[From Bill Powers (930422.1315)]
Bruce Nevin (930422) --
I got the paper you recommended, "Emergent control and planning
in an autonomous vehicle" by Meeden, Mcgraw, and Blank. What an
amateurish piece of fumbling around. One plus is that I got
Ghostscript to work and was able to print out a fairly decent
copy of the paper on my dot-matrix printer, with diagrams. Seven
pages took about an hour. People who post their papers in
PostScript are not aiming to communicate with the hoi polloi.
The "emergent control and planning" advertized in the title were
completely imaginary.
The first clunker is in the first paragraph, where the authors
define an autonomous agent:
"Such an agent is _autonomous_ to the extent that its behavior is
determined by its immediate inputs and past experiences, rather
than by its built-in control [Russel and Wefald, 1991]."
With a literature citation, yet. So an S-R system is autonomous.
The only slightly interesting aspect of this paper is that it
uses a back-propagation neural net to arrive at more or less
useful input-output connections. That approach may prove
interesting some day.
The "carbot" (made of a toy car) can go forward or backward at a
fixed speed, or stop. It can turn left or right on a fixed
radius, or go straight. All the output variables are on-off.
There are four on-off collision sensors, and two analog light
sensors. In one task, the carbot was to maximize or minimize the
light intensity. That was the only task that looked remotely like
a control problem.
The carbot was rewarded for moving, and punished for colliding
with something. The analog task involved rewarding whenever the
light intensity was higher than on the previous step, or lower,
depending on whether maximizing or minimizing was going on. So
that problem was solved e. coli fashion.
"Reward" and "punishment," as nearly as I can figure out, were
administered by the experimenter or an algorithm. The reward and
punishment were for making certain moves. For example, "in the
_avoid and move_ problem, suppose that the carbot has just bumped
into a wall triggering its front sensors. Any action that moves
it away from the wall and clears its sensors should be rewarded,
while any action that persists in bumping into the wall should be
punished."
One common solution to the avoid-and-move problem was for the
carbot to move forward and backward around one point. This kept
it in motion and avoided collisions by not going anywhere.
I really can't write any more about this paper. Reading it
produces feelings much like watching those people being trashed
in Wacko.
···
---------------------------------------------------------
Best,
Bill P.