Decision Making Based on Convex Sets of Probability Distributions: Quasi-Bayesian Networks and Outdoor Visual Position Estimation

Abstract

The thesis advanced by this dissertation is that convex sets of probability distributions provide a powerful representational framework for decision making activities in Robotics and Artificial Intelligence. The primary contribution of this dissertation is the development of algorithms for inference and estimation in two domains. The first domain is robustness analysis for graphical models of inference. Novel results are developed for models that represent perturbations in Bayesian networks by convex sets of probability distributions. The dissertation reports on a system, called JavaBayes, that uniformly handles standard probability distributions and convex sets of probability distributions. This system is publicly available and has been used for teaching and research throughout the world. The second domain explored in this dissertation is outdoor visual position estimation for mobile robots. A novel algorithm for visual position estimation is derived in the context of remote driving for mobile robots in open, natural environments.This algorithm has been implemented in the Viper system, and field tested in a variety of environments, displaying accuracy and functionality levels that surpass previous work.