Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Is there a limit on our ability to compute endgame tablebases?

Author: Tony Werten

Date: 07:07:23 05/29/02

Go up one level in this thread


On May 29, 2002 at 09:18:43, Vincent Diepeveen wrote:

>On May 29, 2002 at 01:07:21, martin fierz wrote:
>
>>On May 28, 2002 at 16:54:22, Vincent Diepeveen wrote:
>>
>>>On May 28, 2002 at 03:30:53, Russell Reagan wrote:
>>>
>>>first of all the size problem is there indeed. whatever compression
>>>you use, it will always be big tables.
>>>
>>>However obviously generating them only has to be performed once,
>>>which means that most generators are not exactly speed based.  a month
>>>a table at a fast big alpha is indeed the truth now i guess.
>>>
>>>Reality is that it can be done a lot of times faster with a faster generator.
>>>
>>>Take mine.
>>>
>>>Also nalimov/heinz format stores positions in 2 bytes a position or so
>>>at 6 men. I just store win/draw/loss/unknown when generating and
>>>win/draw/loss when finished generating.
>>>
>>>that's 16 bits versus less than 2 bits a position. Factor 8 in size
>>>difference directly (practical less than that as i use a less
>>>efficient scheme). However with compression using the superb compression
>>>it compresses much better, so the total size needed to store the EGTBs
>>>is much smaller.
>>>
>>>Doesn't take away that all 6 men represent a total size which is
>>>real *huge*. there are way more 6 men EGTBs than there are 5 men.
>>>
>>>The 5 men is a complete 'few hours' generation joke compared to what
>>>happens with the 6 men. If i have some spare time left this summer
>>>after world champs i will finish the en passant bug i have now and
>>>add the superb compression from andrew kadatch to the
>>>EGTBs and if i win the toto somewhere that i can afford
>>>to buy myself enough ram and a fast U160 SCSI harddisk
>>>
>>>or if someone has
>>>2 GIGAbyte ram (windows or linux is no issue here. also other
>>>OSes no problem as it is ansi-c) and a big harddisk and willing to
>>>generate some 6 men, then please let him either shout loud or email
>>>me. 512mb ram or so which i have myself and many others is simply
>>>a big problem to generate 6 men with pawns on the board. It slows
>>>down the generation process at least 8 times.
>>>
>>>about 40 gigs of diskspace needed to generate a 6 men with a pawn
>>>like KRP KRP is a real interesting egtb. But before that also other
>>>EGTBs are needed.
>>>
>>>Thanks to andrew kadatch for compression. without it, it would have
>>>been hopeless.
>>
>>hi vincent,
>>
>>i recently computed the 8-piece (english) checkers endgame database, also just
>>doing win-loss-draw. i used a 40GB HD / 1GB RAM system. i managed to compress it
>>by about a factor of 9, approximately 20% better than what the chinook guys did
>>- using a similar RLE scheme, which i tested a couple of times and then stopped
>>working on after maybe a few evenings. i was wondering if RLE is in fact the
>>best method to compress such databases, because back in 92 when schaeffer
>>started this database thing, disk access was much faster compared to processors
>>than today; so maybe today a scheme which compresses more but takes longer to
>>decompress might be better. could you give some details of andrew's compression
>>algorithm? i'd be very interested ;-)
>
>Though i have made my own huffman compressor and toyed with fractal
>compression for half a month and also arithmatic compression, i will
>be the first one to admit that i know near zero about the technical
>capabilities of compression.
>
>Please explain some what is RLE?

Run length encoding. What it basicly does is changing 444444 into 46 ( meaning 6
times a 4 ) Works great for drawish tablebases, specially if you change unknown
and broken into draw.

Tony

>
>You can see how andrew kadatch compression works 'easily' perhaps by
>looking to nalimov code. nalimov code uses also kadatch compression.
>
>>aloha
>>  martin



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.