The Tinkertoy Computer and Other Machinations

Puzzles and Games/Computers $22.
95
the Tinkertoy Computer

and other machinations
A. K. DEWDNEY
Matter Computes. Matter Misbehaves.

Mathematics Matters. Computers
Create. These four basic themes of the
electronic age provide the underlying
framework for A. K. Dewdney's latest
book.
Drawn from the pages of Scientific

American and Algorithm, these
columns offer an often whimsical but
always provocative look at some of the
most fascinating topics in computer sci-
ence: chaos and fractals, artificial intel-

ligence, robotics, neural networks, com-
puter-generated art, and more. You
will find descriptions of computer pro-
grams that converse, compose music,
simulate thunderclouds, ace IQ tests,
and produce playable miniature golf
courses on the screen. You will
encounter what was possibly the first
digital computer (a rope-and-pulley cal-
culator dated circa 850 A.D.), as well as

metallic insects with brains of silicon, a
computer created out of an extremely
odd billiard table, and the remarkable
title contraption —a construction of
Tinkertoys that offers formidable com-
petition as a tic-tac-toe player.
But you won't just have to read about

these wonders. THE TINKERTOY
COMPUTER is also a rich recreation-
al, and yes, even educational, resource.
Professor Dewdney has conspired to
smuggle more than a modicum of hard-
core computer knowledge into his dis-
arming essays. Most chapters contain
recipes and algorithms from which
working programs can be constructed.
Digitized by the Internet Archive
in 2014
https://archive.org/details/tinkertoycomputerOOdewd
Tie taw Computer
AND OTHER MACHINATIONS
L
rrm
1
\ I
11 NKERTDY 1 L I
A. K. DEWDNEY
IS
w. h. freeman and company
New York
Library of Congress Cataloging-in-Publication Data
Dewdney, A. K.
The Tinkertoy computer and other machinations.
[A.K. Dewdney].
p. cm.
Includes index.
ISBN 0-7167-2489-8. — ISBN 0-7167-2491-X (pbk.)
1. Computers. I. Scientific American. II. Algorithm.
III. Title.
QA76.5.D458 1993
004— dc20 93-10478
CIP
Copyright © 1993 by W. H. Freeman and Company
No part of this book may be reproduced by any mechanical,

photographic, or electronic process, or in the form of a
phonographic recording, nor may it be stored in a retrieval
system, transmitted, or otherwise copied for public or private
use, without written permission from the publisher.
Printed in the United States of America
1234567890 VB 99876543
Prologue: The Hidden Agenda 1
Theme One: Matter Computes 5

1 The Tinkertoy Computer 7
2 The Rope-and-Pulley Wonder 16
3 The Infinite Brain 26
4 Invasion of the Insectoids 37
5 Building a Brain 46
6 Dance of the Tur-mites 57
Theme Two: Matter Misbehaves 69

7 Microminiature Golf 71
8 Star Trek Dynamics 81
9 Weather in a Jar 93
10 A Portrait of Chaos 100
1 1 Designer Fractals 109
12 The Fractal Workshop 1 19
Theme Three: Mathematics Matters 129

13 Mathematical Morsels 130
14 Golygon City 139

vi Contents
15 Scanning the Cat 148
16 Rigid Thinking 158
17 Automated Math 165
Theme Four: Computers Create 177
1 8 The Computer Composer 1 79
19 Chaos in A Major 191
20 Mark V. Shaney 199
21 Face Space 205
22 Voltage Sculptures 216
23 Latticeworks by Hand 222
Illustration Credits 231
Index 233
I
!i
J TER
PROLOGUE
The term "machinations" in this book's title
refers first to the marvellous machines called

computers — in all their manifestations. The
term refers secondly to my own machinations.
Until this third book of computer recreations,
I have kept my agenda hidden.
During my career as a computer scientist, I

have tried, as many academics do, to express
difficult concepts in simple and engaging
examples with a recreational flavor. Balanced
with more (rather than less) pedestrian mate-
rial, this approach has, I hope, helped to make
computer science a bit more fun and interest-

ing for the few thousand students I have
taught over the years.
In 1984, when I was invited to write the
Computer Recreations column for Scientific
American magazine, the recreational approach
seemed perfect — it already matched that side
of my lecturing style. Moreover, I saw it as a

golden opportunity to bring computer science
to a new set of "students," a worldwide read-
ership that numbered in the millions. But here
the emphasis was reversed. It was recreation
first and education second — and a very distant
second if the columns threatened to become
miniature lectures. I was forced to go under-
ground. Every recreation, from Tinkertoy
computers to programmed cartoons, became a
secret vehicle for teaching anything from
computer logic to graphic programming
methods. That was my hidden agenda.
Later, when Scientific American changed
the name of the column to Mathematical
Recreations, I followed the same secret course,
hiding proof techniques in the alleys of Goly-
gon City and concealing research methods in
the innards of two-dimensional cats.
In both columns the machinations paid off.
Readers seemed to enjoy themselves, even if
they didn't always realize the full importance
of what they had just learned. The trouble lay
in the word put readers at a
"recreations. " It
serious contextual disadvantage. If my regular

students knew that, despite my recreational
ways, lectures in computer science were
serious stuff, too many of my new students
may have been overly influenced by that
word, especially when it came to the heavier,
philosophical ideas. How seriously would you
treat anything Descartes said at a party?
"Here's a good one: I think therefore I am.
Ha ha."
I have always been attracted by certain
philosophical themes that arise from computing

and mathematics. In this latest collection of
columns, I have decided to come clean about
four of the themes, to stop treating them as
schemes. Here they are, boldly stated and
briefly described.
MATTER COMPUTES: Almost any form of matter

may beorganized into a computer. You can use
tinkertoys, ropes and pulleys, even individual
molecules. Living things organize their own
matter, not only into muscle fibers and bone
marrow, but into brains. In the most general
sense, at least, brains compute. From a purely
computational point of view, evolution may turn
out to be nothing less than the emergence of
mind from matter. With a surprisingly small
investment in computation, matter may be made
to behave with something like intelligence.
MATTER MISBEHAVES: We can simulate a great

many physical and dynamical systems by
computer, even a game of golf or a starship
combat. Viewed as predictive models, however,
computer programs have one major limitation.
We have only recently discovered that some
dynamical systems are capable of chaos. This
form of dynamical misbehavior amounts to a veil
that hides a system's future from our view
forever. Even if we cannot look behind it, we
can see the veil itself. It has fractal embroidery.
MATHEMATICS MATTERS: Explicitly or implicitly,

mathematics informs every worthwhile scientific
idea. Besides being an amazing subject in itself,
mathematics determines what is science and
what is not. Computer science, which some view
as a branch of mathematics, derives its major
conceptual machinery from mathematics.
Luckily, even adventures in pure thought may
be disguised as recreations, from the spheres of
Dandelin to the Hexagons of Connelly.
COMPUTERS CREATE: From a sociological point

of view, we are not surprised to find that some
people program computers to play discrete
music, paint pixellated pictures, or to converse
in patchwork speech sewn from human
fragments. Whether the music springs from
chaos, the faces from files or the speech from
random numbers, there is a dilemma: Are
computers invading the arts or are the arts
invading computers? You can have it both ways,
but be warned! The future holds sights, sounds,
and stories more bizarre than we can imagine.
And the creative act will not always be human.
Of the 23 chapters comprising the present
collection, 18 originally appeared as articles in
Scientific American and five appeared in
Algorithm, the recreational programming
magazine. The chapters have been organized
intothemes and each throws a light on its
theme, illuminating it from a special angle.
Most of the chapters contain recipes or
algorithms from which readers can construct
programs. One of the chapters (A Game of
Micro-Miniature Golf) contains a working
program so that beginners in the art of
programming may understand how to convert an
algorithm into a program in the language of their
choice.
Neither the author nor the publisher will
assume responsibility for career-change decisions
inspired by reading this book.
THEME ONE
To say that matter computes means, in part,

that we get it to compute by constructing the
bits and pieces of matter that make up a
computer. But it also means that nature itself
seems able to do the same thing, insofar as
brains amount to computers. If we ever suc-
ceed in constructing what critics would cheer-
fully admit to be a brain, we will have recapit-
ulated our own evolution. Mind will have
crept into matter by a new route.
The first two chapters hint at the astonish-
ing variety of means by which computers may
be constructed. We may build computers not
only from silicon circuits but from Tinkertoy
kits and even from systems of ropes and
pulleys. In a subject called computer logic we
learn that from a few simple logical gates such
as AND, OR, and NOT we may construct
combinational circuits to perform arithmetic
and memory functions. We may even build the
central processing unit whose chief task is to
run programs. There are basic combinational
circuits in rods and spools, also in ropes and
pulleys.
The next on minds, machines, and
chapter,
metaphysics, forms a bridge between the
animate and the inanimate. In it, readers dis-
cover another computer made out of an ex-
tremely odd billiard table. The billiard ball
computer bolsters the argument of a prominent
theoretical physicist that human brains cannot
be computers. The objections are met here
with a determined resistance from the artificial
intelligence community. In the process of
debate, readers learn valuable criteria for de-
ciding when we have met the goal.
Whatever the status of human brains, the
insectoids have already arrived. Metallic in-
sects with brains of silicon do not exactly
refute the critics of artificial intelligence, even
if they (the insectoids) act like insects. But
they show how behavior can be orchestrated
from the interaction of microchips, motors,
and metal legs. The latest progress in robotics
springs from two new evolutionary constraints
of our own: Keep it simple and keep it cheap!
By what machination might we turn a
computer into a brain? Should we be surprised
if neural nets turn out to be the scheme of
choice? The next chapter in this suite takes

the gloves off and shows how to construct a
network of some forty neurons that learns
how to convert between coordinate systems —
a humble but not unimpressive task. In spite
of certain drawbacks to the neural net para-
digm, some manufacturers have begun to con-
struct neural chips.
Because real environments impose such
tremendous computational demands on
intelligent machines, we are still a long way
from putting more advanced brains (neural net
or not) into insectoids. We turn, therefore, to
wholly abstract creatures that contend with a
much simpler environment, an infinite
checkerboard. Tur-mites embody the potential
for full computational power over this simple
terrain. Singly or in pairs, their built-in Turing
machines drive them to paint the screen red and
most intricate mounds imaginable.
to build the
Some chapters suggest extra reading for
those who would like to pursue the nuts and
bolts of specific methods or to probe the key
debate of the twenty-first century.
I firsthad that experience (universality of computation)
before went to school. There weren't any
I
(computers) yet, but we had toy construction sets.

One was called TinkerToy. What's strange is that
. . .
those spools and sticks are enough to make anything.
Marvin Minsky, in preface to LogoWorks
TJ
JL JL ow many of us remember Tinkertoys, those down-home kits of
colored wooden sticks and spools with holes in them? Amid our child-
hood constructions of towers or cranes, how many of us pondered the
outer limits of the Tinkertoy world? Did we conceive of contraptions
that reached the ceiling? Perhaps, butwe lacked the kits or the time to
make it happen. Such a Tinkertoy fantasy took place in 1980 when a stu-
dent group from the Massachusetts Institute of Technology constructed
a computer entirely (well, almost entirely) out of Tinkertoys!
From a distance the Tinkertoy computer resembles a childhood fan-
tasy gone wild or, as one of the groupmembers remarked, a spool-and-
stick version of the"space slab" from the movie 2001: A Space Odyssey.
Unlike the alien monolith, the computer plays a mean game of tic-tac-
toe. A Tinkertoy framework called the read head clicks and clacks its
way down the front of the monolith. At some point the clicking myster-
iously stops; a "core piece" within the framework spins and then with a
satisfying"kathunk" indirectly kicks an "output duck," a bird-shaped
construction. The output duck swings down from its perch so that its
beak points at a number —
which identifies the computer's next move in
a game of tic-tac-toe.
What precisely doesthe read head scan as it feels its way down the
monolith? Nothing less than 48 rows of Tinkertoy "memory spindles"
7
8 Theme One Matter Computes
encoding all the combinations of X's and O's that might arise dur-
critical
ing a game Each spindle is a sequence of smooth spools

(see Figure 1.1).
connected axially by sticks and arranged in nine groups of three each,
one group for each square of the tic-tac-toe board. The presence or ab-
sence of spools from a group indicates that a corresponding square of the
tic-tac-toe board is vacant or is occupied by an X or O.
The Tinkertoy computer is not fully automatic: a human operator
must crank the read head up and down and must manage its input. After
the computer's opponent makes a move, the operator walks to the front
of the machine to adjust the core piece inside the read head, registering
the contestant's move. The operator then pulls on a string to cock the
core piece for impending whirl of recognition. When it discovers a
its
memory that matches the current state of the game, the core piece spins,
and the computer indicates its move.
The best way to understand how the machine works in detail is to re-
count the story of its creation at the hands of the M.I.T. students: Erlyne
Gee, Edward Hardebeck, Daniel Hillis, Margaret Minsky, and brothers
Barry and Brian Silverman. Most of the group has long since graduated
and entered various computer professions. Perhaps the best-known
team member is Hillis. He was the moving force behind Thinking Ma-
chines, Inc., which produces the well-known parallel super-computer
called the Connection Machine. (Perhaps Tinkertoys have something to
teach us.)
In 1975, when and Brian Silverman were in their sophomore
Hillis
year, they participated in a class project to build something digital from
Tinkertoys. The students sat down to play. One made an inverter —
logic device that converts a binary 1 signal to a 0 signal and conversely.
Another made an OR gate; if either of the device's two input signals hap-
pened to be a 1, then its output would also be a 1. It quickly became clear
to the students that Tinkertoys were "computation universal," the theo-
retical term for a set of components from which a fully programmable
computer can be constructed. Theoretical possibility was one thing, the
practical demands of money and time another.
The demands were met in a rather roundabout manner through
Hillis 's interest in robots. From time to time he had mused openly about
building a robot. Word of his idea somehow reached the ear of Harry
Loucks, then director of the Mid-America Center in Hot Springs, Ark.
Would the students like to construct a robot as a display in the center's
museum? The students agreed in principle, but the project seemed too
Figure 1.1 The Tinkertoy computer: ready for a game of tic-tac-toe.
9
complicated. Just then the old Tinkertoy dream resurfaced. Would the
center like a computer made out of Tinkertoys instead?
Hillis and company assemble the first Tinkertoy computer
set out to
in a laboratory at M.I.T. The first model, unlike its successor, was a
bulky cube with sides about one meter long. It was impressively compli-
cated. Packed with logic devices made entirely of wooden sticks and
spools, the machine signaled its moves by waving nine flags from the top
of the framework. The prototype Tinkertoy computer had to be taken
apart for the trip to Hot Springs, and once it was reassembled on site, the
machine never quite worked properly again. On the other hand, it did
make an intriguing exhibit. It has been displayed at the Computer Mu-
seum in Boston.
In1979 Loucks contacted the group again. Could they make a new
Tinkertoy computer, one that worked? By this time Silverman was in
Ottawa and Hillis in Boston, each pursuing a new career. Over the tele-
phone Hillis and Silverman worked out an improved design. It was to be
reliable, and that meant simple. They decided to lay out all the possible
tic-tac-toe boards in a row and devise some kind of reading mechanism
that would move from row to row until it found a pattern matching the
current board. The very act of recognition could trigger a pre-set re-
sponse.
While contemplated ways to represent tic-tac-toe boards with
Hillis
digital Tinkertoy components, Silverman analyzed the game. To appre-
ciate the complexities involved even in this childhood pastime, readers
might consult the game tree shown in Figure 1.2. In the middle of the
tree sits the initial board, a three-by-three grid empty of X's and O's.
From this initial board nine new ones can arise, depending on which of
the nine squares X plays. The figure shows just three possibilities; the re-
maining Each of the three boards at the second
six are rotated versions.
level gives rise to other cases.For example, the board in which X plays
the center square and then O plays another square results in two differ-
ent boards. The remaining two boards at the second level each generate
five new boards at the third level.
I pruned many branches from the tic-tac-toe tree by appealing to a
symmetry argument: the excluded boards are merely rotations or re-

flections of the included ones. Symmetry seems simple to humans, but a
computer must be programmed or wired to recognize it. In a world of
Tinkertoy engineering, symmetry operations would require elaborate
structures.
1
XO o
x X
X
o oX
X
X o X
o o
X X
o o
Figure 1 .2 First three levels of the tic-tac-toe game tree.
Silverman was dealing with a tree, therefore, that was many times
larger than the fragment shown in Figure 1.2. But perseverance paid off,
especially when Silverman employed a computer program that analyzed
the game of tic-tac-toe and discovered that a great many boards could be
collapsed into oneby a forced move. Suppose, for example, that two
row contain O's and the third is blank. The contents of the
squares in a
remaining two rows are irrelevant since an opponent must fill the third
square with an X or lose the game.
Silverman was delighted when he tallied up the final total of relevant
boards: only 48. For each of them he noted the appropriate move by the
machine. The surprisingly short list of possible board positions hear-
tened Hillis. The group converged on Hot Springs, Silverman says,
"with the list of 48 patterns and only a vague idea of how to interpret
them mechanically."
(Readers who have a fanatical bent —
or are stranded in airline
terminals —
may enjoy working out the game tree on a few sheets of
paper. How long does it take, after all, to draw 48 tic-tac-toe patterns?
Four symbols should help sort things out: X, O, blank and a dash for
"don't care.")
Once settled in Hot Springs, the team assembled the raw material for
their spool-and-stick odyssey: 30 boxes of Tinkertoys, each containing
250 pieces. Some team members put together the supporting frame-
work that would hold all 48 memory spindles. To explain precisely how
the spindles were made, I must digress for a moment and describe the
conventions employed by the team to encode tic-tac-toe positions.
First, the squares of a tic-tac-toe board were numbered as follows:
1 2 3
4 5 6
7 8 9
Then a memory spindle was divided conceptually into nine consecutive

lengths in which information about the status of each tic-tac-toe square
was stored from left to right.
Each length was further subdivided into three equal sections, one for
each possible item one might find in a square: anX, an O or a blank. Each
possibility was encoded by the lack of a spool. For example, if anXhap-
pened to occupy a certain square, the memory spindle would have no
spool in the first position, one spool in the second and one spool in the
third. Similarly, a spool missing in the second position denoted an un-
played square, and one missing in the third position symbolized an O.
Finally, if all three spools were missing, it meant that what occupied the
square was irrelevant.
One can hardly mention the subject of memory spindles without
bringing up the core piece, a thing of digital beauty. Here the Latin di-
gituscame into its own, the construction resembling a kind of rotating
claw with nine fingers. The core piece and a sample memory spindle are
shown in Figure 1.3.
MEMORY SPINDLE
x !
Figure 1 .3 A memory spindle, which encodes the X's and O's of a tic-
tac-toe board, prevents the core piece from turning.
The core piece consisted of nine equal sections. Each had its own
finger, a short stick protruding from the rim of a sliding spool. Within
each section the finger could be moved
along the axis of the core piece
into any of three possible positions: one for X, one for O and one for
blank. The core piece could therefore store any possible tic-tac-toe
board by virtue of the positions of its nine fingers as moved by the opera-
tor for each play by human or machine. In Figure 1.3, fingers in the con-
secutive positions 2, 1, 2, 3, 1, 2, 2, 2, 2 would represent the board
shown.
If the current situation of play is stored in the core piece, does the
Tinkertoy computer require any other memory? Could spool-and-stick

logic devices be strung together to cogitate on the position and ulti-
mately to signal a move? Well, yes —
but such a Tinkertoy computer
would be complicated and immense. The memory spindles eliminated
the need for most of the computer's cogitation. All the Tinkertoy com-
puter had to do was to look up the current board in the memory spindles.
The only purpose of the search, naturally, was to decide what move to
make.
A glance at Figure 1.1 makes it clear that each memory spindle was
accompanied by a number written on a paper strip hanging next to its
output duck. These numbers were the machine's responses. As the read
head clicks down the rows of spindles, the core piece wants to turn but
cannot as long as at least one memory-spindle spool blocks one of the
core piece's nine fingers. Only when the read head falls adjacent to the
spindle that matches the current board do all nine fingers miss. Then the
core piece whirls.
By mechanism that would do Rube Goldberg proud, a stick pro-
a
truding from the end of the core piece engages another stick connected
to the output duck. The spinning core piece thus kicks the duck off its
perch to peck at a number writ large on the paper strip.
Computer purists will ask whether the Tinkertoy contraption really
deserves the title "computer." It is not, to be sure, programmable in the
usual sense: one cannot sit at a keyboard and type in a program for it to
follow. On the other hand, one could certainly change the memory spin-
dles, albeit with some difficulty, and thus reprogram the computer for
other games. Imagine a Tinkertoy device that plays go-moku narabe (a
game played on an 11-by-ll board in which one player tries to place five
black stones in a row while preventing an opponent from creating a row
of five white stones). A Tinkertoy computer programmed for go-moku
narabe, however, would probably tower into the stratosphere.
The Tinkertoy computer can teach us resides in a

real lesson the
rather amazing feature of digital computation: at the very root of a com-
putation lies merely an essential flow of information. The computer
hardware itself can take on many forms and designs. One could build
perfectly accurate computers not only of Tinkertoys but also of bamboo
poles, ropes and pulleys (as in Chapter 2), plastic tubes and water
even, strange to think, electrical components. The last-named are pre-
ferred, of course, because of their speed. It would be short-sighted in-
deed to sneer at a computer made of Tinkertoys merely because it is not
electronic. After all, even electrons and wires may not be the best mate-
rials for quick computer processing. Photons and fibers are gaining on
them fast.
Actually, Tinkertoys are well suited to digital computing. For exam-

ple, the memory spindles use a binary principle: the presence or absence
of spools denotes the status of a particular square on a tic-tac-toe board.
The core piece can turn only if all its fingers miss
exhibits digital logic: it
corresponding spools on a memory spindle. Such an operation is called

"and." One can trace the logic for the core piece in Figure 1.3: if the first
spool is absent from the first section of the memory spindle and the sec-
ond spool is absent from the second section and the third spool is absent
from the third section and so on —
only if all nine conditions are met will
the core piece turn. The beauty of the Tinkertoy computer is not just its
clever mechanics but its subtle logic.
Tinkertoy purists will be happy to know that the M.I.T. students
stuck to the original wooden sticks and spools with only a few exceptions
(see Figure 1.4). An aluminum rod runs through the frame-
occasional
work to strengthen it. Two wire cables, an axle and a crank transmit mo-
tive power to the awesome machine for its next move. Finally, the very
joints of sticks and spools were made firm by glue and escutcheon pins
— pieces of hardware that commonly hold commemorative plaques in
place. The team inserted the pins in holes drilled through the rim of the
spool down to the original, central hole and through its stick —a task
they had to repeat more than 1,000 times. (When Hillis walked into a
hardware store to obtain several thousand escutcheon pins, the manager
looked bewildered. " We have," Hillis said with a straight face, "a lot of
escutcheons.")
The Tinkertoy tic-tac-toe computer suffered the fate of most mu-
seum exhibits. It was taken apart and crated. At latest report, it was sit-
ting in storage at the Mid-America Center, waiting to reemerge, perhaps,
Figure 1.4 Edward Hardebeck helps to assemble the Tinkertoy com-

puter.
into the limelight. It may yet click its way to victory after victory, a mon-
ument to the Tinkertoy dreams of childhood.
Further Reading
Charles Babbage et al. Charles Babbage: On the Principles and Development of
the Calculator and Other Seminal Writings. Philip Morrison and Emily Mor-
rison, eds. Dover Publications, 1961.
Sing H. Lee and Ravindra A. Athale, eds. Optical Computing. Special issue of
Optical Engineering, Vol. 28, No. 4 (April 1989).
The Apraphulian excursion fooled few people when it
first appeared in Scientific American in April, 1988.
A. K. DEWDNEY, The Magic Machine
n the island of Apraphul off the northwest coast of New Guinea,

archaeologists have discovered the rotting remnants of an ingenious ar-
rangement of ropes and pulleys thought to be the first working digital
computer ever constructed. Chief investigator Robert L. Ripley of
Charles Fort College in New York dates the construction to approxi-
mately a.d. 850.
The Apraphulians were excellent sailors. Their ships were wonder-
fully built and equipped with the most elaborate rigging imaginable.
Were the Apraphulians led to the digital computer by their mastery of
rope or was it the other way around? Experts continue to debate the
topic hotly.
The ancient rope-and-pulley computer has been partially recon-
structedby Ripley and his team at the Tropical Museum of Marine An-
tiquities innearby Sumatra. Scouring a site that extends through several
kilometers of dense jungle east of the Pulleg Mountains, the group found
faint traces of buried jute fibers and noted the exact position of badly
corroded brass pulleys and associated hardware. The reconstruction has
given me an ideal opportunity to introduce readers to the principles of
digital computing without resorting to tiny and mysterious electronic
components. Here are gates, flip-flops, and circuits made entirely of
rope and pulleys. It is all visible and perfectly easy to understand.
The Apraphulians used a binary system just as we do, but the num-
bers 0 and 1 were represented by the positions of ropes instead of by
electric voltages. Imagine a black blox with a hole drilled in one side. The
16
reader holds a taut rope that passes through the hole. This position of the
rope represents the digit 0. If the reader now pulls on the rope, a creak
and squeal inside the box is heard as a foot or so of rope comes out. The
new position of the rope represents the digit 1.
One can represent numbers with such boxes. Any number from 0
through can be represented by three boxes (Figure 2.1).
7, for instance,
By employing more boxes, larger numbers can be represented. Ten
boxes suffice to represent all numbers from 0 through 1023.
My example of the black box is not arbitrary. The Apraphulians ap-
parently loved to enclose their mechanisms in black wood boxes, small
and large. It may be that the construction of computers was the preroga-
tive of a special technological priesthood. The sight of great assemblages
of black boxes may have kept the masses trembling in awe.
One of the key devices used by the Apraphulians converted a 0 into a
1 and a 1 into a 0. (It is occasionally convenient to speak of 0 and 1 in-
stead of "in" and "out." ) Akin to what modern computer engineers call
an inverter, this interesting mechanism consisted of a box with a hole
drilled in its front and another in its back (Figure 2.2). When someone
(or something) pulled the input rope at the front of the box, an equal
amount of output rope would be played out of the hole in the back. On
BOX 1 BOX 2 BOX 3 NUMBER

IN IN IN 0
IN IN OUT 1
IN OUT IN 2
IN OUT OUT 3
OUT IN IN 4
OUT IN OUT 5
OUT OUT IN 6
OUT OUT OUT 7
Figure 2. 1 How Apraphulians represented numbers.

^SPRING
PULLEY
H
&
1 1
Figure 2.2 The Apraphulian inverter.
peering into the box, the reason is obvious: The ropes entering the box
from front and back pass over two fixed pulleys toward one side of the
box, where they attach to a single spring.
As some readers may have surmised already, the digits 0 and 1 were
not encoded so much by "out" and "in" as they were by the direction in
which the rope moved. The point is best illustrated by a box that has no
mechanism in it whatever. A piece of rope enters a single hole in the
front of the box and leaves by a single hole in the back. If one pulls the
rope from the 0 position to the 1 position at the front of the box, the rope
moves from "in" to "out." The direction of movement is toward the
puller. The rope simultaneously moves from "out" to "in" at the back of
the box, but since the direction of movement is still toward the puller,
the rope at the back of the box also moves from 0 to 1.
Two additional mechanisms almost complete the ancient Apraphu-
lian repertoire of computing components. The first mechanism had two
input ropes entering a box. If either rope was in the 1 position, the single
output rope would also be in the 1 position. The Apraphulians managed

this trick by absurdly simple means (Figure 2.3). Each rope entering the
front of the box passed over a pair of pulleys that brought it close to the
other rope. The two ropes, passing toward the rear of the box, were then
tied to a single ring linked to the output rope. If either or both of the input
ropes were pulled, the ring would be pulled directly. Because the output
Figure 2.3 The Apraphulian OR gate.
of the box was 1 if one input or the other was 1, today's engineers would
call this an OR gate.
The ancient Apraphulians what we would call an AND
fabricated
gate from four pulleys and a spring (Figure 2.4). The two input ropes, in
reality the same rope, passed over three of the pulleys, two of which
acted as guides. The third pulley acted as a numerical divider; if one
pulled one input rope by the amount x, the third pulley would move
toward the front of the box by the amount Vi x. If x should happen to be
one unit, indicating an input of 1, nothing would happen at the output
end owing to a curious linkage between the third pulley and a fourth one
situated in the back of the box. The third pulley was attached to the
fourth by means of two rods (ropes would do equally well) joined by a
weak spring. When the third pulley moved Vi unit toward the front of the
box, the spring would extend and a parallel rope of Vi unit length would
tighten to take up the slack. If the other input rope were now pulled into
the 1 position, the third and fourth pulleys would move in unison V2 unit
toward the front of the box. Since the fourth pulley acted as a two-multi-
plier, multiplying any forward motion by 2 in terms of its associated out-
put rope, the ensemble would convert the second 1 input into an output
of 1.
Figure 2.4 The Apraphulian AND gate.
The name AND gate is derived from the fact that the output of this
device is and only if one input rope and the other are in position 1.
1 if
With these components one can build all the control circuits of a digi-
tal computer. These include circuits that compute arithmetic functions,
interpret program code, and direct the flow of information among the
parts of the computer.
Did the Apraphulians construct their computer along such lines?
The evidence is too fragmentary to reach a definitive conclusion, but ar-
chaeo-computologists working with Ripley maintain they have discov-
ered a simple multiplexer within the half-buried complex. In electronic
computers a multiplexer is essentially an electrical switch that directs
the passage of many signals through a single wire. For example, the sim-
plest multiplexer would have two input wires we might label a and b. At
any given moment each wire could carry a 0 or 1 signal. Which of the two
signals, a or b, will be allowed to pass through the device and out a single
output wire d? The answer to that question is the business of a control
wire, c; if it carries a 1 signal, the signal from wire a will be transmitted
along the output wire. If the control wire carries a 0, on the other hand,
the signal in wire b will be transmitted (Figure 2.5).
This reconstructed double-input Apraphulian multiplexer consists
of two AND gates, an OR gate and an inverter. The whole thing is so sim-
Figure 2.5 An Apraphulian multiplexer: rope c determines whether sig-

nals from a or b reach d.
pie that one dares to believe computer recreationists might build their
own Apraphulian multiplexer at home. Hardware stores might suffer a
puzzling run on rope and pulleys. In any event, one can follow opera-
tions of the multiplexer by referring to Figure 2.5. Ropes a and b enter
the multiplexer from the top left, each going to its own AND gate. Rope c
is split. One branch runs directly to the other input port of the AND gate
to which rope a goes. The second branch of rope c passes through an in-
verter and then runs to the AND gate to which rope b goes. If rope c is
pulled to a value of 1 and held, any sequence of O's and l's sent along
rope a will be faithfully transmitted through the upper AND gate and on
to the OR gate. At the same time any signal sent along rope b will be
stopped at the lower AND gate. If rope c is relaxed to its 0 position, the
inverter creates a 1 at the lower AND gate. In this case any signal sent
along rope b will now be transmitted through the lower AND gate and
signals on rope a will be ignored.
The OR gate merely ties the two output signals together, so to speak.
If the signal from rope a is currently being transmitted, one can easily vi-
sualize exactly what happens directly from the diagram: If rope a is re-
laxed to the 0 position, the rear pulley in the AND box moves toward the
rear of the box. A 0 is thus transmitted along the output rope and into the
OR box. The other input rope to this box is already in the 0 position
(slack). The natural tension on the output rope d immediately pulls it
into the new position, namely 0. If one pulls on rope a again, the pull is
transmitted along the path that has just been described, with the result
that rope d is retracted.
The matter of slack ropes compels me to take up the question of ten-
sion in the Apraphulian computer. Sometimes, as in the OR gate of the
example, a rope will become slack. There is naturally a danger that such
ropes will slip right off their pulleys. Ripley tells me that in such cases the
Apraphulians used a specially modified inverter with an extremely weak
spring to remedy the problem. Wherever a rope was likely to develop
slack, a "weak inverter" was installed to maintain the minimum tension
associated with the signal 0.
No general-purpose computer is complete without a memory. The
memory of the Apraphulian computer consisted of hundreds of special
storage elements we would call flip-flops. Here again the remarkable
simplicity of the Apraphulian mind is immediately evident. In line with
modern terminology, the two ropes entering the mechanical flip-flop are
labeled set and reset (Figure 2.6). The two ropes were connected over a
series of three pulleys in such a way that when the set rope was pulled
away from the box into the 1 position, the reset rope would be pulled
toward the box into the 0 position. The common rope was connected to
a sliding bar at the back of the flip-flop box. The output rope, physically a
continuation of the set rope, had a large bead attached to it that engaged
a slot in the sliding bar. As the set rope was pulled, the bead rode over the
end of the bar, popping into the slot when the set rope reached the end of
In this position, 1 was "remembered."
its travel.
As a consequence the output rope was held in position until the enor-
mous rope computer changed things by pulling on the reset rope. That
SET
RESET
Figure 2.6 The Apraphulian flip-flop served as a memory element.
had the effect of pulling the sliding bar away from the bead, releasing it
and playing the output rope into the 0 position. In this case the flip-flop
would henceforth "remember" 0. How were such memory elements
used in the Apraphulian computer?
Ripley and his team were puzzled to discover in the midst of the vast
Apraphulian computer complex a large overgrown field nearly a kilome-
ter wide. Buried just below the surface of the field were several thousand
rotting flip-flop boxes arranged in rows of eight. Ripley, with the aid of
the archaeocomputologists, eventually surmised that the field repre-
sented the Apraphulian computer's main memory. Each row of eight
boxes would have constituted a single, eight-bit "word" in the same
sense that the three boxes of my earlier example would have constituted
a three-bit word. In that vein, imagine a row of three flip-flops that had
been set to the values 1, 0, and 1. They would have stored the number 5.
The content of this particular memory word would have been ac-
cessed by the rope-and-pulley computer as follows. Each flip-flop in the
row would send an output rope to an associated AND box. The other
input to each AND box would come from a special rope used to retrieve
the contents of the word in question. When the ropes were pulled, the
outputs of the AND boxes would be identical with the outputs of the
flip-flops. The AND box ropes would lead to a large assemblage of OR
boxes and thence into a special array of flip-flops we would call a regis-
ter. A single tug on the rope associated with the word under examination
would place the same binary pattern of rope positions in the register.
The computer's main logic unit undoubtedly would have directed
the flow of information not just from memory to registers but between
registers as well. In particular, by the use of multiplexers and demulti-
plexers (which perform the opposite function of multiplexers), the com-
puter would have sent patterns from register to register. At a specific reg-
ister that we would call the arithmetic register, patterns would have been
combined according to the rules of addition and multiplication.
The Apraphulian computer is believed to have been programmable.
If it was, part of its vast memory would have been used to store the pro-
gram. Program instructions would also have been merely patterns of O's
and 1's retrieved by the same mechanism outlined earlier. Those pat-
terns would in due course have been sent to an instruction register for
interpretation by the computer's logic unit.
It is a pity I can do little more in these pages than to hint at the mar-
velous complexity of the Apraphulian machine. It must have been an

amazing sight when in operation. Because of the enormous lengths of
rope involved, no human being would have had the strength to pull the
input levers into the appropriate positions. The presence of elephant
bones in the Apraphulian complex makes the source of input power im-
mediately clear. At the output end large springs maintained appropriate
tensions in the system. Perhaps flags on the ultimate output ropes en-
abled members of the technological priesthood to read the outcome of
whatever computation was in progress.
The Apraphulian rope-and-pulley computer makes for an interesting
contrast with the recently proposed nanocomputer (see Further Read-
ing). The rope machine, of course, inhabits a distant past, whereas the
nanometer-scale machine dwells in a hazy future. The Apraphulian
computer is relatively massive in scale, covering thousands of acres; the
nanocomputer is incredibly tiny, occupying an area one-thousandth the
size of a human cell nucleus. The mere concept of either machine serves
as a springboard into a speculative realm where recreation blends with
science. Think, for example, of the ongoing dream of artificially intelli-
gent machines. We find it easier to accept the possibility of an electronic
computer that thinks since our own thoughts are to a great extent elec-
tronically mediated. Because any modern computer (and its program) is
conceptually translatable into Apraphulian form, any artificially intelli-
gent device ever realized now or in the future will have its rope-and-pul-
ley counterpart. Can we imagine HAL 9000, the paranoid computer in
61-
A+ B
Figure 2.7 An Apraphulian adding machine.
the movie 2001: A Space Odyssey, being so constructed? Are we willing

to admit that an enormous building full of ropes and pulleys could be just
as smart as we are? (For a partial answer, see Chapter 3.)
We leave the island of Apraphul with just one backward glance at its
misty past: How might the vast digital computer have evolved? From
analog ones, of course. Figure 2.7 shows an analog adding machine made
from two ropes and two pulleys. The two ends of one rope enter the
front of a box through two holes. The rope passes over a single pulley
that is linked with another pulley by an axial connector. One end of the
second rope is attached to the back of the box. The rope passes over the
second pulley and then through a hole in the back of the box. Readers
might find some diversion in discovering for themselves how the ma-
chine adds two numbers; if the two input ropes are pulled a distance a
and b respectively, the output rope travels a distance a + b.
So much is clear. But how did the Apraphulians manage analog mul-
tiplication?
Further Reading
A. K. Dewdney, "Atomic Computers," in The Magic Machine, W. H. Freeman,
1990.
3-
mil
\
j u RAIN
"Are minds subject to the laws of physics? What,

indeed, ore the laws of physics?"
ROGER Penrose, The Emperor's New Mind
T_T
JL JL uman intelligence outstrips artificial intelligence because it ex-
ploits physics at the quantum-mechanical level. That is a provocative
position, but one that Roger Penrose, the noted mathematical physicist,
leans toward in his book, The Emperor's New Mind. Although (as
Penrose readily admits) the proposition cannot be rigorously proved at
present, the intriguing arguments in The Emperor's New Mind have pro-
duced some healthy doubts about the philosophical foundations of arti-
ficial intelligence.
I shall present Penrose's arguments — but because this book follows

its own compass unknown waters, I shall challenge some of
in charting
his conclusions and tinker with some of his ideas. In particular, I shall
expand the question How do people think? to ask whether human
beings will ever know enough to answer such a question. If the universe
has an infinite structure, humans may never answer the question fully.
An infinite regress of structure, on the other hand, offers some unique
computational opportunities.
Before jumping into such matters, I invite readers to explore the re-
cesses of The Emperor's New Mind with me. First, we shall visit the
'
famed Chinese room to inquire whether 'intelligent' programs under- '
stand what they are doing. Next, a brief tour of the Platonic pool hall will
bring us face to face with a billiards table that exploits the classical phys-
ics of elastic collisions to compute practically anything. Moving along to
Erwin Schrodinger's laboratory, we shall inquire after the health of his
cat in order to investigate the relation between classical physics and
26
quantum mechanics. Finally, we shall reach our destination: an infinite

intelligence able to solve a problem that no ordinary, finite computer
could ever hope to conquer.
Watching television one evening several years ago, Penrose felt
drawn to a BBC program in which proponents of artificial intelligence
made what seemed to be a brash claim. They maintained that com-
puters, more or less in their present form, could some day be just as in-
telligent as humans —
perhaps even more so. The claim irritated
Penrose. How could the complexities of human intelligence, especially
creativity, arise from an algorithm churning away within a computer
brain? The extremity of the claims "goaded" him into the project that
led to The Emperor's New Mind.
A methodical exploration of computing theory brought Penrose to
criticize one of its philosophical cornerstones, the Turing test. Many
computer scientists accept the test as a valid way of distinguishing an in-
telligent program from a nonintelligent one. In the test, a human interro-
gator types messages to two hidden subjects, one a person and the other
a computer programmed to respond to questions in an intelligent man-
ner. If, after a reasonable amount of time, the interrogator cannot tell the
differencebetween the typed responses of the human and those of the
computer, then the program has passed the Turing test.
Penrose argues that the test provides only indirect evidence for
intelligence. After what may appear to be an
all, intelligent entity may
turn out to be a mockery, just as an object and its mirror image look
identical but in other details are different. Penrose maintains that a di-
rect method for measuring intelligence may require more than a simple
Turing test.
To strengthen his argument, Penrose wanders into the Chinese

room, a peculiar variation of the Turing test invented by philosopher
John R. Searle. A human interrogator stands outside a room that only
allows the entrance and exit of paper messages. The interrogator types
out a story and related questions and sends them into the room. The
twist: all messages that go into and out of the room are typed in Chinese
characters.
To make matters even more bizarre, a person inside the room exe-
cutes a program that responds to the story by answering questions about
This person exactly replaces the computer hardware. The task would
it.
be tedious and boring but, once the rules of execution were learned,
rather straightforward. To guarantee the ignorance of the human hard-
ware, he or she has no knowledge whatsoever of Chinese. Yet the Chi-
nese room seems to understand the story and responds to the questions
intelligently.
The upshot of the exercise, as far as Penrose is concerned, is that
"the mere carrying out of a successful algorithm does not in itself imply
that any understanding has taken place." His conclusion certainly holds
if it is directed at the executing apparatus, whether flesh or hardware.
After all, whether the program happens to be executed by a human or by
a computer makes no difference, in principle, to the outcome of the pro-
gram's interaction with the outside world.
But for this very reason the human in the Chinese room is something
of a straw man: no one would fault a program because the hardware fails
to understand what the program is all about. To put the point even more
strongly, no one would be critical of a neuron for not understanding the
significance of the pulse patterns that come and go. This would be true
whether or not the neuron happened to be executing part of an algorithm
or doing something far more sophisticated. Any strength in claims for
artificial intelligence must surely lie in the algorithm itself. And that is
where Penrose attacks next.

