Prose samples by Robert Elton Maas <rem@BTR.Com>
General essays and proposals


%% 79807A.ISR (4k, 1979) Considerations needed when designing a
world-wide information system for the general public

WORLD-ISR: Spanning range of needs.

(1) Geographical distance.  Some needs are tightly geographical, for example
a person needing a carpool from Sunnyvale to Palo Alto doesn't have any use
for carpool info for greater Los Angeles or even for the San Jose to Santa
Cruz route;  a lonely girl in Houston probably doesn't want to date a guy
in Menlo Park (CA or NJ).  Other needs are more universal, for example if
a person in Burlingame wants a PET-BASIC program for playing Blackjack (21)
and the nearest such program is in Atlanta, a cassette can be mailed easily;
a person in Melbourne could order a book from New York if he could determine
ahead of time that it would be what he wanted.  A general information system
should classify information according to the geographical range where it
applies.

(2) Time.  Some needs are urgent and become obsolete quickly, for example
a heart-attack must be treated quickly, and a carpool to a concert must
be timely.  Other needs remain valid for a long time, for example a guy
looking for a wife might find her today or next month and the specification
of her qualties desired wouldn't change over that interval, or a recipe
for homemade eggrolls would remain valid for perhaps hundreds of years
except for changing availability of ingredients.  Each item of information
must have attached to it an expiration date, and the system should allow
changing the date if the original data turns out to be wrong.

(3) Expertise.  Some needs are highly technical or professional, for example
diagnosis referral and operation on a punctured lung must be done by trusted
doctors or other trained medical personnel.  Other needs are highly informal,
for example a guy wanting to find out "what's happening" probably doesn't
know exactly what he wants to know but certainly doesn't want his question
treated as a technical query, and he will accept almost any information in
a wide class of "interesting things to do and places to be".  Some users
of the system will have their own terminals and be fluent in talking to
computers directly, while others will be outright paranoid that a computer
has anything to do with handling their question after they submit it to
the friendly librarian over the telephone or in person.  Both the user
interface and the method of parsing incoming questions must be aware of
the varying degree of preciseness and expertise of the users of the system.


%% 79C26B.G14 (whole file here, mid 1970's) Semi-political ad, non-technical

	A CALL FOR PEOPLE TO GET TOGETHER AND WORK TOWARDS A BETTER WORLD

Is there something you want changed about the world?  Is it something you
can't do by yourself?  Perhaps you need to locate several other people who
want the same change, so that you can work together, either to make it happen
yourself or to lobby in the state legislature or the US congress.  Are you
having trouble locating other people to join with, or have a group that is
too small to accomplish its objective, so you need more people?

This may or may not work, but let's try it and see whether it does or not.
Send your name, address and phone number, and a brief description (about
25 words) of what kind of change you want, to the address below.
Your information will be assigned keywords that indicate subject
areas, and will be matched with other people's information to locate people
who want something like what you want.  When a match is found, you and the
other person will be called on the telephone and given each other's names,
addresses and phone numbers.  (Please restrict telephone numbers to those
which are within local dialing from Palo Alto.  This includes from Mountain
View and Los Altos on the south to San Carlos and Belmont on the north.)

It doesn't matter what your dream is, whether it be promoting a new source
of energy, trying to elect a certain person to President, having a child-care
center in your neighborhood, collecting consumer information and publishing
an index of good products to buy and bad products to avoid, or just finding
a friend or lover to share some pasttime you enjoy.  We do, however, request
that illegal activities (addictive drugs, prostitution, gambling, weapons
etc.) be avoided; we refuse to become dealers in such areas, and will reject
any topic which is explicitly illegal or which seems to be written in "street
jargon" which we don't understand but which we can only assume is an attempt
to hide something illegal.

Alone, you can't get much really accomplished.  Together, we can move
the world.  Let's get linked up now!  It only costs you a ten cent postcard.

		REM Information Service
		PO Box 6641
		Stanford, CA 94305


%% ANLGEO.ESS (2k, 1980) Insightful essay

	Analytic geometry and its cousins in space time and electronics
	-C- 1980 by Robert Elton Maas, all rights reserved

Present-day physics and electronics is analogous to analytic geometry
in that it assumes a fixed reference coordinate system.  Let me give
this analogy in more detail.

		Classical method		Analytic method

Geometry	Objects (points, lines)		Coordinates are fundamental,
		and measurements (lengths,	objects are expressed in
		angles) are fundamental.	coordinates, measurements
						are computed from them.

Electronics	Voltage differentials are	Voltages with respct to an
		fundamental.  Kirchhoff's law	absolute "ground" are
		for voltage requires tracing	assumed.  Kirchhoff's law
		around closed circuits.		for voltage is a tautology
						of arithmetic.

In each of these fields we invent a reference frame, and measure
everything else in comparison to this reference frame.  We know, at least
initially, that this frame is a figment of our imagination, put there
merely to make analytic methods (which are so much easier than classical
methods) possible.  Any of several (an infinite number of) frames could
be chosen, but we generally pick one which is fixed with respect to the
largest number of objects or the largest object or the average of all
objects in the system.


