Neural network training for a non-linear dynamical system

The dynamics of a robot are highly nonlinear: for precise positioning of a manipulator, control requires the solution of the inverse dynamics problem for a complex nonlinear system in real time. A neural network offers one of the most promising solutions to these problems. Being a nonlinear dynamical system, a neural network can be trained to learn different relations between variables regardless of their analytical dependency and can provide an approximate solution to the inverse dynamics problem for a robot manipulator. In the case of a robot manipulator, the training patterns are constantly changing. In such cases the training process will never converge, but will approach a temporary optimum. The authors discuss the problem of a neural network adapting to the constantly changing environment