Author: Eugene Nalimov
Date: 17:40:58 04/01/04
Go up one level in this thread
On April 01, 2004 at 19:05:09, Dann Corbit wrote: >On April 01, 2004 at 18:38:59, Sune Fischer wrote: > >>On April 01, 2004 at 18:29:27, Dann Corbit wrote: >> >>>On April 01, 2004 at 17:59:38, Eugene Nalimov wrote: >>> >>>>On April 01, 2004 at 15:16:34, Marc Bourzutschky wrote: >>>> >>>>>The Chessmaster format is indeed better >>>> >>>>What does it mean "better"? :-) >>>> >>>>It stores less information, thus compresses better. >>> >>>I have an idea that I think would be helpful if you should be so kind as to >>>perform it. >>> >>>Write a scanner that reads your wonderful EGTB files and spits out a two bit >>>state only for each position (won/lost/drawn/broke) to create bitbase files. >>> >>>The reason I suggest it is that a bazillion programmers won't have to reinvent >>>the wheel. >>> >>>I suggest the use of the bitbase files early in the search (completely pulled >>>into ram) and then EGTB at the leaves if the bitbase indicates it is worthwhile. >> >>You must mean it the opposite way, bitbases at the leaves and EGTBs a near root? >> >>I think it would be better to use bitbases in the entire search and only use >>full EGTBs when the position is at the root. >> >>Or, if you want the search to eventually return mate scores, probe EGTBs when >>bitbases say it is won and beta>=mate_bound or bitbases says it lost and >>alpha<=-mate_bound. >>Perhaps probing directly into EGTBs when window allows it would be faster, >>matter of tuning of course. > >I guess I had not thought about it carefully enough. I imagined using bitbases >to get a won/lost/drawn opinion (at all nodes). But unless you know the exact >value of the leaves, I don't see how you can choose the best move. > >I imagined something like this: >If the best evaluation is drawn or lost, who cares. Do whatever move is among >the suggested list. >If the best evaluation is won, then: >Examine the bottom leaves that are won and perk the correct values back up. > >How will we otherwise find the true value? I am afraid I don't understand how >it can work. In my "TODO" list. But let me finish 6-men TBs first... Simple way is to keep both w/d/l and full tables. You need to probe full table only when position is OTB. Otherwise you probe w/d/l table. W/d/l tables are smaller, and relevan ones can be always loaded to RAM, so you can probe them everywhere in the search, including Q-search. Probing of the full TBs can be much slower than it is now, probably ~1 sec should be fine. In theory that allows to use better decompression algorithm. And you don't need 2 bits per position. 1.6 bits are enough (5 positions per byte). Thanks, Eugene
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.