The world of algorithms is world of the computable.
essentially the
In Penrose's words, algorithms constitute "a very narrow and limited
part of mathematics." Penrose believes (as I and many other mathema-
ticians do) in a kind of Platonic reality inhabited by mathematical ob-
jects. Our clue to the independent existence of such objects lies in our
complete inability to change them. They are just "there," like mountains
or oceans.
Penrose cites the Mandelbrot set as an example. The Mandelbrot set
was not "invented" by Benoit B. Mandelbrot, the renowned IBM re-
search fellow, but was discovered by him. Like the planet Neptune, the
set existed long before any human set eyes on it and recognized its signif-
icance. The Mandelbrot set carries an important message for those who
imagine it to be a creature of the computer. It is not. The Mandelbrot set
cannot even be computed! Do I hear howls of outrage? Strictly speaking,
Penrose is right.
The Mandelbrot set, while it is but one landmark in the Platonic
world, lies somewhat distant from algorithmic explorers. Readers may
recall that points in the interior of the set can be found by an iterative
process: a complex number c is squared and then the result, z l9 is
squared and added to c, then the second result, z2 is squared and added
,
to c and so on. If the succession of z values thus produced never patters

off into infinity, then c belongs to the set's interior. But here a grave
question emerges. How long does one have to wait to decide whether
the sequence of values remains bounded? The answer is, essentially, for-
ever (Figure 3.1)!
In practice one interposes a cutoff to the computation. In doing so
one inevitably includes a few points that do not belong in the set because
it takes longer for the sequence based on such points to diverge. Difficul-
ties in computing the Mandelbrot set pale in comparison with other limi-
tations on the algorithmic adventure. For example, mathematics itself is
formally considered to be built of axiom systems. Set forth a modest col-
lection of axioms, a rule of inference or two —
and one is in business. A
conceptual algorithm called the British Museum Algorithm generates all
the theorems that are provable within the formal systems of axioms and
inference rules. Unfortunately, the theorems thus produced do not nec-
essarily include all truths of the system.
This discovery, by the mathematician Kurt Godel, dashed hopes of
mechanizing all of mathematics. Penrose takes GodePs famous theorem
as evidence that human intelligence can transcend the algorithmic
method: "... a clear consequence of the Godel argument [is] that the
Figure 3. 1 The Mandelbrot set amid

A
a field of contour lines.
concept of mathematical truth cannot be encapsulated in any formal

scheme." How then could Godel's theorem itself be the result of an algo-
rithm?
As a theorem, GodePs result is itself completely formalizable. It can
be derived, in principle, by a systematic theorem generator from the ap-
propriate axioms and rules of inference. Penrose really should have
stopped to think of that.
Whatever one's opinion on such questions, The Emperor's New Mind
attacks the claims of artificial intelligence on another front: the physics
of computing. Penrose hints that the real home of computing lies more
in the tangible world of classical mechanics than in the imponderable
realm of quantum mechanics. The modern computer is a deterministic
system that for the most part simply executes algorithms. In a somewhat
jolly fashion, Penrose takes a billiard table, the scene of so many classi-
cal encounters, as the appropriate framework for a computer in the clas-
sical mold.
By reconfiguring the boundaries of a billiard table, one might make a
computer in which the billiard balls act as message and their in-
carriers
teractions act as logical decisions. The billiard-ball computer was first
designed some years ago by the computer scientists Edward Fredkin and
Tommaso Toffoli of the Massachusetts Institute of Technology. The
reader can appreciate the simplicity and power of a billiard-ball com-
puter by examining Figure 3.2.
The diagram depicts a billiard-ball logic device. Two in-channels
admit moving balls into a special chamber, which has three out-chan-
nels. If just one ball enters the chamber from either in-channel, it will
leave by either the bottom out-channel or the one at the upper right. If
two balls enter the chamber at the same time, however, one of them will
leave by the out-channel at the lower right. The presence or absence of a
ball in this particular out-channel signals the logical function known as
an AND gate. The output is a ball if, and only if, a ball enters both one
in-channel and the other one.
A computer can be built out of this particular gate type and just one
other, a chamber which a ball leaves by a particular channel if, and
in
only if, a ball does not enter by another channel. Readers may enjoy try-
ing to design such a chamber, bearing in mind that additional balls might
be helpful in the enterprise.
Everyone appreciates the smooth, classical motions of a billiard ball.
Ithas other desirable properties that are hardly given a second thought.
For example, no one ever has to worry whether a billiard ball is in two
Figure 3.2 A billiard-ball AND gate.
places at the same time. Quantum mechanics, however, produces such

anxieties. Quantum systems such as the famous two-slit experiment
leave open the possibility that a photon can be in two places at once.
Briefly, when photons pass through a double slit, they can be re-
garded as waves that interfere with themselves. An interference pattern
emerges on a screen behind the slits unless one places a detector at either
slit. The act of observation forces the photon to decide, in effect, which
hole will pass through! The phenomenon is called a state vector col-
it
lapse. The experiment can be extended to an observation that takes

place at either of two sites that are a kilometer (even a light-year) apart.
The photon can decide which slit it will pass through, many physicists
claim, only if it is effectively in both places at once.
At what point in the continuum of scales, from the atomic to the ga-
lactic, does a quantum-mechanical system become a classical one? The
dilemma is illustrated by Schrodinger's famous cat in Figure 3.3. In this
gedanken (thought) experiment, a scientist who has no fear of animal-
rights activists places a cat and a vial of toxic gas in a room that contains a
laser, a half-silvered mirror, a light detector, and a hammer.
When the room is sealed, the laser emits a photon toward the mirror.
If the photon passes through the mirror, no harm comes to the cat. But if
the photon is reflected in the mirror, it hits the detector, which activates
the hammer, which smashes the vial, which contains the gas, which
kills the cat. From outside the room one cannot know whether the cat
lives or dies.
In the quantum-mechanical world, the two possible events coexist
as superposed realities. But in the classical world,only one event or the
other may occur. The state vector (and possibly the cat) must collapse.
Penrose suggests that current theory lacks a way of treating the middle
ground between classical physics and quantum mechanics. The theory is
split in two but should be seamless on a grand scale. Perhaps the synthe-
sis will come from an area known as quantum gravity.
Now back, finally, to the human mind. For Penrose, consciousness

has a nonalgorithmic ingredient. At the quantum level, different alterna-
tives can coexist. A single quantum state could in principle consist of a
number of different, simultaneous activities.
large Is the human brain
somehow able to exploit this phenomenon? I can hardly explore this
eerie possibility as well as Penrose does. Readers intrigued by the
thought had best buy the book.
I was inspired, however, to investigate a related question. Could

human beings quantify their own intelligence in a universe whose struc-
ture goes on forever? An end to the structure of matter, either as an ulti-
mate particle or a set of particles, seems inconceivable. By this I mean
not just particles but any structure, whether energetic or even purely in-
formational, underlying the phenomena in question.
It seems to me that physics itself may be an infinite enterprise for the
simple reason that as soon as some "ultimate" structure is discovered,
explaining the existence of the "ultimate" laws becomes the next prob-
lem. In any event, would prefer to live in an infinitely structured uni-
I
verse. For one our minds might turn out to be much more power-
thing,
ful than if the structure went only so far.
Computers are constructed to rule out the influence of any physical
process below a certain scale of size. The algorithm must be protected
from "errors." Our brains may or may not be so structured, as Penrose
points out. Physical events at the atomic level might well have an impor-
tant role to play in thought formation. Processes at the molecular level
certainly do. One has only to think of the influence of neurotransmitter
molecules on the behavior of a neuron. Furthermore, it is a well-known
characteristic of nature to take advantage of physical possibilities in the
deployment of biological operations. If physical structures extend to a
certain level, is there some a priori reason to believe that the brain must
automatically be excluded from exploiting it?
What if the brain could exploit all levels of structure in an infinitely

structured universe? To demonstrate in the crudest imaginable way
the potential powers of an infinite brain, I have designed an infinite
computer that exploits the structure at all levels. For the purposes
of the demonstration, I will pretend that the structure is classical at all
scales.
My computer (Figure 3.4) is essentially a square that con-
infinite
tains two rectangles and two other, smaller squares. An input wire enters
the big square from the left and passes immediately into the first rectan-
gle. This represents a signal-processing device that I shall call a substitu-
tion module. The substitution module sends a wire to each of the two
small squares and also to the other rectangle, henceforth called a mes-
sage module.
The structure of the whole computer regresses infinitely. Each of the
two smaller squares is an exact duplicate of the large square, but at half
the scale. When a signal is propagated through the wires and modules at
half the scale, it takes only half as long to traverse the distances involved,
PI
Hi
1st
101
-PUT
III
ill
t».r
Figure 3.4 A fractal computer solves Thue's word problem.
and so the substitution and message modules operate twice as fast as the
corresponding modules one level up.
The infinite computer solves the famous "word problem" invented
by the mathematician Axel Thue. In this problem, one is given two
words and a dictionary of allowed substitutions. Can one, by substitu-
tions alone, get from the first word to the second one? Here is an exam-
ple of the problem: suppose the first "word" is 100101110 and the sec-
ond is 01011101110. Can one change the first word into the second by
substituting 010 for 110, 10 for 111, and 100 for 001? The example is
chosen at random. I deliberately refrain from attempting to solve it.
It might happen that no sequence of substitutions will transform the
first word into the second. On the other hand, there might be a sequence
of substitutions that does the job. In the course of these substitutions,

intermediate words might develop that are very long. Therein lies the
problem. As with certain points in the Mandelbrot set, one cannot effec-
tively decide the answer. There is no algorithm for the problem, because
an algorithm must terminate, by definition. The danger is that the algo-
rithm might terminate before the question is decided. Thue's word
problem is called undecidable for this reason. No computer program,
even in principle, can solve all instances of this problem.
Enter the infinite computer. The target word is given to the com-
puter through the main input wire. It enters the first substitution module
in Va second. The word is then transmitted by the substitution module to

the two substitution modules at the next level. But this transmission
takes only Vs second. Transmission to subsequent levels takes Vi6, then
V32 second, and so on. The total time for all substitution modules to be-
with the target word
'
come 'loaded' ' is therefore half a second.
Next, the three (or however many) substitution formulas are fed into
the computer by the same process and at the same speed. This time, how-
ever, the various substitution modules at different levels are prepro-
grammed to accept only certain substitutions in the sequence as their
own, and they are preprogrammed always to attempt a substitution
also
at a specific place in a word that arrives from a higher level A recital of
the distribution scheme for farming out the substitutions and their
places would probably try the reader's patience, and so I shall omit it.
This should not prevent those who enjoy infinite excursions from imag-
ining how it might be managed.
The computation begins when one sends the first word into the
computer. The first substitution module attempts to make its own sub-
stitution at its allotted place in the incoming word. If the substitution
cannot be made in the required place, the substitution module transmits
the word to the lower square in the next level inward; if it succeeds in
making the substitution, it word to the upper
transmits the transformed
square. If the substitution succeeds and the newly produced word
matches the target word stored in the substitution module's memory,
then the module sends a special signal to the message module: "suc-
cess."
Each square at each level operates exactly in this manner. As I indi-
cated earlier, it is possible to distribute substitutions (and places at which
they are to be attempted) throughout the infinite computer in such a way
that the word problem will always be solved. The question takes at most
one second to decide in all cases: half a second for the computation to
proceed all the way down to infinitesimal modules and half a second for
the message "success" to reach the main output wire. If no substitution

sequence exists, the absence of a message after one second may be taken
as a "no" answer. Readers may enjoy pondering the infinite computer
while exploiting the many (perhaps infinite) structures of their own
brains.
Further Reading
Roger Penrose. The Emperor's New Mind: Concerning Computers, Minds, and
the Laws of Physics. Oxford University Press, 1989.
Robert A. Wilson. Schrbdinger's Cat Trilogy. Dell, 1988.
J IP
H
III IF \\ mm
As Gregor Samsa awoke one morning from uneasy
dreams he found himself transformed in his bed into a
gigantic insect.
Franz Kafka, The Metamorphosis
W T ill the next major advance in robotics spring forth from inex-
pensive machines that crawl, think, and act like bugs? Researchers in the
"insect lab" at the Massachusetts Institute of Technology hope so. They
have spawned a swarm of small robots that behave like your average
arthropod. These insectoids, as I call them, are based on new principles
of robot design and threaten a paradigm shift in the field of robotics.
Until recently, engineers bent on designing a robotic "brain" have
taken a determinedly analytic approach. In this traditional view, they
first decide what the robot will be able to sense; they then consider how
it will analyze sensory inputs and finally how it will plan and take action.
Each step is fraught with complexities that are likely to bog down intri-
cate projects.
Abandoning the traditional approach, Rodney Brooks, director of
the insect lab, has adopted a design philosophy that he calls subsumption
architecture. To apply this philosophy, he starts by designing a network
of processors and hardware that can produce a simple behavior. No be-
havior is added to the system until the behavior it subsumes is up and
running (or walking, as the case may be).
For instance, to design an artificial creature that wanders and avoids
obstacles, Brooks would first assemble a creature that moved randomly
and then add the detectors and processors that would sense objects and
instruct the creature to change direction. In subsumption architecture,
complex behaviors evolve from a variety of simple features.
37
To subsumption architecture, Brooks and a

test the practicality of
team of his graduate students began building and designing many insec-
toids, from Allen, a primitive robot on wheels, to Squirt, a delicate bug
no bigger than a grasshopper. But no creature illustrates subsumption
architecture as well as Genghis, a foot-long assembly of motors, struts,
gears and microchips.
Genghis, who was created in part by graduate student Colin Angle,
has six stiltlike legs, two whiskers and six infrared "eyes" transplanted
from burglar alarms. Each leg is operated by a pair of motors (Figure
4.1). An Alpha Motor moves the leg forward or back; a Beta Motor
swings the leg up and away from the body or moves it down and toward
the body. Between Genghis's legs are microchips that serve as the insec-
toid's nerve center. The microchips contain numerous augmented fi-
nite-state machines (AFSMs). Each AFSM stores numerical information
for controlling various aspects of Genghis' s behavior, such as the move-
ment of legs. The information, or state, of an AFSM may change from
time to time, depending on the input it receives from other modules. The
state will also determine how it reacts to the inputs.
Figure 4. 1 A robotic insect named Genghis is a walking test bed of sub-

sumption architecture. It can avoid obstacles and stalk people.
A robot must walk before it can run. In fact, to do anything worth-

while, Genghis must first Two numbers sent to the AFSMs
stand up.
control the Alpha and Beta Motors on each leg. One AFSM, called Al-
phapos, controls the Alpha Motor, and the other AFSM, Betapos, gov-
erns the Beta Motor. Each number represents vertical and lateral leg po-
sitions when Genghis is standing. As soon as Genghis is powered up in
this simplest of behavioral settings, the motors all run until the leg posi-
tions (monitored by sensors) match the numbers stored in Alphapos and
Betapos. The simple act constitutes what might be called the zero level
of Genghis's architecture.
The next level of behavior, simple walking, is a feat robotics re-
searchers have traditionally judged to be technically difficult. Genghis's
basic walking network, in its simplest form, consists of two master
AFSMs and 30 auxiliary ones, five per leg. Because the circuits for each
of the insectoid's six legs are essentially the same, I will describe what
happens to one leg and the five AFSMs that control it (Figure 4.2).
The key to basic walking is the global controller called Alpha Bal-
ance. This AFSM receives continual reports, in the form of numbers, on
the positions of all six legs. A positive number indicates that a leg is
pointing forward; a negative number, that it is pointing backward. Not
surprisingly, legs that point straight out from the body are represented
by zero. Alpha Balance adds these numbers together, the sum being a
kind of average. If the sum is positive, it means that on average the legs
are pointing forward. If the sum is negative, the average leg projects
rearward.
The whole trick to walking revolves around the fact that if five of the
legs touch the ground and a sixth is raised, then the insectoid may glide
forward by a small amount merely by swinging all its ground legs slightly
to the rear. If the insectoid then swings the upraised leg forward and
places it gingerly back on the ground, it is one small step for an insectoid
but one giant leap for robotics.
When Ghengis swings a leg to the front, Alpha Balance generates a
sum that is positive and then sends a negative signal to all legs that are
currently down. Their motors whine briefly, the insectoid moves for-
ward a bit and the signal is rebalanced. That is all the Alpha-Balance
Module cares about.
The manner in which the various modules interact to create the act
of walking amounts to an electronic ballet among the modules of the six
leg networks. The action begins for a particular leg when its Up-Leg
Module is activated. The activation sets off a chain of coordinated
BETAPOS ALPHAPOS
BETA MOTOR
^ ROBOT CHASSIS
ALPHA MOTOR
ALPHA
ADVANCE
ALPHA
BALANCE
Figure 4.2 Basic circuitry that allows Genghis to stand (top) and walk
(bottom).
events among the modules; the Up-Leg Module then signals the Betapos
Module, sending it a number that reflects an upraised leg position. The
Betapos Module, which controls the Beta Motor, normally receives a
positive number (that keeps the leg firmly planted) from another module
called Down Leg. The new, negative signal from the Up-Leg Module
suppresses the positive signal from Down Leg. Consequently, the Beta
Motor raises the leg to a point where its reported position matches the
new signal.
This event triggers a completion state in the Betapos Module, and it
signals this state to three other modules: Alpha Advance, Up Leg and
Down Leg. The Alpha-Advance Module, which controls the back-and-
forth motion of the leg, sends a strongly positive signal to the Alphapos
Module. The Alpha Motor whirs gently, and the leg waves forward, al-
most as if it were probing the air. When the Up-Leg Module receives the
completion signal, its action is suppressed. When the Down-Leg Module
gets the completion signal, it is activated, and the Beta Motor powers the
leg down to terra firma.
A
master module, called Walk, controls the entire movement by
sending a sequence of signals to the six Up-Leg Modules. But what se-
quence should it use? The triggering pattern most commonly used by in-
sects is called the alternating tripod gait. If the legs are labeled R for right
and L for left, as well as numbered 1,2,3 from front to back, the alternat-
ing tripods are the sets R1,L2,R3 and LI, R2, L3. In normal situations,
an insect like a cockroach will lift the first set, R1,L2, R3, leaving the
other set on the ground. This triangular stance gives stability to the cock-
roach as the first set new positions. Then
of three legs swings forward to
the other set can be raised and swung forward in the same way while the
first set provides stability.
The Walk Module may send out a sequential version of this set of
signals to the six Up-Leg Modules. Or it could send out the pattern some-
times used by stick insects: R3, LI, R2, L3, Rl, L2. There are numerous
possibilities for stable gait patterns.
Perhaps readers can figure out the gait of a millipede machine. If
there are 1,000 legs on each side of an insectoid, devise a gait that will
carry the creature forward without any leg getting dragged along by the
body.
Using the primitive network just described, Genghis could walk but
not very smoothly and not in the manner that Brooks describes as "ro-
bust." For one thing, Genghis wobbled excessively and could not clear
obstacles of even moderate height. The addition of a few more kinds of
AFSM provided a new level of subsumption architecture and a new de-
gree of behavioral competence.
A Beta-Force Module monitors the high strain that develops in a
Beta Motor when its leg has been set down in a position that supports too
much of the creature's total weight. Genghis may have stepped on a
five-centimeter rock, for example. The Beta-Balance Module for that leg
senses the unusually high force and sends a zero message that sup-
presses the leg-down message and makes the offending leg "compliant."
The leg, in other words, gives way a bit, and Genghis compensates for
the high terrain under one of its legs.
But on sloping terrain, the downhill end of Genghis

will take more
become compliant, increasing
force than the upper end, and the legs will
the pitch even more. Correcting this problem required two Pitch Mod-
ules to monitor the outputs from a pitch-measuring device. The Pitch
Modules send messages to inhibit whichever Beta-Balance Modules
have become too compliant.
When Genghis encounters an obstacle while swinging one of its legs
forward, a sensor on the motor picks up the additional strain and sends a
message to an Alpha-Force Module. This AFSM then sends a signal to
the Up-Leg Module, which then results in a higher leg lift.
Among the many sensors used by Genghis are two whiskers and six
infrared sensors.The whiskers send their reports to a feeler module. If a
whisker senses an obstacle, the feeler module resets the Up-Leg Module
for one of the two front legs.
The infrared sensors introduce the next major level of subsumption.
The sensors work in conjunction with a Prowl Module, which gives
Genghis a somewhat sinister mode of behavior. In this mode, Genghis
rests quietly until it detects infrared radiation from, for example, a
nearby human ankle. When that happens, Genghis activates its Walk
Module. The creature then begins to creep forward like some demented
insect toward the hapless human. Of course, there is plenty of time to get
out of its way, but if a Steer Module is added as well, Genghis can be re-
lentless.
A few years ago, when a curious visitor saw the insectoid for the first
time, he asked, "Is it a bug?"
"No," Brooks said, repeating an old programming joke. "It's a fea-

ture." For a while, Brooks insisted on calling the insectoid "Feature."
But later a graduate student suggested "Genghis," which seemed more
appropriate for a creature whose instincts were to stalk and conquer.
Later some new circuits were added to Genghis to test whether self-
organizing behavior might emerge in the absence of a central control
module like Walk. The results were impressive. The microchip ganglion
associated with each leg was given the option of running its own experi-
ments with a set of basic behaviors like lifting or lowering a leg or swing-
ing it forward or back. Each experiment consisted of recording what the
neighboring legs were doing, then trying one of the basic behaviors and
checking whether the body fell down or not. Fascinating to watch, ac-
cording to Brooks, the experimenting insectoid might sit for a while, legs
waving in the air, next thrash for a bit, then begin to move forward with
tentative steps. Within a minute and a half, the network always

"learned" the alternating tripod gait!
The notion of autonomy dominates the subsumption approach to
robotic architecture in the M.I.T. insect lab. Can a robot, no matter how
small, be given a behavior that will enable it to survive in the real world
for extended periods? The insectoid called Squirt will fit inside a one-
inch cube. Too small for legs in the current state of insectoid technology,
it features wheels, a single motor, a microprocessor, two lithium batter-
ies and three sensors. It uses two microphones to listen for sounds and a
single light sensor to gauge the amount of light available.
Squirt will survive, provided it does not get stepped on. For this rea-
son, it has been programmed with several layers of behavior that were
computer to its single microchip. Squirt
transferred electronically from a
hides in the dark while listening for sounds.If it hears nothing for a few
minutes, it ventures out in the general direction of the most recently

heard sound. After wandering for a while, it engages in a spiral search to
find a new hiding place.
In this respect, Squirt resembles the vehicles imagined by the Ger-
man scientist Valentino Braitenberg. The purpose of Braitenberg vehi-
cles, among other things, was to illustrate the thesis that very complex
behaviors could result from very simple control systems. The thesis has
inspired more than one robotics enthusiast to build a behavioral vehicle.
But was Brooks and company, also intrigued in part by Braitenberg's
it
vehicles, who succeeded first. Some might be willing to ascribe emo-

tions such as fear or longing to the behavior of the neurally controlled
vehicles. Is Squirt afraid of people? It certainly acts like it.
Graduate student Anita Flynn, who was part of the team that built
Squirt, sees the future of robotics blossoming in even smaller insectoids
she calls gnats. These creatures would be the size of real insects, not to
mention gnats themselves. Their body parts would be fabricated by the
same techniques currently used to etch microcircuits on silicon surfaces.
The biggest bottleneck is the ultratiny motors that gnats will require.
The field of microengineering has already produced gears that would
scatter at a sneeze. In the meantime, Brooks and company have built
Attila, a more sophistocated version of Genghis (Figure 4.3).
Will the insectoids rule some day? Brooks is cautious about making
claims about the future of subsumption architecture, but he plans to
push the idea to its limits. Will we find that as we add ever more complex
layers of behavior that the subsumption approach will continue to work?
Figure 4.3 Anatomy of an insectoid named Attila is shown above and at

left.
Or will we encounter a barrier that forces us to resort to something like

traditional techniques? I am sure these questions will bug him for the
foreseeable future.
Further Reading
Ivan Amato. "Gearing Down." Science News, Vol. 139, no. 2 (January 12,
1991), pp. 26-27.
A. K. Dewdney. "Braitenberg Memoirs: Vehicles for Probing Behavior Roam a
Dark Plain Marked by Lights." Scientific American Computer Recreations,
March 1987.
A. K. Dewdney. The Magic Machine: A Handbook of Computer Sorcery. W. H.
Freeman, 1990.
Michael C. Smit and Mark W. Tilden. "Beam Robotics." Algorithm, Vol. 2, No.
2 (March 1991), pp. 15-19.
Dill I Bill
In man's brain the impressions from outside are
the imprint of the external world . . .
VICTOR WEISSKOPF, Knowledge and Wonder
N
JL ^ eural networks, assemblies of artificial neurons that
ble of learning an
seem capa-
enormous variety of different tasks, have captured the
imagination of researchers in recent years. But they have also intrigued
recreational programmers to the point of despair. Where were the
neural net programs you could write and run on a personal computer?
One now exists. Developed by Edward A. Rietman and Robert C.
Frye of AT&T Bell Laboratories at Murray Hill, N. J. in the late 1980s,
the program I call POLARNET converts polar coordinates to cartesian
coordinates by trial and error. More than this, POLARNET learns the art
of conversion by training on actual coordinates. The underlying algo-
rithm is reasonably simple and the network's learning process is fasci-
nating to watch. I have altered the algorithm slightly to make it easier for
readers to program.
Special, parallel hardware demonstrates the parallel prowess of
neurons best. But even an old-fashioned serial computer can simulate a
neural net with reasonable electronic elan —
if there are not too many
neurons. The Rietman-Frye network requires only 30 to 40 neurons to

learn coordinate conversions. A modern microcomputer can simulate a
single wave of computation through as many neurons with no noticeable
delay.
All learning networks share the same basic architecture shown in
Figure 5.1. Information enters the network on the through a layer of
left
input neurons. It then passes through one or more layers of medial

neurons and departs, on the right, through a layer of special output
46
neurons. (The illustration, which applies to the Rietman-Frye network

specifically, has only one medial layer.) The network processes the in-
formation as it flows through successive layers, from input to output.
Each neuron in a given layer communicates with every neuron in the
next layer via special synaptic connections. In formal terms, a synapse
amounts to a multiplier, or weight, which modifies the number that is
transmitted from one neuron to the other.
Individual neurons operate, like their biological counterparts, by a
process of summation. Each neuron adds together all the signals that it
receives. If it happens to reside in one of the medial layers, it will modify

the sum of its signals further by applying a special, sigmoidal function
that squeezes it non-linearly into the interval from — 1 to +1: The larger a
sum is, the more closely it will approach +1 or — 1, depending on its sign.
Figure 5.1 displays the general shape of a sigmoidal function just below
the layer of medial neurons.
Sigmoidal functions enable neural networks to respond non-linearly

to their environments. A great variety of actual functions, with names
like the hyperbolic tangent, arc tangent, Fermi function, and so on, play
the role quite nicely. Their ability to keep the outputs of intermediate
neurons bounded (between 0 and 1) is just as important as their non-lin-
ear response. My version of POLARNET will use the Fermi function for
this purpose:
1/(1 -e~x )
When the network embodied in POLARNET has been trained and is
ready to run, the user of the program merely types in the polar coordi-
nates to be converted. The program then gives the numbers to the two
input neurons and the network does the rest, in a manner of speaking.
With one layer of 30 medial neurons, for example, the input signals
would automatically divide into 60 separate signals, 30 from each of the
two input neurons to the medial neurons, via their synaptic connections.
The medial neurons would then add together the two signals that each
receives, apply its sigmoidal function, then send the signal to both out-
put neurons through other synaptic connections. The output neurons
simply add up the signals they receive. These are the two cartesian coor-
dinates desired.
Some readers may be puzzled by the appearance in Figure 5.1 of
three input neurons instead of two. The third neuron contains no coordi-
nate information but provides, instead, a constant value that interme-
diate neurons may add to their other two inputs. The extra number gives
the network an additional degree of freedom to shift signals by a con-
stant or to avoid the unpleasant effects of zero inputs.
How do you educate a neural net? By giving it a lot of examples.
it is not the neurons that you educate, but the synaptic con-
In truth,
nections between them. Two arrays, called synone and syntzvo, contain
all the synaptic weights. The first array consists of the synapses between
the input neurons and the medial ones. The second array consists of the
second layer of synapses, those connecting the medial neurons to the
output neurons. The net 'learns'
' 9
when it changes the synaptic weights
in these arrays as a result of its experience with the coordinate pairs that
it encounters.
Suppose that one of the training examples involves the conversion
of the polar coordinates (15.7, 110°) to their cartesian equivalents,
(—5.37, 14.75). Normally, one would apply the standard sine and cosine
formulas to transform the first set of coordinates into the second. But the
network will split, multiply, sum, and manipulate the coordinates, re-
combining all the numbers at the output end. If the result does not match
the desired cartesian coordinates, the network measures the error in
each.
For example, if the neural network produces (—2.41, 10.82) instead
of (—5.37, 14.75), it will develop two error differences, el and e2, be-
tween the target coordinates and the computed ones: The individual dif-
ferences, —2.96 and +3.93, form the basis for adjustments in the synap-
tic weights all the way through the net, from back to front. The method is
called back-propagation.
In the Rietman-Frye network, the method first calculates how the
weights in the second set of synapses must be changed to reduce the
error if same conversion were to be attempted again. To adjust
the
the synaptic connection between the ith medial neuron and the jth out-
put neuron, for example, POLARNET changes syntwo(i, j) by adding to
it the product of the jth error and the previous output of the ith medial
neuron. Thus,if the jth error is 3.93 and the previous output of the ith
medial neuron was 8.82, the back-propagation method will add the
product
3.93*8.82 = 34.66
to the value of syntwo(i, j), revising it upward by this amount. If the same
coordinate pair were re-submitted to the network revised only in this
one synapse, the new contribution by the synapse to the final sum for
the jth neuron would be 34.66 higher than it was before. If the earlier
sum was 10.82, it would now become 45.48, considerably higher, even,
than the target sum of 14.75.
To avoid problems of overshoot, an additional multiplier must enter
the adjustment formula. A parameter called rate, usually with a value
somewhere in the neighborhood of 0.1, modifies the adjustment to a
kinder and gentler level, say 3.47. Thus, the new value of the jth output
neuron would, in the case of this single adjustment, come to 14.29, very
close (accidentally) to the target value.
Back propagation next alters the values of the first set of weights,
contained in the array synone, by essentially the same method. First,
however, the derivative of the sigmoid function must be applied to the
back-propagated sum of the adjusted synapse values in syntivo. This step

ensures that the error information goes through the same step in reverse
that the processed information does in forward gear.
For each medial neuron, the back-propagation procedure forms the
product between each of the two error terms and their corresponding
synapse values for that particular medial neuron. It adds the two prod-
ucts together and then, pretending that this was the output of the medial
neuron in question, computes the corresponding input error by using the
derivative of the sigmoidal function, in this case, y(l—y), where y is the
signal to be back-propagated.
The weights stored in synone can now be changed in the final phase
of back propagation. The actual adjustment, called delta, is the product
of four numbers: The rate as described earlier, the reconstructed output
as described in the paragraph above, the inverted form of this output,
and the input to the neuron at the front end of the synaptic connection.
The actual formula appears in Figure 5.2.
adjusting the second synaptic layer
for i «- 1 to 2
for j <— 1 to n
syntwo(j, i) <— syntwo(j, i) + rate*medout(j)*error(i)
inverting the sigmoidal signal
for i «— 1 to n
sigma(i) <— 0
for j +- 1 to 2
sigma(i) <— sigma(i) 4- error(j)*syntwo(i, j)
sigmoid(i) «— medin(i)*(l-medin(i)) / add error check
adjusting the first synaptic layer
for i «- 1 to 2
for j
<— 1 to n
delta <— rate*sigmoid(j)*sigma(j)*input(i)
synone(i, j) <— synone(i, j) + delta
Figure 5.2 Back-propagation algorithm.

r
Of course, one cycle of synapse changes does not automatically

make the neural network a perfect coordinate converter. But many such
changes will cause a slow and steady improvement.
The standard combines the two individual error terms (the
error, E,
square root of the sum of the squares of el and e2) to measure the net-
work's accuracy at any given time. Figure 5.3 shows how E changed in
an automated experiment involving 50,000 randomly selected coordi-
nate conversion examples, each involving a point within one unit dis-
tance of the origin.
The network had 40 medial neurons and a rather modest learning
rate of 0.01. The learning curve, in spite of frequent fluctuations, shows
a determined downward trend from 25 percent down to about 4 percent.
It reaches this level of competence shortly after 30,000 trials and does
not improve much thereafter.

Networks with two medial layers may do better. For example, Riet-
man and Frye tested a network having two medial layers of 15 neurons
t rr" 1 1
I
1 1 i 1
I
i 1 1 ————————
I
i i i i I i i
* m *
J 1 1 1 1 1 I I I I I I I I I I I I I I I I L__L
2 3 4 5
4
Number of learning trials x 10
Figure 5.3 Learning curve of the polar converter.

each. When they tested this network on 50,000 conversion examples, it

converged to a competence level of only 1.5 percent error, even by the
25,000th test.
Strangely enough, in the world of neural nets, bigger is not necessar-
ily better. When Rietman and Frye increased the number of neurons in
each medial layer from 15 to 150, the network barely managed a 2 per-
cent error rate!
The foregoing description of the conversion network is just com-
plete enough that readers with a modicum of programming experience
will be able to write POLARNET without the benefit of further advice.
To confirm one detail or another, they only need to consult the two algo-
rithm boxes (Figures 5.2 and 5.4). Others will want to know what the
symbols and variables in these algorithms mean. As usual, the for-loop
stands for the standard iteration loop of which every programming lan-
guage has some version. The left-pointing arrow means the assignment
statement indicated by a " = " or ":=" in most languages.
The algorithm
in Figure 5.4 represents the normal operation of PO-
LARNET. Before the program is run on a single example, however, all
the synapses must be initialized by setting them to a random number be-
tween 0 and 0.1.
The first double for-loop achieves this by running through all possi-
ble combinations of one of the n medial neurons with one of the 3 input
neurons. For each combination, it sets the corresponding value of syn-
one to the product of 0.1 and a random number. Most computer systems
produce random numbers between 0 and 1, so the product of such a ran-
dom number with 0.1 will be a random number that is only one-tenth as
large. The second double for-loop does essentially the same thing for the
array syntwo.
Such small and variable values for the synapse weights are crucial to
getting thenetwork running. If the weights were all the same to begin
with, they would stay the same throughout training and the network
would never change its behavior in a meaningful way. Variety is the
spice of learning.
"Normal operation" refers to the next two double for-loops. The
wave of computation passes through neural nets of this general type one
layer at a time. The first double loop calculates inputs to the medial layer
by adding up the inputs medin to each, as weighted by the entries in syn-
one. The second double loop does the same thing for the output neurons,
adding up the weighted medout outputs. At the end of the second double
loop, POLARNET computes the target values for the cartesian coordi-
initial synapse settings
for i «— 1 to n / n: number of medial neurons

for j
<— 1 to 3 / 3: number of input neurons
synone(j, i) 0.1*random
for i «- 1 to 2 / 2: number of output neurons

for j
<— 1 to n
syntwo(j, i <— 0.1*random
coordinate conversion
input input(l), input(2)

for i <— 1 to n
medin(i) <— 0 / medial input
for j «- 1 to 3
(medin(i) <— medin(i) + synone(j, i)*input(j)
medout(i) <— l/(l-exp(-medin(i))) / add error check
for i<— 1 to 2
output(i) «- 0
for j <— 1 to n
output(i) <— output(i) + syntwo(j, i)*medout(j)
compute target(i) / using input(l) and input(2)
error(i) <— target(i) — output(i)
Figure 5.4 Normal operation.
natesby using the standard sine and cosine formulas mentioned earlier.
Readers may easily expand this macro-step into the appropriate code.
Thus, target ( 1) is the cartesian x-coordinate and target(2) is the cartesian
y-coordinate of the point the user inputs at the head of the conversion
procedure. The two error terms, error( 1 ) and error(2), play a crucial role
in the next procedure.
The algorithm in Figure 5.2 is the back-propagation algorithm as ex-
plained earlier in the paragraph on educating the synapses. The explana-
tion is already complete enough that readers who compare each section
of the algorithm with the corresponding description should have no dif-
ficulty producing working code. But programmers are hereby warned of
a special problem that may occur when POLARNET computes the sig-
moidal function during forward propagation or its derivative during
back propagation. Extremely small or extremely large numbers can pro-
duce overflow errors. Specifically, during forward propagation, the pro-
gram must test medin(i) to ensure that exponentiation does not exceed
the available precision. It must also check that the denominator of the
sigmoidal expression is not so close to zero that division will also cause
overflow. The latter test must also be made during back propagation
when the derivative of the sigmoidal function is being generated.
A version of POLARNET that simply sets up initial synaptic
weights, performs one wave of computation, then one wave of back
propagation, would be of little use. But if the blocks following the initia-
lization steps are placed inside a loop, readers may enjoy the privilege of
typing in pair after pair of co-ordinates, watching as the network be-
comes more and more accurate.
Unfortunately, it takes thousands of pairs of polar coordinates to
train the net adequately. For this reason, the same loop should be made
automatic according to the following little recipe in which all steps but
the second refer to blocks within the two main algorithms:
initialization
for cnt <— 1 to 10,000
select random point
convert coordinates
back propagation
output error E
To generate random points (r, a) from within the unit circle, first gen-
erate a standard random number (between 0 and 1) for the radius r. Then
generate a second random number and convert it to an angle a, multiply-
ing by 360 (for degrees) or 2n (for radians). The resulting points will not
be uniformly distributed in the unit circle, but will tend to cluster around
harm the experiment but purists may want
the origin. This bias will not
to remove it by applying a correction function to the first coordinate, r.
Outputting the error E could mean anything from printing the value
to plotting it on the screen. A plot of error values such as those in the
learning curve in Figure 5.3 not onlymakes the program more informa-
tive and absorbing to watch, it becomes an indispensible tool for moni-
toring the progress of the network in many experimental situations.
Rietman and Frye have developed an important visualization tool

that some readers might wish to emulate. At each stage in the learning
process, one can stop the experiments and examine visually how well
the network, as currently configured, performs conversions. Why not
convert the whole unit circle at once, then plot the result? If the polar
grid of rays and concentric circles survives the conversion more or less
intact, thenetwork is performing well. But if the grid comes out
squashed or distorted in some way, the network's education will not be
complete.
Readers may alter POLARNET to taste. Not only may they change
the totalnumber of experiments (from 10,000, above), but they are also
free to change the number of input neurons (3), output neurons (2) or
medial neurons (n) to any number they like. They may even install two
or more medial layers to see how much difference they make to the
speed of learning. In such a case, however, both the forward compution
and backward propagation procedures will be in for something of an
overhaul.
The ability to alter the basic size and shape of the network raises a
number of interesting questions that readers might like to investigate:
First, how does the size of the rate parameter affect the network's learn-
ing ability? With higher and higher values of rate, learning may converge
more quickly but speed has it risks: learning may suddenly diverge
wildly! Another interesting question concerns the effect of the initial
synaptic weights on the network's ability to learn quickly. Are there
some choices for initial synaptic values that throw the neurons off-track?
Or ones that bring the network up to speed sooner? The question is
worth asking because neural nets usually amount to what computer sci-
entists call hill-climbing routines. The initial direction specified by the
first synaptic values will determine how fast the network finds the opti-
mum value (or hill).

