Uri Blass wrote:diep wrote:
Actually i invented reductions back in 1998 and entire 1999 Diep used reductions. LMR and History pruning is a subset of what i experimented with back then. Using history moves for reductions or even as a selective move mechanism was used in the 90s massively also by Said Koudache in his 10x10 international checkers program.
Many authors independant from each other invented reductions and some used it with more succes for their engine than others. Back then in 90s the nps of diep wasn't high enough to really profit from what i invented.
Vincent
LMR helps at all time control including 1 second per move with single cpu hardware that is not the fastest single cpu hardware.
It means that
one million nodes per move are clearly enough to make LMR productive.
I am sure that Diep at 3 minutes per move searched more than 1 million nodes even in 1998 or 1999 and it means that what you invented that did not work for you was different than LMR that top programs use today(for example different conditions when to use LMR).
I believe that some authors like the author of shredder invented some productive LMR independently and if you did not find LMR to be productive for you it means that you did not invent it.
Uri, back in 2004 you got with your own engine 12 ply or something at a time control of 2+ minutes a move and your engine also has a branching factor that is so bad, and so inefficient that any form of pruning seems to work for you.
LMR is a gift from heaven for programmers who do not have efficient branching factors without it. If you turn off LMR in Rybka and the clones, then you end up with an engine that has a far worse branching factor than Diep without.
Diep searched in 1996 at a pentiumpro 200Mhz between 5k and 10k nps.
Search depths at world champs 1997 at a 300Mhz PII were around 7 ply to 10 ply in endgame.
Branching factors of 10.0 were very common back then. You can still see this in some old programs from then like Schach 3.0.
In 1998 i did improve the branching factor of diep bigtime. Better move ordering, which with a big evaluation function is more complicated, and so on.
Sure crafty didn't have a b.f. of 10.0 back then, but was tactical extremely weak. Also its qsearch was really lobotomized and picked up zero tactics. Not even being mated.
Doing claims now that LMR in 1999 at a 450Mhz PII which i owned by then would have worked is just showing your simplistic insight in computerchess which is totally flawed as you have no clue about search.
You posted yourself positions back then where reductions are total contraproductive against and your only judgement back then was 100% based upon whether engines found this quickly or not.
As for Diep with reductions back then i could win about 2 ply, yet many positions, not to mention tactical positions, it lost 4-5 ply.
So there is a break even trajectory there.
It was tested playing games simply.
I remember how Stefan Meyer-Kahlen and i at a hotelroom turned off for a test the reductions of Shredder.
It just got 9 to 10 ply suddenly. This was in 2004 at his 3ghz laptop.
Branching factors of most of the beancounters are so bad without reductions that they can't even remotely compete if you kick it out of those engines.
A huge test of Diep in 2001 using a more selective prunings mechanism like Ed Schroeder has in Rebel, and which he didn't clearly publish (if i remember well the official reason given for that is that Tiger also is using it and that he therefore doesn't want to publish that), which did search a lot deeper, in fact it won 3 ply, using of course reductions in a more sophisticated manner than LMR is doing, it is finding the tactics actually.
So his scheme has the advantage of searching deeper like LMR, yet you do not miss too much tactics.
Jan Louwman, who did do several tests for Diep, at the time played 1000 games at 36 computers. Time control around 3 hours a game at the fastest machines, which was back then k7's and at P3's and laptop P3's of around 600-800Mhz or so we put the time control up to 9 hours a game and auto232 was used against another machine. So 18 sets of auto232.
It took quite a while to play 1000 games.
500 with this form of search and 500 games without.
The version with the selectivity scored 20% less.
You simply are not intelligent enough to imagine other realities than the only thing you notice at 1 program. You just see 1 reality of yourself toying with 1 program.
I bet you never do experiments with other programs, as only the thing right now winning matters for you.
You probably didn't even figure out why history pruning worked for Fruit.
As for Diep, LMR which is turned on right now is basically working big crap. It hardly gives elo. So i intend to replace it with a different form of selective search that i self invented - though i'm sure Stefan MK and others have invented something similar as well. I'm going back to an algorithm i invented in fact around 1996 which didn't work back then as it had an overhead of 20% which was too much back then. I hope it works now and gets diep a bigger search thanks to getting a bigger nps nowadays.
You really have no idea about search Uri.
Vincent