[Hans Blom, 941206]
(Tom Bourbon [941202.1110])
Assuming adaptation occurs through a random Ecoli reorganization,
different systems will "set themselves up" in different ways, but all
will achieve the same end. If the control systems were human, an
implication of such a state of affairs might be that they disagree
about how the world works, or about the best way to produce a
particular result in that world.
Considering this, it is a miracle that we sometimes understand each other!
An observation: I notice a practical difference between people's opinion
"about how the world works" on the one hand, and "about the best way to
produce a particular result in that world" on the other hand. Let me give
you an example. When I build an expert system, it is my goal to translate
the knowledge/expertise of a human expert (usually someone from the softer
sciences; I mostly work with MDs) into formal rules and procedures that a
computer program can use, partially mimicking the human. Acquiring the
human's knowledge, called "knowledge engineering", is a difficult process,
most of all because experts cannot very well express what they know. There
is a saying in these circles that goes "an expert does not think, he
acts", meaning that an expert does not do on-the-spot learning but is a
fully converged state-of-the-art "pattern matcher" or "control system",
however you might wish to express it. After having done a lot of "know-
ledge engineering", I discovered that experts disagree so much about how
the world works that consulting more than one expert is usually a waste of
time. On the other hand, experts are usually quite unanimous about the
best way to produce a particular result. That is, expert ACTIONS are a lot
more in agreement than the REASONS experts give for their actions. This
leads me to the conclusion that in order to be able to control well, many
different "world models" do the job about equally well. Or: a coarse model
is usually sufficient.
For me, this has deep implications for everyday life. One is: if you want
to understand people, look at what they do, not at what they say. Another
one: it is not the theories that are important, but the results of those
theories.
An extension of those thoughts would be that control systems with
similar intrinsic reference signals, but with somewhat different
environments would almost certanly end up with differences in their
parameter settings, and probably in their ideas about what the world is
and how it works. Hence, some of the differences between control
theorists in engineering and those who study living systems?
You can be even more general: since everybody lives in -- and has learned
from -- a different environment, differences in opinion will be ubiqui-
tous. My opinions on certain themes are different from those of other
control theorists, if only for the sole reason that I followed a different
curriculum in school and a different research path after that.
By the way, even engineering control theorists study living systems,
though from a different perspective. We design, after all, systems that
one can consider to be TOOLS to be used by living systems. And even if
those tools are smart and have their own built-in "machine intelligence",
they are still extensions of the living system's control hierarchy; they
add, one could say, another layer to the hierarchy. Note that in the PCT
hierarchy a lower level goal is a TOOL to accomplish a higher level goal;
it is a MEANS, a "how to", not a goal in itself.
If I want to design a successful (control) system, I need to be acutely
aware of what goes on in the living system whose tool it is going to be.
I'm not certain
what you mean when you say a system, reaches ". . . convergence in its
estimation of the correlation between e.g. an action and its effect."
Could you elaborate a little on that idea? For example, where, or how,
would the system calculate such a correlation?
I answered this in my contribution dated 941124.
Would that calculation require a
"world model" of the iconic kind, rather than a set of parameter
settings in the system?
I have no idea what you mean by a "world model of the iconic kind". The
kind of world model that I proposed (and use in some of my designs) con-
sists of formulae with free variables, the values of which are adjusted
through a process of correlating perceptions. The engineering science
behind this approach is called "systems identification".
Or are you talking about something more like the
procedure we have tried in a few adaptive PCT models: an adaptive loop
in the model makes small random adjuetments in some parameter (e.g., a
gain factor) until the magnitude of error the adaptive loop senses in the
main control loop falls below a criterion level?
Similar, but not the same. There is no "until": the model is adjusted
continually. The model is also not adjusted based on the magnitude of some
control error, but on something like a "best line fit" criterium. An
accurate world model "explains" (describes) regularities between the
perceptions of our sensors. It does not imply that we can control those
regularities. I can see the sunrise, not control it.
I'm just not sure what you mean here when you speak of a "correlation."
Refer back to my contribution dated 941124. An example: a computed corre-
lation between the forces developed in my leg muscles and my perception of
the speed of a football after the kick can provide a "model" of the mass
of the ball. After some trials, I can use that estimate and accurately
kick that ball. But substitute the ball with another one with a different
mass and control is lost (but may be relearned, of course).
To me, the poker example looks like the result of a process in which the
system adjusts itself until sensed error is at or less than the specified
level. That would make it a "satisficing" system, rather than an
"optimizing" one. Is that at all close to what you had in mind?
No. If you want to be an expert poker player, a "satisficing" system is
not good enough, because you would only adjust your playing level such
that you win from the opponent(s) that you are playing with in the current
game. That is not good enough to maximize your profits, which I assume a
professional player would want to do. PCT's reorganization is satisficing;
if there is an error, reorganize it away. This is, as I understand it, a
one-shot affair without memory. The correlational type of learning accumu-
lates perceptions, the more the better, just like a best line fit becomes
more accurate when the number of points the line is fitted through increa-
ses. This is an "optimizing" system in the sense that all relevant infor-
mation is remembered and can be used.
I have a few questions and comments about your reply (to Bill P. [Hans
Blom, 941130a]), but they must wait until after I run some experimental
subjects.
Fine. Are we starting to understand each other?
Greetings,
Hans