This work presents a new multivariable control strategy using neural networks. The proposed control strategy uses past and present process information to design the best controller, as well as to generate the new control actions. At each sampling time the controller is optimized, using the future error of the closed loop, generated by a neural model of the process. The proposed control algorithm was tested in the control of a fixed bed catalytic reactor, which has a complex dynamic behavior. Such system presents inverse response and it is a distributed parameter system, so that its control is not a trivial task. The results have shown the potential of the controller to deal with the non-linearity of the process for the several tested disturbances.