Many researchers do not appreciate the problems in building high-dimensional fuzzy models or control surfaces, yet this task has occupied researchers in several fields for the past thirty years. The problems occur due to the lack of both available training data and the required computational resources necessary for building and calculating the response of the model. This paper outlines several techniques for partially overcoming the curse of dimensionality associated with high-dimensional data modelling problems and compares and contrasts them with several algorithms developed in the statistical community. The work is intended to outline both conventional concepts which can be usefully applied in neurofuzzy models and new developments in this field. >