>DNAunion: … So, after all this, what are others' views on whether a chess
computer possesses some form of intelligence? Have human intelligent agents
already created a new form of intelligence that is not biological?
>Brian: Being an avid chess player I cannot resist answering this one :).
The short answer is that chess playing computer programs are not intelligent.
I'm not sure whether a program would pass the Turing test you outlined. I
have a feeling it would not given that a strong player (grandmaster) familiar
with how computers play were the one checking.
DNAunion: Surely 10 years ago even most average chess players could tell the
difference. In fact, companies advertising chess computers made a point of
emphasizing how much more "human-like" theirs played. Back then, for
example, one could confound most chess programs by simply opening 1. a3, 1.
e3, or 1. d3. Even though these are considered weak opening moves and a
human playing black would equalize, at least, quickly and easily, when the
same moves were played against a computer, they practically guaranteed an
opening advantage. This is no longer true as computers have been programmed
with hundreds of thousands of opening moves, and even when you get them out
of their stored book lines, they still have a much better "understanding" of
the goals of the opening and handle it more "intelligently". Also in the
past, computer programs were masters of tactics, but were terrible at
positional play. Just lock up the position and watch the opponent flounder:
that could be a strong indicator that you were playing a computer. But this
too has changed drastically: closed positions are no longer lethal weapons
against the better chess programs, and in fact, chess programs have pulled
off brilliant *positional* sacrifices in many of their tournament games.
With their strong positional games, the elimination of holes in opening
books, the strengthening of their opening strategy, and much better play in
the endgame, today's chess computers play very human like: I am not sure that
even grandmasters could tell the difference accurately when not viewing 100
different opponents, 50 of which were advanced chess computers and 50 of
which were international grandmasters.
>Brian: But since this is not clear, let me give another experiment more in
the spirit of a Turing test. Here a knowledgeable chess player gets to ask
"questions" where the questions are board positions with instructions like
"white to move, what's the best move?". The positions are carefully selected
so that they are highly tactical and intricate but lead to a forced win for
white after 8 or 10 moves. Give a sequence of these positions and then
conclude that the computer is the one with the best score.
DNAunion: I feel confident that great human tacticians - either
international grandmasters known for their tactical prowess, or players who
compose chess puzzles - would score very high under these conditions. Are
you sure this would accurately distinguish between computers and humans? For
example, humans often times have an "intuition" about what move is the best,
even before doing any explicit mental calculations (Bobby Fischer was said to
have great "sight of board", meaning that he could walk up to a position
without any prior knowledge of it and pick the best move intuitively in a
heartbeat): especially when they know the position is a problem (they see a
loose piece, or a weakened king position, or one side having better control
of the center or more pieces positioned in proximity to the opponent's king,
etc.). If you hand some tactical chess puzzles to a master tactician, he or
she will probably see the correct move within the first couple seconds, and
would need more time only to analyze the position to make sure that the move
he or she mentally flagged as the best candidate is in fact the best (if not,
then candidate number 2 would be analyzed). But I don't think an
international grandmaster known for tactics or a chess puzzle composer would
score worse than a computer on these kinds of positions.
>Brian: Oh, to make it clearer you have a time limit of 5 or 10 minutes per
position.
DNAunion: As I said above, I believe the best humans would be able to pick
out the best move within that time frame. In fact, a few years ago there was
a foreign player that typically spent only 30 minutes or so - in total: for
the entire game - in his games and matches against other international
grandmasters (who themselves spent several hours); and he was a contender for
the world championship and did not deviate from his breakneck pace even in
those matches (sorry, when I got out of chess, I sold all my FIDE Informants
and no longer have access to the player's name). I imagine he - and others -
might be misclassified as a computer under your test conditions.
>Brian: To make it even easier, have the first 3 or 4 moves of the winning
variation be very "odd" (untypical) moves.
DNAunion: You mean like sacrificing a queen or the exchange? If so, this
would not throw the best human players - espcially those that compose chess
puzzles themselves. Perhaps you meant some kind of a tempo move - a waiting
move where any move the opponent makes worsens his/her position, with that
waiting move also setting up the win. Again, these are typical in chess
puzzles and most humans that are good at solving such positions might be
classified as computers under your test conditions.
>Brian: Several years ago one the older grandmasters (I want tosay Robert
Byrne) dug up one the old masterpieces
played by Morphy or Alekhine or someone similar. There was a truly brilliant
combination in this game which the author claimed a computer would never
find. Next month he had to eat his words since many readers wrote to him
telling him that their programs found the winning move in just a few seconds.
This shows that the author really didn't appreciate how it is that computers
play so well. But, if we do understand this, we can easily construct a Turing
test that will find the computer.
