Hello Chess Friends,
As usual, I'm pleased to announce that,I've managed to organize another EvalFile competition!Note: the previous EvalFile tour can be viewed: ResultsWhere 1520.bin managed to perform with best performance!And since previous tour...several months have passed,So now let's see which NN files are the strongest...
Note that as Book Openings,I used a new Balsa suite, which is based on previous NNUE winsSure my main target was simply to reduce the draw percentageAccording to current testings: approx. 5% are reduced, not badAnyhow, the overall draws are still very high, in case of NNUEI mean in case of close in strength 3700 Elo + same EvalFileBtw, to be more clear, here the draw problem is not in openings,The real problem comes/related to Top NNUE chess engines...E.g via NNUE: minus plus 15% are higher...that's too much!Usually, we can see wins only with very critical openings...For example, the overall NNUE draw percentage is close to 95%That means in every 100 games, we can see minus plus 5 winsSure I am referring for Blitz, but with Slow time controls:I expect to see almost 100% draws, e.g with Slow TC 40/120:Maybe with very weak openings may appear a few wins...
In other words,I hope to see a real 'Hero' programmer,Who will manage to release a Top NNUE engine, which will produceMinus plus 70-75% draws, sure under Blitz, Rapid... conditions!And then there is no doubt that,3700 Elo NNUE competitions will be much more exciting...!
Just I'd like to point out that too,I've realized the EvalFiles to be played by two NNUE engines:Cf EXT 291120 and SugaR AI 1.80The reason about why I preferred older CF EXT engine is that,Cf EXT 291120 produces lower draw percentage than all Top...Also in this Evalfile tour,CF EXT 291120 produced approx. 3% less draws than SugaR AI 1.80Note: These stats are based on Top 10 Evalfiles (played each other..)But please bear in mind this too, in case of NON-NNUE: ResultsCf EXT 291120 produces close to 70% draws (e.g playing itself...)!That means approx. 20% less draws...unbelievable draw numbers!
Meanwhile, a new NNUE Duel (2m+1s): CF EXT vs SugaR:And as we see,CF EXT 291120's engine's strength is still not dead...!What does it mean ? simply no much Elo progress by the latestSF based chess engine releases...at least under these conditions!
1 CF EXT 291120 +6/-6/=214 50.00% 113.0/2262 SugaR AI 1.80 +6/-6/=214 50.00% 113.0/226
Some more notes about the above Duel match:Both engines played via same nn-62ef826d1a6d.nnueSugaR AI 1.80 engine is played with its default ContemptWhere CF EXT 291120 is used to play via Contempt 40Duel's overall draw percentage (based on 226 games) 95%Plus, I run another Draw test: CF EXT (playing itself)Which can be useful, for anyone who is interested:CF EXT 291120 C40 (via 62ef826d1a6d) produced: 93%These Draw value statistics are based on 366 gamesAnd,as I mentioned before,Via NN file: CF EXT's draws are approx. 20% higher...!
And now let's back to the current NNUE tour,As Evalfile players,I've picked mainly the latest NN releases (per author) NetBtw, I have no any idea why, but some NN files didn't work:FatFritz2_v1, 2edce6db11bb, 9ac267288b8f, b700127fe34120f7dbe4b33b, 57bacd2f041e, 9aaa2abe4015, b20e264c6f88Note: I mean via both: CF EXT 291120 and SugaR AI 1.80
However, I have good news,Mostly which I've tried to test... worked flawlessly...And here is the final 'EvalFile' ranking:
The Winner: nn-62ef826d1a6d.nnue - Congrats to SFisGOD!Note also that nn-62ef826d1a6d nnue released on 28.10.2020
# PLAYER : RATING POINTS GAMES (%) 1 62ef826d1a6d : 3692.7 313.0 576 54.3% 2 371e294c6117 : 3687.1 307.0 576 53.3% 3 66a8461e0dd6 : 3686.4 307.0 576 53.3% 4 cb26f10b1fd9 : 3685.1 305.5 576 53.0% 5 ae5925b37cc9 : 3681.4 302.5 576 52.5% 6 82f30097537a : 3680.5 301.5 576 52.3% 7 20200914-1520 : 3675.0 296.5 576 51.5% 8 30f6ddc579c2 : 3664.9 288.0 576 50.0% 9 9ed9e3a70a1b : 3660.2 283.5 576 49.2%10 1de20d6fd769 : 3659.4 282.5 576 49.0%11 011f4b2f4629 : 3601.8 163.5 396 41.3%12 c193f6be9438 : 3534.0 125.5 396 31.7%
Games
More Details:All EvalFiles are played by CF EXT 291120 and SugaR AI 1.80Due to parallel matches, so for maximum performance/strength:Several Arena Chess GUIs used, so for each Evalfile player isInstalled separate copy of engine, for more information: Details011f4b2f4629+c193f6be9438 are played less games, the reason:Tour main target is testing NN files close to 3700 Elo points!Current overall draw percentage (based on 3276 games) is 85%Here I think, Balsa (Wins) Opening Suite played a BIG role..!Plus another factor is that, some of the NN files are weaker...In case of same NN file...the draw numbers are going higher...And as a last note,Running these kind of testings are full with pain..not so easy...I mean a lot of details, attention, free time etc. are required....Even publishing the current online results take a lot of free time!But for a chess LOVER:All the efforts should not be counted as a BIG problem, right ?!)
And many thanks for your interest...!)
