| The research in this thesis contains two main topics: (a) design and analysis of recurrent neural networks for synthesizing feedback control systems via pole assignment and their control applications, (b) design and analysis of dual neural networks for online solving constrained quadratic optimization problems and their robotic applications.; A multiplayer recurrent neural networks called pole-assignment neural networks is applied to online synthesizing and tuning feedback control systems. Analytical results including global stability, exponential convergence and parameter-selecting rules are presented for the pole-assignment neural networks.; Furthermore, a novel recurrent neural network model is established by absorbing the first-order time-derivative information to solve the Sylvester equation with time-varying coefficient matrices.; A recurrent neural network called the dual neural network is developed and analyzed to handle quadratic programs subject to linear equality, inequality and bound constraints.; For readability and as preliminaries, the Lagrangian neural network and the primal-dual neural network have been first reviewed for minimum two-norm and minimum infinity-norm kinematic control of redundant manipulators, respectively. Then, by reformulating bi-criteria kinematic control as a constrained quadratic program, the single-layer dual neural network is applied to the real-time path-following inverse kinematic control of redundant manipulators. This application is able to remedy the discontinuity defect of pure minimum-effort solutions.; Another important issue in the motion planning and kinematic control of redundant manipulators is the obstacle avoidance. In the thesis, an improved problem formulation is proposed in the sense that the collision-avoidance requirement is represented by dynamically-updated inequality constraints, and that physical constraints such as joint physical limits are also incorporated directly into the formulation. The dual neural network is thus developed for the online solution to kinematic control of redundant manipulators in the presence of point obstacles and window-shaped obstacle.; As a topic of dynamic control, the dual neural network is exploited for the real-time joint torque optimization of kinematically redundant manipulators. (Abstract shortened by UMI.)... |
