New engine releases & news H1 2023

[Dariusz]

Berserk 11 for Linux arm64, Linux x64, Macs, etc. binaries to download from:
https://www.chessengeria.com/files

Tomorrow it will also be downloadable from the home page of my blog and from the page dedicated to this engine.

Jay, many thanks for such a universal source code that allows compilation for different systems.

[Rebel]

http://rebel13.nl/windows/rebel-16.html

Update REBEL 16.2

1. Bugfix version time control and playing with multiple threads.
2. No engine or network improvements, some elo might come from the time control changes.

3. New UCI options.

• Polyglot book support. Default setting is off controlled by the UCI parameter Bookdepth=0 Changing the value (for instance) to 5 Rebel will limit the book moves till 5 moves. Occasionally this might be handy to avoid long book move lines that end up in positions where there is noting to play for any longer. For full use of the book set Bookdepth=100.

The book is created by Jeroen Noomen, book creator since Rebell 5.0 released 1986. Many voluminous experimental books are available from the Rebel website. [ 1 ] [ 2 ]

• Evalcorrect - This option can be used to change the playing style of the engine. The default value is 202. Changing the value to 256 (the maximum) will increase the engine aggressiveness, a value of 1 (the minimum) will change the playing style to boring. In general values between 128 and 202 are good without much loss in elo.

[peter]

Berserk 11 for Linux arm64, Linux x64, Macs, etc. binaries to download from:
https://www.chessengeria.com/files

Tomorrow it will also be downloadable from the home page of my blog and from the page dedicated to this engine.

Jay, many thanks for such a universal source code that allows compilation for different systems.

Berserk 11 for Linux arm64, Linux x64, Macs, etc. binaries to download from:
https://www.chessengeria.com/files

Tomorrow it will also be downloadable from the home page of my blog and from the page dedicated to this engine.

Jay, many thanks for such a universal source code that allows compilation for different systems.

Thanks for the compiles and the info!
Gauntlet between the latest version you gave (230130) and 11.0 is running.
To get out of error bar asap VSTC (15" for whole game +0.25" incr.), single thread, ponder off, 15 concurrencies of 16x3.5GHz CPU, axv2pext- compile from github for vs11.0 against your latest avx2- one of 10.x, Pohl's UHO- set 6mvs 110-119.

Here's a preliminary result:

Score of berserk-11-x64-avx2-pext vs Berserk-20230130b_Windows_x86-64-avx2: 51 - 41 - 170 [0.519]
... White vs Black: 91 - 1 - 170 [0.672] 262
Elo difference: 13.3 +/- 24.9, LOS: 85.1 %, DrawRatio: 64.9 %
262 of 500 games finished.

If 500 games should still not be enough to get out of error bar with Elo- difference, I'll stop it and have two more runs against SF dev. of 200 games each with incidentally same conditions as above.

Edit: 500 already finished:
Score of berserk-11-x64-avx2-pext vs Berserk-20230130b_Windows_x86-64-avx2: 93 - 86 - 321 [0.507]
... White vs Black: 175 - 4 - 321 [0.671] 500
Elo difference: 4.9 +/- 18.2, LOS: 70.0 %, DrawRatio: 64.2 %
500 of 500 games finished.

As expected, much too little LOS still, probably not possible to get out of error bar <3000 games, so back to plan B (two runs against SF dev. instead, should have gone that way from very start to spare hardware- time), regards

[peter]

As expected, much too little LOS still, probably not possible to get out of error bar <3000 games, so back to plan B (two runs against SF dev. instead, should have gone that way from very start to spare hardware- time), regards

Yet had rather a round robin between the three engines, SF, beserk 10 and berserk 11, conditions as described above:

Code: Select all

Rank Name                                      Elo     +/-   Games   Score    Draw 
   1 stockfish_23020319_x64_avx2               151      23     409   70.4%   50.4% 
   2 berserk-11-x64-avx2-pext                  -62      23     410   41.2%   52.7% 
   3 Berserk-20230130b_Windows_x86-64-avx2     -82      23     411   38.4%   53.0% 

And yet the gauntlet too for comparison of differences between amounts of games, Elo- discrimination and error bar:

Code: Select all

Rank Name                                      Elo     +/-   Games   Score    Draw 
   0 stockfish_23020319_x64_avx2               155      18     631   70.9%   52.1% 
   1 Berserk-20230130b_Windows_x86-64-avx2    -152      26     315   29.4%   51.7% 
   2 berserk-11-x64-avx2-pext                 -157      25     316   28.8%   52.5% 

646 of 1000 games finished.

