New release Revenge 4.0

[Fabio Gobbato]

Happy new year!

I have released a new version of my engine. I have changed my work recently and I have less time for the development of the engine, so I cannot reach an high level of play and I have decided to release it for free. If you like the engine or you want to support the development there's the chance to donate what you want, of course there's no need and you are free.

Revenge 4.0 is 38 elo stronger than 3.0 tested in selfplay at 1:00+0.5 after 10000 games with balanced openings.

You can find further informations on the website: https://www.revengechess.com/

[matejst]

Thanks, Fabio, and happy new year!

[Dariusz]

Whoa, what a great gift for the beginning of 2025.

Congratulations on the latest version of Revenge and thank you beautifully for sharing

Fabio, all the best!

[Pvt. Ryan]

Thank you and happy new year.

[peter]

Happy New Year and many thanks for the new and free version!

Here I have some results so far.
3'+1" single thread with UHO 2024 900-990cp 6mvs:

Score of Revenge4 vs Revenge3: 154 - 71 - 275 [0.583]
Elo difference: 58.2 +/- 20.4, LOS: 100.0 %, DrawRatio: 55.0 %
500 of 500 games finished.

128 anti engine puzzles:



excerpted out of a list created with EloStatTS from Frank Schubert and 30"/pos. and 30 threads of a 16x3.5GHz CPU or 2 threads of it and a 3070ti GPU for Lc0

Code: Select all

    Program                                    Elo   +/-  Matches  Score   Av.Op.   S.Pos.   MST1    MST2   RIndex

  6 ShashChess34.6-MuPV4                     : 3554    3  10906    58.0 %   3498    99/128    4.0s    9.9s   0.65
  
 12 Crystaldev-20240503-MuPV4                : 3551    3  10797    57.6 %   3498   102/128    5.3s   10.3s   0.65
 
 
 47 Lc0v0.31.1-6147500PT-MuPV4               : 3512    3  10696    51.6 %   3501    87/128    5.9s   13.6s   0.48

 49 Stockfishdev-20240709-MuPV4              : 3511    3   9928    51.5 %   3500    84/128    4.9s   13.5s   0.55

 74 Stockfish17-MuPV4                        : 3488    3   9670    48.0 %   3501    80/128    6.6s   15.4s   0.47

 95 Patricia4.0                              : 3383    4   9177    32.8 %   3507    41/128    7.3s   22.7s   0.21

 97 Revenge4.0bmi2-MuPV4                     : 3345    4   9300    28.2 %   3507    55/128   23.8s   27.3s   0.03
 98 Monty20241122                            : 3337    4   8771    27.0 %   3509    25/128    4.4s   25.0s   0.16
 99 Igel3.6.0                                : 3335    4   8649    26.9 %   3509    30/128    7.8s   24.8s   0.15
100 Patricia3.1-MuPV4                        : 3307    4   9054    23.9 %   3509    32/128   24.5s   28.6s   0.03
101 Revenge3.0-MuPV4                         : 3302    4   9051    23.3 %   3509    29/128   23.5s   28.5s   0.02

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

And with this one MEA- suite



(2112 to about equal parts mixed forced and unforced positions, thus with single best move or with mulitple solutions and 1"/pos at 4 threads or 2 CPU- threads + 3070ti GPU, error bar in this one list is about 1% of Total Rate):

Code: Select all

    EPD  : epd\2112.epd
    Time : 3000ms
                                                 Max    Total   Time   Hash          
    Engine           Score   Found  Pos   ELO   Score   Rate    ms      Mb   Cpu      

 4  Corchess241222   110878   1835  2112  4018  124155  89.3%   1000    32    4
 5  Stockfish241222  110766   1831  2112  4014  124155  89.2%   1000    32    4
 6  Lc0v0.32-6147500 108776   1790  2112  3942  124155  87.6%   1000   100    2
 
10  Berserk241119    100699   1661  2112  3649  124155  81.1%   1000    32    4
11  Igel3.6           99232   1612  2112  3595  124155  79.9%   1000    32    4
12  Revenge4          98366   1622  2112  3564  124155  79.2%   3000    32    4

14  Revenge3          96392   1591  2112  3492  124155  77.6%   3000    32    4

                                    Created with MEA
                                          by
                                       Ferdinand
                                         Mosca

[peter]

(2112 to about equal parts mixed forced and unforced positions, thus with single best move or with mulitple solutions and 1"/pos at 4 threads or 2 CPU- threads + 3070ti GPU, error bar in this one list is about 1% of Total Rate):

Code: Select all

    EPD  : epd\2112.epd
    Time : 3000ms
......