How much does a second layer of medial neurons improve perform-
ance of the learning network? A third layer? Do diminishing returns
eventually set in?
What effect does the choice of training examples have on the net-
work's performance? Perhaps, as a practical matter, it would be best to
have specific training sets of coordinate pairs that would bring the net-
work up to its optimum performance levels after only a few hundred
trials, not many thousand! Grid points might work very well.
Although it cannot be guaranteed that neural netwoks will converge

to useful behavior in every potential application, many problems have
yet to be investigated in this respect. Suddenly, recreational computing

looks more and more like original research. Here is one suggestion.
Can neural nets learn to do arithmetic? Using exactly the same ar-
chitecture as the coordinate conversion network, readers can investi-
gate the ultimate competence of networks at addition and multiplica-
tion. In both cases, the networks will have two inputs as before, but they
will have only one output. The only other modification involves one
small part of the program POLARNET: The error terms will involve the
difference between the real sum (or product) and the one computed by
the net. Does the network learn addition more quickly than multiplica-
tion? How accurate does it ever get at either task?
Neural networks have found a small niche in our software toolkit.
How much larger that niche becomes will depend, ultimately, on the
range of tasks that they can reliably learn to perform.
One can be mildly pessimistic that layered networks will find appli-
cations in any but the simplest problems. If the solution space has more
than one local optimum (hill) for example, the network may well go
rushing up the wrong optimum. But if it's just a question of time, parallel
computers with neural elements will make solvable problems increas-
ingly practical. In the meantime, thanks to Rietman and Frye, readers
find themselves once again at the forefront —ready to tackle the latest
computational paradigm for themselves.
Further Reading
E. A. Rietman. Explorations in Parallel Processing. Tab Books, 1990.
D. E. Rumelhart and J. L. McClelland, eds. Parallel Processing: Explorations in
the Microstructure of Cognition, Vol. 1. MIT Press, 1986.
m
(Termites) . . .forge the complexion of a landscape

like no other organism except man.
Walter Linsenmaier, Insects of the World
A
JL Ajiyone who has ever seen a termite mound must have been
pressed by the complex patterns of tunnels built by the industrious but
im-
mindless insects. Paradoxically, artificial forms of life that make ter-

mites look like geniuses in other departments can produce equally as-
tounding creations. Take tur-mites, for example. They are squarish, cy-
bernetic creatures that have the most rudimentary of brains. And yet as
they move about on the infinite plane on which they live, they trace out
strange patterns that appear to reflect an underlying intelligent design.
The tur-mites were inspired in part by Greg Turk while a graduate
student at the University of North Carolina at Chapel Hill. In 1988 Turk
firstexperimented with a special type of Turing machine, a basic model
of computation. A Turing machine is usually assumed to operate on an
infinite linear tape that is divided into cells. Turk, however, studied
Turking machines that operate on a kind of two-dimensional tape —
essentially the same plane on which the tur-mites roam. Converting a
two-dimensional Turing machine into a tur-mite is simple and painless:
abstract rules are replaced straightforwardly by a neural network. Such a
conversion highlights an important theme in the theory of computation:
one computational scheme often turns out to be equivalent to another,
seemingly unrelated, one.
Turing machines are named after the British mathematician Alan M.
Turing, who first proposed them as a way to define computation. In ef-
fect, a Turing machine is the ultimate digital computing machine. It can
57
compute anything that a modern computer can — as long as it is given

enough time.
One can visualize a Turing machine as it is shown in Figure 6.1: a
black box equipped with a device that reads a symbol in a single cell of an
infinitely long tape, writes a new symbol in the cell and moves the tape
either forward or backward in order to examine the symbol
in an adja-
cent cell. What is inside the black box?
does not
It really matter, as long
as the box adheres strictly to a given table that lists what the Turing ma-
chine must do for every symbol read and for every one of the machine's
possible "states." These may change with each cycle of operation. A
cycle consists of the following three steps:
1. Read the symbol currently under the read/write device.

2. Look up the table entry given by the machine's current state and
the symbol just read.
3. Write the symbol given by the table entry, move the tape in the
direction indicated and enter the state shown.
Each table entry therefore has three parts: a symbol to be written on the
current cell, a direction in which to move the tape and a state to enter.
To a Turing machine the tape's motion is relative. One could just as

easily arrange for the tape to remain fixed and the machine to move itself
from cell to cell. In fact, once one contemplates the idea of moving the
Figure 6. 1 Standard visualization of a Turing machine.

Turing machine and not its tape, it does not take much imagination to
envision a two-dimensional "tape" on which the machine may move
about freely in not one but two independent directions.
Regardless of whether it has a one- or two-dimensional tape, a Tur-
ing machine's table is what ultimately determines its behavior. It is
closely analogous to the program that controls a modern digital com-
puter. In terms of computational capability, two-dimensional Turing
machines are not more powerful than one-dimensional ones. They just
have more interesting patterns of movement over the cells. The pattern
shown in Figure 6.2, for example, was made by a single-state two-
dimensional Turing machine. Its internal table is:
BLACK RED
(RED, LEFT, A) (BLACK, RIGHT, A)
The machine's single state has been labeled A.

A slightly more complicated two-dimensional Turing machine that
was discovered by Turk has two states, designated A and B, and it fol-
lows this internal table:
Figure 6.2 One of Turk's patterns.

BLACK GREEN
(GREEN, LEFT, A) (BLACK FORWARD, B)
(GREEN, RIGHT, A) (GREEN, RIGHT, A)
According to Turk, a two-dimensional Turing machine programmed

with this table produces a marvelous spiral pattern. The machine creates
"larger and larger patterned regions that are placed in orderly fashion
around the starting point."
Any pattern generated by a two-dimensional Turing machine can be
reproduced exactly by a tur-mite. A tur-mite's behavior, however, is not
controlled by a mysterious black box. It is controlled by what can be
loosely described as a brain. The fact that one can dissect and examine a
tur-mite brain what makes the creature so fascinating.
is
Judged only by its appearance and ethology, a tur-mite certainly is

not fascinating (see Figure 6.3). Its body is roughly square, so that it fits
snugly into the squares that divide the infinite plane on which it lives. It
has a flat bottom equipped with some form of locomotory apparatus. (I
Figure 6.3 A tur-mite occupies one square at a time.

do not know what makes a tur-mite go, since I have never turned one
over.) The apparatus enables the creature to rotate and to move exactly
one square in the direction in which it happens to be facing. Actually, a
turmite's face has no purpose except to let us know which way is for-
ward; its "eyes" do not function. When a tur-mite changes direction, it
merely swivels 90 degrees on its current square before moving to a new
one.
Initially all the squares on the plane, including the one the tur-mite
occupies, are black. Before it moves, however, a tur-mite may change
the color of the square it currently occupies. (The tur-mite's color-chang-
ing organ is locomotory apparatus.) A tur-mite that
as mysterious as its
duplicates the pattern shown in Figure 6.2, for example, must be capable
of painting the square one of two colors (in this case, red or black). To
produce the pattern shown in Figure 6.4, however, a tur-mite has to have
more colors at its disposal.
How does a tur-mite know when to move or when to change the
color of its square? Those actions are controlled by its brain, which con-
sists of a collection of "neurodes," simplified versions of the neurons in
our own brain. A neurode receives signals along fibers that originate at
sensors (which are found on a tur-mite' s underside) or at other neurodes
and sends signals along fibers to effectors (such as the tur-mite's loco-
motory apparatus or its color-changing organ) or to other neurodes.
A neurode fires (sends a signal along its output fiber) if the number of
incoming signals equals or exceeds the neurode's threshold, which is
given by the number written on the neurode. Otherwise, it does not fire.
Because time in the tur-mite' s world proceeds in discrete steps, all excit-
atory and inhibitory signals are sent or received in discrete steps as well.
To illustrate how a tur-mite actually makes a decision, I shall dissect
the brain of two specimens and middle parts of Figure 6.5), both of
(left
which produce exactly the same pattern as that generated by the single-
state two-dimensional Turing machine described earlier. The brain on
the left contains two neurodes that are not connected to each other. Each
neurode has just one input fiber and one output fiber. When the tur-
mite's color sensor detects red, it sends a single signal to the left neurode,
causing the neurode to fire. The neurode 's output two fiber splits into
parts,one going to the color effector (which then colors the entire
square) and the other going to the locomotory apparatus (which then
swivels the creature 90 degrees to the right and advances it one square in
the new direction). On the other hand, when the tur-mite's color sensor
detects black, it sends a signal to the right neurode, causing it to fire. The
Figure 6.4 Multicolored tur-mite pattern.
BLACK RIGHT RED LEFT RED LEFT BLACK LEFT FORWARD RIGHT GREEN
Figure 6.5 Three tur-mite brains, two of which (left and center) do the
same job.
neurode's output, in turn, causes the tur-mite to paint the square red be-
fore turning and heading to the square on its left.
In short, when the tur-mite finds itself on a red square, it colors the
square black and then moves one square to the right. And when the tur-
mite occupies a black square, it changes the color of the square to red,
then turns left and advances one square in its new direction.
The second tur-mite brain is more complicated, but it does exactly
the same job as the first. It was derived by a method I shall presently de-
scribe. The two neurodes both have threshold 2; neither will fire unless
it receives two input signals during the same time increment. Once the
brain is set in motion, one neurode will always fire at each step in time.
The simple behavior embodied in the two neurode circuits just de-
scribed results in the complicated image in Figure 6.2: a red cloud of tiny
squares from which an intricate structure extends straight to infinity.
What causes that sudden sense of purpose in the tur-mite after what
seems a great deal of pointless meandering? The answer has to do with
the pattern of colored squares in the cloud. At a certain point, part of that
J
pattern, in combination with the tur~mite s neurode-based rules, locks
the creature into a repetitive sequence of moves that weaves the struc-
ture. (I wonder any readers can discover the triggering pattern.)
if
Life is like that for tur-mites. Sometimes a seemingly random mean-

dering turns into an almost deadly determinism. Of course, the appear-
ance of randomness is purely illusory. All tur-mites are decidedly deter-
ministic at all times.
Nonetheless, there are mysteries to be found in the tur-mite' s world.
Consider, for example, the pattern shown in Figure 6.4. The tur-mite
that made that pattern is outfitted with four effectors that change the
color of a square to black, red, yellow or green. It abides by the following
rules:
Square color Action
black paint red, turn right

red paint yellow, turn right
yellow paint green, turn left
green paint black, turn left
This tur-mite also has a very simple brain. It consists of four neur-
odes that are not interconnected. Each neurode executes one of the four
behavioral rules in the manner of the first tur-mite's simplest brain. Turk
is puzzled by the fact that this particular tur-mite produces a pattern
having bilateral symmetry. Readers can explore different versions of

this machine in which the move sequence (right, right, left, left) is ro-
tated to (right, left, left, right) and beyond.
How exactly does one get a tur-mite from a particular two-dimen-
sional Turing machine? The technique is actually quite simple. One
merely replaces each entry of the machine's internal table with a thresh-
olds neurode that receives input signals from a sensor for the color cor-
responding to the entry's column and perhaps from other neurodes as
well. Each neurode's output fibers go to the effectors necessary to exe-
cute the moves and color changes listed in the table entry.
For example, suppose that a certain neurode corresponds to a table
entry in a column labeled "red" and a row labeled "B." According to the
conversion scheme, the neurode would have an input fiber from the
sensor that detects red. If the table entry happened to be (black, left, B),
then the neurode would send an output fiber to the effector that colors
the occupied square black and to the effector that enables the tur-mite to
execute left turns.
The various states of a particular Turing machine are realized by the

connections between neurodes in a tur-mite brain. Because the table
entry in the example requires the tur-mite to adopt state B, the neurode
representing that entry would extend output fibers to each of the neur-
odes making up row B of the table.
In this context, such a neural network is nothing more (and nothing
less) than a form of hardware embodying a behavioral table. A sample
conversion is displayed schematically in the right-hand part of Figure
6.5. It shows how one would go about constructing the brain of the tur-
mite that mimics the behavior of Turk's spiraling two-dimensional
Turking machine.
It is fun to watch a tur-mite (or a two-dimensional Turing machine)
wander about on a cellular plane. To follow the action, however, the

reader must write a program that simulates the tur-mite 's movements.
How does one go from a table to a program?
Luckily, the process is nearly as simple as designing a tur-mite 's
brain. A program that I call TURMITE consults a Turing-machine table
in the form of three separate arrays: color, motion and state. Each array is
indexed by two variables, c and s. The variable c indexes the color of the
present square, and the variable s indexes the Turing machine's (or the
equivalent tur-mite's) current state. Because the indexes have to be as-
signed integer values, the colors and states used in the simulation must
be numbered.
For example, the colors black and green can be assigned to the vari-
able c by the numbers 1 and 2, respectively. Similarly, the states A and B
can be designated respectively by the values 1 and 2 of the variable s. In
this case, a simulation of a spiraling tur-mite would require the following
arrays:
: :
1 2 1 2
2 1 1 2
2 2 1 1
COi,OR ST/
Directions of motion must also be coded in terms of numbers. Hence,

forward, backward, left and right could be indicated respectively by the
numbers 1, 2, 3 and 4, which are contained in the array motion.

The main purpose of TURMITE is to color small squares (perhaps in-
dividual pixels) on the computer's display screen that highlight a tur-
mite 's peripatetics. The program keeps track of the displayed squares'
colors in a two-dimensional array called pattern. Initially only one
square is lit —
the one lying at the center of the screen.
The value of c at any given time is provided by the entry in the array
pattern corresponding to the turmite's current coordinates on the
screen, say i and ;. With c and s in hand, the program simply looks up the
array entries color(c,s), motion(c,s) and state{c,s).
The program changes the color encoded in the entry in pattern(i,j)
and then alters either i or/, depending on the value of motion{c,s). Here
the program must translate the relative movement encoded in motion
into an absolute movement by consulting another variable, dir, which
contains the last direction moved: up, down, left or right. The final step
in TURMITE 's operating cycle consists merely of changing s to the num-
ber given by state(c,s). The rest can be left to the imagination and inven-
tive skill of those readers who own programs.
like to write their
Readers might even want to continue a new investigation begun by
Norwegian computer scientist Odd Arild Olsen. Since Turk's initial tur-
mite experiments in 1988, the creatures have enjoyed a certain endemic
popularity with programmers who treasure behavioral exotica, not to
mention those who probe the relationship between computation and
"behavior," broadly defined. Olsen, who places not one but two tur-
mites on the plane at a time, reports amazing relationships developed by
a
d
aaa a a
DO
d aaa a a
aaa a aa
a aaa a a
aaa a aa
a aaa a a
aaa a aa
a aaa a a
aaa a aa
a aaa a a
aaa a aa
a aaa a a
aaa a aa
a aaa a a
a a aa
m aaa a a
mmm a aa
m
in a
a
aaa a a aa
mmm a aa at
mma
mm a m aam m
m ma aa m m amm
mm m aaa a mm
o aa a
am a
mmm m
mm
mm m
mm
mm
m mmmt
m iiii a
Figure 6.6 Two tur-mites build a nest and a burrow.
the pair. In particular he has systematically explored the behavior of the

one-state tur-mite (described at the beginning of this chapter) when
paired with itself.
The screen shown in Figure 6.6 captures just one of 1,681 behaviors
Olsen has examined. In this case the tur-mites "cooperated" to build a
"nest." Then one of them, perhaps growing restless, built a complicated
burrow, seemingly toward infinity. The second tur-mite found the bur-
row, followed it to the end, and forced the first tur-mite to follow the
burrow back to the nest. The second tur-mite also worked its way back,
slowly recoloring all the squares! The 1,681 experiments performed by
Olsen amounted to nothing more complicated than placing the first tur-
mite at the origin of the grid, then placing the second tur-mite at one of
—
the 1,681 grid locations with x- and y-coordinates ranging from — 20 to

+ 20. In all experiments, both tur-mites began with the same orienta-
tion. Olsen suggests that readers might explore what happens when tur-
mites of different species are put on the plane.
Whether one is constructing a tur-mite's brain or simulating its be-
havior on a computer, it is on the fact that, since
interesting to reflect
tur-mites can carry out any computation a Turing machine is capable of
executing, tur-mites can be just as powerful as some computers. If, as
some claim, the human brain amounts to nothing more than a kind of
digital computer, then some tur-mites could be just as smart as we are
if not smarter!
Further Reading
Martin Gardner. "Mathematical Games." Scientific American, Vol. 216, No. 3
(March 1967), pp. 124-129.
Martin Gardner. "Mathematical Games." Scientific American, Vol. 229, No. 5
(November 1973), pp. 116-123.
THEME TWO
Matter Misbehaves
To say that matter misbehaves may serve to

catch the reader's eye but it does a disservice
to matter itself. Matter does not break its own

laws the way humans break theirs. Since the
arrival of computers, we have learned to
program the laws we know and to simulate
the behavior of matter in many of its
manifestations.
The science of simulation enables us to
predict weather (up to a point), to design new
structures, and to probe the details of
molecular interactions, all without recourse to
the laboratory. But the magic of this science
has been pulled up short by the discovery that
some physical systems will evade the power of
any computer to predict. Chaos has arrived.
The theme opens mildly enough with two
computer games, each of them involving a
dynamical system. A golf ball, gently tapped
by an abstract putter, rolls across the green
and narrowly misses the hole. A spaceship
changes its orbit around the sun when a
computer programmer types in a new thrust
factor. Here, students of the new science learn
how to turn Newtonian dynamics into
behavior on the screen. But chaos lurks in the
background already.
The next dynamical system, Lorenz's water
wheel, captures all the dynamics of his famous
weather-in-a-jar model, historically one of the
first models of chaos. It turns out that the
simplest weather system of all, air circulating

in a heated jar, exhibits the exquisite
sensitivity to initial conditions that we call
chaos. But it's easier to see buckets revolving

chaotically on a water wheel than to spot
invisible air in a jar. A peek at the system's
attractor immediately reveals its fractal nature,
the Lorenz attractor.
Lyapunov space maps the behavior of a
choatic system. For every combination of
parameter values that determine the system's
behavior, the color of this space reflects whether
chaos is present or absent. The ultra-simple
logisticsystem gives rise to a space filled with
strange, fractal swallows and fantastic fairy
bridges over alien cities. The Lyapunov exponent
not only enables probers of systems to detect
chaos, it enables them to map all the possibilities
at once.
One of the most notable advances in the
science of fractals, the iterated function systems
of Michael F. Barnsley, enable us to custom-
design our own fractals, from infinite fern leaves
to smoky clouds. The principles, spelled out as
an innocent game of fractal tennis, become
specific programming advice in the very next
chapter.
The foregoing chapters develop the theme of
misbehaving matter in that special order that
clarifies among three key
the relationships
subjects. First come dynamical systems, then
chaos, then fractals. Readers who understand
dynamical systems will have an easier time
appreciating what chaos really means. At the
same time chaos itself has a shape. The real
significance of fractals lies not in the happy
shapes of puffy clouds but in the nightmarish
forms that lurk in the "phase-space" of their
dynamical systems. The first step in dealing
effectively with chaos is to inspect it in the
glaring light of the computer microscope.
I have once, had the distinction 'of making a
it is true,
hole in That is to say, after had hit, a ball
one'. . . . I
was found in the can, and my ball was not found. It is

what we call circumstantial evidence the same —
thing that people are hanged for.
Steven LEACOCK, "Mathematics for Golfers," in The

World of Mathematics, Volume 4
TJL he ordinary golfer deploying an armory of clubs tries desperately

to drive a small ball into a distant hole hidden in several acres of grass
and guarded by ponds, trees, and sand traps. Players of miniature golf
use but one club on a much smaller obstacle course of bumps, tunnels,
rotating blades and swinging pendulums. Now players of micro-minia-
ture golf, without swinging a club at all, can attempt a hole in one on the
screen of a microcomputer. There are hazards here, too, including some
that violate the laws of physics.
Although the microgolfer is confined to an armchair, he or she can
putt with an electronic club and even enjoy a dimension not available to
players of the larger games: programming. How many of us are able to
design, build, and play a course all in the same day? Not even Jack
Nicklaus —unless he knows how to program. In fact Nicklaus would not
even have to be a professional programmer to set up a microgolf course.
I shall lay down plans for three holes: one for beginners and two for
more advanced players. At the end of the day everyone who plays by the
rules will have made a microgolf course on a par with his or her talents.
The project is illustrated in its simplest form in Figure 7.1. Here the
program I call hole in one has displayed a single hole on a rectangular
green. A putter appears as a short line segment behind a tiny ball at one
71
72 Theme Two Matter Misbehaves
ENTER SPEED AND AIM? 8,-20

TYPE 0 TO PUTT BALL? 0
GREAT SHOT!
Figure 7.1 A successful putt on the hole in one green.
end of the green. A player angles and positions the putter to aim and
strike the ball toward its target, the cup at the other end of the green.
The simple hole in one version of microgolf is all hit or miss. If the
ballgoes past the cup, it will cross over the course's edge as though it
were not there, disappearing off the screen and rolling into the com-
puter's memory, never to reappear. Even this version of the game has a
certain pleasurable tension to it.
It is actually possible for an amateur to prepare hole in one as a kind

of software springboard to the more advanced versions. Fragments of
the program hole in one can be inserted into the programs I call birdie
and eagle, which are depicted in Figures 7.2 and 7.3, respectively.
birdie features a hazard near the cup, a circular twilight zone of
sorts. If the ball enters this zone, it changes direction and speed in a com-
pletely unpredictable manner. More treacherous hazards plague the
eagle green, but I will withhold the horrifying details for now.
7 Micro-Miniature Golf 73
Figure 7.2 A
twilight zone and a demon ball make life difficult for the in-
termediate putter.
In this chapter have provided a more detailed program description

I
for those hesitant or beginning programmers who need just a bit of extra
information to get started. The following description of hole in one
adopts the tried-and-true method of starting with a clear description of
the computation to be performed, usually given in steps as an algorithm.
From there it moves, as fast as possible, to actual fragments of program.
A reader who puts all the pieces together is just a few keystrokes away
from the Micro-Miniature Open!
hole in one first displays the cup ready for play, and then it requests
that a player adjust the putter and putt, hole in one then draws and re-
draws the ball as it rolls across the electronic turf, perhaps into the hole.
From these specifications a crude algorithm for hole in one can be
devised: draw the layout, prompt the user to putt, move the ball many
times (to animate the direction and speed of the putt) and decide each
time whether the ball has fallen into the cup. The algorithm can then be
refined to create the program hole in one. To begin with, the initial
Figure 7.3 Three hazards on the advanced green.
layout can be displayed in a few steps that draw the green, cup, ball, and
putter. These steps can be programmed easily.
In order to be helpfully explicit, I shall assume that the reader is writ-
ing a program in the basic language on an IBM PC or a PC-compatible
computer. (Do not be discouraged if you lack the hardware. The pro-
gram is easily adapted to other computer systems.) To be even more ex-
plicit, I shall pretend that all readers have a screen at least 300 pixels
(screen dots) wide and 200 pixels high. All distances on the screen must
be measured from the origin, which is the top left corner of the IBM
screen.
On a 300-by-200 pixel screen it is quite reasonable to draw a rectan-
gular green 240 pixels long and 160 pixels high. To center the layout
more or less on the screen, hole in one places the green 30 pixels from
the left side of the screen and 20 pixels from the top. In short, the hori-
zontal coordinates of the green run from 30 to 270 and its vertical coor-
dinates run from 20 to 180. The green takes shape from the following in-
structions:
10 SCREEN 1
20 LINE (30,20) - (270,20)

30 LINE (30, 180) -(270, 180)
40 LINE (30,20) -(30,180)
50 LINE (270,20) -(270,180)
Note by numbering basic statements with increasing (but not nec-

that
essarily consecutive) values, theprogrammer designates the exact order
in which the computer should execute those statements. The first com-
mand, CLS, clears the screen of any images that may previously have
been drawn on it. This feature is employed when one wishes to restart
the program.
hole in one next represents the cup as a circle at the far right end of
the rectangle.The center of the cup is 240 pixels from the left and 100
pixels from the top. The BASIC instruction that will draw the cup is there-
fore
60 CIRCLE (240,100),5,1
The centered on the coordinates (240,100) and has a radius of

circle is
five pixels. The "1" at the end of the line specifies the color white.
Just to keep the shot from being boring, hole in one places the ball
at a random location along the tee-off line, which is 30 pixels from the
left end of the green. Because the ball's position changes throughout the
program, the variables XBALL and YBALL are created to keep track of its
coordinates
70 XBALL = 60
80 YBALL = 160*RND + 20
90 CIRCLE (XBALL, YBALL),4,1
Line 70 restricts the

ball's x coordinate to coincide with the tee-off line.
Line 80 selects a random number between zero and one called RND,
scales it up to the number between zero and 160 and then adds
size of a
20 to the result. This means that YBALL will lie somewhere along the
tee-off line. Line 90 draws the ball with its center at (XBALL, YBALL).
The ball has a radius of four, a snug fit in the radius-five hole.
A line just behind the ball represents the putter, or more accurately
the blade of the putter. Its resting state is given by the following:
100 LINE (54,YBALL + 8) - (54,YBALL - 8)
This vertical line is tangent to the left side of the ball and is 16 pixels long.
The next step of the hole in one algorithm — to 'prompt the user to
'
putt" — must be refined programming continues. "To

just a little before
prompt" means not only to send a message that asks a player to putt but
also to acquire information about the orientation of the putter and the
speed of the swing. It also means that hole in one must redraw the put-
ter in its new position so that a player can judge the shot by eye.
Two lines of basic are added to hole in one, both to print a message
and to accept input.
110 PRINT "ENTER SPEED AND AIM"

120 INPUT SPEED, AIM
The variable called SPEED stores the distance in pixels that the ball
moves every time HOLE IN ONE updates its position in the course of a
putt.
The variable AIM stores the ball's direction of motion. Many players
will finddegrees to be the natural unit for entering an angle for AIM.
hole in one therefore accepts an angle between + 90 and — 90 de-
grees as measured from a horizontal line. The extremes represent
strokes that send the ball straight up or down.
Unfortunately, most versions of basic handle angular measure-
ments in units called radians rather than in degrees. Therefore, hole in
one requires a small calculation that converts degrees into radians.
130 RADAIM = (AIM*3. 141 6)/l 80
The conversion is based on the fact that 180 degrees equals n radians.
The values of RADAIM and SPEED can be inserted into a formula
that determines the position and orientation of the putter, hole in one
requires the formula to draw the putter poised either to stroke the ball
into the cup or to whack it off the green.
The formula positions the center of the putter behind the ball along
the stroke line and at a distance proportional to the value of SPEED. The
putter should just touch the ball's surfacewhen SPEED equals zero, that
should be at least three pixels to the left of the ball's center. The x
is, it
coordinate of the putter's center is therefore discovered by subtracting

(SPEED + 3) times the cosine of RADAIM from XBALL. The y coordi-
nate will be found by adding (SPEED + 3) times the sine of RADAIM to
YBALL.
To angle the putter, hole in one requires a little more trigonometry.

Since the putter is 16 pixels long, its end points are eight pixels away
from the center. The displacement along the x axis is then eight times the
cosine of RADAIM, and along the y axis it is eight times the sine of RA-
DAIM.
These quantities, denoted by A, B, C and D, respectively, in the pro-
gram fragment below, are all computed separately on lines 140 to 170.
The actual coordinates of the putter's "top" end (XTOP,YTOP) and
"bottom" end (XBOT,YBOT) are then computed. Finally, in line 220 the
putter is drawn as a line connecting these points.
140 A = 8*COS (RADAIM)

150 B = 8*SIN (RADAIM)
160 C = (SPEED + 3)*COS (RADAIM)
170 D = (SPEED + 3)*SIN (RADAIM)
180 XTOP = XBALL - C + B
190 YTOP = YBALL + D + A
200 XBOT = XBALL - C - B
210 YBOT = YBALL + D - A
220 LINE (XTOP,YTOP) - (XBOLYBOT)
It might appear at this point that the program is ready to enter the
third phase of its operation, to produce the animation of the ball heading
toward the cup. Has anything been left out? What if a player decides that
the angle of the putter looks wrong. Surely the fallible human being
should be given a second chance. This is done by branching back to line
10 at the player's option.
230 PRINT "TYPE 0 to PUTT BALL"

240 INPUT PUTT
250 IF PUTT < > 0 THEN GOTO 10
Here, if the player types any number but zero, the program will branch
back to line 10, where it will clear the screen, redraw the green and
prompt the player again for new values of AIM and SPEED. If the player
types zero, hole in one will proceed to the final phase of its operation,
the one specified much earlier by the phrase "move the ball many times
to animate the direction and speed of the putt."
To move the ball, however, hole in one must increase the x and y
coordinates of the ball independently. To this end, the variable SPEED is
decomposed into two new variables, one called XSPEED in the x direc-
tionand one called YSPEED in the y direction.
260 XSPEED = SPEED*COS (RADAIM)

270 YSPEED = SPEED*SIN (RADAIM)
The final section of the program consists of a loop within which two
operations will be performed continually. The ball will be moved and
then checked to see whether it happens to be in the cup.
280 FOR K = 1 to 300

290 XBALL = XBALL + XSPEED
300 YBALL = YBALL + YSPEED
310 CIRCLE(XBALL,YBALL),4,1
320 IF ABS(XBALL - 240) > 4
THEN GOTO 360

330 IF ABS(YBALL - 100) > 4
THEN GOTO 360
340 XSPEED = 0
350 YSPEED = 0
360 NEXT K
The instruction on line 280 sets up the simplest kind of loop. A variable
K counts from one to 300 to ensure enough move-and-draw cycles for
the ball to make it to the cup even at a speed of one. Within the loop the
new ball coordinates are updated, and at line 310 the ball is drawn.
Lines 320 and 330 test the ball's coordinates separately to find out
whether the ball is in the cup. If it is not, the program skips down to line
360. If it is, the ball's two speed coordinates are reset to zero, effectively
freezing the ball.
Professionalsmay object to this "for" loop because the program
continues to run even when the ball drops into the cup. hole in one tests
the ball's coordinates until the count X reaches 300. To remedy this, the
amateur golf programmer might add three instructions to the end of the
program. One would test whether the speed of the ball is zero. If the test
is affirmative, a second instruction could print a pat-on-the-back mes-
sage such as "GREAT SHOT!" A final instruction, required by most

versions of basic, marks the end of the program with the word "END."
This completes hole in one. Readers who type in the program and
run it will notice a peculiar effect as they try to putt the ball at different
SPEED'S. The unusual discrete physics that appears to operate in the mi-
crogolf world of hole in one will produce a hole in one only if the speed
of the ball divides the distance to the cup evenly!
What might be considered from one point of view may be

a defect
regarded as a "feature" from another; if the physics is wild by accident,
perhaps it can also be so by design, birdie features some hazards not
seen on any grass turf. In addition, the ball bounces around the green.
To describe these golfing wonders, I will shift expositional gears to
exclude IBM PC's and basic, reverting to algorithmic language. The bir-
die course shown in Figure 7.2 is L-shaped. How does birdie ensure that
the ball stays within the borders as it rolls? Even in the strange world of
microgolf the angle of reflection must equal the angle of incidence. This
,
is not difficult to set up for the six segments of border, which may be
called walll through wall6 without worrying too much about which wall
is which. Each time the main display loop calculates a new position for
the birdie will check to see whether any of the walls has been
ball,
crossed. For example, if walll has been crossed, birdie should reflect
the ball.
When the test and its(possibly) resulting reflection have been exe-
cuted, theprogram checks walll and then wall3 and so on. An interest-
ing problem surfaces at this point. What if the ball, in "striking" one wall
and then receiving its subsequent reflection, now finds itself outside an-
other wall? Is there not a possibility that if the ball is destined to hit a wall
near a comer, it will bounce only from one wall but not from the other?
Readers who ponder the point properly will develop a successful solu-
tion.
Once the ball has done all the rebounding
it is destined to do in the
current cycle, birdie displays it and

whether the ball is in the cup
tests
just as hole in one does. What if the unlucky putter misses the cup? The
ball will continue to bounce around within the confines of the green until
it either enters the cup or falls into an endless cycle of rebounds. Some
duffers will find this a wonderful proposition; others will sneer at the
lack of realism.The game lacks that all-important factor, friction!
To allow the ball to be slowed to a stop, birdie multiplies SPEED by
some constant that is less than one, say .995, each time the program goes
through its move-and-test loop. Because SPEED decreases by a factor of
.995 each time through the loop, both XSPEED and YSPEED will also be
multiplied by this constant within the loop.
It may of course happen that the ball eventually slows to zero with-
out having dropped into the cup. In this case birdie must call a halt to the
loop, and so the loop must be not of the "for" type but of the "while"
type. The second kind of loop keeps the cycle going as long as (while)
some condition holds. In this case the condition is that SPEED be greater
than some rather small number such as 0.5. At such a point birdie sends
the user back to the interactive part of the program where he or she is
prompted for a putt.
What hazards does the hapless golfer encounter on the second hole?
One hazard is referred to as a circular twilight zone. Having entered the
zone, the ball suffers a random change in its current direction. What
happens is that birdie tests whether the ball lies within the charmed cir-
cle and, if it does, changes the angle RADAIM by a random number be-
tween — 10 and + 10. By planning a careful series of bounces, a good
putter can send the ball around the twilight zone.
The second hazard is harder to avoid. A demon ball orbits the cup at
a fixed distance. It completes an orbit for every 10 cycles of the loop. If
the player's ball happens to touch the demon ball, the player's ball reap-
pears back at the tee-off line. Most readers who have followed the course
this far will probably think of a way to manage this awe-inspiring event.
eagle, as the elaborate illustration in Figure 7.3 shows, has a more
complicated green than birdie. Largely a glorified version of birdie, it
features some advanced hazards. For example, whenever the ball tra-
verses the neck connecting one end of the layout to the other, it experi-
ences a force pulling toward the neck's center. The force is related to the
ball's distance from the centerline, as though the passage were gracefully
banked. There is also a sand trap, where the SPEED is decreased not by a
factor of .995 but by .9. Only the most careful of strokes will get one out
of the trap without sending the ball careening around the layout like a
runaway bullet. The final hazard involves a bit of "rough." Here the ball
becomes lost, disappearing before it even rolls to a stop. How to find the
ball? If you reposition the putter in the right location, the ball will appear
beside it.
Iwas inspired to take readers on a tour of microgolf greenery by way

of a game called Zany Golf, distributed by Electronic Arts in San Mateo,
California. One can buy the game, of course, but that might mean miss-
ing the fun of building one's own. Readers are by no means limited to
golf, in any case. The techniques described here lend themselves readily
to games of micropinball and electronic billiards.
Further Reading
Bruce A. Artwick. Microcomputer Displays, Graphics, and Animation. Prentice
Hall, 1984.
" -7
\3
,
s Til KthiIII!ICS
in the early 1970's I often worked late at my university office in the

hope of avoiding the interruptions of students. Unfortunately, just down
the corridor in the computer-graphics research laboratory was the
nightly gathering for a favorite student game variously called Star Trek
or Space War. "Get him! Get him!" came the cries that echoed through
my closed office door. "Watch out for the missiles!" The noise was just
enough to disrupt my train of thought. Unable to beat Star Trek, I would
usually join it. With a resolve to return to less frivolous matters as soon
as the stretch had refocused my attention, I would stroll down to the lab
to watch the action.
The game Star Trek is loosely based on the television series of the
same name; the competition is a battle between the starship Enterprise
and a Klingon battle cruiser. Aficionados of the television program will
recall that Enterprise traveled where human beings had never ventured
before. The exploration was done on behalf of a collection of cooperat-
ing races called the Federation, which was apparently dominated by
humans. Klingons were the hirsute rivals of the Federation for the domi-
nation of the Milky Way, if not of the entire universe.
Although Star Trek originally required a powerful graphics-research
computer, it can now be programmed comfortably on a personal ma-
chine. The two starships orbit a central sun, launching missiles at each
81
other and dodging the return fire. Both ships and missiles are subject to
gravity, and motion

orbital determines much of the action. Pilots with-
out a reasonable feel for celestial mechanics incinerate their ships in the
sun or unintentionally cross the deceptive trajectory of a missile's orbit.
The side whose ship blows up first is the loser.
At one time versions of Star Trek had appeared at hundreds of uni-
versities and other institutions; the game was officially frowned on but
secretly enjoyed. I find it ironic that years after those late-night interrup-
tions I still recall the game with fondness. Star Trek was only one of the
many games developed by students, and most such games soon found
their way into commercial packages that were largely responsible for the
revolution in home computers. Commercial versions of Star Trek have
only recently been introduced, although arcade versions have been
available for some time.
The version of Star Trek described here takes the reader back to the
clandestine romance of the earliest computer games. Furthermore, it
serves to introduce the arcane subject of arcade programming. One goal

is keep the screen alive: the computations that create the action in the
to
Star Trek world must be as fast and as simple as possible. A second pro-
gramming goal is to create a realistic gravitational environment, but here
too readers have encountered the same problem in somewhat different
guise. In The Armchair Universe I introduced a star cluster in which stars
dance about in accordance with their mutual gravitational attraction.
Gravity of a simpler kind warps the tracks of ships and missiles in Star
Trek. Only the central sun exerts a perceptible force.
To engage in their orbital duel two players sit at the keyboard of a
computer and press the keys assigned to control their ships. One player,
perhaps the hairier one, takes command of the Klingon ship; the other
guards the fortunes of the Federation. At first only the central sun and
the two ships occupy the screen. The sun is a circle and the ships are
icons with just enough detail for distinguishing friend from foe (see Fig-
ures 8.1 and 8.2).
When the game begins, both ships are in free fall toward the sun. The
players immediately turn their ships away from the fall line and fire their
A ship that touches the
rocket engines to bring the ships into safe orbit.
sun instantly vaporizes and of course the game is lost.
As soon as the orbits are established, each player begins trying to
eliminate his opponent. One extreme tactic is to lie in wait until the
enemy ship passes nearby. A quick salvo may then finish it off. Another
extreme is trickier: one can try to aim the shots from a position on the
opposite side of the sun. On the screen bright points of light — called
photon torpedos on the television program —
fly outward from the fir-
ing vessel and burn their way around the sun in a gently curving array of
menace. Unless the enemy ship is commanded by a pilot of extraordi-
nary skill, the ship is destined to meet one of the missiles and explode in
a burst of interstellar debris. The screen signals the event with a brief
cloud of dots and announces either victory for the federation or
VICTORY FOR THE KLINGON FORCES.
Because the action near the sun is so intense, cautious players prefer
The main disadvantage of the strategy is the need to
to orbit farther out.
recharge a solar energy cell. The control of each ship depends on its en-
ergy cell. When the cell is spent, the ship must quickly move closer to the
sun to replenish its supply of solar photons. In a distant orbit the photon
stream is weak and the ship runs the risk of becoming a sitting duck.
There is another tricky feature to be aware of in high orbital flight:
the battle space in Star Trek is "toroidal." In other words, if the ship
moves too close to one edge of the screen, it will disappear there and
reappear near the opposite edge.
Each ship has an infinite supply of missiles, but there can never be
more than 10 of them in flight at the same time. The missiles obey the
same laws of physics as the ships do. A missile lasts either until it strikes
a ship (including its own) or until it runs out of fuel. So much for the
game.
The program I call trek is the most ambitious one presented in this
book, not so much because of its complexity as because of its length.
Some of the more standard routines can only be sketched in. Program-
ming neophytes may nonetheless try the project with some hope of suc-
cess and considerable entertainment.
trek cycles through six major sections of code as long as both ships
are operational:
Read the keys.

