Tuesday, July 14, 2015

Computers achieve autonomy

In the unofficial world-championship for engines (TCEC) the role of an opening-book has been restricted. The goal of the competition is to measure the pure strength of the algorithms so excluding any influence of manual corrections. A programmed opening-book can assure a significant advantage at least this was the view many shared until recently.

Today this theory is by more and more people questioned. Nowadays so many improvements are discovered in old openingbooks that the decision to early disconnect the engine from the openingbook , becomes increasingly attractive. Programmers start to prefer that the engine finds independently the moves in an opening. So the program plays better without additional human input but not yet without mistakes. To avoid the same errors reoccurring, statistics of the results are stored in a special openingbook. The engine learns autonomously a repertoire.

Nothing new some people will think as this already exists more than a decade. Correct but there is a big difference with the past. This special opening-book was considered only useful for an engine while today I see more and more applications for the otb and correspondence-player. Several current engines play stronger than Magnus Carlsen so it is not optimal anymore to only study the games of the top-grandmasters.

A nice example is the recent theoretical developments of the Fraser defense. The critical move 9.Rg1 which I already mentioned in my previous article, does not pop up in the big database 2015 neither in my correspondence-databases (updates till June 2015). However if we check my database containing a collection of computer-games published on sddfccrl and tcec then it transpires that Rg1 is not a novelty either, at the contrary.
36 computer-games with 9.Rg1
I am a chess-manic in my analysis so I am never satisfied with the number of found games. Therefore the next logical step is to create your own computergames played from a specific position using the available HW and SW. Though such standard games take too long without even thinking about the unattractive energy-bill. By playing shortened games as explained in my article analyze with a computer the process is accelerated. That allows me to achieve 18 games in 6 hours. Not bad but still too slow to my taste. I learned on chesspub from Vass that top-correspondence-players allow their engines to analyze at the speed of 10 seconds per move so I decided to test that strange looking method myself for this specific position. The final-result of several hundreds of games can be viewed below.
[Event "Ponziani 5...Bc5"] [Date "2015"] [Round "?"] [White "?"] [Black "?"] [Result "*"] [ECO "C44"] [PlyCount "41"] 1. e4 e5 2. Nf3 Nc6 3. c3 Nf6 4. d4 Nxe4 5. d5 Bc5 {(The Fraser-defense in the Ponziani.)} 6. dxc6 Bxf2 7. Ke2 Bb6 {(The old more popular bxc6 does not guarantee equality after Qa4.)} 8. Qd5 Nf2 9. Rg1 {(A novelty according to the Bigbase 2015 and my correspondence-database but already known from many computer-games.)} (9. cxb7 $6 {(Rejected by the theorists but it remains an interesting choice in practical play.)} Bxb7 10. Qxb7 Nxh1 11. Qe4 $5 {(Lets check also covers this move which is not easy to answer without any study. The alternative a4 can be answered by a5 or 0-0 with a small advantage for black.)} (11. Be3 Bxe3 12. Kxe3 O-O 13. Nbd2 d5 14. Bb5 Rb8 $19) 11... O-O (11... f6 12. Be3 c6 13. Nbd2 O-O 14. g3 f5 15. Qh4 Bxe3 16. Qxd8 Raxd8 17. Kxe3 f4 18. gxf4 exf4 19. Kd4 Nf2 $13) 12. Ng5 g6 13. Qh4 h5 14. Na3 d5 15. g4 Nf2 16. gxh5 Qe7 17. Bh3 $15) (9. Qxe5 $6 Kf8 10. Rg1 f6 {(Ng4 is also fine.)} 11. Qf4 g5 12. Nxg5 dxc6 13. Be3 Nd3 14. Qf3 Ne5 15. Bxb6 Nxf3 16. Bc5 Kg7 17. Nxf3 $15) 9... O-O $5 {(The critical test but dxc6 looks playable too.)