[Hans Blom, 941205]
re: learning and performance
(Bill Powers (941201.0950 MST))
... anything for which there is a
built-in intrinsic reference level significantly greater than zero is
the definition of "good" for that organism; anything with a zero
reference level is "bad."
I have some problems here. What makes a reference level of zero so much
different from a non-zero level? No control system schematic that I know
of shows a difference; a control loop will operate regardless of the value
of its setpoint (within limits, of course). Moreover, in organisms zero
reference levels frequently have a corresponding non-zero antagonist
reference level. I would propose the following:
1. for an elementary control system "good" is a perception that matches
its reference level; "bad" is a perception not matching a reference level.
An elementary control system will attempt to control its perception in
such a way that it will closely match its reference level, i.e. it will
attempt to transform "bad" into "good".
Note a problem in terminology. The term "reference level" is used in
two different meanings. The first one refers to a specific setpoint,
or one to be specified, and is used in answers to questions like
"WHICH reference level ...". The second one concerns the level
(value) of a setpoint, and is used in questions with numerical
answers like "AT WHICH LEVEL ...". This double meaning frequently
confuses me.
2. for an elementary control system a perception is "indifferent" if there
is no reference level (in the first meaning of the term) in the system for
that perception.
Moreover, what is "good" for one elementary control system may not be good
for the organism as a whole. I recently gave an example of this: ad libi-
tum feeding, which supposedly reduces the error of one elementary control
system to zero, leads to premature death and inferior health in all
animals investigated so far, compared to animals whose body weight is kept
at 80% of ad libitum weight.
It is, in my opinion, important to realize that at any moment in time we
have a great many different goals that we want satisfied. Being limited in
our actions, we may not be able to satisfy all at the same time. We there-
fore may need to focus the "attention" of the system as a whole on those
goals that are most important and that also can be fulfilled given the
perceived environment. This may require us to (temporarily) abandon some
of our (minor) goals. Thus it appears that wild rats need to tolerate
error in their hunger-satisfying subsystem because other goals such as
perceived safety (bodily integrity) are more important, although this goal
cannot be fully satisfied either. PCT would call this an inescapable "con-
flict". I would rather call it a (in wild animals probably nearly optimal)
compromise between sub-goals.
The answer to "why" at one level is "how" at a higher level.
Very insightful. The "why" and "how" explanation subsystems that are
implemented in many expert systems show this as well. In an expert system
consultation session the "why (are you asking this question)" is answered
by indicating the higher level goal that is being evaluated: "this
question is being asked because I am trying to establish conclusion X";
and the "how (do you know this conclusion)" is answered by referring to
lower level conclusions that have already been evaluated, e.g.: "I know X
because X = Y AND Z, and both Y and Z have been evaluated as being true".
I am frequently struck between the close parallel between the "semantic
networks" (or whatever they are called) of AI and PCT's hierarchical con-
trol systems, even though one works with logical levels and the other with
numerical values. The parallel is even closer when the logic employed is
fuzzy, e.g. admits "probability" or "possibility" values anywhere between
e.g. -1 (false) and 1 (true).
Simulation is especially important where the
system is so complex or nonlinear that there are no analytical
techniques for solving the system equations.
I fully agree. Through simulations we may come to (partly) "understand" a
non-understandable system, i.e. a (partial) answer to what are the conse-
quences of our hypotheses.
I think your view of simulations is too limited. I have discovered many
things from simulations that I didn't know already. There is more to a
simulation than simply reproducing a behavior.
Yes. This is related to a deep philosophical question. Some philosophers
maintain than the whole of Euclidean geometry contains no more knowledge
than the few basic axioms on which it rests. Every theorem can, indeed, be
reduced to those axioms. And those axioms are assumptions anyway. Others
maintain that combining chunks of old knowledge generates new knowledge.
Who is right? When we do simulations, we adhere to the second interpreta-
tion.
Of course the contingency acquires no new properties. I thought I men-
tioned that. Of course it was the _cat_ that changed. We do not
disagree at all, although you seem to think so. But you do not go far
enough. There are no "causal relationships" in the physical world; they
are internal in us. Why is _this_ so hard to see?
Come on, now you're just looking for revenge. In the context of
practical epistemology, it's just a mistake to say that a consequence
can select a behavior. Of course you can always flee to a higher level
of abstraction and reply that it's a mistake to say anything, but my
point was much too simple to warrant such a strategy.
I was not looking for revenge but trying to see things through the cat's
eyes. Cats presumably do not know about causes and effects. But what a cat
CAN discover is that, if it wants something, some random action that it
executes leads to fulfilment of the desire. Somehow, that action will
start to be more and more correlated with the perception of the situation
the cat finds itself in. This type of learning is, in my opinion, building
new, useful correlations between perception and action. In SR-psychology
this could be called building a new stimulus-response path, in engineering
terminology a transfer function; in PCT-terms it would be building a new
control system.
By the way, I have never yet come across an epistomology that is practical
(useful) in all circumstances. I frequently find that each new application
requires its own, new epistomology. This is related to my observation
that there is no such thing as logic; there are (a great many different)
logicS.
What about reference signals? AAARRRRGH!
How about the above example? Did I have a reference signal for
following a shorter route to the university?
You did after you adopted it.
The crucial word is "after": a perceived consequence of an action led me
to install a new goal (or to change an existing one, however you look at
it). What I wanted to stress in this example is that this was PASSIVE dis-
covery: there was no control system or reorganizational system at work
that wanted to try to minimize some error.
In control engineering terminology, one might say, perhaps, that this can
be modelled with a dead-band controller that tolerates -- and does not
attempt to regulate away -- errors below a certain magnitude. But when it
discovers that the tolerance band can be set tighter, it will do so,
ratchet-like. On second thought I find this analogy less than convincing;
I wouldn't know how to model this type of learning.
Some learning seems not to depend on reference levels that are a priori
given, but leads to the installation of a new reference level (or a new
control system?) _after_ a discovery.
All this will be clearer if you think in terms of levels of control. The
reference signal for which route you take has to be changed _before_ you
change routes -- otherwise you would go by the old route.
I don't understand you here. Going the other route was motivated by a
_different_ goal: to accompany a colleague. Or rather, by a combination of
goals: going to the university (along any not too long route) AND accompa-
nying a colleague. So it seems that, by persuing one goal, one may disco-
ver, quite accidentally, a better solution for a different goal. The
reference signal for which route I took definitely did not change _before_
I went the new route for the first time; I even did not know which route
we would follow; I merely followed someone else's route.
This example may illuminate your concept of "discovery" learning. In
choosing a new route, you are not learning any skill you didn't have
before, nor are you fine-tuning a skill. Instead, you're selecting among
different ways of using existing skills, at a level higher than the
skills. We could think of a collection of routes about which you know,
only one of which, of course, can actually be employed on any one trip.
So the higher system, concerned with controlling things that are
affected by which route you take, has to pick one reference-route for
the lower systems to execute.
This is definitely not how I experienced things. What I experienced is
that one goal (following my old route) was at that time less important
than the other goal (accompanying my colleague). So it seems that fol-
lowing my old route was not really a goal -- it was easily disturbed.
Following my old route was merely a sub-goal of a goal that I did NOT
relinquish -- going to the university in a reasonable time. So it seems
that (some of) these sub-goals are easily replaced when a better alter-
native to fulfilment of a higher level goal somehow offers itself. But the
point I wanted to stress was that in my case this new sub-goal was ACCI-
DENTALLY discovered, NOT through an active process of error-minimization.
That is, I do not see an (active) CONTROL process at work.
Greetings,
Hans