And STS- MEA with Mosca's SF15- evaluation set:

Code: Select all

    EPD  : LAN_vs5.epd
    Time : ms
                                                Max     Total   Time    Hash          
    Engine           Score   Found  Pos   ELO   Score   Rate     ms      Mb  Cpu      
 1  SF230203         144406   1193  1500  4342  149703  96.5%    200     8    8
 2  lc0221229-325M   144040   1189  1500  4329  149703  96.2%    500     2    2
 3  Dragon3.2        143310   1180  1500  4306  149703  95.7%    200     8    8
 4  Berserk11        139297   1094  1500  4185  149703  93.0%    200     8    8
 5  Berserk230130    138822   1081  1500  4171  149703  92.7%    200     8    8
 6  Berserk10        138643   1081  1500  4167  149703  92.6%    200     8    8

                                    Created with MEA
                                          by
                                       Ferdinand
                                         Mosca

[Dariusz]

Berserk 11 for Linux arm64, Linux x64, Macs, etc. binaries to download from:
https://www.chessengeria.com/files

Tomorrow it will also be downloadable from the home page of my blog and from the page dedicated to this engine.

Jay, many thanks for such a universal source code that allows compilation for different systems.

Berserk 11 for Linux arm64, Linux x64, Macs, etc. binaries to download from:
https://www.chessengeria.com/files

Tomorrow it will also be downloadable from the home page of my blog and from the page dedicated to this engine.

Jay, many thanks for such a universal source code that allows compilation for different systems.

Thanks for the compiles and the info!
Gauntlet between the latest version you gave (230130) and 11.0 is running.
To get out of error bar asap VSTC (15" for whole game +0.25" incr.), single thread, ponder off, 15 concurrencies of 16x3.5GHz CPU, axv2pext- compile from github for vs11.0 against your latest avx2- one of 10.x, Pohl's UHO- set 6mvs 110-119.

Here's a preliminary result:

Score of berserk-11-x64-avx2-pext vs Berserk-20230130b_Windows_x86-64-avx2: 51 - 41 - 170 [0.519]
... White vs Black: 91 - 1 - 170 [0.672] 262
Elo difference: 13.3 +/- 24.9, LOS: 85.1 %, DrawRatio: 64.9 %
262 of 500 games finished.

If 500 games should still not be enough to get out of error bar with Elo- difference, I'll stop it and have two more runs against SF dev. of 200 games each with incidentally same conditions as above.

Edit: 500 already finished:
Score of berserk-11-x64-avx2-pext vs Berserk-20230130b_Windows_x86-64-avx2: 93 - 86 - 321 [0.507]
... White vs Black: 175 - 4 - 321 [0.671] 500
Elo difference: 4.9 +/- 18.2, LOS: 70.0 %, DrawRatio: 64.2 %
500 of 500 games finished.

As expected, much too little LOS still, probably not possible to get out of error bar <3000 games, so back to plan B (two runs against SF dev. instead, should have gone that way from very start to spare hardware- time), regards

Interesting results. Berserk 11 is also currently playing on my computer. I'm going to test it on 1000 games with a short time control 1m+0.6s move. Probably the results will be out this afternoon.

[peter]

http://rebel13.nl/windows/rebel-16.html

Update REBEL 16.2

1. Bugfix version time control and playing with multiple threads.
2. No engine or network improvements, some elo might come from the time control changes.

3. New UCI options.

• Polyglot book support. Default setting is off controlled by the UCI parameter Bookdepth=0 Changing the value (for instance) to 5 Rebel will limit the book moves till 5 moves. Occasionally this might be handy to avoid long book move lines that end up in positions where there is noting to play for any longer. For full use of the book set Bookdepth=100.

