A three-stage procedure for design of a neural-net controller for nonlinear plants is developed. The design is based on linearisation of the plant at each operating point. First, the input-output data of the plant is used to train a neural-net model of the plant. Next, the plant is controlled using a time-varying linear controller based on the linearised plant model at each operating point. The neural-net model of the plant obtained in the first stage is used to obtain the linearised models. In the third stage, a special structure neural net is trained to replace the time-varying linear controller. Results of simulation study are also presented.