Bio-robotics meeting

Last Thursday I was in London at a seminar on Bio-robotics: the
collaboration of robotics and biology. Over coffee I casually mentioned to
one of the delegates that I had a 24-jointed six-legged insect simulation
that could control its body position in all six degrees of freedom
simultaneously, against random disturbing forces and torques, even to the
extent of surviving having legs shot off, without any inverse kinematic
computations, but without drawing any spark of interest.

Richard Dawkins and Francisco Varela had been down to speak, but withdrew,
a pity. Some of the talks were on robotics, and some on biology. There
wasn't much about how to make a robot balance or walk. The emphasis was
mainly on getting robots to perform more complicated tasks, and experiments
on how insects perform such tasks, e.g. navigating back to their nest.

On the robotics side, all the thinking was either input -- make plan --
output, or input -- neural nets -- output. One demo would have been an
example of the behavioral illusion, if anyone there had known what that
was. A film was shown of three robots running around an arena a couple of
metres square, in which there were scattered a number of styrofoam cubes.
Eventually, all the cubes were pushed together into a few clumps. What,
asked the presenter, are these robots doing? Suggestions included
"cleaning up", "making space to move around in", "making clumps". In fact
the robots were simple S-R machines with a few rules: turn left if there's
something to the right; turn right if there's something to the left;
otherwise go straight ahead. (Which suggests that you can't even tell what
an S-R system is doing by watching what it's doing.)

There was a talk by a behavioral psychologist on incentive learning and
behavioral shaping, and one by a roboticist on using these to train robots.

I found the robotics papers pretty shallow: just "we programmed this with a
neural net/genetic algorithm/subsumption architecture, ran it on a toy
problem, and here's a cute video". Reminiscent of where AI was twenty
years ago -- is robotics just the latest fae of AI? Some of the videos
were certainly cute, e.g. a two-legged robot doing somersaults. I found
the biology more interesting. Here are some feats that some animals can
do:

There is a type of desert ant that can leave its nest, search in a very
convoluted path for food, and then run back to the nest in an almost
straight line.

If, before the ant starts running back, it is picked up and moved, without
being able to see where it is being moved to, then when it is released, it
runs in the same direction as it would have if it hadn't been moved. It
runs as far as would have been required to reach the nest, and then starts
wandering about.

If it is moved, and then prevented from running for several hours, it still
runs back in the same direction when released. It appears to be using the
polarisation of sunlight to provide a reference for measuring direction,
but that changes with time, and it is able to allow for that.

These ants can run at about 1 metre/second, and turn with an angular rate
of 4000 degrees/second.

When a rat is placed in a rectangular box, it is possible to record
individual cells in the hippocampus which exhibit increased firing when the
rat is in a given position in the box. These cells together appear to map
out the box.

A quote: "organisms are driven by desires, not by [evolutionary] fitness."
A good observation. The speaker was making it in the context of the use of
genetic algorithms to evolve robot control algorithms. He raised (but did
not answer) the question of how much in the way of desires (references of
control loops, as we would call them but he didn't) should be built into
robots.

-- Richard Kennaway, jrk@sys.uea.ac.uk, http://www.sys.uea.ac.uk/~jrk/
   School of Information Systems, Univ. of East Anglia, Norwich, U.K.

[From Bill Powers (980218.1502 MST)]

Richard Kennaway (980218) --

Over coffee I casually mentioned to
one of the delegates that I had a 24-jointed six-legged insect simulation
that could control its body position in all six degrees of freedom
simultaneously, against random disturbing forces and torques, even to the
extent of surviving having legs shot off, without any inverse kinematic
computations, but without drawing any spark of interest.

etc.

This is why I turn down all invitations to such meetings (not that I've
received anything but the usual form letters). I've reached the end of the
line on this sort of thing -- maybe you're smarter than I am and will give
up sooner.

I don't mean give up on PCT. I mean give up on looking for anything but
competition from those who consider themselves in the forefront of
behavioral modeling, robotics, and all that.

Here's what I think.

I think we've established that the PCT model is basically right. I think we
have a lot of simulations and demos that give us an idea of where to go
next. I don't think that these hot-shots from MIT and other such places
have a clue about living control systems, nor do they want to hear from
anyone outside their circle unless it's to pat them on the back and tell
them how right they are.

I also think that we have a pretty good number of smart people who are
starting to do research and modeling in PCT. We don't have to ask other
people what is interesting about human behavior, what problems need
solving, or what research projects are worth doing. We can decide for
ourselves what's important, and set up our own research agenda. We can
write for each other, criticize each other, and try to understand each other.

In short, we can just go ahead and reinvent behavioral science. We don't
need anyone else's approval. We don't need grant money or institutional
support. Most especially, we don't need permission from the self-proclaimed
leaders in psychology or any other field. We can just go ahead and do it.

Let's get that bug working, and then apply what we've learned to a model of
a walking, running, jumping, and standing still person. Then let's take the
working simulations around to people who want to make money by building
exploration robots and humanoid robots, and sell the model to them for
further research support. Let's start a counseling system using the method
of levels and whatever else makes sense, and sell it to insurance companies
with a stake in efficient psychotherapy. And let's write all about this,
not for journals but for magazines and in books, to let people know what
we're doing and to get popular support.

Let the rest of them catch up when they wake up.

Best,

Bill P.

[From Bruce Abbott (980218.1935 EST)]

Bill Powers (980218.1502 MST) --

I think we've established that the PCT model is basically right.

Bill, this might be a good time for you to explain what you mean by "the PCT
model." In this forum we talk about PCT one moment and about HPCT the next,
as if there were a distinction to be drawn between them. Another time
someone will state that PCT doesn't predict something, and the reply will
come, "ah, but you forget, this is a hierarchical system of many levels," or
"what you are overlooking is reorganization." But that is HPCT. _Is_ there
a distinction to be made there, or has PCT become just a shorthand for HPCT?

When you say "we've established that the PCT model is bacially right," to
what PCT model do you refer?

Regards,

Bruce

[From Bill Powers (980219.0347 MST)]

Bruce Abbott (980218.1935 EST)--

Bill, this might be a good time for you to explain what you mean by "the PCT
model." In this forum we talk about PCT one moment and about HPCT the next,
as if there were a distinction to be drawn between them. Another time
someone will state that PCT doesn't predict something, and the reply will
come, "ah, but you forget, this is a hierarchical system of many levels," or
"what you are overlooking is reorganization." But that is HPCT. _Is_ there
a distinction to be made there, or has PCT become just a shorthand for HPCT?

PCT is the basic control loop. HPCT comes from asking "what generates
reference signals?" and the reorganzing system comes from asking "how does
the specific organization of a control loop come into being?"

The PCT model is a direct application of control theory. When there is a
closed causal loop, the ordinary cause-effect interpretations of behavior
become inadequate for explaining what is happening, particularly in terms
of quantitative predictions. Control theory shows us how to analyze a
closed causal loop so we can understand its operation correctly and make
correct predictions about its behavior. That analysis is the basis of PCT.

Best,

Bill P.