Artificial neural control for nonlinear systems

Research was conducted to test a new automatic design methodology for nonlinear controllers that can be used when an accurate conventional computer model exists for the plant (the nonlinear system which is to be controlled). The real-time recurrent learning algorithm was employed to train an artificial neural network to perform an unknown nonlinear control function which obtains the desired behavior from the plant. Two different methods are employed to compute partial derivatives for the real-time recurrent learning algorithm. Derivative arithmetic was used to compute the partial derivatives of the state variables at the next time step with respect to the state and control variables at the current time step. The backward error propagation algorithm was used to compute the partial derivatives of the control variables with respect to the current state variables and the control system parameters which are simply the network biases and connection weights.; The methodology was applied in the design of a flight test maneuver autopilot for high performance fighter aircraft. An artificial neural controller was installed in a realistic computer flight simulator and trained to fly coordinated 2 g turns while maintaining altitude and airspeed. It should be possible to teach an artificial neural controller to fly almost any standard flight test maneuver but training would probably require considerable engineering insight and knowledge of the system to configure the network and learning algorithm. Of course, this compromises the ultimate goal which is to completely automate the controller design process.