} (9... dxc6 $5 10. Qxd8 Kxd8 11. Be3 Ng4 $1 12. Bxb6 axb6 13. h3 $5 Nf6 14. c4 $5 (14. g4 $5 Nd5 15. Nbd2 $5 f6 16. Kf2 $5 Nf4 17. a3 $5 g5 $13) 14... Ne4 15. Nxe5 $5 Re8 16. Nxf7 Kd7 $11) (9... bxc6 $6 10. Qxe5 Kf8 11. Nd4 Ng4 12. Qf4 c5 13. Nf5 $14) 10. cxb7 Bxb7 11. Qxb7 Qf6 12. Na3 {(I still found some computer-games with Qa6 and Qd5 but only Na3 looks sufficient for maintaining the balance.)} (12. Qd5 $6 c6 13. Qd2 e4 14. Nd4 e3 15. Qc2 Rae8 $1 {(A computer has the biggest difficulty to find the right moves in this position so impossible for a human to find independently the right track at the board.)} 16. Nf3 (16. Qf5 Qd6 17. Na3 Bxd4 18. Nc4 Qxh2 19. cxd4 g6 20. Qxd7 Nd1 21. Qxc6 Qxg1 22. Qf3 Re4 $15) 16... g5 $1 17. g4 (17. g3 g4 18. Nh4 Ne4 19. Rg2 Re5 20. Ke1 Nd2 21. Bd3 Nf3 $17) (17. Ke1 g4 18. Nd4 g3 19. hxg3 Bxd4 20. cxd4 e2 21. Bxe2 Qxd4 $17) (17. h4 gxh4 18. Na3 d5 $1 $17 {(An improvement on the analysis published on a chess-forum.)}) 17... Ne4 $1 18. Na3 Qf4 $1 19. Bh3 h5 $5 {(Re6 is another way leading to Rome.)} 20. gxh5 f5 21. Bg2 g4 22. Nc4 gxf3 23. Bxf3 Kh8 24. Nxb6 Nd2 25. Rg3 Nxf3 $19) (12. Qa6 $2 e4 13. Nd4 c5 14. Nc2 d5 15. Be3 Ng4 $17) 12... e4 {(I consider this as the mainline but black has interesting alternatives to avoid a preparation.)} (12... Rab8 $5 13. Qd5 c6 14. Qxd7 (14. Qd2 e4 15. Nd4 e3 16. Qc2 Bxd4 17. cxd4 $13) 14... e4 15. Nd4 Nd3 16. Be3 Rfd8 17. Qxc6 Bxd4 18. Qxf6 Rxb2 $11) (12... Qf5 $5 13. Qd5 ( 13. Qa6 Rfe8 14. Qc4 Ng4 15. h3 e4 16. hxg4 Qxg4 17. Kd1 d5 $13) 13... c6 14. Qc4 d5 15. g4 Qd7 16. Qa6 f5 17. g5 f4 18. Rg2 e4 $13) 13. Nc4 Rab8 {(This allows black to avoid an immediate draw.)} (13... exf3 $5 14. Qxf3 Rae8 15. Be3 Bxe3 16. Nxe3 Rxe3 17. Qxe3 Ng4 18. Qxa7 $11) 14. Qd5 exf3 {(Again black has 2 fun alternatives.)} (14... Ng4 $5 15. Kd2 (15. Bg5 Qg6 16. Nd4 Bxd4 17. cxd4 c6 18. Qa5 d5 19. Be3 dxc4 $13) (15. Nxb6 Rxb6 16. Bd2 exf3 17. Qxf3 Qg6 18. Kd1 $11) 15... exf3 16. Nxb6 f2 17. Rh1 Rxb6 18. Kc2 Ne5 19. b3 $11) (14... Nh3 $5 15. Rh1 exf3 16. Qxf3 Rfe8 17. Kd2 Nf2 18. Qxf6 gxf6 19. Nxb6 Rxb6 20. Rg1 Rd6 21. Kc2 Re1 22. b4 Rdd1 $11) 15. gxf3 Rfe8 { (I have a slight preference for this rook-move but probably it makes little difference in the end.)} (15... Rbe8 $5 16. Kd2 Ne4 17. fxe4 Bxg1 18. Bd3 c6 19. Qf5 Qxf5 20. exf5 Bxh2 $13) 16. Kd2 Ne4 17. fxe4 Bxg1 18. Bd3 { (It is noteworthy that black had a number of playable lines while white had to play each time the only move to avoid a disadvantage.)} (18. Qf5 $6 Qxf5 19. exf5 Bxh2 20. Rb1 Bf4 21. Kc2 Bxc1 22. Rxc1 $15) (18. Be2 $2 c6 19. Qd6 (19. Qf5 Qxf5 20. exf5 d5 21. Na3 Be3 22. Kd3 Bxc1 23. Rxc1 Rxb2 $17) 19... Qxd6 20. Nxd6 Re6 21. Kc2 Bxh2 22. Nf5 Rxe4 23. Bd3 Rg4 $17) 18... c6 {(This allows exchanging the queens. There is also the alternative Qf2 to keep the position complex tactically.)} (18... Qf2 $5 19. Kd1 Bxh2 (19... c6 {(Or anyway exchanging the queens.)} 20. Qf5 Qxf5 21. exf5 Bxh2 22. Bd2 Bc7 23. b3 d5 24. Nb2 $13) 20. Be3 Qg2 21. Rb1 Rbd8 $13) 19. Qf5 {(Again white has not really a choice.)} (19. Qd6 $6 Qf2 20. Kd1 Bxh2 21. e5 Bg3 22. Rb1 f6 23. Be3 Qh2 24. Qc7 $15) 19... Qxf5 20. exf5 Bxh2 {(This is logical but not mandatory.)} (20... d5 $5 21. Na5 c5 22. b3 Bxh2 23. Kd1 Rb6 24. Bd2 Bc7 25. c4 Ra6 26. Nb7 $13) (20... c5 $5 21. Nd6 Re7 22. h4 Bh2 23. Nb5 d5 24. b3 a6 25. Na3 $13) 21. b3 {(The first white move for which I found a playable but risky alternative: b4. This position requires extra research and definitely is worth testing in practice. Komodo prefers slightly blacks chances but Stockfish let us believe that it is an easy draw. Anyway a clear path to an advantage for black is not immediately clear for me.)} *