Edit- time over, sorry for the typo above, in MEA- list the time per position- value is wrong as for the header and as for the Revenge- lines. 1000msec is correct as for the all of the runs and their real consumption of time per position. Revenge just needs to be set to 3000 in batch- file to use same 1000 as the other A-B-engines do. Lc0 needs 2000 for 1000, at this one older run I corrected it in result, with the latest Revenge- runs and the header of the list I forgot to do so, regards

[Paloma]

....
128 anti engine puzzles:

excerpted out of a list created with EloStatTS from Frank Schubert and 30"/pos. and 30 threads of a 16x3.5GHz CPU or 2 threads of it and a 3070ti GPU for Lc0

Code: Select all

    Program                                    Elo   +/-  Matches  Score   Av.Op.   S.Pos.   MST1    MST2   RIndex

  6 ShashChess34.6-MuPV4                     : 3554    3  10906    58.0 %   3498    99/128    4.0s    9.9s   0.65
  
 12 Crystaldev-20240503-MuPV4                : 3551    3  10797    57.6 %   3498   102/128    5.3s   10.3s   0.65
 
 
 47 Lc0v0.31.1-6147500PT-MuPV4               : 3512    3  10696    51.6 %   3501    87/128    5.9s   13.6s   0.48

 49 Stockfishdev-20240709-MuPV4              : 3511    3   9928    51.5 %   3500    84/128    4.9s   13.5s   0.55

 74 Stockfish17-MuPV4                        : 3488    3   9670    48.0 %   3501    80/128    6.6s   15.4s   0.47

 95 Patricia4.0                              : 3383    4   9177    32.8 %   3507    41/128    7.3s   22.7s   0.21

 97 Revenge4.0bmi2-MuPV4                     : 3345    4   9300    28.2 %   3507    55/128   23.8s   27.3s   0.03
 98 Monty20241122                            : 3337    4   8771    27.0 %   3509    25/128    4.4s   25.0s   0.16
 99 Igel3.6.0                                : 3335    4   8649    26.9 %   3509    30/128    7.8s   24.8s   0.15
100 Patricia3.1-MuPV4                        : 3307    4   9054    23.9 %   3509    32/128   24.5s   28.6s   0.03
101 Revenge3.0-MuPV4                         : 3302    4   9051    23.3 %   3509    29/128   23.5s   28.5s   0.02

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

....

Hello Peter,

is the score of Patricia 4 ........ 41/128 in the list correct? (95 placed)

above SF 17 ........................80/128 (74 placed)

below Revenge4.0 ..................55/128 (97 placed)

Or Revenge's score is to high ?

[peter]

is the score of Patricia 4 ........ 41/128 in the list correct? (95 placed)

above SF 17 ........................80/128 (74 placed)

below Revenge4.0 ..................55/128 (97 placed)

Or Revenge's score is to high ?

Score of SF17 is 3488 in EloStatTS- list, 105 more than that of Patricia4 and 143 more than that of Revenge4. Ranking with that tool comes from Elo- count, not from solved positions, what you can attach value to the least of course is the absolute numeric placement in list because the pool is very inhomogenous. I just let it grow as big as that with such a widely spread range of engines to get the error bar as small as that, the more engines scoring within this range is the smaller, the error bar gets.

Probably you doubt the higher Elo- count with the (in this case much) lower number of solved positions of Patricia compared to Revenge, well, that comes from the (in this case much) better MST (Mean solution time)- values of Patricia. EloStat compares each an every single run engine to engine and position to position, solved and unsolved ones, every new run again to every old one stored, and if an engine even has higher counts of solved positions now and then it happens, that it's yet listed below because of weaker MST- values. Cases like the one of Revenge 4 with rather much solved positions and yet rather low rating and ranking is seldom, that's right.
On principle with such small suites the results get less and less comparable to each other, the less positions are solved within TC, so the results at the top of the list have the more reliable performances than the ones at the bottom, even if error bar given by tool is near to each other. Problem of small sample size matters here of course too, at end of list especially as for transitivity and reproducibility, at top the more as for discrimination between the engines, branches and settings.
Yet you see the speciality of the time- usage of the engine in MEA too, Revenge, as written in second posting too, has to get 3000msec/pos. to use same amount of time in full run as other engines do with 1000.

[peter]

is the score of Patricia 4 ........ 41/128 in the list correct? (95 placed)

above SF 17 ........................80/128 (74 placed)

below Revenge4.0 ..................55/128 (97 placed)

Or Revenge's score is to high ?

Added Patricia now in MEA- list too:

Code: Select all

    EPD  : 2112.epd
    Time : 1000ms
                                                Max    Total   Time   Hash          
    Engine           Score   Found  Pos   ELO  Score   Rate     ms     Mb  Cpu      
 
12  Igel3.6          99232   1612  2112  3595  124155  79.9%   1000    32    4
13  Revenge4         98366   1622  2112  3564  124155  79.2%   1000    32    4

15  Revenge3         96392   1591  2112  3492  124155  77.6%   1000    32    4

18  Patricia4        92172   1518  2112  3339  124155  74.2%   1000    32    4

                                    Created with MEA
                                          by
                                       Ferdinand
                                         Mosca

The anti engine puzzles rather are a world of their own anyhow, but that's just why I'm yet still interested in them too, and as for such suites I like ElosStatTS especially, giving more discrimination than numbers of solved positions only do.

MEA- suites like the one above is nearer to game playing- results, as for those Patricia4- Revenge4 head to head, again with UHO 2024 900-990cp 6mvs and 3'+1" single thread, I already stopped after 260 games:

Score of Patricia4 vs Revenge4: 23 - 146 - 91 [0.263]
Elo difference: -178.6 +/- 35.7, LOS: 0.0 %, DrawRatio: 35.0 %
260 of 500 games finished.

[bastiball]

+90 elo in FRC
https://computerchess.org.uk/ccrl/404FR ... evenge_4_0