%% DOC.LEA (9k, 1981) Discussion of meaning of "learning"

	LEARNING ASPECTS OF ROBOT'S BEHAVIOR
	by Robert Elton Maas

One of the major goals of Artificial Intellgence research is a program
that will learn things. But when we obtain a program that seems to
learn, if we look at its internal workings to see how it learns we often
see what's doing isn't really learning, or at least nothing that we want
to call learning. Usually "learning" is really something simple that we
call by another name if we understand how it works.

One might claim that a thermostat learns how to keep a room the right
temperature, but all it really does is turn the heater on and off depending
on whether the temperature as measured is lower or higher than a threshold
we've set. One can "teach" it the right temperature by adjusting the
thermostat upward whenever it feels cold but the heater is off, and by
adjusting it downward whenever it feels warm but the heater is on.
There are thus two stages of "learning", first when the air in the room
isn't yet warmed at all (when you first turn on the system after returning
from vacation) and it needs to bring it all up to the established "correct"
temperature, and second when the room temperature is stabilized at the
"correct" temperature but the meaning of "correct" needs adjusting according
to human desires.

Hardly anyone would call either type of adjustment by the word "learning",
but what if a program that played some game had some parameter that was
adjusted by the program itself.  What if initially it picked random values
for that parameter, but it got feedback about whether it was winning or
losing and kept a record of [value, win/lose] pairs, and did a simple
regression on those pairs to try to find the optimum value of the parameter,
and then picked random values in a normal distribution (bell-shaped curve)
centered on the computed optimum value with standard deviation equal to
the expected error in the optimum given by the regression calculation?
What if by that relatively simple means it automatically adjusted this
one parameter until it remained close to the truely optimum value for
winning the game more often than not.  Would we call that "learning"?

What if the program had lots of parameters, and it did a multiple
regression to attempt to guess which parameters affected the outcome
most and what their optimum values seemed to be?  What if it adjusted
them all to be near-optimum, i.e. given by the normal distribution
around optimality that is given by the multiple regression calculation?
What if it simultaneously adjusted them all? What if it even discarded
those which seemed not to have hardly any effect on the outcome of
the game, or fixed those whose optimal value had been computed very
accurately (narrow normal distribution, small expected error), and
replaced them by new trial parameters from a pool provided by the
progammer? Arthur Samuel's program to play Checkers did something like
this.  Is this really "learning"?  Like children learn in school or
during play?  Or is it just a complicated bunch of optimization of
parameters sort of like a thermostat in each room of a building?

What about rote learning?  In the old days teachers would give the
answers to the student and insist the student repeat them back.
	"four times seven gives twenty-eight"
If the student got it wrong, a beating with a stick would get the
poor student's attention and then that particular fact would be
repeated until the student was repeating it back correctly.  With
computers it's easier.  Computers have a wonderful memory, and we
have tools for directly modifying their memory, sort of like doing
surgery on the student's mind.  To get a computer to learn a fact
or an algorithm, all we have to do is start up a text-editor on the
relevant data or program file, invoke a keyword search command to
find the correct spot in the file, issue a few edit commands, issue
a save-the-file command, and presto the computer has learned something
it didn't know before, and it never forgets it unless we go in there
and delete the item (or the whole file), except when it "drops a bit"
and mungs our item or loses our entire file.  We can thus teach it
facts.  Initially when we do a keyword search for some fact, it doesn't
turn up any matches.  After we've entered the fact, a subsequent keyword
search turns up that fact.  Has it learned that fact any less so than
the student in school has?  Initially the computer doesn't know how
to do some task we'd like it able to do, ...

When we say some person "knows how to bake a cake", do we simply mean
that person has stored somewhere in that person's mind an algorithm for
baking a cake, retrievable by the key phrase "bake a cake" and any other
phrases that are considered synonyms to that phrase?

Our current project, to build a "robot" program that somewhat emulates
a child learning arithmetic algorithms and other grade-school tasks,
will use simple "learning" methods such as those above (parameter
optimization and rote memorizing) and others described below.  Perhaps
when the robot is finished, somebody looking at its external behavior
will claim it's "intelligent", that it "really learns", while anybody
knowing its internal mechanisms will say "but all it does is a few
simple memorizing and optimization things, none of which is really
learning".  Perhaps what we call "learning" is nothing more than
a very large set of obvious trivial things that individually would
never be called "intelligent" but when considered en masse are too
complicated for us to understand and thus appear to be something different.
Perhaps that's all intelligence is anyway, a very large complicated
biological computer doing massive amounts of simple things that affect
one another so much it's hard to separate out their individual mechanisms.


%% SUBPRO.PRO (6.8k, late 1970's) Not too technical, essay about computing

There are several useful computation tasks that are suitable for and
cost-effective on microprocessors.  The latter is especially so because
most microprocessors are left idle most of the time, making the market
value of their CPU time essentially zero.  Although a single micro-
processor can't solve a giant problem in a reasonable time, the
thousands of microprocessors sitting idle in homes around the country
could be a formidable computing machine if put to bear on small factors
of a large factorable compute-task.

To be suitable at the current time (semi-manual downloading of software
and task-specific paramaters), a problem must satisfy these criteria:
It must be factorable into arbitrarily-small (relative to the original
problem) subproblems.  Each subproblem must have a tiny natural
representation that can be transmitted from the master task dispatcher
to the particular processor that will run it.  Each subproblem must run
in small memory.  It must be trivial to combine the results of the
subproblems back into a solution for the overall problem.


%% End of samples of general essays