Author: Vincent Diepeveen
Date: 11:33:23 12/29/02
Go up one level in this thread
On December 28, 2002 at 14:55:16, Gerd Isenberg wrote: >On December 28, 2002 at 14:00:05, Vincent Diepeveen wrote: > >>On December 28, 2002 at 12:31:01, Alessandro Damiani wrote: >> >>>On December 28, 2002 at 12:15:46, Vincent Diepeveen wrote: >>> >>>>On December 28, 2002 at 11:18:58, Alessandro Damiani wrote: >>>> >>>>>On December 28, 2002 at 10:10:53, Ricardo Gibert wrote: >>>>> >>>>>>It seems Ed Schröder has added a bit more to his web page: >>>>>> >>>>>>http://members.home.nl/matador/chess840.htm#SEARCH >>>>>> >>>>>>I have the feeling that ES will be getting a lot of thank yous for quite awhile >>>>>>for his fine contributions to computer chess. >>>>>> >>>>>>Thanks again! >>>>> >>>>>It seems to me that there are two typos in the following code: >>>> >>>>it's not about the source code. It's about the idea. >>>>Any sort of pseudo code gets accepted then. Definitely >>>>by me. >>> >>>I am talking about the pseudo code Ed published on his homepage (pseudo code is >>>still code). I don't understand your statement in this context. >>> >>> >>>> >>>>Of course do not forget that these reductions are very dangerous >>>>to use in combination with nullmove and that as Ed describes them >>>>they completely rape your hashtable. A depth stored as 'n depth left' >>>>might be in reality n+1. You need to add a bit and some code for >>>>transpositions to the hashtable in order to fix that. >>>> >>> >>>My new variant of ABC uses depth reductions instead of extensions. So I am >>>looking at the difference between Ed's reductions and mine. I don't use >>>conventional null-move yet. >>> >>>There is a doctoral thesis by Thomas Barth which describes how depth reductions >>>work fine with a hashtable. His work is from 1988. >>> >>>Alessandro >> >>I do not know the work of Barth here, but i know that without >>modifying the hashtable you run into trouble a lot. >> >>In the 90s several publications from Feldmann ignored the hashtable >>problem with his Fail High reductions. >> >>Note that though the implementation detail of Ed seems small compared >>to the FHR concept of Feldmann, the implementation of Ed is practically >>working a lot better. > >FHR is based on nullmove observation. Rather than doing a nullmove-cutoff they >reduced the depth. Isn't Ed's approach more like Heinz's "limited razor"? > >Gerd FHR is buggy allright. No It's not similar to Heinz's limited razor. The major difference is where you prune. Heinz is doing most of his pruning near the leafs. That influences the evaluation bigtime there. Ed is doing this single reduction within the main search so your leafs don't hurt too much from it, except that we must of course not forget that most of us replace the nullmove search last 3-4 ply with a nullmove. Further the reality is that pruning comes down to which conditions you use. If you do not post 100% what conditions you used you can conclude anything about the pruning. This is a major problem which some people will never understand. Ed has it perfectly tuned. That 5 pawns is a very clever thing. I have written down on paper here many positions now and so far everything should get get found on the same search depth. The very clever thing is to do the pruning *after* you just make a move and demand conditions on that move itself. So compared to pruning on the top of a search you now first reduce depth and then try to transpose in hashtable. So the vaste majority of prunings that try to prune on the top of a search already get beaten here. Then a big majority of positions which could feal near the leaves don't misues this pruning either because you demand 3 conditions on giving check, being in check and not being a capture. Then the 5 pawn value takes care in combination with the conditions that i so far didn't find much which suffers in theory already a ply. >> >>It's conceptual the same thing of course. I can't remember so quickly >>whether Feldmann did them also non recursively.
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.