Feature Dynamic Bayesian Networks

Mayank Daswani (ANU)

CS HDR MONITORING AI Group

DATE: 2011-04-20
TIME: 12:00:00 - 12:30:00
LOCATION: Ian Ross Seminar Room, R214 with Pizza
CONTACT: JavaScript must be enabled to display this email address.

ABSTRACT:
The aim of Feature Reinforcement Learning is to automatically reduce a complex real-world problem into a useful representation. phiMDP achieves this goal by constructing an unstructured Markov Decision Process from the agent's history. This works well for small problems. However, without structure state spaces can become very large and hence harder to learn. The aim of phiDBN is to instead use structured Dynamic Bayesian Networks. This allows complex problems to be represented in a relatively smaller number of nodes, but has its own challenges. This talk will explain briefly my former work on phiMDP, and my plans for working with phiDBN.