Update ship and missile positions.
Check for contacts.
Update energy of ships.
Manage missiles.
Display.
With the possible exception of the first section, many readers will
find that trek is relatively straightforward to write. Reading keys will be
•
•
<<
*
Figure 8. 1 Enterprise and the Klingon battle cruiser exchange missiles.
new to some, but the facility is indispensable for arcade programming.

Most high-level languages have statements that enable a program to test
whether a particular key has been pressed.
For example, in the language Micro-soft basic the relevant com-
mand is
On key(/c) gosub n.
When the "On key" command is executed, the program checks at the
beginning of every subsequent command to determine whether or not
key k was pressed. If it was, the program branches to the command at
line n. At line n a subroutine begins whose purpose is to record the

pressing of key k, by assigning some value to a "flag" variable.
generally
There are 14 specific keys on the keyboard of the IBM PC that can be
checked in this way: the 10 function keys and the four cursor-control
keys. One must assign four keys to each player — say function keys Fl,
F2, F3 and F4 to the Federation forces and the four cursor keys to the
Klingon forces. The numbers for this assignment are respectively 1
through 4 and 11 through 14. A manual is indispensable here.
For each side the first of the four keys controls thrust, the next two
control direction and the last one controls the firing of missiles. The
thrust key simply imparts a fixed thrust, at full throttle, for one program
Figure 8.2 The Klingon ship is not able to dodge the barrage.
cycle. Each press of a direction key turns the corresponding starship

through an angle of 10 degrees. A tap on the missile key fires a single
missile. Before playing the game it is useful to stick small square labels
on the controlling keys. Appropriate symbols marked on the labels can
then remind the combatants which key does what.
The program trek must include an "On key" command for each of
the eight designated controlling keys. "On key" is not executed unless it
is preceded by the command
"Key(&) on."
All the subroutines to which trek branches from "On key" com-
mands are simple and essentially the same. Each subroutine is made up
of two instructions. The first instruction sets a flag variable to 1 for later
consultation by the program; the second causes a return of the program
execution to the line number of the "On key" command that invoked
the subroutine. The might as
flag variables for the control of Enterprise
well be called fdgo, fdrt, fdlt and
fdfr. They stand go"
for "Federation
(turn on thruster), "Federation right" (turn clockwise), "Federation
left" (turn counterclockwise) and "Federation fire" (shoot a missile).

Similarly, the variables kngo, knrt, knit and knfr record the thrust, direc-
tion and missile firing of the Klingon battle cruiser.
When a flag variable is set to 1 within a subroutine, it triggers a
change in one of the starships. For example, when fdgo is 1, the position-
updating segment of the program adds a small acceleration (between 2
and 5, according to taste) to the current acceleration of Enterprise, trek
must then reset the flag to 0.
Updating the positions of two starships and a handful of missiles is
much easier than managing an equal number of massive stars. The com-
bined mass of warring hardware is trivial compared with the mass of the
central sun, and so the mutual gravitational attractions of ships and mis-
siles are assumed to be zero. Given the distance of each object from the
sun, trek simply calculates the acceleration of the object caused by the
centrally directed solar gravity, updates the velocity of the object and fi-
nally revises its position.
Even for such a conceptually simple calculation there is substantial
computational overhead. Sums, products and square roots are needed to
carry out each calculation. To avoid slowing down the game with exces-
sive arithmetic, trek consults a table; for every possible distance from
the sun an array called force gives the predetermined acceleration expe-
rienced by an object (see Figure 8.3). Since the Star Trek universe is tor-
oidal, the new position derived from the acceleration found in the force
table must be calculated modulo the horizontal or vertical distance
across the rectangular display.
Two arrays give the current velocity and position for the two star-
ships and as many as 20 missiles; the arrays are called vel and pos. Each
array has two columns and 22 rows. The first two rows hold starship data
and the next 20 are devoted to missiles. Thus in the first row of vel the
two entries vel(l,l) and vel(l,2) are respectively the velocities in the x
(horizontal) direction and the y (vertical) direction of Enterprise. Simi-
larly, vel(2,l)and vel(2,2) give the two mutually perpendicular velocity
components of the Klingon battle cruiser. A special variable called mis-
num enables trek to keep track of the number of missiles currently in
flight. Misnum ranges from 2 (no missiles) through 22 (20 missiles).
By lumping starships and missiles together in the arrays one creates a

shorter and slightly more efficient program. Only one loop is needed to
update both positions and velocities. Before the update, however, two
variables are needed to keep track of the orientations of the two space-
craft: fdor and knor. Their values are expressed in degrees, where the
DISTANCE FORCE
10 8.000
11 6.612
12 5.556
800
FORCE =
(DISTANCE) 2
178 .025
179 .025
180 .025
Figure 8.3 Sample force table and its formula.
angle 0 indicates a ship is pointing east, 90 indicates north and so on.
Every time a player presses either direction key, one of the two variables
is increased or decreased by 10 degrees as appropriate.
The loop for updating positions is indexed by the letter z, which

ranges from 1 through 22. For each value of z the program calculates the
acceleration: the current values of the x and y coordinates of the zth row
of pos are squared and added together, trek then finds the square root of
the sum, and that number is truncated to the greatest integer less than or
equal to it. The truncated intetger square root is the approximate dis-
tance from the sun to the position of the zth object. That integer is then
taken as an index to the acceleration table, where the solar attraction can
be looked up.
For the two starships the components of solar acceleration must
then be added to the components of thrust. If either flag variable fdgo or
kngo has been set equal to 1 (in other words, if the thrust key has been
pressed) and if neither starship is out of fuel, trek must multiply the
thrust constant by its horizontal and vertical components. For my moni-
tor I take the thrust constant to be 3, which gives reasonable maneuver-
ability to the spacecraft. For Enterprise the horizontal component is then

three times the cosine of fdor; the vertical component is three times the
sine of fdor. For the Klingon ship the thrust components are derived
from the angle knor. As soon as the calculation is done trek resets both
fdgo and kngo to 0; the throttle is turned off until the next time the thrust
key is pressed.
When the index i becomes greater than 2, the calculation of thrust
can be bypassed because missiles have no thrust. The rest of the updat-
ing loop is devoted to calculating new velocities and positions for the
moving objects on the screen. For each object the numerical magnitude
of the acceleration is added to that of the velocity, and the magnitude of
the velocity added to that of the position. Such a simplistic calculation
is
is made possible by adjusting the thrust and solar attraction to reflect a
system of units that assumes the passage of one unit of time for each pro-
gram cycle.
To check for contacts among the various objects the program must
first determine for each ship whether the ship lies on or within the
boundary of the sun. Since trek has already calculated the updated dis-
tance between each ship and the sun, the program needs only to com-
pare that distance with the solar radius, say 10 units. If either ship has
collided with the sun, trek responds with an appropriate screen mes-
sage, such as klingon vaporized; the program then branches to its dis-
play segment.
A second check for contact must determine for each missile whether
the missile lies within a certain small distance of either ship. Here trek
uses a simple but effective shortcut: it finds the difference between the x
coordinates of the missile and a ship, and it does the same for the two y
coordinates. Finally it adds the two differences; the process avoids both
squaring and taking square roots, and the result is nearly as good as the
usual distance calculation. If the sum of the two differences is less than,
say, 4, theprogram scores a hit. A message appears on the screen, such
asenterprise hit by A missile, klingon wins. A single loop carries out
the test for each missile. Its index starts at 3 and ends at misnum, the
number of missiles in space plus 2.
To update the energy levels of the ships the program divides the
solar acceleration obtained from the table by 60. Since the acceleration
increases as the ship moves closer to the sun, such a ship can receive a
more concentrated stream of energetic solar photons. The energy is then
added to a fuel variable called fdft or knfl, depending on which starship is
involved. Each of these variables is decreased by .1 when thrust is ap-
plied; the decrements are made in the position-updating

segment of
trek. A starship is considered out of fuel if it has no more than one unit
of solar energy left in its tank. The tank begins life with 10 units of fuel.
Missile management requires an array called time, which stores the
number of program cycles in the life of each missile. When a cycle count
reaches 25, the corresponding missile is removed from the array pos, the
count is reset to 0 and misnum is decreased by 1. The missile can be re-
moved from pos in one of two ways. The first method is easier to pro-
gram but may slow down the game, trek runs through the array from the
index value at which the missile is removed and decreases the row num-
ber of each entry by 1. Thus the last entry to be shifted lies at index mis-
num. It is moved to row number misnum-1. The same operation is car-
ried out on the arrays vel and time.
A faster technique takes advantage of the observation that the oldest
missiles have the smallest indexes; they were the first missiles added to
the One can therefore keep track of the missiles without shifting
list.
their indexes; missiles whose age has reached 25 program cycles must all
be found at the beginning of a contiguous group of missiles in each array,
and so only they must be removed. Similarly, new missiles are always
added at the end of the contiguous group.
Introduce two new variables called old and new, which serve as
pointers to the oldest and newest missiles in each array. As missiles are
removed and added, only the values of old and new must be changed.
One can then apply modular arithmetic to keep the contiguous group of
missiles cycling around in each array. When a new missile is to be added
at index value 23, trek reduces the index modulo 23 to 0 and then adds
3 to avoid replacing one of the starship coordinates by a missile coordi-
nate. The variables old and new undergo the same process. Such a data
structure is called a circular queue. If this arcade trick is used, the posi-
tion-updating segment of the program must be modified: split each sin-
gle loop into two smaller loops, one for ships and the other for missiles.
When a player presses a key to fire a missile,trek first checks a
count of missiles currently activated by that side. If the count is less than
10, trek consults the values of the flag variables fdfr and knfr. If fdfr is 1,
for example, the program adds 1 to the missile count for the Federation,
increases misnum by 1 and then loads the position and velocity coordi-
nates of Enterprise into the appropriate slots of pos and vel. In the pro-
cess the program should add four units to the position coordinates and
two units to the velocity coordinates; in both cases the addition is made
along the same direction as the ship is currently moving. For example,
the horizontal position coordinate of a missile fired by Enterprise is four
times the cosine of ) dor plus the horizontal position coordinate of Enter-
prise; the vertical coordinate increases by four times the sine oijdor. The
initial position of the missile is thereby kept clear of the ship, preventing
immediate destruction by the ship's own photon torpedo. The same
operation applied to the missile's velocity coordinates reflects a relative
launch speed of two units per cycle: a missile travels two units per cycle
faster than the ship that launches it.
The last major section of trek displays the sun, two ships and what-
ever missiles are active at a given time. The program draws a circle of
radius 10 in the center of the screen and then works its way through the
array pos. The Federation and Klingon ships are represented by icons.
One icon is essentially a circle that recalls the famed discoidal Enterprise
with its twin engine booms. The Klingon icon is angular (see Figure 8.4).
Readers are free to attempt any miniature variation on these ships as
long as two things are reasonably clear at a glance: which ship is which
and where each one is headed. In drawing the icon for either spacecraft,
trek calls on a list of display points that must be translated and rotated to
reflect the position and orientation of each object. For these operations
the program consults the array pos and the variables fdor and knot.
Missiles are simpler. The display program draws each missile as a
point within a single loop, consulting pos as it goes along.
trek is one disadvantage of the display screens now in
subject to
popular use. Such screens operate in storage mode: an object drawn on
the screen remains there. To avoid a confusing welter of remnant ships
and missiles trek must draw each object twice. It first draws the object
in its old position in black. Then it redraws the object in its new position
in the normal color.
I program to the Trekkers
shall leave the details of initializing the
who attempt it. some of you may find
In spite of arcade programming,
the game too slow; you may be tempted to call it Star Truck. For better
performance try compiling your program, or impose arms limitations on
the number of missiles allotted to each side.
I have described only the bare-bones version of Star Trek. Fancier
but private editions have been built and they continue to propagate;
games have appeared that allow three or more spacecraft, laser guns,
color graphics and status displays. I must thank Jonathan N. Groff of
Clearwater, Fla., for reminding me of this underground classic and for
introducing me to a version of the game that includes an automated
Klingon. Earthlings representing the Federation are continually wiped
out by Groff's program.
, M , — —— 1
Weather ii a Jar
The butterfly effect was the reason. For small pieces of

weather . prediction deteriorates rapidly. Errors
. .
and uncertainties multiply . . .
JAMES Gleick, Chaos
T JL he famed Lorenz attractor seems an apt symbol for the field of dy-
namics and chaos. Its gracefully interfolded wings remind us of the but-
terfly that flutters in Venezuela only to cause a typhoon in Taiwan. But
how many readers know the arcana behind the Lorenz attractor, that it
represents a miniature weather system confined to ajar? I will show how
to simulate the system, not with Lorenz's differential equations, but
with an equivalent dynamical system of equal fascination Lorenz's —
water wheel.
James Gleick tells the story behind the Lorenz attractor in his book,
Chaos. (See the Further Reading section at the end of this chapter.)
More than 30 years ago the MIT weather theorist Edward Lorenz
constructed a primitive weather simulating program that ran on his
Royal McBee computer, "a thicket of wiring and vacuum tubes
that . . broke down every week or so." The program incorporated
.
equations that governed the interactions of air pressure, temperature,

humidity, and so on. He would begin a run of his program by feeding
numbers that represented meterological data at a number or "stations"
throughout his miniature world. Then he would press the start key and
the system would churn through the equations. Storms and calms fitted
through the vacuum tubes as several "days" of weather passed. Then
Lorenz would end the run. He was watching for consistencies in the
data. If he could develop predictive methods with his ultra-simple
weather simulation program, there might be hope for the real thing. He
93
hardly expected that his model would hint so strongly in the other direc-
tion.
It happened one day when he typed in data for yet another run and
went off for a coffee to let the Royal McBee buzz along on its own. When
he came back to the office, he discovered to his horror that the run was
for nothing. He had accidently omitted several decimal digits in the
input numbers. A careful experimenter, Lorenz re-entered the data,
fully expecting that the results of the more accurate numbers would
differ very little (if at all) from those of the original run. He was aston-
ished to find that the weather patterns prevailing in his miniature world
at the end of the new run were completely different. He repeated both
runs and got the same result. And again.
Searching for the reasons for the divergence, Lorenz refined and
simplified his weather model. In the end it did not span the whole world,
nor a single country, nor even a county. It consisted of a single convec-
tion cell. Such volumes are fairly common in weather systems: Warm
air, heated by ground radiation rises. As it goes up the air cools. Could a
single convection cell hold the key to strange behavior in his weather
model?
Figure 9.1 shows a single cylindrical volume of circulating air. Heat
applied at the bottom of the cylinder causes the central air to expand and
rise, cooling as it goes. At the top of the cylinder, the air spreads out to
the sides of the cylinder and begins to descend, cooling even more and
completing the circulation. Lorenz used three differential equations to
model events inside the convection cell. When he solved these by com-
puter under different initial conditions and for different values of the pa-
rameter that represented the rate of heating, it did not take him long to
discover an extraordinary phenomenon. In certain cases, a very slight
change in the initial inputs would cause completely different behavior in
the weather-in-a-jar system. The phenomenon now goes by the name
"sensitivity to initial conditions." It is the hallmark of chaos.
The famous Lorenz attractor (Figure 9.1) summarizes all possible
behaviors to which the weather-in-a-jar is "attracted" as it runs. The
lines that wind around and around within the attractor represent trajec-
tories in phase space, a three-dimensional space in which every point
represents a unique combination of velocity, temperature, and rate of
temperature change. As the computer solves the equations iteratively,
the point moves through phase space. When the heat is turned up, so to
speak, the point traces out the strange, nested curves of the Lorenz at-
tractor.Here, you can see sensitivity to initial conditions explicitly.
There are pairs of curves that pass arbitrarily close together within the
Figure 9. 1 Weather in a jar and its attractor.
attrac tor's central portion, yet diverge to different wings of the butterfly.
Within a short time, jar weathers that seem nearly identical develop in
completely different ways.
To make your own Lorenzian weather model and watch it run, you
do not have to build a little jar with air circulating inside it (although a
French physicist did something very similar). Nor is it necessary to write
a computer program that solves the three equations iteratively. All you
have to do is write a program that mimics a certain water-wheel, a pre-
cise mechanical analog of weather in a jar.
Figure 9.2 shows a most peculiar water wheel in the act of turning.
Eight leaky buckets hang from the wheel at evenly spaced points. Above
the center of the wheel, a steady stream of water pours into each bucket
as it swings underneath. Each bucket leaks at the same rate and, as it fol-
lows the wheel around, loses water.
The Lorenzian water wheel has no direct application that I am aware
of — except to illustrate the behavior of the convection cell. Here, water
is the analog of heat. Each bucket represents a portion of the cell's air
that loses heat as it circulates within the cylinder. The analog does not
suffer much from this discrete form of air and, provided there are
enough buckets, the exact number is not important. The varying speed
of the water wheel recreates the behavior of the air within the weather
cell.
And what strange behavior it is! With water pouring slowly onto the
wheel after a discreet clockwise nudge, it will rotate gently clockwise,
speeding up to some limiting rotational velocity and maintaining it

thereafter. But turn on the tap and let the water pour in. The wheel may
speed up for a time, then mysteriously begin to slow, stop, and even re-
verse direction! Then it may (or may not) reverse direction again. The
behavior seems random, even perverse. This is chaos. The reason for
any particular behavior is never far to seek. If the wheel slows down and
reverses, it is only because the buckets ascending the return side are
fuller than those descending. The latter went by the tap more quickly.
Readers with a mechanical bent are of course free to build their own
Lorenzian water wheel but others will resort to their computers, all on
fire to see the wheel rotate.
The program I call WEATHER WHEEL captures the Lorenzian
water wheel and its behavior within a computer. The eight buckets com-
prise an array by the same name and time passes in discrete steps that
vary with the speed of the wheel. The algorithm in Figure 9.3 consists of
four parts, an input segment, a dynamical computation segment, a rela-
belling segment, and a display segment.
The user of WEATHER WHEEL
inputs the rate / at which water
enters the system through the spout. The initial velocity v simply sets the
wheel turning. Note that the initial velocity must be non-zero.
input filling rate f, leaking rate g

input velocity v
initialize buckets to 0 (empty)
repeat
***compute dynamics***
dt <r- abs(l/v)
bucket(O) <- bucket(O) + f *dt
for k <- 0 to 7
if bucket(k)>=g*dt
then bucket(k) <- bucket(k) - g*dt

else bucket(k) <— 0
rightwt <— 0
for k <- 1 to 3
rightwt <— rightwt + bucket(k)
leftwt «- 0
for k «- 5 to 7
leftwt <- leftwt + bucket(k)
acc «— (rightwt — leftwt) *const*dt
v <— v + acc
if v = 0 then v <- 0.001
***leak and relabel buckets***

if v > 0 then inc «- +1
< 1
else inc
for k 0 to 7
j
«— (k + inc) mod 8
bucket(j) <— bucket(k)
***display wheel***
(display procedure)
until key pressed
Figure 9.3 Water wheel algorithm.

new state of the water wheel is computed on

Inside the repeat loop a
the basis of the previous state.The time constant ^reflects how quickly
the wheel has been turning. It controls how long the top bucket (always
labelled 0) will remain under the spout. Immediately, the top bucket re-
ceives j*dt gallons (or any unit you want to imagine) of water.
Before the wheel turns in this discrete rotational world, the buckets
must all leak a little. A loop runs through all eight buckets, removing a
quantity of water that is proportional to the leak rate g applied over the
time interval dt. Next, contributions to any change in velocity must be
computed by adding up the weights of the three right-hand buckets and
those of the three left-hand buckets. The acceleration, a, will be the dif-
ference between the two weights multiplied by some constant, const,
and again by dt, the period over which the acceleration will act. Readers
may want to start with const = 1 and then alter it later, if necessary.
WEATHER WHEEL next updates the velocity by adding the accelera-
tion increment. This step is kosher if I argue that the constant that would
otherwise multiply acc has been absorbed by the units chosen.
Because the time increment uses II v, the program must prevent v
from becoming 0 by testing for this condition and then making it slightly
non-zero, if necessary. The generation of a 0 by the program is highly un-
likely, but possible.
It may seem eccentric to relabel the buckets but that simplifies the
programming somewhat. The alternative involves a rather messy com-

putation of which buckets currently contribute to the right-hand and
which to the left-hand weights. But watch the modular arithmetic! It re-
quires that the index values for the bucket array start at 0. Depending on
which way the wheel happens to be turning (v is positive or negative),
the contents of each bucket are transferred, in effect, to its clockwise or
counter-clockwise neighbor.
The display procedure opens up a host of possibilities. At a mini-
mum, any decent graphic rendering will show
eight apparently station-
ary buckets with differing amounts of water circulating one way or the
Each bucket may be represented by a square in which either a) a
other.
number representing the current contents is printed, or b) a special 'fill" '
routine covers the lower area of the square to an appropriate height with
graphic water. Two options enter the picture in this case: How much
water should buckets be allowed to hold?
Readers who place no limits on bucket capacity must, nevertheless,
limit the graphic fill to the top of each bucket. Readers who limit the
amount in their programs will not face this problem. On the one hand,
the heat capacity of air is and so should the ca-

theoretically unlimited
pacity of buckets be. But I suspect it will not make any difference which
option one pursues except at very high fill rates.
One final and crucial piece of advice for the display section: Add a
timing loop that freezes the display for a period of time that is propor-
down, of course, but only in a
tional to dt. This will slow the simulation
physically realistic way. Who wants to watch the Lorenzian wheel
"spinning" always at the same rate? The timing loop holds up the pro-
gram by executing some trivial or pointless instruction while it counts
toward the loop limit, int(lOOO^)- M the 1000 factor does not produce
an appropriate increment, it may be increased or decreased by factors of
10.
This brings to mind some related issues concerning physical simula-
tions. I have played fast and loose with the physics of the Lorenzian
water wheel. Consider what happens in the algorithm: The wheel is es-
sentially stationary while water is added to the top bucket and then
leaked out of all the buckets. Then the weight contributions are calcu-
lated to produce a new wheel speed and the buckets all move, in effect,
to their next positions. The newly calculated speed determines how long
the top bucket will fill and all the buckets will leak during the next cycle
of computation. All of this seems reasonable if you imagine that during
each cycle the wheel turns steadily at precisely the predetermined rate,
v. During that period, the top bucket will receive exactly the amount
computed. Andduring that period some of the buckets, at least, may
well leak the amount of water so computed. Unfortunately, real leakage
(like real heat loss) is proportional to the amount present. I have not built
this into the model on the assumption that it makes little or no differ-
ence. Careful simulators may want to alter the loss from "g*dt" to
"bucket(k)*dt", allowing the constant g to alter its meaning.
These simplifications, due mainly to the discrete nature of the
model, will nevertheless allow chaos to appear. Readers with a yen to
explore sensitivity to initial conditions may experiment by isolating
what they suspect to be chaotic behavior, having made a careful note of
the initial conditions that led to it. Now alter those conditions by how-
ever slight an amount. Does a similar scenario evolve or an entirely dif-
ferent one?
Further Reading
James Gleick. CHAOS Making a New Science. Viking, 1987.
n
1 PORTRA J
I feel that art has something to do with the

achievement of stillness in the midst of chaos.
Saul bellow
w, Mario Markus of the Max Planck

hile investigating digestion,
Institute for Nutrition discovered beauty in chaos. He and his collabora-
tor Benno Hess have studied several mathematical models in an attempt
to simulate how enzymes break down carbohydrates. By adjusting a pair
of parameters, they found they could make the simulated enzymes be-
have in either an orderly or chaotic manner. To illustrate the chaos in-
herent in the model, Markus created a series of portraits of chaos that
provided not only food for thought but also a feast for the eyes.
The images are based on a formula named after the Russian mathe-
matician Aleksandr M. Lyapunov. The formula generates a single num-
ber for any dynamic system. Known as the Lyapunov exponent, the
number indicates how chaotically the system is behaving. When the
Lyapunov formula is applied to Markus 's model, it produces an expo-
nent for each pair of parameters. By representing each pair as a point on
a computer screen and assigning colors to each point depending on the
value of the exponent, Markus created what I call Lyapunov space. In
this virtually unexplored territory, order inhabits colored shapes, and
chaos lurks in black regions.
Not long after his work appeared in academic journals, Markus
rushed several pictures to an art gallery for exhibition. He can hardly be
blamed for doing so. The pictures could just as easily have been made by
an apprentice to Salvador Dali.
The model developed by Markus is based on a variation of the so-
called logistic formula, the simplest-known formula that describes a cha-
100
otic dynamic system. The formula contains a variable, x, whose value is

always somewhere between 0 and 1. It also involves a system parame-
ter, r. The formula can be written:
x «- rx(l — x)
The arrow indicates that once r, x, and 1 — x have all been multiplied to-
gether, the resulting number becomes the new value for x, that is, it re-
places the old value. The process can be repeated so that the formula
continually spews out new values for x.
The resulting sequence holds great fascination for dynamicists, but

what does it all mean? The logistic equation gets its name from the logis-
tics of animal populations. In the equation, x represents the number of
animals in an isolated region divided by the maximum number that the
region could ever be expected to support. The amount of food available
is therefore proportional to 1 — x. In other words, as the number of ani-
mals (x) approaches the maximum (1), the food supply (1 — x) dwindles
to nothing (0). The parameter r expresses the proportionality. It may be
thought of as the fecundity of the population. The higher the value of r,
the more quickly the population will rebound from any disaster.
Strangely enough, higher values are precisely the ones that lead most
quickly to chaotic populations.
Although the equation is too simple to represent real animal popula-
tions, it can serve as a rough approximation to population dynamics.
If the parameter r is less than 2, the sequence of numbers produced
quickly homes in on a single value.It makes no difference what number
the formula uses for x initially.The population always converges to a

stable value. In the jargon of chaos theory, a system whose dynamics
stabilize at a single value is said to have a one-point attractor.
If the parameter r is greater than 2 but less than about 2.45, the logis-
ticformula generates numbers, that eventually alternate between two
values. The system then converges on a two-point attractor. In some
sense, when fecundity is high, the population pays a price: its size fluctu-
ates over time.
If is cranked up to values greater than 2.45, the
the fecundity factor
logistic formula produces numbers that converge on a four-point attrac-
higher values of r lead very quickly to eight-point attractors,
tor. Still
then 16-point ones and so on. But if the value of r is greater than 3.57 (or,
to be more precise, 3.56994571869), chaos reigns.
At that level of fecundity, the formula seems to generate values at
random, though, to be sure, the formula is deterministic. The reason for
this strange happens to be a one-dimen-

behavior lies in the attractor. It
sional fractal. Like when any small piece of it

all fractals, it is self-similar:
is magnified, the enlarged region looks very much like the whole.
The fate of the hypothetical populations is clearly portrayed in Fig-
ure 10.1. The diagram is produced by plotting r against the values to
which the logistic formula converges. The result is a kind of tree. One-
point attractors make up the trunk; two-point attractors, the first pair of
branches. At an r value of 3.57, the onset of chaos can be seen clearly:
the branches suddenly proliferate.
Chaos can be characterized using the Lyapunov formula. For each
dynamic system, the formula produces a single number, the Lyapunov
exponent. If the number is less than 0, the system is stable. If the number
is greater than 0, the system is capable of chaotic behavior.
The Lyapunov formula is complicated, but it can be translated into a
series of simple steps. In the case of the logistic system, we start with one
particular value of r. The logistic formula is iterated, say, 600 times, so
that the numbers converge to whatever attractor is present in the sys-
tem. After the numbers settle down, it is safe to compute the Lyapunov
exponent. The following recipe outlines the computation:
POPULATION LEVELS (x)
Figure 10.1 Population levels, x, can fluctuate over a wider range of

values as the fecundity factor, r, increases.
total <- 0
forn «- 1 to 4,000
x <— rx(l — x)
fofo/ <- total + ((log|r - 2rx|)/log 2)
//op <- total /4.000
The algorithm first sets fota/ = 0 and then iterates the logistic formula
4,000 times. On each iteration it computes a new value for total by add-
ing the old value of total to the logarithm of r |
—
2rx\ divided by the loga-
rithm of 2. (The vertical bars indicate the absolute value of r — 2rx.) The
quantity |r — magnitude of the rate at which succes-
2rx\ represents the
sive values are growing or shrinking.When it has added up all 4,000 log-
arithms, the algorithm divides the sum by 4,000. The result, which has
been assigned to the variable lyap above, is something like an average
logarithm of change. The result closely approximates the Lyapunov ex-
ponent.
Readers who demand precision can more accurately estimate the
Lyapunov exponent by increasing the number of iterations and, at the
end of the procedure, by dividing the sum of logarithms by the number
of iterations.
Iencourage readers to use the algorithm above to calculate the Lya-
punov exponent for r equal to 3. Then compare the result with that ob-
tained when r equals 3.57. The first number should be slightly negative,
indicating a stable system, and the second number should be positive, a
warning of chaos.
Figures 10.2 through 10.4 are all based on the logistic equation.
Markus merely adds one twist of his own. To produce his pictures,
Markus used periodic forcing. This means that r systematically changes
its value, alternating between two fixed numbers, a and b. In other
words, the logistic equation is iterated with rvalues of a, then b, a, b, a, b

and so on. The resulting system may or may not develop chaotic be-
havior. The issue can be settled only by calculating the Lyapunov ex-
ponent.
For that reason, Markus plotted the value of the exponent for each
possible combination of a and b values. A two-dimensional image in
Lyapunov space emerges when the points {a, b) are colored in some con-
sistent fashion. Markus assigned black to all points {a, b) that produce a
non-negative value for the Lyapunov exponent. Chaos is black. Shades
of a single color, such as yellow, appear everywhere else in Lyapunov
space. From a Lyapunov exponent of zero down to minus infinity, the
Figure 10.3 A Lyapunov "jellyfish."
shade ranges continuously from light to dark. At zero, there is an abrupt

discontinuity as the color suddenly turns from bright yellow to black.
The resulting images are maps of chaos in the forced logistic system.
In particular, the upper image in Figure 10.2 depicts the straightforward
system just described. The parameter r alternates in completely regular
fashion between a and b.
Figure 10.4 Zircon Zity.
The crossing point of two or more spikes in any of the accompanying

images reveals the coexistence of periodic attractors. This means that, at
a point (a, b) where such a crossing occurs, the corresponding dynamic
system in which r alternates between a and b will have two attractors.
Which attractor operates depends, strangely enough, on the initial value
that one chooses for x before iteration.
If the Lyapunov exponent is plotted for a succession of initial x
values, it may take on a specific value, say, 0.015 for a number of these
initial values. Then the exponent may suddenly switch to another value,
0.142, to which it may stick for several more successive initial x values
before reverting to the first value. The switching back and forth can be-
come quite frequent.
Lyapunov space often contains darkish bands that run along the
spikes. These represent superstable regions in which the underlying
forced logistic systems exhibit the most regular behavior.
The Lyapunov space generated by alternating a and b values con-
tains a tiny fleck that resembles a swallow. The fleck is enlarged in the
image at the bottom of Figure 10.2. There, just off the swallow's tail, lies
another little fleck. Readers are free to guess just what it might turn out
to be when it is similarly enlarged.
The appearance of self-similarity in the figure should not surprise
students of chaos. Structures that exhibit self-similarity are more often
than not produced by chaotic processes.
The methods used to create the mother swallow and its offspring can
The images in
be varied slightly to generate a host of different creatures.
Figures 10.3 and 10.4 differ only in the a and b value sequences that
were used. For example, a jellyfish —
the yellow tentacled blob shown
in Figure 10.3 —
is spawned from a sequence that begins b, b, a, b, a and
that repeats over and over again.
Figure 10.4 resembles the cover of a science fiction magazine from
the 1950s. I call it Zircon Zity because it is obviously the futuristic me-
tropolis of the Zirconites (whoever they are). The underlying sequence
of the zity is bbbbbbaaaaaa. By repeating this sequence while calculating
the Lyapunov exponent, a computer can build the city with all its deli-
cate bridges, golden spaceships, and interplanetary walkways.
What does all this have to do with enzymes, carbohydrates, and nu-
trition? At best, a small region in some Lyapunov map might actually de-
scribe the dynamics of enzymes breaking down carbohydrates. But per-
haps more to the point, Markus's work makes it possible to visualize the
dynamics of periodic forcing. One might say he has made chaos easier to
digest.
A Lyapunov Program
For those readers who wish to explore Lyapunov space, I can
give a few hints about how they might create the appropriate
program. The heart of the program should be a double loop
that runs through values of a and b. To compute the Lyapunov
exponent for each combination of a and b, the program should
include a routine based on the algorithm given on page 103.
The routine should allow various sequences of a's and b's to be
stored in an array. The routine could enable the user to specify
the sequence by filling the array with 0's (representing a) and
l's (representing b).
With each iteration of the Lyapunov loop, a counter called
index may be used to step through the array. Each time the
inner loop produces a new value for index, it will look up the
current value of either a or b, depending on whether it finds a
0 or a 1 in the array at the value of index. Finally, the program
should include a table of logarithms to help speed the
computations. Indeed, it may take hours for a personal
computer to generate a single image in Lyapunov space.
Further Reading
Freeman, 1990.
Mario Markus. "Chaos in Maps with Continuous and Discontinuous Maxima/'
Computers in Physics, September/October 1990, pp. 481-493.
Designer Fractals
So, naturalists observe, a flea Hath smaller fleas that

on him prey; And these have smaller still to bite 'em;
And so proceed ad infinitum.
JONATHAN SWIFT, On Poetry. A Rhapsody
S
V^*J o, mathematicians observe, if fleas are all the same except for size,
then all their hopping and rotations reduce to affine transformations.
What exactly is this high-sounding term? It is nothing more than a for-
mula for scaling, turning, displacing, and sometimes even distorting an
object geometrically. As in the case of fleas, a single affine transforma-
tion can be applied repeatedly to produce miniature replicas of the origi-
nal. People who prefer not to waste their talents on propagating fleas can
apply these rather simple geometric formulas to generate images as in-
tricate as the paintings in museums or landscapes in nature.
A small set of affine transformations can create such abstract works
as the Sierpinski triangle in Figure 11.1. A larger group of transforma-
tions can re-create landscapes such as the Monterey coastline shown in
Figure 11.2. In fact, any image whatsoever can be reproduced from a
series of affine transformations. The trick is knowing which ones to
choose. Along these lines, Michael F. Barnsley of Iterated Systems, Inc.,
in Norcross, Ga., has discovered a general procedure for reducing an
image into a series of affine transformations. His technique has opened
up some exciting possibilities for the transmission of television and
computer images. Before I describe his work, let me say a bit more about
affine transformations.
When an affine transformation is applied to a figure such as a triangle
or a leaf, it moves the points that make up the figure to new locations. In
the process, the transformation may translate, scale, rotate, and stretch
109
Figure 11.1 The Sierpinski triangle.
the original figure. If one starts with a triangle, an affine transformation

could shrink the triangle and move it to the left somewhat, thus creating
a second triangle. If the same transformation is applied to the smaller,
displaced triangle, it will produce a third triangle that bears the same re-
lation in size and proximity to the second triangle as the second does to
the first. One can apply the transformation repeatedly and watch the tri-
angles trace a path into infinitesimal oblivion.
If one applies an infinite series of affine transformations to an object,
the result has the property of being self-similar, that is, a magnified por-
tion of the result looks like the whole. Hence, a series of affine transfor-
mations can create a self-similar object, better known as a fractal.
All affine transformations have the same kind of formula for moving
the points of a figure around in a plane. The original point can be defined
by two coordinates, which I will call x and y. The new point has the coor-
dinates (x\ y'). The transformation is then defined by two equations:
x' = ox + by + e
y' = cx + dy + f
The symbols a, b, c, d, e, and /represent a set of numbers that determine

the character of the affine transformation.
What would happen, for example, b were equal to .5, c were — .5
if
and a, d, e and / were all 0? The two equations that define the affine
transformation would become:
Ill
Figure 1 1 .2 Photograph of Monterey coastline (top) and the same image

