Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Distance-to-mate vs Best-move tables

Author: Robert Hyatt

Date: 15:25:15 10/23/01

Go up one level in this thread


On October 23, 2001 at 18:09:50, Ratko V Tomic wrote:

>>1 bit per position value, your savings are not that good.  The
>>current tablebase method takes around 1.3 bits per position
>>after compression.
>
>Is the 1.3 bit/position valid only for archiving database
>on disk, or is it a working size (used directly during
>search)? I.e. to lookup the data in the search tree, does
>one have to expand the blocks of tables? If it does need
>block level expansion, then that refutes the Bob's 1-lookup
>per leaf node assertion, since if you need to decompress
>entire block with thousands of distance-to-mate values
>for one lookup, that's no different than having to follow
>up a sequence of say 100 best moves.

You have to read in some "blocksize" anyway.  Our blocksize is 8K
bytes.  But that is one I/O to get the byte I want.  I don't have to
read in 99 _more_ blocks to track down to the mate-in-N score...

>
>>Totally wrong.  DTM tables actually have a quite high level
>>of predictable patterns, the most common ones being that positions
>>whose indexes differ by 1, 8, or 64  tend to have the same value.
>>This is because these positions differ only by one piece being
>>located one square away.  These predictable patters provide a
>>significant boost to the compression rate as compared to
>>huffman by itself.
>
>These are relatively crude patterns in the same category of
>pattern as the simple board symmetries (all of these I assumed
>already factored in, i..e they're available equally to any
>encoding method).
>
>The patterns I was talking about (that the best-move tables can
>use for compression, but not the distance tables) are the
>nontrivial endgame rules/knowledge built into its move
>generator. If you imagine a program posessing the patterns
>you're mentioning (in addition to the bare legal move generator),
>it would, for all practical purposes play random chess, weaker
>than a human novice ten minutes after learning how the pieces
>move and what is the objective. This is in a different legaue
>compared to even a simple move generator which knows only the
>plain captures in the next move, to say nothing about a generator
>which knows various quick-to-compute endgame rules. All such
>patterns are far outside the reach of distance-to-mate table
>methods for reasons mentioned earlier.
>
>Note also that the strength (rating/points) are not strictly
>correlated to the shortest mate but to the ratio of points won
>to the points lost. In other words, you pay in large overhead
>with distance-to-mate tables for knowledge which doesn't
>contribute to the strength of the program (however
>interesting it may otherwise be in some endgame reasearch).



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.