This has probably been thought of several times, but here goes anyway.

I enjoy playing a computer when it plays more like a human than a computer. It
seems to me that we are not quite at the point (yet) to make a computer play
like the best players in the world (while we can make them play at their level,
we can't make a computer "think" like they do). Programs like Chessmaster 9000
attempt to do this with personalities but only in trying to "guess" what the
player would choose based on material and positional ideas, not actual data.

So why hasn't there been a program which can load up a set of PGN files (played
by a human opponent) and then create a learning database off of them? Like if I
download all of David Janowski's games, have a program analyze all of the games
and create a player (perhaps within 90-99% accuracy) which plays and thinks like
Janowki did.

I wouldn't care if it took the program 2 days to analyze all of the games
carefully before the personality could be created, at least could it be
possible? Maybe a program can already have a few pre-stored?

I'm not just talking about simply "playing" what Janowski plays at a certain
position, but rather playing "what he could" play based on this data and his
philosophy looking at certain positions. Sort of like "The Sims" meets
"Shredder". :-)

If anything, such a program would be more realistically
compared to (75%-95% accuracy)

It's an interesting idea, however it tells me that
"calculation" power of a computer would become less
of an interest.

A lot of human players calculate different depths at
different parts of the game. A computer calculates the
same depth at every spot of a game (assuming it has
either a fixed ply or same amount of time each move).

Janowski could look at a position with 3 plies and then
the next position at 8 plies, then the next one at 5 plies...
all while incorporating his "philosophy" as you put it. The
computer's consistent ply depth (which is usually the same
for every move) would need to be adjusted along with the
correct attacking (or defending) ideas from his games.

-Wayne

Sidney wrote:

So why hasn't there been a program which can load up a set of PGN files (played
by a human opponent) and then create a learning database off of them? Like if I
download all of David Janowski's games, have a program analyze all of the games
and create a player (perhaps within 90-99% accuracy) which plays and thinks like
Janowki did.

To play chess like Janowski, we need to know how he thinks, not what
moves he makes.

A human might see the moves and infer things about his style, but that
is because humans players probably think about the games in similar
terms, the computers would need to know what sort of concepts Janowski
has about chess before it could even start.

I've read Nimzowitch and studied his games and I don't play like him
athough there was this one game where my opponent hit a kind of
positional Zugzwang that felt like Nimzowitch could have played it

The closest we have is weight tuning such as Deep/(er) Blue did, but
some concepts may turn up very rarely in a chess game or it's analysis
and still influence a GM's thinking.

-----BEGIN PGP SIGNATURE-----
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFAmC1JGFXfHI9FVgYRAjuKAKCahZk1IOA+aNcV7wr3AV a+vXamZgCgz2As
wsfBZgBaHUwmg5iuqXnE4Gc=
=Hl/M
-----END PGP SIGNATURE-----

Simon Waters wrote:
Sidney wrote:

So why hasn't there been a program which can load up a set of PGN files (played
by a human opponent) and then create a learning database off of them? Like if I
download all of David Janowski's games, have a program analyze all of the games
and create a player (perhaps within 90-99% accuracy) which plays and thinks like
Janowki did.


To play chess like Janowski, we need to know how he thinks, not what
moves he makes.

A human might see the moves and infer things about his style, but that
is because humans players probably think about the games in similar
terms, the computers would need to know what sort of concepts Janowski
has about chess before it could even start.

Anyone interested in these topics should do searches for a man named de
Groot as well as Gobet and Simon (together). Gobet and Simon continued
study and further developed de Groot's theories and observations. Many
scientists have studied how humans think about chess and what makes the
difference between novice and master players.

The difference? Experience. Humans think of chess in 'chunks'. They
immediately see recognisable components of play in a position that they
use to guide their search and evaluate positions. These chunks, and the
moves associated with them, are developed through experience in study
and play. To go beyond that basic description is not within my level of
knowledge.

These concepts have been used by several computer scientists in chess AI
programs. They use pattern matchers and databases to either guide the
search or evaluate positions. Some learn their patterns, others are
hard coded. There has been MUCH study in this area but it is under the
surface because brute force methods have so far been the stronger.

Interested people might start by reading "The game of chess" which is
online in the citeseer database.

--
"I'm a war president. I make decisions here in the Oval Office
in foreign policy matters with war on my mind." - Bush

Thanks to everyone who replied.

"Simon Waters" wrote:
To play chess like Janowski, we need to know how he thinks, not what
moves he makes.

Is it possible to make a software program analyze all of his games, and then
create some sort of knowledge base that can make assumptions of how he attacks
when pieces are in certain positions? Is that creating an "idea" of how he
thinks before he plays? It might not be completely accurate but they could make
a "smart" program that knows how to properly learn by studying his games and
looking for his patterns.

Also, it should be noted that if the same program would happen to study a weaker
player (say a 1500 club player) then the computer would in fact "play" at that
strength as well.. making the same mistakes.

