Author: Vincent Diepeveen
Date: 09:22:43 09/01/01
Go up one level in this thread
On September 01, 2001 at 09:55:25, Robert Hyatt wrote: >On September 01, 2001 at 08:48:00, Vincent Diepeveen wrote: > >> >>You are entirely right Ed, i have singular extensions inside diep now >>and play with them turned on tournaments now. first tournament i played >>with them turned on was back in 1994 the dutch open championship, >>but my implementation sucked bigtime there. Then in paderborn 2001 i >>used a better implementation with big reduction factor (R=3) and >>also in combination with other extensions. The reason i have them >>inside diep now is >> a) diep doesn't search very deeply so overhead isn't too big then >> b) to solve testpositions quicker otherwise i need days to solve things >> >>However what i notice is that in mainlines in complex positions the >>value of singular extensions is very limited. Of course if i'm already >>won i see mates way before my opponent sees them (in middlegame) or >>i see a win way before my opponent sees it, as well as that Rxf7 move >>which ferret played against gandalf i see within seconds with within 90 >>seconds the right score, but after all the only impact of singular extensions >>is that they give a psychological good feeling "i'm not going to lose >>because of a cheap trick if my program messes up". Of course combinations >>can only be there and getting outsearched is only important if a program >>plays completely anti positional chess. >> >>In normal game play and in sound positions, there the value of SE and similar >>extensions gets hugely overrated i think. > > >I don't believe they _hurt_ when done right. The _last_ report by the DB >team suggested that SE was worth maybe 10-20 Elo rating points at most. Their >first report suggested far greater improvement, but this was because it did far >better on test suites. In matches, it won, but the margin of victory was much >less than the test suite results suggested... The problem from SE is that i can use them right now with diep, but in the future with bigger search depths, they simply cost too much. My branching factor from 2.5 to 3.5 normally spoken gets with SE hugely above 3 and goes up to 5 or something sometimes. A big problem is when the rootscore changes. Researches are simply too expensive. Normally the hashtable is that powerfull that it doesn't take too much. At 9 ply with R=3 i do for 5 ply singular extensions measured from root. So the root i don't do singular extensions, ply 2 from root i do, but then when search depth is 9 in total that means that i try testing at 2,3,4,5,6 and not at 1,7,8,9 Here it's for me about 10%, but at 10 ply it already gets way bigger, especially if its about king safety. At depths 11,12, then it can cost simply several ply these extensions, because if you extend for singular reasons, then usually after that there is an extra ply which again can trigger extensions. But if you go search 13, 14, 15 ply, then singular extensions are simply too expensive. You can watch the search depths from diep at the world champs easily that it gets to 12 ply quick but then it dies! Of course some failing high positions are not representive here. The real hard problems are positions where score goes up and down a bit. The classical NOLOT position there are loads of tactics but score goes only up for the side to move! >My tests in CB produced the same results. Modest improvements in games, some- >times wild improvements in test positions. But even though the improvement in >games is very modest, that one game here or there is _still_ very important. >That could be your game vs Kramnik in a big tournament, for example. This is why i have it still turned on, in order to give that push to an opponent who digs lots of holes and is busy avoiding his own dug holes.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.