Selecting, Learning

[From Rick Marken (960301.1400)]

Stefan Balke (960301.17.00 MEZ) to me:

you say: You consciously try different things... you select references...

My question is: How do I select them? Is there a model that describes the
process of selection?

The current model of this selection process is random reorganization; it is
assumed that we randomly select references and maintain a selected reference
as long as the variable we are trying to control continues to move towards
the reference state; we randomly select a new reference (behavior) when
the variable moves away from the reference state.

I agree, however, that subjectively it seems like I select references non-
randomly when I am reorganizing; when I was dating (and thus, constantly
reorganizing), for example, I tended to take my date to a nice retaurant, a
play or a concert rather than a porno movie (as DeNiro did in "Taxi Driver")
in order to achieve my not altogether wholesome goals. I don't know whether
any of these behaviors actually helped me control any better, but I did have
a bias away from the "Taxi Driver" option.

I don't know much about the E. coli approach, except the demo, but there
is a german children game which came up to my mind while watching the demo.

That's it! It's called the "hot and cold" game here in the US.

Shannon Williams (960301.13:30) --

I am saying that learning has the following two aspects:

1) Learning involves associating perception A and perception B...

2) Learning involves associating your output with changes in your
perceptions.

Learning #1 allows a creature to reflexively influence situations in the
future.

Learning #2 allows a creature to intelligently influence situations.

Do you agree with (any of) this so far?

It sounds like learning # 2 is what I would call "learning". I don't know
what learning # 1 is about; I don't agree that creatures ever "reflexively"
influence situations. I think what you might be describing as "learning # 1"
is just what we would call "memorizing"; when someone asks us, in the future,
whether we expect to see cat food when we open the refridgerator, we can say
"yes" if we have stored the association between the perception of cat food
and of refridgerator.

Actually, memory may be useful as a means of biasing the reorganization
(learning) process. Is I were stuck in my house alone for days and I was
starving Imight happen to play back a memory that reveals a behavior that
_might_ reduce this intrinsic error: "opening the refridge"; of course,
this behavior will only reduce error if I had actually stocked up on cat
food before I got stuck.

As I keep saying (with no apparent impact) what would help this discussion of
learning enormously is some DATA. It's hard to know how HPCT fails as an
explanation of a phenomenon (learning) without knowing what phenomenon is to
be explained. It would be best if we could see some quantitative "learning"
data and some examples of the non-HPCT models that account for this data.

Best

Rick

[From Shannon Williams (960301.1815)]

Rick Marken (960301.1400)--

I am saying that learning has the following two aspects:

1) Learning involves associating perception A and perception B...

2) Learning involves associating your output with changes in your
perceptions.

I think what you might be describing as "learning # 1"
is just what we would call "memorizing";

I agree.

As I keep saying (with no apparent impact) what would help this
discussion of learning enormously is some DATA. It's hard to know how
HPCT fails as an explanation of a phenomenon (learning) without
knowing what phenomenon is to be explained.

HPCT does not model learning. Consider the following scenario:

Cat has a reference.
Cat's perception does not match reference.

Does cat:
1) Output behavior based on its current perception/reference mismatch?
2) Learn a new reference?
3) Revert to an old reference?

Now draw an HPCT diagram. On your diagram point to: Cat's reference,
Cat's perception, the mechanism by which Cat determines what it will do.

It would be best if we could see some quantitative "learning"
data and some examples of the non-HPCT models that account for this data.

OK. The diagram that I sent earlier is just a PCT diagram with some
neural nets thrown in. Does that constitute a model? If not, then tell me
what I need to do to make that a model. I think that it can explain
anything that HPCT can explain, and it can also explain learning/planning.

Tell me what data has been explained by HPCT, and I will tell you how it
is explained by a PCT diagram with neural nets.

-Shannon