Author: Vincent Diepeveen
Date: 11:16:01 12/08/99
Go up one level in this thread
On December 08, 1999 at 01:26:31, blass uri wrote:
>On December 07, 1999 at 21:42:42, Vincent Diepeveen wrote:
>
>>On December 07, 1999 at 02:17:27, Dave Gomboc wrote:
>>
>>>On December 07, 1999 at 01:21:46, Tina Long wrote:
>>>
>>>>On December 05, 1999 at 19:47:14, Vincent Diepeveen wrote:
>>>>
>>>>>On December 05, 1999 at 14:25:36, Thorsten Czub wrote:
>>>>>
>>>>>>it isn't "astounding".
>>>>>>Torstein spoke about Fritz6.
>>>>>>You talk about fritz5.32.
>>>>>>we don't know so far where fritz6 will rank.
>>>>>>
>>>>>>also we don't know if fritz5.32 would react in the same
>>>>>>way than fritz6 in torsteins experience.
>>>>>
>>>>>fritz6 looks tactical stronger and a bit more aggressive tuned
>>>>>a la nimzo. Hard to say whether tactical stronger is derived
>>>>>from the more aggressive tuning or from seeing more...
>>>>>...anyway nothing changed in its eval seemingly, just a few parameter
>>>>>settings as usual...
>>>>
>>>>Hi Vincent,
>>>>I'm not clear what you are meaning here.
>>>>
>>>>"nothing changed in its eval seemingly" are you meaning the Fritz6 engine seems
>>>>to be the same as the Fritz5.32 engine?
>>>>
>>>>"just a few parameter settings" such as we change in CM6k or RebelCentury
>>>>"personalities?
>>>>
>>>>"as usual" that this not-new engine but tweaked old engine is usually what is
>>>>sold? How many versions are the same engine in Fritz?
>>>>
>>>>This worries me, as I (and my Father) spend a large amount on what are sold as
>>>>New ewngines.
>>>>
>>>>Thankyou,
>>>>Tina Long
>>>
>>>Don't be worried. It's just Vincent's standard hyperbole.
>>>
>>>Dave
>>
>>Dave you should go to a world championship, say nothing there and listen
>>when programmers talk without press near them, or dudes that can quote
>>them in a forum.
>>
>>Quote from 1997 from the aegon tournament which after i talked to
>>other programmers was repeated:
>>
>> "My program is using piece square tables. I'm not happy about it.
>> when searching more than 10 ply i can't measure any positional
>> progression. the only progression it makes is seeing a few more
>> shots at testsets".
>
>If the programmer was not happy with it then the natural thing to expect is that
>he is going to change it.
>
>This quote is from 1997 so I do not believe that the programmer did nothing to
>change it.
>
>I also do not believe a top program in 1997 was based only or mainly on piece
>square table.
>
>There are many things that you cannot evaluate by piece square tables(for
>example passed pawns) and it is possible that some fast searchers evaluate
>things that you did not think to evaluate.
What many piece square table programs do is make in advance an exchange
table, passed pawns get done by some programs even incremental.
that exchange table is obviously updated at the leafs, but basically in
the root.
For non-chessprogrammers it's very hard to believe that this causes a
style of play that does very well at all kinds of testpositions, even in
a lot of games it generates a very consequent form of playing.
However the deeper one searches the more those piece square tables
are away from the reality of the board. Especially pawnmoves can
drastically modify a boardposition.
This explains why many of these programs can't finish games, they
just don't know how to exchange to an endgame. if they want to
exchange then they always exchange.
This of course also applies to programs that do not only rely on
piece square tables but have a very limited set of rules to define
their play.
>Uri
To quote Amir Ban (what he said on icc to me some time ago):
"Piece square tables work! Brought me a world title!".
I'm obviously not a supporter of the opinion though
that piece square tables work. The way the search works in junior
already indicates obviously that PSQ doesn't work at bigger
depths.
Junior basically seems to use a static exchange evaluation (which
can be done if your evaluation is simple) and its search seems to
recognize a normal move as 2 ply, and special moves are seen as a 1 ply
move.
though i obviously don't know the fine implementation of junior
it seems that it sees next moves as special moves:
- good captures
- covering attacked pieces (if not attacked by minor piece)
- moving away attacked pieces
- recaptures
- checks
- attacks to piece (not sure whether all attacks,
only interesting attacks)
When i tried this in diep i also experimented with next
to see next as special moves:
- passed pawn moves
- interesting non losing pawn moves, such as centre pawn pushes
and levers.
- the highest evaluated move if it's more or less singular
and there is no good captures, or the 2 moves as evaluated
best. highest evaluated move could never be a capture obviously,
those were evaluated already above by a SEE
Note that one must limit the number of moves drastically if one
wants so search deeply, as searching at a depth of say 14 ply means
that you basically see 7 ply positional and if in the interesting
lines there are some moves that are considered tactical you see like
9 ply. However all tactical shots you see in a lightflash.
Of course in DIEP a big problem using this approach was that i don't
do a simplistic evaluation, i'm using a big quiescencesearch,
and this whole approach means you MUST get a big search before you
see anything.
After this version was setting new high scores for DIEP at the internet,
it took me many months to think objectively.
After many months of fiddling with this i 'invented' trying the best
evaluated moves also. I noticed that just adding a few more moves
to the special moves completely blew up the search.
If you have a program that normally would search say 8 plies
and now you search suddenly 12 plies, that means that every line you
search your average depth may not get over 8 plies.
However i made next calculation.
Suppose that for the lines i want my program to see, there is a 90%
chance that a move, from the line you want it to see, is considered
a special move.
This means you only need to search on average roughly 10% deeper
to see on average the same trick/positoinal shot.
Getting normally 9 ply with DIEP at tournament level, this belle/junior
approach is very tempting to use, as you suddenly get 13 or 14 ply,
and hardly miss a trick.
13 or 14 ply obviously is way over 10% more.
Now i know that some of the programmers here will say: "but there is
a chance you miss from a big trick the complete line".
There is a counter argument here. The deeper one searches the more
unlikely that you miss the trick, as the combination is then made up
from many moves which again have that 90% chance to get in the special
move selection.
So tactical obviously the belle/junior method is great.
For my diep there were however a few problems which were so
gigantic to me that i dropped this search methods.
- when comparing the searches at infinite depth i measured a worse
branching factor for the belle/junior method of searching. However
because of the major pruning (every nonspecial move is pruned by
1 ply exactly) you of course start already with a depth of say 12 ply
for the belle/junior method.
- when i compared at tournament level the positional insight of the
belle/junior version with the normal version, i clearly saw that
the normal version was playing positional better moves, even though
there was seemingly a big depth difference.
- nullmove worked a lot worse. it didn't give the speedup i got at normal
move. this might be obvious to most of you, but consider that when
searching a lot deeper you can nullmove also more, and i'm used to
get at bigger depths way better performance of nullmove than at small
depths.
most important argument:
- the idea of just searching tactics deeper and not positional
lines (obviously that 90% chance was only true for tactical moves)
i considered not a good idea for a program with a good evaluation.
In this way you don't take advantage of a good evaluation, but
assume more or less that your evaluation is that bad that you only
need to search tactical deeper.
Vincent
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.