One of the main applications of neural networks in control of nonlinear systems is in feedback linearization. In the latter, a neural network trained to approximate the nonlinear dynamics is used in the control law that forces the closed-loop system to behave linearly. The drawback of this approach is that the linearized systems are usually very sensitive to the error in the neural network approximation of the nonlinear dynamics. This paper presents a combination of an appropriate neural network training technique and a linear controller design procedure that minimizes the influence of the linearization error to the stability and performance of the resulting closed-loop system.