From: Michael Roy Ames (michaelroyames@yahoo.com)
Date: Mon Dec 30 2002 - 12:23:45 MST
Ben,
I agree with your comment that a non-AGI program could probably figure
out the earlier games using GA or brute-force methods. However an AGI
with cognitive abilities based solely on these techniques will quickly
'run out of steam'. The early lessons are designed for the learning
benefit of the non-brute-force processes - processes that can *create*
new methods of reaching correct answers, and then apply those methods in
future.
I imagine an AGI is going to have a variety of ways of reaching correct
answers. Some will be 'canned' methods like sequential-search,
random-walk, heuristic-lookup and so on. Others will be 'learned'
methods that are created on-the-fly and found to be useful. These
'learned' methods will be few (or even non-existant) at the start of
teaching, and will require much more processing than brute-force
techniques for the first games. But as the lessons and microdomains
become more complex, the 'learned' methods should come into thier own.
Such methods might come up with correct answers all by themselves, or
narrow the search-space for brute-forcing. I am sure there are many
ways it could be programmed to work... haven't done it yet though ;>
IRT:
> Where things get a bit more interesting is with your idea of games to
> teach conditionals, loops and so forth.
That particular lesson "Lesson 16 - Algorithmic Fundamentals" is too
large right now. It should be split up into two or three separate
lessons, each with a number of games. However, the incremental step
from the previous games to the games in lesson 16 is not great. With
each step I have attempted to make the correct answer 'findable' by a
combination of
a) applying lessons learned in previous games and
b) a bit of semi-random guessing.
I do think this is still true of all the games in lesson 16, but it is
currently not obvious to the reader -- *that* I can correct by adding
more detailed explanations... not only of what is to be learned, but why
it should be 'easy' to learn it.
IRT: "...basal arithmetic..." Yes, that one was already on my list of
games :) but it would *have* to come *after* the learning of symbols
(lesson 16). I decided to cut out a *lot* of possible games by
introducing a new microdomain right after lesson 16. My reason for this
was to expand the learning space in a new direction, a new dimension.
Too much learning within a single microdomain would lead to very narrow
understanding. Beyond a certain point there are diminishing returns
within a given microdomain.
There is a lot more work to do with the doc, and even more fun afterward
creating an automated teacher and all the game data.
Michael Roy Ames
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:41 MDT