I already cleaned up the analysis so it became readable but anyway I am very satisfied of the results for this specific analysis. I used Stockfish 6 supported by the Fritz 14 interface. A further automation of the analysis is possible with the tool Aquarium already 5 years available on the market. That tool allows to execute several projects automatically in a parallel or sequential mode and automatically structures the analysed positions. Another important feature is that no limitation exists in the number of lines you want to analyze as opposed to the old "deep position analysis" mode of the Fritz 14 interface.

So computers achieve more and more autonomy for which we again pay a price. This price is paid e.g. in correspondence-chess as top-players due to the influence of the engines don't succeed anymore to win games.
The wc-finale which ended in 2002 and was won by Gert-Jan Timmerman

  Compare with the still running wc-finale which started end of 2013.
The still running wc-finale which started end 2013.

In 1 decade the drawing-rate went up till almost 100%. We should ask ourselves if such championships still make sense. The former worldclass-player Nikolai Papenin is the only player not performing as expected likely because of an early drop out. Recently the strong German correspondence-grandmaster Arno Nickel published an article on chessbase to get attention for this problem.

We are today entering a phase in which engines neutralize creativity by extensive analysis which can detect the risks in time. On the Utrecht chess-forum in the thread about Norway Chess  the Dutch programmer and fide-master Vincent Diepeveen rightly remarked that he can today in most endgames or advanced middle-game positions tell which moves are the best after a couple of hours extensive plugging on the computer.

Fortunately we aren't allowed to use engines at the chessboard so we should not fear the dead of chess by draws in standard-chess. Neither will our game be solved in the nearby future. However further progression of the understanding of chess will be managed increasingly by autonomous computers.


No comments:

Post a Comment