Speaker：
Prof. Martin Muller (University of Alberta)

TITLE:
Feature Learning in Computer Go using Latent Factor Ranking: A
Detailed Case Study Using the Fuego Program

ABSTRACT:
Supervised learning of evaluation features from human game records has
been used with great success in Monte Carlo-based computer Go programs
by Remi Coulom and many others. An effort to define, learn and use
such features has recently been undertaken for the open source Go
program Fuego. The learning method used is Wistuba’s recent Latent
Factor Ranking (LFR) algorithm. This talk describes the experiences in
converting a largely rule-based program into a feature-based one, the
ongoing work on defining and refining features, and the results
obtained for feature learning and for this new version of Fuego.