Members - Faculty, students, and collaborators |
News - Recent news and publicty from members of the IDEA lab |
| Theses and Dissertations - Publications and code releases for student theses and disserations |
| Publications - Recent technical papers and presentations |
Software - Recent software releases |
Abstract
We introduce a method of learning spatial probability distributions for discrete gridded domains. We develop a decision tree based approach, which we call Spatial Probability Trees (SPTs).
SPTs split the data by asking questions about statistics calculated upon local regions in the domain. The leaves of each tree are local probability density functions. By aggregating these regional probabilities across the entire space, SPTs can be used to represent global probability density functions. We also introduce an ensemble of SPTs called a SPT forest, which significantly improves performance.
We apply SPTs to computer Go to predict expert moves. The learned SPTs can then be used to provide advice to Monte Carlo based computer Go players. We specifically demonstrate that SPTs and SPT forests can be used to significantly improve the playing strength of a state-of-the-art computer Go player that uses Upper Confidence Bounds for Trees.
Thesis
Zachery Tidwell (2012) Expert Move Prediction for Computer Go using Spatial Probability Trees. Master's Thesis, School of Computer Science, University of Oklahoma.
SPT Code
The code for the thesis can be released on request.