Feedback-Error Learning Scheme Using Recurrent Neural Networks for Nonlinear Dynamic Systems

Because of their parallelism, functional approXimation and learning capabilities, artificial neural networks can be effectively employed to apjmximate nonlinear functions, and to synthesize controllers for nonlinear dynamic systems. The use of dynamic neural networks to model and control dynamic systems is of great importance in the control paradigm. The intent of this paper is to use one such dynamic neural structure namely the recurrent neural network to drive unknown nonlinear systems to follow the desired trajectories. The leaming scheme employed for this task consists of a conventional proportional-plus-derivative (PD) controller in the feedback loop and the recurrent neural network in the feedforward path. Once the convergence is achieved, the recurrent neural network approximates the inversedynamics model of the plant under control. The PD controller, on the other hand, guarantees the stability of the learning scheme. The effectiveness of this leaming scheme is demonstrated through computer simulations and an experimental set up that demonstrates the balancing of a two-wheeled robot.