🚀 go-pugleaf

In article <23107@edsews.EDS.COM> clairday@edsews.EDS.COM (matt clairday) writes:
>I wonder if anyone saw the article two days ago in the Wall Street Journal
>concerning an AI project at UCLA?  The article described the work of
>David Jefferson, a CS person there.  He is working on the idea of "growing"
>software.
>
>He desinged a simulated environment for these "ants" where they were judged
>on their "performance" in navigating a simulated trail.  Only the best 10%
>survived during each generation. "After creating 200 generations of ants,
>Mr. Jefferson had a strain of super ants who follow the trail with ease."


The problem with this work is that the ants only learned how to solve the
traversal of a specific path.  If they were given a different path with
the same sort of characteristics, they were not able to traverse it.  Also,
learning one path did not help in subsequent learning of other paths.  If
I remember his talk correctly, learning one path actually hinderend learning
other paths.  It appears that much more work must be done so that 
the techniques he presented are applicable to multiple instances of the
same sort of problem.


kurt

-----------------------------------------------------------------------
Kurt Thearling                               1101 W. Springfield Avenue
University of Illinois                                Urbana, Il  61801
Coordinated Science Lab                          kurt@bach.csg.uiuc.edu 
-----------------------------------------------------------------------