CCRL adapted TC to current hardware

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Uri Blass
Posts: 8771
Joined: Wed Mar 08, 2006 11:37 pm
Location: Tel-Aviv Israel

Re: CCRL adapted TC to current hardware

Post by Uri Blass » Thu Jan 02, 2020 11:00 am

Laskos wrote:
Thu Jan 02, 2020 10:18 am
Guenther wrote:
Thu Jan 02, 2020 9:21 am
Gabor Szots wrote:
Thu Jan 02, 2020 8:28 am
I am very much in favour of incremental clock which in my case (I currently test at 40/2) would be 4+2, with a ratio of 120, while approximately keeping total game time the same.

...
I guess this was a typo and you meant 2+2 instead of 4+2?
40/2 is similar (just a bit weaker IIRC) compared to 2+1 strength-wise, but 2+1 usually consumes less total time per game. Also, the time usage curve is smoother and more efficient with 2+1.
I think that in order to know if 40/2 is a bit weaker than 2+1 we need to optimize existing time management for both time controls.

I doubt if engines are optimized for 40/2 time control(all the tests with stockfish are done with incremental time control and not with time control like 40/2 so I suspect engines are not optimized for this time control and may be too slow in the first moves(or too fast in the first moves).

I am not sure if using optimal time control for 40/2 is really weaker than using optimal time control for 2+1

I wonder if it is not possible to increase the playing strength of many engines by modifying the time control to time control that in theory is not better
for example
replace 40/2 repeating by 20/1:10+20/50+40/2

It is clear that with optimal time control it is not worse to have 40/2(because the engine may assume the second time control and get the same result) but it is not clear for me that practically engines are not going to perform better with the second time control.

Guenther
Posts: 3282
Joined: Wed Oct 01, 2008 4:33 am
Location: Regensburg, Germany
Full name: Guenther Simon
Contact:

Re: CCRL adapted TC to current hardware

Post by Guenther » Thu Jan 02, 2020 11:14 am

Uri Blass wrote:
Thu Jan 02, 2020 11:00 am

...

I doubt if engines are optimized for 40/2 time control(all the tests with stockfish are done with incremental time control and not with time control like 40/2 so I suspect engines are not optimized for this time control and may be too slow in the first moves(or too fast in the first moves).

...
There are >800 programs around with a name, which doesn't start with 'Stock...'
Sure 'some' of them are optimized for different time controls.
https://rwbc-chess.de
Greg Strong@ovyron wrote: What I think is that you have a very, very, VERY strong desire to post, and very little self-control.
All the time. In every conversation. You must post. Something. It's like you just can't help yourself.

User avatar
Ovyron
Posts: 4360
Joined: Tue Jul 03, 2007 2:30 am

Re: CCRL adapted TC to current hardware

Post by Ovyron » Thu Jan 02, 2020 11:28 am

Gabor Szots wrote:
Thu Jan 02, 2020 8:28 am
I wonder if anyone would notice any difference between the quality of the games played under different TC types.
It'll be trivial as repeating time controls will have a dip in move quality before move 40 and a spike in move quality after it is reached.

User avatar
Laskos
Posts: 10328
Joined: Wed Jul 26, 2006 8:21 pm
Full name: Kai Laskos

Re: CCRL adapted TC to current hardware

Post by Laskos » Thu Jan 02, 2020 11:42 am

Uri Blass wrote:
Thu Jan 02, 2020 11:00 am
Laskos wrote:
Thu Jan 02, 2020 10:18 am
Guenther wrote:
Thu Jan 02, 2020 9:21 am
Gabor Szots wrote:
Thu Jan 02, 2020 8:28 am
I am very much in favour of incremental clock which in my case (I currently test at 40/2) would be 4+2, with a ratio of 120, while approximately keeping total game time the same.

...
I guess this was a typo and you meant 2+2 instead of 4+2?
40/2 is similar (just a bit weaker IIRC) compared to 2+1 strength-wise, but 2+1 usually consumes less total time per game. Also, the time usage curve is smoother and more efficient with 2+1.
I think that in order to know if 40/2 is a bit weaker than 2+1 we need to optimize existing time management for both time controls.

