Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: suggestions for (a new generation of) chess knowledge using engines

Author: Ingo Lindam

Date: 13:54:01 10/23/02

Go up one level in this thread


A feedback I got on the rec.games.chess.computer newsserver from Jaques Ruiz and
my answer on his questions might enrich the discussion:

Hello Jacques,

thank you for your reply and your questions.

"Jacques_Ruiz" <JacquesRuiz@wanadoo.fr> schrieb im Newsbeitrag
news:ap69fi$m5d$1@news-reader10.wanadoo.fr...

> 1/ what is a pattern?
>    eg: static pattern like "Black Knight at G8 and Black Queen at D8";
>         boolean pattern like "(nG8 || nF6) && qD8 && (KE1 || KD1)"
>         chess-specific dynamic patterns like "a black knight attacking
> square E4 and an undefended white pawn at E4";
>         something else?

all of the 3 examples are pattern, but the second one seems to be next to a
definition of the pattern I was thinking of. All pattern should be boolean
pattern in the meaning of that a position either fullfills/contains this
pattern or not. So all pattern themselves can be a part of a new pattern:

if a,b are pattern then also a||b and a&&b are pattern (candidates)

So also the 3rd example you gave is a more complex pattern just given in
words instead by a formula based on elementary one bit patterns. Also the
1st example of a static pattern might be a pattern. Ofcourse we are not
interested in all patterns but in patterns that are useful in practical
chess.

To derive plans you could also build patterns of moves and move sequenzes or
combinations of positional patterns and move patterns:

e.g. it might be derivable that with black castled to the kingside (with
f7,g6,h7,Bg7) and white to the queenside, h2-h4-h5-hxg6 is a much better
plan than a2-a4-axb5. But at first step I would rather think of positional
patterns to evaluate positions than of patterns of moves to evaluate plans.

The aim ofcourse is not only to have patterns but to use them estimating
P+,P=,P- of the position to evaluate.

What is a useful set of pattern?
Ofcourse the pattern you can find in my (or the computers) future games
should with a high probability apear in my set of patterns. Further P+,P=,P-
estimated over all positions containing a pattern should significantly
deviate from P+,P=,P- estimated over all positions.

Some simple estimations made with chessbase on a given set of games e.g.
lead to
patterns like

(P+,P=,P-)(Pe4,Pg4,Nf5,pe5,pg5) = (0.5,0.3,0.2)
(P+,P=,P-)(Pf3,Pg4,Nh2,pg5)     = (0.2,0.29,0.51)

As initial one bit items of the positions of a given set of games you could
also use some items like
white has a pair of bishops, no queens on the board, material is equal,
white has 2 pawns for a knight,... to accellerate the process of finding
complex patterns. But all these initial pattern might also be derived in a
later stage of the process of generating patterns (chess knowledge).

Generating pattern should not be much more difficault then to generate
moves/positions of a tree in the search, but it could be done outside of
plaing a game and lead to knowledge the computer is able to use in all
future games. All pattern easily may permanently reevaluated on growing
databases of games.

> 2/ how do you extract a pattern from a set of N games?

I could use know cluster algorithmns:
a) top down to cluster the elemantary 1 bit items of the N games to my set
of pattern
b) buttom up to obtain my set of pattern out of the positions of my N games.

Also I could find a way to generate systematically pattern and evaluate the
on the set of N games.

It is also possibel to cluster a set of pattern first on the games of e.g.
Karpov and then evaluate that set of pattern on the whole set of games.

> 3/ how do you use these patterns?
>    eg: assign a weight to each of them and use them as a positional
> heuristic in the evaluation;
>         assign a hint move to each of them, like "pattern 82 -> should
play
> Knight at F3";
>         something else?

I was first of all thinking of using them to evaluate P+,P=,P- of a given
position. The pattern not nesessaryly distingish between positional and
tactical patterns, but it might be a strategy to sort out obviously tactical
motives if you are sure the tactical possibilities are all catched by the
search itself.

As I said earlier the pattern might also be used to evaluate plans, but not
in the way
if (pattern82 ) then always play Nf3.

Best regards,
Ingo



This page took 0.02 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.