rule learning for go an overview roy de bruijn. rule learning process convert board state to text 5...
Post on 20-Dec-2015
216 views
TRANSCRIPT
Rule Learning for Go
An Overview
Roy de Bruijn
Rule Learning Process
Convert Board State to Text
5 5 4 1 9
5 5 11 5 9
6 9 10 9 8
8 8 8 8 8
8 8 8 8 8
Rule Learning Process
Feed Text to RIPPER to learn rules
5 5 4 1 9
5 5 11 5 9
6 9 10 9 8
8 8 8 8 8
8 8 8 8 8
goodMove 428 15 IF position12 = 5 position22 != 5 position14 != 5…
Rule Post-Processing
Rule Scope Length of Time from Activation to Used/Broken
Rule Usage Statistics Is the rule:
Used by the player it’s in effect for Broken by the opposing player Broken by the player it’s in effect for
Rule Evaluation
Explorer vs. GNUGo Test Games Static Move Prediction
Rule Weighting Center Point Weight = (Used + OBroke) / SBroke Tested Points’ Weight = SBroke / (Used + OBroke)
Best Results
Ruleset Accuracy
9x9c-mna2-1third 18.78%
5x5-mnl2-1third 17.25%
5x5-mna2-1third 17.02%
5x5-mfa2-1third 16.69%
5x5-mra2!-1third 16.57%
5x5-mna2!-1third 15.81%