Abstract: In using the epsilon-support vector regression (epsilon-SVR) algorithm, one has to decide on a suitable value of the insensitivity parameter epsilon. Smola et al determined its ``optimal'' choice based on maximizing the statistical efficiency of a location parameter estimator. While they successfully predicted a linear scaling between the optimal epsilon and the noise in the data, the value of the theoretically optimal epsilon does not have a close match with its experimentally observed counterpart. In this paper, we attempt to better explain the experimental results there, by analyzing a toy problem with a closer setting to the epsilon-SVR. Our resultant predicted choice of epsilon is much closer to the experimentally observed value, while still demonstrating a linear trend with the data noise.
Proceedings of the International Conference on Artificial Neural Networks (ICANN), Vienna, Austria, August 2001.
Postscript: http://www.cs.ust.hk/~jamesk/papers/icann01.ps.gz