Author: Robert Hyatt
Date: 08:25:18 09/28/04
Go up one level in this thread
On September 28, 2004 at 01:01:04, Stuart Cracraft wrote: >On September 27, 2004 at 23:56:17, Robert Hyatt wrote: > >>On September 27, 2004 at 23:48:56, Stuart Cracraft wrote: >> >>>On September 27, 2004 at 17:00:28, Robert Hyatt wrote: >>> >>>>On September 27, 2004 at 14:21:51, Stuart Cracraft wrote: >>>> >>>>>So when in PVS I've searched my first move, and it is a PV >>>>>move as well, with search returning value then I do a depth-2 >>>>>search on all other moves and if none -search(depth-2,-beta-MARGIN,- >>>>>alpha-MARGIN) <= -beta-MARGIN, where MARGIN is set to 3/4 of a pawn, >>>>>then the PV move is singular and I re-search it with depth instead of depth-1 >>>>>and use the returned value as my score against which to >>>>>measure all other non-PV moves against in the normal part >>>>>of the search, searching them to depth-1. >>>>> >>>>>Is the above wrong? >>>>> >>>>>Stuart >>>> >>>> >>>>It is wrong. >>>> >>>>Search the first move with the normal window. Search the _remaining_ moves with >>>>an offset window alpha-w, beta-w. If all still fail low (which they should do >>>>if the first move is best) then the first move is singular. re-search it again >>>>with a deeper search. If one of the remaining moves fails high on the offset >>>>window search, now you have a problem. Is this move better than the best move? >>>>If not the best move is not singular. But, this move could itself be singular >>>>so you have to test that hypothesis by re-searching the first move with a window >>>>lowered by the usual offset from the score returned by the second move that >>>>failed high. Repeat until sick or finished. >>>> >>>>You are describing what is done for determining singularity at fail-high nodes >>>>where only one move is normally searched before returning, but you do the >>>>"cheaper" searches to try to prove singilarity anyway and extend even a >>>>fail-high move one ply... >>> >>>Ugh -- sounds like a coding mess. I've saved your comments in a futurefile >>>but probably won't do much on SE until my computing resources are faster. >>>Giving up a ply for long-shots against humans is enticing of course as long >>>as the ply loss won't kill me due to very deep searching anyway -- that >>>won't happen until my box is 5x faster and that won't happen for awhile. >>> >>>So at this point not sick but "pending". >>> >>>Stuart >> >> >>It _is_ a mess. By the time you do both PV and FH singular tests, add in the >>sticky transposition table stuff, you see why I occasionally question a "certain >>poster" that claims (depending on the day of the week) that (a) SE is no good or >>(b) he has implemented SE and uses it. It is non-trivial to do if you do _real_ >>SE. > >I don't know about a poster but it certainly looks complex enough to make >me only want to implement the "cheapo" version that I heard Chess Genius >uses which I assume is some variant on the PV-only singular extension piece >and not for at least a year or few. I've never heard anyone but Vincent claim to know what Lang did/does in his chess program. I have heard he used some sort of singular extension. I have never seen any description of what he did. Dave Kittinger did PV singular in Wchess, but I believe he was doing the sort of "cheapo" approach I have described in the past, based on lots of conversations he/Bruce/I used to have on ICC several years ago...
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.