From newshub.ccs.yorku.ca!ists!torn!utgpu!cs.utexas.edu!swrinde!mips!mips!decwrl!mcnc!aurs01!throop Tue Jun 23 13:21:20 EDT 1992
Article 6324 of comp.ai.philosophy:
Path: newshub.ccs.yorku.ca!ists!torn!utgpu!cs.utexas.edu!swrinde!mips!mips!decwrl!mcnc!aurs01!throop
>From: throop@aurs01.UUCP (Wayne Throop)
Newsgroups: comp.ai.philosophy
Subject: Re: 5-step program to AI
Message-ID: <60848@aurs01.UUCP>
Date: 19 Jun 92 17:25:19 GMT
References: <60840@aurs01.UUCP> <1992Jun18.022002.29912@mp.cs.niu.edu> <60842@aurs01.UUCP> <1992Jun18.205639.3093@mp.cs.niu.edu>
Sender: news@aurs01.UUCP
Lines: 91

> rickert@mp.cs.niu.edu (Neil Rickert)
> There is no question that computers use symbols very differently from
> the way humans use them.  I never argued otherwise.

Ah.  My misinterpretation, then.  Back to chess.

>>> The computer chess program and the human chess player both proceed in a
>>> somewhat similar manner.  They construct possible sequences of continuation
>>> moves and evaluate the result. 
>>I  think this is incorrect.  [...]
> So far this does not disagree with my point, that both generate
> continuations and evaluate the result.

Sorry, I think perhaps I didn't emphasize the differences in
approach strongly or clearly enough.  Let me try again.

Computers generate many positions that are guaranteed to be reachable
by legal play from the current position, and then evaluate them for
strategic and tactical "goodness".

Humans, on the other hand, (seem to) generate a few positions that
are guaranteed to have strategic and tactical "goodness", and then
evaluate the possible paths from the current position towards them.

Note that humans do NOT choose nearby, generic positions that are
"good".  They choose quite specific positions (or small classes of
positions) quite far in the "future" of the game, and then steer
towards them.  More on this below.

Humans: choose positions and evauate paths.  
Computers: choose paths and evaluate positions.

We can try to see whether these guesses as to human chess method are
reasonable by considering, for example, how well humans and computers
play in differing phases of the game.  In the midgame, the evaluations
are tenative and there are many, many possibe reachable moves to
consider.  By the above hypothesis, humans should do better here, since
they "skip" the hard parts.  In the endgame, the position evaluations
are relatively easy, and the tree of choices of legal moves (and hence
reachable positions) is far less bushy, so the computer should do much
better.  Indeed, the relative skill between humans and computers do
follow this pattern.

How it is that humans settle on the few positions that they do, and why
it is so likely that they will indeed prove to be reachable from the
current position without any untoward side-effects is a mystery.  It
also doesn't seem to me to involve anything like "pattern recognition"
except in the vaguest of ways.

This is also analogous to the differences between how computers and
humans approach "looking for math theorems".  The computer attempts at
this I have seen described essentially generate lots of theorems and
evaluate the "intreresting-ness" of the results.  (Granted, there is
severe pruning, and attempts are made to only generate
probably-interesting theorems, but the analogous situation is true of
chess also.)  Humans, on the other hand, choose some interesting result
by some means, and then try several strategies to show that the result
is reachable from axioms.

The computer and human approaches are as night and day.

> [...] Who said anything about tree searches?  I certainly didn't.  You are
> reading things into my reply which were never intended to be there. [...]

Yes, indeed I was.  I think I follow your point more clearly,
but still don't agree with it, as I'll detail below.

> All that is necessary for chess playing is to generate the next move.
> Generating extra moves beyond that is only necessary when needed to
> evaluate the possible next move.  The human method, based on pattern
> recognition (of such patterns as control of the center of the board,
> momentum, etc) is generally quite distinctly superior, so very few
> continuation moves are needed for an adequate evaluation. 

To the contrary, human master players evaluate far *DEEPER* than do
computer players (at least, in the mid-game).  They are playing 13, 15,
18 and more plys ahead.  Computers (at the time I read the articles in
Science News (and other such places) about this a year or two ago) were
just starting to reach 8-12 plys.  (At least, according to my
perhaps-faulty memory).  Computers can afford to look far deeper
in the end-game, and equal or exceed human capacity.  But not yet in
the midgame.

We know how computers are overtaking human capabilities in this area.
Sheer brute force.  We have no idea how humans do as well as they do.
Computer chess and computer mathematics have posed some interesting
questions for AI, but I see no clear answers.

But then, I'm not a worker in the field.  "I am only an egg."

Wayne Throop       ...!mcnc!aurgate!throop


