Based on few cultivations of the microorganism Paenibacillus polymyxa the yield of the antibiotic macrolactin is maximized via a nonlinear optimization of the feeding strategy. To this end, an approach to automatically build up process models is used to generate many small-size structured models that satisfactorily describe the identification experiments. In the early stages of the model building process, the right model can only be selected with a high uncertainty. Therefore, a robust multi-model planning is applied that addresses this uncertainty and prevents large deviations between predicted and measured results. After presenting the applied methods, four identification experiments and one optimized cultivation are used to demonstrate the concept.