Author: Vincent Diepeveen
Date: 08:46:57 09/01/05
Go up one level in this thread
On August 31, 2005 at 00:14:32, Robert Hyatt wrote: >On August 30, 2005 at 11:49:22, Gian-Carlo Pascutto wrote: > >>On August 30, 2005 at 10:53:22, Alvaro Jose Povoa Cardoso wrote: >> >>>On August 30, 2005 at 08:43:56, Gian-Carlo Pascutto wrote: >>> >>>>On August 30, 2005 at 08:15:12, Alvaro Jose Povoa Cardoso wrote: >>>> >>>>>- You mean index = (from,to) pair taken from the current move at the parent >>>>>node? >>>> >>>>Yes. >>>> >>>>>- In the 64x64 table do we store complete moves (as they come from the move >>>>>generator) or just the (from,to) information? >>>> >>>>Doesn't matter. Your choice. You can test with noting the piece type that moved >>>>and encforcing that, or not doing that. >>>> >>>>-- >>>>GCP >>> >>>One more thing, how do we age this entries? >>>Do we simply let the search overwrite the entries? >> >>Why would aging be needed??? >> >>I guess you could make a countermove1, countermove2 etc just like one does with >>killers, if you really want to... >> >>-- >>GCP > > >Are you using this? > >I remember when it came out in the JICCA several years back (don't remember >exactly when). I tried it but never got any sort of improvement with it. Of >course I was using killers and the history heuristic already... > >Are you getting a tree reduction with it? Dappet team figured this idea out of countermove. It's pretty interesting idea. In those days it didn't work. Just like history heuristic at the time didn't work. With todays search depths of 12-15 ply it's time to redo a lot of that research again. Like Ed i'm doing a lot of tactical move ordering in Diep. Things like history heuristics and countermove seem to give less in that case. Note Ed is confusing killermoves with countermoves. This is no problem however, it's all the same type of ideas. Ed is speaking about a tactical killermove idea, in his case for mate scores. In Diep in the move ordering there is extensive mate ordering code. In Diep probably a bit more extensive than in other software is the case i'm focussing upon move ordering. I have to, Diep is just 100k nps at a single cpu. That's pretty little. A small disadvantage of all that tactical code is that it works especially well for positions where there is a trick to be found. Some years ago of course the focus was only at those positions. What i do nowadays is simply let Diep run automatic at 213 positions from games (world champs 2004) and then automatically calculate whether something works or not. That's more safe. Vincent
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.