NetNews Usenet Archive 1992 #18/NN_1992_18.iso/spool/bit/listserv/csgl/703

Text File | 1992-08-15 | 5.1 KB | 101 lines

Comments: Gated by [e-mail]
Subject: (no subject given)
Path: sparky!uunet!paladin.american.edu!auvm!FAC.ANU.EDU.AU!ANDALING
Message-ID: <[e-mail]>
Newsgroups: bit.listserv.csg-l
Date:         Sun, 16 Aug 1992 16:55:28 EST
Sender:       "Control Systems Group Network (CSGnet)" <[e-mail]>
From:         Avery Andrews <[e-mail]>
X-To:         [e-mail]
Lines: 89

[from Avery Andrews (920815)]
 (penni sibun 920814, on marken 920808)

> on my view, you talk like a cognitivist.  partly cause you oppose pct
> to behaviorism.  more tellingly, cause you explicitly locate control
> in the head, just as cognitivists do.  ...

What I'd say to this is that PCT is not cognitivist, because the *phenomenon*
of control is only found when an organism (or other `Agent') is put in an
appropriate environment (e.g. if the road is covered with an oil-slick,
one ceases to observe control of the direction in which the car is going).
The phenomenon of control is just what Rick says it is: when you introduce
disturbances that you would expect to change some aspect of the environment,
but that aspect doesn't change, because the Agent consistently does things
that `mysteriously' counteract your attempted disturbances.

There is a terminological difficulty here in that in normal usage,
`control' does *not* imply any capacity to cope with unpredictable
disturbances as they arise (I have retained a copy of one of Randy
Beer's postings that documents this point extensively, which I could send
to anyone who'se interested in looking at it).  So for example,
most people would call the spikey cylinder in a music box its
`control unit', but this object does not effect control in the PCT
sense, since it doesn't counteract disturbances in any systematic way.

PCT goes on to say (at least) two further things:

  a) an Agent that is efficacious in the real world must effect control

  b) the only way control can be effected is by means of certain kinds
     of internal arrangments  - perceptors, comparators  & effectors
     appropriately connected so as to constitute negative feedback loops
     in the context of the actual environment.

The argument for (a) is that the sorts of things that serious
Agents might need or be designed to achieve are going to be subject to
constant and unpredictable disturbances, which will then need to be
counteracted as they arise, and the argument for (b) is that no alternative
arrangments that effect control have been proposed (for example, the reason
that Sonja could cope with conveyerbelts, whirlwinds, and, let us say,
crunchbirds that would occasionally eat her monster-killing missiles in
mid-flight, is that her central architecture is full of little control
systems, even if Chapman didn't think of them that way).

So I'd see PCT as a subtype of interactive AI, which makes some additional
claims about how to make sense out of how the internal structure of
an Agent is related to its `behavior' (e.g. what happens in environments
that contain it).  I'd take what I've come up with from cogitating about Sonja
so far as supportive of these claims (and am encouraged that Penni finds it
interesting) but there's still a lot I don't know, & there might be all
sorts of genuinely efficacious but not-PCT organization in Sonja that I have
missed.  But my experience so far is that PCT ideas really are
helpful in figuring these things out (recall that I got on to Chapman
& Agre when I showed Penni my proposed control system organization for
getting beer, & she said that it was like their stuff).

What I would like to do is design and/or analyse some relatively large-scale
agent along PCT lines - the crowd people and astro are a start, but they
are obviously too rudimentary to convince people of the utility of
the approach, in part because they don't have any `tactics'.  Crowd
people, for example, avoid obstacles, but they don't really navigate
in the way that Sonja does.  if you set things up so that there is a
wall of stationary people between a mobile crowd person and her goal:


                               P
                               P
             G                 P                            Goal
                               P
                               P
                               P
                               P


(you can position stationary people by making them `active', but setting
their gains to zero), G will head straight for the Goal until she is about
to crash into the row of P's, and will then veer off one way or the
other, not consistently taking the shortest way.  But Sonja (or my dog)
would head immediately to whichever end of the wall would give the shortest
path around.  I conjecture that the way to get the crowd people to navigate
would be to give them more sophisticated perceptual functions (maybe even
tuning the ones they have would do the trick - tho I haven't managed
to do this), but to say this is not of course the same as to actually
do it (Sonja in effect controls for not having a wall between where she
is and where she wants to be going, according to my current understanding).



[e-mail]