The book is created by Jeroen Noomen, book creator since Rebell 5.0 released 1986. Many voluminous experimental books are available from the Rebel website. [ 1 ] [ 2 ]

• Evalcorrect - This option can be used to change the playing style of the engine. The default value is 202. Changing the value to 256 (the maximum) will increase the engine aggressiveness, a value of 1 (the minimum) will change the playing style to boring. In general values between 128 and 202 are good without much loss in elo.

Very fine!
Even with VSTC I didn't see any time losses here so far anymore (haven't checked all 200).

15"+0.25 single thread 3.5GHz, 16Mb hash, ponder off, Pohl's UHO 6mvs 110-119:

Score of Ethereal-14.00-pext-avx2 vs Rebel-16.2: 83 - 29 - 88 [0.635]
... Ethereal-14.00-pext-avx2 playing White: 72 - 1 - 27 [0.855] 100
... Ethereal-14.00-pext-avx2 playing Black: 11 - 28 - 61 [0.415] 100
... White vs Black: 100 - 12 - 88 [0.720] 200
Elo difference: 96.2 +/- 36.4, LOS: 100.0 %, DrawRatio: 44.0 %
200 of 200 games finished.

Ferdy's version of the STS- set:

Code: Select all

    EPD  : LAN_vs5.epd
    Time : 200ms
                                                   Max   Total    Time   Hash          
     Engine           Score   Found  Pos   ELO   Score   Rate     ms     Mb  Cpu      
  1  SF230203         144406   1193  1500  4342  149703  96.5%    200     8    8
  2  lc0221229-325M   144040   1189  1500  4329  149703  96.2%    500     2    2
  3  Dragon3.2        143310   1180  1500  4306  149703  95.7%    200     8    8
  4  Berserk11        139297   1094  1500  4185  149703  93.0%    200     8    8
  5  Kovisto9.2       139203   1089  1500  4185  149703  93.0%    200     8    8
  6  Berserk230130    138822   1081  1500  4171  149703  92.7%    200     8    8
  7  Berserk10        138643   1081  1500  4167  149703  92.6%    200     8    8
  8  Ethereal14       138604   1087  1500  4167  149703  92.6%    200     8    8
  9  Rebel16.2        137776   1053  1500  4140  149703  92.0%    200     8    8
 10  Rebel16          137656   1070  1500  4140  149703  92.0%    200     8    8

                                    Created with MEA
                                          by
                                       Ferdinand
                                         Mosca

[peter]

15"+0.25 single thread 3.5GHz, 16Mb hash, ponder off, Pohl's UHO 6mvs 110-119:

Score of Ethereal-14.00-pext-avx2 vs Rebel-16.2: 83 - 29 - 88 [0.635]
... Ethereal-14.00-pext-avx2 playing White: 72 - 1 - 27 [0.855] 100
... Ethereal-14.00-pext-avx2 playing Black: 11 - 28 - 61 [0.415] 100
... White vs Black: 100 - 12 - 88 [0.720] 200
Elo difference: 96.2 +/- 36.4, LOS: 100.0 %, DrawRatio: 44.0 %
200 of 200 games finished.

As for Evalcorrect indeed maximum (256) instead of default 202 above didn't do much harm to the result of the same match as above:

Score of Ethereal-14.00-pext-avx2 vs Rebel-16.2: 80 - 21 - 99 [0.647]
... Ethereal-14.00-pext-avx2 playing White: 69 - 4 - 27 [0.825] 100
... Ethereal-14.00-pext-avx2 playing Black: 11 - 17 - 72 [0.470] 100
... White vs Black: 86 - 15 - 99 [0.677] 200
Elo difference: 105.6 +/- 34.2, LOS: 100.0 %, DrawRatio: 49.5 %
200 of 200 games finished.

