Design of fixed (non-adaptive) neural controller for nonlinear plants described by input-output representation has been presented. The control scheme is based on an error plus output feedback. Integrators are used in the control loop to ensure set point following. A training scheme for the neural controller is derived that directly minimizes a versatile loss function. The training is conducted offline for a class of set points conforming to the normal operating condition of the plant. Results of simulation studies are presented.