Author: Jay Scott
Date: 17:10:48 02/19/04
Go up one level in this thread
>5. Limitation of the seach node count to a mean of one thousand. While this by
>itself is not a sufficient condition to prove human-like reasoning, it is a
>necessary one.
>
>6. Explicit and extensive use of pattern matching.
>
>7. Explicit and extensive use of planning.
I take it that these features are intended to be human-like.
>8. Natural language output including construction of a search narrative
>describing the reasoning used for a particular search.
This feature is not human-like. Humans can generally say what they looked at,
but cannot say much about why. Humans can give reasons and explanations that at
first seem fully-worked out ("it exploits the light-square weakness"), and
somehow behind the scenes must be, but if you probe ("why that move rather than
another one?") you quickly come to a level where the answers are "because it
looked right" or "that's the one that came to mind".
Another way to say it is that humans use pattern matching and planning in a way
that's "extensive" but not "explicit". The knowledge is implicit, somehow
encoded in low-level routines (perhaps largely perceptual routines) that people
don't have conscious access to the details of, only the results of.
That is, of course, why we don't have human-like programs today: Nobody knows
what humans do.
Since so much human chess thinking happens at the perceptual level (or perhaps
you could vaguely say the "intuitive" level; below the cognitive level, anyway),
I'm doubtful that a symbolic approach is human-like. I suspect that human chess
skill may have a lot to do with exploiting high-level statistical patterns in
the data, which is not a traditional Newell-and-Simon-General-Problem-Solver
kinda activity.
But at another level, that doesn't matter. It's research, what matters is how it
turns out. Go for it!
>10. To perform a search with a mean time of less than one minute.
Because that's how patient you are when debugging?
If I were working on a similar chess program, one of my goals would be to
reproduce the human profile of chess strength versus time spent per move.
Traditional programs find that humans become tougher opponents as the time
control increases. This appears to be true at all time controls from blitz
through correspondence.
Full-width searches slow down exponentially with depth. I interpret the time
control finding as meaning that the human "algorithm" may have asymptotic
behavior better than exponential. That's speculative, but it's true across the
data range we have available today. If it's true it's deep and powerful; it
means traditional chess programs are on the wrong track in the long run, and
it's a hint that computers may be able to play far more strongly than we realize
today.
If Symbolic can ever play even (while trying its best) with one single human, of
any rating, across all time controls, that would be strong evidence that its
play is human-like in one essential respect. I would choose that goal ahead of
any time control or rating goal.
>15. To include a kibbutz facility.
That would be an amazing feature indeed, entirely human-like and well worth the
great investment, but perhaps you meant "kibbitz"?
Jay
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.