Letting the one and the other one setting run these 128 hard tactical positions

https://www.dropbox.com/s/804b7chwli13laf/1284.epd?dl=0

Evalcorrect 256 helps significantly with 15"/position and 30 threads of the 16x3.5GHz CPU, Rebel 16.2-ECMax at nr. 37, default at end of list:

Code: Select all

    Program                                    Elo   +/-  Matches  Score   Av.Op.   S.Pos.   MST1    MST2   RIndex
 .....      
 25 Berserk11-MV4                            : 3483    7   2961    47.1 %   3504    48/128    4.5s   11.1s   0.32
 26 Berserk20230130b-MV4                     : 3483    7   3041    47.2 %   3503    50/128    4.8s   11.0s   0.29
 27 Dragon3.1byKomodoChess-MV4               : 3477    7   3000    46.3 %   3502    47/128    4.7s   11.2s   0.28
 28 SlowChessBlitz2.9avx2-MV4                : 3459    7   2894    43.5 %   3504    43/128    5.5s   11.8s   0.23
 29 Ethereal14.00(NNUE)-MV4                  : 3451    7   2828    42.1 %   3506    40/128    5.5s   12.0s   0.22
 30 Dragon3.1byKomodoChess                   : 3450    7   2810    42.1 %   3506    39/128    5.0s   12.0s   0.26
 31 Minic3.30-forceNNUE-MV4                  : 3447    7   2820    41.6 %   3506    40/128    5.7s   12.1s   0.21
 32 Dragon3.1byKomodoChess-MCTS-MV30         : 3445    7   2843    41.2 %   3507    37/128    4.0s   11.8s   0.13
 33 Ethereal13.75(NNUE)-AnalysisMode-MV4     : 3438    7   2722    40.0 %   3508    34/128    4.3s   12.2s   0.22
 34 Berserk20220725-MV4                      : 3433    7   2688    39.4 %   3508    35/128    4.8s   12.2s   0.20
 35 RubiChess20220813(bmi2)-MV4              : 3423    8   2702    37.9 %   3509    29/128    3.4s   12.4s   0.19
 36 PowerFritz18AVX20-MV4                    : 3406    7   2611    35.4 %   3510    30/128    6.1s   12.9s   0.14
 37 Rebel-16.2-ECMax                         : 3386    8   2560    32.5 %   3513    21/128    4.6s   13.3s   0.13
 38 Koivisto9.2                              : 3380    8   2499    31.6 %   3514    21/128    5.0s   13.4s   0.12
 39 Revenge3.0b-MV4                          : 3376    7   2708    31.7 %   3510    30/128   12.4s   14.4s   0.03
 40 Koivisto8.16                             : 3373    8   2449    30.7 %   3514    19/128    3.5s   13.3s   0.13
 41 Koivisto9.0                              : 3373    8   2456    30.7 %   3515    18/128    3.4s   13.4s   0.14
 42 Rebel-16.2                               : 3369    8   2487    30.3 %   3514    18/128    4.0s   13.4s   0.11



MST1  : Mean solution time (solved positions only)
MST2  : Mean solution time (solved and unsolved positions)
RIndex: Score according to solution time ranking for each position

MV4 means MultiPV=4, regards

[jhonnold]

Please note to testers using Berserk with a cyclical TC - there seems to be a serious time management bug resulting in random moves. I’m looking into it and will release a bug fix version.

[Wolfgang]

15"+0.25 single thread 3.5GHz, 16Mb hash, ponder off, Pohl's UHO 6mvs 110-119:

Score of Ethereal-14.00-pext-avx2 vs Rebel-16.2: 83 - 29 - 88 [0.635]
... Ethereal-14.00-pext-avx2 playing White: 72 - 1 - 27 [0.855] 100
... Ethereal-14.00-pext-avx2 playing Black: 11 - 28 - 61 [0.415] 100
... White vs Black: 100 - 12 - 88 [0.720] 200
Elo difference: 96.2 +/- 36.4, LOS: 100.0 %, DrawRatio: 44.0 %
200 of 200 games finished.

