases slightly, it is possible that you still form all of the linear models the same as before, since the upper bound on the allowable norm of the weights along with the margin may not influence on their choice very much within a small range of C.