Best Wishess,Sedat Canbaz
As usual, I'm pleased to announce that,I've managed to organize another EvalFile competition!Note: the previous EvalFile tour can be viewed: ResultsWhere 1520.bin managed to perform with best performance!And since previous tour...several months have passed,So now let's see which NN files are the strongest...
Note that as Book Openings,I used a new Balsa suite, which is based on previous NNUE winsSure my main target was simply to reduce the draw percentageAccording to current testings: approx. 5% are reduced, not badAnyhow, the overall draws are still very high, in case of NNUEI mean in case of close in strength 3700 Elo + same EvalFileBtw, to be more clear, here the draw problem is not in openings,The real problem comes/related to Top NNUE chess engines...E.g via NNUE: minus plus 15% are higher...that's too much!Usually, we can see wins only with very critical openings...For example, the overall NNUE draw percentage is close to 95%That means in every 100 games, we can see minus plus 5 winsSure I am referring for Blitz, but with Slow time controls:I expect to see almost 100% draws, e.g with Slow TC 40/120:Maybe with very weak openings may appear a few wins...
In other words,I hope to see a real 'Hero' programmer,Who will manage to release a Top NNUE engine, which will produceMinus plus 70-75% draws, sure under Blitz, Rapid... conditions!And then there is no doubt that,3700 Elo NNUE competitions will be much more exciting...!
Just I'd like to point out that too,I've realized the EvalFiles to be played by two NNUE engines:Cf EXT 291120 and SugaR AI 1.80The reason about why I preferred older CF EXT engine is that,Cf EXT 291120 produces lower draw percentage than all Top...Also in this Evalfile tour,CF EXT 291120 produced approx. 3% less draws than SugaR AI 1.80Note: These stats are based on Top 10 Evalfiles (played each other..)But please bear in mind this too, in case of NON-NNUE: ResultsCf EXT 291120 produces close to 70% draws (e.g playing itself...)!That means approx. 20% less draws...unbelievable draw numbers!
Meanwhile, a new NNUE Duel (2m+1s): CF EXT vs SugaR:And as we see,CF EXT 291120's engine's strength is still not dead...!What does it mean ? simply no much Elo progress by the latestSF based chess engine releases...at least under these conditions!
1 CF EXT 291120 +6/-6/=214 50.00% 113.0/2262 SugaR AI 1.80 +6/-6/=214 50.00% 113.0/226
Some more notes about the above Duel match:Both engines played via same nn-62ef826d1a6d.nnueSugaR AI 1.80 engine is played with its default ContemptWhere CF EXT 291120 is used to play via Contempt 40Duel's overall draw percentage (based on 226 games) 95%Plus, I run another Draw test: CF EXT (playing itself)Which can be useful, for anyone who is interested:CF EXT 291120 C40 (via 62ef826d1a6d) produced: 93%These Draw value statistics are based on 366 gamesAnd,as I mentioned before,Via NN file: CF EXT's draws are approx. 20% higher...!
And now let's back to the current NNUE tour,As Evalfile players,I've picked mainly the latest NN releases (per author) NetBtw, I have no any idea why, but some NN files didn't work:FatFritz2_v1, 2edce6db11bb, 9ac267288b8f, b700127fe34120f7dbe4b33b, 57bacd2f041e, 9aaa2abe4015, b20e264c6f88Note: I mean via both: CF EXT 291120 and SugaR AI 1.80
However, I have good news,Mostly which I've tried to test... worked flawlessly...And here is the final 'EvalFile' ranking:
The Winner: nn-62ef826d1a6d.nnue - Congrats to SFisGOD!Note also that nn-62ef826d1a6d nnue released on 28.10.2020
# PLAYER : RATING POINTS GAMES (%) 1 62ef826d1a6d : 3692.7 313.0 576 54.3% 2 371e294c6117 : 3687.1 307.0 576 53.3% 3 66a8461e0dd6 : 3686.4 307.0 576 53.3% 4 cb26f10b1fd9 : 3685.1 305.5 576 53.0% 5 ae5925b37cc9 : 3681.4 302.5 576 52.5% 6 82f30097537a : 3680.5 301.5 576 52.3% 7 20200914-1520 : 3675.0 296.5 576 51.5% 8 30f6ddc579c2 : 3664.9 288.0 576 50.0% 9 9ed9e3a70a1b : 3660.2 283.5 576 49.2%10 1de20d6fd769 : 3659.4 282.5 576 49.0%11 011f4b2f4629 : 3601.8 163.5 396 41.3%12 c193f6be9438 : 3534.0 125.5 396 31.7%
Games
More Details:All EvalFiles are played by CF EXT 291120 and SugaR AI 1.80Due to parallel matches, so for maximum performance/strength:Several Arena Chess GUIs used, so for each Evalfile player isInstalled separate copy of engine, for more information: Details011f4b2f4629+c193f6be9438 are played less games, the reason:Tour main target is testing NN files close to 3700 Elo points!Current overall draw percentage (based on 3276 games) is 85%Here I think, Balsa (Wins) Opening Suite played a BIG role..!Plus another factor is that, some of the NN files are weaker...In case of same NN file...the draw numbers are going higher...And as a last note,Running these kind of testings are full with pain..not so easy...I mean a lot of details, attention, free time etc. are required....Even publishing the current online results take a lot of free time!But for a chess LOVER:All the efforts should not be counted as a BIG problem, right ?!)
And many thanks for your interest...!)
Best Wishess,Sedat Canbaz