Sensational! Y.Ossipov talking about STRELKA

[Dr.Wael Deeb]

Maybe the same applies to Rybka.

[mjlef]

He also used correlation analysis to try to figure out how Rybka worked, trying to closely match how it functioned. All that seems fine to me. Copying the tables is theft, however. I wonder if Vasik even put a copyright warning in the EXE? Hmmm..

So this does look like very clever reverse engineering. I am pretty impressed. A pity though. Something original would have been much more interesting.

[Dr.Ex]

So there you have it chess programmers.

The road to success:
1. Rewrite Fruit to bitboard.
2. Include improved tables of material imbalances.
3. Do some fine tuning.

Now get to work!

[mjlef]

The author hinted that Vasik had also used Fruit (and Vasik has admitted to readin Fruit and that it has some benefit). I am not sure how he would know that from a compiled program. For the guys who reviewed Strelka source code...was it similar in structure to Fruit? As I recall they stated the code was original and not like anything else they had seen. Of course, the move generator and evaluation would look quite different. Basic search code should look very similar. I guess if Strelka is ever made public we can all take a look.

[mclane]

material imbalance stuff is what ChessSystemTal has in it.
CSTAL can handle positions with different material.
its specialized on this.

Its funny. i remember when in the hague Don Daily and i let CSTAL (Dos)
play against his program on his laptop.

those were funny games.

[Event "Revange for Paderborn 95"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "XXXX2"]
[Black "CSystem Tal"]
[Result "0-1"]

1. e4 b6 2. d4 Bb7 3. Bd3 e6 4. Ne2 c5 5. c3 d5 6. exd5 Qxd5 7. Nf4 Qd7 8. dxc5
Bxc5 { oh - this is of course a position CSTal likes very much!} 9. O-O Nf6 10. Re1 O-O 11. c4 e5 12. Nh3 Rd8 13. Be2 Qc8 {oops - white gets into trouble....} 14. Nd2 Bxg2 { ! Ha ! played with a score of 3.63 !! and from the first second of the search-process !! Standard Tal-move ....you could see the flickering in Martins' eyes when I told him Tals evaluation-....} 15.
Kxg2 e4 16. Qc2 Qf5 {as discussion started if this sac was accurate. Evaluation of XXXX2 said: no this sac was stupid....} 17. Nf1 Nc6 18. Ng3 Qc8 19. Nxe4 { chess system tal says: this move was bad, no we can hit back... } Nxe4 {cstal evaluation is +1.63} 20. Qxe4 Re8 21. Qd3
Nb4 { many moves have gone since the first sac. Nothing is forced. But this is the problem: a search based program can only defend against something, that can be SEEN in the tree. Here nothing can be seen, therefore XXXX2 is not overprotecting much. + 0,76 evaluation} 22. Qd1 {evaluation +0,79 after this move. Chess System Tal feels well in this position because it knows about the fact that the king is exposed heavily and it is just a few steps to die here, other programs counting material would see it different} Qf5 23. Bf3 Nc2 {+1,12} 24. Rxe8+ Rxe8 25. Bd2 Nxa1 {CSTals eval says +1,73 } 26. Qxa1 Qg6+ {and still the problem is : the king is exposed. Moves and moves have gone, but the main problem that was build with Bxg2 is on the board. CSTal increases evaluation to +2.68 and higher because another defending piece - the queen on a1 - is far away from the main-point of the game} 27. Ng5 {horizont}
h6 {CSTal says +3,09} 28. h4 Qd3 {+4,47 !!} 29. Qc1 hxg5 {+4,42} 30. hxg5 Re5 {found late, before Rd8 was considered with 5.03, now Re5 was played with +5,80} 31. Bh5 {Martin still cannot believe, I guess.} Qe4+ {+8,51 says CSTal.}32. f3 Qh4 {+7,93} 33. f4 Re4 34.
Qh1 Qf2+ {Martin must have nerves like iron CSTal says +10,48 pawns} 35. Kh3 Rd4 36. Qa8+ {no lost game without revange-check! CSTal says +12,62 and plays...} Bf8 37. Bxf7+ Kxf7 38. Qb7+ Be7 {+13,77 pawns up and here Martin has enough... } 0-1

Wasn't this an unbelievable game ???


