|
I've just implemented a version of the support vector regression algorithm. This is the primal optimization problem
Since it is the first time I really use this algorthm I'm wondering how to estimate the epsilon parameter. The epsilon parameter refers to the epsilon insensitive loss function, which is implicitly contained in the problem above i.e.:
So, is there some rule how to set this parameter? |
|
I think the usual "solution" is to do cross validation. There might be good heuristics, though. If your output range is normalized to [0, 1], I would think something < 0.1 should be reasonable. |
