Shogi and Realization Probability search

1
Shogi and Realization Probability search

• Hashimoto Tsuyoshi
• JSPS Research Fellow

2
Chess, Shogi and Go

• Average number of possible moves
• Go gt Shogi gt Chess
• Chess simple (brute force was enough for
  beating human champion!)
• Shogi complex ( ?reuse of captured pieces)
• Go shogis experience is useful!! (I wish)

3
My Shogi program Tacos

• Realization-Probability Search
• Tsuruoka, 2001
• Df-pn (mating search)
• Nagai, 2000
• SPH (brinkmate search)
• my Ph.D Thesis, 2001

4
My current 2 problems

• When call mating search?
• What is good category?
• ( To find good shape of search)

5
Realization-Probability (R-P) Search Tree
Threshold 0.1
Move probability
Deeper Wider
Shallower Narrower
Realization probability Position probability
6
Move Probability
Our program 2500 category

• Calculate from many games records
• Category (Shogi)
• Check, capture piece, attack, promotion
• Category (Go)
• Atari, Take stones, Tsuke, Kakari, ?

7
Why R-P search is effective?

• Systematic and statistical
• Simple search algorithm
• ? Complex games tuning parameters by hand is
  too difficult!
• Search deeply after important moves
• Drastically reduce search after unimportant moves
• ? Complex games Courage to stop searching is
  important!!

8
Advanced category 1 Just before group

• Just before capture, just before escape, just
  before attack
• ex Capture with high gain (pawn)
• normal 40.3
• just before 85.9

9
Advanced category 2 History Analysis

• Renda no Fu (re-attack pawn)
• 3.1?15.4
• Hashizeme Fu tarashi (put pawn at edge after
  sacrifice of pawn)
• 0.2?9.3

10
Advanced category 3Okuri no Tesuji

• ?2??(22gold)
• Before After
• ( check with loss) (Okuri no Tesuji)
• 0.592 ? 5.44

11
Advanced category 4 Area Analysis

• Move of pieces on board
• Come close to opponent king, soppo (keep off from
  opponent king)
• Put captured pieces
• Around opponent king, Around self king, normal
• Normal ? Soppo 12?00.3

12
Experiment

• With History Analysis Okuri v.s.
• Without History Analysis Okuri
• 61-38-1
• With Area Analysis v.s. without
• 58-42

13
Conclusion

• RPsearch good for Go?
• Advanced categories
• Pick up good category from big category of low
  probability!
• (or Pick up bad category from big category )

14
Future (Current) work

• Make programs stronger for tournaments (Computer
  Olympiad, Nov. GPW cup, Nov. The World Computer
  Shogi Championship, May.)

15
About Me

• Date of birth  September 17, 1970
• Place of birth  Kyoto City, Japan (Old capital,
  794 1869)
• Language Japanese gtgt Chinese(mandarin) gt English
• Hobby Shogi (4 dan, Rating 2100-2200 level of
  amateur champion in a city) , Travel, Camp
• Ph.D March 2002, Shizuoka University, Hamamatsu,
  Japan (about Shogi Endgame, brinkmate search)
• Programs Shogi (Tacos), LOA ((T-T)),
  AMAZONS(ASKA), 4 persons Shogi (HIRA)
• Big Goal Defeat Meijin (Champion of proffesional
  Shogi player)