1. Rule Learning for Go Introduction Data Extraction Bad Move Problems
Ripper Explanation
Example
Experimental Results

2. Objective Increase the playing strength of Explorer through
the use of machine-learned rules for local play
Why Learn Rules?
How We Learn Rules
Cohen’s RIPPER Algorithm
(Repeated Incremental Pruning
to Produce Error Reduction)
Local Move Generation
Speed

3. Overview Professional Games Data Extraction Ripper Explorer
Tests Vs. GNUGo

4. Data Extraction Ripper Format 0 – Off-Board 1 – Own Stone
2 – Own Territory
3 – Possible Own Territory
4 – Near Own Territory
5 – Neutral
… up to 11
Comma-Separated Lists
5,5,1,9,5,
1,5,1,10,5,
5,5,5,5,5,
5,5,5,9,5,
goodMove.

5. Processing Options Rotation Mirroring Nearby Bad Moves
5,5,1,9,5,
1,5,1,10,5,
5,5,5,5,5,
5,5,5,9,5,
goodMove.
Rotation
Mirroring
Nearby Bad Moves
Farther Bad Moves
Random Bad Moves
Stone Data Only
Limited Territory Data
All Territory Data
Extra Weighting
5,9,1,5,5,
5,10,1,5,1,
5,5,5,5,5,
5,9,5,5,5,
goodMove.
5,5,5,5,5,
5,5,5,9,5,
1,5,1,10,5,
5,5,1,9,5,
goodMove.
5,5,5,5,5,
5,9,5,5,5,
5,10,1,5,1
5,9,1,5,5,
goodMove.
Y – Axis
X – Axis
X & Y

6. How Ripper Works
Make GrowSet, PruneSet with 2/3 & 1/3 of each class respectively
Do { // Generate Ruleset
While NewRule covers negative examples in GrowSet
For All Attributes of the data
For All Possible Values of the Attribute
Evaluate GainOf( Attribute == Value )
If Negation is Allowed
Evaluate GainOf( Attribute != Value )
Add Condition for which Max. Gain occurred to NewRule
For All possible deletions of conditions from NewRule
Evaluate ((correctPositives - incorrectNegatives) /
totalCovered ) On the PruneSet
Implement deletion with maximum value
Add NewRule to Ruleset
} While #GoodMoves in GrowSet != 0 &&
Ruleset desc. length < min desc. length + D bits

7. How Ripper Works (Continued)
Repeat #Optimizations Times {
// Optimize Ruleset
For Each Rule in Ruleset
Create Replacement of Rule by removing Rule from
Ruleset and growing a new rule then pruning it so as
to minimize the error of the entire ruleset with the
replacement instead of the original on the PruneSet
Create Revision by greedily adding conditions to the
original
Evaluate Rev. and Rep. by evaluating the description
length of the rulesets having them switched with
the original, the choice with minimum description
length is kept.
Add new Rules again // As in previous slide. }
Description Length: Number of bits required to encode the information in a binary string.
Gain: (# moves covered by old rule still covered by new rule)*(Information of Old Rule –
Information of New Rule)
Information: - log2(goodMoves covered / total # moves covered by rule)

8. Ripper Example
Input Filter
Negative Input Data
Positive Input Data

9. Example Data Position1 = ‘1’ Position1 = ‘1’ Position4 = ‘1’
Valu e
Position
0.555
0.121
0.679 2
1.181
0.925
2.670 1
0.263
0.625 13 12 11 10 9 8 7 6 5 4 3
Position1 = ‘1’
Valu e
Position
0.541
-0.125
0.195
1.120 2
2.01
1.548 1
0.180
0.765
0.443 13 12 11 10 9 8 7 6 5 4 3
Position1 = ‘1’ Position4 = ‘1’
Valu e
Position
-0.176
-0.589
0.789
1.660 2
2.339 1
0.678
1.263
-0.058 13 12 11 10 9 8 7 6 5 4 3
Position1 = ‘1’ Position4 = ‘1’ Position2 = ‘1’

10. Position1 = ‘1’ Position4 = ‘1’ Position2 = ‘1’ Position3 = ‘2’
Example Data
Valu e
Position
-0.387
2.712 2 1
0.678 13 12 11 10 9 8 7 6 5 4 3
Position1 = ‘1’ Position4 = ‘1’ Position2 = ‘1’ Position3 = ‘2’
No negatives covered, all positives removed, attempt at making second rule wil fail.
Revision: Nothing occurs as all good moves are covered.
Replacement: Same rule learned.
Final hypothesis is:
goodMove :- pos1='1', pos4='1', pos2='1', pos3='2' (12/0).
default badMove (48/0).

11. Current Results 18 Rulesets Covering a Range of Processing Options
Only one completed test vs. GNUGo, no improvement seen.
Ruleset Used: 5x5 Square
Mirrors Added
Random Bad Moves
All Territory Data
Double-Weighted