In order to build an efficient nearest neighbor classifier three objectives have to be reached: achieve a high accuracy rate, minimize the set of prototypes to make the classifier tractable even with large data bases and finally, reduce the set of features used to describe the prototypes. Irrelevant or redundant features are likely to contribute more noise than useful information. These objectives are not independent. This chapter investigates a method based on a hybrid genetic algorithm combined with a local optimization procedure. Some concepts are introduced to promote both diversity and elitism in the genetic population. The prototype selection aims to remove noisy and superfluous prototypes and selects among the others only the most critical ones. Moreover, the better the selection the faster the algorithm. The interest of the method is demonstrated with synthetic and real chemometric data, involving a large number of features. The performances are compared to those obtained with well known algorithms