Learning algorithms have been developed to construct fuzzy rule-based models from training data. The quality of the resulting model is affected by the decomposition of the input and output domains and by the number and precision of the training examples. This paper investigates the robustness of fuzzy models produced from training data. The objective is to analyze the effects of increasing complexity on the off-line performance of the learning algorithm and the on-line performance of the model, where the complexity is measured by the number of variables describing the problem domain and the number of rules in the model. A hierarchical model is proposed to reduce the complexity in high dimensional systems.