Author: Bruce Moreland
Date: 10:16:42 06/03/01
Go up one level in this thread
On June 03, 2001 at 08:12:09, Gian-Carlo Pascutto wrote: >On June 02, 2001 at 17:44:38, Bruce Moreland wrote: > >>I do a test to see if there is apparently only one move that doesn't suck. I >>do a reduced depth search. If the first few don't drive the score to something >>close to alpha, I leave, because I assume this is a fail-low node. If one of >>them drives the score pretty close to alpha, I check the rest to see if there >>are any more that do this. > >Ok, I assume you do this before starting the normal search, but >after nullmove etc? Without looking at the source, I don't know. It's fair to try it both ways, if that makes sense. >Before starting the normal search, you search every move with R=2/R=3 >or something similar, and check whether there is one single move >that produces a score close to alpha, whereas the remainder is >clearly worse. You speed this up by exiting if neither of the first >few moves gets near alpha. You can probably also speed this up >(I have no idea whether you do it or not) by, after you have found >a 'potential singular' move, exiting if neither of the following >moves gets near alpha too. I don't remember what R I use, but I picked an R through experimentation which others can also do. The philosophy is simply to see if there is only one obvious way to avoid disaster, and if so, to investigate this further. >>If there's only one move that looks decent, I call it singular and extend it a >>ply. > >A full ply? No fractional increments? This should also be the subject of experiments. >>This doesn't work in the endgame, but in the middlegame I found it >>improved solution times in the ECM suite dramatically, while costing a ply in >>the general case. > >Intresting. So you found a definite tactical strength increase? >Any numbers on WAC? WAC isn't a good suite once you've made your engine capable. If you run it at a minute per test, on decent hardware, you'll get at least 296 of them right. Reducing time doesn't change this much. I use the ECM suite because it is harder, so making the program faster or better results in more solutions. ECM has bugs, which I ignore, since even with them it's still the best tactical suite. >>This extension made my search pretty unstable so I had to do some of the hash >>table stuff mentioned in the DT article on singular extension published in the >>ICCAJ some years ago. > >Ok. Do you also hash nodes where you did not find a singular move, >i.e. the ICCAJ paper only mentions hashing positive results with >stickyness, so even if the move is no longer singular we still >pretent it is after a while. You could do the same for non-singular >nodes. Do you? I do something approximating what is mentioned in that paper. I don't mean to be a butthead, but I'm not going to go look at my stuff and lay it out here in pseudo-code. If I'm going to do that, I'd at least do it for a paper or something. I don't mind helping amateurs with stuff, since they will credit what they get from me. There are others who will take without crediting, and I don't want to be responsible for inflating their myth. >>The version that played in the WCCC 1999 used this. Don Beal came around >>collecting program information and I told him then, so it may have already been >>published in an ICCAJ. > >I don't have access to those, unfortunately :( > >Sorry for all the questions, but you're the only person I know >of that is currently using singular extensions without seeing >a noticable strength decrease. Since I've been trying to get >them to work in Crafty for a while, this is very interesting. > >Mike Byrne sent me some code from Robert from 2 years ago that, >as far as I can tell, pretty much seems to do what you describe, >but it's got some weird things in there too, and I don't fully >understand them. I think that Bob started messing with it because I told him that I was messing with it and liked it. I am not sure if he had success with it or not. It's hard to get it to do anything other than suck plies. I wrote a fairly direct implementation and it worked alright after some fiddling. I didn't do everything the DT guys did. I did it as cheaply as I could. bruce > >-- >GCP
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.