We discussed much in Den Haag. I met Don Daily, who is a nice guy working on the CILK-CHESS project. I guess this program has more chances ever to win against Kasparov than any other program.

Don Daily likes to implement chess-knowledge into his parallel-program. Although he had only a small notebook with him, we played some games, I wanted to know about cilk-chess - of course - and the best way to find out about a program is to watch the games, and watch the evaluations and main-lines.

[Event "?"]
[Site "15'game"]
[Date "????.??.??"]
[Round "1"]
[White "CSTal v231"]
[Black "Cilkchess 1CPU"]
[Result "1-0"]

1. e4 c5 2. Nf3 Nc6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 e5 6. Ndb5 {+0,50} d6 7. Bg5 {+0,39} a6 8.
Na3 b5 9. Bxf6 {+0,25} gxf6 10. Nd5 {+0,20} f5 11. Bd3 {first seconds CSTal considered about Bxb5, I was very nervous. NONONO I tried to hypnotize it} Be6 12. Qh5 {of course here we come again in a kind of position Tal likes more... +0,95} Bg7 13. O-O f4 14. c4 {+1,02} bxc4
15. Bxc4 {+1,39} O-O 16. Rac1 Rb8 17. b3 {cstal did not decided to eat with Bxa6} Qa5 18. Nb1 Kh8 19. Nbc3 f5 20. Rfd1 {temptation here was Ne7 Bxc4 and difficult stuff. I don't know why Rook-move came in the end...} Bd7 21.
Rd2 fxe4 22. Nxe4 {+2,40} Bf5 23. Ndc3 Nd4 24. Ng5 {+2,05} Bg6 25. Qh3 Bf6 26. Nge4 Bxe4 27.
Nxe4 d5 28. Nxf6 {+2,61} Rxf6 29. Bd3 Rf7 30. Rcd1 {+1,62} Rg8 31. Qh5 Rfg7 32. Bxh7 {+4,41 !! It looks that cilk has overseen some tactical stuff that is TAL's favourite job: fishing threads arround the king....never try to do king-attack stuff with cstal as opponent.} Rxg2+ 33.
Kh1 {+6,81} Rg1+ 34. Rxg1 Rxg1+ 35. Kxg1 Qxd2 {cstal evaluated finally +7,42 main-line says:Bd3+ Kg7 Qxe5+ Kf7 Qxd5+ Ke7 Qc5+ Ke6 Qb6+ Kd5 Bc4+ Ke5 Qc7+ Kf6 Qf7+ Ke5 Qe7+ Kf5 ...and strange stuff} 1-0

What a funny position on board!

As you now see Rolf, CSTal is almost always good for a spectacular sac or a promising undercover-job.
Here comes another sacrificial game....

[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "2"]
[White "CSystem Tal"]
[Black "Cilk-chess 1CPU"]
[Result "1-0"]

