In this paper Rough Set Theory (RS) is employed to discuss “rule+exception” modeling, which will have fewer rules compared with rule-based modeling and fewer exceptions compared with example-based modeling. An attribute reduction strategy based on discernibility matrix is described. We attempt to consider what kind of data sets are suitable for the model, and how to distinguish exceptions within the data sets. To illustrate the principle the psychological model of Nosofsky’s category learning is simulated, and three more complex examples are provided.