Revenge 3.0 produce a contempt problem?

[Frank Quisinsky]

Hi there,

a question to the programmer of Revenge:
Draw-Score = 0 for Revenge 2.0 and Revenge 3.0 (UCI-settings).
I made here no changes and used default settings for "draw-score".

But have a look here ...
white draw = 19 moves higher (average) ... v.3 to v.2
black draw = 12 moves higher (average) ... v3 to v.2
= average for white / black = 16 moves higher between v2.0 and v3.0 without that I changed "draw-score".

Is this a "secret" contempt parameter or perhaps a small error for "Draw-Socre = 0).
Perhaps "Draw-Score = x, 10 for an example" is more right?

Best
Frank


40 moves in 6 minutes (test-tourney)

Code: Select all

10. Revenge 3.0 NN                 (467+,   531=,   23-)  71.7%

          CATEGORIES   MOVE_AVE    GAMES    TOTAL  GAME_RATE
           all games         91     1021
           white won         78      289
           black won         84      178
          white draw        109      216
          black draw         94      315
          white loss         95        5
          black loss         86       18
            all wins         81      467     1021      45.7%
           all draws        100      531     1021      52.0%
           all loses         88       23     1021       2.3%

40 moves in 20 minutes
FCP Tourney-KI, last offical run-6

Code: Select all

26. Revenge 2.0 NN                 (437+,   713=,   50-)  66.1%

          CATEGORIES   MOVE_AVE    GAMES    TOTAL  GAME_RATE
           all games         84     1200
           white won         81      286
           black won         85      151
          white draw         88      305
          black draw         82      408
          white loss         91        9
          black loss         92       41
            all wins         82      437     1200      36.4%
           all draws         84      713     1200      59.4%
           all loses         92       50     1200       4.2%

Forgotten ...
Games I produced are always without resign-mode!

[Frank Quisinsky]

Shortly ...

At the moment I delete all engines produced a high move average from my private test tourneys.
If I can't set contempt = 0

Examples:
For Dragon by Komodo 3.0 I can set Contempt = 0 and the move average is fine, works great.
Standard setting is with contempt = x

For current Stockfish dev versions "Contempt" parameter is since a while no longer available.
SF produced very long and boring draws with resign = off.
Engine delete from my group of test-engines I like to use for testing.
But in case of Stockfish I can use the last SF without NN (interesting to see).
Here I can set Contempt = 0.

Koivisto I can not use ... no contempt paramter ... move average is clearly to high.
Engine cancelt from the group of test-engines.

My topic is:
Energy is expensive these times!!
Games running 150-300 moves in clearly draw position isn't in my interest to see again and again.

Revenge 2.0 works fine here!
So I can use Revenge 2.0.
But perhaps it's really a small mistake in Revenge 3.0?

In case of Revenge ...
One of my favorite engines, I hope the programmer can fix that or can give a tip.
Perhaps a negative draw-score?

Best
Frank

[jhonnold]

I think you should consider enabling draw adjudication.

[Frank Quisinsky]

Hi Jay,

yes, that is possible.
In hard testing I try to find out without such parameters.

---

With Berserk I made also some experiments.
At first, I believe in one of the first v9.0 dev versions you added contempt parameter.

I can changed from Contempt 12 to 0 and this works great!

Note:
I am looking at the moment not on Elo, only on short wins and move average.

That is a blessing in case of Berserk and move average (contempt).
Attacking stats are also where I am looking ... all is OK, good to have for me!

I am thinking ...
Strongest attacker in a field of test-engines ...

The quantity of miniatures is low but the quality of such produced games is higher.
When the strongest attackers among themselves produced short won games.
All this with a low move average and I am happy.

Example:
Velvet 4 have the lowest move average from all TOP-50 engines (games without resign-mode).
71-73 ... only 3 or 4 engines have around 77-79 moves.
Wasp have for an example around 83 moves and is here around on place 14 in the World.

