How to decide which type of kernel to use in SVM




While trying to implement SVM in R I noticed that with the linear kernel the accuracy was about 84% and with the rbf kernel it went upto 93% but when I used the sigmoid kernel the accuracy fell below 10%.
So how do we decide which type of kernel to use?
Is is based on some patterns we see in the data or something else.
Can someone please help me with this??


@pagal_guy- It completely depends upon the data in which we are applying the SVM classifier, for example, this data will be best classified by a non-linear classifier.

Hope this helps!