Dragon 2.6 elo so far

[lkaufman]

Yes true. SF has won EVERY 16 white games vs Ethereal!

If in the future rating lists use opening books like this, we'll continue to see large elo gains even at long time limits for years to come. Ratings like 5000 will be possible. What that means is a matter of opinion; it's basically like removing the draw factor from chess. Imagine if 100 yard races were run with the rule that any result within 10% would be called a draw. All the best runners would be tied with each other almost every time. Then someone might suggest that they take turns running 110 yards vs 100 yards. Suddenly half the races would have a winner. Then a runner who was 1% faster than another would consistently win when he had the shorter distance to run and draw when he had the longer distance. He would score 75%, +192 elo! This is almost exactly what we are seeing here. We have found a solution to the draw problem in computer chess. Whether that makes it a different game than normal chess can be debated. The only other problem with this is that the choice of openings is arbitrary, dependent on some arbitrary rules and decisions. It would be nice to have opening chosen by some simple rules that don't involve an engine's judging the positions. FRC (960) is an example of this, but unfortunately those positions aren't near the win/draw line. Some similar rules that did produce positions near the line would be ideal.

I like it. Does this "make it a different game than normal chess..."? No - it's the same rules, piece placings, etc. It's those abominations (variants) that are truly different games.

I would like to see more 'black wins' in normal chess between engines though. Can they 'only' win with White? This means you let the engines play decidedly sub-optimal (but not outright stupid) White set ups with sometimes very short opening lines. OTB and online, I've backed into reversed Benoni's and Benko Gambits more times than I can count. I win a lot simply because I understand the structures better than many.

Well, obviously you can create openings where Black has the same advantage as White does in these events, just by wasting a move for White. More generally, White has to play more or worse bad moves to get to a nearly-losing position than Black does. It seems better to me to minimize the number of bad moves required, and also it is just much easier to keep track of results and statistics if we know that White is always the one starting with the advantage. White can even blunder a pawn and still be near the win/draw line (I think that 1.b4 e6 2.Nf3? Bxb4 is likely near the win/draw line for example), but this feels less like real chess than the lines we actually see in these events.

[Chessqueen]

Yes true. SF has won EVERY 16 white games vs Ethereal!

If in the future rating lists use opening books like this, we'll continue to see large elo gains even at long time limits for years to come. Ratings like 5000 will be possible. What that means is a matter of opinion; it's basically like removing the draw factor from chess. Imagine if 100 yard races were run with the rule that any result within 10% would be called a draw. All the best runners would be tied with each other almost every time. Then someone might suggest that they take turns running 110 yards vs 100 yards. Suddenly half the races would have a winner. Then a runner who was 1% faster than another would consistently win when he had the shorter distance to run and draw when he had the longer distance. He would score 75%, +192 elo! This is almost exactly what we are seeing here. We have found a solution to the draw problem in computer chess. Whether that makes it a different game than normal chess can be debated. The only other problem with this is that the choice of openings is arbitrary, dependent on some arbitrary rules and decisions. It would be nice to have opening chosen by some simple rules that don't involve an engine's judging the positions. FRC (960) is an example of this, but unfortunately those positions aren't near the win/draw line. Some similar rules that did produce positions near the line would be ideal.

I like it. Does this "make it a different game than normal chess..."? No - it's the same rules, piece placings, etc. It's those abominations (variants) that are truly different games.

I would like to see more 'black wins' in normal chess between engines though. Can they 'only' win with White? This means you let the engines play decidedly sub-optimal (but not outright stupid) White set ups with sometimes very short opening lines. OTB and online, I've backed into reversed Benoni's and Benko Gambits more times than I can count. I win a lot simply because I understand the structures better than many.

The only way you will see more Black wins in normal chess between engines that are almost equal in strength or within 50 Elo gap is to let the Black pieces play first, for 2 reasons, first because it has the first move and 2nd because there is no Opening to follows

[pgn][Event "Computer chess game"]
[Site "MININT-UB2PIMJ"]
[Date "2021.12.21"]
[Round "?"]
[White "Stockfish_14.1_win_x64_avx2"]
[Black "Dragon-2.5-64bit-avx2"]
[Result "0-1"]
[BlackElo "3600"]
[Time "16:41:22"]
[WhiteElo "2000"]
[TimeControl "180+1"]
[SetUp "1"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR b KQkq - 0 1"]
[Termination "adjudication"]
[PlyCount "106"]
[WhiteType "program"]
[BlackType "program"]