And Velvent won a lot of fast games! Ok, lost vs. strongest opponents the same quantity of games.
But the for Elo Velvet have is that sensational!

Such things I like most ... low move-average and many fast won-games.
I am thinking ... if 71-73 move average is possible maybe other programmers have interest to try such optimizations also.

...

Thanks for Berserk!

Best
Frank

[Frank Quisinsky]

Other example = Stockfish

If I add Stockfish 11 (last official without NN) in my TOP-41 tourney and looking for quantity of fast won games only ...
compare that with Stockfish 15 or a current dev version ...

The final result is:
Stockfish 11 produced around the same quantity of fast won games but is 150 Elo weaker as Stockfish 15 NN (40 moves in 20 minutes).

Interesting is to looking in detail in the games.
Won games from SF 11 are more speculative as the won games from SF 15.

From this point of view ...
Stockfish 11 is more interesting to have in a field of strong attackers as Stockfish 15.

Best
Frank

...
TOP-4 with Velvet
Berserk is not inside this 21er tourney, but in my last 40 / 6 tourney Berserk is on place 5. Stats are great!
From Stockfish is playing the last dev version without NN, around 140 Elo weaker as a current dev version.

A still running tourney on different on my systems:
40 moves in 6 minutes!
At the moment all engines played around 1050 games.
Here only the games 20-59 moves.

Code: Select all

   1 Dragon 3 NN (Komodo)      :  3358.73    215    68.6   80   135     0   147.5   62.8  25.11  3209.75  20.49   19.0
   2 Revenge 3.0 NN            :  3334.37    167    63.8   46   121     0   106.5   72.5  24.89  3223.94  20.53   18.4
   3 Stockfish 310720 dev      :  3330.06    190    65.5   61   127     2   124.5   66.8  24.21  3210.76  20.50   19.4
   4 SlowChess Blitz 2.9 NN    :  3306.31    179    62.0   44   134     1   111.0   74.9  22.70  3215.63  20.14   19.1
  11 Velvet 4.0.1 NN           :  3224.27    197    52.0   27   151    19   102.5   76.6  17.91  3216.02  20.28   18.3
 

[Fabio Gobbato]

I have added a small random score for the draws by repetitions in this way we reduced by 3-4% the number of draws by repetition and add some elo against low rated opponents.

[Frank Quisinsky]

Hi Fabio,

thanks for your answere!

think so ...
But it's possible to create an UCI parameter for such things?
So I can switched to off.

Average of draw games is 16 moves higher!!

In the same time I can produce 1.000 40/20 games with Revenge 2.0 I can produce 860 games with Revenge 3.0 only!
That is indeed a big problem if I try to find out, with engines strong in king attacks, best openings for test-sets.
I lose to many time.

With other words:
Elo vs. Playing style

Both should be possible with UCI settings.
Max Elo with such ideas you added in Revenge 3.0
And to try to reduce average of moves for draw games!
Both is important!

No doubt about it, the short won-games stats are again improved ... from 2.0 to 3.0.
That's great!

Best
Frank

Short information for you v2 to v3 ... Elo gain:
- +37 Elo with 40 moves in 20 minutes
- +53 Elo with 40 moves in 6 minutes
- +74 Elo with 40 moves in 3 minutes

[Fabio Gobbato]

Unfortunately is not possible to switch off this draw score. If you want I can create a new build for you with this kind of repetition scoring disabled so you can use this version for your tests.

[Frank Quisinsky]

Fabio,

please ... Revenge is extremly important for me.
"Must have engine" for the ideas I have to find out best openings for test-sets and v3.0 is clearly improved in aggressiveness. If I can reduce the move average for draw games I can produced more games.

If you like to do that ...
Big thanks!!!

Best
Frank

[Frank Quisinsky]

Hi Fabio,

thank you, have it!
TOP-Service!!

A great day!

Best
Frank