1. e4 c5 2. Nf3 Nc6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 e5 6. Ndb5 d6 7. Bg5 a6 8.
Na3 b5 9. Nd5 {+0,49} Qa5+ 10. Bd2 {+0,62} Qd8 11. Nxf6+ Qxf6 12. Be3 Be6 13. Nxb5 {oh no - here it comes, +0,17 for this tiny little sac} axb5 14.
Bxb5 Kd7 {here the king is there where we wanted him to be...} 15. Qd2 {+2,51} Qg6 16. f3 {+2,46} Be7 17. a3 {+1,24} f5 18. Bxc6+ {+3,06} Kxc6 19. exf5 Bxf5 20. g4 {+0,88 ugh - its not easy to fight against cilk-chess. Score goes up and down...}
Be6 21. O-O-O {but Tal is watching us in heaven and sometimes sending HELPING hands...} Ra4 22. f4 {sometimes you must give a pawn to bluff, and sometimes to fool the opponent} Bxg4 {Ha ! Bluff worked!!} 23. Qd5+ {+2,39 the bishop e6 made anything so safe in blacks area, now, having taken the poisened pawn, the game turns into the right direction, so that IN THE END the sac on b5 was alright - he - only joking here...} Kc7 24. Bb6+ {another piece is thrown in his mouth. Will he eat it too ?? +0,60 fail-high. Thats indeed the game of the feeding Tal.} Kxb6 25. Qb3+ Kc7 26. Qxa4
Qh6 27. Qa7+ { +2,20 and Tal feeling in his element....} Kd8 28. Qb8+ {+2,22} Bc8 29. Kb1 {not to forget the king-safety. +3,02 for this little move...} exf4 30. Rhe1 {+4,16} Re8 31. Qb6+ {+10,02 Just a few little moves and the game is over...} Kd7 32. Qa7+ {+9,10}
Kd8 33. Qa5+ {+9,75} Kd7 34. Qb5+ {don't worry, we will not give check-check-check and draw like in Paderborn against the spanish program, we fixed this bug !! } Kd8 35. Rd3 Bg4 36. Qb8+ {cannot tell you evaluation, must be mate announced or something....} Kd7 37. Qb7+ Kd8 38. Rc3 1-0


You see, with chess system tal the boring days of computerchess are over.

We will never see games where no side of the players have an idea, this is the step into funny games....

Of course it is much more complicated to win with black....

[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "3"]
[White "Cilk-chess 1CPU"]
[Black "CSystem Tal"]
[Result "0-1"]

1. d4 Nf6 2. Nf3 d5 3. c4 c6 4. e3 e6 5. Bd3 Nbd7 6. O-O dxc4 7. Bxc4 Bd6 {-0,41} 8.
Nc3 Nb6 9. Bd3 O-O 10. e4 Be7 11. Bg5 Ng4 {-0,80} 12. Bxe7 Qxe7 13. h3 Nh6 14. Rc1 Rd8
15. Re1 e5 {-0,47} 16. dxe5 Kh8 {+0,03} 17. Re3 Be6 18. Re2 Qb4 {-0,02} 19. Rd2 Qa5 20. b3 Nd7 21. Be2
Nxe5 {-0,16 } 22. Nxe5 Rxd2 23. Nxf7+ {he cilk - we are Tal, not you !!} Nxf7 24. Qxd2 Rd8 {-0,26} 25. Qe3 a6 26. f4 Qb4 27. Kh2
Qe7 {the programs are fishing arround, nothing special...CSTal tries to activate it's queen. A queen needs a king to mate...} 28. g3 Qf6 29. e5 Qh6 30. h4 Qg6 {-0,90, CSTal is not satisfied with the position. Will Cilkchess win this game?} 31. b4 Nh6 {0,00 ! Tal feels something coming} 32. Qb6 {Ha - I told Don that this is brilliant! Now CSTal feeds Cilk with UNIMPORTANT pawns. The Bronstein-method of handling with computerchess. Don laughed, but maybe he did not believed it right here ?} Rd2 {+0,53 CSTal on it's way making a win!} 33. Qxb7 {unimportant pawn number 1. +1,18 says CSTal. } Qd3 34.
Qa8+ Ng8 {+1,34} 35. Qxc6 {unimportant pawn number 2} Bg4 {+2,03 and Tal in good mood! 3 pawns less but having an attack...} 36. Qg2 Bxe2 {+2,80} 37. Nb1 Rb2 {+2,88} 38. a3 Qe3 39. Re1 Ra2 40. h5
Nh6 41. Qa8+ Ng8 42. Qg2 Qd3 {+2,91} 43. Nc3 Qxc3 {+2,16} 44. Rxe2 Rxa3 45. e6 {+2,38} Ne7 46. Rc2 Qf6 {+1,94}
47. Qa8+ Ng8 48. Qe4 Qd8 49. Qe2 Ra1 {+2,17} 50. h6 {Hu - the trouble comes nearer ! Good to know that Tal works also to defend own king-position, not only to attack with it....+2,13} Qd5 51. hxg7+ Kxg7 52. Qg4+ Kh8 53.
Kh3 h5 {+4,44} 54. Qg6 Rh1+ {+4,48} 55. Rh2 Qf3 {+2,97} 56. Qc2 Rg1 {+3,93} 57. Qg6 Qf1+ {+4,39} 58. Kh4 Qxf4+ {+4,09} 59. gxf4
Rxg6 60. f5 Rg4+ {+1,05} 61. Kxh5 Nf6+ {+1,19} 62. Kh6 Rg3 63. Rh1 Ng8+ 64. Kh5 Kg7 {+1,91} 65. Kh4 Rg2
66. Rf1 Kf6 {+4,14} 67. Kh3 Rb2 68. Rf3 Rxb4 69. Ra3 Rb6 70. Ra5 Ne7 71. Kg4 Nc6 {+4,75} 72.
Ra4 a5 {+5,15} 73. Re4 Rb4 {+6,69} 74. Kg3 Rxe4 {+19,79} 0-1

