Svmlight warning: Relaxing epsilon on KT terms
I am trying to explore a dataset using Svmlight. I always see the following warning during the optimization process.
WARNING: Relaxing epsilon on KT-Conditions (0.010000).
.....................................................
WARNING: Relaxing epsilon on KT-Conditions (0.100000).
.....................................................
WARNING: Relaxing epsilon on KT-Conditions (1.000000).
.....................................................
WARNING: Relaxing epsilon on KT-Conditions (10.00000).
....................................................
Is it important and what does it mean? The training has not been successful at this time. Could it be related to this warning? Thank.
source to share
KT conditions are part of the optimization problem at the heart of SVM. Without going into too much theory, there are a number of conditions that, when you have an optimal solution, will be true. However, going through enough iterations for the KT conditions to be absolutely true, it takes more time than just making them true within espilon (a very small number) to be true. Essentially, instead of saying that A should equal B, we are saying that A should equal B inside + or - epsilon. Research has shown that this works as well as taking extra time to get conditions exactly right.
SVM Light first tries with a very small epsilon, when it cannot solve this problem, it relaxes epsilon, making it larger to try to find it, but clearly, if the problem is not solvable, it tries several times before deciding that it is not solvable.
So I would say the warnings are because your problem is not solvable, not the other way around.
source to share