Author: Vincent Diepeveen
Date: 06:37:39 12/07/99
Go up one level in this thread
On December 05, 1999 at 22:35:23, Michael Fuhrmann wrote:
>Bob Hyatt stated in a recent post that all chess programs have "absolutely
>glaring holes." What are the specific weaknesses of some of the top programs.
We must see 3 different types of programs
- piece square table programs
for example junior
- simple eval programs but not PSQ
for example crafty
- full eval programs having 'complete knowledge'
for example diep
I consider 'complete knowledge' still a subset of
what a human in a lightflash sees positional.
However what i refer to is having a full package
True piece square table programs rely fully on
their tables that are pregenerated
- positional rules
nearly all programs except the true PSQ ones
have some positional rules like: bishop on d3 is
bad if there is a pawn on d2
even having a few thousands of these rules should not
slow down an evaluation *too* much. most are simply
patterns
- complex positional rules
complex positional rules i see as rules that require some
form of scanning. like trapped rooks in the corner (something
that's obvious not in fritz, lost a world title on it some months
ago against ferret).
- tactical rules
simple tactical rules like: put pressure on a pawn
are also easy to implement. Tiger is a good example of a program
putting extreme pressure on static weak pawns of the opponent
to my big surprise not many programs evaluate more complicated
tactical rules like pins.
- space rules
space is the space you occupy in general, very important in go
also very important in chess. i refer a lot to it as 'mobility'
or 'activity', though both can be seen as different forms of
space.
now we get into an area which is in my opinion very important
both historically seen as well as for the future. Programs
search at different depths. We can assume a lot of the above
rules influence the play of a program. However a lot of programs
are lacking space rules more referred to as mobility by my side.
I see sometimes programs at the internet (basically crafty as this
program obviously plays the most there against my DIEP) crafty
having several choices to either do a concession or get an advantage.
In 50% of the cases it's bad luck then as crafty takes its choice
without looking to the space it has after making its move.
SO THERE IS A GOOD CHANCE that if you play on that it gets into
problems because of this.
In general programs lacking a clear sense of mobility/activity
are completely dead when you outsearch them, as there is always a time
they have to chose and they chose wrong then. I think this historically
has made a lot of people do believe that just getting another ply
would remain the biggest problem.
- meta rules
meta rules are rules that deal with previous outcome of evaluation parts
so rules about previously generated knowledge. this is an extremely
tough terrain and very interesting for AI researches as this is
exactly the terrain what is so far hidden from research.
my own experience with meta rules is that it is extremely dangerous
to use it without care, as a few lines of code might influence the
final evaluation of a position more than a couple of hundreds of
patterns that applied to the position.
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.