[bob]

The author hinted that Vasik had also used Fruit (and Vasik has admitted to readin Fruit and that it has some benefit). I am not sure how he would know that from a compiled program. For the guys who reviewed Strelka source code...was it similar in structure to Fruit? As I recall they stated the code was original and not like anything else they had seen. Of course, the move generator and evaluation would look quite different. Basic search code should look very similar. I guess if Strelka is ever made public we can all take a look.

the problem is the conversion to bitboards. That makes surface changes _everywhere_. It would take months of careful study to determine that "A using mailboxes" and "B using bitboards" are the same identical algorithm. The code would just look so different. (Compare any of Crafty's code to a mailbox program to see what I mean).

[mjlef]

I suppose it all depends on how the program was structured. Tord went from mailbox to bitboard yet all the specifics are well placed in functions. So the search part of the code and even the evaluation looks remarkably similiar.

I am still impressed by this reverse engineering job. But I feel bad. SHould I be impressed by a copy of the Mona Lisa? I guess I am impressed technically, although I suppose many of us could have done the same thing, given enough time...and sleepless nights.

[Alexander Schmidt]

he was saying that he believe Vasik Rajlich walked the same way as him...

Hi,

I just picked up a mail I wrote at the time when the first Rybka beta was released. See below, no comment.

Best,
Alex

----------------------------------------

I just took a quick look on Rybka, as you know I am always sceptical with new and very strong engines.

My first view is always the look on the debug output, and there are several similaries to Fruit/Toga:

Rybka:

2275625<1:info depth 3
2275627<1:info depth 3 score cp 7 time 8 nodes 235 nps 30080 pv b1c3
2275634<1:info depth 3 time 15 nodes 328 nps 22391
2275650<1:info depth 4
2275652<1:info depth 4 score cp 3 time 23 nodes 491 nps 21860 pv b1c3
2275656<1:info depth 4 time 28 nodes 826 nps 30208
2275658<1:info depth 5
2275677<1:info depth 5 score cp 8 time 45 nodes 1117 nps 25417 pv b1c3 b8c6
2275680<1:info depth 5 time 51 nodes 1352 nps 27146
2275683<1:info depth 6
2275703<1:info depth 6 score cp 3 time 69 nodes 1913 nps 28390 pv b1c3 b8c6 g1f3
2275746<1:info depth 6 time 112 nodes 4005 nps 36617
2275748<1:info depth 7
2275784<1:info depth 7 score cp 5 time 160 nodes 6216 nps 39782 pv b1c3 b8c6 g1f3 g8f6
2275837<1:info depth 7 time 216 nodes 9565 nps 45345
2275839<1:info depth 8
2275926<1:info depth 8 score cp 7 time 298 nodes 14615 nps 50220 pv b1c3 b8c6 g1f3 g8f6 d2d4
2276015<1:info depth 8 time 380 nodes 19009 nps 51224
2276017<1:info depth 9
2276586<1:info depth 9 score cp 5 time 961 nodes 47418 nps 50526 pv b1c3 g8f6 d2d4 d7d5 c1f4 c8f5
2276769<1:info depth 9 time 1150 nodes 57842 nps 51504
2276771<1:info depth 10
2277664<1:info depth 10 score cp 3 time 2028 nodes 93647 nps 47285 pv b1c3 g8f6 d2d4 d7d5 g1f3 b8c6 d1d3

