Re: Innoivationm and the Curse of Knowledge, etc

casey wrote:
On Dec 31 2007, 3:56 pm, "Wolf K." <wolf...@xxxxxxxxxxxx> wrote:
...

Well, actually what my reading about simple
machines has shown me is biological machines
are rather more complex than I had imagined.

I first thought that it would be simple to
build a conditionable machine when I saw the
turtles that the Engineering students had
built (this was some 50 years ago). You know
the ones I mean: Two or more photocells are
coupled to motors through a balancing circuit
so that when one cell receives more light
than the other the turtle turns until all
photocells receive the same amount of light.
I was pretty sure that the Engineers were
only a few circuits away from building a
trainable turtle, one which followed you
rather than someone else's flashlight.

The problem with building a robot that can
follow you instead of a simple light source
is recognizing you as a stimulus. A simple
photocell will respond to a variable light
source but will not respond to something as
complex as you.

Instead you need an array of photocells the
patterns of which have to be processed to
produce an invariant output called "Wolf".
The interesting question is how a system
might learn to recognize "Wolf" as a stimulus.

I don't see the complexity in connecting two
stimuli, I see the complexity in getting the
two stimuli to connect.

One of the first attempts at Pavlovian type
conditioning I believe was Dr. Grey Walter's
Cora. Its response unit was a neon tube which
flashed to a light stimulus. If a whistle
was sounded as a prelude to the light then
the whistle alone would start to cause the
neon tube to flash.


But that's because I didn't understand that
simple machines like phototropic electric
turtles aren't complex enough to reproduce
even simple conditioning.

And what do you mean by _simple_ conditioning?

Not even "association" (quotes because I don't
believe any more that I know what AI engineers
mean by it.)

Association simply means two things are connected
in some way by whoever or whatever is doing the
connecting. It is about how two things are bonded
together.

Two things might be connected by occurring together.
Two things might be connected by having some parts
in common. Whatever you imagine association means
a program is always a definition of what the word
means with respect to the program. So if you want
to know what an AI engineer means by the word you
only have to read the program or study the machine.


Anyhow, it's a two-street: GS notices something
that advocates of "simple RL machines" haven't
noticed: that "trainable artificial neural
networks" (for example) do not reproduce even
classical conditioning.


Why? Because you can't condition a system (animal)
unless it has at least two built in behaviors.
Why? Because conditioning means connecting a
stimulus to a response that was not connected to
that stimulus. But that implies that the desired
response was connected to another stimulus...


All the trainable ANNs I've seen or read of were
attempts to reproduce "association" of a stimulus
to a response, on the incorrect assumption that
conditioning consists of making such associations.
What the TANNs have shown is that conditioning is
rather more complicated than that. As I've tried
to explain.

Not sure how to untangle the above. I take it you
are saying that current TANNs do not show the input
output behaviors you would call "conditioning"?

It may be helpful to drop the term "conditioning"
and replace it with the actual observations and
a different set of terms.


When I first read about the Skinner box it
seemed to me that the rat was being replaced
by simple measurable variables such as lever
pushing. The rat could be put in a 'black box'
and the relevant input/output be all that was
seen. You could then play the imitation Turing
game with another 'black box' containing
electronic components and play spot the
difference.

The difference would be easy to spot. Just change
one or more environmental factors, such the color
of the light, or the timing between lever press
and pellet release, or randomly not release a
pellet, or whatever. Come to think of it, an
electronic box that shows the typical strengthening
of the conditioned response with random changes in
timing would be an interesting challenge to build,
don't you think? Or program...

Programming is building a machine using the components
of a general purpose computer.

Of course similarity of behavior does not mean the
mechanics for producing the behavior is the same.
You might change your schedules of reinforcement
and find the black box with the rat inside and your
black box program cease to show the same behavior.

Machines can be understood and studied in their
own right without reference to biological machines
which may have inspired their construction.

So a simple program might be given hand written
characters as input which have to be converted to
the appropriate output name. If it produces the
wrong output name we press the "punish" button.
After a while if it starts producing the right
output we might say it is showing learning behavior.

You can collect your experimental data showing how
the number of correct responses changes over time.
The resulting graphs may not be the same as a human
would produce if learning to classify a set of
previously unknown patterns but it may still be
appropriate to say they both exhibit learning
behaviors.



To me the behaviorists were not studying
animals "in all their complexity" but rather
restricting the environment and selecting
a set variables such as lever pushing with
lights, bells, food, over time, to _simplify_
the system they were studying.

You might just as well say the physicists don't
study nature in all its complexity, but rather
restrict the environment which they study and
select set variables such as time of rolling
down a slope, or dimming curve of a star or ....

That is exactly what they do. They have to find
a set of variables that change over time in a
predictable way. Given any state at time t defined
by the values of the variables at time t the next
state (set of values) would show a single valued
transformation. It would be a determinate system.


Of course you have to simplify the system you want
to study. When observing animals in their habitat,
it's damn hard enough figuring out which of X, Y, Z...
_might_ be a discriminant ...

I understand all that. An analogy might be complaining
that a computer model of two interacting planets based
on Newton's laws did not explain the Solar System "in
all its complexity". This seems to be GS's complaint
with regards to simple machines. But complex machines
are built out of simple machines.


I'll give you an example of an animal behaviour
"in all its complexity" which reduces to a few
simple rules of behaviour: the flight of a flock
of birds. < snip explanation>

BTW, the above explanation of the flight of a flock
of birds was AFAIK first proposed or taken seriously
by computer animators who were trying to simulate
the appearance of a flock of birds. Seems to me they
"reduced" the complexity of that flight to a very
elegant, simple model.

Yes, same problem, _maybe_ similar solution although
you would have to study real birds to know for sure.

And as you pointed out this complex behavior results
from some simple behavior from each bird. The very
point I make all the time. Stop worrying about the
"complexity" and play around with simple machines
we can fully understand.

Game programmers like to experiment with decentralized
group behavior programming based on a few simple rules
with an eye toward implementing enemy AI into their games.
It can be fun to watch some of the more complex versions
involving predator/prey behaviors where they are just
moving dots on the screen of different colors, they look
very life like as they reproduce, die, and change numbers.

A thoroughly behaviourist one, please note.

In what way? Do they show this "conditioning" you
talk about? Programs put together determinate machines.
Their states change in time. Everything "behaves" and
like the word "conditioned" the word "behavior" is a
very overworked IMHO.


--
JC

.