Latest TCEC has 379 moves draw

[Jouni]

https://tcec-chess.com/#div=sf&game=84&season=25. I play normally with 100 moves limit . Saves a lot of time.

[Frank Quisinsky]

Hi Jouni,

no sensation ...
In my older FCP-Rating-List I found games with more than 450 moves (about 120.000 played games).

From the first 9.960 90-minutes-games (move average = 94,7 from 41 of the TOP-44 CPU-engines) on 4.0Ghz without resign-mode.

Statistics from the still running FCP-Tourney-2024 ...
= 3.641 / 9.960 games with more than 100 moves = 36,5%.

All contempt parameters for nothing or maybe today for 1-3 Elo more. For different engines you can set contempt, but in reality if you set contempt=0 the engine used a non-adjustable "secret" contempt.

Here is a graphic of the longest of the 9,960 games...



It's nice to have so many nice combinations in a database (games without resing-mode = for me nice-to-have).
With such databases you can write 100 written chess books if you have the time.
The other side of the gold medal are all the long games and a lot of test time runs in Nirwana ... gone to the happy (in this case unhappy) hunting grounds.

Best
Frank

[Frank Quisinsky]

Maybe all the bad contempts comes from Stockfish dev ideas.
Since Contempt is not adjustable for Stockfish so many other programmers do the same.

More stats from the still-running FCP-Tourney-2024 ...

Move average for the first 75 Stockfish 16 NN games = 102 moves

Let me compare it with Dragon 3.3 NN (Komodo) with Contempt = 0 (Contempt is working fine)
Move average for the first 75 Dragon 3.3 NN (Komodo) games = 86 moves

Logical, that I like more to looking in Dragon games while the current updates are still running.
I have here a lot of fun with Dragon ... the aggressivess after the opening moves = for me ... sensational!

Best
Frank

[RubiChess]

Maybe all the bad contempts comes from Stockfish dev ideas.
Since Contempt is not adjustable for Stockfish so many other programmers do the same.

I don't know what contempt has got to do with this 379 moves draw. When you look at the game you can see that Stockfish agrees to a draw at move 163 with eval < 0.25 and keeps this until end of game while Leela keeps the eval high preventing adjudication.

My guess is that in general long games are not caused by what you call contempt (most of the engine probably don't have implemented anything like that) but by wrong/bad eval of one of the engines seeing some advantage (material or advanced pawn that leads to nothing or whatever) that is none.

And if SF games/draws are longer in average than probably because its evaluation is superior to most/all other engines and it saves many games into drawn positions that the opponent doesn't accept to be drawn.

Keeping the position complex, keeping pieces on the board until advantage is visible is indeed part of SF strategy, will probably raise average move number and gains Elo so there's nothing wrong with it.

Will download your pgn when there are Stockfish games in it to look for some examples.

Regards, Andreas

[Frank Quisinsky]

Sorry, this is nonsense in most cases.

To avoid a draw after 49 moves with a pawn move. When I like to replay the long games, very often is to see ... 0.06 (and 0.07) and the other engines like to give 0.00, no contempt settings adjustable. And all this in clear different piece combinations on board. If adjustable with a contempt parameter in clearly most the move average is 15 moves less. Test it with 8 of the engines contempt = 0 is possible, exactly it's 14.7 moves less. Easy statistic!

The reality is that engines like CSTal or Wasp forced a clearly draw very intelligent (sensational move-average with 83 moves) and others awoid a draw with "brute-force". With the final result that the move-average of games goes clearly over 100 moves ... very easy to see in all the games. For nothing if I am searching the 1:0 end-games in database.

[Sopel]

You're right. If you could adjust contempt by 10000 most games would end pretty quick.

[Frank Quisinsky]

Hi Tomasz,

I wish programmers would add such a contempt parameter. I have a feeling that the parameter has been out of Stockfish for around two or three years. I like ShashChess, but also removed from ShashChess.

On the other hand have a look in CSTal ... no contempt parameter and engine produced clearly the lowest move average of TOP-20 (games without resign=on).

Thinking about the discussion my FCP-Tourney-2020 is still running or it was before with a rating list. The difference between rank 1 and rank 41 for FCP-Tourney-2020 is clearly over 400 Elo. The stronger engines lost many Elo against the weaker engines with all the draws and contempt was important. But today, in neural network times ... the time for contempt parameters is over. The results are just very long games for absolutely nothing.

I keep hearing that people are losing interest in computer chess. The reason is not more draws ... the reason is unnecessary playing in clear draw endgame positions. All the long games chessplayers will not see.

OK, I can produce enough interesting games on my systems to look at.

Best
Frank

[connor_mcmonigle]

Hi Tomasz,

I wish programmers would add such a contempt parameter. I have a feeling that the parameter has been out of Stockfish for around two or three years. I like ShashChess, but also removed from ShashChess.

On the other hand have a look in CSTal ... no contempt parameter and engine produced clearly the lowest move average of TOP-20 (games without resign=on).

Thinking about the discussion my FCP-Tourney-2020 is still running or it was before with a rating list. The difference between rank 1 and rank 41 for FCP-Tourney-2020 is clearly over 400 Elo. The stronger engines lost many Elo against the weaker engines with all the draws and contempt was important. But today, in neural network times ... the time for contempt parameters is over. The results are just very long games for absolutely nothing.

I keep hearing that people are losing interest in computer chess. The reason is not more draws ... the reason is unnecessary playing in clear draw endgame positions. All the long games chessplayers will not see.

OK, I can produce enough interesting games on my systems to look at.

Best
Frank

Rather than demanding that the world change, you could just configure draw adjudication. Engines won't force a draw in a comfortable position/a position where they maintain some slight edge as doing so would obviously lose Elo.

[Gabor Szots]

All of the long games were draws. They should simply stop the game and award a draw after a move limit has been reached. The limit could be chosen by taking the longest game not ending in a draw and chosing its number of moves (in fact some lower number will do because it is unlikely that a decisive game will ever even approach that number).

[Frank Quisinsky]

Hi Connor,

You wrote:
"Rather than demanding that the world change ..."

OK, you are right here!
My fight for Winboard, UCI and other things has been over for a few years.
I'm often in "winboard-fighting mode" when I don't like this or that.

I know that!
My computer chess heart is powerful.

I have a different opinion on the subject.
I search the database for games where engines lost Elo.
Easy, I have to look in the end of games of about 500 of 10,000 games goes over 180 moves. 41 engines are in the tournament, it seems that three of the engines lost games in late endgames. All three engines without Syzygy support.

But all the files are online and I am sure the others who have looked in detail will have the same opinion.

Martin Blume and I have been working on the Arena GUI for a long time.
We always have the same opinion.

A GUI should never affect a game between two engines. Not with endgame tables and not with other things. That is the job of the engine programmers. I have never changed my opinion!

Once again ...
All the long games are indeed for me clearly the bad point in computer chess. So I talk about it. We can also talk about the highlights. Many new and very interesting and very powerful engines available. Lots of young programmers with great ideas. The highlights are in the background, but in my opinion ... my impressions ... the neural network is more in the background a detail optimisation for endgames. The database is full of games with wrong bishop-endgames, engines do not understand ... also a reason for all the long games. syzygy support helps here a bit.

But with time also the "newer" engines will be more and more optimated.
I've always been very impatient.

Best
Frank