Learning and evolution

[From Bruce Abbott (951104.0945 EST)]

Shannon Willyams (951103) --

Bruce Abbott (951102.1115 EST)

If the cat lacks the capacity to reason it out, to develop and test
hypotheses, what other mechanism might lead to a workable solution?

My guess would be some kind of "coincidence detector."

Maybe. But (I think) learning is only evolution. The mechanism for
learning needs to follow the same general principles as other evolution
mechanisms.

Yes, variation and selection. The variation part isn't too difficult to
visualize, but what does the selecting? Some mechanism must exist to
identify which action is the effective one, and that mechanism must be
active only when there is a need to identify an effective action, i.e.,
there must be some system "attempting" to being some perception under control.

In any case there is no way to test the "coincidence detector" concept yet.
You need to at least determine a method by which a coincidence is
identified, and then determine if this method is used.

Yes, exactly right. But before one can think up a mechanism to test, one
must first identify how that mechanism is supposed to function: what inputs
does it require in order to carry out its function, and what must it do with
them? I'm working on that first step in the process.

I cannot yet visualize the mechanism by which control loops develop, but
call it mechanism X.

When the cat is in the box, he already has a working set of control loops
at his disposal. All of these loops have their own assessment of the
situation. Each loop has an (some) associated output. One by one the
cat tests these outputs, and then retests them... When the door opens,
mechanism X is invoked which begins the formation of another control loop.
Each time that the door opens, mechanism X is invoked which optimizes the
new control loop. Eventually, as soon as the state of being caged is
recognized, the cat will have developed a control loop for handling the
situation.

Reasonable but fuzzy--as you said, you cannot yet visualize the mechanism.
Note, however, that your "mechanism X" is hypothesized to be activated when
the door opens. Why? What's significant about the door opening? And how
does the cat identify which of its actions is the effective one?

Depending upon what inputs that the cat uses to recognize his situation,
the cat now applies his new control loop to other situations. For
example, if the defining input for his new loop is "I do not want to be
enclosed", then anytime he is enclosed, he will activate this control loop.
If the defining input is "I need to move that obstacle between me and the
outside", then anytime something is between him and his goal, he will
activate this control loop.

Ah, generalization. Very nice.

The testable part of this theory is the conglomerate control loops. We
should be able to determine if these exist.

That should present no real difficulties. Let's say we've done that. What
next?

Regards,

Bruce

[From Shannon Williams (951105)]

Bruce Abbott (951104.0945 EST)--

The mechanism for learning needs to follow the same general principles
as other evolution mechanisms.

The variation part isn't too difficult to visualize...

I would say that the outputs of the different active control loops
constitute the different variations. And somehow, the new control loop
that is being formed is going to 'remember' which active control loops lead
to a solution.

but what does the selecting?

Some mechanism must exist to identify which action is the effective one,
and that mechanism must be active only when there is a need to identify
an effective action, i.e., there must be some system "attempting" to
bring some perception under control.

When the finches' beaks are selected to be longer, what is the 'coincidence
detector'? Is this 'coincidence detector' only active for a short time?

I would say that when a control loop exists, it always exists. But it is
only emitting output when it is activated. When its reference level is
restored, it stops emitting output. There does not need to be a
'coincidence detector'.

But before one can think up a mechanism to test, one must first identify
how that mechanism is supposed to function: what inputs does it require
in order to carry out its function, and what must it do with them? I'm
working on that first step in the process.

Ok.

Each time that the door opens, mechanism X is invoked which optimizes the
new control loop. Eventually, as soon as the state of being caged is
recognized, the cat will have developed a control loop for handling the
situation.

your "mechanism X" is hypothesized to be activated when the door opens.
Why? What's significant about the door opening?

When the door opens is when the new control loop is restored to its
reference level.

And how does the cat identify which of its actions is the effective one?

That is mechanism X. Somehow the control loop must *remember* what outputs
worked. It must select for these and inhibit others.

The testable part of this theory is the conglomerate control loops. We
should be able to determine if these exist.

That should present no real difficulties. Let's say we've done that.
What next?

(Are you always in such a hurry?) Once we start testing loops, then we
will develop a picture of them. We will start describing them. We will be
in 'functional specification' phase again.

-Shannon

[From Bruce Abbott (951106.1230)]

Shannon Williams (951105) --

Bruce Abbott (951104.0945 EST)

but what does the selecting?

Some mechanism must exist to identify which action is the effective one,
and that mechanism must be active only when there is a need to identify
an effective action, i.e., there must be some system "attempting" to
bring some perception under control.

When the finches' beaks are selected to be longer, what is the 'coincidence
detector'? Is this 'coincidence detector' only active for a short time?

Do control systems reproduce? Does "like beget like"? What does it mean to
be "successful"? How does it happen that successful variants are "preserved"?