(bottom) reproduced from a few affine transformations.
X' = .by
Y'= - .5 x
To determine its effect on a specific point, say (1,2), one merely applies
the formulas. Thus, becomes (.5 X 2), which equals 1, and similarly y'
becomes — .5. If one carries out this operation for a great many points
in a triangle, a general pattern emerges. The entire triangle seems to have
rotated 90 degrees clockwise and simultaneously to have shrunk to half
its former size. If e and /were both equal to 1 instead of 0, then not only
would the triangle be reduced and rotated, it would also be shifted one
whole unit up and to the right.
This kind of transformation is called contractive because its effect on

any collection of points is to shrink the distances between them. The
transformation also preserves shapes. Contraction and shape preserva-
tion are key properties of the affine transformations employed in the
technique that Barnsley calls an iterated-function system.
The fun begins when several transformations of this kind are applied
many times to any figure the mind can imagine. The transformations,
along with their continued reapplication, make up an iterated-function
system. One might think initially that solving the formulas for iterated-
function systems involves an extraordinary amount of arithmetic. For
instance, if a sequence of transformations were applied to a figure com-
posed of 1,000 points, each transformation would have to operate on
1,000 points at a time. This would yield 8,000 arithmetic operations.
The astute reader will realize that to determine the effect of an iter-
ated-function system on a figure such as a triangle, it is necessary to
operate only on the three corner points. (The figure can be completed by
simply connecting the points with lines.) In this case the transformation
would only have to operate on three points at a time, and yet no savings
can be guaranteed for a figure that has an irregular shape whose outline is
defined by many points.
Do not despair. Barnsley has come up with a clever idea so that even
complicated figures can be efficiently transformed many times. The ad-
vantage of his technique is that only one point at a time is transformed.
His idea can be appreciated by playing the game I call fractal tennis.
This unusual racket sport requires four mathematically minded
players and an umpire. The players — Abby, Bob, Carla and David —
stand around a square court, separated by two nets that divide the court
into four square quadrants. The umpire assigns each player to a home
quadrant. To start the game, the umpire tosses a ball into Abby's home
quadrant and shouts "Bob." Abby lets the ball bounce only once, swings
her racket and hits the ball over the net so that it lands in Bob's quadrant.
Abby is not allowed to hit the ball to just anywhere in Bob's quad-
rant, however. To make the perfect shot as prescribed by the rules, she
must first judge where the ball landed in her own quadrant. This task is
rather simple because the tennis ball, which has been soaked in black
ink, leaves a mark on the court. Abby creates a mental map of the ink
mark within the entire court. She superposes the map on Bob's home
quadrant by shrinking it to half its dimensions. The position of the su-
perposed ink mark on Bob's quadrant is where she must hit the ball. If
the ball had landed in the center of the whole court, Abby would have
been required to hit the ball to the center of Bob's quadrant. In this case,
however, the ball landed two meters north and six meters west of the
center of the whole court. Because the dimensions of Bob's quadrant
and the others are half those of the court as a whole, Abby should hit the
ball one meter north and three meters west of the center of Bob's quad-
rant.
Abby's great shot represents an affine transformation. She has cre-
ated a second ink mark, which has been displaced to Bob's quadrant and
is closer to its boundaries.
After Abby's return, the umpire calls out "David." Bob rushes to
catch the ball on the bounce and makes a perfect shot into David's home
quadrant. The umpire, perhaps a bit maddened by the sun, then starts to
shout names at random. Yet Abby, Bob, Carla, and David, being con-
summate calculators, play a flawless game. Each player always hits the
ball to just the right point. After a while, however, the ink marks left by
the tennis ball create a fractal pattern on the court. In fact, the marks
eventually blacken the court uniformly. That is when the umpire calls a
halt to the play.
Figure 11.3 shows the early stages of the game. After Bob made his
first shot to David, the umpire called "David" for a second time. This did
not present a problem, except for David of course. He hit the ball toward
the southeast corner of his own quadrant because his quadrant is in the
southeast corner of the whole court. After the ball bounced in David's
own quadrant, the umpire called "Carla," and David hit the ball to her
quadrant according to the rules. If the umpire had yelled "David" again
and again, David would have had to direct the ball ever closer to the
southeast corner.
Fractal tennis illustrates a key feature of iterated-function systems.
A single point mapped repeatedly by a random sequence of affine trans-
THE UMPIRE SHOUTS

•BOB' 'DAVID'
CARLA'S DAVID'S
QUADRANT QUADRANT
'DAVID' 'CARLA'
Figure 1 1 .3 Opening shots in a game of fractal tennis.
formations will eventually "fill in" a certain region. The umpire's calls,
or transformations, are what determine the ultimate image. In practice,

the game ends when a satisfactory density of black dots emerges.
Readers who wonder why the game is called 'fractal" tennis will see
'
the game come into its own when it is played on the leaf of a black
spleenwort fern. The game still involves four players, but unlike the
classic smooth square of the practice court, the fern-leaf court has a
jagged outline. As Figure 11.3 shows, when the umpire calls "Abby,"
one of the players must hit the ball to Abby's leaflet. The point where the
ball must land depends on the position of the last bounce relative to the
whole leaf. In this way, the call "Abby" corresponds to an affine trans-
formation. The call "Bob" is also an affine transformation that sends the
"ball" to the corresponding point in the leaflet at the left side of the leaf
near the base. The call "Carla" does the same thing in relation to the leaf-
let at the right side of the base. Finally, the call "David" sends any point
of the leaf as a whole into the straightline segment representing the stem
at the base of the leaf.
When the point is put into play, the umpire begins to call the names
of these four players in arandom order. The point might start in the mid-
dle of the leaf,hop to the middle of Bob's leaflet, then shift to a point in
Carla's leaflet and so on. The game goes on for 10,000 hits. As it con-
tinues, an image of the fern leaf, delicate and organic-looking, emerges
(see Figure 11.4).
There is one element of the game that I have yet to describe. The se-
quence is not quite random: the judge favors certain players as he makes
the 10,000 calls. The case of the spleenwort leaf provides a perfect ex-
ample. Abby's leaflet has the largest area. Therefore, if the umpire is just
as likely to shout "Abby" as any of the other three names, Abby's leaflet
will fill in more slowly than the others'. To adjust for this, the umpire
gives Abby the lion's share of play. In fact, the amount is proportional to
Abby's share of the total leaf area.
The umpire calls "Bob" and "Carla" roughly the same number of
times, the numbers in both cases being proportional to the areas of their
respective leaflets. Because the stem has the least area of all, David will
get to play the least.
Perceptive readers may have noticed that the four transformations
associated with the spleenwort leaf change the basic outline into four re-
Figure 1 1 .4 An iterated-function system fills in the leaf of a black

spleenwort fern.
gions that approximately subdivide the outlined area. Barnsley calls this
subdivision a collage. He and his colleagues have found a theorem that
guarantees good fractal reproduction. The collage theorem says that the
more accurately the outline of a fractal shape is approximated by a col-
lage of a certain number of affine transformations of the shape, the more
accurately the iterated-function system will reproduce the original frac-
tal.
The collage theorem brings us to a fork in the expository road, that

between creation and re-creation. To what purpose will an iterated-func-
tion system be put?
Creatively speaking, almost any fractal can be constructed by means
of one iterated-function system or another. Take, for example, the Sier-
pinski triangle shown in Figure 11.1. This is a triangle from which a cen-
tral triangular area has been removed, leaving three smaller triangles.
Given the fractal nature of our topic, it will surprise no reader to find that
each of the three triangles has its own central triangle removed.
Three affine transformations participate in the production of the
Sierpinski triangle. They arise from a collage composed of three identi-
cal right triangles that all have the same orientation. The triangles in the
collage are positioned to form a right triangular hole of the same size as
one of the three triangles. Each of the triangles is associated with a par-
ticular affine transformation. When the game of fractal tennis is played
on this strange surface, the original triangle begins to fill in, except for
the hole in its middle. Each of the corner triangles has a hole in its mid-
dle, of course, and so do the triangles in their "corners. " The final object,
insofar as any finite scheme can reproduce it, is literally full of triangular
holes at all visible scales of magnitude.
Readers who have computers at their command can reproduce the
Sierpinski triangle by following an algorithm for the appropriate iter-
ated-function system. I will describe the algorithm in general terms. It
begins by setting the coordinates x and y equal to 0. Then three main
operations are repeated 10,000 times: First, one of the affine transfor-
mations is chosenrandom. Second, the chosen affine transformation
at
is applied to the current coordinates of the point, namely, (x, y). The re-
sult is a set of new values that are now deposited in the x and y symbols,
so to speak. Third, a test is made to determine whether 10 iterations
have been carried out.
The third step ensures that the ball has been in play long enough to
be somewhere in the court. In a general scheme such as this, one does
not know in advance the best place to start the ball bouncing. Hence, one
starts it from the origin and assumes that after 10 iterations it has pretty
well "settled into" regular play.
This algorithm will work for any iterated-function system if one adds
an extra feature. Because an affine transformation must be chosen at a
rate that depends on the area it must cover, the algorithm must select
each transformation with a certain probability.
What are the formulas for the affine transformations that produce
the Sierpinski triangle? At the beginning of this chapter, I described the
type of formula one needs, and I mentioned that six coefficients deter-
mine the transformation's character. The coefficients for the Sierpinski
triangle are given below.
a b c d e f
(1) .5 0 0 .5 0 0
(2) .5 0 0 .5 .5 0
(3) .5 0 0 .5 .5 .5
Each row of the table represents one of the three transformations.

One can tell what these particular transformations are up to almost
at a glance. All three have the primary effect of shrinking any geometric
figure to half its size, but transformation number 1 also shrinks the figure
toward the origin, whereas transformations 2 and 3 move the shrunken
form one unit to the right and one-half unit up and to the right, respec-
tively. Thus, the original triangle is transformed into three smaller trian-
gles, and the miniature triangles are in turn transformed into minuscule
ones.
From these simple ingredients, readers with moderate programming
experience should be able to reconstruct the Sierpinski triangle on their
display screens.
Up to this point I have addressed the creation of forms. There is not
as much the major application of
recreation in re-creation, but that is
iterated-function systems. Thanks to Barnsley and his colleagues it is

now possible to convert any scene into an iterated-function system.
Normally it takes thousands of bits of information to store the image
of a natural scene in a computer file. For example, an image might con-
sist of a 300-by-300 grid of pixels, or picture elements, each requiring
several bits to specify a gray level or color at that point. Ordinary pixel-
by-pixel storage might therefore take up a million bits or more. One can
apply standard compression techniques to such pictorial data to store
the same information in a smaller space, but iterated-function systems
promise compression factors of 500 to 1 —

or even better! The key:
store the iterated-function system rather than the image it produces. A
glance at the pictures of the Monterey coastline in Figure 11.2 enables
one to compare the veracity of the coded image (on the right) with the
original photograph (on the left).
The method begins with a computer image drawn directly from a
photograph or video camera. The image is analyzed and broken up into
connected patches, large and small, in which the gray level (or color) is
relatively constant. The pieces in this collage become the basis for a
search through a large library of standard affine transformations that
map some pieces into others. should happen, for example, that one
If it
of the transformations maps many other

a droplet of spray into a great
droplets, then that transformation would make a valuable addition to the
iterated-function system under construction. Actual images can then be
reconstructed in certain computers at video rates, that is, 30
faster than
whole images per second.
The commercial potential of the discovery so impressed Alan D.
Sloan, a mathematician formerly at the Georgia Institute of Technology
and a close collaborator of Barnsley 's, that the two researchers started
Iterated Systems. Among the company's current products is a video
modem that produces an iterated-function-system code for a 512-by-
512 pixel image in just under three seconds. At the receiving end the
images can be reconstructed at video rates. In the future Barnsley sees
full-color video transmission over telephone lines as a definite possibil-
ity. The technology may also be applied to automated pattern recogni-
tion and other projects that are as yet a fractal gleam in Barnsley' s eye.
Further Reading
Michael F. Barnsley. Fractals Everywhere. Academic Press, 1988.
Benoit B. Mandelbrot. The Fractal Geometry of Nature. W. H. Freeman, 1983.
The number of distinct scales of length of natural
patterns is for all practical purposes infinite.
Benoit B. Mandelbrot, The Fractal Geometry of Nature
Q
W-^ince Benoit Mandelbrot first coined the term, we have learned
B.
to look for fractals everywhere, especially in nature. Although nature
contains no true fractals, in which every part contains an infinite regress
of similar parts, it abounds with structures that produce the illusion. One
may find self-similar structure in certain leaves and plants, in shorelines,
perhaps even in the structure of galaxies and galactic clusters. But a
chapter on the elements of fractal generation may as well practice on the
traditional elements of nature: earth, air, fire and water.
Earth may be symbolized by a jagged mountain, air by a puffy fractal
cloud, fire by a fractal flame and water by an infinitely curly wave. Air
and water are shown in Figures 12.1 and 12.3 but readers can supply
earth and fire for themselves, based on what they glean from this chap-
ter.
Chapter 11, which describes the iterated function systems (IFS) of

Michael F. Barnsley and his co-worker Alan D. Sloan, makes the method
clear but provides little in the way of concrete details.
A set of simple equations and an iteration rule that uses them sends a
point hopping madly about the plane. If the point is plotted in each of its
successive positions, thousands of them in fact, a fractal image slowly
fills in. In this chapter readers will find an algorithmic description that is
complete enough to begin programming: you may float away on a fractal

cloud or ride a fractal wave. Such is the power of computers!
119
Figure 12.1 An IFS produces a fractal cloud.
Suppose you have a description of a fractal object (whether real or

fanciful) in the form of a map. The object will normally be composed of
parts, each similar to the whole. The first step in setting up an iterated
function system, or IFS, is to prepare by hand what Barnsley calls a "col-
lage." This is essentially a map of the object as a region which is divided
into sub-regions that are similar to the region as a whole. Consider, for
example, the upper collage shown in Figure 12.2. Here, the region that
represents the circular outline of a cloud as a whole has been subdivided
into six sub-cloud regions.
The next step in the preparation of the program is to set up what
mathematicians call affine transformations, one for each part of the
whole. In the case of the cloud and its six sub-clouds, six affine transfor-
mations are called for. Each transformation has three parts: A shrinkage
of the original cloud down to the size of the sub-cloud, a rotation of the
Color Plate 1 Early model of the Tinkertoy computer,
on display at The Computer Museum, Boston,
also plays tic-tac-toe.
Color Plate 2 Insectoid prototype Genghis strides over rock (above),
production model Hannibal awaits action (below).
Color Plate 3 Patterns demonstrate the amazing variety of
tur-mite behaviors
Color Plate 4 A zoo of alien creatures illustrates some simple
graphic tricks.
12 The Fractal Workshop 121
Figure 12.2 Working collage for a cloud.
sub-cloud, then a translation (shift) of the sub-cloud into its target posi-
tion.The effect of such a transformation is to move any point in the big
cloud into the corresponding point of the sub-cloud. Thus a point near
the top of the big cloud will end up close to the top of the sub-cloud.
To set up an affine transformation, you must first determine a rea-
sonable coordinate system, one that makes the transformations simple.
By placing the origin at the center of the big cloud, the transformations
are easy to obtain, especially if the exercise is carried out on graph paper.
At the end of this chapter, I will explain exactly how I obtained them.
The general form of an affine transformation is:
x'«-ax + by + e and
y'«-cx + dy + f
The values of a, b, c, d, e, and /are best presented in a table. For the
cloud the table has six rows, one for each sub-cloud in the collage. The
sub-clouds are labelled with capital letters that have nothing to do with
lower case letters just mentioned. For each sub-cloud, naturally, there is
a corresponding transformation that maps the whole cloud into that par-
ticular sub-cloud:
a b c d e f
A .643 0 0 .643 0 -2.5

B .08 -.199 .199 .08 -5.6 -0.4
C .303 -.189 .189 .303 -3.7 3.0
D .386 -.068 .068 .386 1 4.3
E .187 .754 -.754 .187 4.7 1.6
F .026 .213 -.213 .026 3.5 -1.7
It is two equations from each line of the

a simple matter to produce
table. Simply substitute the values of a, b, and e into the first equation
and the values of c, d, and /into the second one. I have used the algorith-
mic replacement symbol ( <— ) in place of an equals because that is how a
fractal-generating program must ultimately use these numbers. Inciden-
tally, I have placed the symbols x' and y ' on the left-hand sides of these
equations instead of x and y. The reason lies in the preservation of infor-

mation during a computation. If I used x and y instead of x' and y ', the
program would destroy the old value of x before the second equation got
to use it. Once both equations have been computed, the program may
replace old values by new values by using the following simple assign-
ments:
x <— x'
In any event, the coordinates (x, y) represent a point before it is

f
moved and the coordinates (x y') represent the point after a particular
,
transformation hasmoved it.

Once we have set up the transformations for an IFS, we may set a
point in motion by an iterative process; wherever the point happens cur-
rently to be, the process will plot it, then move it to a new location within
the cloud as a whole. In the collage that we will use, the sub-clouds are
represented by circles of different sizes. In most IF systems, the iterative
process sends a point in the object as a whole into a sub-object with a
probability that corresponds to the sub-object's area. In the case of the
cloud, it turned out (by calculating the area of each circle, then reducing
all the numbers proportionally until they added to 1) that the resulting
probabilities must be:
A: .47, B: .06, C: .15,

D: .18, E: .09, F: .05
Unfortunately, the use of these probabilities results in a cloud that

looks too bottom-heavy. To compensate, I greatly reduced the probabil-
ity of a point landing in the large sub-cloud at the bottom of the collage.
Giving cloud A a probability of .21 instead of .47, the result was defi-
nitely more cloud-like. Readers should feel equally free to tinker with
the probabilities to obtain different effects.
At the toss of an abstract coin, the program selects one of the sub-
clouds and maps the current point (x, y) into it by means of the appropri-
ate transformation. Each of the transformations shrinks the large cloud
down to the size of the sub-cloud, rotates it slightly one way or the other,
then translates it into the right position.
The basic IFS algorithm is simple and short enough to state right here
and now:
1. x «- 0, y «- 0
2. forn <- 1 to 20,000
Z <- random (A,B,C,D,E,F)
(x,y) «- Z(x,y)
if n > 10 then plot {x,y)
On this basis, readers may construct the program I call CLOUD. In

the first step, the variables x and y are set to 0. The second step involves
a loop that repeats the basic iteration process 20,000 times. Readers are
free to change this number if the images that result look either too dense
or too light.
Within the loop, I have condensed what will ultimately be many pro-
gram instructions into single lines. For example, the 'instruction" Z <—
'
random (A,B,C,D,E,F) really means, "Choose one of the six transfor-

mations A, B, C, D, E, or F based on their relative probabilities of .21,
.09, .22 and .27, .13 or .08." How is this done? Imagine a one-unit line
segment that is subdivided into consecutive lengths equal to these prob-
abilities. The first division point would lie at .21, the second at .30 and
the third at .52 and so on. In each case, merely add the current probabil-
ity to the ongoing sum.
If a random number between 0 and 1 is chosen, it must lie some-
where on the unit line. The segment that it happens to fall into will be the
basis on which the program CLOUD chooses which transformation to

apply next.
A simple algorithm for this process would involve six separate tests.
I have displayed the first three below so that readers will get the general
idea:
1. r <— random
2. if r < .21 then transform A
3. if r > = .21 and r < .30 then transform B
4. if r > = .30 and r < .52 then transform C etc.
Even this algorithm is not quite complete. What, after all, do the
commands "transform A," and mean? The structure that I have
so on,
chosen for CLOUD must expand each of these commands into little al-
gorithms of their own, algorithms that correspond, in fact, to the second
line of the loop.
To actually carry out transformation B, for example, requires two
steps:
1. x' <- .08* - .199y - 5.6

2. y'^.199x+.08y-.04
These instructions will look a little friendlier to readers who glance at

the second line of the IFS table on page 122: The numbers .08, — .199,
.199, .08, — 5.6 and — 0.4 appear in the second row, the one for trans-
formation B.
Each of the four algorithmic modules "transform A," "transform
B," and so on, must be translated into similar two-line algorithms, each
followed by the assignments of x' and y' back into x and y.
In the final stage of the loop, the point (x, y) must be plotted. This is
not quite as simple as saying, "plot (x,y)." Indeed, the programmer of

CLOUD must use an appropriate screen coordinate system. Suppose
that your display screen happens to be 200 pixels wide and 100 pixels
high. The coordinate system used in the four transformations described
places the origin in the middle of the cloud. But in most computer
screens, the origin lies on the left side of the screen, either at the top or
the bottom. How
do you "plot" the point (x, y)?
The answer lies in using two systems of coordinates, one for compu-
tation of the cloud, the other for display purposes. The first system of co-
ordinates is used for a virtual cloud, so to speak, that we never see. The
second system of coordinates is used strictly for display purposes. Thus,
every time a new point (x,y) in the virtual cloud is computed by the pro-
gram, it must convert these coordinates into a new point (w,z) in the dis-
play cloud.
The virtual cloud happens to be 14 units (of graph paper) in diame-
ter.To create a display cloud that is, say 112 pixels in diameter, the vir-
tual coordinates must each be multiplied by 8. But that is not enough.
They must also be translated from the left edge of the screen over to the
middle. The following statements will handle the case of a 20- by-100
screen that has its origin in the upper left corner. The variables w and z
represent the display coordinates.
1. w <- 8x + 100
2. z «- 8y - 50
This completes the description of CLOUD. I will not claim for a mo-
ment that I have selected the best possible collage for my cloud. Others
may find a better arrangement of sub-clouds that fill the overall shape in
a more natural or aesthetically exciting fashion. In general, the more
open space within the collage that is not covered by sub-figures, the
more open and diffuse the resulting graphic image will be.
Figure 12.3 IFS-generated fractal wave.

The second IFS fractal represents the element water. This time I will
use not a two-dimensional collage but a one-dimensional one. Figure
12.4 shows a pair of spirals joined together at their apexes to produce a
shape that, overall, looks somewhat like a wave. The shape may be re-
garded as a one-dimensional "region" that may be sub-divided into
spheres of influence for three separate transformations, A, B, and C.
The first transformation, A, maps the entire double-spiral wave into
a smaller version of itself still centered on the apex. This transformation
not only shrinks, but it rotates the figure so that it still fits the smaller
portion of itself. The second transformation maps the entire wave into
the small wavelet labelled B Readers will see at once that
in the figure.
this is simply a smaller version of the entire double spiral placed at the
trailing edge of the large wave.
But what about the other half of the wave, the trough into which it
threatens to crash with a fractal roar? The third transformation, C, sim-
ply rotates the entire figure by 180 degrees so that whatever gets gener-
ated on one side of the image has a chance of being mapped over to the
other side in order to complete the picture. In fact, because the right half
of the wave has half the area, the third transformation gets half the ac-
tion. The probability with which the program I call WAVE uses transfor-
Figure 12.4 Working collage for a wave.

a b c d e f p
A .606 .35 -.35 .606 0 0 .35
B .207 -.174 .174 .207 -7.9 1.6 .15
C -1 0 0 -1 0 0 .50
mation C is exactly .5. The table above shows the IF system I developed
to produce the wave image. Readers are free to write their own version
of WAVE, using these equations or others they might wish to derive for
themselves.
This raises the question, of course, "Just how do you find those
equations?" Insofar as they involve only contractions, rotations, and
translations, they can be calculated on the basis of measurements taken
directly from the figure you have drawn on graph paper.
Consider, for example, the subcloud B in the cloud collage in Figure
12.2. Its diameter, as measured it on squared paper, was 3 units while
I
the cloud as a whole had a diameter of 14 units. This meant that the pro-
gram would have to shrink the large cloud by a factor of 3/i4 or .214.
Next, I decided that the small cloud should reproduce the whole cloud at
an angle of 68 degrees in the positive sense. This would turn it well to the
left so that the same surface would be reproduced in the cloudlet at the
part where it seemingly emerged from the main cloud.

The formulas for a, b, c, and d are based on the theory of affine trans-
formations which, luckily, needn't concern us here. As it happens, sim-
ple trigonometric functions do all the work:
o = .214cos(68) b= .214sin(68)
c = .214sin(68) d= .214cos(68)
In each case, the contraction factor is multiplied by the cosine or sine

of the angle of rotation. we calculate these numbers on a hand calcula-
If
tor, the first four numbers in row B of the cloud's IFS table emerge.
The translations, e and /, turn out to be especially easy to calculate.
The shrunken and rotated sub-object has the same relation to the origin
as the whole object does. Where does the origin go when the sub-object
is at last translated to its final resting place? The horizontal distance is e
and the vertical distance is /.

By now, some readers will be pondering some elementary creations

of their own. How to make a mountain or fire? Should a two-dimen-
sional or one-dimensional collage be used? The answer to the second
question depends on whether the fractal appearance aimed at is essen-
tiallyone- or two-dimensional. Take mountains, for example. I can
imagine a mountain in outline as a series of sub-mountains that mount to
the summit. I can also imagine it as a more or less solid object with a
fractal outline.
As for imagine a collage with a large flame composed of three
fire,
subflames, two at the base of the flame and one at the top.
This much completes the excursion into the elements of IFS sys-
tems, not nature's workshop, but your own.
Further Reading
Michael F. Barnsley. Fractals Everywhere. Academic Press, 1988.
Benoit B. Mandelbrot. The Fractal Geometry of Nature. W. H. Freeman, 1983.
THEME THREE
Mathematics Matters
Mathematics, once described by the great

popularizer, Eric Temple Bell, as "the queen
and servant of the sciences," displays some
tiny perfect gems from her crown jewels in the
first chapter of this suite. The spheres of
Dandelin provide a startling proof that points
on an ellipse enjoy the equidistance property:
the sums of their distances from the two foci
of the ellipse are all equal! Other gems illus-
trate other proofs, laying bare the real motiva-
tion of those who become mathematicians. To
know and appreciate the process of proof is to
have the ear and speech of the queen herself,
to hold a few of her crystalline secrets in your
hand.
Golygon City and the art of cat-scanning
illustrate the search for truth as most mathe-
maticians experience it. Conjectures are made,
then either proved or disproved. It astonishes
most people, including mathematicians, that
results that seem to apply only to some ab-
struse or toy system can suddenly turn out to
have important, even life-saving applications.
Cat scans turn this situation around by show-
ing how image reconstruction gives rise to
interesting mathematics even when there are
only a handful of x-rays available.
A whole branch of mathematics, the theory
of rigidity, addresses the problem of determin-
ing when an arbitrary structure of rods and
joints is rigid and when it will flex. A bracing
excursion into the problem of making cubic
and rectangular frameworks rigid illustrates
real problems faced by rigidity theorists and
the theories they have discovered to solve
those problems.
A book of this nature cannot stray long
from the computational realm. Mathematical
thinking has long been thought to comprise an
important area of human IQ. Whether the
term "intelligence quotient" has any definite
meaning, IQ tests certainly abound with math-
ematical examples. Can an IQ test-writing
program succeed? This one completes se-
quences and solves analogy problems with
frightening success.
To end with a program that solves IQ test
problems merely serves to introduce the next
section, where computers enter the arts.
s
T>
JL %oss Honsberger has spent more than two decades collecting
mathematical morsels for general consumption. Attending one of his
rare public lectures, I had the pleasure of sampling one of his mathemati-
cal feasts. I found it not just palatable but downright delicious.

Honsberger, who teaches mathematics at the University of Water-
loo in Canada, will stop at nothing to demonstrate mathematical princi-
ples. To illustrate a topological tidbit one day, he wandered into a col-
league's lecture wearing his trousers inside out. He announced that he
would turn them right side out without really taking them off.
To show that the trousers would not "really" be removed, he tied his
own ankles together with an eight-foot length of rope. He proceeded to
pull both trouser legs down onto the rope, revealing his favorite heart-
spotted undershorts. The students watched closely for any deception as
he first twisted the pants through the rope, then turned them right side
out and finally wriggled back into them. I can testify that Honsberger'
trick requires neither magic nor mirrors.
I listened to Honsberger's lecture near my hometown in Ontario. He
served to the audience such delights as spheres in a cone, checkers on a

board, dots on a dish and beans in a Greek urn.
Honsberger began by describing the marvelous spheres of Germinal
Dandelin, a 19th-century Belgian mathematician. Dandelin discovered
131
132 Theme Three Mathematics Matters
an amazing connection between the classical and modern concepts of

the ellipse. The Greeks conceived of an ellipse as the figure that results
when a plane cuts obliquely through a cone. Since the time of Descartes,
however, the ellipse has been described analytically in terms of two spe-
cial points called foci. The sum of the distances from the two foci to any
point on the ellipse is constant.
Honsberger introduced Dandelin's spheres by drawing our attention
to a projected transparency of a plane cutting a cone. (Readers can follow
Dandelin's argument with occasional glances at Figure 13.1.) I cannot
swear that what follows are Honsberger's exact words, but he readily
admits to a certain, broad similarity:
"It takes no genius to see that the plane divides the cone into two
pieces. But it was Dandelin's idea to insert a sphere into each piece. Like
an over-inflated balloon, each sphere contacts the wall of the cone and
touches the elliptical plane at a certain point. But where? One can imag-
ine Dandelin's heart leaping at the thought that the two spheres might
touch the plane at the two focal points of the ellipse."
Honsberger places his marking pen on the transparency. He labels
the two points of contact by the symbols F and G. Are these the foci of
the ellipse?
Figure 13.1 An ellipse in a cone separates Dandelin's spheres.

"Let's take a look atwhat clever old Dandelin did. First, through any
point P that we care to on the ellipse, we may draw a straight line
select
that runs up the side of the cone to its tip. Second, the line will touch the
two spheres at two points, say, A and B. No matter where we pick P to lie
on the ellipse, the length of AB will be the same.
"Ah, but that gives it away! The distance from the point F to P equals
the distance from P to A. After all, both PF and PA are tangents to the
same sphere from the same point. By the same reasoning, the distance
from the second point, G, to P equals the distance from B to P. Are we
not finished? PF + PG = PA + PB, and the latter sum is just the (con-
stant) length of AB.
"Now isn't that the darndest thing?" Honsberger sounds rural.
As I look around the lecture hall, students appear stunned. Profes-
sors alternately smile and frown. One of them behind me murmurs,
"Well, I'll be."
Without pausing for a breath, Honsberger serves up the next morsel.
We find ourselves staring at a slide of a peculiar board game.
"Here's a simple little exercise in checker-jumping. I imagine that
such a clever audience will have no trouble figuring this one out." A dev-
ilish gleam invades his eye, a warning that something unusual is about to
happen.
The slide shows a grid of squares with a line drawn through it (see
Figure 13.2). Honsberger explains the rules: solvers can arrange a given
number of checkers any way they like behind the line. Checkers can be
jumped and removed in the vertical or horizontal direction, but the final
jump can leave only one checker. The problem is to decide how many
checkers it will take, at a minimum, to "propel" the last checker a given
distance d beyond the line. In deciding this, solvers must also devise an
arrangement that allows the frenzy of jumping to take place.
"Does anybody want to take a crack at this problem?" Honsberger
looks up at the audience, sees there are no takers and smiles ingen-
uously.
"Well, then, let's try a few examples."
He places two
checkers adjacent to each other on the board, jumps
the back checker over the front one, then strides triumphantly to the
blackboard, where he writes
d number of checkers
1 2
-—
—
J
-
1
Figure 13.2 Four checker-jumping arrangements.
Next he places four checkers on the board. He jumps the checkers until a
single checker is left in the second row.
"I'm just saving you people the trouble of figuring it out. Believe me,
this is the best that anyone can do." He writes "2" under "d" and "4"
under "number of checkers." He creates another configuration of eight
checkers and manages to propel one checker to the third row. He scrib-
bles "3" and "S" on the blackboard.
"Anybody want to guess how many checkers it takes to send one
checker four units beyond the line?"
Somebody volunteers the figure 16. No. The answer turns out to be
20 checkers, at a minimum.
The audience now getting somewhat worked up. Could the rela-
is
tionbetween distance and the number of checkers be described by one

of those superexponential functions? Perhaps it will take a million
checkers to send one checker five units beyond the line. When Hons-
berger reveals the answer, members of the audience look at one another,
smiling uncomfortably.
"Alas! A million checkers will not be enough, nor a billion. It is sim-
how many checkers you assemble behind the
ply impossible, no matter
line orhow you arrange them. It was John Conway, the Cambridge
mathematician, who proved the task impossible."
Hh
1 1
— •
••
1
• #
•
•
H
Honsberger does not stop the lecture to describe Conway's difficult
proof, although he would not dream of discouraging anyone from at-
tempting it. Instead he quickly moves on to a discussion of the pigeon-
hole principle.
This famed principle simply states that if I build 9,999 pigeonholes
for 10,000 pigeons, at least one of the holes would house more than one
bird. The pigeonhole principle has been used to prove many theorems in
combinatorics, the branch of mathematics that deals with finite collec-
tions.
"The next mathematical morsel is one of the strangest applications
of the pigeonhole principle ever made. Imagine that someone has placed
650 points inside a circle of radius 16 units. You have been given an an-
nulus, a flat ring in the shape of a washer. The outside radius of the
washer three units, and the inside radius is two. You are then chal-
is
,,
lenged to place the washer so that it covers at least 10 of the 650 points.
"Impossible," whispers an impetuous undergraduate behind me.
"What if all the points are in a tiny area?"
"Then he can cover all of them with one washer, you idiot\" retorts
another student.
Is it really possible to cover 10 points with the washer? Honsberger
begins the proof by drawing a diagram shown in Figure 13.3. He invites
Figure 13.3 The washer problem.
us to imagine that a copy of the washer has been centered at each of the
650 points inside the circle.
Some of the points may be near the edge of the circle, in which case
some of the annuli will extend beyond its circumference. But because
each point lies inside the circle and because the washer has radius three,
all the annuli will lie within the larger circle having the same center and
having radius 19, that is, the sum of 16 and three. The area of the washer
is the difference between the area of a circle of radius three and one of
radius two. This comes to five times n.
"The 650 washers blanket the large circle with a total coverage of
650 times 5n, that is 3,250tt. Of course, much of the coverage will be
overlapping, but suppose for the moment that no point of the inner circle
is buried under more than nine washers. In such a case, the total amount
of area covered within the larger circle could not come to more than
3,249ft, nine times the area of the circle. But because 3,249ft is less than
3,25071, some point, x, must be covered by at least 10 washers. The pi-
geonhole principle strikes again."
Honsberger pauses to catch his breath. "You see it now, don't you?"
Then he feigns surprise. "You don't?"
The application of the pigeonhole principle is clear enough, but
some of us are confused. We thought the idea was to cover 10 points by
one washer, not to hide some point x under 10 washers. Suddenly, our
minds are turned inside out like a pair of trousers.
"Look at the point x. If you take away the 10 washers that cover it
and replace these by a single washer centered at x, then that washer
alone must cover the centers of the 10 washers that we took away. Each
of these centers is one of the 650 original points!" The morsel is digested
as we hear a faint gulp from somewhere at the back of the lecture room.
The piece de resistance of Honsberger's menu arrives in the form of
a Grecian urn adorning his next transparency (see Figure 13.4). "How
much can a Grecian earn?" quips Honsberger.
When the groans have died away, he explains the problem. An urn is
filled with 75 white beans and 150 black ones. Next to the urn is a large
pile of black beans. The beans are removed from the urn according to
certain rules.
Figure 1 3.4 What is the color of the last of 75 white beans and 150 black
ones in the urn?
"Here's how it works. Remove two beans from the urn at random. If
at least one of the two beans is black, place it on the pile and drop the
other bean, whether white or black, back into the urn. But if both of the
removed beans are white, discard both of them and take one black bean
,,
from the pile and drop it into the urn.
"Each time a Greek or anyone else dips into the urn to remove two
beans at random, either operation ensures that there will be one fewer
bean in the urn that there was before the move. Slowly and steadily, the
original supply of black and white beans dwindles. At last there are three
beans left in the urn, then two, then one. What color is the last bean?"
The simple and startling answer is white. By figuring out why, a
Greek can earn intangible delights worth more than a hill of beans.
Further Reading
Ross Honsberger, ed. Mathematical Gems: From Elementary Combinatorics,
Number Theory, and Geometry. Dolciani Mathematical Expositions, No. 1.
Mathematical Association of America, 1973.
Ross Honsberger, ed. Mathematical Morsels. Dolciani Mathematical Exposi-
tions, No. 3. Mathematical Association of America, 1979.
Ross Honsberger, ed. Mathematical Plums. Dolciani Mathematical Expositions,
No. 4. Mathematical Association of America, 1979.
—— -
km City
A journey of a thousand miles must begin with a

single step.
LAO-TZU, The Way of Lao-tzu
JL ^llow me to start you on a journey in Golygon City. You can take a

similar trip in New York, Kyoto, or almost any large city whose streets
form a grid of squares. Here are your directions. Stroll down a city block,
and at the end turn left or right. Walk two more blocks, turn left or right,
then walk another three blocks, turn once more and so on. Each time
you turn, you must walk straight one block farther than you did before. If
after a number of turns you arrive at your starting point, you have traced
a golygon. If you do not need the physical exercise, you can easily simu-
late the journey by moving a pencil along a piece of graph paper with a
square grid. If you become lost, you may refer to Figure 14.1.
A golygon consists of straight-line segments that have lengths (mea-
sured in miles, meters, or whatever unit you prefer) of one, two, three
and so on, up to some finite number. Every segment connects at a right
angle to the segment that is one unit larger —
except the longest seg-
ment, which meets the shortest segment at a right angle. Golygons are
more than just a geometric curiosity. They have inspired some delightful
puzzles as well as some intriguing problems for research. What better
way to develop insight into the research process than to take a recre-
ational journey?
It seems that golygons were first conceived by Lee Sallows, a re-
doubtable engineer of the Catholic University Nijmegen in the Nether-

lands. I first featured Sallow's work in The Armchair Universe (see Fur-
ther Reading) where I described his search for pangrams, sentences
139
Figure 14.1 A map of Golygon City.
containing each letter of the alphabet. Since then Sallows has invented
many new recreations butnone so engaging as golygons.
In the fall of 1988 Sallows began his search for golygons. It did not
take him long to find an eight-sided golygon. Yet he could find no such
objects that had fewer than eight sides. Nor did he discover a golygon
with nine sides, nor 10, nor 11, . . until at last he stumbled on a 16-
.
sided golygon.
Wondering whether he had missed any, Sallows wrote a computer
program to automate the search. It spewed out no less than 28 different
16-sided golygons (see Figure 14.2) before moving on to higher orders.
The program did not find golygons with from 17 up to 23 sides, but it
generated numerous 24-sided golygons —
2,108 to be exact.
By then Sallows had a hunch that the number of sides in a golygon
must be a multiple of eight. Yet his program, which was already laboring
in the 20's, was unequal to the task of discovering any 32-sided goly-
gons. Frustrated, Sallows appealed to Martin Gardner, the supreme au-
14 Golygon City 141
Figure 14.2 All 28 golygons that have 16 sides.

thority in matters of recreational mathematics. Could Gardner prove

that the number of sides in a golygon must be a multiple of eight?
We can creep up on Gardner's proof by starting with the simpler task
of proving that the number of sides must be a multiple of two. To ex-
plain, allow me to take you back to Golygon City. We start our journey
by walking one block to the north. (I choose north only for conve-
nience.) Hence, the first side of the golygon will run north for one block.
Next we can turn either left or right, west or east. Therefore, the
that is,
second side of the golygon will run either two blocks to the west or two
blocks to the east. It follows that all odd sides in the golygon (the first,
third, fifth, etc.) are an odd number of blocks long and run north or
south; all even sides (the second, fourth, sixth, etc.) are an even number
of blocks long and run east or west. Because the last side of the golygon
meets the first at right angles, the last side must run east or west. There-
fore, the last block is an even side of the golygon, and the number of
sides in a golygon must be a multiple of two.
To show that the number of sides must be a multiple of four, we
begin by calculating the total distance we traveled to the north of the
starting point. To do this, we simply add the number of blocks we walked
to the north and subtract the number of blocks we traveled to the south.
(A net negative distance to the north simply means that we are south of
the starting point.)
Because all north and south sides are an odd number of blocks long,
we are essentially adding and subtracting consecutive odd numbers. It
takes only a little tinkering to convince ourselves that the result is always
even when we add or subtract an even number of odd lengths for ex- —
ample, 1+3— 5 + 7 = 6. By the same token, an odd number of odd
lengths always gives an odd sum. Therefore, the distance walked north is
an even number of blocks if and only if we have walked along an even
number of north and south sides.
Now if we walk from the starting point around Golygon City and re-
turn, the distance to the north of the starting point equals zero. Because
zero is an even number, the golygon must have an even number of north
and south sides. The total number of sides is twice the number of north
and south sides, because for every north or south side there must be an
east or west side. Therefore, the number of sides in a golygon is a multi-
ple of four.
How on earth did Gardner show that the number of sides must be a
multiple of eight? Let us accompany the master of his journey.
14 Golygon City 143
The horizontal sides of a golygon may be divided up into an east set

and a west set. The sums of numbers in each set are even and equal. Sup-
pose for a moment that a golygon has n sides, where n is a multiple of 4
but not 8. If you add up the horizontal legs in such a golygon, you will
obtain a number in the sequence 6, 42, 110, 210, 342. But then either
the sum of east sides or the sum of west sides (it doesn't matter which)
will turn out to be a number in the sequence 3, 21, 55, 105, 171, and so
on. All of these numbers are odd, contradicting the fact that the sums of
the east and west sides are both even. The only way to get out of this dif-
ficulty is to realize that n must be a multiple of 8.
Gardner's proof may seem rather peculiar because even numbers
played a major role and odd numbers hardly entered into the discussion.
The same conclusion can be reached, however, by making a similar ar-
gument on behalf of the odd lengths. In fact, the two kinds of lengths
play independent roles in the construction of any golygon. One may ob-
tain, for example, a new golygon from two others of the same size by
using the even lengths from the first golygon and the odd lengths from
the second.
Our recreational path has so far taken us from the question of exis-
tence to the question of conditions on that existence; the number of
sides in a golygon must definitely be a multiple of eight. For convenience
I will call this theorem the Sk condition for two-dimensional golygons. It
is what mathematicians call a necessary condition. If a golygon exists,
the number of its sides is necessarily a multiple of eight. But is this con-
dition sufficient? In other words, if we merely ask that the number of
sides, n,be a multiple of eight, is that sufficient to guarantee the exis-
tence of a golygon?
We can find a golygon that has eight times any number of sides by
applying a set of simple instructions. The first step is to decide how large
a value of n one wants. To illustrate the method, I will choose a 16-sided
golygon. step is to write down consecutive numbers from 1
The second
to n, in this case,from 1 to 16.
The third step would have us place plus and minus signs in front of
the numbers. The first quarter and the last quarter of the numbers will
receive plus signs, and all the numbers in between will receive minus
signs:
+ 1, +2, +3, + 4, -5, -6, -7, -8, -9,

-10, -11, -12, +13, +14, +15, +16.
To translate this sequence of numbers into a tour of Golygon City, you

need only know that positive odd numbers run to the north, negative
odd numbers to the south, positive even numbers to the east and nega-
tive even numbers to the west. The golygon that corresponds to these
numbers looks like a snake. A 32-sided "snake" is shown in Figure 14.3.
Having settled the question of sufficiency of the 8k condition,
Gardner and Sallows began to wonder just how many golygons should
exist. They had found one of order eight, 28 of order 16 and 2,108 of
order 24. How many would there be of higher orders? Such questions
are most frequently asked by mathematicians who specialize in combin-
atorics, the study of discrete mathematical objects such as sets and
graphs. Gardner decided to check with some experts. He wrote letters to
computer scientist Donald E. Knuth of Stanford University and to math-
ematician Richard K. Guy of the University of Calgary in Alberta. Could
either of them assist in the matter of enumerating golygons?
Before long, Knuth had written a computer program that counted all
the golygons up to 64 sides. The total number of 64-sided golygons, for
instance, is 127,674,038,970,623. Unfortunately, Knuth's program be-
came unwieldy at larger sizes.
Meanwhile Guy was able to develop a formula that expressed the
approximate number of golygons for each possible value of k, where k
equals the number of sides divided by eight. The formula consists of a
fraction, the principal parts of which are 2 raised to the power 8k — 6 in
the numerator and k cubed in the denominator. Because, as k increases,
the exponential term increases much more rapidly than the cubed term,
the growth and the number of golygons is basically exponential. Guy's
formula has the further important property of being asymptotic, that is,
it gives a value that, expressed as a percentage of the true value, comes
increasingly close to 100 percent as k gets larger and larger.
Figure 14.3 One golygon snake that has 32 sides.

