[From Bill Powers (930507.0800 MDT)]
I've seen a couple of references to a post by Ken Hacker -- have
I missed something with the subject "Pseudo-science?" My mail
sometimes arrives with inexplicable (by me) delays.
Chuck Tucker (930507) --
Re: Problem Solvers' Associates
I wonder how many of your students will realize that they can
apply PCT to Problem Solvers Associates itself. What is a
"problem?" It's a difference between the way things are perceived
and the way they are supposed to be, in someone's opinion.
Suppose there's a group problem like rivalry that is threatening
to disrupt the group. Before this can even become a "problem"
someone has to have a reference level for existence of the group
(if nobody cares whether the group remains intact, then
disruption of the group will proceed without objection). In other
words, the problem itself must represent an error in at least one
person's control systems, or there will be no attempt to solve it
(remove the error).
Once the problem with the group exists as a control-system error
in one or more persons, one or more of them can try to analyze
it: it is the result of a mismatch of perceptions about what the
group is trying to do? Is it a difference in goals? Is it a
direct conflict between actions? If the group is a football team,
a coach may be quite aware that there is rivalry (for, say, the
position of quarterback), but perceive that one person interprets
this as a chance for personal glory while the other sees it as a
way to rescue the team from bad quarterbacking by the rival. The
coach would try to get the players to agree on what is best for
the team instead of what is best for the person, or show how what
is best for the team is, at a higher level, also what is best for
the person (this isn't necessarily true, but the coach probably
believes it is). Or both quarterbacks might genuinely want what
is best for the team, but perceive what is best differently. Then
the coach could work on teaching football strategy so that all
his quarterbacks would understand it the same way.
So, PCT could be used to design the strategies of the operatives
of PSA in approaching client problems. Problem-solving is just an
example of PCT.
ยทยทยท
-------------------------------------------------------------
Greg Williams (930507) --
Having HPCT models which behave won't be very impressive to the
skeptic who will wonder about how much of their performance
depends on the SPECIFIC arrangement of control loops postulated
by Bill (higher- level errors altering lower-level reference
signals).
The performance never depends on the SPECIFIC connections. There
are always other possible connections that do the same thing. The
trick is to think of ways to decide between the alternatives
until you can only think of one possibility. That doesn't mean
that other possibilities don't exist; it just means that you
can't think of any. Of course the alternatives have to involve
differences that make a difference; otherwise they're trivial.
Alternatives to HPCT (some perhaps hierarchical themselves, but
not structured exactly as Bill Powers postulates) exist which
DO behave to various extents (the Beer bug behaviors appear a
lot closer to reality than do the Rick spreadsheet behaviors,
in my opinion, and there are lots of models in BEHAVIORAL
CYBERNETICS with rich behaviors).
I suppose that what most people mean by a model's "behaving" is
showing graphics moving around on a screen, or some kind of
mechanical model moving around in a room. I suppose I'd better
work more on that sort of thing. Basically, however, when you've
given the bug or crowd element specific perceptions and have
decided how those perceptions will depend on consequences of
outputs, you've defined the environment as far as the acting
system is concerned; how it looks to someone else is rather
irrelevant (except for salesmanship). The numbers in the model
tell the whole story that matters.
As in the crowd program, people looking at the behavior see a lot
more than the behaving system does, and read all sorts of
spurious organization and intelligence into the behaving system;
I wonder if this doesn't just promote the old externalized way of
understanding behavior. Would I be working against myself?
Would it be worth my while to subscribe to BEHAVIORAL
CYBERNETICS? How expensive is it? Unfortunately, Fort Lewis'
library doesn't have this kind of journal.
If you just want behaving bugs, like Beer's, you can sit down and
design one that will do what you want. This is just mechanical
engineering and circuit design, created by a brain more
interested in seeing some pattern than in how it is achieved.
Different engineers would accomplish the same result by different
means.
I remember that at one time you claimed that Beer wasn't just
doing circuit design, but was using the properties of neurons
that are actually there. In fact, all the neurons and circuits he
used were simply made up to do what he imagined is actually done
(see p. 49 ff in _Intelligence as adaptive behavior_). The
circuits weren't particularly elegant, but they did produce
visible behavior. He didn't take advantage of a lot of known
neurons, like those that respond to leg position and velocity (he
used only the hair cells that detect, binary fashion according to
him, limits of motion). His motor systems were pretty much open-
loop with respect to control of propulsion (that's why he used
such a weird model of the physics, in which velocity was
proportional to force. With a correct physical model, his
circuits wouldn't have worked, He let the strange environment do
what his circuits couldn't).
Above the level of generating the leg movements, Beer did come up
with some crude control systems. They tended to be on-off
systems, so he had to use separate subsystems to stop movements
as well as to create them; he used a lot more circuitry than was
actually necessary. He even had to use a separate system to
create random searching behavior, instead of just letting this
apparent randomness emerge from the interaction of different
control systems with varying degrees of error.
