Navigation with Artificial Neural Networks
Abstract
The objective of this dissertation is to explore the applications for Artificial Neural Networks (ANNs) in the field of Navigation. The state of the art for ANNs has improved significantly so now they can rival and even surpass humans in problems once thought impossible. We present different methods to augment, combine, or replace existing Navigation filters with ANN. The main focus of these methods is to use as much existing knowledge as possible then use ANNs to extend the current knowledge base. Next, improvements are made for a class of ANNs which provide covariance called MDNs. MDNs are necessary since covariance is required for navigation problems. Finally the improvements and framework are demonstrated in a VLF signals navigation problem. Without ANNs, our VLF signals navigation problem would be very difficult. We conduct two VLF navigation experiments with an indoor and outdoor environment. The ANNs used for these problems provide confidence with probabilistic estimates of position either through class probabilities or probability distributions parameterized by the output of MDNs. ANNs need a measure of confidence in their estimates to work with thexC;filters since navigation xC;filters require a confidence of their estimates. In our problems we achieve an indoor localization accuracy of 86.7% for 50 discrete locations, and a 2D RMS error of 63m for a 1km2 area of navigation.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 13, 2018
- Accession Number
- AD1063265
Entities
People
- Joseph A Ii Curro
Organizations
- Air Force Institute of Technology