| For several classes of nonlinear systems, neural network (NN) controller structures and weight update rules are studied for closed-loop control applications. Design techniques guaranteeing stability and tracking are derived using the Lyapunov theorems and extensions. Throughout this work, we have focused on designing a controller working on-line under less knowledge, reasonable conditions, and mild assumptions.; The relation between adaptive control and NN control is studied. It is argued that the NN controller performs very well when compared with adaptive control, which needs to know a complete regression matrix of the system. However, in the presence of any unmodeled dynamics the NN controller outperforms adaptive control. For a serial-link rigid arm, functional-link and multilayer NN controllers with guaranteed tracking are developed. Standard robot control notions, such as filtered tracking error and robot properties, together with passivity are used to derive the structure of the NN. Novel tuning rules which guarantee boundedness of closed-loop signals without a persistency of excitation condition are derived using Lyapunov theory. These rules suggest necessary changes to the backpropagation algorithm to guarantee stability of the overall system. Then, these results are extended easily to other robotics systems, such as flexible-link manipulators and force/motion control problems.; This design technique is next applied to a general class of nonlinear systems. Using the same multilayer NN structure with an additional robustifying control input we have shown boundedness of all signals. For a larger class of systems, a feedback linearization technique is chosen for control; two three-layer NN controller are implemented to estimate the feedback linearizing control without using knowledge of the nonlinear system dynamics. As the NN learns its actual weights on-line, the proposed controller based on the NN weight estimates avoids zero division and remains bounded.; Finally, simplified tuning algorithms which yield similar stability results with much less computation are developed. Tuning algorithms are modified-Hebbian rules which perform on-line tracking for the classes of systems studied above. A great deal of computational advantage is gained without sacrificing stability results.; In all of the work, initialization of the NN weights is straightforward (we simply set them to zero) and no stringent condition like persistency of excitation or off-line training is required. Thus, these NN controllers are applicable to a wide range of realtime applications. |
