Michael Buro's Reserach
- Search and inference in imperfect information games. We are using
the trick-based card game Skat as research testbed. Our program
kermit is the strongest in the world.
- Machine Learning, heuristic search, and planning in multi-player
real-time strategy (RTS) games. This is a hot AI topic which lies in
the intersection of commercial computer games and military research on
computer generated forces. The goal of this project is to use
sophisticated learning and planning techniques to construct an AI
system capable of defeating human experts in this domain. This is a
long-term project which uses the ORTS system as test domain.
- Evaluation function learning in heuristic search. Constructing
evaluation functions automatically is a hard problem. We understand
how to tune thousands of numerical parameters, but we don't know how
to generate feature sets. The idea behind this research is to combine
inductive feature generation with least squares weight optimization (GLEM).
It has been very successfully applied to the game of Othello. Now we
are looking for other applications and model improvements.
- Design and implementation of a hack-free real-time strategy
game environment (ORTS). Issues: handling thousands of
moving objects in real-time in a server-side simulation, fast
collision detection, fast view computation, efficient network
communication), ergonomic GUIs, AI plugins.
- Selective mini-max search. How to exploit statistical properties
of evaluation functions to improve the efficacy of look-ahead search?
Multi-ProbCut (MPC) is an algorithm that exploits evaluation correlations.
This project focusses on how MPC can be improved.