Classification and clustering are, without doubt, among the most frequently encountered data analysis tasks. This thesis provides a comprehensive synopsis of the main approaches to solve these tasks that are based on (point) prototypes, possibly enhanced by size and shape information. It studies how prototypes are defined, how they interact, how they can be initialized, and how their parameters can be optimized by three main update methods (gradient methods, alternating optimization, competitive learning), which are applied to four objective functions (sum of squared distances, sum of squared errors, likelihood, likelihood ratio). Besides organizing these methods into such a unified framework, the main contributions of this thesis are extensions of existing approaches that render them more flexible and more robust or accelerate the learning process. Among these are shape and size parameters for (fuzzified) learning vector quantization, shape and size regularization methods, and a transfer of neural network techniques to clustering algorithms. The practical relevance of these extensions is demonstrated by experimental results with an application to document structuring.