[From Bill Powers (920302.1500)]

Randy Beer (920202) --

Don't feel in a rush to answer anything. I have plenty to do, too, and
quite understand.

... you state that one advantage of your proposed
control system over the current appetitive and consummatory circuits
is that it would allow the insect to move as food is eaten or moved
out from under it. The implication seems to be that the present
circuits do not have this capability. But, in fact, they do.

I agree, they do.

I seem to detect a strong aversion to having "superordinate" control
systems turning off subordinate control systems in your comments.

It's a mild aversion. I like to see as much as possible emerge from a model
without being explicitly put into it. I don't think you would like a model
in which the direction from the animal to the food acted to turn the animal
in that direction (i.e., if the food were at azimuth 60 degrees, an ad-hoc
neuron would rotate the animal's body by 60 degrees). I think the
underlying idea is to try to keep the modeler's intelligence out of the
model as completely as possible (an ideal which, I realize, all modelers
are forced to violate frequently for lack of data or clever enough
designs). The model should operate strictly from its own organization,
using only information available to it in the form of sensory signals. I
think your model does fairly well in this regard, although some parts of it
(particularly obstacle avoidance and edge following) seem a bit awkward to
me (one of my scientific criteria).

With regard to the "search" neuron, I just think the model would be
prettier if it didn't have discrete modes. Maybe they're unavoidable. But
by using the approach circuitry I suggested, you can leave the food-
approach system turned on all the time, with hunger bringing it
automatically into action, and proximity to food stopping the forward
motion without any need to recognize the abstract condition "Now I'm at the
food and don't need to move any more." In the same way, "wandering" doesn't
have to be a special condition any more: when the animal is satiated
(energy error is zero), other motivations can make it move, with the noise
level in the food-direction sensors providing the random bends in the path.
By leaving all these systems on all the time, you can get some interesting
(and I think real) conflict situations, and see how the bug resolves them.
In my model of crowd behavior, the individuals use goal-seeking and
obstacle-avoiding circuits that run concurrently, and the resulting
behavior as a "person" finds a way through a crowd of other people to a
goal looks far more intelligent than it actually is. The "people" will even
retrace paths to get out of traps. That's the effect I like to see: minimum
design, maximum apparent complexity of behavior.

So you can see that my scientific criteria for models rest in good part on
such subjective notions as niceness, prettiness, and simplicity. This
applies, of course, only when the data don't constrain us to a particular

I'm still working on simplifying the locomotive system. The data on the
real cockroach's gait, p. 82, show that the forward swing wave begins as
the leg moving backward on the other side passes the midpoint of its travel
during the stance phase (this seems to be true at all speeds). A circuit
that detects this midpoint position of the rear leg and generates a pulse
can trigger the swing wave on the contralateral side. Can I assume a sensor
that responds continuously to leg angle, or do I have to do this open-loop
using the motor driving signal? I'm going to make leg angle proportional to
the driving signal, by the way, rather than using a pseudo-force output
with some rather odd physics involved. If leg angle sensors exist (other
than the limit sensors), I can make leg angle a controlled variable.

If the triggering of a swing wave can be made automatic, then the speed
control circuit becomes very simple: just a time integrator. I don't know
if this is going to work out, but it looks promising so far.

My swing-wave generator will produce ALL gaits, with the tripod gate as a
natural limiting case at high speeds. It continues to work at all speeds
down to zero. You were right, by the way, in pointing out that more than
one leg can move at a time on one side: I didn't look closely enough at the

Reversal of direction is going to be interesting. In real cockroaches, is
the swing phase still initiated in the rear legs while traveling backward?
Or does it start at the front?



Are you interested in how nervous systems control behavior? Are you
interested in designing artificial autonomous agents? Are you interested
in how the artificial insect works? Are you interested in a rational
reconstruction of the artificial insect using HCT?

All of the above, but the emphasis is on modeling the behavior of real,
particularly human, organisms. The "hierarchical" aspect of the modeling
may be less important in simple organisms, because the goals are going to
be pretty simple and there won't be many levels of organization. More
important is the concept of control -- the idea that the system VARIES its
actions to CONTROL variables defined by its input apparatus. As an example,
instead of thinking of the food patch as causing turning behavior via the
odor sensors, think of varying the direction of motion as controlling the
unbalance of the sensor signals in an odor gradient field. This shifts the
viewpoint from the observer-centered laboratory system to the organism-
centered control system. In simple systems this shift of viewpoint doesn't
make much difference, but when things get complex it can make the behavior
much more understandable. It also helps one avoid putting too much of the
observer into the model -- when you think of control of perception, the
world that matters is the one represented by sensor signals, not the one
the observer sees. Most of the "behavior" we observe is a more or less
irrelevant side effect of what the control system is really doing.

When I get a working model, by the way, I'm just going to dump it in your
lap to do with as you please. I have no desire to publish in this field. I
hope you'll see some principles in it that will interest you. I'd like to
see you start using the CT orientation in your work, but I have no urge to
compete in your area. I figure the best way to recruit you is to
demonstrate the CT approach using something dear to your heart like your
pet cockroach. When we get past this preliminary stage, I think you'll
start seeing some real power in the CT approach. Then you can start
teaching us things about real neurons in real control systems.

Best regards,

Bill Powers