I get best score in WAC test when Skipper only counts material. (Actually I only use fifty test positions)
All other evaluations only give a slow down or don't give a better score.
Time to create a better quick test or is counting material only, the best way to go.
WAC again
Moderators: hgm, Rebel, chrisw
-
- Posts: 570
- Joined: Mon Jul 20, 2015 5:06 pm
Re: WAC again
Do you have source code available? Maybe it is time for you to get some help with your Skipper program.Henk wrote:I get best score in WAC test when Skipper only counts material. (Actually I only use fifty test positions)
All other evaluations only give a slow down or don't give a better score.
Time to create a better quick test or is counting material only, the best way to go.
-
- Posts: 7216
- Joined: Mon May 27, 2013 10:31 am
Re: WAC again
No then I better quit.D Sceviour wrote:Do you have source code available? Maybe it is time for you to get some help with your Skipper program.Henk wrote:I get best score in WAC test when Skipper only counts material. (Actually I only use fifty test positions)
All other evaluations only give a slow down or don't give a better score.
Time to create a better quick test or is counting material only, the best way to go.
-
- Posts: 7216
- Joined: Mon May 27, 2013 10:31 am
Re: WAC again
Adding Skippers center control term to evaluation makes it perform significantly better on STS(v12.0) Center Control test. It makes it perform worse on WAC test but not that much.
So I think I keep the center control term.
So I think I keep the center control term.
-
- Posts: 1357
- Joined: Wed Mar 08, 2006 10:15 pm
- Location: San Francisco, California
Re: WAC again
First, I can't imagine why you would only use 50 positions when running the entire 300 positions takes less than an hour, even if you use 10 seconds per move.
Secondly, your issue does make some sense, since almost all of WAC's positions are tactical in which the best line wins material or gives checkmate. So a material only eval COULD perform better.
Thirdly, I can understand why you don't want to publish your source. I also haven't done so simply because I prefer not to invite ridicule from the entire computer chess community. However, privately sending your code to one or two people from this board will prove immensely helpful. I did that and was able to quickly get several hundred ELO thanks to the MANY helpful comments.
Secondly, your issue does make some sense, since almost all of WAC's positions are tactical in which the best line wins material or gives checkmate. So a material only eval COULD perform better.
Thirdly, I can understand why you don't want to publish your source. I also haven't done so simply because I prefer not to invite ridicule from the entire computer chess community. However, privately sending your code to one or two people from this board will prove immensely helpful. I did that and was able to quickly get several hundred ELO thanks to the MANY helpful comments.
-
- Posts: 759
- Joined: Fri Jan 04, 2013 4:55 pm
- Location: Nice
Re: WAC again
There is a problem when you use just a set of positions instead of playing thousands of games : the engine never improve well .JVMerlino wrote:First, I can't imagine why you would only use 50 positions when running the entire 300 positions takes less than an hour, even if you use 10 seconds per move.
Secondly, your issue does make some sense, since almost all of WAC's positions are tactical in which the best line wins material or gives checkmate. So a material only eval COULD perform better.
Thirdly, I can understand why you don't want to publish your source. I also haven't done so simply because I prefer not to invite ridicule from the entire computer chess community. However, privately sending your code to one or two people from this board will prove immensely helpful. I did that and was able to quickly get several hundred ELO thanks to the MANY helpful comments.
Skipper's programmer do that since more than 2-3 years and his engine never improve .
Also no releases yet .Releases may help when more than one guy use your engine lol.
He won't release his code , it's a choice, it's also my choice for the moment .But sure it can be good to send his code to one or two experts and of course , listening their feedback
Isa download :
-
- Posts: 1357
- Joined: Wed Mar 08, 2006 10:15 pm
- Location: San Francisco, California
Re: WAC again
Absolutely agreed that games are far better than test positions. But the WAC test, among other test suites, are still useful. Particularly WAC because it's fast to run and can help you easily spot when you've broken something significant.
-
- Posts: 12538
- Joined: Wed Mar 08, 2006 8:57 pm
- Location: Redmond, WA USA
Re: WAC again
Improving WAC scores probably won't make a chess engine play stronger.
The main purpose for a test like that is to show missing or poorly implemented features in analysis like king safety, x-ray, etc.
If an engine does better just doing a wood-count, then something is wrong with the eval, since a large percentage of the moves require a sacrifice (though search may compensate for that if {for instance} king safety triggered an extension).
Turning off eval features when testing with WAC is what I would describe as its anti-purpose.
The main purpose for a test like that is to show missing or poorly implemented features in analysis like king safety, x-ray, etc.
If an engine does better just doing a wood-count, then something is wrong with the eval, since a large percentage of the moves require a sacrifice (though search may compensate for that if {for instance} king safety triggered an extension).
Turning off eval features when testing with WAC is what I would describe as its anti-purpose.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
-
- Posts: 7216
- Joined: Mon May 27, 2013 10:31 am
Re: WAC again
Trying to find quick test that improves Skippers play. Of course best to quit for there are other things to do and we already have a Stockfish.
-
- Posts: 490
- Joined: Tue Feb 04, 2014 12:25 pm
- Full name: Colin Jenkins
Re: WAC again
Hi Henk,
I hesitate to say this because Lozza is no champion, but watching skipper play at the last HGM tourney, to me, it looked like something more fundamental was wrong. Skipper moved her queen about 10 moves in a row at the opening, which seemed weird.
Just a thought...
When I was developing Lozza, I initially used the Simplified Evaluation Function only, essentially Material+PSTs, but designed to be used alone - and added move ordering, Hash, ID, Null move, LMR, futility etc etc (all the control stuff) until Lozza could consistently beat FairyMax.
Given Lozza is slow (JIT compiled Javascript) it follows that if you simplify Skipper's eval to just the Simplified Evaluation Function and then bug fix you should be able to get to the same (consistently beat FairyMax) - and *then* fiddle with the evaluation function.
https://chessprogramming.wikispaces.com ... n+function
Lozza: http://op12no2.me/toys/lozza/lozza.js
Or I can make a copy of Lozza that is just the simplified eval if you llke and using the same logic, a bug-free Skipper using the same simplified eval must be able to beat Lozza over a couple of hundred games say (because Lozza is inherently slow). Then when it can, start tweaking the eval for real.
I hesitate to say this because Lozza is no champion, but watching skipper play at the last HGM tourney, to me, it looked like something more fundamental was wrong. Skipper moved her queen about 10 moves in a row at the opening, which seemed weird.
Just a thought...
When I was developing Lozza, I initially used the Simplified Evaluation Function only, essentially Material+PSTs, but designed to be used alone - and added move ordering, Hash, ID, Null move, LMR, futility etc etc (all the control stuff) until Lozza could consistently beat FairyMax.
Given Lozza is slow (JIT compiled Javascript) it follows that if you simplify Skipper's eval to just the Simplified Evaluation Function and then bug fix you should be able to get to the same (consistently beat FairyMax) - and *then* fiddle with the evaluation function.
https://chessprogramming.wikispaces.com ... n+function
Lozza: http://op12no2.me/toys/lozza/lozza.js
Or I can make a copy of Lozza that is just the simplified eval if you llke and using the same logic, a bug-free Skipper using the same simplified eval must be able to beat Lozza over a couple of hundred games say (because Lozza is inherently slow). Then when it can, start tweaking the eval for real.