I doubt if engines are optimized for 40/2 time control(all the tests with stockfish are done with incremental time control and not with time control like 40/2 so I suspect engines are not optimized for this time control and may be too slow in the first moves(or too fast in the first moves).

I am not sure if using optimal time control for 40/2 is really weaker than using optimal time control for 2+1

I wonder if it is not possible to increase the playing strength of many engines by modifying the time control to time control that in theory is not better
for example
replace 40/2 repeating by 20/1:10+20/50+40/2

It is clear that with optimal time control it is not worse to have 40/2(because the engine may assume the second time control and get the same result) but it is not clear for me that practically engines are not going to perform better with the second time control.
It was studied.

First a note.
40/2 is on average some ~15% longer in games than 2+1, so even if they come very close Elo-wise, 2+1 is better.

The time usages of engines are not arbitrary. For the same amount of spent time, the decaying curve after move 20 or so to some almost constant small increment was shown to be optimal time usage for engines. The experiment is doable. Take for the move interval 20-30 double time for one engine versus single time for the same engine. The same time for both engines for the rest of the moves. See the outcome in self-games. Compare it to total doubling in time Elo difference. This way, on intervals, I found that IIRC moves 10-20 are the most important in games, moves 3-10 just a bit less important (this was somehow a surprise, and I am not sure of it), then importance after the move 20 decreased, and to move 50-60 became small and close to (diminishing a bit) constant to moves 100+. This shape is not arbitrary and not dependent on a particular engine and its time usage implementation.
The time control which closely resembles this type of curve, independently of the engine used, is "base + inc".
40/X simply cannot be adapted to play that well, the moves 41-80 are FAAAR less important than the moves 1-40. Seeing the result, independently of engine, "base + inc" comes as no-brainer as optimum time usage at fixed total time. The details sure might depend on the engine, how is the shape in moves 1-20, how fast is the decay of time usage towards the tail function of material on the board, complexity of positions, etc. With a very good time usage implementation, even "total time per game" can become very competitive, to being equal to the best of "base + inc". But not 40/X, one can do nothing with the time usage implementation in his engine in this case to approach the efficiency of a well chosen and implemented "base + inc" time usage.

mwyoung
Posts: 2122
Joined: Wed May 12, 2010 8:00 pm

Re: CCRL adapted TC to current hardware

Post by mwyoung » Fri Jan 03, 2020 1:19 am

Guenther wrote:
Wed Jan 01, 2020 11:07 am
It seems I missed this change or was it reported before?
40/40 => 40/15 and 40/4 => 40/2
(The links itself are not yet changed though)

This reflects more the real tc played on newer hardware - I would still welcome adding the real tc always
in the pgn header and not just the generic one, at least for the future, as it is probably impossible w/o huge work
to have it for past games too.

http://ccrl.chessdom.com/ccrl/404/rating_list_all.html

Image
Your welcome, search.php?st=0&sk=t&sd=d&sr=posts&keyw ... s&start=70
Professing themselves to be wise, they became fools,
take on me. Foes 0.

Adam Hair
Posts: 3220
Joined: Wed May 06, 2009 8:31 pm
Location: Fuquay-Varina, North Carolina

Re: CCRL adapted TC to current hardware

Post by Adam Hair » Fri Jan 03, 2020 3:18 am

Laskos wrote:
Wed Jan 01, 2020 7:19 pm
kasinp wrote:
Wed Jan 01, 2020 7:06 pm
Guenther wrote:
Wed Jan 01, 2020 4:19 pm
Alayan wrote:
Wed Jan 01, 2020 4:03 pm
Thanks for the explanations, both of you.

...

My understanding is that CCRL is sticking to the old repeating time control (which at equal total time used leads to lower quality compared to standard increment,
...
I see it exactly the opposite, usually inc games lead to low quality in the endgame, or even long before, if a game lasts much longer
than 40 or 50 moves. IMO it is ugly to have a game finally decided by the low inc for dozens of moves.
I completely agree.

