Dragon 3.3 Released at KomodoChess.com

[lkaufman]

KomodoChess.com has released Dragon 3.3, expected to be the final version of Dragon, primarily to fulfill obligations to subscribers who were still due for another upgrade. It is free for anyone who bought Dragon 3.2 or who received Dragon 3.2 via subscription. New purchasers of Dragon should receive 3.3 from now one.

Although Mark and I retired from running KomodoChess at the start of this year, turning it over to chess.com, we have continued to improve its search on a part-time basis, while Dietriche Kappe was able to upgrade the net to a 50% larger size with increased strength, still trained on Dragon moves. In addition, the Elo levels should be much more accurate now (aimed at 10' + 5" Rapid chess), due to a new methodology and also to correcting an issue that made the low Elo levels weaker than intended. We welcome feedback on the accuracy of the new levels for human opponents.

Dragon 3.3 is a normal size upgrade; with "UHO" openings we got about 12 elo plus (against Dragon 3.2) in bullet and 15 elo plus in our blitz testing, with the increased gains with longer time controls presumably due to the larger net. With normal openings, we got a +8 elo result in 40,000 single-thread games at 1' + 0.6", and have confirmed this with a +8 elo result after 11400 games at 5' + 2". Gains on 4 or 8 threads appear to be a bit better, at least not worse. MCTS and FRC play should also be a bit improved, although these were not specifically targeted since version 3.2. Considering how difficult it is to make further Elo progress from the current lofty levels, we feel that these gains fully justify this final release.

There may be a short technical delay in getting copies to testers. We hope it will be fully tested as the final Dragon version.

[peter]

Thanks for the new version!

Single thread (3.5GHz) 3'+1", UHO 6mvs 110-119cp:

Score of dragon-3.3-64bit-avx2 vs dragon-3.2-64bit-avx2: 108 - 91 - 301 [0.517]
Elo difference: 11.8 +/- 19.2, LOS: 88.6 %, DrawRatio: 60.2 %
500 of 500 games finished,
regards

[dkappe]

Try that match with 8 threads now.

[pohl4711]

Thanks for the new version!

Single thread (3.5GHz) 3'+1", UHO 6mvs 110-119cp:

Score of dragon-3.3-64bit-avx2 vs dragon-3.2-64bit-avx2: 108 - 91 - 301 [0.517]
Elo difference: 11.8 +/- 19.2, LOS: 88.6 %, DrawRatio: 60.2 %
500 of 500 games finished,
regards

Try the next UHO stage: UHO 6mvs 120-129cp. I use this one for my UHO-Top15 Ratinglist. The draw-ratio should be lower and still not too many 1:1 gamepairs...

[Fritz 0]

KomodoChess.com has released Dragon 3.3, expected to be the final version of Dragon, primarily to fulfill obligations to subscribers who were still due for another upgrade. It is free for anyone who bought Dragon 3.2 or who received Dragon 3.2 via subscription. New purchasers of Dragon should receive 3.3 from now one.

Although Mark and I retired from running KomodoChess at the start of this year, turning it over to chess.com, we have continued to improve its search on a part-time basis, while Dietriche Kappe was able to upgrade the net to a 50% larger size with increased strength, still trained on Dragon moves. In addition, the Elo levels should be much more accurate now (aimed at 10' + 5" Rapid chess), due to a new methodology and also to correcting an issue that made the low Elo levels weaker than intended. We welcome feedback on the accuracy of the new levels for human opponents.

Dragon 3.3 is a normal size upgrade; with "UHO" openings we got about 12 elo plus (against Dragon 3.2) in bullet and 15 elo plus in our blitz testing, with the increased gains with longer time controls presumably due to the larger net. With normal openings, we got a +8 elo result in 40,000 single-thread games at 1' + 0.6", and have confirmed this with a +8 elo result after 11400 games at 5' + 2". Gains on 4 or 8 threads appear to be a bit better, at least not worse. MCTS and FRC play should also be a bit improved, although these were not specifically targeted since version 3.2. Considering how difficult it is to make further Elo progress from the current lofty levels, we feel that these gains fully justify this final release.

There may be a short technical delay in getting copies to testers. We hope it will be fully tested as the final Dragon version.

Larry, a few questions:

Why did you switch from 15'+10'' to 10'+5'' in the Elo levels?

Can you tell us more about that new methodology? Is it still based on node count and randomness?

Does Dragon 3.3 still has Regular Eval option? (just asking this because Stockfish removed it)

[peter]

Try that match with 8 threads now.

Too much hardware- time at the moment for me, it (higher hardware- TC) would raise the error bar even more anyhow and as for Stefan's proposal of next higher unbalanced level of UHOs, I don't think that way I'd get out of it (error bar) with less then 1500 games neither, at least not without using even lower hardware-TC. So I'll rather wait for his results, just wanted to get a first impression of my own special interest, regards

[lkaufman]

Thanks for the new version!

