Development of a potential field estimator for a path-planning application using neural networks

Abstract

This thesis presents the development of a potential field estimator for a locally constrained autonomous path-planning application. The potential field estimator was developed using back-propagation neural networks, which have been shown to be useful for solving classification problems. It is shown that a back-propagation neural network may be used to classify a three-dimensional obstacle field on the basis of geometrical statistical data. The results of the classification may then be used to find a set of weighting factors that are used to reduce the computational effort needed to solve a path-planning problem. Using the method described here, a reduction of approximately seven percent of the worst-case total computational time for a path-planning problem is observed. The computational expense is still higher than desired for path-planning problems, providing opportunities for further research. Computational expense is measured as the total central processing unit (CPU) time spent solving a path-planning problem. The mathematical framework for the potential field estimator and the path-planning application it supports is presented. This thesis is the product of research performed for Texas A&M University and United Space Alliance.