"Sidney" wrote in message:
....
I enjoy playing a computer when it plays
more like a human than a computer. It seems
to me that we are not quite at the point
(yet) to make a computer play like the best
players in the world (while we can make them
play at theirlevel, we can't make a computer
"think" like they do). Programs like Chessmaster
9000 attempt to do this with personalities but
only in trying to "guess" what the player would
choose based on material and positional ideas,
not actual data.

So far, the most "human" playing computer software
I've ever seen (that plays good) is Chess System Tal
II. I believe on modern systems, it can play between
2300-2400 ELO, but plays "frighteningly" human, and
loves to attack.

It's not available "by itself" anymore, however there
is a game pack called: "10 Pro Board Games" which comes
with "Checkers, Go, Connect 4, Chess" etc.. and the
chess program on the disc is "Chess System Tal II", so
pick it up.

Here is a game. I had my copy play against "Chess Tiger 15
Gambit Suicidal" and even though Chess System Tal II
lost, take a look at some of the amazing moves it came up
with:

[White "Chess System Tal II"]
[Black "Chess Tiger 15.0 - Gambit Suicidal Style"]
[Result "0-1"]
[Opening "B46 Sicilian: Taimanov Variation"]

1.e4 c5 2.d4 cxd4 3.Nf3 e6 4.Nxd4 Nc6 5.Nc3 a6
6.Nxc6 bxc6 7.Qd4 d5 8.Bf4 c5 9.Qe5 f6 10.Qh5+ g6
11.Qf3 e5 12.exd5 exf4 13.d6 Rb8 14.O-O-O Qb6 15.Qe4+ Kd8
16.Na4 Qa5 17.Qc4 f3 18.Nxc5 Qb6 19.Qd4 fxg2 20.Bxg2 Qb4
21.Rhe1 Qxd4 22.Rxd4 Bd7 23.Nb7+ Rxb7 24.Bxb7 Nh6 25.Bxa6 Nf5
26.Rd5 Bxd6 27.h3 Bc6 28.Rdd1 Re8 29.Rxe8+ Kxe8 30.Bc8 Bf4+
31.Kb1 Nd6 32.Re1+ Kd8 33.Be6 Ne4 34.Bg8 h6 35.Bf7 g5
36.Re2 Nd2+ 37.Ka1 Nf3 38.Bg6 Bd7 39.Re4 Ne5 40.Bh5 Bxh3
41.a4 Bg4 42.Rxe5 Bxh5 43.Rc5 Be2 44.Rd5+ Kc7 45.b4 h5
46.Rc5+ Kb7 47.Rd5 h4 48.Rd8 h3 0-1

If anything, I REALLY hope they continue diving into this area
because there is a great deal of promising improvement to be
made.

Theoretically yes, but statistically you don't have anywhere near enough
games on record to do this. And that's not allowing for human
variability - how he felt when he got out of bed that morning. You can
only deal with very vague generalities. Such an analysis does not
really touch the thought process - the 'how he got there' part.

Also, over the course of a chess players career, this process matures,
improves, and declines. Not static. Even if you had every single game
a player played from the day he started, you'll find a dynamic, moving
target. You are limited to, at best, very vague generalities.


Is it possible to make a software program analyze all of his games,
and then
create some sort of knowledge base that can make assumptions of how he attacks
when pieces are in certain positions? Is that creating an "idea" of how he
thinks before he plays?

Sidney wrote:
"Simon Waters" wrote:
To play chess like Janowski, we need to know how he thinks, not what
moves he makes.

Is it possible to make a software program analyze all of his games, and
then create some sort of knowledge base that can make assumptions of how
he attacks when pieces are in certain positions?

The difficulty of that approach (working out which parts of the position
are relevant and which aren't) is the very reason that we're stuck with
brute-force approaches.


Dave.

--
David Richerby Edible Mouldy Book (TM): it's like a
www.chiark.greenend.org.uk/~davidr/ romantic novel but it's starting to
grow mushrooms and you can eat it!

Take Paul Morphy's famous opera game in Paris against the Duke of Brunswick and
Count Isouard.

1. e4 e5 2. Nf3 d6 3.d4 Bg4 4. dxe5 Bxf3 5. Qxf3 dxe5 6. Bc4 Nf6
7. Qb3 Qe7 8. Nc3 c6 9. Bg5 b5 10. Nxb5 cxb5 11. Bxb5+ Nbd7 12. O-O-O Rd8
13. Rxd7 Rxd7 14. Rd1 Qe6 15. Bxd7+ Nxd7 16. Qb8+ Nxb8 17. Rd8# 1-0

8. Nc3 is a very human move. Is it even possible for a computer (without being
in book) to even see a move like 8. Nc3 continuing development? Most computers
would just die to play 8. Bxf7+ instead I would imagine.