From: Ben Goertzel (email@example.com)
Date: Mon Dec 30 2002 - 11:00:22 MST
I think this is a great direction to be thinking in, and a good .5'th
Now a few specific comments..
I think that everything up thru section 13 in your document could be done
fairly straightforwardly by a genetic programming system with a reasonable
number of training examples and a reasonable amount of processing power. By
"reasonable" I mean: Hundreds of training examples for each game, and
perhaps a few dozen machines of processing power.
Where things get a bit more interesting is with your idea of games to teach
conditionals, loops and so forth. But this part of the document is very
sketchy, and I'm looking forward to seeing some details and examples of this
stuff in the next draft... For all the games prior to this algorithmic
stuff, I see nothing that distinguishes AGI systems from machine learning
algorithms that learn rules based on training examples. [Note that it is
precisely the "microworld" nature of the tasks that will, I conjecture,
allow simple ML algorithms to do very well on them.]
I'd like to see more tests specifically intended to require *abstract*
learning. For instance, suppose one teaches a system to do arithmetic in
bases 2 and 5. Does it then learn more quickly how to do arithmetic in base
4, if given some training examples? I.e., when being taught arithmetic in a
number of bases, at what point does it shift from learning a separate model
of arithmetic in each base, to learning a general model of basal arithmetic
with a "base" parameter as an input? Probably more tests could be
formulated along these lines....
-- Ben G
> -----Original Message-----
> From: firstname.lastname@example.org [mailto:email@example.com]On Behalf Of Michael
> Roy Ames
> Sent: Saturday, December 28, 2002 11:20 PM
> To: firstname.lastname@example.org
> Subject: Curriculum for AI
> Let the games begin!
> A few weeks ago I started trying to put together a teaching curriculum
> for Friendly AI. After making a couple abortive attempts, I realized
> (even more than before) that there was one hell of a lot of things that
> an AI needed to learn before ve could understand the first thing about
> Friendly content. So, I punted the idea of Friendly content up a level
> (or five), and attempted to figure out how to teach an AI what it needs
> to know before that. I created some lesson outlines, game descriptions,
> teaching goals and microdomain definitions. I now have about 40 pages
> of (rather incomplete) notes. They could easily be fleshed-out to 240
> pages... but... before doing that perhaps I should take a reality-check.
> These are links to three formats of the notes:
> HTML - Readable stuff starts after a long table of contents.
> RTF (Works with Abiword)
> Described within is one way (out of many) to teach an AI from 'scratch'.
> The teaching format may seem a little restrictive or 'constrained', but
> I made it that way so programming an automated teacher would be fast and
> The basic idea is to teach an AI through the playing of games. Starting
> with very simple games, and gradually getting more complex. By playing
> the games in various microworlds it is intended that the AI build a
> variety of concepts without being overwhelmed with complexity. These
> concepts can then be applied to real-world real-time data, and could be
> used to acquire a small subset of human language (and various other
> skills). Education through game playing... hey, it works for kids!
> What problems do you see with this approach?
> Comments, new game ideas, and new microdomain ideas welcome.
> Michael Roy Ames
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:41 MDT