1. ... e5 2. e3 d5 3. d4 Nc6 4. Nf3 Bg4 5. dxe5 Nxe5 6. Be2 Bxf3 7. gxf3
Nf6 8. Nd2 Qd7 9. f4 Ng6 10. h4 h5 11. Nf3 O-O-O 12. a3 c5 13. b4 Qc7 14.
Qd3 Kb8 15. Bb2 c4 16. Qf5 c3 17. Bc1 Bd6 18. Nd4 Rhe8 19. Kf1 Qb6 20. Rh3
a6 21. Qd3 Be7 22. f3 Nd7 23. Qxc3 Bf6 24. Bb2 Re7 25. f5 Nxh4 26. Qe1 g5
27. f4 g4 28. Rxh4 Rxe3 29. Rd1 Bxh4 30. Qxh4 Rde8 31. Rd3 R3e4 32. Qf2 Qd6
33. Kg2 Qxf4 34. Rd1 Qxf2+ 35. Kxf2 h4 36. Bd3 Rf4+ 37. Kg2 Ne5 38. Bf1 Re4
39. Bc1 Rc8 40. Kh1 h3 41. a4 f6 42. Bb2 Nc4 43. Bc1 Ne3 44. Bxe3 Rxe3 45.
b5 axb5 46. axb5 Rh8 47. b6 g3 48. Kg1 Rh4 49. Ne6 Kc8 50. Nf8 g2 51. Bd3
Rb4 52. Ra1 Kd8 53. Kh2 Ke7 54. Ne6 {User Adjudication} 0-1[/pgn]

[carldaman]

Same version has 10,5 - 19,5 against SF dev in CCC?

I think it's because CCC uses unbalanced openings, which magnify elo differences. With more threads, the effect of unbalanced books becomes larger and larger, as with normal books the draw percentage starts to approach 100% with the best engines on many threads (or long time controls). With opening positions near the win/draw line, a slightly better engine might consistently win with White and draw with Black, scoring 75% = +192 elo.

Maybe it depends on how we actually classify 'unbalanced', but I thought unbalanced openings generally compressed the differences in Elo, since they potentially led to too many dreaded ' book wins', including for the weaker side, assuming there is one. Perhaps, since there is no clearly weaker side between SF and Dragon, that effect is not present.

[lkaufman]

Same version has 10,5 - 19,5 against SF dev in CCC?

I think it's because CCC uses unbalanced openings, which magnify elo differences. With more threads, the effect of unbalanced books becomes larger and larger, as with normal books the draw percentage starts to approach 100% with the best engines on many threads (or long time controls). With opening positions near the win/draw line, a slightly better engine might consistently win with White and draw with Black, scoring 75% = +192 elo.

Maybe it depends on how we actually classify 'unbalanced', but I thought unbalanced openings generally compressed the differences in Elo, since they potentially led to too many dreaded ' book wins', including for the weaker side, assuming there is one. Perhaps, since there is no clearly weaker side between SF and Dragon, that effect is not present.

If we see "book wins" it means that the openings were poorly chosen. I assume you mean that the book ends with one side winning with perfect play. A good unbalanced set ends with positions that aren't clearly winning or clearly drawing, but near the line, preferably on the drawing side of the line. Of course no one really knows exactly where that line falls in the opening, but we can estimate pretty well, ideally by having the same engine play it out hundreds of time (with some variety added perhaps just by using MP) at a level high enough to draw nearly every game with normal openings. If this isn't practical (due to the number of openings needed), we have to trust the eval of a strong engine. Stephen Pohl is the most noted author of such opening sets, and he relies on Komodo Dragon for the evals (can't fault his choice of engine!). He uses several methods to create imbalance, such as removal of a pawn, restricting castling rights, giving White an extra move in normal openings, or (my personal favorite) simply selecting openings from human play that give the desired White advantage. Typically he might use openings that result in something like 60% draws and 40% White wins with Dragon or Stockfish playing both sides at say CCRL blitz TC on four threads (for variety). Those are the sets that CCC uses often. This is pretty much optimal in terms of increasing elo differences while keeping the game somewhat normal-looking. If you only need a 100 positions or less, I would pick from "one-mistake" positions where Black makes one early bad move where human masters actually do make those moves occasionally, with the resultant position evaluated just slightly above plus one pawn by Dragon 2.6. Some examples: 1.e4 h6?; 1.e4 a5?, 1.e4 b6? (this one is a surprise, I would never guess it to be so bad), 1.e4 d5 2.exd5 c6?. All of these give results in the 70% ballpark for White, whether based on human master play or from engine playouts. Probably not winning for White, but very easy to lose as Black. At long time controls this is apt to triple the elo difference compared to a book representative of usual human GM play.