Fruit:

2237122<1:info depth 3
2237124<1:info depth 3 seldepth 3 score cp 54 time 8 nodes 148 pv b1c3 b8c6 g1f3
2237128<1:info depth 3 seldepth 3 time 9 nodes 186 nps 0
2237130<1:info depth 4
2237133<1:info depth 4 seldepth 6 score cp 0 time 10 nodes 300 pv b1c3 b8c6 g1f3 g8f6
2237136<1:info depth 4 seldepth 6 time 13 nodes 976 nps 0
2237151<1:info depth 5
2237157<1:info depth 5 seldepth 9 score cp 48 time 52 nodes 1729 pv b1c3 b8c6 g1f3 g8f6 d2d4
2237163<1:info depth 5 seldepth 9 time 54 nodes 1933 nps 0
2237165<1:info depth 6
2237167<1:info depth 6 seldepth 12 score cp 0 time 59 nodes 3331 pv b1c3 b8c6 g1f3 g8f6 d2d4 d7d5
2237189<1:info depth 6 seldepth 12 time 94 nodes 9447 nps 0
2237191<1:info depth 7
2237226<1:info depth 7 seldepth 14 score cp 42 time 119 nodes 15332 pv b1c3 b8c6 g1f3 g8f6 d2d4 d7d5 c1f4
2237231<1:info depth 7 seldepth 14 time 127 nodes 16243 nps 0
2237233<1:info depth 8
2237296<1:info depth 8 seldepth 17 score cp 0 time 190 nodes 35078 pv b1c3 g8f6 g1f3 b8c6 d2d4 d7d5 c1f4 c8f5
2237419<1:info depth 8 seldepth 20 time 313 nodes 72286 nps 0
2237421<1:info depth 9
2237580<1:info depth 9 seldepth 20 score cp 15 time 479 nodes 125215 pv b1c3 g8f6 g1f3 b8c6 d2d4 d7d5 d1d3 c6b4 d3b5 b4c6
2237667<1:info depth 9 seldepth 21 time 562 nodes 151298 nps 0
2237669<1:info depth 10

You will find no other UCI engine with the same similaries: 3 lines/depth, first line only with "info depth", same order of the infos. This looks _very_ strange to me. Look on the output of other UCI engines and you will see what I mean.

In Rybka the info seldepth is missing, info nps is added but there where such minor changes in most other Fruit clones.

Rybka starts the output at depth 3, but the PV line looks like the depth is incrased by 2, from depth 4 by 3.

From depth 1-5 (4-8) we have the same PV in the startposition, not in other positions:

Rybka:
2275926<1:info depth 8 score cp 7 time 298 nodes 14615 nps 50220 pv b1c3 b8c6 g1f3 g8f6 d2d4

Fruit
2237157<1:info depth 5 seldepth 9 score cp 48 time 52 nodes 1729 pv b1c3 b8c6 g1f3 g8f6 d2d4



From depth 10 both, Rybka and Fruit show the currmove infos:

Rybka:

