Plausible Move Generation Using Move Merit Analysis in Shogi

Plausible Move Generation Using
Move Merit Analysis in Shogi
Reijer Grimbergen (Electrotechnical Laboratory)
Hitoshi Matsubara (Future University Hakodate)
Full-width Search and
Plausible Move Generation
Full-width search

World championship level programs in chess, checkers
and Othello.
Plausible Move Generation


Important in the early days of chess research.
Good alternative in domains where full-width search
can not search deep enough:
Large average number of legal moves (e.g. Go, Shogi).
 Single agent search problems with extremely long solution
sequences (e.g. Sokoban).


Interesting for cognitive science.
Why Plausible Move Generation
in Shogi?
High average branching factor (80)
compared to chess (35).
Average branching factor increases as the
game progresses.
Time constraints are strict.
Legal Moves in Shogi
250
Legal Move Avg
200
Highest Legal Move
Number
Average Legal Move
Number
Lowest Legal Move
Number
150
100
50
0
0 10 20 30 40 50 60 70 80 90 00 10 20 30 40 50
1 1 1 1 1 1
Move Number
Plausible Move Generator Set for Shogi
PMG-Goal


Capture Material
Promote Piece
PMG-Th





Check
Attack king
Attack material
Discovered attack
Threaten promotion
PMG-DefTh





Defend checks
Defend king
Defend material
Defend discovered attacks
Defend against promotion
PMG-Pim






Defend pins
Tie improvement
Defend undefended pieces
Defend against exchange
Cover squares in own camp
Develop pieces
PMG-DefPIm



Pin piece
Cover squares in enemy camp
Avoid development
Move Merit Analysis
In our approach it is possible that moves are generated by
more than one PMG.
Knowledge about which PMGs generated a move is used
for analyzing the merit of a move.
Move Merit Analysis: Every PMG gives a value to the
generated move based on the importance of the PMG.

E.g.: Winning a rook is more important than improving the
position.
Possible performance improvements:


Move ordering is improved, aiding alpha-beta search.
Additional cuts are possible: discard all moves with a negative
MMA value.
Experimental Results
We compared the behavior of:


PMG-All: Generate all moves from the plausible
move generator set.
PMG-MMA: Generate all moves with a positive
MMA value.
Four experiments:
1.
2.
3.
4.
Plausible move generation test.
Move ordering test.
Search comparison test.
Self play experiment.
Experimental Results (1)
Plausible Move Generation Test
70
60
50
40
30
20
10
0
Game Number
10
0
90
80
70
60
50
40
30
20
PMG-ALL
PMG-MMA
10
0
Save Percentage
Savings of PMG-ALL and PMG-MMA in 100 Test Games
(12097 positions)
Experimental Results (1)
Plausible Move Generation Test
Savings and accuracy of PMG-All and
PMG-MMA in 100 test games:
Version
Not
Accuracy Savings
Generated
(%)
(%)
PMG-ALL
81
99.4%
23.7%
PMG-MMA
144
98.8%
46.5%
PMG-MMA: savings much better and only slightly less accuracy
Experimental Results (2)
Move Ordering Test
100
90
80
70
60
50
40
30
20
10
0
10
0
90
80
70
60
50
40
30
20
Cumulative
Absolute
10
0
Percentage
Move Ordering Results
Rank in Move Ordering
Only very few moves are ordered low by MMA
Experimental Results (3)
Search Comparison Test
Performance in 298 tactical shogi problems from Shukan Shogi:
Cat Total Pos
Full Width
PMG-ALL
PMG-MMA
1
50
17
21
22
2
50
9
8
13
3
50
10
10
11
4
50
9
7
8
5
50
5
4
4
6
48
4
5
6
Tot
298
54
55
64
Only for PMG-MMA there is a significant improvement
Experimental Results (4)
Self Play Experiment
We played 20 game matches between shogi programs that
use PMG-MMA, PMG-All and Full-width search:
No
Version
1
2
3
Result
1
PMG-MMA
X
15-5
(75%)
20-0
(100%)
35-5
(87.5%)
2
PMG-ALL
5-15
(25%)
X
16-4
(80%)
21-19
(52.5%)
3
Full-width
0-20
(0%)
4-16
(20%)
X
4-36
(10.0%)
PMG-MMA outplays PMG-All and Full-width search
Conclusions
Plausible move generation deserves further
investigation.
Plausible Move Generation with Move
Merit Analysis gives important
improvements of the search performance in
shogi.
Move Merit Analysis is vital for our method
of plausible move generation.