In order to reduce computation and storage costs of learning methods, we present a prototype selection algorithm. This approach uses information contained in a connected neighborhood graph. It determines the humber of homogenous subsets in the R space, and uses it to fix the number of prototypes in advance. Once this number is determined, we identify prototypes applying a stratified Monte Carlo sampling algorithm. We present an application of our algorithm on a simulated example, comparing results with those obtained with other methods.