I mean tuning the rbf_sigma, boxconstrain and other parameters to get best answer.I have 116 feature vectors and I use RBF kernel and I need to optimize rbf_sigma and boxconstraint for better results. I'm using the following Matlab code:
svmtrain(xtrain,ytrain,'Kernel_Function','rbf', 'rbf_sigma',3.03,'boxconstraint',10, 'kernelcachelimit', 10000);
Do you suggest any other ways to optimize this classifier for my data? any other parameters?
Currently I'm trying "fminsearch" and "patternsearch" and "GlobalSearch" in Matlab for optimizing just rbf_sigma and boxconstrain. Which one might work better? They don't work well and I'm not sure about their performance. I think they stop in local minimum.
Thanks for your time
I descretize my svm parameters and perform grid search. I choose a cmination that give low error on my testing set as well as relatively few support vectors to keep parsimony.
In general you can't find the optimum parameters, you just search for a good enough combination, that is my suggestion.
Thanks Dimitris, so there is no optimum parameters for classifiers,
but it's important for me to improve recognition rates,
Since you have 116 features involved, I might suggest you to do a PCA (Principal Component Analysis). Although this does not guarantee better classification, it could provide you a better insight to the data set + also lead to better classification. I have done this on one of my research projects and was very helpful.
MATLAB has a function that optimizes theses parameters [link below].
https://www.mathworks.com/help/stats/optimize-an-svm-classifier-fit-using-bayesian-optimization.html
Dear sina,
A very applicable suggestion for your question is using genetic algorithm for svm parameter optimization.
The classificarion rate could be considered as GA objective function, while the kernel free parameters formed its chromosoms.
Regards
- Badges
- Science topic