Ciekce wrote: ↑Wed Dec 27, 2023 9:43 amSo you would have us waste ridiculous amounts of electricity making the millions of games that happen in testing longer, so that you can save the odd bit of electricity in tournaments?
{snip}
Literally just use adjudication in your tournaments if you are this obsessed with shorter games. It is not our job to fix a problem of your own making.
Yes - save time, electricity, and mental bandwidth: adjudicate the opening position to be drawn, then go and do some actual living!
Human chess is partly about tactics and strategy, but mostly about memory
Let me have a look at the link ...
Ah, the Booot vs. SF game (in Booot there are different problems with endgames). So Booot very often think they have an advantage on the board in clear draw position and in most cases the games ended with the 50 move rule. In your example I can't give an answer, no answer is possible. But all the other engines have such endgames against Boost and the move average is not a random product. If I read what you have written in the past on this topic, I think you are more of the opinion that the higher move-average is a random product.
I think it was RubiChess, the last game of my tournament against Slow yesterday in the evening ... I remember me on over 250 moves. Slow with KNR and Rubi with KR. Both programs with 5-man syzygy and at least I remember (I did not look at the game much yesterday) your Rubi made the moves directly and all the time 0.00. Try 3fold different times. Not Slow, gave a little advantage for the position with syzygy ... seems to be a swindle mode with 0.50 ... maybe the opponent create a blunder. Nothing to do with contempt in this case.
It should be clear that longer games are in every database, all engines have them. But if the draw rate is too high, there must be other reasons and they are often very easy to see. Revealed by the engine with the same small evaluations.
All this is not necessary in neural network times and produces a higher move average. In the case of Booot an order for the programmer of Booot what to do in bad bishop endgames. OK, at the moment he have other problems in Ukraine, should be clear.
107 moves average for 14 Engines in a tournament.
Incredible ... for two years 1/50 engines I tested are light in the near of 100 moves.
Let us not speak about such a game you wrote before (Booot-SF), let us speak about 30% of games from the engines with such an high move-average, the games runs and runs and runs in draw positions. I search the position on SF site where the developers gave up Contempt. The explanation ... after a long time ... all what we try out is not working with Neural-Network (with my words, have the exactly text not in the head). And why the others have it?
The different between Stockfish and RubiChess to that topic is ... each SF draw need 17 moves longer as a draw from Rubi.
Again, what is wrong here and in all the other programs produced such high move-averages.
Best
Frank
Forderst mich dazu lange am PC zu sitzen und Aufsätze in English zu verfassen.
Normal sprechen Statistiken Bände und diese von mir dargestellten sind offensichtlich.
You know Andreas ... a lot of interesting chess programs are available.
Who I am that "cramp-draw-games" are running on one of my own PCs.
I have to stopped that directly.
So before I shoot the programs on the moon, I have interest to know what are the genial-ideas from the programmers who create that.
Hope, that I must not shoot all the programs on the moon because a week later I have to move on the moon. I've been annoyed watching it for weeks. OK, in a few hours a new tournament starts. Rarely have I been so annoyed ... 107 moves ... the average ... from 14 engines end of the year 2023 with all the knowledge we have about chess programmings. May to much of the knowledge is available.
That is an affront against the game chess.
107 moves the average ... should told me a person that for 20 years ... enough!
I somehow understand how it could be nice to have shorter games in principle. But this goal becomes completely random if you forbid resignation and draw offers. Then you will have zillions of moves in a decided position by setup - just as you created the scenario for this development yourself.
What should the engine that has 0.00 evals for hours on end do to avoid annoying you - blunder the game to a loss?
Peter Berger wrote: ↑Wed Dec 27, 2023 4:55 pm
I completely fail to get your point, Frank.
I somehow understand how it could be nice to have shorter games in principle. But this goal becomes completely random if you forbid resignation and draw offers. Then you will have zillions of moves in a decided position by setup - just as you created the scenario for this development yourself.
What should the engine that has 0.00 evals for hours on end do to avoid annoying you - blunder the game to a loss?
Really curious.
Peter
+1 Good point, well presented.
How about... shorten the thinking time that each player is allowed? Then you'll get more blunders, and hence shorter games. The games might be more fun as well, depending on your point of view.
Human chess is partly about tactics and strategy, but mostly about memory
Shortly, I have to work on my new tournament ... and on TV a nice documentary will start in a view time.
Please give me a short explanation why 99,9% of available chess programs produce a normal move average (around 85-90) with standard resign=off games. And suddenly 14 programs produced 107 moves? Again, I have never seen this in all my years of computer chess. If a programm is close to 100, that is a gigantic high result. 14 at the same time, well over 100!