DNAunion: I still disagree (but not emphatically). I believe that IGMs who
are master tacticians - or players who compose chess puzzles themselves -
would be misclassified under your test conditions.
SIDE COMMENT. By the way, several years ago my chess friend and I composed a
forced mate in 16 that no computer we have tried has so far solved: that
includes my Fritz-engine-based computer program running on its highest level
for 2 days on a dedicated 300 MHz Pentium II system with 128 MB RAM. Here is
the position, white to move and mate in 16:
White: King on h2, Queen on c4, Rooks on c8 & f7, Bishop on c2, Knight on
d6, Pawns on e4, g2, & h4
Black: King on g6, Queen on a1, Rooks on a8 & g4, Bishop on f6, Knights on e5
& g8, Pawns on g7 & h5.
I think this shows that chess computers are not as "all-powerful" tactically
as some might believe.
>Brian: I think the problem with your argument is that it is an argument by
analogy and these types of arguments are notoriously weak. The argument
seems to go basically like this. (1) Humans play chess well because they are
intelligent. (2) Therefore it requires intelligence to play chess well. (3)
Computers play chess well. (4) Therefore computers are intelligent. The flaw
is with statement (2). The problem is that when computers are playing chess
they are not doing anything remotely similar to what a human is doing. They
play chess well for the same reason that they perform finite element
simulations well. They are very fast.
DNAunion: I disagree to some degree. Computers do not simply examine all
possible moves over many plies and then pick the one with the greatest
mathematical increase in "piece score" - that leads to very limited tactical
abilities as the number of moves that must be examined increases
exponentially while time does not. For example, assuming 30 legal moves per
player per position, consider the following. A *very simple*
white-to-move-and-win-a-piece-on-his-3rd-move tactic would require looking at
30^5, or over 24 million, positions. That's fine so far. But for a *modest*
combination where white wins a piece on his 5th move, the number of positions
that need to be analyzed jumps to 30^9, or almost 20 trillion, positions.
Moving on to a lengthier combination where white wins a piece on his 7th
move, the number skyrockets to 30^13, or over 15 quintillion - that is, 15
million trillion - positions! Even the fastest computers cannot handle
examining that many positions under tournament time limits (or even years!).
They must invoke something other than pure brute force to find such winning
moves (and they do: intelligently-programmed selection algorithms that prune
sequence branches). Or consider positions in which neither side can force
the win of material after 5 full moves. What does the computer do? There is
no mathematical advantage to one ending position over all the others, so does
it just find all those millions of possible ending positions that share the
top mathematical spot, pick one of them at random, and then play the move the
began that series? No. Computers evaluate the resulting leaf positions
based on standard considerations: control of the center, mobility, potential
threats, king safety, pawn formations, etc., just as humans do, and then
select the root (first ply) move that lead to it. Computers evaluate
positions to save on processing: if a computer runs across a position in
which material is equal at that time, but its pawn position has been
compromised, its king has been exposed, its knight is on the rim, its only
bishop is sealed in by its own pawns, it has opened up the position by
exchanging pawns in the center even though the opponent is better developed,
etc., it will judge that position as inferior to others that do not share
these negative attributes, even though the piece score could be the same.
This borders on some form of intelligence: some form of thought process
(whether the computer possesses it itself, or is merely a stand-in for the
programmers is a different issue).
>Brian: Consider for example that deeper blue can analyze about 200 million
positions per second.
DNAunion: Yes, but as I pointed out above, even at that rate, Deep Blue
would be unable to solve a combination leading to the win of a piece on its
7th move if it used simple "piece sums" in evaluating positions. At 2 x 10^8
(200 million) positions per second, and approximately 1.5 x 10^19 (15
quintillion) positions to examine, that would require 7.5 x 10^10 seconds, or
over 2,000 years!
>Brian: How could any human possibly survive even 20 moves against such a
monster? The reason is that humans are very intelligent while the computer is
merely fast :).
DNAunion: But computers are not "merely fast" - they must also "know" what
they are doing: they must still have some form of "intelligence" - involved
algorithms that do far more than simply sum up the values of remaining pieces
- to find the best move - no matter how fast they operate (it does not matter
how fast you can run, if you run in a circle, you will get nowhere). They
still must understand about a minority attack, the standard attacks when
opposite-side castling occurs, attacks against a fianchettoed bishop, value
of passed pawns and of protected passed pawns, advantages of knights over
bishops, advantages of bishops over knights, how the relative values of
bishops and knights change as the position opens up or closes or as endgame
approaches, that the king should be tucked away during the opening and
middlegame, but should become active and centralized during the endgame,
about the concept of blockade, etc. The output of such chess programs, if
not intelligent, at the least mimics the intelligent output produced by
humans.
This archive was generated by hypermail 2b29 : Mon Oct 16 2000 - 01:12:39 EDT