On Thu, Oct 24, 2013 at 11:05 PM, meekerdb <meeke...@verizon.net> wrote:
> On 10/24/2013 12:08 PM, John Mikes wrote:
>
> Craig and Telmo:
> Is "anticipation" involved at all? Deep Blue anticipated hundreds of steps
> in advance (and evaluated a potential outcome before accepting, or
> rejecting).
> What else is in "thinking" involved? I would like to know, because I have no
> idea.
> John Mikes
>
>
> Learning from experience.  Actually I think Deep Blue could do some learning
> by analyzing games and adjusting the values it gave to positions.  But one
> reason it seems so unintelligent is that its scope of perception is very
> narrow (i.e. chess games) and so it can't learn some things a human player
> can.  For example Deep Blue couldn't see Kasparov look nervous, ask for
> changes in the lighting, hesitate slightly before moving a piece,...

Bret,

Even in the narrow domain of chess this sort of limitation still
applies. Part of it comes from the "divide and conquer" approach
followed by conventional engineering. Let's consider a simplification
of what the Deep Blue architecture looks like:

- Pieces have some values, this is probably sophisticated and the
values can be influenced by overall board structure;
- Some function can evaluate the utility of a board configuration;
- A search tree is used to explore the space of possible plays,
counter-plays, counter-counter-plays and so on;
- The previous tree can be pruned using some heuristics, but it's
still gigantic;
- The more computational power you have, the deeper you can go in the
search tree;
- There is an enormous database of openings and endings that the
algorithm can fallback to, if early or late enough in the game.

Defeating a grand master was mostly achieved by increasing the
computational power available to this algorithm.

Now take the game of go: human beings can still easily beat machines,
even the most powerful computer currently available. Go is much more
combinatorially explosive than chess, so it breaks the search tree
approach. This is strong empirical evidence that Deep Blue
accomplished nothing in the field of AI -- it did did accomplish
something remarkable in the field of computer engineering or maybe
even computer science, but it completely side-stepped the
"intelligence" part. It cheated, in a sense.

How do humans play games? I suspect the same way we navigate cities
and manage businesses: we map the problem to a better internal
representation. This representation is both less combinatorially
explosive and more expressive.

My home town is relatively small, population is about 150K. If we were
all teleported to Coimbra and I was to give you guys a tour, I could
drive from any place to any place without thinking twice. I couldn't
draw an accurate map of the city if my life depended on it. I go to
google maps and I'm still surprised to find out how the city is
objectively organised.

If Kasparov were to try and explain us how he plays chess, something
similar would happen. But most AI research has been ignoring all this
and insisting on reasoning based on objective, 3rd person view
representations.

My intuition is that we don't spend a lot of time exploring search
trees, we spend most of our time perfecting the external/internal
representation mappings. "I though he was a nice guy but now I'm not
so sure" and so on...

Cheers,
Telmo.

> Brent
>
> --
> You received this message because you are subscribed to the Google Groups
> "Everything List" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to everything-list+unsubscr...@googlegroups.com.
> To post to this group, send email to everything-list@googlegroups.com.
> Visit this group at http://groups.google.com/group/everything-list.
> For more options, visit https://groups.google.com/groups/opt_out.

-- 
You received this message because you are subscribed to the Google Groups 
"Everything List" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to everything-list+unsubscr...@googlegroups.com.
To post to this group, send email to everything-list@googlegroups.com.
Visit this group at http://groups.google.com/group/everything-list.
For more options, visit https://groups.google.com/groups/opt_out.

Reply via email to