14 Golygon City 145
Sallows, Gardner, Knuth, and Guy decided to report their various

results in a paper. The word "golygon" does not appear in the title. In-
stead it is 90 Degrees/ a name that rings more
called "Serial Isogons of '
solemnly in the halls of mathematics than "golygons." The reference to

90 degrees implies that other kinds of golygons might be assembled. In-
deed, some of the authors have looked at golygons in which every angle
is not 90 degrees but 60 or 120 degrees. To construct such golygons,
readers may need hexagonal-grid paper, which can be found in most

graphics-supply stores.
Would have been a good place to stop? Research (like recre-
this
ation) never stops. Often, as research in one question progresses, other
questions suggest themselves. Sometimes it is very hard to decide just
what question to work on.
f3f%
Figure 14.4 The smallest golygon tiles the plane.

Research problems are no sooner solved than the recreationist in us

thinks of new avenues of inquiry: Can we find prime-sided golygons?
The lengths of consecutive sides here increase by the sequence of odd
primes: 1, 3, 5, 7, 11, 13 and so on.
Another offshoot of golygons deserves mention. Readers are urged
to look at the eight-sided golygon once again. It is unique, the only mem-
ber of its class. Eight-sided golygons will fit together very nicely to make
tiling patterns. They will, as the professionals say, "tile the plane." Fig-
ure 14.4 shows an attractive tiling.
Properly speaking, the region bounded by an eight-sided golygon
should be called a polyomino. Not all golygons produce polyominoes
because some golygons intersect themselves in their meandering march
around the plane. In fact, polyominid golygons (a formidable expression
by which readers can lose friends and influence people) are probably in-
creasingly rare in relative terms as n gets large. Only four of the 28 goly-
gons shown in Figure 14.2 are the boundaries of polyominoes. None of
these appears to tile the plane.
But in the matter of tiling, Sallows has a challenge for readers: sup-
pose the eight-sided golygon represents the boundary of one's kitchen
floor. Given 13 L-shaped tiles, can readers cover the kitchen floor ex-
actly with these eccentric tiles? For the purpose of this question, the
FOUR TEE
W|E|L|V|E
Z E m E
fio|u|r|e1
F
vl
T IE
TgThTt
Figure 14.5 A logolygon crated by Sallow.

14 Golygon City 147
floor may be thought of as divided into 52 squares, like the squares of the
special paper on which this enquiry started. Each square is one step on a
side. The L-shaped tiles consist of three squares in a row and an addi-
tional square comprising the foot of the L. Anyone who finds the task im-
possible must, of course, prove it to be so.
I close with some final words from Sallows, the engineer who started
all this. The "words" are ZERO, ONE, TWO and so on, up to FIFTEEN.
They can be arranged in a special golygon all their own in which the let-
ters of these words determine the sides of what can only be called a logo-
lygon (see Figure 14.5).
Further Reading
A. K. Dewdney. The Armchair Universe: An Exploration of Computer Worlds.
W. H. Freeman, 1987.
Lee Sallows, Martin Gardner, Richard Guy, and Donald Knuth. "Serial Isogons
of 90 Degrees." Mathematics Magazine, Vol. 64, No. 5 (Dec. 1991), pp.
315-324.
Scanning
This time (the Cheshire Cat) vanished quite slowly,

beginning with the end of the tail, and ending with
the grin, which remained some time after the rest of
it had gone.
Lewis Carroll, Alice's Adventures in Wonderland
T
a
JL he medical practice called CAT scanning herein lends its name to
new recreationin which readers attempt to reconstruct unknown ob-
jects on the basis of what might be called digital X rays, or D rays. A
Cheshire cat may not be reconstructive from its grin, but some cats can
be reconstructed from their D rays.
In 1979 Allan M. Cormack and Sir Godfrey N. Hounsfield shared
the Nobel prize in medicine for the invention of the CAT scanner. Short
for computed axial tomography, the machine combines a computer and
a series of X-ray "slices" of a patient to reconstruct a cross section of his
or her anatomy, replete with bones, vessels, and other tissues.
fat,
An ordinary X-ray machine creates a two-dimensional image that is

akin to a shadow. In sunlight, my shadow on the wall shows clearly an
outline of my body. But in the "light" of an X-ray beam, my outline is
filled in by bones and other internal organs. CAT scans are not like this at
all. Here the X rays do not spray everywhere through my body. Instead
they are confined to a single plane. On the other side of my body they
produce what can only be called a one-dimensional shadow, a strip of
dark and light that betrays the presence of bones and organs only within
that plane and only as seen from that angle.
As shown in Figure 15.1, a medical technician positions a patient in-
side a CAT scanner so that an X-ray source may project its beam through
the patient within a specific plane. On the other side of the patient a row
148
Figure 15.1 A patient inside a CAT scanner.
of detectors records the radiation that went through the patient along
that particular angle. From here, the one-dimensional shadow is relayed
in digital form to a computer. All of this conspires to make but one snap-
shot, so to speak.
The next snapshot is taken from a slightly different angle. The X-ray
detector and source are rotated a bit, and a new beam of X rays slices
through the patient within the same plane. The new array of readings is
also sent to the computer.
When enough snapshots have been taken, the computer begins to
reconstruct the patient's cross section in the plane of all the X rays. How
is thisdone? It hardly takes a Nobel laureate to understand it. In fact, we
shall do precisely the same kind of thing as the computer on a humbler
scale. Rather than asking readers to spend half a million dollars on a CAT
scanner, I shall provide some recreational CAT scanning that requires
nothing more than pencil and graph paper.
The recreational version I call simply the catscanner reconstructs
two-dimensional images composed of small black-and-white squares in
an image grid as in Figure 15.2. For those who see no challenge in this
exercise, let me hasten to add that the image is normally not provided,
only the information from the so-called D rays.
The image is traversed, from two directions only, by D rays. A beam

of horizontal D rays penetrates the object from right to left and is re-
corded as a sequence of numbers. Each number is the sum of the black

squares in a particular row of the image grid. A beam of vertical D rays
streams through the object from bottom to top. The beam is also re-
corded as a sequence of numbers with the same interpretation, namely,
the number of black squares, one number per column. In Figure 15.2,
for example, the number beside the second row is 2 because the second
row contains two black squares. Each black square adds one to each
number. From such a simple observation springs an important fact for
cat scanners: the sum of the row readings equals the sum of the column
readings.
Theoretical niceties aside, can the image be reconstructed from just
two beams? The answer depends on the particular pattern of black-and-
white squares that makes up the object. For convenience, I will describe
those patterns that we may logically reconstruct as "categorical" and
those that cannot be constructed as "catastrophic."
Before I get too involved in what characterizes a catastrophic pat-
tern, I offer an example that will prove to be categorical. The example is
the spiral shown in Figure 15.2. After the catscanner has bombarded the
spiral pattern with D rays, the sensors next to the rows reveal the se-
quence 8,2,6, 4, 5, 3, 7, whereas the sensors over the columns read 7, 1,
6, 3, 4, 5, 2, 7.
To reproduce the spiral from the two sequences, the catscanner
could generate every combination of black-and-white squares until it
found a pattern that would produce the original sequences. Luckily, we
can develop a better technique to reconstruct the spiral.
In any set of sensor readings derived from a categorical pattern, one
of the readings in one of the sequences will always match the number of
nonzero readings in the other sequence. For the spiral pattern, the sen-
sors next to the rows indicate seven nonzero readings, and the sequence
associated with the first column contains a "7." It can therefore be de-
duced that the first column contains seven black squares in all seven
rows.
These seven black squares contribute one unit to each of the read-
ings associated with the rows. To determine the positions of other black
squares, therefore, you change the 7 to a 0 in the column sequence and
subtract one from each nonzero reading in the row sequence. Two new
sequences result:
— —
0 0 7 1 6 3 4 5 2 7 0 0
t t t t t t t t t t t t
0 —
0
8 —
2 —
6 «4
i 1
5
^_
M—
''"'it —
3 M—
7 —
0
0
—
Figure 1 5.2 Catscanner readings of a categorical spiral.
Rows: 7, 1,5, 3, 4, 2, 6
Columns: 0, 1,6, 3, 4, 5, 2, 7
Notice that at least one reading in one of the sequences is equal to the
number of the nonzero readings in the other sequence. In fact, because
there are more than one, readers may take their pick of which one to
work on next. The final outcome will be the same. If, for example, you
pick the 7 in the first sequence, fill in the appropriate squares and carry
out the arithmetic, the next sequence is:
Rows: 0, 1, 5, 3, 4, 2, 6
Columns: 0, 0, 5, 2, 3, 4, 1, 6
Once all the numbers have been reduced to zero, the picture of the spiral
will be staring you in the face.
The catscanner does not decipher all patterns as easily as it does the
spiral.A catastrophe would occur if two patterns that produced the same
sequences of numbers were placed in the catscanner. By trial and error,
you can probably come up with a few examples.
If you examine catastrophic patterns in a systematic way, however,
you may notice that all such patterns have features that I call chains.
Suppose that two different patterns —
call then A and B —
produce the
same sequences of numbers (see Figure 15.3). Then pattern A must have
the same number of rows and columns as pattern B. If pattern A is su-
perposed on pattern B, there must be at least one black square in A that
overlaps a white square in B. (Otherwise the patterns would not be dif-
ferent.) This square is the first link in the chain.
one square were the only difference between patterns A and
If this
B, however, the sensor readings of A and B would not agree. Therefore,

the row that contains the first square must also include a second square
that is white in A and black in B. The second square is the second link in
the chain. Likewise the column that contains the second square must
also include a third square that is black in A and white in B. This is the
third link. A similar procedure can be repeated to find other links in the
chain. Ultimately, the chain must be linked back to the first square, guar-
anteeing that there are no disagreements in the readings of A and B.
A study of catastrophic patterns leads to two theorems. The first
states that a catastrophic pattern can be transformed into another by
changing the color of each square in its chain. Such a transformation
might be useful to describe ambiguous reconstructions. Indeed, the sim-
ple technique for reconstructing categorical patterns can be modified to
partially reproduce catastrophic patterns. I invite readers to attempt this
modification.
The second theorem says that a pattern is categorical if, and only if, it
has no chains. Using the theorem, I wonder how many readers can now
justify the method I gave earlier for reconstructing categorical patterns.

How can I always be so confident of finding a number in one of the se-
quences that counts the nonzero numbers in the other sequence? A hint
shall suffice: What would be the consequence for the pattern if no such
numbers existed?
The second theorem can also be applied to determine the number of
categorical patterns of a given size. For instance, although 512 patterns
can be created by coloring the squares of a three-by-three grid, only 99
of them are categorical. The three-by-three categorical patterns are
0 4 5 4 6 5 6 4 4 5 3 7 0
4 4 4 4 4 4 4 4 4 4
1. 1 t
10
a
o
.
o ;
J
10 1: 1 1 i
1
11
1
0 4 5 4 6 5 6 4 4 5 3 7 0
t t ,i
I
J 1i .1. .L I ,1 J, L
0
1
6
2 f - =
10 I i
A-
3 i
T
7 1
10 I I
11
I I 1 I
Figure 1 5.3 Patterns A and B and a chain that converts one into another.
shown in Figure 15.4. To save space, I have omitted those patterns that
are reflections or rotations of others.
Readers who have access to computers can write a simple program
that simulates the catscanner. The program I call grin (Figure 15.5) then
stores the readings in two arrays that I cannot refrain from calling x-
array and y-array. Next the program requests the number of non-zero
entries in the two arrays and stores them in xcount and ycount.
GRIN searches through the arrays of catscanner readings until it finds
the number xcount in the y-array or the number ycount in the x-array.
The program carries out the appropriate arithmetic, draws the corre-
sponding squares on the screen and then sets a special variable called
done to 1. The last step prevents the program from repeating the pre-
vious instructions until it has had a chance to recompute xcount and
ycount. When xcount reaches 0, the non-zero numbers have all been pro-
cessed, done is set to 0 and the computer takes a catnap. A somewhat ab-
breviated version of the algorithm is given above.
Here are some on your catscanner.
categorical patterns to try out
Blending the medical with the recreational for a moment, the following
categorical pattern was recently found in a CAT scan of a cat. What had it
eaten?
input x-array and y-array

input xcount and ycount
repeat
done <— 0
for each array position I
if x-array (I) =
ycount and done = 0
then x-array (I) <— 0
for each y-array (J) > 0
decrement y-array (J)
color square at (I,J)
done <— 1
if = xcount and done
y-array (I) = 0
then y-array (I) <— 0
for each x-array (J) >0
decrement x-array (J)
color square at (J,I)
done <— 1
compute new xcount and ycount
until xcount = 0
Figure 15.5 Outline for the GRIN program.
Rows: 9, 5, 7, 5, 6
Columns: 5, 3, 5, 0, 5, 2, 5, 0, 1, 5, 1
Recently someone brought me a box that he could not open, having

lost the key. Could I determine the contents of the box? A catscanner re-
vealed an object with an ambiguous reconstruction, but I immediately
knew what it was.
Rows: 1 1, 5, 4
Columns: 3, 2, 3, 1, 1, 1, 1, 2, 3, 2, 1
Finally, here is the grand catscanning challenge of The use of

all.
square grids would appear to limit the reconstruction game to just two
sets of D rays, one horizontal and the other vertical. But what about diag-
onals? If diagonal D rays are permitted, four sets of D rays are possible,
and the class of objects that can be reliably reproduced is dramatically

increased.
The following D rays were taken first horizontally, then diagonally,

then vertically, then diagonally again.
Horizontal: 1, 13, 17, 17, 14, 14, 12, 6, 6, 6

First diagonal: 1,2,5, 6, 4, 4, 6, 7, 7, 8, 8, 5, 6, 6, 6, 6, 5, 5, 3, 3, 1
1, 1
Vertical: 2, 1, 1, 3, 3, 6, 6, 8, 9, 7, 5, 5, 5, 5, 6, 8, 8, 8, 4, 4, 2
Second diagonal: 1 , 2, 2, 2, 2, 4, 5, 4, 5, 5, 4, 5, 6, 6, 9, 9, 7, 5, 5,
4, 2, 3, 3, 3, 2, 1
The horizontal was recorded on the left of the object, the first diago-
nalbeam was recorded on the lower left, the vertical beam was recorded
below and the second diagonal beam was recorded below and to the
right of the object. In each sequence, one of the numbers appears in bold
type. This number represents the ray that passed through the center of
the object.
The reconstruction process may use any insights contained herein,
but there is no substitute for good old-fashioned logical thought. Readers
may even try one of the first reconstruction techniques employed in
medical imaging, the back-projection technique.
You can determine a minimal outline by drawing the region in which
all four D rays are present over the grid. Then back-project: because
each number corresponds to a specific ray (either horizontal, vertical, or

diagonal),one can distribute the number back over every square in the
ray as thoughit encountered a uniform shade of gray. That shade is cap-
tured by an average: divide the number of the ray by the number of

squares in it.
The shade of gray can be entered in the squares numerically (but one
must have a steady hand) or, better still, can be reflected by a number of
dots. If the quotient turns out to be .8, for example, place eight dots
in each of the squares of the ray. A quotient of 1.2 would then require
12 dots. The process is time-consuming but perfect for that long flight,
quiet lunch, or slow evening. When a square is dotted in for a second,
third, or even fourth time, room must be found for additional dots. At
the end of the process, the object will be plainly visible, if not precisely
rendered. It can be cleaned up by erasing all those squares that contain
fewer than a certain number of points. This practice is known as thresh-

olding.
To speed up the task, some readers might be able to adapt the grin
program. Logicians will of course insist on deducing which squares are
dark. This can bedone by a process of elimination like the technique for
reconstructing categorical patterns. Yet I think that logicians and pro-
grammers alike will find the solution to be the cat's meow.
Further Reading
Gabor T. Herman. Image Reconstruction from Projections: The Fundamentals of
Computerized Tomography. Academic Press, 1980.
in Train
London Bridge is broken down
My fair lady.
ANONYMOUS
1 1 helps to be flexible when you think about rigidity. I learned this les-
son in the summer of 1978 as my father, my son, Jonathan, and I fixed

up our cabin in the Canadian North. To patch the leaky roof, my father
had built a scaffold from freshly cut spruce poles. When my father and
Jonathan climbed to the top of the scaffold, the rustic framework
groaned and swayed. I mentioned that the scaffold looked a little shaky,
but my father scoffed, "Why this thing will hold 10 men —
and I used
the absolute minimum number of poles."
Who was I to argue with my father, an expert woodsman and an ama-
teur mathematician to boot? I returned to my chores inside the cabin.
Less than a minute later I heard a whoosh, a thump and two startled
cries.Racing outside, I found Jonathan and my father sprawled on the
moss. The scaffolding had scaf-folded, so to speak. The two stood up, and
my father grinned sheepishly, exclaiming, "Isn't that the damnedest
thing!"
can hardly blame my father for building an unstable scaffold.
I
Mathematicians and engineers have been struggling for centuries with

the theory and practice of constructing rigid frameworks. Mathemati-
cians call the subject rigidity theory. I investigated this topic, hoping that
a few insights might save my family and others from further injury. My
research has also uncovered a host of amusing puzzles to flex the mind.
Rigidity theorists prefer not to make frameworks out of spruce poles
and nails. Instead they have a mental construction set that consists of
abstract bars that cannot be stretched, compressed, or bent by any
158
amount of force. Such bars come in all conceivable lengths, and if the
ends of two or more of them touch, an instant universal joint is formed.
The joint allows the two bars to swivel and twist unless other connecting
bars constrain their motion.
Imagine, for example, a framework of 12 bars of equal lengths ar-
ranged into a cube. The cubic framework is not rigid. Placed on a table, it
would flop over in an instant. Indeed, if such a framework were rigid,
bridges and towers would not need diagonal girders.
Recently I attempted to brace a cube by adding diagonal bars to
some of its six square faces. How many bars would it take to make the
cube rigid?Four bars, judiciously placed, seemed enough until I found a
way to make the cube flex. Even a cube that had diagonal bars added to
five of its faces turned out not to be rigid (Figure 16.1). The five bars,
along with six of those already present in the cube, form two tetrahe-
drons that are hinged along a common bar. If the two tetrahedrons are
folded against each other, two of the four joints that define the unbraced
square face move toward each other as in Figure 16.1. The other two
joints move outward. No matter how five diagonal bars are added to the
faces of a cube, there will always be a way to flex it. No fewer than six
bars are needed.
Instead of bracing a cube on its faces, what if it were braced by diago-
nal bars that run from one joint right through the center of the cube to the
opposite joint? (With the careless elan of theorists, readers may ignore
the intersections of the diagonals.) A cube braced by four interior diago-
nals has a strange kind of flexibility that theorists call an infinitesimal
flex. In some sense, an infinitesimal flex is a motion of one part of a
framework relative to another. The motion is so small, however, that it
does not even exist.
Figure 16.1 The braced cube at the left has an ordinary flex (center),
whereas the one at the right has an infinitesimal flex.
Let me explain. The diagonally braced cube shown in Figure 16.1

below has arrows that indicate a tiny rotation of the top face in relation
to the bottom face. Because all bars making up the cube are made of ideal
materials that will not suffer the slightest change in their length, the top
face cannot be truly rotated, even by a tiny amount. Yet one may start to
rotate the top face and the bottom face in opposite directions. During
this vanishingly tiny moment, there is no resistance from any other part
of the cube because all bars that connect the upper face to the lower one
make right angles with the direction of rotation.
If this diagonally braced cube were made of real materials, it would
be distinctly vulnerable to small but measurable rotations. The structure
would wobble. (My father avoided this particular style of bracing.)
Frameworks that have only infinitesimal flexes are considered rigid, but
those that have no flexibility whatsoever are called infinitesimally rigid.
Besides their mental construction sets, rigidity theorists also have a
mental tool kit containing a great many theorems and techniques that
can be applied, among other things, to bracing a cube. One of the sim-
plest and most effective tools was discovered by 19th-century engi-
neers. A framework that has /joints must have at least 3/ — 6 bars to be
infinitesimally rigid. This theorem can be applied to the cube; its eight
joints mean that / = 8. The corresponding magic number computed by
the formula is (3 X 8) - 6 = 18.
To show that a cube composed of 18 bars (12 edges and six braces) is
actually infinitesimally rigid, one might appeal to a theorem discovered
by the Russian geometer A. D. Alexandrov in the 1940s. Alexandrov
studied rigidity in frameworks based on a convex polyhedron. These fa-
ceted surfaces include everything from cubes to cut gems to the geodesic
domes of R. Buckminster Fuller. Alexandrov proved that any frame-
work based on one of these shapes can be made infinitesimally rigid by
adding bars to the framework so that every face is composed of triangles.
As far as Alexandrov's theorem is concerned, then, a triangular bracing
of each face of the cube (one bar each) will make it strong.
I sympathize with any readers who have problems visualizing the
cubic bracings. Even the diagrams in Figure 16.1 are a bit complicated.
Perhaps it is time to descend from the three-dimensional space that gave
birth to the theory down to the plane, a two-dimensional space inhabited
by a vast panoply of various flat frameworks. Although readers can eas-
ily figure out that a square can be made rigid with a single diagonal, they
will find it rather challenging to figure out how to brace a grid of squares.
For example, how many diagonals must be added to make a four-by-four
grid of squares immune to flexes? Figure 16.2 shows two ways to brace
such a grid with only seven diagonals. But one of the braced grids is not
Can readers tell which one?
rigid.
The answer can be deduced in the following manner. Make up a dia-
gram composed of two sets of dots. The first set represents the four rows
of the grid, one dot per row. Likewise, the second set corresponds to the
four columns. For each of the seven diagonal bars in the grid, connect the
appropriate dots in the diagram. For example, if there is a diagonal bar in
the square situated in the third row and the fourth column, then draw a
line from the third dot in the first set to the fourth dot in the second set.
Whether the grid is now rigid can be answered by asking the follow-
ing question: Is the dot diagram connected? In other words, is there a
continuous path from any dot in the diagram to any other? If (and only
if) so, the grid is rigid. This elegant theorem —
first proved by Henry
Crapo of the inria near Paris and Ethan D. Bolker of the University of
Massachusetts at Boston —
can help readers quickly determine whether
a grid will flex. The diagram for the grid on the left in Figure 16.2 is con-
nected, but the other is not. As an exercise in rigid thinking, I will leave
readers with the problem of using the Crapo-Bolker theorem to decide
why seven is the minimum number of bars necessary to brace the grid.
As far as I know, there is no corresponding theory to advise readers, or
my father, about how to brace scaffolds or other cubic grids.
Figure 16.2 Dot diagrams reveal whether a grid is rigid (left) or not
(center and right).
Sometimes the search for rigidity requires flexibility in the literal

sense. No story better illustrates the point than the history of the famed
rigidity conjecture. In the 17th century the French mathematician Au-
gustin-Louis Cauchy wondered whether all convex polyhedral surfaces
were rigid. Such surfaces include the triangulated polyhedrons of Alex-
androv's theorem and many more. Their facets, or faces, are bounded by
plane polygons with any number of sides. Being convex, they have no
indentations or hollows of any kind. In 1813 Cauchy proved that a con-
vex, polyhedral surface is rigid if all its faces are triangles. The theorem
meant that any convex surface one could construct from triangles, each
triangle sharing each of its bars with one other triangle, would be rigid.
Despite the restriction of Cauchy' s theorem — that the surface be
convex — mathematicians were beginning to wonder whether all sur-
faces composed of triangles were rigid — even those surfaces that were
not convex. Such surfaces may appear to be folded, twisted, or con-
torted in quite crazy ways. The only requirement was that they be sim-
ple in the topological sense. If suddenly converted to rubber and in-
flated, they must be (more or less) spherical. Additionally, a simple
surface required that no part of it touch another part of the same surface.
Mathematicians conjectured that if a surface had all these properties,
then no matter how deformed it happened to be, a version composed of
triangles would suffer no flexes.
For more than a 100 years, no one was able to prove this so-called
rigidity conjecture, nor could anyone disprove the conjecture by finding
a flexible, nonconvex surface made of triangles. The strongest support-
ing evidence for the conjecture came in 1974, when Herman R. Gluck of
the University of Pennsylvania showed that "almost all" such surfaces
were rigid. In other words, examples counter to the conjecture, if they
existed, would be rare indeed. Even a contrarian would find this much
evidence in favor of a conjecture discouraging.
But Robert Connelly of Cornell University was convinced in some
corner of his being that the rigidity conjecture was false. After visual-
izing surface after surface that looked asthough it should flex, Connelly
realized one day that he was working against Gluck's theorem. His office
was full of models sent to him by amateur mathematicians who claimed
flexibility for them. Gluck's theorem said, in effect, "Not likely!" Faced
with the same difficulty, Connelly decided to examine mechanisms,
namely, frameworks that he knew would flex.
Starting with a very simple flexible framework, he employed his
knowledge of topology, spanning parts of the framework with simple tri-
Figure 16.3 How to construct a Connelly-Steffen surface.
angles. Then one day he felt close. Before him was a nonconvex surface
that flexed. But it was not quite what topologists call a sphere. Two edges
within the surface touched each other, like a deflated basketball in which
one side is pressed against the other. The thing was distinctly annoying.
So near and yet so far.
It was then that the idea of a crinkle came to him. He suddenly
thought of a way to introduce a subdivision of the annoying edges and

surrounding triangles that amounted to a fold —
enough to take the two
lines out of contact. The model he built flexed!
The counterexample to the rigidity conjecture appeared in the litera-
ture in 1978. Shortly after, the German mathematician Klaus Steffen
found an even simpler surface, based on Connelly's idea, that flexed.
Readers who would like to flex their own version of the Connelly-Stef-
fen surface will find it laid out in Figure 16.3. To obtain a size that is easy
to work with, readers should interpret the edge numbers as centimeter

lengths. Arrows that connect the edges in pairs indicate attachments to
be completed by armchair rigidity theorists.
When the Connelly-Steffen surface is completed, the two central tri-
angles make a fold by which one hand may grasp the surface from above.
With the other hand, it will be possible to reach up under the model and
then (delicately!) to flex the bottom vertex from side to side, but only by
a small amount, roughly 10 degrees.
When this tiny flex is performed, the surface bounds the same vol-
ume. These days Connelly ponders whether the constant volume prop-
erty holds true for all flexible, nonconvex surfaces made of triangles. If
he conjectures that they do, he himself may have to be flexible. Some
young upstart may find a counterexample.
As something of an upstart myself, I gave my father some trouble
over the collapse of the scaffolding. But within a few hours of the acci-
dent, the scaffolding was up again. It was identical to the previous struc-
ture, except for one extra spruce pole. My father climbed the scaffold
confidently. I am sure the tiny wobbles I detected were merely infinites-
imal flexes.
Further Reading
Jay Kappraff. Connections: The Geometric Bridge Between Art and Science.
McGraw-Hill, 1991.
— ^
Automated Mat
The whole machinery of our intelligence, our general

ideals and laws . .are so many symbolic,
.
algebraic expressions.
George Santa y ana, The Sense of Beauty
TJ. here an old vaudeville comedy routine that pokes good fun at
is
the strong-man act familiar from the circus and the state fair. A heavily
muscled man takes the stage with his not so heavily muscled female as-
sistant. The man strains mightily against an enormous weight, and after
tremendous effort he manages to lift it above his head. The spectators
cheer, but the cheers turn to laughter when the assistant casually picks
up the weight in one hand and carries it offstage.
There are two computer programs that leave one with a similar sense
of comic deflation over the mental "muscle" allegedly displayed by a
high score on the traditional I.Q. Both programs perform at or near
test.
genius level on two tasks widely used in the tests, the completion of nu-
merical sequences and the perception of visual analogies. Yet both pro-
grams are simple to understand, and it is startling to realize just how
dull-witted they are.
Although I have no wish to offend readers who suppose themselves
plentifully endowed with mind stuff, I am twitting the I.Q. test with a se-
rious purpose. The stated intent of the test is to measure intelligence,
and few human qualities evoke such pride in their presence or anxiety
over their absence as intelligence does. Nevertheless, the concept of in-
telligencepresupposed by the traditional I.Q. test is seriously mis-
guided.The reasoning behind this assessment is cogently set forth by
Stephen Jay Gould of Harvard University in his book The Mismeasure of
Man (see Further Reading). What it comes to is this: The traditional I.Q.
165
test rests on the unstated and erroneous assumption that intelligence,

like strength, is a single quality of human physiology that can be mea-
sured by a graded series of tasks.
Numerical-sequence completion is a good example: What is the next
number sequence 2, 4, 6, 8, ... ? In the sequence 2, 4, 8,
in the
14, . ? In the sequence 1, 2, 6, 24,
. . ? The percentage of correct
. . .
responses to a set of such questions measures your "general intelli-

gence/' just as a strain gauge measures the weight you can lift and there-
fore the strength of your arm muscles. Note that if the results of the I.Q.
test are to be interpreted as a measure of "general intelligence/ there '
must be some core ability, or some small set of core abilities, that pro-
vides an index of what one means by general intelligence. Because the
very idea of general intelligence presupposes a strong correlation among
the core abilities, the precise kind of graded task adopted by the I.Q. test
is relatively unimportant. One good as another.
task is as
One of the I.Q. programs presented here is derived from a more
elaborate program written by Marcel Feenstra, a student living in Rot-
terdam. Feenstra's program is called hi q, and it solves two kinds of nu-
merical problems that often appear on standard I.Q. tests: sequence
completion and numerical analogies. Feenstra tested hi q on a number
of sample I.Q. tests that appear in a book by Hans J. Eysenck of the Uni-
versity of London, Know Your Own I.Q. (see Further Reading). The I.Q.
of Hi Q is apparently in the neighborhood of 160. Although the experi-
ment was not exactly a carefully controlled one, it leaves little doubt that
the program would score quite well under real test conditions.
The program I have in mind is called se Q, and it duplicates Hi q's per-
formances on numerical-sequence completion. Readers who write and
run SE Q may consider their own numerical intelligence amplified, as it
were, by proxy. The main idea of the program is straightforward. When
one is given a sequence of numbers and told to find the next number in
the sequence, one does not search for the number directly. Instead one
searches for the rule that led to the numbers already present. There is a
mathematical aside to be made here: For any finite sequence of num-
bers, there are infinitely many rules that give rise to it. The search thus
boils down to finding a simple rule for generating the sequence.
There are two kinds of rules considered by se Q: additive and
just
multiplicative. For example, to find the next number in the sequence 2,
4, 8, 14, ... one might look for an additive rule, and the best way to
,
find the rule is to construct what I call a difference pyramid (Figure

17.1). At the bottom of the pyramid is the given sequence, and the pyra-
1
17 Automated Math 167
2
2468 2
4
2
8
> 2
14
\
\
[ii] 1
2345 1
2
1
6
> 1
24
\
\
[
1 20
2, 11, 48, 189, [T] 1, 1, 17, -607, [7]
Figure 17.1 Numerical-sequence completion by the pyramid method.

Can the reader solve the lower two?
mid is up from bottom to top by finding the differences between

built
successive numbers in the preceding level or row of numbers. Thus the
first number in the second row of the pyramid is obtained from the first
two numbers in the first row, namely 2 and 4. Their difference is 2, and
so 2 is the first number in the second row. Similarly, the other numbers
in the second row are 8 — 4, or 4, and 14 — 8, or 6; the second row is the
sequence 2, 4, 6. Continuing the same process to a third row of the pyra-
mid gives a sequence with only two numbers, and they are both 2's.
The equality of all the numbers in some row of the pyramid is the
signal, so to speak, to stop building the pyramid upward and to start
building it sideways. For example, suppose the third number in the third
row is also 2. It is then reasonable to suppose the next number in the sec-
ond row is obtained from the preceding number, namely 6, by adding 2:
the sum is 8. the newly derived number in the second sequence can then
be added to the last number given in the first sequence: 14 plus 8 is 22,
and 22 is indeed given a perfect score by the test makers. New numbers
in each sequence percolate down the pyramid once a constant sequence
is derived at the top.
A great many questions on I.Q. tests about numerical sequences
yield to this simple procedure. Readers who have more than a nodding
acquaintance with algebra will recognize the signature of a polynomial in
the exercise. Any polynomial evaluated for consecutive integers yields a
sequence that generates a difference pyramid. Given enough values of
the polynomial, a row of identical numbers will eventually top off the
difference pyramid. The number of rows needed to build the pyramid up
to a constant row, minus 1, is the degree of the polynomial. The se-
quence 2, 4, 8, 14, which gives rise to a constant row of 2's in the third
level of the difference pyramid, is generated by successive values of the
quadratic, or second-degree, polynomial x — x + 2.
2
Unfortunately one cannot solve all sequence questions by making

difference pyramids. For example, the sequence 1,2, 6, 24, . . . yields
a difference pyramid with the numbers 3 and 14 in its top row. The rapid
growth of the numbers, however, strongly suggests a geometric series:
the consecutive terms of a geometric series are related by multiplication
instead of addition. Hence it seems reasonable to construct a set of quo-
tients from the sequence instead of a set of differences (Figure 17.1). By
taking quotients of successive pairs in the sequence 1, 2, 6,
24, .one obtains the second row in a pyramid, the sequence 2, 3,
. .
4, . The second sequence hints at an abrupt rule change: the third

. . .
row in the pyramid must be obtained by taking differences, not quo-

tients. Who can doubt that the intended solution requires a 5 at the end
of the second row? The solution itself is thus 120: the product of 24, the
last given number in the first row, and 5.
The sequence-solving program SE Q attempts to build pyramids by

considering both the consecutive differences and the consecutive quo-
tients of successive pairs of numbers in a given row. Even more, it exam-
ines successive pairs of numbers in a sequence for more general additive
and multiplicative rules. In the additive rule the first member of each
pair may be multiplied by a constant k before the usual addition is done,
and in the multiplicative rule the constant k may be added just after the
usual multiplication. Here is an easy piece for programming novices
(Figure 17.2).
The few simple formulas that give the general rules make up the core
of SE Q. Suppose the given sequence has already been assigned to the
four variables a(l), a(2), a(3), a(4). To obtain the second row, b{l), b(2),
b{3), se Q tries substituting either a generalized difference, of the form
b(l) <— a(2) — k X a(l), or a generalized quotient, of the form b(l) <—
[a(2) — k]/a(l). In both examples k stands for any integer in some pre-
determined range. The program also tries analogous substitutions for
b(2) and b(3), each for the same value of k: for b{2) it tries a(3) — k X
a(2) or [a(3) - k]/a(2), and for b(3) it tries a(4) -kX a(3) or [a(4) -
k]/a(3).
The developed even more simply: SE Q tries
third row, c(l), c(2), is
substituting only simple differences, c(l) <— b{2) — b{l) and c(2) <—
b(3) - b(2), or simple quotients, c(l) *- b(2)/b(l) and c(2) <- b{3)l
b{2). Apparently it is rare for sequence-completion questions on I.Q.
tests to get more complex than the formulas allow for.
When SE Q develops a pyramid, it tries each generalized substitution
for the set of with each simple substitution for the set of c's. Concep-
fr's
tually, therefore, se q is made up of four major segments. Each segment

1 7 Automated Math 169
INITIALIZING AND INPUT INSTRUCTIONS

I
TRY DIFFERENCE FORMULAS
ON FIRST AND SECOND ROWS
LOOP FOR k
LOOP FOR i
b(i)<-a(i + 1)-[kxa(i)]
c(i) <—b(i + 1 - b(i))
COMPUTE SOLUTION?
I
TRY DIFFERENCE FORMULAS
ON FIRST ROW AND QUOTIENT
FORMULAS ON SECOND ROW
LOOP FOR k
LOOP FOR i
b(i) *-a(i + 1) - [k x a(i)]
c(i) <-b(i + 1)/b(i)
COMPUTE SOLUTION?
i
TRY QUOTIENT FORMULAS
ON FIRST AND SECOND ROWS
LOOP FOR k
LOOP FOR i
b(i)«-[a(i + l)-k]/a(i)
c(i)<-b(i+.1)7b(i)
COMPUTE SOLUTION?
I
TRY QUOTIENT FORMULAS
ON FIRST ROW AND DIFFERENCE
FORMULAS ON SECOND ROW
LOOP FOR k
LOOP FOR i
b(i)<-[a(i + 1)-k]/a(i)
c(i) <—b(i + 1) - b(i)
COMPUTE SOLUTION?
Figure 17.2 Conceptual flow chart for se Q.
is a loop with one combination of substitution formulas in it. For exam-

ple,one such segment of the program applies the three generalized first
quotient formulas, of the form b(l) «— [a(2) k]/a(l), to compute the —

values of b(l), b(2) and b(3) that make up the second row of the pyramid.
The program segment then two simple difference formulas,
applies the
of the form c(l) <- b{2) - and c(2) that
b(l), to obtain the values of c(l)
make up the third row. The complete set of five formulas in this segment
is embedded in what might be called a try-everything loop, in which dif-
ferent values of k are tested. Feenstra recommends allowing k to take on

all integer values from — 5 to 5.
Within the loop, each time new values for c(l) and c(2) are computed
they are tested for equality. If they prove to be equal, their common
value is stored in a variable called c and the current value of k is saved in a
variable called kk. Just after the loop there are instructions that construct
the solution to the original sequence (if one has been found) from the
values of c and kk. In the example I am describing one obtains b(4), the
new member of the second row, by adding c to b(3). The solution, a(5), is
then obtained by multiplying a(4) and b(4) and adding kk to the product.
Two instructions at the end of each loop thus suffice to recover a so-
lution from a successful search within the loop. The instructions appro-
priate for each loop depend on the formulas used in it, and I shall leave it
to those who write SE Q to discover the instructions for themselves. Use
a bit of algebra to isolate the variable of interest. In each case, if one of
the loops in the program finds a solution, there must be an instruction to
printit. The program may then skip the remaining loops and stop, or it
may execute all the loops in an effort to find more than one solution. By
executing all the loops Feenstra has detected several "bad" I.Q. ques-
tions that have more than one solution. If none of the four loops finds
any solution, it is reasonable to include an additional output statement

following the entire lot. The message it prints can vary according to
taste; those who like to invest their programs with a little personality can
have it print "Help!"
One can try out SE Q on questions from sample I.Q. tests found in
several widely available books. In Eysenck's book there are eight com-
plete I.Q. tests, which allegedly enable the reader to discover his or her
own The tests incorporate several different kinds of questions that
I.Q.
appear on standard I.Q. tests, including questions that involve a missing
number, a missing letter, a missing word, an odd-man-out, scrambled
words and visual analogies (Figure 17.3). A few sub-types are usually
found within each major category. For example, in Eysenck's book there
are three kinds of questions that ask for a missing number, namely the
ordinary numerical-sequence problems I have already described and
two other kinds typified by the following examples:
164 (225) 286

224 ( ) 476
8 3 21
6 5 25
12 2 —
1. Insert the missing number. 3 7 16 35
2. Insert the missing letter. N Q L S J U
3. Insert the word that completes the first word and starts the second. GRO ( ) PER
4. Underline the odd man out.

