[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: computer-go: perfect players
Of course how far we are away from perfect play is opinion and subject
to huge amounts of error.
My personal feeling is that we are much farther than we think. I
base this on these observations:
1. Human arrogance. This is not directed to anyone as an individual
but as a class we are all a bit like this. We think we are much
better at various tasks than we really are in my opinion. We compare
to other humans, or to the "average" human and are too easily impressed.
We also romanticize the abilities of our hero's, and make them bigger
than life.
The problem is that we don't have a reasonable frame of reference, we just
can't imagine someone playing a lot better than our own "great" players.
We define greatness and perfection in these terms even though we don't
really mean to.
There will be players that come along later who will prove this, and
yet they will be limited too. This is similar to the great Roger
Bannister, running a mile as fast as humanly possible ... or so it was
thought at the time.
2. Comparing checkers to chess is kind of like comparing chess to go.
It's "easier" to reach perfection in checkers than in chess, and
it's easier to reach perfection in chess than in go as far as writing
programs that play these games. And yet, you will see that Chinook,
the worlds top checker playing program is far from perfect. It's may
very well be the case that Chinook is the best player in the world,
maybe even the best player ever. There is room for disptute here,
but I think anyone would agree that it is a close call. Any yet
if you talk to the author of Chinook (or read his book on the subject)
you will see that he doesn't consider Chinook particularly close to
perfection, he is aware of numerous improvements that need to be
made, and he would love to have a much faster computer to do a few
extra ply. If he considers the "best player" in the world as
needing so much improvement, in a game tremendously simple compared
to chess or go, then I believe it says something about our own human
limitations.
I did my own experiments along these lines. I created an even
more simplified game. It is checkers played on a 6x6 board. I
run a huge series of autotest games at various depths of search.
What I wanted to do was to analyze how hard it would be to create
a perfect player. Although I did not get a definitive answer,
I kept being surprised at how useful a single extra ply of lookahead
was. I was going almost 40 ply deep in my "most fit" individual.
Remember, this is a very tiny game where just a few moves puts you
deep into the endgame. In the end, I could not predict how much
depth I would need to play perfectly, because even at the highest
level I tested (I think it was 38 ply) it was beating the previous level
by several percent (like 54%.) It was tapering off (2 vs 1 was
a much bigger crush than 38 vs 37) but only very slowly. If I played
versions 2 or more levels apart, I would get huge crushing victories.
(I played 200 game series to muffle most of the statistical noise.)
The deep level were mostly drawn games, an indication that perfection
was being approached, but every few games would end in a victory,
usually for the side looking ahead a single measly move deeper.
It's hard to draw firm conclusions from this, becaues I don't
have a simple way to relate this to human skill.
Still, I felt like I was exploring a bottomless
pit of possible skill levels, and I am pretty convinced that at
the bottom (perfect play) would be a player very much better than
any human is likely to be. This is a trivial game compared to
chess and go and even 8x8 checkers.
In Chinooks experience, Chinook would often play a losing move
and not know it for many many moves and then suddenly wake up.
In these cases, the human master it was playing usually sensed
the error right away. Even though this was happening, Chinook
was winning tournaments and beating the best players. The lesson
here is that even though humans were better at some things, Chinook
was better at other things and neither was very close to perfection!
Don
Date: Wed, 2 May 2001 17:24:15 +0100
From: Nick Wedd <Nick@xxxxxxxxxxxxxxxxx>
References: <CE7CC0FC87E4D411BCB50060B0689C173B8245@xxxxxxxxxxxxxxxxx>
MIME-Version: 1.0
X-Mailer: Turnpike Integrated Version 4.02 U <2JQI4ZxDiDMfZZvAyGkNIZYSCT>
Sender: owner-computer-go@xxxxxxxxxxxxxxxxx
Precedence: bulk
Reply-To: computer-go@xxxxxxxxxxxxxxxxx
Content-Type: text
Content-Length: 1449
Fant, Chris <chris.fant@xxxxxxxxxxxxxxxxx> writes
>It has been said by at least one chess player that they could beat a perfect
>player, given a one pawn advantage. Has this ever been tried against a
>computer player? I am interested in knowing how close computer players are
>to a perfect player, how close the best human players are to the perfect
>player and how close the best Go players are to the perfect player.
This has been debated on rec.games.go.
The consensus seems to be that the correct handicap for the best human
to receive when playing a perfect player is about three stones.
The question is complicated by the need to define "the perfect player".
Is it one that makes the best move on the board in each position? Or
one that can read its opponents mind, and take advantage of the
misconceptions that it finds there?
> It is
>my conjecture that human and computer chess players are very close to the
>perfect player relative to how close the best Go players are to the perfect
>player. I believe that Go programs will eventually surpass human players
>and go on to absolutely destroy the best of the best.
They show no signs of doing this so far.
> I believe this
>because of the complexity of the game. Computers can't be stopped. They
>will continue to approach perfect player performance. What does the group
>think?
They will continue to get better, yes. There's a way to go.
Nick
--
Nick Wedd nick@xxxxxxxxxxxxxxxxx