| |
Abstract:
We consider the problem of learning a dynamic deterministic
finite automaton with noisy inputs and outputs. The graph of the
DFA can be thought of as a topological map, which we are trying to
learn using a robot which has unreliable sensors and actuators.
Furthermore, the underlying topology can change over time, to
reflect the fact that, e.g., doors can open and close. We therefore
consider the topology as a random variable (instead of a fixed
parameter), and try to learn it using Bayesian inference (instead
of EM). We use sample-based methods to do the inference
approximately. We give results for a simulated dynamic grid
world.
|
