SVM for regression

I like the explanation of SVM regression in the nu-svm paper. The idea is to use a margin-based loss for regression. The most popular one is the epsilon-insensitive loss function: max(0, |x - f(x)| - epsilon). There is an implementation of it in libsvm.

As a complement to Alexandre's answer and your reviewer's comment: for non-linear regression you should probably also include a comparison with Gaussian Processes and Random Forests that are also state-of-the-art AFAIK although completely different from one another and from SVR and NN.

I have also got a review from one journal--

"why is SVM used, instead of other regression methods? This question should be answered from both theoretical and practical perspectives. There are a number of other more advanced regression approaches that were explored in the CAD community (e.g., linear regression with regularization, sparse regression). These methods are not compared against the proposed SVM method."

Please tell me how can I compare these methods? Is there any document which compare these methods???