Beer in the Brooks

[From Rick Marken (920628.1600)]

Martin Taylor's diagram:

This is, indeed, a nice way of showing that the environment
part of a hierarchical control system is a hierarchical structure
to the extent that the systems observing that environment are also
hierarchical control systems. I tried to make the same point,
verbally, in my "Hierarchical behavior of perception" paper, that
everyone is rushing to avoid publishing.

The diagram is particularly relevant, I think, to the observation
of human behavior (since the behaviors that we see controlled
by people are likely to correspond to perceptions we can have, even
if we can't control them -- though this is not necessarily the
case; it is possible to imagine watching a human behave and have
no idea what s/he is controlling. I think this happens when
we watch certain skilled behaviors -- like that of a surgeon.
It's often hard to see what aspect of the surgeon's environment is
controlled. There can also be intentional obfuscation of a controlled
variable -- as in magic.)

With organisms other than humans, many of the variables that are
controlled by the organism are not normally part of the human
perceptual hierarchy -- like the high frequency chirps of the
bat. Nevertheless, these variables can be detected by artificial
sensors -- but all the human really perceives in this case is
the "meter reading" as the controlled variable.

Technically Sweet <thinman@NETCOM.COM> says:


Welcome to CSG-list, Mr. Sweet.

I've recently discovered Rodney Brooks' work at the MIT AI Robotics'
Lab. If you folks aren't familiar with this, you should be.

We are familiar with him.

He is leading a reaction to AI's traditional robotics goals:

We (some people on CSGNet) just had a long interaction about
the work of a fellow named Randall Beer whose appoach seems
similar to Brooks' (he makes insects, uses "independent behavior
producing agents", etc -- I think the term "subsumption architecture"
is TM Brooks -- but it looks like Beer does the same thing). Beer
and Brooks seem completely uninterested in PCT (our model) apparently
because they are modelling a different phenomenon than we are.
They are modelling "behavior" which to them means certain observable
"outputs" generated by the organism -- like "finding soup cans" and
"collecting them". We are interested in modeling "control" --
the ability or organisms to keep aspects of their own
perceptual experience in internally specified states (the
specifications may be fixed or variable). "Control" and "output
generation" LOOK the same (on casual inspection) but they are not.
We have spent PAGES trying to communicate the difference -- so don't
expect it here in a sentence. If you're interested, stay tuned to CSGNet.

The architectures we use to model control are completely different
than those used by Brooks to model behavioral output. The focus of
our models is the design of the perceptual function and the design
of methods for specifying the intended level of the outputs
of these functions.

I'm personally interested in this work because I want a
method of implementing interesting autonomous non-boring
robots in my VR window system without hooking up a Cray.

Brooks's is probably as good an approach as any if that is your goal
(incidentally, I passed a toy store yesterday and saw some nice
little "robot" puppies barking away in a cage -- poking their noses
through the bars as though they were trying to lick the hands of
passersby. I could barely resist the urge to pet them. These dogs
were not controlling anything -- but they were exhibiting some
pretty interesting behavior. I'm sure the behavior exhibited by
Brooks' insects could be even more "non-boring" -- but it is mostly
NOT control -- though some may be because at the lowest levels
of these architectures there is a closed loop).

It is not easy to understand the difference between behavior as
an interesting display of "outputs" and behavior as "control".
That is one of the main things CSGNet is about. It is a VERY
difficult concept to get -- and it is virtually impossible to
get if you don't believe there is such a distinction in the
first place. Beer seemed quite uninterested in PCT; I believe
Brooks is similarly inclined. They seem to be happy (and famous
too, at least Brooks) and quite disinclined to discuss the issue.
More power to them.

He's going at consciousness from the
bottom up with the same architecture that you folks are
using from the top down.

Actually, we're going at "purpose" (control) -- usually
from the inside out but we're trying to get at from the outside
in too. We're not really very close to putting "consciousness"
into the model yet.

Best regards




Richard S. Marken USMail: 10459 Holman Ave
The Aerospace Corporation Los Angeles, CA 90024
(310) 336-6214 (day)
(310) 474-0313 (evening)