Incremental vs repeating TC

[AndrewGrant]

Don't hold back, Frank - tell everyone what you really think....

I've no dog in the hunt currently...and maybe I should feel bad about your power consumption and cost; but maybe Larry and team Komodo/Dragon have even more stats than yours regarding their product.

With the ccrl rapid elo gap between sf15 and dragon3 only seven elo and likely to shrink more with dragon 3.1, 3 elo is a big deal. For testing vs other NN engines I don’t object to Contemt 0, but many testers still test vs much weaker engines. Giving them draws in equal endings will hurt elo of course. In my view, testing with increment rather than with repeating time controls is the way to save electricity and testing time. Repeat time controls are obsolete in both human and engine play. No one likes to waste time playing out dead draws at a slow pace.

I test solely with increment and no repeating TC. From time to time I'll make a big change to the time management that gains elo for X+Y TC. The _most_ I would do, is ensure non-regression with 40/X

Based on Larry's comment, it stands to reason that Komodo does not work towards 40/X. Stockfish does not work towards 40/X. None of the open-source engines on OpenBench (17 of them!) work towards 40/X

[Frank Quisinsky]

Hi Larry,

testing with 5+1 for an example made no sense (if I understand your comment).

The longer the game the higher the Elo for stronger engines vs. weaker engines.
After 100 or 150 moves weaker engines like to play the blunder and the game is over (not enough time on clock).
With x moves in x minutes such problems you will never see = hardcore testing!!

Furthermore, from TOP-41 around 8 engines have not a strong time management.
In this case results from Dragon with 5+1 or 3+2 vs. others are only interesting for the trash.
You can't see how many stronger Dragon is if the opponents produced problems with the time control you like.

Harmless ...
Ponder = on produced bigger problems for an additional example!

If I like to see differences between engines, Fischer time controls make no sense for me.

Best
Frank

With standard settings ... draw games runs 40 moves longer ...
That is indeed hard to explain because no explanation possible.

[Modern Times]

In my view, testing with increment rather than with repeating time controls is the way to save electricity and testing time. Repeat time controls are obsolete in both human and engine play. No one likes to waste time playing out dead draws at a slow pace.

No disagreement from me. At blitz/rapid, long games and electricity aren't such a big problem with repeating time control, but nevertheless still obsolete.

[Graham Banks]

In my view, testing with increment rather than with repeating time controls is the way to save electricity and testing time. Repeat time controls are obsolete in both human and engine play. No one likes to waste time playing out dead draws at a slow pace.

No disagreement from me. At blitz/rapid, long games and electricity aren't such a big problem with repeating time control, but nevertheless still obsolete.

Many older engines don't accommodate x minutes with x seconds increments.

[Modern Times]

Many older engines don't accommodate x minutes with x seconds increments.

Sure, then they don't participate.

[Uri Blass]

Hi Larry,

testing with 5+1 for an example made no sense (if I understand your comment).

The longer the game the higher the Elo for stronger engines vs. weaker engines.
After 100 or 150 moves weaker engines like to play the blunder and the game is over (not enough time on clock).
With x moves in x minutes such problems you will never see = hardcore testing!!

Furthermore, from TOP-41 around 8 engines have not a strong time management.
In this case results from Dragon with 5+1 or 3+2 vs. others are only interesting for the trash.
You can't see how many stronger Dragon is if the opponents produced problems with the time control you like.

Harmless ...
Ponder = on produced bigger problems for an additional example!

If I like to see differences between engines, Fischer time controls make no sense for me.

Best
Frank

With standard settings ... draw games runs 40 moves longer ...
That is indeed hard to explain because no explanation possible.

You can win a chess game only by mistakes of the opponent so I see no problem with the fact that after 100 or 150 moves weaker engines play a blunder and the game is over.

For every weak engine that you test there are weaker engines and if weaker engines that are still above rating 3000 can blunder with 1 second increment then I am sure that there are a lot of strong engines that can blunder also with repeating time control of x minutes/40 moves
so the decision to make the games long with a lot of moves is justified because you continue if there are practical chances to win.

Saying the engine should know it is a draw and make a fast draw is not fair if there are practical chances to win.
Something that I dislike about analysis of engines is 0.00 scores for positions that are draws but only one side has practical chances.

When I analyze game that humans played and somebody(in a theoretical drawn position) did a practical mistake that lost practical chances to win then I want chess engines to tell me it and not say that there is nothing wrong because the position was a draw with perfect play also earlier.

I am not sure if top engines are better in winning drawn endgames against humans relative to carlsen who won a drawn endgame in the olympiad.
If their evaluation is that everything is a draw(0.00) then I see no reason for them to choose moves that give better practical chances to win.

https://2700chess.com/games/carlsen-mei ... 2022-07-30

[lkaufman]

You can win a chess game only by mistakes of the opponent so I see no problem with the fact that after 100 or 150 moves weaker engines play a blunder and the game is over.

