Interesting read
Posted: Thu Jul 12, 2012 11:33 pm
"i'm the last to even have time to reverse engineer anyone. First of all i'm one of those who invented a bunch of algorithms himself and several of my tricks boosted the elo of Rybka,
And no i didn't like that.
To give one example of what i invented somewhere start 21th century and which i used in Diep for nullmove:
if( static eval >= beta + S ) then R = R+1;
I'm the one inventing that. I never had the machine power to test whether it worked well at bigger search depths. First of all diep never got those depths.
I just put it in Diep and for example world champs 2004 and onwards i played with it.
How did it end up in Rybka?
I don't know. Someone illegally took it over from me obviously.
In end 90s i invented extension, for some who were around back then, tactics was important mid 90s. It solved lots of tactical positions back then for Diep. Realize back then Diep got just a few plies. So did nearly everyone. Crafty got bigger depths end 90s, but didn't see anything tactical (cheapo quiescencesearch).
One extension i used there to big effect, it's very interesting to mention that 1 year after i put it into effect, that suddenly Rebel also had this extension.
I never posted about it.
Oh rybka rybka. How surprisingly. Ordering in root. Whether it delivers elo i don't know, as i never did do it differently since mid 90s, but how i order moves there, that's also how Rybka 3 is doing it. Not sure whether the clones took it over from rybka3 - didn't check. I never posted this anywhere.
Oh back onto Ivanhoe. It's doing up to a ply or 9-10 futility potentially.
Practical most futility will happen last 3-4 ply or so.
As it's exponential bonus to search depth, it will keep reducing of course and win plies.
It's total trivial that doing this futility to 9 deep loses elo.
It's needless to say that loses massive elo. I never understood all those guys doing futility. In Diep i can't use it at all. If you want to do something,
then just do a cheaper nullmove last 3 - 4 plies. Not deeper than that or you will miss tactics. This is on my todo list to build for Diep as well. A big attempt there failed in 2005 when i tried to replace my entire search last few plies with a fast highspeed simple search. It's easy to see for me that this is where ivanhoe messes up versus houdini.
Now of course you search deeper doing that futility yet it backfires.
As for the extension. Houdini sees tactics faster where it involves giving check.
As the evaluation is 100% similar, very obvious to see for chessplayers, and it solves all kind of the same positions like 4-5 ply sooner than ivanhoe, meanwhile ivanhoe searches 2 ply deeper.
It's obviously tactical better.
If you go play your own engine against another version of your own engine and one of the engines is tactical better, it will win always in selfplay.
That's kind of what happens at most rating lists now, as most 'top engines' are total clones.
This will change when one guy has fixed his evaluation function enough and gets some reasonable depth without messing up too much in search.
Ivanhoe just messes up in search further it's identical to Houdini, as evaluation is the same.
It's total trivial to me that biggest difference of Diep versus the clones is in opening. If i manage to fix that, a few plies more than the turtle gets now is enough for world domination - and i'm not interested in running single core contests nor interested in rapid levels at a quadcore from half a decade ago of course.
Fixing all the old junk knowledge to modern well debugged knowledge is NOT easy.
In a given position positionally the thing is doing great, yet in opening, just simply developing quick is strategically nearly always better. That's where the deep searchers pick up massive elo, including KOmodo of course.
Now you shouldn't lose that advantage - yet it's obvious that if you want besides just deep mainline checking more elostrength you also must pick up tactics.
With simple methods you can do that.
Now i didn't check komodo, but if your material values are DIFFERENT from the rybka/fruit clones, so not nearly exactly the same, and if it also has knowledge that's NOT inside ivanhoe and clones, then you'll already see a DIFFERENT behaviour there than Houdini will have versus the other clones that are so much the same in evaluation.
The easiest way to beat someone who is identically evaluating things, is by being tactical stronger.
As soon as someone has a different evaluation function, that tactical stronger matters less.
What still wins search depth meanwhile not getting tactical too much stronger is a cheapo nullmove last 3-4 plies. Cheapest thing you can possibly do there is razoring.
if( !incheck && eval >= beta + 1 pawn ) then return eval;
making things tactical stronger there- really everything has been invented there already.
A simple trick is to check whether you were in check +2 plies direction root.
Then just do normal nullmove with qsearch, provided that nullmove also detects giving checks.
With diep's small search depths, such things lose it too much elo for now.
So i just use normal nullmove there.
