Re: Deep Blue is officially lame

From: Eliezer S. Yudkowsky (sentience@pobox.com)
Date: Tue Jul 31 2001 - 12:14:25 MDT


Brian Atkins wrote:
>
> There's a new little blurb/article on kurzweilai.net today about a new
> book coming out called Blondie24: Playing at the edge of AI. It shows
> how properly combining neural networks, genetic algorithms, and the
> traditional brute force methods of computer game playing allows this
> software to beat traditional game playing software on the same hardware,
> even though the traditional software was able to look a lot farther
> ahead.
>
> http://www.kurzweilai.net/meme/frame.html?main=/articles/art0234.html

Deep Blue isn't lame quite yet. This is about checkers, rather than
chess. Impressive checkers-playing ability was achieved very early on in
AI - 1963 - and not as the result of a traditional tree-searching
program. It was instead a program that memorized a lot of board
positions. For example, suppose the program looked three moves ahead from
a certain position, and assigned that position a certain value based on
the minimax of the lookahead. That value was then stored in memory and
future lookaheads which touched on that position would use the stored
value, rather than the value assigned by the primitive position-evaluation
function.

The winning program thus also had the property that it was very impressive
in opening games and midgame, but tended to wander a bit during the
endgame. The combination of the fascinating (well, back then) algorithm
and its ability to actually beat a human opponent (a huge milestone for AI
back in 1963) made Arthur Samuel's checkers program one of the classics.
Although the game wasn't as "solved" as it was then thought, since it
still took until 1994 for a new program called Chinook to become world
champion.

So it's been known for a long time that learning is the key to checkers.
Furthermore, a very simple learning algorithm in checkers sufficed to
yield impressive performance using very little computing power. It isn't
at all surprising that a sophisticated combo of neural nets and genetic
algorithms can beat a checkers program based on blind lookahead; a
checkers program based on an extremely simple, almost "blind" form of
learning can do the same thing.

I'm not saying that the neural/genetic combo had less effect than was
claimed - just that, from the description, and from the choice of problem
domain, it's pretty hard to evaluate how much of an advance, if any, has
been made. Furthermore, the checkers domain is much smaller and more
regular than the chess domain, which may have helped them to develop
neural nets acting on it, but also makes those neural nets less
impressive.

-- -- -- -- --
Eliezer S. Yudkowsky http://intelligence.org/
Research Fellow, Singularity Institute for Artificial Intelligence



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:37 MDT