Single thread (3.5GHz) 3'+1", UHO 6mvs 110-119cp:

Score of dragon-3.3-64bit-avx2 vs dragon-3.2-64bit-avx2: 108 - 91 - 301 [0.517]
Elo difference: 11.8 +/- 19.2, LOS: 88.6 %, DrawRatio: 60.2 %
500 of 500 games finished,
regards

Try the next UHO stage: UHO 6mvs 120-129cp. I use this one for my UHO-Top15 Ratinglist. The draw-ratio should be lower and still not too many 1:1 gamepairs...

60% draws is already pretty good, no great need to reduce it towards 50%. Personally I think that the optimum range to use is somewhat lower than whatever range produces the largest elo-spreading, because there is also the consideration that the more unbalanced the initial position is, the less it resembles normal chess. So there is a tradeoff here, both factors should be considered. We still use the 110-119 range in our testing.

[lkaufman]

KomodoChess.com has released Dragon 3.3, expected to be the final version of Dragon, primarily to fulfill obligations to subscribers who were still due for another upgrade. It is free for anyone who bought Dragon 3.2 or who received Dragon 3.2 via subscription. New purchasers of Dragon should receive 3.3 from now one.

Although Mark and I retired from running KomodoChess at the start of this year, turning it over to chess.com, we have continued to improve its search on a part-time basis, while Dietriche Kappe was able to upgrade the net to a 50% larger size with increased strength, still trained on Dragon moves. In addition, the Elo levels should be much more accurate now (aimed at 10' + 5" Rapid chess), due to a new methodology and also to correcting an issue that made the low Elo levels weaker than intended. We welcome feedback on the accuracy of the new levels for human opponents.

Dragon 3.3 is a normal size upgrade; with "UHO" openings we got about 12 elo plus (against Dragon 3.2) in bullet and 15 elo plus in our blitz testing, with the increased gains with longer time controls presumably due to the larger net. With normal openings, we got a +8 elo result in 40,000 single-thread games at 1' + 0.6", and have confirmed this with a +8 elo result after 11400 games at 5' + 2". Gains on 4 or 8 threads appear to be a bit better, at least not worse. MCTS and FRC play should also be a bit improved, although these were not specifically targeted since version 3.2. Considering how difficult it is to make further Elo progress from the current lofty levels, we feel that these gains fully justify this final release.

There may be a short technical delay in getting copies to testers. We hope it will be fully tested as the final Dragon version.

Larry, a few questions:

Why did you switch from 15'+10'' to 10'+5'' in the Elo levels?

Can you tell us more about that new methodology? Is it still based on node count and randomness?

Does Dragon 3.3 still has Regular Eval option? (just asking this because Stockfish removed it)

The switch in target time control was due to the fact that most online Rapid play is now at just 10' with no increment, so this is a compromise, and to the fact that the top level online events are now 15' + 3" instead of 15' + 10". Also it seems that most people who play against bots don't spend more than 15' per game, usually less (with the bot moving in one second or so).

The new methodology refers more to how we test rather than to how the levels are defined. However we did raise the minimum number of nodes substantially (with more randomness for the low levels to compensate), because the low elo levels in Dragon 3.2 and prior would often just repeat position even when up multiple queens, making test results nonsense. This is now fixed. For testing the levels were run against varying fixed depth levels of a chosen unrelated conventional engine with the relative ratings determined from those results, no direct level vs level testing which inflates the differences (compared to human results), with a few data points vs. humans at higher levels to set the overall level. This should be much more accurate.

Yes, we still have Regular Eval. Although Stockfish has reduced UCI options to a minimum, we have not gone that route.

[Fritz 0]

KomodoChess.com has released Dragon 3.3, expected to be the final version of Dragon, primarily to fulfill obligations to subscribers who were still due for another upgrade. It is free for anyone who bought Dragon 3.2 or who received Dragon 3.2 via subscription. New purchasers of Dragon should receive 3.3 from now one.

Although Mark and I retired from running KomodoChess at the start of this year, turning it over to chess.com, we have continued to improve its search on a part-time basis, while Dietriche Kappe was able to upgrade the net to a 50% larger size with increased strength, still trained on Dragon moves. In addition, the Elo levels should be much more accurate now (aimed at 10' + 5" Rapid chess), due to a new methodology and also to correcting an issue that made the low Elo levels weaker than intended. We welcome feedback on the accuracy of the new levels for human opponents.

Dragon 3.3 is a normal size upgrade; with "UHO" openings we got about 12 elo plus (against Dragon 3.2) in bullet and 15 elo plus in our blitz testing, with the increased gains with longer time controls presumably due to the larger net. With normal openings, we got a +8 elo result in 40,000 single-thread games at 1' + 0.6", and have confirmed this with a +8 elo result after 11400 games at 5' + 2". Gains on 4 or 8 threads appear to be a bit better, at least not worse. MCTS and FRC play should also be a bit improved, although these were not specifically targeted since version 3.2. Considering how difficult it is to make further Elo progress from the current lofty levels, we feel that these gains fully justify this final release.

There may be a short technical delay in getting copies to testers. We hope it will be fully tested as the final Dragon version.

Larry, a few questions:

Why did you switch from 15'+10'' to 10'+5'' in the Elo levels?

