Path Following Using A Learning Neural Network

This thesis is a development ofprevious works done by [2} on capability of
neural controller to efficiently track prescribed paths. Equipped with knowledge on
optimal preview control obtained from [1], the initial weights of linear and nonlinear
neural controller are initialized to the optimal gains. The implemented neural
controller will in turn minimize a performance index, which includes the lateral and
attitude angle errors ofvehicle models with respect to the paths.
The thesis differs from [2] in a sense that different types of neural controller
are established to achieve a better path following accuracy. Two algorithms, gradient
descent and quasi-Newton which utilize a batch training method, are introduced as
comparison to the gradient descent method that incorporates the online (or
incremental) training method. The class of learning (whether good or bad) of the
neural controllers is evaluated from the obtained percentage of average weight
change, maximum path and yaw attitude angle errors as well as the maximum steering
wheel angle. The behaviors oflearning rates and updated weights are given special
attention in this thesis. To conduct the specified works, the MATLAB programs
written by [2] have been extended and modified.