In fact, this is my issue with tournaments like TCEC. Amazing hardware to ensure "highest game quality", but only until engines reach bare increment stage. At that point quality actually falls below what many enthusiasts can afford at home, provied they stick to long time control.
Computer chess needs creative solutions to game adjudication. Today we simply sacrifice engame quality to avoid trolls.

Peter
It was studied. For the same amount of total time spent, the strongest TC setting Elo-wise is increment in the form of base + base/100 or even base + base/200. Fixed time per move is the weakest Elo-wise for the same total time spent. And 40/X somewhere in between. The large amount of time spent on moves 41-60 with most engines is partly wasted.
This thread involves one such study. I believe Kai has performed more extensive experiments.

From that thread:

Code: Select all

  
   # PLAYER                       : RATING    POINTS  PLAYED    (%)
   1 SF 141130 1'+1"              :   28.5     332.0     600   55.3%
   2 SF 141130 2'+0               :    6.7     307.5     600   51.2%
   3 SF 141130 40/1'              :   -8.0     291.0     600   48.5%
   4 SF 141130 2" per move        :  -27.1     269.5     600   44.9%

Code: Select all

Engine                 Depth       Time   Games     Moves  Average Forfeit
SF 141130 1+1          22.88   19:57:32     600     35336    2.03     0
SF 141130 2sec_move    24.18   19:48:40     600     35734    2.00     0
SF 141130 40/1         23.67   17:16:46     600     35346    1.76     0
SF 141130 2+0          20.92   16:29:08     600     35201    1.69     0


Time control comparison between engines

Depth     : Average search depth
Time      : Total time engine used
Moves     : Total moves engine played
Average   : Average time per move in centi-seconds
Forfeit   : Games engine lost due to time forfeit

List is sorted on Average Time indicating the engine that uses the most time tops.

Image

Adam Hair
Posts: 3220
Joined: Wed May 06, 2009 8:31 pm
Location: Fuquay-Varina, North Carolina

Re: CCRL adapted TC to current hardware

Post by Adam Hair » Fri Jan 03, 2020 3:38 am

mwyoung wrote:
Fri Jan 03, 2020 1:19 am
Guenther wrote:
Wed Jan 01, 2020 11:07 am
It seems I missed this change or was it reported before?
40/40 => 40/15 and 40/4 => 40/2
(The links itself are not yet changed though)

This reflects more the real tc played on newer hardware - I would still welcome adding the real tc always
in the pgn header and not just the generic one, at least for the future, as it is probably impossible w/o huge work
to have it for past games too.

http://ccrl.chessdom.com/ccrl/404/rating_list_all.html

Image
Your welcome, search.php?st=0&sk=t&sd=d&sr=posts&keyw ... s&start=70
Thank you?

User avatar
Ovyron
Posts: 4360
Joined: Tue Jul 03, 2007 2:30 am

Re: CCRL adapted TC to current hardware

Post by Ovyron » Fri Jan 03, 2020 3:41 am

Laskos wrote:
Thu Jan 02, 2020 11:42 am
the moves 41-80 are FAAAR less important than the moves 1-40.
Wow, I always critiqued Zappa Mexico II for its terrible time management on repeating time controls, always using too much time and being at a clock disadvantage against their opponents after move 40. It's only now that I find out it was using its time where most mattered.

Now I wonder if people can improve the performance of their engine at repeating time control by having the engine lie to itself and make it think it's on some base+inc time and manage their time like that, a time control that doesn't exist that keeps changing and slowly converges to repeating time control near the 40th move. There's got to be a way to exploit how opposing engines manage their repeating time control.

Uri Blass
Posts: 8771
Joined: Wed Mar 08, 2006 11:37 pm
Location: Tel-Aviv Israel

Re: CCRL adapted TC to current hardware

Post by Uri Blass » Fri Jan 03, 2020 8:03 am

Laskos wrote:
Thu Jan 02, 2020 11:42 am
Uri Blass wrote:
Thu Jan 02, 2020 11:00 am
Laskos wrote:
Thu Jan 02, 2020 10:18 am
Guenther wrote:
Thu Jan 02, 2020 9:21 am
Gabor Szots wrote:
Thu Jan 02, 2020 8:28 am
I am very much in favour of incremental clock which in my case (I currently test at 40/2) would be 4+2, with a ratio of 120, while approximately keeping total game time the same.