*How did it end up in Rybka?
I don't know. Someone illegally took it over from me obviously. *
Vincent Diepeveen"
And no i didn't like that.
To give one example of what i invented somewhere start 21th century and which i used in Diep for nullmove:
if( static eval >= beta + S ) then R = R+1;
I'm the one inventing that. I never had the machine power to test whether it worked well at bigger search depths. First of all diep never got those depths.
I just put it in Diep and for example world champs 2004 and onwards i played with it.
How did it end up in Rybka?
I don't know. Someone illegally took it over from me obviously.
In end 90s i invented extension, for some who were around back then, tactics was important mid 90s. It solved lots of tactical positions back then for Diep. Realize back then Diep got just a few plies. So did nearly everyone. Crafty got bigger depths end 90s, but didn't see anything tactical (cheapo quiescencesearch).
One extension i used there to big effect, it's very interesting to mention that 1 year after i put it into effect, that suddenly Rebel also had this extension.
I never posted about it.
Oh rybka rybka. How surprisingly. Ordering in root. Whether it delivers elo i don't know, as i never did do it differently since mid 90s, but how i order moves there, that's also how Rybka 3 is doing it. Not sure whether the clones took it over from rybka3 - didn't check. I never posted this anywhere.
Oh back onto Ivanhoe. It's doing up to a ply or 9-10 futility potentially.
Practical most futility will happen last 3-4 ply or so.
As it's exponential bonus to search depth, it will keep reducing of course and win plies.
It's total trivial that doing this futility to 9 deep loses elo.
It's needless to say that loses massive elo. I never understood all those guys doing futility. In Diep i can't use it at all. If you want to do something,
then just do a cheaper nullmove last 3 - 4 plies. Not deeper than that or you will miss tactics. This is on my todo list to build for Diep as well. A big attempt there failed in 2005 when i tried to replace my entire search last few plies with a fast highspeed simple search. It's easy to see for me that this is where ivanhoe messes up versus houdini.
Now of course you search deeper doing that futility yet it backfires.
As for the extension. Houdini sees tactics faster where it involves giving check.
As the evaluation is 100% similar, very obvious to see for chessplayers, and it solves all kind of the same positions like 4-5 ply sooner than ivanhoe, meanwhile ivanhoe searches 2 ply deeper.
It's obviously tactical better.
If you go play your own engine against another version of your own engine and one of the engines is tactical better, it will win always in selfplay.
That's kind of what happens at most rating lists now, as most 'top engines' are total clones.
This will change when one guy has fixed his evaluation function enough and gets some reasonable depth without messing up too much in search.
Ivanhoe just messes up in search further it's identical to Houdini, as evaluation is the same.
It's total trivial to me that biggest difference of Diep versus the clones is in opening. If i manage to fix that, a few plies more than the turtle gets now is enough for world domination - and i'm not interested in running single core contests nor interested in rapid levels at a quadcore from half a decade ago of course.
Fixing all the old junk knowledge to modern well debugged knowledge is NOT easy.
In a given position positionally the thing is doing great, yet in opening, just simply developing quick is strategically nearly always better. That's where the deep searchers pick up massive elo, including KOmodo of course.
Now you shouldn't lose that advantage - yet it's obvious that if you want besides just deep mainline checking more elostrength you also must pick up tactics.
With simple methods you can do that.
Now i didn't check komodo, but if your material values are DIFFERENT from the rybka/fruit clones, so not nearly exactly the same, and if it also has knowledge that's NOT inside ivanhoe and clones, then you'll already see a DIFFERENT behaviour there than Houdini will have versus the other clones that are so much the same in evaluation.
The easiest way to beat someone who is identically evaluating things, is by being tactical stronger.
As soon as someone has a different evaluation function, that tactical stronger matters less.
What still wins search depth meanwhile not getting tactical too much stronger is a cheapo nullmove last 3-4 plies. Cheapest thing you can possibly do there is razoring.
if( !incheck && eval >= beta + 1 pawn ) then return eval;
making things tactical stronger there- really everything has been invented there already.
A simple trick is to check whether you were in check +2 plies direction root.
Then just do normal nullmove with qsearch, provided that nullmove also detects giving checks.
With diep's small search depths, such things lose it too much elo for now.
So i just use normal nullmove there.
*How did it end up in Rybka?
I don't know. Someone illegally took it over from me obviously. *
Vincent Diepeveen"