The ab initio prediction of the structure of a polypeptide from its sequence necessarily requires the detection of the lowest energy forms which correspond to the native state of the polypeptide. A potential for modeling the energy hypersurface of polypeptides using a hybrid level of description is optimized for the structures of four training peptides which have been shown experimentally to adopt alpha, beta beta, alpha beta, and beta beta alpha conformations in aqueous solution. This potential is then used in diffusion process-controlled Monte Carlo simulations to predict the native structures of this training set of peptides and a test set of 20 peptides which were not themselves used during the optimization of the potential. Starting from various fully extended conformations, all simulations lead to an ensemble of conformations compatible with experimental results. These conformations include simple motifs such as coil, alpha helix, beta-turn, beta-hairpin, beta alpha, and coil-alpha conformations, but also more complex motifs such as turnlike, beta beta alpha, beta beta beta, and alpha-helical hairpin conformations.