Suzie a chessprogramm which plays Go

1
Suziea chessprogramm which playsGo

• Chrilly Donninger
• Nimzowerkstatt OEG

2
Suzie Contributors

• Peter Woitke
• Chrilly Donninger
• Stefan Mertin
• Gert Schnider
• Disclaimer This lecture reflects the opinion of
  Chrilly. Peter, Stefan and Gert will see things
  different.

3
Chess programm Hydrasponsored by Pal-Computing
UAE

• Same architecture as Deep Blue
• 64-processor cluster
• Special purpose hardware (FPGAs)
• Evaluation is based on sophisticated
  mobility-concept
• gt 3000 Elo, never lost to human
• Hydra-M.Adams 5.5 0.5.
• Game over

4
Suziefirst plan

• Replace mobility with influence
• Use Hydra search
• Make prototype in C
• Find Go-Sheikh
• Implement it in FPGAs and build Cluster
• Small Problem Influence is not sufficient
• Big Problem There is NO Go-Sheikh

5
Branching Factor ist NO problem

• Chess programms do NOT search brute-force. Search
  is selective
• Hydra Depth 18 7-40 Plies
• Optimal exponential factor 6-7
• Real with extensions 3-4
• Without extensions 2-3
• 9x9 Go has similar branching-factor
• 19x19 Real branching factor lt 10
• The problem is the efficiency of the search and
  NOT the branching factor. 12x12 chess would be
  easy.

6
Technical problems

• Subtile differences to chess.E.g. Hashcode gt
  nasty bugs
• There are several rule sets.
• Symmetry of board is in chess no topic, in Go it
  is.
• Chess Board-representation and efficient
  data-structures are an own science.
• Go No topic at all. Speed does not matter. One
  has to start at AdamEve.
• Nasty special cases at border and corner. Go on a
  sphere would be much easier.
• For humans simple questions are difficult to
  programm in an efficient way. E.g. Is point
  within area or outside.

7
A knight is a knight

• Chess A bad knight is worth 3 pawns, a good
  knight is 4 pawns, but a knight is a knight.
• Go Material value of a stone pointless
• Which pieces aka groups are on the board at all?
• Search efficiency depends on move-ordering.
  Chess Capture highest piece.
• Quiescence-Search. Do all profitable captures and
  stop if nothing to capture.
• There is not such a simple rule in Go gt Search
  is much less efficient.

8
The kermis problem(on can only dance on one)

• Several local games/fights.
• Who moves in which subgame first?
• Value of the right to move?
• Mathematical SubGame Theory is useless.
• Evaluation needs local tactic-solver. Makes Eval
  slow. No problem in FPGA. Can be done in
  parallel.

9
The shark problem

• White Group B can kill black Group A. B can be
  killed by black C... Who will survive?
• Rules like counting liberties have many
  exceptions. In fact effective liberties.
• Determination of group strength must be done
  recursive. Slow and no fix-point.

10
Handling the bizarre

• Most positions generated by an Alpha-Beta search
  are bizarre. E.g. quattro-pawn in chess happens.
• Evaluation must handle the bizarre cases. E.g. A
  long ladder is worse (for the prey) than a short
  one. Otherwise the program knows it is in a
  ladder, but it does not worry.
• Conventional programms generate only moves which
  the programm understands. Ladder-move is only
  played if ladder-breaker exists.

11
Horizon-effect

• Chess Greatest problem in punchcard-age.
  Programm sacrificies smaller pieces to push the
  capture of a major one over the search-horizon.
  But delaying moves have severe costs.
• Go Delaying move has almost no cost. Plays
  Auto-Atari-Atari in opponents territory.
• Programm pushes not only heavy tactical problems
  over horizon, but also positional ones.
• Looks terrible and gives away many small
  advantages.

12
A Ply is still a Ply

• An additional ply of search is in Go at least as
  effective than in chess.
• There are just too less plies. The advanced
  search techniques of chess work only in deep
  trees.
• Current conventional hardware is too slow for
  Alpha-Beta.
• There is existing hardware around which is fast
  enough (e.g. Hydra Cluster with FPGAs).
• One could build within 3 years and a budget of
  500.000 Euro a Deep Go
• Deep-Go would beat all humans in 9x9 and the best
  Austrian player on 19x19.

13
UCT - the balance of stupidity

• 2 apes are playing many games from a position. If
  the black ape is winning more games, the black
  position is better.
• Apes do not like a waste of time, replace them by
  random player.
• Try all moves, let aps play. Move with highest
  winrate is best.
• Concentrate effort on most promising moves,
  replay from time to time also bad moves. Maybe
  they had just bad luck.
• Do the same recursively in a game tree.
• We have an UCT Go-Engine.
• UCT is currently the most successfull approach.
• Plain-vanilla UCT (like UCT-Suzie) is very easy
  to programm.
• The knowledge is the search. Speed is very
  important.
• Complete Antithesis to AI-approach Yet another
  Drosophila lost.

14
The real problem

• There is no money in computer Go.
• Industrial programming 60 Euro/h.
• 2000h x 60E 120.000 E.
• This equation is on both sides wrong
• 20.000E/5000h 4E/h.
• Austrian minimal wage is 6E/h.
• There is no interest for Deep Go or Hydra-Go.
• Working for 4 instead of 60E/h is professionally
  completly stupid.

15
The real real problem

• Number of Go players in Japan (Shukan Igo
  Shimbun) 1981 12 Million 2005 3,5 Million
• Chrillys Hypothesis Slow games like chess, go,
  bridge will disappear from the social mainstream.
  They do not fit to the ever faster rythmn of
  life. Waste of time.
• Every mainstream creates a counter-stream. The
  slow games will survive within a
  subculture-elite.
• Poker fits much better to the modern times.
• Technology is driven by action-game playing kids
  Cell processor in Playstation 3, NVIDIA-graphics
  card. These new technologies will have great
  impacts on a number of fields (e.g.
  Computer-Tomography).

16
Conclusion

• Computer-Go is a waste of time for professional
  programmers.
• Go is the next big challenge is the usual
  AI-blabla.
• There are for programmers - much more
  interesting fields around.
• In Europe Go was never Mainstream, it has a
  bright future as an elitist-cult-game.
• Best of all The players will not be disturbed by
  programms and their ignorant programmers.