It is more likely to be struck by lightning.
It is more likely that l think various people do not understand me for the reason that most of the computer chess people I know like to play with resign=on. Well, I have liked resign=off from the beginning and know what I'm talking about.
Draws are fully OK, more wins are better but ...
The move-average should not be a problem in computer-chess.
Thinking on the following situation:
Your wife / girlfriend, whatever you have ... has invited you to a dinner. You are sitting in front of your PC. You are looking the last chess game from two engines. You say to hear ... 10 minutes and we can go. 6 hours later you are searching your girlfriend / wife because you are ready for the dinner. The game is over after 400 moves.
You understand know?
In that time the 14 programs are most of time playing chess with 40 moves in 20 minutes ...
in that time ... for one game only ... I can give the house here a new roof. And if ready the last 100 moves are still running.
And all this for nothing, draw after 60 and after 400 moves in 99,9% of cases.
Config a tourney with resign = off ...
Take a week's holiday
And start the 50 games.
Only slightly exaggerated
Best
Frank
Again, enough ... have annoyed me enough about it.
I will not looking more in the evening in this thread ... closing time with move-average for a while.
Nr Avg Name
1 106.30 Booot 7.2 NN
2 104.26 Berserk 12 NN
3 104.06 Superultra 1.0 NN
4 103.31 Obsidian 8.0 NN
5 102.83 Koivisto 9.2 NN dev (hotfix)
6 102.42 Stormphrax 3.0.0 NN
7 102.26 rofChade 3.1 NN
8 100.97 Revenge 3.0 NN
9 100.36 SlowChess Blitz 2.9 NN
10 100.10 Stockfish 16 NN
11 99.42 Viridithas 11.0.0 NN
12 99.32 Texel 1.10 NN
13 99.10 Black Marlin 8.0 NN
14 98.35 Carp 3.0.1 NN
15 98.08 Clover 6.1 NN
16 97.93 Alexandria 5.1.0 NN
17 97.81 Willow 3.1 NN
18 96.41 Fire 9.2 NN
19 95.51 Altair 6.0.0 NN
20 95.28 BlackCore 6.0 NN
21 94.97 Akimbo 0.7.0 NN
22 93.64 Minic 3.39 NN
23 93.44 Uralochka 3.40a NN
24 92.97 Fritz 19 NN (Gingko)
25 92.78 Smallbrain 7.0 NN
26 91.93 Devre 4.0 NN
27 91.83 Seer 2.7.0 NN
28 91.83 Nemorino 6.11 NN dev
29 91.43 Pawn 2.0 NN
30 91.07 RubiChess 20230918 NN
31 90.84 Dragon 3.3 NN (Komodo)
32 90.67 Halogen 11.4 NN
33 90.55 Velvet 5.3.0 NN
34 90.54 Caissa 1.14.1 NN
35 90.49 Arasan 24.0 NN
36 89.67 Marvin 6.2.0 NN
37 89.49 Igel 3.5.0 NN
38 88.45 Wasp 6.61 NN dev
39 86.88 Rebel EAS NN
40 86.05 CSTal 2.00 NN
41 83.22 Chess.cpp 4.0 NN
So I don't think Halogen, Smallbrain, Akimbo and Willow should be on your top-14 list. Instead the following engines should be on the top-14 list:
About Revenge:
After the release of Revenge 3 I saw that the Contempt parameter is no longer included. After my first games I saw the high move average of Revenge last year and asked the author of Revenge directly in TalkChess. He sent me a new version in which I can adjust Contempt. The problem is that I forgot to add the "special" version in my first tournament and Revenge plays my first tournament with the standard release version. In the new, still running tournament I use the special version with Contempt = 0 ... ah, called DrawScore = 0.
To SlowChess:
For KI-ratings I forgot to set both contempt settings to 0 for Slows own GUI. Set contempt 2x to 0 in the UCI settings. Same for my first tournament. It seems that contempt does not work, so for the still running tournament I set both contempt settings for the GUI and from UCI settings to 0. Maybe the long games of SlowChess comes from a secret swindle mode, I don't know. I never asked the programmer.
I can't remove SlowChess from the tournament, because I need the engine for the ELO calculation for both tournaments ...
in the condition file you can see ...
SlowChess Blitz 2.9 NN plays with ~3380 Elo (reference).
But all in all ... I am looking for "Move-AVG wins" + "Move-AVG draws" ...
The table I added here are the games of the 14 engines after round 15, not after round 16.