Unfortunately, approaching such problems the way I do doesn't
much impress people who are only interested in the outside view
of behavior. I'm sure many people have been disappointed by
Little Man Version 2 -- if it's such a radical improvement over
Version 1, why does the Little Man still just reach out and touch
the target in the same way Version 1 did? The fact that the model
incorporates actual arm dynamics, a muscle model, a visual map,
and a model that explains, for the first time, how the stretch
and tendon reflexes work together, remains invisible to the
observer who only looks at the actions of the model. For such an
observer, version 1 is just as good as version 2, and maybe a
little better because it behaves more smoothly (and faster on a
slow computer).
I think that the skeptical physiologist confronted with ONE
model which matches (more or less) behavioral reality should
insist that (to begin to make choices about which model is
"best") EITHER HPCT models be made for behaviors currently
modeled in nonHPCT ways (Bill began doing this for the Beer bug
but didn't finish ...
One reason I haven't gone on with the cockroach model is that I
haven't solved the propulsion problem yet. Neither, in fact, did
Beer. Beer solved a put-up problem, how to produce gaits. What he
didn't solve was the problem of how the bug could slow down,
stop, and reverse, with all the proper phase relationships
remaining pretty much the same, and how the bug could move
sideways (crablike) with the same phase relationships and the
same ability to control speed reversibly and smoothly through
zero. Solving that problem will also solve the gait problem.
My real Soldier Beetle pet could do all that beautifully; I
presume a cockroach could, too, but nothing at that level of
detail was reported in the materials I saw. Beer accomplished his
gait control using oscillators, and oscillators can't do what the
bug actually does. However locomotion occurs in bugs, it isn't
accomplished by relaxation oscillators, not as Beer designed
them. But if all you want is for the bug to walk over some range
of speeds, a relaxation oscillator or a shift register or coupled
integrators or a tiny computer program will do the job just fine.
But Beer did made his bug walk, so he wins the impress-the-
observer contest.
So I think what is needed is a RANGE of models for ONE kind of
behavior, to see if HPCT can do the job "better" (match the
data of real organisms) better than the alternatives.
It's hard enough to come up with just one model that does exactly
what the real system does. The only way you get multiple models
is by not looking very hard at the actual behavior. When you
observe the system closely, there are enough details to consider
that the range of possible systems gets pretty narrow. In the
case of bug locomotion, the range is zero in my head just now.
But I have a feeling that there is a simple and elegant solution
trying to tell me about itself. That's why the bug model is still
sitting on a shelf in the back of my head, ticking. I almost have
it; I have parts of it, like how integrators can control forward
speed by varying the rate of charge and discharge, under control
through a single speed reference signal. The rest of the circuit
is still too clumsy and ad-hoc; it requires limit detectors and
flip-flops. In the real system those things are going to happen
because of the way sensors in the feet detect the weight of the
bug and contact with the ground (those sensors exist), and
through interactions between control systems that maintain the
body at a controlled height above the ground and those that
control forward -- and sideways -- speed, as well as differential
speed for turning.
My feeling is that when you consider the behavior you're trying
to explain carefully enough, there won't be a "range" of models
that "explain" it. There will be only one possible (minimal)
model. I exclude, of course, details such as whether three
signals are added together by a three-input neuron, or by three
two-input neurons. Below a certain level of neural detail there
are always alternate ways to accomplish a given function. I
wouldn't be surprised to find that many of the alternatives could
be found in different individuals of the same species.
You only get a range of feasible models when you're looking
superficially at a behavior. Looking more closely quickly
eliminates many alternatives; in fact, I think one should always
keep looking more closely until only one alternative is left.
That's pretty much what I do. If I can think of two ways of doing
something I try to see what there is about behavior that might
eliminate one of them. Sometimes I eliminate all of them, which
is where I am now with the bug.
Tom Bourbon asked
Do you think the spreadsheet and arm models offer any
support for speculations that PCT can be expanded to HPCT?
And you replied
I do, but the question is whether a skeptical physiologist
would count these as more than examples of HPCT NOT failing,
with the question of what other kinds of models ALSO might not
fail left open.
This makes me wonder whether I really care what a skeptical
physiologist would think. It's hard enough to come up with one
specific and compelling model that fits the facts. Having some
person hanging over my shoulder saying "Yeah, but I'll bet there
are lots of other models that would work, too" makes me wonder
how my cup of coffee would look all over the jerk's face. If he
thinks there are other possibilities, let him produce them.
Otherwise, stuff a rag in his mouth.
Perhaps what we need to do is issue a challenge for an open
competition to model a particular kind of behavior?
That sounds good to me. If we get to pick the behaviors that are
modeled, and the conditions under which that behavior has to
work, it will produce a lot of blessed silence .
-----------------------------------------------------------
Best,
Bill P.