3620642<1:info depth 10
3621603<1:info depth 10 score cp 3 time 2085 nodes 93647 nps 45992 pv b1c3 g8f6 d2d4 d7d5 g1f3 b8c6 d1d3
3621618<1:info currmove b1a3 currmovenumber 2
3621651<1:info currmove d2d3 currmovenumber 3
3621691<1:info currmove d2d4 currmovenumber 4
3621791<1:info currmove b2b3 currmovenumber 5
3621820<1:info currmove a2a3 currmovenumber 6
3621844<1:info currmove a2a4 currmovenumber 7
3621875<1:info currmove b2b4 currmovenumber 8
3621972<1:info currmove c2c4 currmovenumber 9
3621995<1:info currmove c2c3 currmovenumber 10
3622009<1:info currmove e2e4 currmovenumber 11
3622075<1:info currmove e2e3 currmovenumber 12
3622131<1:info currmove f2f4 currmovenumber 13
3622146<1:info currmove f2f3 currmovenumber 14
3622150<1:info currmove g2g4 currmovenumber 15
3622195<1:info currmove g2g3 currmovenumber 16
3622227<1:info currmove h2h4 currmovenumber 17
3622259<1:info currmove h2h3 currmovenumber 18
3622291<1:info currmove g1f3 currmovenumber 19
3622296<1:info currmove g1h3 currmovenumber 20
3622326<1:info depth 10 time 2817 nodes 128456 nps 46694
3622335<1:info depth 11

Fruit:

4837296<1:info depth 10
4837771<1:info time 1017 nodes 310000 nps 304818 cpuload 973
4837784<1:info hashfull 14
4837845<1:info depth 10 seldepth 22 score cp 12 time 1087 nodes 330356 pv b1c3 d7d5 d2d4 c8f5 g1f3 g8f6 f3h4 f5g4 h2h3 g4d7
4837851<1:info currmove g1f3 currmovenumber 2
4838032<1:info currmove g1f3 currmovenumber 2
4838156<1:info depth 10 seldepth 22 score cp 15 time 1391 nodes 427686 pv g1f3 b8c6 d2d4 d7d5 b1c3 g8f6 d1d3 g7g6 c1f4 c8f5
4838166<1:info currmove d2d4 currmovenumber 3
4838210<1:info currmove e2e4 currmovenumber 4
4838266<1:info currmove d2d3 currmovenumber 5
4838276<1:info currmove e2e3 currmovenumber 6
4838295<1:info currmove b1a3 currmovenumber 7
4838314<1:info currmove g1h3 currmovenumber 8
4838317<1:info currmove b2b3 currmovenumber 9
4838336<1:info currmove b2b4 currmovenumber 10
4838356<1:info currmove g2g3 currmovenumber 11
4838374<1:info currmove g2g4 currmovenumber 12
4838377<1:info currmove a2a4 currmovenumber 13
4838396<1:info currmove h2h4 currmovenumber 14
4838399<1:info currmove c2c4 currmovenumber 15
4838401<1:info currmove f2f4 currmovenumber 16
4838420<1:info currmove c2c3 currmovenumber 17
4838423<1:info currmove a2a3 currmovenumber 18
4838431<1:info currmove h2h3 currmovenumber 19
4838446<1:info currmove f2f3 currmovenumber 20
4838449<1:info depth 10 seldepth 24 time 1692 nodes 504717 nps 298296
4838451<1:info depth 11

There are some additional lines in Fruit, if you remove them everything looks similar.

For me all this looks like there is something taken from Fruit code, but it is also possible that Fruit was the example on how to do the engineoutput.

I could not find anything else, in my movegenerator testsuite Rybka plays always different moves, the errormessages of Fruit are not in Rybka (But Rybka crashes at the points where I get an errormessage of Fruit).

[Nid Hogge]

Wow I just came back after a long absence and all that mess..

So anyway it looks like he affirmed what many(including myself) have thought for a long time.

That he somehow reverse engineered parts of Rybka and combined those ideas into his private engine (Or that the based his code around this idea).

Now, What makes less sense :

1. Successfully reverse engineer Rybka's secrets and plant them in your chess engine .

2. Admit to the whole wide world you did just that.

I'm having hard time comprehending all this.

Dazed and Confused,

[Guetti]

http://kasparovchess.crestbook.com/view ... 578#p93578

Hi and thanks for the information.

This is a clear violation against the GPL.
As a member of the FSF, i will send a message to the FSF to verfiy possible measures in this case.

Best,
Daniel

Well, the similarity to Fruit can not be very high anymore, else Dann Corbit etc. would have recognised it.
I'm more concernd about the evaluation and tables he used from Fruit or Rybka, which took hundereds of hours of testing to optimize in the first place.