Learning discrete Bayesian models for autonomous agent navigation

摘要

Partially observable Markov decision processes (POMDPs) are a convenient representation for reasoning and planning in mobile robot applications. We investigate two algorithms for learning POMDPs from series of observation/action pairs by comparing their performance in fourteen synthetic worlds in conjunction with four planning algorithms. Experimental results suggest that the traditional Baum-Welch algorithm learns better the structure of worlds specifically designed to impede the agent, while a best-first model merging algorithm originally due to Stolcke and Omohundro (1993) performs better in more benign worlds, including such model of typical real-world robot fetching tasks.