...
I guess this was a typo and you meant 2+2 instead of 4+2?
40/2 is similar (just a bit weaker IIRC) compared to 2+1 strength-wise, but 2+1 usually consumes less total time per game. Also, the time usage curve is smoother and more efficient with 2+1.
I think that in order to know if 40/2 is a bit weaker than 2+1 we need to optimize existing time management for both time controls.

I doubt if engines are optimized for 40/2 time control(all the tests with stockfish are done with incremental time control and not with time control like 40/2 so I suspect engines are not optimized for this time control and may be too slow in the first moves(or too fast in the first moves).

I am not sure if using optimal time control for 40/2 is really weaker than using optimal time control for 2+1

I wonder if it is not possible to increase the playing strength of many engines by modifying the time control to time control that in theory is not better
for example
replace 40/2 repeating by 20/1:10+20/50+40/2

It is clear that with optimal time control it is not worse to have 40/2(because the engine may assume the second time control and get the same result) but it is not clear for me that practically engines are not going to perform better with the second time control.
It was studied.

First a note.
40/2 is on average some ~15% longer in games than 2+1, so even if they come very close Elo-wise, 2+1 is better.

The time usages of engines are not arbitrary. For the same amount of spent time, the decaying curve after move 20 or so to some almost constant small increment was shown to be optimal time usage for engines. The experiment is doable. Take for the move interval 20-30 double time for one engine versus single time for the same engine. The same time for both engines for the rest of the moves. See the outcome in self-games. Compare it to total doubling in time Elo difference. This way, on intervals, I found that IIRC moves 10-20 are the most important in games, moves 3-10 just a bit less important (this was somehow a surprise, and I am not sure of it), then importance after the move 20 decreased, and to move 50-60 became small and close to (diminishing a bit) constant to moves 100+. This shape is not arbitrary and not dependent on a particular engine and its time usage implementation.
The time control which closely resembles this type of curve, independently of the engine used, is "base + inc".
40/X simply cannot be adapted to play that well, the moves 41-80 are FAAAR less important than the moves 1-40. Seeing the result, independently of engine, "base + inc" comes as no-brainer as optimum time usage at fixed total time. The details sure might depend on the engine, how is the shape in moves 1-20, how fast is the decay of time usage towards the tail function of material on the board, complexity of positions, etc. With a very good time usage implementation, even "total time per game" can become very competitive, to being equal to the best of "base + inc". But not 40/X, one can do nothing with the time usage implementation in his engine in this case to approach the efficiency of a well chosen and implemented "base + inc" time usage.
I think that things may be dependent on the engine(at least in theory).
moves 10-20 are the most important in the game if the engine can improve the moves by using more time.

Engines with some late opening knowledge(I do not know if we have engines with enough knowledge at moves 11-20) may usually find the best moves at move 10-20 even without a lot of time so for these engines later moves may be more important.

lczero is relatively better in the opening and I guess also at moves 11-20 so I will not be surprised if for lc0 it is not so clear that moves 11-20 are the most important.

Gabor Szots
Posts: 505
Joined: Sat Jul 21, 2018 5:43 am
Location: Szentendre, Hungary
Full name: Gabor Szots

Re: CCRL adapted TC to current hardware

Post by Gabor Szots » Fri Jan 03, 2020 9:25 am

Guenther wrote:
Thu Jan 02, 2020 9:21 am
Gabor Szots wrote:
Thu Jan 02, 2020 8:28 am
I am very much in favour of incremental clock which in my case (I currently test at 40/2) would be 4+2, with a ratio of 120, while approximately keeping total game time the same.

...
I guess this was a typo and you meant 2+2 instead of 4+2?
No, it was not, only I mixed up the old and the current TC's.
The approximate length of my games is 70 moves. With 4+2 it adds up to 6 minutes and 20 seconds, a bit faster than with 40/4, the old reference time control. With the current reference TC of 40/2 I would use 2+1.
Gabor Szots
CCRL testing group

Post Reply