Neuro-fuzzy systems combine the theory of two popular computational intelligence techniques: neural networks and fuzzy logic systems. In this thesis, we study Continuous Inference Network (CINET), a neuro-fuzzy classifier, developed at Applied Research Lab, Pennsylvania State University. Our work is to make some enhancements of CINET classifier. We prove that the problem of learning the range of input membership function of CINET classifier is a Linear Mixed Integer Programming (LMIP) problem. Moreover, the necessity and sufficiency functions for fuzzy inference are simplified to satisfy some algebraic properties and facilitate using back-propagation algorithm to adjust system parameters. To deal with the randomness in input measurements we also define the ambiguity degree and give out its calculation method.