ANIMAL ENGINE IDENTITY OCTAGON UNICORN
5. Underline which of these towns is not in Italy.
NORLEFEC DARDIM SAIP LIMNA
6. Which of the four numbered figures completes the top line?
Figure 17.3 I.Q. minitest based on questions from Hans J. Eysenck's

Know Your Own I.Q.
In each case the would-be genius must supply the missing number in ac-
cordance with some perceived rules. Feenstra's hi q program handles
such questions by procedures that draw on the same kinds of formulas as
the sequence-completion program does. I encourage readers to try
them. Answers to the problems above, as well as others, appear at the
end of this chapter.
Although hi Q answers only one major kind of I.Q. test question, the
solutions to other kinds of questions can also be mechanized. In fact, a
program that solves visual analogies was written more than 25 years ago
by Thomas G. Evans as part of a Ph.D. dissertation done at the Massa-
chusetts Institute of Technology. Heavy as it sounds, the essential ideas
of Evans' program are easy to understand.
The visual analogies it solves all have the same form: figure A is to
figure B as figure C is to one four figures listed as potential an-
of, say,
swers. The program selects the analogous figure by first determining a

simple set of rules that can transform figure A into figure B (Figure 17.4).
It then repeats the procedure with figure C and each of the four potential
answers; in each case it generates a set of rules that can transform figure
C into the potential answer. The figure obtained from the transforma-
tion rules that most closely resemble the rules for transforming figure A
into figure B is selected as the solution.
Evans' program repeats essentially one operation five times. Two
figures at a time, a source figure and a destination figure, serve as input.
For each pair of figures the program then develops a three-part tabular
description of how the source figure becomes the destination figure.
First the program lists the spatial relations among the parts of the source
figure; then lists the spatial relations among the parts of the destination
it
figure. Both descriptions consider only three spatial relations, above, left
and inside. Finally, the program describes how parts of the source figure
change into parts of the destination figure in one of four basic ways: each
part can be altered in size, rotated, reflected, or deleted.
Suppose figures A, B and C each have three parts, a circle, a square
and a triangle. In figures A and B the program may label the triangles a,
the squares b and the circles c, but it makes no attempt to label the parts
of figure C in the same way. Instead it arbitrarily assigns the labels x, y,
and z to the three parts of figure C. It then develops its three-part tabular
description for the pair of figures A and B and four more descriptions,
one for each pairing of figure C with a potential solution. The last four
tables all employ the labels x, y, and z throughout.
The final operation of Evans' program is to make every possible
substitution of a, b, and c for x, y, and z. Since x, y, and z can be permuted
in only six ways, there are six substitutions to be tried. One of the substi-
tutions may convert the tabular description of the pair of figures A and B
into the corresponding tabular description of figure C paired with one of
the potential answers. The figure in this pair is the solution. Even if no
perfect matches are found, however, the program can score the relative
success of an analogy and so pick the substitution that yields the best
match.
Patrick Henry Winston describes the visual-analogies program in his
book Artificial Intelligence (see Further Reading.) Winston states that the
program "works well," and he attributes its success to the use of an ef-
fective framework for representing knowledge about the geometric fig-
a: UNCHANGED
IS TO
b: REDUCED
a ABOVE b b ABOVE c
c: ENLARGED
a ABOVE c b ABOVE a
c INSIDE b a INSIDE c
x: UNCHANGED
y: UNCHANGED
x LEFT y
z: UNCHANGED
z LEFT y
z INSIDE x
x: REDUCED
y: ENLARGED
y LEFT x
z: UNCHANGED
z LEFT x
z INSIDE y
IS TO
x: REDUCED
y: UNCHANGED
z: ENLARGED
x ABOVE y
z ABOVE y
x INSIDE z
x: REDUCED
y: ENLARGED
LEFT z z: UNCHANGED
y LEFT z
x INSIDE y
Figure 17.4 Solution of visual analogies in Thomas G. Evans' disserta-

tion.
173
ures considered by the program. For example, instead of specifying how

relations such as above, left, and inside change from picture to picture,
the program might have described how one figure in the first picture gets
transformed into another figure in the second. Such a program might be
extremely cumbersome, not to say ineffective, because it would have to
check a much larger number of potential substitutions than Evans' pro-
gram does. Indeed, the search for a good representation is a major theme
of artificial intelligence: it is often the key that enables a computer to
mimic some aspect of human problem solving.
By analogy the representation of objects in the mind is also the sub-
ject of much discussion among cognitive scientists. In this context the
study of artificial intelligence is often justified as an attempt to exhibit an
"existence proof" for the mechanistic description of human capabilities.
Thus, goes the argument, if a computer program can be made to simulate
some aspect of human behavior, the representation of the behavior
adopted by the program at least could serve as the underlying represen-
tation adopted by the brain. Nevertheless, it often seems that successful
simulations of such behavior give little insight into how people do the
same things. For all one knows there may be no relation whatever be-
tween the way Feenstra's or Evans' I.Q.-test programs perform and the
way people solve the same kinds of problems. Presumably human intel-
ligence deploys more general strategies in attacking particular problems.
This point brings me full circle to the reconsideration of human in-
telligence: what it is and how it is measured. As I have noted, Stephen
Jay Gould has characterized I.Q. as a mismeasure of man. His criticisms
carefully document two major fallacies that underlie the concept: the
uncritical reification of an abstraction and the ranking of the reified ab-
straction along a single scale. Language itself accounts in large part for
our tendency to make things of what are at best nebulous abstractions.
Moreover, once we have persuaded ourselves that we are dealing with a
thing, our reflex is to measure it.
In demanding a single numerical measure we succumb to the second
fallacy, namely ranking. We want to reduce complex phenomena to a
single scale. Such practices have led to excellent physics, but they have
also led to some poor social science. I.Q. testing is a case in point; it is to
the 20th century what craniometry was to the 19th. In both instances
entire racial groups found themselves mismeasured not only because
the measure was almost meaningless to begin with but also because
there were biases introduced (either consciously or unconsciously) in
the process of measuring.
Gould vigorously attacks biological determinism, the idea that

human behavior is determined by genes, and he warns against viewing
the capacities of our brains as direct products of natural selection. "Our
brains are enormously complex computers," he writes. "If I install a
much simpler computer to keep accounts in a factory, it can also perform
many other, more complex tasks unrelated to its appointed role. These
additional capacities are ineluctable consequences of structural design,
not direct adaptations. Our vastly more complex organic computers
were also built for reasons, but possess an almost terrifying array of ad-
ditional capacities — including, I suspect, most of what makes us
human."
With this metaphor Gould has put his finger on what I find most
last
unsettling about a relatively simple computer program that can score at
the genius level on an I.Q. test. Does the score on the test measure the
intelligence of the computer? If it does not, just how does one go about
measuring the intelligence of a computer, whether it is made of silicon
and plastic or of carbon and tissue? The answer: Probably not by running
some I.Q. program through a battery of tests.
Answers to the Minitest and Other Questions
The first problem on the minitest is a good example of the

ambiguity typically found in such problems. The problem was
to complete the sequence 3, 7, 16, 35,. Each term minus
. . .
twice the preceding term gives the sequence 1, 2, 3, the

second row of a pyramid. By this reasoning the missing term
must be twice 35 plus 4, or 74. On the other hand, if a simple
difference pyramid is constructed with three rows, the third
row gives the sequence 5, 10, and it seems reasonable to
complete the sequence with 15. The missing term must then
be 69, but the programs described in this chapter would have
missed this answer. The other answers to the test: H is the
missing letter; the missing word is "up"; the odd man out is
"identity"; the unscrambled name of the town not in Italy is
Madrid, and the correct visual analogy is number 2.
The two numerical sequences are completed by 350 and 22
respectively. The first sequence can be solved by applying a
generalized difference rule, with k equal to 3, and then a

generalized quotient rule; the missing term is 324. The second
sequence ought to defeat but the most patient puzzle
all
solvers who did not try to write SE Q. It can be solved by two

quotient rules; the value of k in the first rule is 5. The missing
term is — 65,551.
Further Reading
Hans J.Eysenck. Know Your Own LQ. Penquin Books, 1982.
Stephen J. Gould. The Mismeasure of Man. W. W. Norton, 1983.
Patrick Henry Winston. Artificial Intelligence, 2nd ed. Addison- Wesley, 1984.
THEME FOUR
LOMPUTERS LREATE
In the chapters that follow, computers com-
pose canons and play chaotic improvisations,
cultivate the art of disjointed conversation,
cartoonize human faces, and paint scenes from
other planets. The degree to which humans
intervene in these creations varies from exam-
ple to example. The one, creating those
last
fascinating arabesques of centuries past, is left
tohumans —
so far.
The composition of a strict, first-order
canon by the rule of renaissance music in the
first chapter, for example, allowsno human
intervention whatever. The computer searches
systematically through all possible melodic
sequences, matching it against itself. This

process is not exactly the one taken by the
great masters, from Byrd to Victoria. Nor are
humans allowed to intervene in the logistic
equation as each churns out surprisingly
human-sounding musical fragments. Some
might argue that the human has already inter-
vened in the form of a programmer. Does not
he or she get the real credit (if credit there be)
for the composition? Only if one allows "com-
position' to include music that the program-
'
mer hadn't the faintest idea was coming and

took not the slightest part in directly determining.
A surprisingly simple conversational pro-
gram called Mark V. Shaney speaks strangely
to other humans over a computer network.
Some think it is a secret Al project that got out
of control, others think it is a demented person
trying to get attention by saying strange things.
But Mark V. Shaney merely echoes the word
groups of earlier conversants. It weaves these
together into a fabric of speech that reflects
our thoughts fragmentarily, like a bad memory
of something we said at a party (and now
regret). What fascinates people about this pro-
gram is its marvellous economy of means. It
has only a tiny fraction of the length of much
more ambitious —
and hardly more successful
conversational programs. Perhaps it is a come-
dian.
The next two programs presented here
depend heavily on humans. From a massive
file of actual faces, the first program may
compile an average face, then use the average
face, to create a cartoon of any face in its file.
The human, at least, makes the selection. A
cartoon in this case is just the face
extrapolated in "face space" through point-by-
point comparison with the average face, an
exaggeration by formula. The second program,
really a package of programs, uses geometry
and a very powerful graphics computer to
render creatures that never were and, one
hopes, never will be.
At the end lie the graceful latticeworks
from medieval Islamic tombs, monuments, and
mosques. How did their artisans discover
these brilliant figures? Before anybody
programs them, readers have a perfect right to
follow the advice of the final chapter to create
a special pattern of his or her own. Is this art?
This depends on how well it's done.
As for the earlier computer-generated
sounds, words, and images, the question is
harder to answer. For one thing, no one seems
exactly sure what art is or isn't. But critics
who commit themselves to gung-ho
postmodern cybernetic realism will
undoubtedly cheer each new creation from the
sidelines. Others, made of sterner stuff, will
take an extraordinarily long time to convince.
II est certain que la premiere qualite d'un compositeur,
c'est d'etve mort.
Arthur Honegger, Je suis compositeur
1 1was in 1965 that I first heard the sound of computing. An IBM 7090
at the University of Michigan Computing Center had been fitted with a
simple electromagnetic pickup connected to a loudspeaker. As a com-
puter program ran, a specific register in the machine changed its con-
tents many thousands of times per second. The resulting pattern of min-
iature clicks was heard as an astonishing rush of alien sound that
alternated among buzzing, screaming, burping, rumbling, and whining.
At times a grinding noise would change from bass to treble. This may
have been the sound made by a double loop in the program; perhaps the
inner loop executed ever faster, creating a tone of increasing pitch. It all
sounded like a humpback whale.

The experience has suggested a variety of programs that explore
aspects of melody, harmony, and rhythm. Even though home computers
do not normally come equipped with electromagnetic pickups poised
over accumulator registers, the majority have small loudspeakers con-
nected to a primitive tone generator. There are instructions in most pop-
ular computer languages that call forth a variety of sounds when they are
executed. Such a simple facility can be exploited both to make programs
audible and to make audible programs. In the former case a sound-gener-
ating instruction is added to a program originally intended for computa-
tional purposes. I shall discuss this topic briefly later. In the latter case a
program is deliberately constructed to produce melodic, harmonic, or
rhythmic effects. Two of the programs described here resemble the
179
180 Theme Four Computers Create
former kind; a few sound-generating instructions are inserted in other-

wise normal-looking code.
Melody, considered merely as a succession of notes, is easily gener-
ated by a program. In fact, the program can consist of a single loop. A de-
cision process embedded in the loop decides what note to play next and
how long it should be played. The process itself may be arbitrarily com-
plicated. For all I know there are already programs of this type that gen-
erate convincing melodies in a given traditional style. If such programs
exist, I should be glad to hear of them. In the meantime I shall take the
tack of letting available algorithms shape the melody.
The humblest applicants for the job of melody maker are simple ar-
ithmetic algorithms. The numerical output of such algorithms is easily
converted into notes under an enormous variety of possible encodings.
The simplest of the encodings uses the linear congruential assignment, a
process that is shorter than its own name:
x <- (a x •
+ b) mod m
Here, when one specifies the parameters a, b and m
in advance, an initial
value of the variable x is converted into a succession of values by the
continued iteration of the assignment. The expression "mod m" is an ab-
breviation for "modulowhich means that the number computed in-
ra,"
side the brackets should be treated like the hours of an ra-hour clock. For
instance, 10 modulo 8 equals 2. Thus if m is 8 and if a, b, and the initial
value of x are all integers, one obtains a sequence of numbers ranging be-
tween 0 and 7.
The resulting sequence of numbers is readily converted into a suc-
cession of notes by a simple table:
0 12 3 4 5 6 7
do re mi fa sol la ti do
In the program SOLFEGGIO (from the practice of sightsinging) the

I call
quantities symbolized by the names of the notes are replaced by the fre-
quencies, in cycles per second, of the C-major musical scale beginning at
middle C (Figure 18.1):
0 12
3 4 5 6 7
262 294 330 349 392 440 494 523
18 The Computer Composer 181
Note C C# D D* E F F# G G# A A# B C
Frequency 261.6 277.2 293.7 311.1 329.6 349.2 370.0 392.0 415.3 440.0 466.2 493.9 523.3
Frequencies of notes above or below this octave are obtained by multiplying or dividing by 1 .05946 and
rounding off as appropriate. The number is ife the twelfth root of two.
,
Figure 1 8. 1 Frequencies of the semitone scale from middle C to one oc-

tave above middle C.
The complete program can be summarized in the usual algorithmic

form:
input a x to,
for / = 1 to 100
x «- (a x + to) mod 8
•
note <- notes (x)

play note
In my version of solfeggio the numbers in the figure called notes are

employed directly by the instruction that generates a tone of the appro-
priate pitch. Most languages also allow the duration of the tone to be
controlled. For the present it will suffice to set the duration to, say, half a
second.
A world of maniacal melody now awaits those with an "ear" for ar-
ithmetic. Depending on what numbers are selected for the three param-
eters, one hears either boring staccato monodies, odd, repetitive melo-
dies on a few notes, or strangely wild music full of leaps and sudden runs.
In the last category I have been caught short on occasion by not having
made a note of the parameters responsible for a haunting little piece for-
ever lost. The algorithm I have suggested produces melodies 100 notes
long. Explorers of this new musical terrain may wish to shorten the
length in order to investigate more possibilities.
solfeggio can and should be enhanced. Choose a larger value for
the modulus m, large enough to embrace two or more diatonic octaves.
There is no modal restriction as such; instead of diatonic scales one can
choose 12-tone scales consisting of half-note steps. One can even choose
the ultra-modern form of linear-congruential music in which the num-
bers x produced by the algorithm specify frequencies more directly by
the addition of a constant, say 100. If at one point x is 183, then 283 is
the frequency of the note the program will play. I wonder what other
possibilities readers might invent for themselves. In the meantime here
an interesting question about boredom. For given values of a, b, and m,
is
how many notes will be played before the melody begins to repeat itself?
Harmony is now withinthe vocal competence of many home com-
puters. Those equipped with two or more speakers will be able to play
the entire repertoire generated by a program I call canon. Indeed, a few
enthusiasts will undoubtedly carry the project much further. Some
readers may already have acquired a MIDI. The acronym stands for Mu-
sical Instrument Digital Interface, an electronic black box that converts
signals generated by one's computer into commands for electronic in-
struments such as musical keyboards and individual synthesizer chan-
nels. (Readers interested in learning more about MIDI may write the In-
ternational MIDI Association, 12439 Magnolia Boulevard, Suite 104,
North Hollywood, 91607.)
Calif.
canon generates two-part harmony involving two almost identical
melodic lines. Rounds such as "Row, Row, Row Your Boat" and "Frere
Jacques" exemplify the type if not the species, canon generates a canon
in the academic tradition known as first-species imitation, a first stage in
the serious study of counterpoint.
Such a harmony has two melodic lines that satisfy four criteria. First,
all notes have the same duration. Second, one line begins after the other.
Third, both lines are identical except that one line is transposed upward
by some standard musical interval (unison, perfect fourth, perfect fifth,
or octave). Fourth, the two lines together must satisfy, note for note,
certain rules of first-species imitation. All the rules are found in standard
texts. Such rules generally include the allowed note-for-note harmonic
intervals (Figure 18.2) and establish the shape of participating melodic
lines. In the interest of simplicity the latter has been omitted.
An example of a harmony in first-species imitation is shown in Fig-
ure 18.3. The example was written by a human being, not by a machine.
A glance makes it plain that after a certain point the composer tires of the
strenuous demands imposed by all four criteria. Rule three is usually the
first to go; enough if the second line imitates the first in spirit only.
it is
A computer program, on the other hand, is undaunted by the con-

straints, canon will grind on tirelessly until an entire piece is generated.
The usual output of the program is tedious, however. To avoid this
problem I have generated a great many short canons and have catalog-
ued the more interesting ones. Some of these may then be strung to-
gether to make longer pieces (Figure 18.4). The experiment was done a
I
18 The Computer Composer 183
b
UPPER NOTE NUMBER
OF HALF-
c C# D D# E F F# G G # A A# B C NAME STEPS
X X X X X X 1 0
c# X X X X X X 1+/2° 1
D X X X X X 2 2
D# X X X X 3- 3
E X X X X 3 4
F X X X X P4 5
F# X X X X 4+/5° 6
LOWER NOTE G X X X P5 7
G# X X 6- 8
A X X 6 9
A# X X 7~ 10
B X 7 11
C 8 12
Notes sounded together are consonant, according to table a, if the combinations do not fall on
an X. The names and sizes (in half-steps) of the resulting harmonic intervals are found in table b.
The intervals of unison (1), minor and major third (3~ and 3), perfect fourth (P4), perfect fifth
_
(P5), minor and major sixth (6 and 6) and octave (8) are considered consonant in first-species
imitation. The corresponding numbers of semitone (half) steps are 0, 3, 4, 5, 7, 8, 9, and 12.
Figure 18.2 Table of consonant harmonic intervals.
1 —
o*ti—o
° "
1_0 O & 1
Figure 18.3 Example of first-species imitation.

Figure 18.4 Two computer-generated canons are spliced to make a

larger one.
decade ago with the assistance of Gregory Utas, one of my students at

the University of Western Ontario.
canon counts its way into a piece. Suppose, for example, it has been
told to generate a miniature canon six notes long. For reasons that will
soon be obvious, the tonal range spanned by each musical line is small,
say six chromatic steps up and six steps down from the tonic, or begin-
ning note. The set of 13 possible notes can be thought of as digits in a
base-13 number system. The digit 0 represents the tone six steps below
the tonic and the "digit" 13 represents the tone six steps above, canon
proceeds quite simply by counting its way through all six-digit base-13
numbers. Each one, after all, encodes a melody of a kind.
Every time canon generates a new melodic line it makes a copy,
transposes it upward by a fifth, and translates it forward by a specified
number of notes, say two. It then compares the pairs of notes brought
into temporal juxtaposition by the operation. If none of the resulting
pairs violates the rules of harmony, the line is accepted as canonic and is
either printed out directly to the waiting composer or saved in a file for
later printing.
1 8 The Computer Composer 185
In particular, canon requires that the composer specify three pa-

rameters before running the program: int, the interval of imitation, del,
the delay in starting time for the second melodic line, and num, the num-
ber of notes in each line. The notes of the line being generated are stored
inan array called melody, or melior short. Here is an algorithmic outline
of canon:
input int, del, num

mel(l)«-7
for i
= 2 to num
mel(i) <- 1
found <- false
while found false
increment mel
for i
= 1 to num — del
compare mel(i) and
mel(j + del) + int
ifharmonious then found <— true
output mel
option to continue
The algorithm starts by assigning to mel(l), the first note of the
canon, the value 7. This is the tonic note, and it not change. The re-
will
maining entries in mel all start at the value 0. A while loop tests a Boo-
lean, or logic, variable called found, which is initially set false. Found is
made true if discovered by the instruc-
a valid canonic melodic line is
tions in the body of the loop. First, the array mel is incremented. This can
be done by scanning the array from right to left. In the process of scan-
ning, the counting procedure looks for an array entry that is less than 13.
On finding such an entry, it adds 1 to it and changes all the entries to the
right (if there are any) to 0. This is precisely what happens in ordinary
counting, where 9 replaces 13. For example, 3572 + 1 = 3573,
3579 + 1 = 3580 and 3599 + 1 = 3600.
The next job of the loop is to compare each note mel(j) with the note
mel(j + del) + int. In other words, the program adds int to the note that
is del notes after mel(j) and looks up the difference of the two note
values in the table of rules for first-species imitation. If the difference is
considered harmonious in all cases from 1 to num — del, the Boolean

variable found is made true. (The notes beyond num — del are played
alone with no accompanying harmony. Thus there are no note differ-
ences to look up in the table.) Once such a melodic line is found and
printed out, the program asks the composer, "Do you want to con-
tinue?" If the answer program branches back to the found <—
is yes, the
false instruction and the count picks up where it left off.
One can, of course, play the melodic lines discovered by canon
through the tiny loudspeaker of one's home computer. Readers of a mu-
sical bent, however, will develop the knack of humming the line or of
transcribing it, along with its canonic companion, onto sheet music. The
canon can then be tested at another keyboard in all its harmonic glory.
Rhythm is a more sophisticated musical form that some readers may
realize; traditional Western music has never been very elaborate rhyth-
mically. Popular musical culture, on the other hand, has embraced an ex-
traordinary variety of rhythmical forms (Figure 18.4). Most of them
originate either directly or indirectly from traditional African or Asian
music. This includes most rock music, jazz, Caribbean, and Latin-Amer-
ican music. Westerners are also increasingly aware of the complex con-
tribution of the tablas (a pair of drums played by the fingers) to Indian
musical forms such as the raga.
The program I call beat enables one to spedify simple rhythms as
sequences of O's and I's (Figure 18.5). These are translated into sounds
by the simple expedient of running through the sequence repeatedly.
Each time through, the presence of a 1 triggers a brief tone pulse. A 0
triggers nothing.
Actually beat is simple enough to describe without further ado.
Structurally it is rather similar to solfeggio (the program that plays lin-
ear-congruential music). A single array called pulse holds the rhythm as

specified by the programmer at the start of the run:
input pulse, num, dur

for = 1 to 25
i
for = 1 to numj
k<- 1
while k > dur

k <- k + 1
if pulse(j) = 1
then sound
The variables called num and dur refer respectively to the size of the
input array and the duration between sounds. The outer loop specifies
that the basic rhythmic interval determined by pulse will be played 25
X
1 8 The Computer Composer 187
CYMBAL
ROCK: SNARE DRUM
BASS DRUM
1010101010101010
00001001 01 001001
CYMBAL
SNARE DRUM
1011 0000101 0100 1 BASS DRUM
HI HAT
REGGAE: SNARE DRUM
BASS DRUM
1 0 HI HAT
0 1 SNARE DRUM
0 0 BASS DRUM
CYMBAL
JAZZ: SNARE DRUM
BASS DRUM
10 0 0 0 10 CYMBAL
110
1 1
110
1 1
0 0 0 1 0 SNARE DRUM
10 0 0 1 1 0 0 0 0 1 BASS DRUM
X— SNARE DRUM
SAMBA: SNARE DRUM RIM
BASS DRUM
HI HAT
1 o 1 1 0 1 1 1 1 1 o 1 SNARE DRUM
0 1 0 0 1 0 0 0 0 0 1 0 SNARE DRUM RIM
0 0 0 0 1 0 0 0 0 1 BASS DRUM
0 1 0 1 0 0 1 1 0 HI HAT
Figure 18.5 Four sample rhythms for beat.
times. This number can easily be altered by readers who stumble onto
rhythms they would like to hear for a longer period of time. The next
inner loop controls the array index; the algorithm will consider each
entry in turn in order to decide whether or not to play a tone. How long
to wait between sonic events? That much is determined by a special
waiting loop that simply counts up to the specified duration, dur. Then, if
pulse(j) is 1, the program beat will play a tone, a buzz, or whatever may
be available. If dur is small, the rhythm will be fast. If dur is large, the
rhythm will be slow.
Like its predecessors, beat can be enhanced. Something like a

also
percussion ensemble is possible if low ones
different tones (preferably
representing various kinds of drums, bells, or cymbals) are employed.
Although the sounds may be far from realistic, the rhythms will be the
real McCoy. One can complicate the array called pulse by using integers
such as 0, 1, 2 and 3 to represent silence, a high drum playing alone, a low
drum playing alone, and both drums playing together. In each case the
entry of pulse being examined in the inner loop must be decoded by a set
of if statements that control the playing of no sound, one sound, or both
sounds. In the last case two notes are sounded consecutively as a substi-
tute for simultaneity. If it is found that playing two consecutive sounds
throws the timing off, then the waiting loop must be moved inside the if
statements so that there is one copy for each of the four possibilities. In
the case of no sound and two sounds, the limit dur must be replaced by
new limits, one longer and one shorter than dur. In both cases the differ-
ence would be the time value assigned to an individual note.
A more sophisticated version of beat will certainly be able to play
some of the interesting rhythms displayed in Figure 18.4. The various
scores shown there have been interpreted as sequences of discrete
events strung along a common time base. They are easily translated into
the contents of pulse under encodings arranged by the aspiring com-
poser.
In addition to the preceding forays into melody, harmony, and
rhythm, I continue to be fascinated by the notion of endowing every
computer program with two rules. In one role the program computes
what it computes. In the other role the program is fitted with a tone or
two next to its inner loops, outer loops, and conditional (if) statements.
The program has a song to sing for each problem it is given. Those who
listen regularly to their favorite program (be it recreational or commer-
cial) will develop an ear for its performance. I have no doubt that some
bugs in new versions of a program may be detected by ear. The program-

mers will thank me for putting this bug in their ear.
Chaos ii i In
Extraordinary how potent cheap music is.
NOEL COWARD, Private Lives
F
JL ew readers have heard of Victor Chaosky His music has an other-
.
worldly quality that is hard to describe. But critics have yet to call it un-
musical. His April One Suite, though played by a single instrument that
sounds at times suspiciously like the tone generator on a home com-
puter, leaves no doubt of the composer's brilliant, chaotic talent.
Some people have attempted to imitate the music of Chaosky, all
with amazing success. One of these is Arthur Davidson, a low tempera-
ture physicist at IBM's Yorktown Research facility in Yorktown
Heights, NY. He uses the famed logistic equation, the simplest chaotic
system known, to produce the notes from his personal computer, an
IBM PC running turbo Pascal. He notes that more than one visitor to his
office has remarked (in almost the same words on each occasion), "Say!
Isn't that Philip Glass?"
They may have heard of the composer Glass, but not, evidently, of
From time to time, Davidson runs the program he calls
Victor Chaosky.
CHAOS IN A MAJOR (or CHAOS-A for short), listening intently. This
is the sound of chaos reduced to a manageable scale. Interesting musical
patterns seem to repeat themselves with slight variations when, sud-
denly and unaccountably, they change gears as if Chaosky had aban-
doned his theme in mid-passage, inspired by a new musical idea. But the
patterns are not merely musical. Can Davidson hear in them a hint of or-
Among his missions at IBM
ganization to be exploited or learned from?
Research, he studies non-linear dynamics, the behavior of physical or
mathematical systems that do not have a linear response to their state
variables, but a non-linear (e.g. quadratic) one.
189
As a child, Davidson had studied the violin and still likes to "fool
around" on the piano. "Anyone who has heard me will tell you that I am
not a musician/ he says. A few years back, Davidson was heavily in-
'
volved with the logistic equation, staring at the famed map by day and,
as usual, playing around with musical scales in the evening. "Eventually,
I two things together." The rest is history of a sort
got the idea to put the
that some future student of simultaneous invention may want to ex-
plore: At the same time, others were trying exactly the same idea.
There are a few versions of the logistic equation, all equivalent. The
one favored by Davidson has the fewest symbols in it:
x <— x(a — x)
I have put the equation form of an algorithmic assignment

in the
statement so that readers may see more clearly what the equation
"does." The current value of the variable x is subtracted from the param-
eter a and the result is then multiplied by x. In the final step, the value
just computed replaces the previous value for x and the whole process
begins anew. To make a short story miniscule, when you put this state-
ment inside a loop and translate the resulting sequence of values into
musical notes, a wonderful new composition by Chaosky emerges.
The logistic equation does not represent a real-world physical sys-
tem as such, although it vaguely echoes the behavior of a predator-prey
dynamic; the more there is of x, the less there is of a — x. The logistic
equation represents only the simplest version of an infinite number of
non-linear dynamical equations some of which describe the behavior of
real-world systems and some of which do not. Since the phenomena to
be probed are in a sense common to all such systems and since the phe-
nomena are primarily mathematical in any event, it is not particularly
important that they be "real." Chaos is chaos.
In the logistic system, the presence or absence of chaos depends on
the choice of value for the parameter a. Ranging from 1 to 4, this parame-
ter alone will determine a remarkable number of different behaviors. At
a =1, for example, no matter what starting value of x less than a is cho-
sen, the sequence determined by the logistic equation will converge rap-
idly to a single value (Figure 19.1). This represents an equilibrium value
for the system, like a vibrating string coming to rest in a single position.
There is a way to visualize the process precisely. In Figure 19.2 readers
will see a set of nested parabolas. Each parabola represents the quadratic
equation that results when a different value of a is frozen. Take out a

pencil and ruler and try the following experiment on the smallest parab-
ola, the one labelled 2.9.
This parabola represents the dynamics of the logistic system when
a = any value you like on the jc-axis between 0 and
2.9. Starting with
2.9, draw a line straight up to intersect the curve, then from the point of
intersection rule another line across to meet the diagonal. This defines
another point, and from that point draw a third line vertically to intersect
the parabola again. Continuing to draw this sequence of horizontal and
vertical lines that connect the diagonal and the parabola, you will see a
pattern emerge. A square spiral of line segments will converge on a spe-
cific point of the parabola. All of the vertical lines in the sequence repre-
0 12 3
Figure 19.2 Each parameter value produces

4
new musical possibilities.
sent successive values in the logistic formula. The horizontal lines repre-
sent iterations of the equation. In the parabola labelled 3.1 two points
are marked. These comprise a two-point attractor representing the equi-
librium behavior of the system when a = 3.1. Successive parabolas
yield a geometric interpretation of what is happening in the logistic sys-
tem at certain values of the a-parameter. The largest parabola is a sample
of chaotic behavior. Only 100 points are shown but if the iteration had
continued, the parabola would soon be black with them! Chaos.
This situation corresponds to a slice across the bifurcation diagram
in Figure 19.1 at the level a = 3.88. On the way to chaos, readers who
simply write a program that produces and plots successive values of the
formula as points on a horizontal line will notice the well-known period-
doubling phenomenon.
If your program is written to accept different values for a from the
keyboard, then to run the equation for, say, 100 iterations, it will pro-
duce values that converge on a one-point attractor for any a up to ap-
proximately 3. After that a two-point attractor shows up. For example,
at 3.1 successive numbers bounce back and forth between two limiting
values, converging to them. At a = 3.5, there is a four-point attractor. At
still higher values there is an eight-point attractor, a 16-point attractor,

and so on. The period doubling builds up at such a frantic pace that long
before a gets very much bigger than 3.6, the system begins to behave
very strangely. Suddenly, the attractor consists not of some power of
two but of an infinite number of points. The attractor is no longer simple
but "strange." Here is where music enters the picture literally! —
If you play with the program I have just described, you will notice, as
Davidson did, that when a is increased above 3.57, the chaotic points
spread out to cover more and more of the parabola. Davidson's descrip-
tion of what he saw is colorful enough to repeat.
"Now in playing around with these things on a PC, I noticed that
iterating the map in the chaotic regime would not exactly repeat a value,
but it would frequently almost repeat a value. In fact it would almost re-
peat whole sequences of values." (see Figure 19.3.) "Aha! This is where
I heard the bell ring. It was exactly how I had come to regard music! To
maintain your interest, music must represent some sort of pattern that
you recognize. . . .
As Davidson goes on to say, variations in the pattern will sustain the

interest even longer. This made an entirely new program necessary. He
divided the range of x-values up into subranges, one for each note in the
diatonic (major) scale from A
to A. Then he set to work.
A bare bones chaos composition program can be written on the basis
of a seven-line algorithm. Inside a loop which may be one of several
types, including even a simple for-loop, iterate the logistic equation
once, translate the value produced by the equation into a frequency,
then play the tone that corresponds to the frequency. Figure 19.4 shows
the essential steps based on a repeat loop that can only be terminated
when the user presses a key.
Within the loop, when the next value of x is calculated, it is con-
verted into an integer between 0 and 12 called step, then into a fre-
quency called tone.
The notation in which this last calculation is expressed leans slightly
in the direction of turbo Pascal. Essentially, a base frequency of 220
Herz (middle-A on the musical scale) is multiplied by 2 to the power of
step 112. The function called exp denotes taking a power of the transcen-
dental number e, then correcting with a factor of ln(2), the natural log of
2. So much amounts to a power of 2 and what follows is just the power
taken, namely, step/12. If step = 0, the base note ,
A, remains the same.
If step = 1, 220 multiplied by 2 to the power of V12. This is
the result is
one-twelfth of the way up the semitone scale from A to the A one octave
higher. Thus, when step = 12, the frequency of 220 is multiplied by 2 to
yield a frequency of 440, an octave above 220.
input x and a
repeat
x <— x*(a — x)
step <— int(12*x/a)
tone «- int(220*exp(ln(2)*step/12)
play tone
until key pressed
Figure 19.4 Chaotic music algorithm.

