I asume you need jernel extention and hyperplane to be adequite to discriminate two sets of data. First try some tools that do kernel exension automaticly so you dont have to aproximate youreself. Second if you have to do it by hand, use disrtibution density algorithms( chose one adequite for you, bayesian is popular) to define where are most members located and by that decide which gaussian function youll use (where will local and global extremes be located)