Although previous papers have focused primarily on the functional mapping between an approximate reasoning algorithm and the neural network approach, here we are concerned mainly with structural mapping between those two paradigms. Our objective is to deal with three different but correlated issues, rule-base acquisition, computational representation, and reasoning, under a unified framework of CPN-based neural networks. In particular, we introduce a simple and systematic scheme capable of self-organizing and self-learning the required control knowledge in a real-time manner for a multivariable process. The starting point of the approach is to map structurally a simplified fuzzy control algorithm (SFCA) developed by the authors into a counterpropagation network (CPN) in such a way that the control knowledge is explicitly represented in the form of connection weights of the nets. Then, by extending original training algorithms at both the Kohonen and Grossberg layers into highly self-organized and completely unsupervised versions, the control rule-base, initially empty, is gradually self-constructed to satisfy the prespecified performance requirements. Finally, the approximate reasoning is carried out by replacing a winner-take-all competitive scheme with a soft matching cooperative strategy. The approach is very generic in the sense that, in principle, an arbitrary dimensional control rule-base can be constructed automatically to satisfy arbitrary desired but physically achievable performance requirements. The learning process involved is extremely fast owing to the simple network topology and the efficient learning algorithms. A problem of multivariable control of blood pressure has been studied as a demonstration example. Extensive simulations have been performed, aimed at investigating adapting, learning, and reasoning capabilities of the system with respect to a variety of situations in the sense of defined performance measures.