Looking at this http://www.top-5000.nl/eval.htm
It seems :
- PSQT are worth 70 elo
- king safety 43 elo
- mobility 118 elo
- pawn eval (?) 15 elo
- passed pawn 97 elo
- piece pair / outpost / center control 33 elo
- other stuff 63 elo
So evaluation features seems to worth 440 elo
Let's say a depth more in search is worth 60 elo, and BF is 2. Making a x4 speed up lead to another 120 elo.
So starting from a simple, reasonably fast engine around elo 2000, I feel that 2600 is already quite difficult to achieve (rebel / prodeo is indeed around 2700)! How can some engine reach 3000 ? Is it just years and years of fine tuning and speed up adding 5/10 elo each ?
About a rebel evaluation study
Moderators: hgm, Rebel, chrisw
-
- Posts: 1871
- Joined: Sat Nov 25, 2017 2:28 pm
- Location: France
-
- Posts: 276
- Joined: Sat Mar 04, 2017 12:24 pm
- Location: Hungary
Re: About a rebel evaluation study
Yes, tuning is very important, even including search.
Search must also be in balance with eval.
eg:
Aspiration window is significant better with a "perfect" eval. (Score does not fluctuate)
I dont no what use Rebel, but the modern search give lots of Elo.
Threats is more important, than rebel says..if you make that well.
tomitankChess written in JavaScript. JavaScript is ~10x slower than C languages..
Nonetheless the next version it will be over 2600 Elo.
(With manual tuning...machine learning would be better.)
Search must also be in balance with eval.
eg:
Aspiration window is significant better with a "perfect" eval. (Score does not fluctuate)
I dont no what use Rebel, but the modern search give lots of Elo.
Threats is more important, than rebel says..if you make that well.
tomitankChess written in JavaScript. JavaScript is ~10x slower than C languages..
Nonetheless the next version it will be over 2600 Elo.
(With manual tuning...machine learning would be better.)
-
- Posts: 2204
- Joined: Sat Jan 18, 2014 10:24 am
- Location: Andorra
Re: About a rebel evaluation study
Yes. Also less than 5 elo patches. For Andscacs maybe I have done 8000-10000 versions since sep 2013.xr_a_y wrote: How can some engine reach 3000 ? Is it just years and years of fine tuning and speed up adding 5/10 elo each ?
Daniel José - http://www.andscacs.com
-
- Posts: 10314
- Joined: Thu Mar 09, 2006 12:37 am
- Location: Tel-Aviv Israel
Re: About a rebel evaluation study
I think that you cannot do a math like this.xr_a_y wrote:Looking at this http://www.top-5000.nl/eval.htm
It seems :
- PSQT are worth 70 elo
- king safety 43 elo
- mobility 118 elo
- pawn eval (?) 15 elo
- passed pawn 97 elo
- piece pair / outpost / center control 33 elo
- other stuff 63 elo
So evaluation features seems to worth 440 elo
Let's say a depth more in search is worth 60 elo, and BF is 2. Making a x4 speed up lead to another 120 elo.
So starting from a simple, reasonably fast engine around elo 2000, I feel that 2600 is already quite difficult to achieve (rebel / prodeo is indeed around 2700)! How can some engine reach 3000 ? Is it just years and years of fine tuning and speed up adding 5/10 elo each ?
The value of new stuff is dependent on the value of older stuff that you have so there is no meaning to the claim that PST worth 70 elo.
Edit:I can add that the best PST is dependent on the other stuff that you have and is not constant.
Edit 2:I can add that I believe that only piece square table with good search should have already 2600 elo
-
- Posts: 349
- Joined: Sat Aug 06, 2016 8:31 pm
- Location: United States
Re: About a rebel evaluation study
I really hope that's not true considering all the work, including eval tuning, I've put into creating an almost 2400-rated engine.Uri Blass wrote: Edit 2:I can add that I believe that only piece square table with good search should have already 2600 elo
-
- Posts: 10314
- Joined: Thu Mar 09, 2006 12:37 am
- Location: Tel-Aviv Israel
Re: About a rebel evaluation study
I feel sure it is correct.zenpawn wrote:I really hope that's not true considering all the work, including eval tuning, I've put into creating an almost 2400-rated engine.Uri Blass wrote: Edit 2:I can add that I believe that only piece square table with good search should have already 2600 elo
I remember that in the past I changed strelka's evaluation to only piece square table evaluation and I believe that I got something that was better than 2400.
It could easily beat Joker that is 2300 elo engine in most games(so I think it was 2400-2500 elo)
http://www.computerchess.org.uk/ccrl/40 ... ker_1_1_14
Note that I did not try to optimize Strelka's piece square table and I believe that strelka's search is also weaker than the search of today top programs.
-
- Posts: 349
- Joined: Sat Aug 06, 2016 8:31 pm
- Location: United States
Re: About a rebel evaluation study
Depressing.
-
- Posts: 388
- Joined: Sun Dec 21, 2008 6:57 pm
- Location: Washington, DC
Re: About a rebel evaluation study
You can say that again!zenpawn wrote:Depressing.
-
- Posts: 1871
- Joined: Sat Nov 25, 2017 2:28 pm
- Location: France
Re: About a rebel evaluation study
Wow 2600 just with psqt ... So the search prunings / razorings must be highly tuned. Is there such an opensource engine with a basic evaluation but that is higher than 2500 elo to get some inspiration ?
-
- Posts: 2204
- Joined: Sat Jan 18, 2014 10:24 am
- Location: Andorra
Re: About a rebel evaluation study
Take Andscacs or Stockfish and remove most of the stuff in the eval function.xr_a_y wrote:Wow 2600 just with psqt ... So the search prunings / razorings must be highly tuned. Is there such an opensource engine with a basic evaluation but that is higher than 2500 elo to get some inspiration ?
Daniel José - http://www.andscacs.com