As for Evalcorrect indeed maximum (256) instead of default 202 above didn't do much harm to the result of the same match as above:

Score of Ethereal-14.00-pext-avx2 vs Rebel-16.2: 80 - 21 - 99 [0.647]
... Ethereal-14.00-pext-avx2 playing White: 69 - 4 - 27 [0.825] 100
... Ethereal-14.00-pext-avx2 playing Black: 11 - 17 - 72 [0.470] 100
... White vs Black: 86 - 15 - 99 [0.677] 200
Elo difference: 105.6 +/- 34.2, LOS: 100.0 %, DrawRatio: 49.5 %
200 of 200 games finished.

Letting the one and the other one setting run these 128 hard tactical positions

https://www.dropbox.com/s/804b7chwli13laf/1284.epd?dl=0

Evalcorrect 256 helps significantly with 15"/position and 30 threads of the 16x3.5GHz CPU, Rebel 16.2-ECMax at nr. 37, default at end of list:

Code: Select all

    Program                                    Elo   +/-  Matches  Score   Av.Op.   S.Pos.   MST1    MST2   RIndex
 .....      
 25 Berserk11-MV4                            : 3483    7   2961    47.1 %   3504    48/128    4.5s   11.1s   0.32
 26 Berserk20230130b-MV4                     : 3483    7   3041    47.2 %   3503    50/128    4.8s   11.0s   0.29
 27 Dragon3.1byKomodoChess-MV4               : 3477    7   3000    46.3 %   3502    47/128    4.7s   11.2s   0.28
 28 SlowChessBlitz2.9avx2-MV4                : 3459    7   2894    43.5 %   3504    43/128    5.5s   11.8s   0.23
 29 Ethereal14.00(NNUE)-MV4                  : 3451    7   2828    42.1 %   3506    40/128    5.5s   12.0s   0.22
 30 Dragon3.1byKomodoChess                   : 3450    7   2810    42.1 %   3506    39/128    5.0s   12.0s   0.26
 31 Minic3.30-forceNNUE-MV4                  : 3447    7   2820    41.6 %   3506    40/128    5.7s   12.1s   0.21
 32 Dragon3.1byKomodoChess-MCTS-MV30         : 3445    7   2843    41.2 %   3507    37/128    4.0s   11.8s   0.13
 33 Ethereal13.75(NNUE)-AnalysisMode-MV4     : 3438    7   2722    40.0 %   3508    34/128    4.3s   12.2s   0.22
 34 Berserk20220725-MV4                      : 3433    7   2688    39.4 %   3508    35/128    4.8s   12.2s   0.20
 35 RubiChess20220813(bmi2)-MV4              : 3423    8   2702    37.9 %   3509    29/128    3.4s   12.4s   0.19
 36 PowerFritz18AVX20-MV4                    : 3406    7   2611    35.4 %   3510    30/128    6.1s   12.9s   0.14
 37 Rebel-16.2-ECMax                         : 3386    8   2560    32.5 %   3513    21/128    4.6s   13.3s   0.13
 38 Koivisto9.2                              : 3380    8   2499    31.6 %   3514    21/128    5.0s   13.4s   0.12
 39 Revenge3.0b-MV4                          : 3376    7   2708    31.7 %   3510    30/128   12.4s   14.4s   0.03
 40 Koivisto8.16                             : 3373    8   2449    30.7 %   3514    19/128    3.5s   13.3s   0.13
 41 Koivisto9.0                              : 3373    8   2456    30.7 %   3515    18/128    3.4s   13.4s   0.14
 42 Rebel-16.2                               : 3369    8   2487    30.3 %   3514    18/128    4.0s   13.4s   0.11



MST1  : Mean solution time (solved positions only)
MST2  : Mean solution time (solved and unsolved positions)
RIndex: Score according to solution time ranking for each position

MV4 means MultiPV=4, regards

This thread is for announcement of new engines/versions only. So PLEASE stop discussing, open a new thread if you want to do so

[jhonnold]

My sincere apologies for all testers who potentially wasted CPU hours on Berserk 11.

Berserk 11.1 with a major bug-fix for anyone testing without an increment or in cyclical TCs.

https://github.com/jhonnold/berserk/releases/tag/11.1