Can you tell us more about that new methodology? Is it still based on node count and randomness?

Does Dragon 3.3 still has Regular Eval option? (just asking this because Stockfish removed it)

The switch in target time control was due to the fact that most online Rapid play is now at just 10' with no increment, so this is a compromise, and to the fact that the top level online events are now 15' + 3" instead of 15' + 10". Also it seems that most people who play against bots don't spend more than 15' per game, usually less (with the bot moving in one second or so).

The new methodology refers more to how we test rather than to how the levels are defined. However we did raise the minimum number of nodes substantially (with more randomness for the low levels to compensate), because the low elo levels in Dragon 3.2 and prior would often just repeat position even when up multiple queens, making test results nonsense. This is now fixed. For testing the levels were run against varying fixed depth levels of a chosen unrelated conventional engine with the relative ratings determined from those results, no direct level vs level testing which inflates the differences (compared to human results), with a few data points vs. humans at higher levels to set the overall level. This should be much more accurate.

Yes, we still have Regular Eval. Although Stockfish has reduced UCI options to a minimum, we have not gone that route.

I have just run a 15000 games match between Dragon 3 and Dragon 3.3, both set to 2000 Elo. Dragon 3.3 won by 32 points. Taking into account that Dragon 3 Elo levels were aimed at 15'+10'', this is very strange. Is it possible that Dragon 3 Elo levles were so much weaker than they should be?

[lkaufman]

KomodoChess.com has released Dragon 3.3, expected to be the final version of Dragon, primarily to fulfill obligations to subscribers who were still due for another upgrade. It is free for anyone who bought Dragon 3.2 or who received Dragon 3.2 via subscription. New purchasers of Dragon should receive 3.3 from now one.

Although Mark and I retired from running KomodoChess at the start of this year, turning it over to chess.com, we have continued to improve its search on a part-time basis, while Dietriche Kappe was able to upgrade the net to a 50% larger size with increased strength, still trained on Dragon moves. In addition, the Elo levels should be much more accurate now (aimed at 10' + 5" Rapid chess), due to a new methodology and also to correcting an issue that made the low Elo levels weaker than intended. We welcome feedback on the accuracy of the new levels for human opponents.

Dragon 3.3 is a normal size upgrade; with "UHO" openings we got about 12 elo plus (against Dragon 3.2) in bullet and 15 elo plus in our blitz testing, with the increased gains with longer time controls presumably due to the larger net. With normal openings, we got a +8 elo result in 40,000 single-thread games at 1' + 0.6", and have confirmed this with a +8 elo result after 11400 games at 5' + 2". Gains on 4 or 8 threads appear to be a bit better, at least not worse. MCTS and FRC play should also be a bit improved, although these were not specifically targeted since version 3.2. Considering how difficult it is to make further Elo progress from the current lofty levels, we feel that these gains fully justify this final release.

There may be a short technical delay in getting copies to testers. We hope it will be fully tested as the final Dragon version.

Larry, a few questions:

Why did you switch from 15'+10'' to 10'+5'' in the Elo levels?

Can you tell us more about that new methodology? Is it still based on node count and randomness?

Does Dragon 3.3 still has Regular Eval option? (just asking this because Stockfish removed it)

The switch in target time control was due to the fact that most online Rapid play is now at just 10' with no increment, so this is a compromise, and to the fact that the top level online events are now 15' + 3" instead of 15' + 10". Also it seems that most people who play against bots don't spend more than 15' per game, usually less (with the bot moving in one second or so).

The new methodology refers more to how we test rather than to how the levels are defined. However we did raise the minimum number of nodes substantially (with more randomness for the low levels to compensate), because the low elo levels in Dragon 3.2 and prior would often just repeat position even when up multiple queens, making test results nonsense. This is now fixed. For testing the levels were run against varying fixed depth levels of a chosen unrelated conventional engine with the relative ratings determined from those results, no direct level vs level testing which inflates the differences (compared to human results), with a few data points vs. humans at higher levels to set the overall level. This should be much more accurate.

Yes, we still have Regular Eval. Although Stockfish has reduced UCI options to a minimum, we have not gone that route.

I have just run a 15000 games match between Dragon 3 and Dragon 3.3, both set to 2000 Elo. Dragon 3.3 won by 32 points. Taking into account that Dragon 3 Elo levels were aimed at 15'+10'', this is very strange. Is it possible that Dragon 3 Elo levles were so much weaker than they should be?

Yes, it is quite possible. The relative strength of the different elo settings should be much more accurate now, so even if say 2600 remains the same, other levels will be different. I would suggest that the chess.com Rapid ratings for humans make for a better comparison with Komodo levels than FIDE ratings for amateur players, because the lower FIDE ratings seem to vary widely depending on the age of the players and the nationality. A 1600 young FIDE player today from a young country will be much stronger than a middle aged 1600 FIDE player from an older country who earned his rating 15 years ago. Online Rapid ratings are a much more consistent basis for comparison, and on average there is little difference between chess.com rapid ratings and FIDE ratings.