We target the problem of predicting resource usage in situations where the modeling data is scarce, non-stationary, or expensive to obtain. This scenario occurs frequently in computing systems and networks, mostly due to the high dynamicity of the underlying processes. Utility computing environments are an important example for such a scenario, as their frequent reconfiguration reduces the amount of training data available for modeling. We propose an approach based on a genetic algorithm and fuzzy logic which allows for creation of robust prediction models even with scarce training data. The method is evaluated on demand usage traces collected from 41 servers in a business data center. The results show in the setting of scarce training data amount our method has a significantly higher prediction accuracy compared to other non-linear techniques such as decision trees or support vector machines.