# How to select best parameter for polynomial kernel?

I am using LibSVM library for classification. For my problem I am using polynomial kernel and I need to select best parameters (d = degree of polynomial kernel, and C = soft margin constant). The LibSVM guide suggests for grid search for this job. In this library there exists a parameter selection tool (grid.py), but it seems to me this is implemented to tune parameters of RBF kernel (gamma). My questions are:

1. Is there any other good solution for selecting suitable parameters rather than Grid search?
2. Can anybody give me some hints / sample code for selecting parameters for polynomial kernel?

Grid search is a sensible procedure as @JohnSmith suggests, however it is not the only stable technique. I generally use the Nelder-Mead simplex algortihm, which I have found to be very reliable and more efficient than grid search as less time is spent investigating areas of hyper-parameter space that give poor models. If you are a MATLAB user, you can get my implementation of this method here. Nelder Mead simplex methods are attractive as they don't require gradient information (I suppose you can think of the gradient of the simplex as being an approximation of the local gradient) and is very easily implemented.

Also gradient descent optimisation of the span bound is a good way of optimising the hyper-parameters, see Chapelle et al. (also investigate the other papers that Olivier has written, there are some real gems).

One advantage of grid search however is that it is quite easy to over-fit the cross-validation error (or span bound etc.) when optimising the hyper-parameters, especially if there is little data and many hyper-parameters. This means you can get a very poor classifier if you tune the hyper-parameters to convergence, so a coarse grid search can often perform better (I have a paper in preparation on this).

• does this method (Nelder–Mead method) really works good for SVM parameter search?I don't know the properties of the method, I am just a little sceptic about applying methods for local optimum search to such tasks. – Dmitry Laptev Jun 7 '12 at 11:04
• yes, I use it all the time. If you only have a couple of hyper-parameters (any more and grid search is impractical anyway) then the model selection criterion is likely to be fairly unimodal, so local minima are generally not a problem. – Dikran Marsupial Jun 7 '12 at 17:33
• @Dikran Marsupial - Thanks. I've also heard that Grid search is sensitive, thats why I was curious about other options. I will try to use your procedure. – Raihana Jun 9 '12 at 7:56

Why don't you want to use grid search? It seems to me the only stable technique for SVM parameter estimation. Of course you can use adaptive search, but it can save you not so much time, but lead you to some strange results.

And the implementation for CV procedure is pretty standart:

Divide your dataset into two parts: $X_{train}$ and $X_{test}$.
Run cycle through the parameter $d$ {
Run cycle through $C$ {
Train SVM with $X_{train}$, $d$ and $C$;
Test on $X_{test}$;
Remember the parameters, leading to the best result;
}
}

• thank you very much for the help. Actually I'm not against Grid search, I wast just wondering if there is other better solution. But your reply gives me a clear idea. – Raihana Jun 7 '12 at 10:01

You can try using a random search algorithm to get near equivalent results as grid search with a lot less work depending on the granularity of your grid. You can read more about it at Dato Blog along with other hyperparameter tuning techniques. This method is relatively quick if you set it up to run in parallel.