Passive Model Learning under Positional Uncertainty
       ---------------------------------------------------

Sven Koenig		Weh7220			3:30 pm, June 14

I have developed a probabilistic navigation approach that is centered
around a partially observable Markov model. The task of the Markov
model is to perform position estimation in the absence of complete
metric information about the environment and in the presence of noisy
effectors and sensors. The Markov model contains uncertain distance
information and probabilistic sensor models.

One advantage of this navigation approach is that the navigation
module does not need to know a complete metric map. The question
remains, though, how the Markov model should be refined to improve the
performance of the navigation module. This involves both learning
better distance estimates and adapting the sensor models to better
reflect the characteristics of the current environment.

In this talk, I will discuss how the Baum-Welch algorithm can be used
for this purpose, even in the presence of substantial positional
uncertainty. Advantages of this algorithm include that it learns
purely by observing (it does not need to control Xavier at any point
in time) and that it can take simple geometric knowledge into account
(one might for example know that two parallel corridors are equally
long). Disadvantages include that the Baum-Welch algorithm is usually
rather training data and memory intensive.

I have implemented a simple version of the Baum-Welch algorithm and
interfaced it with the navigation and position estimation module. In
the talk, I will review the Baum-Welch algorithm and describe how the
disadvantages of the Baum-Welch algorithm can be overcome by using
time windows and state classes.