Author: Robert Hyatt
Date: 05:12:46 11/24/98
Go up one level in this thread
On November 24, 1998 at 07:35:38, Vincent Diepeveen wrote: >On November 24, 1998 at 03:02:51, David Blackman wrote: > >>On November 23, 1998 at 09:20:43, Robert Hyatt wrote: >> >> >>>It's not as bad as it sounds, because remember that I said to search the *other* >>>moves to D-2. If the fail high search costs you X nodes, then the extra tests >>>on the moves you would normally look at will only cost you w/w^(D=2) nodes >>>which lets you constrain the cost quite a bit. It certainly isn't free, but >>>it doesn't double the size of the tree or anything close to that... >>> >>>ie if your normal branching factor (W) is 36, then the tree will grow by a >>>factor of about 1/36... not too bad.. >> >>Are you sure about that? Given that you're still using alpha-beta i'd have >>thought the cost of checking the extra moves was about >>w * w^((D-2)/2) >>compared to finding the initial fail high at >>w^(D/2) >>ie, about the same. >>Although i could be missing something due to failing high and failing low having >>different costs. >> >>Certainly, when i attempted singular-move extensions a few years back i found >>that detecting singular moves was often more expensive than doing the extension. >> >>However i only did a one-ply extension. I think Deep Blue does a 2 ply extension >>in most cases. >> >>I briefly considered always doing the extension, without checking if it was >>singular, but then i decided that was silly. >> >>I guess you could reduce the cost of detecting singular moves a bit by cutting 3 >>or more plies off the depth, instead of just 2. >> >>I think i tried that at one stage. It's nice for tactical problem sets, but not >>clear if it actually helps in real games. > >Doing singular extensions reducing depth too much doesn't make sense, >because for tough problems you need a huge depth left to detect whether >it is a singular move. > >The most difficult singular moves are only singular after a huge depth left >to search. > >Meaning that if you reduce a lot, that you will miss an awfull lot of >interesting singular moves. > >A good example is in the bs2830 test set the move Qb7-d5, which with >singular extensions depth reduction = 0, could be found at say 6 ply. >Please turn off all other extensions, just use singular extensions, and >report back how deep it takes then! > >This problem is found by most programs not because of singular extensions, >but because of other extensions such as passed pawn extensions, recapture >extensions, and check extensions. > >If the combination of all extensions find it soon, then how much would it >take to find it WITHOUT singular extensions and with the extensions used? > >So there are 4 ways to test > a) all extensions off, SE on > b) all extensions on, SE on > c) all extensions on, SE off > d) all extensions off, SE off (which is the most uninteresting) > >The most interesting is to compare b with c. In fact S.E. should find this >really quickly, but when i test this with S.E. on, then i'm suffering horrible >from the reduction factor, which misses the advance of the c-pawn. > >In fact i find this move easier then without S.E. than with, as it eats >up too many plies extending nonsense. > >here is the position (bs2830-26): > >1r4k1/1q2pN1p/3pPnp1/8/2pQ4/P5PP/5P2/3R2K1 b - - Qd5 Did you implement the "real" singular extension algorithm as explained in the JICCA along with the "sticky transposition table" stuff to avoid losing a singularity at odd times? I did this in Cray Blitz and found it worked pretty well in the right positions...
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.