Decision-theoretic planning is about getting computers to synthesize strategies for acting in an uncertain environment.
State-of-the-art planners and solution algorithms are terribly inefficient, because they do not reason about or learn good "control knowledge". In other words, they are inefficient because: (1) they do not leverage symbolic knowledge provided about their environment, and (2) because they do not learn from their mistakes
The project will focus on one or more of the following types of control knowledge.
- Landmark detection and/or exploitation
- Nogood detection and/or exploitation
- Representation and exploitation of nogoods
Your supervisor has (co-)authored a number of decision-theoretic planning systems, in C++. If you choose to focus on empirical work, you can leverage those code bases if you so choose.
Theoretical: demonstrate exponential separation
Empirical: demonstrate compelling efficiency gains in 1 or more benchmark problems
Advanced knowledge about artificial intelligence planning.
artificial intelligence, decision-theoretic planning, simulation, nogoods, uncertainty