For every weak engine that you test there are weaker engines and if weaker engines that are still above rating 3000 can blunder with 1 second increment then I am sure that there are a lot of strong engines that can blunder also with repeating time control of x minutes/40 moves
so the decision to make the games long with a lot of moves is justified because you continue if there are practical chances to win.

Saying the engine should know it is a draw and make a fast draw is not fair if there are practical chances to win.
Something that I dislike about analysis of engines is 0.00 scores for positions that are draws but only one side has practical chances.

When I analyze game that humans played and somebody(in a theoretical drawn position) did a practical mistake that lost practical chances to win then I want chess engines to tell me it and not say that there is nothing wrong because the position was a draw with perfect play also earlier.

I am not sure if top engines are better in winning drawn endgames against humans relative to carlsen who won a drawn endgame in the olympiad.
If their evaluation is that everything is a draw(0.00) then I see no reason for them to choose moves that give better practical chances to win.

https://2700chess.com/games/carlsen-mei ... 2022-07-30

Of course you are correct that sometimes the top engines do give 0.00 scores when only one side may win in practice, but in general endgames where one side has the better side of a draw, such as rook plus 3 pawns vs 2 pawns on one wing, or even rook plus 2 vs rook plus 1, or even bishop plus two vs bishop plus one on same side (even with BOC!), get evals from both Stockfish and Dragon (without Contempt) that imply about a 51% expected score. That's not much, but it is enough to make the engines go for these positions rather than just repeating moves for example. So in practice they are extremely strong in playing endgames against 2800 level humans, far better than Carlsen in general, although there will of course be specific positions where he might score better than the engines against 2800 humans. In general, I think that top engines will win more drawn positions against 2800 humans if they don't search too deeply; maybe something like 22 plies or so is optimal for that purpose. I demonstrated that when giving knight odds, Komodo's results top out at somewhere around this level, more depth actually hurts. The same is probably true if you force it to give draw odds to 2800 level opponents (human or engine) from rather drawish initial positions. Once you outsearch your opponent by somewhere in the ballpark of ten plies or so, it doesn't pay to worry about things he is unlikely to see anyway.

[Frank Quisinsky]

Hi Modern Times,

in the past testers are thinking (Winboard times):
Fischer time controls for computer chess is more or less a toy.
Interesting for people like to have a big number of games in a short time, or for people like to have more "fun".

The playing strength is influenced too much.

Example DGA:
decisively gained advantage

Average of DGA for most 1:0, 0:1 games is between move-number 54-66!
Games ended with a average from 86 moves!

We are thinking in the past: What is the best time control for testing engines?
With 60 moves in x minutes to many engines are playing moves to fast in most critical phase of game (before time control).

Best time control for testing engines:
40 moves in x minutes (not 60 moves in x minutes, or Fischer time controls) ... the final results.

Equal opportunities are most important for testing engines!!
Like one core (factor for 2, 4, 8 cores isn't the same for all the engines), like all with same hash, like all with 4-pieces, 5-pieces or without endgame databases and so one.

With Fischer time controls engines, stronger in endgames, get a booster.
For better moves they need lesser time.

But too often the time-management is bad and endgame-booster and these problems is to much!
All what is artificially created is ... BAD!

I like Fischer time-controls for looking still running games also much more!
But for looking in real ElO differences, Fischer time-controls are more or less a toy in computer-chess hardcore-testing.
No hardcore-tester I like to switched information since more as 20 years used Fischer time controls.

Programmers I can understand here ... need game in 1 minute or 1+1 but for hardcore-testing with many engines ... contempt, games with resign= on, Fischer time controls, more as 1-core and so one is really bad ... if you like to see the real differences between a big group of available engines and if you like to looking in statistics.

Best
Frank

[Modern Times]

The only reason I haven't switched to Fischer time controls is that it would mean abandoning everything I've done and starting again. I'll just continue on with repeating time control until I cease ratings list work completely.

[Frank Quisinsky]

OK, plausible!

Thinking the same before I started FCP Tourney-2020 because I like more the Fischer time controls.
But during the tourney is still running I am thinking ... all the efforts for that long-time of tournament with 41.000 games are for the trash. After I am looking in detail and stats of engines. So I switched back to 40 moves in x minutes.

Most I like:
Fischer time controls in combination with ponder = on and contempt = off.

But here the problems are much bigger as to use ponder = off.
To many mistakes in engines and time-management.

But maybe I should kill all engines with a bad time-mangament, engines where I can't set contempt = 0 or engines produced "lost on time" games with "ponder = on". The question is how many from TOP-41 engines are end of the day in my tourney.

Means end of the day I have only:

Frank Quisinsky 1
Frank Quisinsky 2
Frank Quisinsky 3
and no engines!

And this is more as boring!