The twelve-tone semitone scale will sound just a bit too "modern"
for some readers. To get a more pleasing diatonic scale, the basic calcu-
lations must be changed. First, the value of x will have to be transformed
into an integer between 0 and 8 (0 and 7 will serve about as well). The
resulting number, step, will then be converted to a tone by adding an in-
teger that causes the appropriate semitones to be skipped. Under this ar-
rangement, the needed tones are: 0 (A), 2 (B), 4 (C#), 5 (D), 7 (E), 9 (F#),
11 (G#), 12 (A). The problem of writing the one or two lines of code that
will convert step into tone can, I think, safely be left to most readers.
How do you get sound out of your system? Davidson uses the turbo
Pascal command called "Sound." Specifically, Sound (220) or more gen-
erally, Sound (freq) will produce a tone of middle A in the first case and a
tone of whatever frequency happens to be specified by the variable freq
in the second case. In order to hold the tone for a time, Davidson inserts
a delay loop (" delay (n)" in turbo Pascal, where n is in milliseconds). The
sound must, after this delay, be turned off. The next step in CHAOS A
will therefore be the turbo Pascal command "Nosound." Anyone pro-
gramming in BASIC can use a somewhat simpler command: SOUND
freq, dur. By this I mean that SOUND is a function that has two argu-
ments. The first argument, whatever you call it, is either a number or a
variable that specifies the frequency in Herz of the desired note. The
second argument specifies the duration in "clock ticks," each BASIC
clock tick lasting approximately Vis of a second.
There are a dozen useful features that can be added to the basic
CHAOS A program to make it a genuine tool for the intellect, not to
mention a flexible music maker. Here is what Davidson's version of the
program does:
1. makes a musical graph of the notes being played

2. interactively allows new values of a to be entered
3. gives the user the option of supressing the music
4. gives the user the option of transposing the music up or down
Davidson is the first to admit that the program can be developed well
beyond this point. In fact, there is no need to stick to the logistic equa-
tion. There is a wonderful book by Francis Moon (see Further Reading)
that lays out a number of other simple equations, including one based on
the circle. I await the results of reader programs with bated breath, as
does Davidson.
Figure 19.5
Who else has tried a chaos music program? Ralf D. Tscheuschner at

the Institute for Theoretical Physics in the University of Hamburg, for
one. Tscheuschner, a physicist, has collaborated with Steffan Schindler,
a computer science graduate student at Hamburg, to produce an interac-
tive music composition system based on the logistic map. Of all the cha-
otic music I have heard, these compositions seem the most musical.
After all, human touch, getting the best of both
they incorporate the
worlds from chaos and the human musical mind. I have included a sam-
ple composition of theirs (and the program's) entitled Fuguette Vegetar-
ienne (Figure 19.5). Not Chaosky's work but, one would have to say, in-
spiredby it.
With such profound musical developments, who can doubt that the
music of Chaosky (in one form or another) will soon reverberate from
thousands of homes and offices!
Further Reading
Francis C. Moon. Chaotic Vibrations. John Wiley & Sons, 1987.
20
When meet someone on a
I professional basis, I want
them to shave their arms.
Mark V. Shaney, net. singles conversation
in Boston on a chilly November day in 1991 a press of media milled

about the auditorium at the Computer Museum, a converted warehouse
in the Boston harbor district. The museum, which houses the world's
largest exhibition of computer history and hands-on displays of com-
puter technology, was hosting the first annual Loebner Prize Competi-
tion. In another room of the museum ten judges communicated via type-
written messages with entities that were either computers or humans.
Their job: Decide which. It was the first public trial of artificial intelli-
gence based on the famed Turing test.
The Loebner prize, established by New York businessman Hugh G.
Loebner, was to be awarded to the program that fooled more of the
judges than any of its competitors. In this restricted form of the Turing
test, the topics of discussion were strictly limited. The judges, ordinary
people off the street, were policed by referees who were experts in com-
puting and artificial intelligence.
In the grand auditorium, ten screens displayed the conversations as

they progressed. Each screen displayed the text flowing through a partic-
ular terminal. And whether the terminal was connected at the other end
to a distant human or computer, the domain of discourse appeared as a
label on the screen: fishing, women's clothing, martini making, personal
relationships, small talk, the plays of Shakespeare, whimsical conversa-
tion, and so on. Every eighteen minutes, each judge got up, moved to a
new terminal and began a new conversation. He or she carried a score
199
pad on which each entity would ultimately be ranked from most human
to most computerlike. Behind the scenes museum officials wrung their
hands and muttered. There had been an hour delay setting up some of
the lines to long distance participants. What if no program fooled any of
the judges?
The audience moved from screen to screen, laughing here and
groaning there. Some people had no doubt that the screen for terminal
five (labeled whimsical conversation) belonged to a program:
JUDGE: Hello! You seem to be feeling quite witty today.
TERM: Hello, judge, what is your mood today? happy? sad? ready for
—
some fun? Hello how's the weather up there today? Let's talk about
psychology!
JUDGE:It has been a long day! Now think I'm ready for some fun. The
I
weather isn't very pleasant today. Do you have any jokes that could lift
the clouds?
TERM: Everybody talks about the weather but nobody seems to do much
about it. Why do you need to tell me you think you're ready for some
fun? Groannnnnn. . . .
JUDGE: Well you did ask me!
TERM: Now you have really outdone yourself in asking stupid things!
Woof . . . woof . . . woof. . . . Why do you think did ask you?
I
Other members of the audience who watched this screen weren't so

sure. Perhaps it was a human going "woof, woof." Some humans, after
all, are better than others at whimsy.
If the entity behind screen five was a computer, who could doubt
that the entity behind screen four (the plays of Shakespeare) was a
human?
JUDGE: Tell me what your favorite Shakespeare play is.
TERM: The Tempest, how about you?

JUDGE: Julius Caesar.
TERM: Ah, of course, that play would better be called Brutus, but the title
of Julius Caesar was and is more saleable.
JUDGE: Why do you think it would better be called Brutus?

TERM: Becauseit's about Brutus! Caesar dies pretty quickly, and except
death scene he hardly has any scenes. Maybe Shakespeare

for his
played Caesar, he always wrote himself a nice little role to play.
Here it seemed obvious that the entity behind screen four was meet-
ing the questions andcomments of the judge head on, so to speak. As this
and other conversations with judges revealed, the entity was quite
knowledgeable about Shakespeare in a way that expert systems, even
equipped with all the same facts, could hardly express in fluent English.
By the same token, however, audience members who were less conver-
sant with computers and AI were more willing to believe that its exten-
sive knowledge betokened something of the superhuman, something of
the computer, in other words.
Museum officials needn't have gnawed their nails about the out-
come. Five of the judges ranked the whimsical conversant on the human
side of a line that was supposed to separate the humans from the com-
was revealed at the end of the Loebner Prize competition as
puters. All
cameras whirred and flashbulbs crackled. Headlines could now read:
COMPUTER FOOLS HUMANS
The winner was a program called PC Therapist, the creation of Jo-

seph Weintraub of Woodside, N. Y. Specializing in whimsical conversa-
tion, the program is meant to beguile the hours of lonely or depressed
people with a kind of rollicking if somewhat evasive companionship.
The cash award on this inaugural event was a mere $1500. In a later
competition, when the full Turing test is to be implemented, the award
will climb to $100,000. But only if a program takes at least one (expert)
judge the distance in a more or less unlimited topic area. But never mind
the small award. As one reporter whispered to another at the edge of the
crowd, "He'll clean up on this!"
PC Therapist resembles the famous program Racter. Racter, short
for "raconteur," is program that uses immense files to create
a Fortran
chains of association that, over the short term at least, give the impres-
sion of topic orientation. But like Racter and like Eliza, the famed psy-
chotherapeutic program before it, PC Therapist frequently gets the
grammar wrong when it attempts to mirror statements made by the hap-
less user.
"Oh, you say you think I are evading the question!"
In view of these shortcomings, it seemed a little unfair to some of the
AI experts who refereed the contest that there should have been a cate-
gory called whimsical conversation in the first place. It gave an unfair
advantage to any program entered in it. Whimsy, after all, is a very broad
and forgiving category. Who is to say that grammatical errors shouldn't
be a part of whimsy? Evasion and frequent topic changes also take a
more or less natural place here and an inexperienced judge is all the
more prone to fall prey to a program that asks about the weather and
goes "woof woof."
But the potential of whimsical conversation is something that future
entries in the Loebner Prize Competition might want to take advantage
of. Enter a potential candidate.
Some years ago, the bulletin board service called Net. Singles was in-
vaded by an obstreperous program called Mark V. Shaney, the brain-
child of Don P. Mitchell and Bruce Ellis, then with the AT&T Bell Labo-
ratories in Murray Hill, New Jersey. Ellis and coworker Rob Pike had
noticed the banalities of typical Net. Singles postings and thought it
might be fun if Mark V. Shaney were unleashed as a participant.
First, however, Mark had to be fed substantial volumes of typical
postings. This was quite easy to do since they were already in machine
readable form. Mark scanned the text of these conversations, building a

rather large table in the process. For every pair of words that occurred in
the conversations on dating, make-up, mixed marriages, and related
topics, the table developed a series of entries, once for every word that
followed the given pair.
For example, the pair "find a" might be followed, here and there in
this large volume of text, by words such as "date," "person," "dollar,"
and so on. Some follower words would occur more frequently than
others and for each of these the table would contain a probability based
on frequency. Thus if the word pair "find a" occurred 7 times in the text
and it was followed 3 times by "date," then the table entry opposite the
ordered triple (find, a, date) would be 3/7 = 0.429. And so with all other
pairs of words and their followers in the text.
Simple in principle, Mark V. Shaney consists of two parts, a table
builder and a text generator. After scanning an input text and construct-
ing the table of follower probabilities, Mark V. Shaney is ready to "talk."
It begins with a single pair of words. The generating algorithm is simple.
repeat
r «- random
determine pair follower
output follower
<— second
first
second <— word

until someone complains
When a random number r is selected, it determines a follower by the

process of adding together the probabilities stored for each of the words
that follow the given pair until those probabilities first equal or exceed r.
In this way, each follower word will be selected, in the long run, with a
frequency that reflects its frequency in the original text. And in this way,
the text so generated bears an eerie resemblance to the original:
"When I meet someone on a professional basis, I want them to shave
their arms. Whileconference a few weeks back, I spent an interest-
at a
I wouldn't dare take them seriously! This
ing evening with a grain of salt.
brings me back to the brash people who dare others to do so or not. I love
a good flame argument, probably more than anyone . . .
"I am going to introduce a new topic: does anyone have any sugges-
tions? Anybody else have any comments experience on or about mixed
race couples, married or otherwise, discrimination forwards or reverse,
and eye shadow? This is probably the origin of make-up, though it is

worth reading, let alone judge another person for reading it or not? Ye
gods!
Mark V. Shaney can continue virtually forever in this vein, spewing
out an endless Markov chain (hence the name) of probabilistically de-
termined word followers.
Network opinion did not take long to crystallize. Mark V. Shaney
was either a severely disruptive individual, not to mention a psycho-
path, or else was someone's pet AI project that had got out of hand. As
it
one complainer put it: "Will someone please pull the plug on Mark V.
Shaney?' Given Mark's relative success on Net. Singles, might there be
'
some hope for a near relative in next year's Loebner Prize Competition?
It all depends on what I mean by a "near relative."
The table of letter followers created by the Mark V. Shaney program

is three-dimensional because every pair of words in a received next has
one or more followers that are also words. But what is true of pairs of
words is also true of triples. Every triple of consecutive words that occur
in the text is followed by a fourth word. This suggests that one might con-
struct a four-dimensional table and generate, in consequence, text that is
even closer to the language of the received text.
There is no reason, of course, why a second-order Mark V. Shaney
could not be constructed to operate on triples of words rather than single
ones. Let us call her Shirley Shaney. I expect that Shirley will sound
somehow more sensible than her disruptive brother. Readers with a
modicum of programming skill have enough advice to build a Shirely
Shaney. Warning: Once Shirley is written and debugged, she will require
rather large volumes of text to get started. An entertaining form of input
would involve conversations between Shirley and her maker. Cycles of
table building would alternate with sentence generating while Shirley
slowly "learned" her creator's conversational pattern. At such a pass her
creator could then watch for the announcement of the next Loebner
Prize Competition.
Fi Sri
I never forget a face, but in your case I'll be glad to
make an exception.
GROUCHO MARX in Leo Rosten's People I hove Loved,

Known or Admired
T JL he face is unmistakable. There are the low, floppy ears, the prom-
inent cheekbones, the high pompadour. Ronald Reagan's face is familiar
around the world, but somehow it is even easier to spot his likeness in a
caricature than it is in a photograph (Figure 21.1). Surely the art of cari-
cature calls for deep insight into human nature. If this be the stuff of
computation, surely the computer is a trivial adjunct —little more than a
sketch pad —that merely stores the highly subtle renderings of the cari-
caturist in a visual form.
Or is it? The caricatures on these pages were all generated by a pro-
gram devised by Susan E. Brennan, a staff scientist at the Hewlett-Pack-
ard Laboratories in Palo Alto, Calif. To run the program a mouse, a light
pen, or some other analogue of a pencil might be convenient, but they
are certainly not essential. The results depend hardly at all on a steady
hand or a practiced eye. Instead, once a photographic likeness of the face
is entered into the computer, the program takes over and draws the cari-
cature. How is it done? A short answer is deceptively simple: the pro-

gram compares the photograph of the target face with an average face
stored in the memory of the computer. The features that differ most
from the average face are scaled up in size.
Brennan' s program followed naturally from her own considerable
abilities as a caricaturist and her interest in the cognitive processes un-
derlying face recognition. Such processes have long baffled psycholo-
gists and cognitive scientists, and caricatures seem to play a special role
205
Figure 21.1 Fromrealism to "facelessness" in FACEBENDER, a pro-

gram based on the work of Susan E. Brennan, with stops for caricatures
of Ronald Reagan in between.
in the process because when they are recognized, they are recognized
almost instantly. Could it be that instead of remembering a friend's face,
we remember a caricature of it? To address these issues Brennan in-
vented her simple technique for generating caricatures, and she de-
scribed it in her master's thesis at the Massachusetts Institute of Tech-
nology. She continues that interest in her spare time; in her working
hours she now experiments with new forms of communication between
21 Face Space 207
human and computer that rely in part on natural language understand-

ing.
Conceptually Brennan's technique is closely related to a trick that
computer animators call in-betweening. Imagine two drawings of famil-
iar objects, such as an apple and a banana, both done by connecting dots
with lines (Figure 21.2). Each dot on the apple is then paired with a dot
on the banana. If the paired dots are also connected by lines, the mid-
points of the lines depict a brand-new fruit that splits the difference be-
—
tween the apple and the banana- a banapple, of course.
The same lines that connect the apple and the banana can also give
rise to an extreme form of the banana— from the point of view, so to
speak, of the apple. Extend each line beyond the banana by half its origi-
nal length and then place dots at the end of the lines. When the dots are
connected, the banana emerges in caricature. Similarly, by projecting
the connecting lines beyond the apple, one can obtain a caricature of an
apple —
from the point of view of the banana. Faces can be treated in
much the same way. Each pair of faces defines two mutual caricatures.
The best caricatures, however, arise from comparison with a norm, or
average face.
The norms in Brennan's program are made up from sets of several
dozen real faces in a data base of several hundred. Points are chosen that
outline the features of each face, and the points are labeled with respect
to a set of matrix-based coordinate axes. The origin is at the upper left of
the image plane, and the coordinates increase from left to right and
downward. The scale is adjusted so that the left pupil is at the point (135,
145) and the right pupil is at the point (190, 145). The coordinates of
corresponding points on each face are averaged to give the norm for that
point. For example, the combined coordinates of the outer corner of the
left eyebrow give the average coordinate for the outer corner of the eye-
brow of the average face. Three norms are constructed in this way: there
is an average male face, an average female face, and an average, overall
plain-vanilla face. It is no surprise that the plain-vanilla face looks some-

what androgynous; it establishes the norm for most caricatures (Figure
21.3).
To draw a caricature based on the norm, the program must be sup-
plied with a digitized version of a real face. In practice the face begins as
a photograph, and the program prompts the user in turn for each of the
186 key points on the photograph. For instance, when the program calls
make up the left eyebrow, the user can respond by
for the six points that
Figure 21.2 How to turn an apple into a banana, and beyond.
208
21 Face Space 209
Figure 21.3 The androgynous average face.
moving a mouse to successive points on the left eyebrow of the photo-

graphic image on the screen.
It is useful to think of Brennan's program as a fast shuttle for explor-
ing what she calls face space. The entered coordinates for the points de-
fining a photograph can be strung together in a predetermined order.
The result is a list of numbers that can be treated as coordinates of a sin-
gle point in a high-dimensional space. For example, both the average
faces and the photographic face are represented by 186 points, each of
which has two coordinates. The resulting list of 372 numbers for each
face is a point in a 372-dimensional space. In principle every face can be
assigned to a point in face space, and any two faces in face space can be
connected by a straight line.
There is no need to be mystified over the concept of a higher-dimen-
sional space. Face space is merely a handy abstraction for describing dif-
ferences and similarities among faces. The familiar concepts of the
straight line and the distance between two points have straightforward
analogues in any higher-dimensional space. All the points along a
straight line in face space represent proportional changes in each coordi-
nate value. The distance between two points in face space is a measure
of their similarity: similar faces are close neighbors in face space, and
dissimilar faces are literally farther apart.
In face space one can imagine the norm as being near the center of a
cloud of points representing realistic images of real faces. A line joins
each real face to the norm. The points along the line correspond to a suc-
cession of intermediate faces that look increasingly like the real face.
Beyond that face are the caricatures, but there is a natural limit to recog-
nizable exaggeration: the caricatures eventually lose their human quali-
ties and degenerate into a chaotic state Brennan calls facelessness.
The idea that every face is a point in face space suggests another fas-
cinating transformation. Since any two faces in face space can be joined
by one can ask the program to generate a transitional se-
a straight line,
quence from one face to another. Brennan finds such sequences particu-
larly intriguing when the two endpoint faces are male and female: the
program effortlessly transforms Elizabeth Taylor into, say, the late John
F. Kennedy (Figure 21.4).
The reader can duplicate some of Brennan's feats of caricature by
writing a smaller version of her program; I facebender. It re-
call it
quires the user to supply at least two faces: a norm and the target face to
be caricatured. I have referred above to the norm, whose coordinates
have been generously provided by Brennan (Figure 21.5). The user must
then convert the target face into the same form. In the absence of so-
phisticated digitizing equipment the reader can, with relatively little
pain, convert a photograph of a loved one (possibly oneself) into a list of
coordinates. Brennan warns, however, that the face in the photograph
must have a bland, neutral expression; even a slight smile will grow to a
monstrous grimace. The face must also be fully frontal; if the head is
turned, facebender will turn it even more.
To determine the scale for the axes, assume the coordinates of the
left and right pupils are the same as the norms: the left should be at (135,
145) and the right at (190, 145). (Remember that horizontal coordinates
increase from left to right and vertical coordinates increase downward.)
Once the distance scale is established the user must find the rest of the
coordinates by careful measurement. In Brennan's digitizing scheme the
points on the face are organized into 39 facial features; each feature is a
Figure 21.4 Elizabeth Taylor (as Cleopatra) meets former President

Kennedy in face space.
21 Face Space 211
succession of connected points. The order of the points depends on the

orientation of the feature: for features that are mainly horizontal the
points are listed from left to right, and for features that are mainly verti-
cal they are listed from top to bottom.
Brennan admits that identifying the key points on a face is governed
mostly by trial and error, but the former president's face can be used as a
guide. For this reason it is important that the same person carry out the
conversion from photograph to list for each face entered into a data
base.
facebender stores the two and
digitized faces in arrays called face
norm. A needed to crate a display. All three
third array called disp is
arrays have 186 rows and two columns: one face point per row and one
coordinate per column. Points are arranged in the serial order given in
the list The advantage of this ordering is that all lines in the
for norm.
final picture can then be drawn between successive points in the array;
of course, lines are not drawn between successive array points when one
feature is complete and another is about to be drawn.
The first feature the program draws is the pupil of the left eye; the
second feature is the right pupil. Each pupil can be rendered as either a
dot or a small circle; somehow the circles look friendlier. For the re-
maining features, however, lines are drawn to join consecutive points in
the array. A special array called features is needed to skip the line be-
tween the last point in one facial feature and the first point in the next.
The array gives the number of points in each feature, and a double loop
supervises the skips (Figure 21.6).
Because the two features have already been drawn, the display
first
routine begins with the third feature, namely the left iris. The first point
in the left iris is the third point of the array disp, which is indexed by the
variable i; hence the value of i is initially set equal to 3. The array features
is indexed by another variable, and it ranges from 1 to 37 because there
are 37 features left to draw. Within the ; loop another variable called
count keeps track of the number of lines drawn for each feature; it in-
creases by 1 with each passage through the ; loop. The index i is also in-
creased with each passage through the loop; it identifies the point in the
array disp thatis currently participating in the frantic expercise of con-
nect-the-dots.
Inside the loop is a second loop called a while loop; it compares the
number of points joined so far in feature ; with the total number of points
in that feature. The program leaves the while loop when the two num-
bers are equal; the feature is complete. If there are still points to connect
CM CM CM CM NO IS ON CO
00
rH
00
rH
00
rH
00 co CO
rH rH rH
On
rH
©
CM
rH rH rH co" cm" rH is" oo"
SO no NO NO m rH IS ts
rH rH rH rH' rH CM
V —
rH
^
rH
—
f s ^ s ^^-^ ^ s
rH rH ON ON ON ON rH CM IS CM
rn X 1
t> S> IN
_h
rn n CO CO
V i
n ON
rn
©
CN
in CD no" is" co" o" o" oo" oo"
©"
co on
rH rH
m NO
rH rH srH rH'
m m ©
rH CM
SO
rH
IS
rH
—
v —
s — >w >>w' s —
^ N ^ ^
CM CM CM CM 00 00 NO IS ON CO
_h
rH ; j
t> IS 00 00 CM CM CO co ON
rn _J n n oH
rn rH n
rH ri n i
©
CN
©" NO oo" oo" On" On" CM
©" cm" co"
On
tH
m SO CO ON rH NO
rH rH rH rH rH rH rH CM rH
m SO
rH
ON Q\ IS SO CO CM CO O mSO 00 00 IS ON CM NO CM
rH pH rH
-or <Oh ©
rH , —
WO
— in sO l® IS is CM CM CO co ON
rn _in — ~
:
1
i
1— — —
< rr rn
.
!
i
©
CN
rn oT Is" rH ©o" no" so" so" no" Is" m" NO" oo"
©" no" so"
co 00 o ON
rH rH rH CM rH 00 rH CM rH rH
e o m so
rH rH
IS CM 00 co in
rH rH rH rH rH CM rH
© m
rH
rH © © ©
is to CO CO co CM CM CM CM ON CO
^> HiO CO CO in in an is CO CO CO co On o
rH
oo"
rH rH rH rH rH rH
00 ©" co" rH CN rH in
© rH rH rH rH rH rH rH rH rH rH
is" so" so" Is" no" CO" co" is" On"
CN
oo"
(N GO m ON © m © ©o m CO I** rH IS CM 00
rH rH ?H rH
rl rH^ rH d rj^ rH .rH rH^ rH, rH rH^ r-^
CM
rH^
rH rH S* SS S. co fO rH On ON is NO © O © 00 SO CO CO
©
CM
LO in © ©
NO NO CO CO
rH rH rH rH rH rH rH rH rH rH rH rH rH rH rH rH
© CO
rH
CO ©
rH CM is" CM
©
00
ON © rH
CO On CO ON rH IS CM
00 NO IS 00 NO NO
is cm is no
CO CM
is rH IS m ©
m CM
rH
rH
co CO —
s>
ooooooooooooooooooooo
PhPuPhPhPhPliPliPliPhPhPliPhPhPuiD-iPhPhPmPLiPhPli
o
Ph
rHrHmincOCOCOCOCOCOCOCOsOsOsOsOsOsO'^-Tt-lS IS
o
a 3 pci
PQ
O pq Ph
00!
pq o&
PQ
pq
w H
[Tj
O
co
CO p^
PQ pq
H DC OZ pq >^
>< pq
X pq
Ph
W O PJ
a Ph
Ph
hJ S3 H E-h
2 P^
D
H H
p^ pq
P0 PH o w Ph Ph Ph Ph
d E O O
pq
Q pq
O hJ oD Ph
o
Ph D 5 PJ
Ph Ph
9 oo
Ph S PH
H Oooo CO
o o oo o
H o o H DC H H E-h
H H H Ph Ph Ph O Ph H H Ph E-h
OOOOO w O O Oo O
PQ H E-h 2 H PQ PQ PQ
212
' ' H
00
-Itn
CN
o
(N
o
CN
"^t
O 00
rH
ro
CN
^ so
CN CN 00"
§
On ON rH ^ CN-
CN
CN
NO
rH
w
rH
CO
rH rH
PCS
00 00 rH <N
CN
o
(N
rH
CN
ON IN
ON, rH (N ON
of
CN
00
v£)
^ ~ rH
O rH d>
—
(N
o
N. co '
ro _o
rH £ cn W^
rH
ON CN
m
in <n
ON
CO
o
^—1
rH
^ O
ON VO
ON ^
(N <N CN "2 CO
CO CO <N
SO NO ro
T~H rH1
O
o m CN
CN
o o
rH CO
ON
in rH
04 OrH ON
ON ON ON 00
O ON
rH
CN CN ON rH r^ rH rH CN
ON ©~ CN
so"
m in
m o ^ rO
in rs
rH ^rH
vo" tn
CO ^0 On
co
IN
00 w
rH rH^ CN
o
00 ^ 0> CN
00
m Tt CN CO «fr CO CO CO 00 SO
©
CN
o
CN
NO
rH
ON
ro
\£i 00 00 00 00 On ON CO rH
rH CN CN
« « * «
on oo~
-Iro
rH
o w
t^s
rH ON
CO 00
m O
On "3-
rH
m" cn" CN
NO NO NO
rH ^
CO
rH
ro
CN (N _ W
rH
ON
^ £\
co
00
CN
HHM
g (N 5" <N t+ LO NO 00 00 NO NO CN 00
CD
CNrs JO
00
m rH OS
53 ^ IN IN IN
00 ^"1 ^"1
00 00 CO rH
rH^ rH CN CN
00 rH
ro >— $5i ON CN
ON CN
fN
o w S-iro"
ON
^
r-\ in
00" in 00"
5J- IN
rH rH rH CN
rH O oT m"
m
rH
so no
cm" co"
m
00 CO CO CO CO CO CO CO CO cocococococococo
H H hh hh H H H E-hHHHHE-hHH
2 2 2222 2 2 2 22222222
O
PL,
O
Oh
OOOO
pL, Ph Ph Ph
00
Ph Ph
0
Ph
00000000
PhPhPhPhPhPhPhPh
IN IN CO CO IN rH CO CO COCOCOCOCNCNCNCO
Ph
w 2
pc
w
u pq 53 Ph
<
Ph
2 pq 2 J Ph
3 2 O
Ph*
w o Ph
Q O PQ & 3
O PQ M pc5
PJ H
o
Ph
O PJ
Q p< 2 w Ph Ph
< Ph Ph
DU J
-7
s
Ph s H 00 PJ 2 Ph
EC
u -I
o o H O H
H X H X 2 2
Ph H Ph CJ o I— Ph Ph
O O pq a < O PJ 2 55 £
H PQ hJ p< PC H Ph U U
213
ni<;PT AY ROIITTNF FYAOriFR ATTOM ROTTTTMT7

i <— 2 f or1
1 VJ y
t
<— i
JL to
t v/ 186
XOU
fftr y < — I f ft -2J ue7iu\i,±j * juce\i,±) i
i «- / + 1 f*[face(i,l) - norm (z,l)]

count <— 1 bend(i,2) <- face(i,2) +
while count < features (j) f*[face(i,2) — norm (i,2)]
draw line from disp(i)
to i + 1)
co«?z£ «— count + 1
Figure 21.6 The heart of FACEBENDER.
in the feature, the program draws a line from point z in the array rfzsp to
point i + 1. My notation is merely shorthand. A real display command
would call for a line from the point whose coordinates are disp(i,l) and
disp(i,2) to the point with coordinates disp(i+l,l) and disp(i+l,2).
The heart of facebender is its exaggeration routine. Its structure is
even simpler than the display routine I have just outlined (Figure 21.6).
For each of the 186 facial points in the arrays face and norm, the loop
calculates a new array called bend. The new array encodes the carica-
ture-to-be. Each coordinate of the array bend is calculated by adding the
corresponding coordinate of the array face to a quantity that exaggerates
the differences between norm and face. The exaggeration factor / is
typed in by the user; / then multiplies the difference between the hori-
zontal coordinates of face and norm, and it also multiplies the difference
between the vertical coordinates.
The only things left to do are to organize the program and, option-
ally, to tune up the drawing routine. A simple, nonprocedural approach
to organization is to place both the display routine and the exaggeration

routine inside an interactive loop that asks the user: "Want to try an-
other?" The program must also prompt the user for the exaggeration
factor. Arrange the prompt so that a number of different exaggeration
factors can be tried without having to reenter the array face; their effect
on the caricature is then easy to compare.
The drawings can be somewhat enhanced if the dots are connected
with so-called spline curves instead of with straight lines. Splines avoid
zigs and zags and connect the dots smoothly; Brennan's program usually
draws spline curves to form the smooth contours of facial features. Nev-
ertheless, I was aware that splines might prove sticky to explain in a
21 Face Space 215
book that is devoted largely to easy programs. I asked Brennan for an al-
ternate method. Could straight lines be used instead? Much to her sur-
prise and mine, caricatures drawn with straight lines proved almost as
good as the onesdrawn with splines. Indeed, all her images appearing
here were drawn using straight lines. With only a small loss in aesthetic
value the programmer can avoid a most troublesome technique. One can
immediately set about digitizing a favorite photograph.
Brennan's caricature generator has been applied in several studies of
facial recognition. Faces from her program have been transmitted over
telephone lines as part of an experiment in teleconferencing at the Mas-
sachusetts Institute of Technology Media Laboratory. In 1985 she did
an experiment with Gillian Rhodes of the University of Otago in New
Zealand, who was then a graduate student working with Roger N. Shep-
ard of Stanford University. First she generated caricatures of faculty
members and students in the psychology department at Stanford. The
caricatures were then tested for recognizability against standard line
drawings.
Brennan has summarized the findings: "The
caricature generator
was particularly useful for this study because enabled us to generate
it
stimuli that varied in a continuous and controlled way; previous percep-

tual studies have had to compare caricatures with photos or other kinds
of not-so-similar images, and have therefore not been free of represen-
tational effects. Caricatures were not found to be particularly better as
recognizable representations (the 'best' representations were only mod-
estly exaggerated), but when the highly exaggerated caricatures were
recognized, they were recognized significantly faster — about twice as
fast, in fact,as the realistic line drawings of the same people."
Brennan suggests a number of other experiments with the caricature
generator. For example, it would be fascinating to recover the "norm"
assumed by human caricaturists. Handed a caricature of a given subject
by a given artist, she would try reversing the exaggeration to determine
the normal face, presumably lodged somewhere in the artist's uncon-
scious mind, from which the exaggeration was derived. Would the re-
constructed norm be much the same from one subject to the next?
Would different artists assume different norms?
Further Reading
Freeman, 1990.
Voltage Sculptures
Sometimes consider myself a fisherman. Computer

I
programs and ideas are the hooks, rods and reels.

Computer pictures are the trophies and delicious meals.
Clifford A. Pickover, Computers Pattern Choos and

Beauty, St. Martin's Press, New York, 1990.
(^^n the day of the earth year 2991, the Armstrong interstellar
first
spacecraft touched down on the fourth planet orbiting the star Tau Ceti.
The Armstrong's crew detected movement from the northeast and fo-
cused the ship's camera on a distant rocky cliff. There on a ledge was a
nest made of rock crystals and an egg that resembled a fried pastry, a
French cruller to be exact. The egg began to dissolve, and from it
emerged a snakelike creature composed of two intertwined rings. The
mission biologist quickly dubbed it a "gorgonoid." As the probe moved
closer to get a better look at the gorgonoid, the creature stiffened in
fright and bounced off the cliff into an acetylene river.
To be sure, the world of the gorgonoid is science fiction, but its image
resides in a computer at the IBM Thomas J. Watson Research Center.
Clifford A. Pickover, a graphics wizard at IBM, created the alien I call
the gorgonoid to demonstrate powerful, new tools for computer
graphics. He has developed the techniques as part of his mission to help
other scientists visualize the intricate shapes produced by physical phe-
nomena or derived from theories. Pickover, whose microscopic bio-
morphs appeared in my book, The Magic Machine, describes his cre-
ations as "graphics from an unseen world" (Figure 22.1 and color
insert).
Although the gorgonoid egg looks like an alien life-form, it is actually
a model based on physical principles discovered in terrestrial laborato-
ries. If one could peel away the "shell" of the gorgonoid egg, one would
216
Figure 22.1 Computer creatures from an unseen world.

find a frame composed of two "wires." One is bent into a circle; the
other winds around the circle in a spiral that rejoins itself. If the wires
were charged with a certain voltage, they would generate an electric
force that would be stronger at points close to the wire frame than at
points farther away. Pickover' s computer program finds all the points
representing a given strength of the force and then plots them to form
the shell of the gorgonoid egg. Pickover calls this imaging technique volt-
age sculpture.
Pickover engages in the art of voltage sculpture to depict a variety of
atomic structures from single molecules to the complex spiral of DNA.
Because the voltage sculpture displays the electric forces surrounding
the molecules, investigators may be able to deduce how some molecules
produced by living cells can fit certain receptor sites in other cells.
The young gorgonoid is not a voltage sculpture but what might be
called a worm necklace. Like the egg, the gorgonoid is based on two wire
loops, one winding around the other. To make the body of the gorgon-
oid, Pickover adorns the wires with spherical beads: large ones for the
circular wire, small ones for the spiral one. The beads are spaced evenly
along the wires, and consecutive beads overlap.
The mature gorgonoid is a worm necklace made of three wires: the
wraps around the second, which in turn curls around the third. The
first
mature gorgonoid also has an eye made from three nearly concentric
spheres that intersect to form an iris from one sphere and a pupil from
another.
A mature gorgonoid can spot a predator a mile away through an am-
monia haze —
an important survival strategy when it is being stalked by
a pacmantis. This cup-shaped creature spends half its time basking in the
rays of Tau Ceti. But when the pacmantis gets hungry, it rolls on the
ground, opening and closing its mouth like the computer sprite known as
Pac-man.
The anatomy of the pacmantis is no more complicated than the mor-
phology of the gorgonoid. To bring the pacmantis to life, Pickover
creates a computer pendulum. He simulates a ball that is tied to one end
of a rigid wire; the other end is connected to a pivot that allows the wire
and ball to swing freely in all directions.
Initially, the pendulum is pushed sideways with a certain velocity
and swings down under the influence of gravity. After it swings back and
forth, it arrives at a point that is a certain distance away from its starting
point. In the course of its subsequent swings, the ball covers most of the
available space within the sphere of possible positions.
As the pendulum swings, Pickover's computer periodically takes

snapshots of the ball. When many images of the ball are displayed simul-
taneously, they form a shell. By rotating the shell 90 degrees, one sees
the exoskeleton of a pacmantis in its proper orientation.
Although the pacmantis will occasionally munch on a gorgonoid, it
prefers to dine on the tubanides that live in ammonia oceans. These suc-
culent shellfish resemble certain ammonoids that flourished on the earth
during the Mesozoic era. The tubanide has an attractively striped shell,
which begins as a straightforward open spiral but subsequently curves
back on itself — like the product of a demented tuba maker. As a result
of its twisted shell, the tubanide tumbles as it swims, making it easy prey
for the pacmantis.
Tubanides were spawned from Pickover's collaboration with Aus-
Pickover and Illert studied a bizarre
tralian conchologist Chris Illert.
ammonoid called Nipponites mirabilis. Most ammonoids, like the mod-
ern Nautilus, have regular, logarithmically spiraling shells that allow the
animal to move smoothly through the water. During the early stages of
growth, the shell of N. mirabilis grows much like that of other ammon-
oids, but later it twists and turns in all directions. Illert hoped to investi-
gate such unusual growth patterns by searching for a mathematical de-
scription of the irregular spiral.
He came up with a formula that has a simple interpretation. The ori-
entation of the opening of the shell determines the direction of the
shell's growth. In ordinary spiral growth, the orientation of the opening
would remain fixed in relation to the adjacent rings of the shell. But the
growth of N. mirabilis can be nicely simulated if the shell opening is ro-
tated according to an exponential rule: as the shell grows, the opening
rotates more and more. This hypothesis produces a proper ammonoidal
appearance for young N. mirabilis and yields a twisting, irregular-looking
spiral for the older animal.
Pickover and demonstrated that the tubanide is a good model
Illert
for the adult N. mirabilis. To render tubanides in living 3-D, Pickover

used the worm-necklace technique. He colored the tubanide by alternat-
ing between crimson and white spheres.
The concepts behind worm necklaces and voltage sculptures are as
simple as advertised, but I am tempted to add the performer's warning:
"Do not try this at home!" After all, Pickover has access to computers
specifically designed for graphics. His computer system can automati-
cally shade and hide surfaces; it can show light from several sources re-
flecting off a surface; and it can produce, in an instant, a view of any
three-dimensional object from any angle. The skin of Pickover's crea-

turesis therefore only a few keystrokes away.
Although such facilities are not available in home computers, Pick-

over would not discourage amateur programmers from creating some
exquisite alien graphics called spherical Lissajous figures (Figure 22.2).
In 1857 the French mathematician Antoine Lissajous first described
these sinusoidal figures that today parade on the screens of oscillo-
scopes. A single Lissajous curve is traced out on the screen as a bright dot
moves up and down and side to side any number of times and eventually
returns to its starting point.
Spherical Lissajous figures have the same properties as their two-
dimensional relatives, except that they lie on the surface of a sphere. To
represent this three-dimensional curve in three dimensions, one needs
Figure 22.2 A spherical Lissajous figure.

three separate equations, each involving a single variable, t, which one

can think of as time.
x = Rs\n(At)cos(Bt)
y = /?sin(vAf)sin(Bf)
z = Rcos(At)
R, A, and B are constants. For each value of t, the three formulas collec-
tively specify a single point in three-dimensional space. As the value of t
is incremented (as time passes), the formula produces a succession of
points that generate the spherical Lissajous curve.
By setting values for R, A, and B, the rates at which the curves oscil-
late, one may generate fascinating figures. The curve will close back on
itself unless the ratio of A to B is an irrational number, not a likely event
in a computer.
Readers can devise a simple computer program to view a spherical
Lissajous curve on a two-dimensional screen. The program should first
ask for the values of R, A, and B. The program should then enter a loop
where the value of t is increased from, say, 1 to 1,000. For each value of
t, the program should calculate x and y according to the formulas. The x
coordinate, for example, will multiply R by the sine of A times t, then by

the cosine of B times t. Finally, the program should plot the point (x,y).
Some caveats accompany this algorithm. First, the numbers x and y
may have to be specially modified so that the point being plotted will ap-
pear on the screen. If necessary, add a suitable constant. Second, the
values of t may have to change more gradually to produce a solid-looking
curve rather than a string of widely spaced points.
Pickover is delighted at the potential of his tools and similar tech-
niques to help not only scientists but also artists. As examples of artists
who have already exploited such possibilities, he cites William Latham
of the IBM U.K. Scientific Center, John Lewis of the New York Institute
of Technology and Donna J. Cox of the National Center for Supercom-
puting Applications at the University of Illinois at Urbana-Champaign.
As electronic devices shrink and get faster still, even computers like
Pickover's sophisticated system will find their way into smaller and
more affordable boxes. The implications for science and art will be
equally great as more and more graphics emerge from unseen worlds.
Further Reading
Clifford A. Pickover. Computers, Pattern, Chaos, and Beauty. St. Martin's
Press, 1990.
23
j r y Han
A lthough computers are creeping tone by tone and pixel by pixel

endeavor still remain open to human beings.
into the arts, vast areas of
The designs I call latticeworks are not exhibited on computer screens
but rather on ordinary sheets of paper. The latticeworks are not pro-
duced in a few seconds by a computer program; they emerge slowly
from a ruler and compass guided by the human hand.
The designs begin when imagination meets a grid of circles. Anyone
can play and possibly produce an amazing interlocking network of lines
that confuses and delights the mind. The designs are similar to those
found on ancient tombs, mosques and palaces from Samarkand to Se-
ville.Rectilinear ornaments from the Islamic world of medieval times
hint at infinity within a finite space. Did the artisans use the methods I
shall describe? Experts are uncertain what method was employed, but
the mathematical reverberations of the patterns echo from geometry to
topology.
There are many ways to view a latticework, but perhaps a good
starting place is to look for symmetry. The finished version of the speci-
men in Figure 23.1, for example, has a high degree of symmetry: one can
rotate it by 120 degrees about certain points and end up with an identical
pattern. One can also reflect it across certain lines and get an almost
identical pattern. The chief elements of the pattern are gold bands that
weave like small highways of thought through a potentially infinite land-
scape. Throughout the design circles are spaced according to the under-
lying symmetry of the pattern. The bands travel from one circle to an-
other, converging in 12's and then abruptly bending away. The angle of
reflection equals the angle of incidence. The bands themselves have
shapes, in this case either hexagonal or zigzag. Everywhere one looks,
overpasses alternate with underpasses. By what means was this ar-
ranged?
222
Figure 23.1 Various stages in the construction of latticework.
Latticeworks belong to a wide class of infinite patterns that have at

least two independent symmetries of translation: if one translates, or
slides,the pattern over a copy of itself in either of the two directions,
eventually the pattern and its copy will line up with each other. No mat-
ter how cleverly one lays down an infinite pattern, as long as it has two
independent translations it will inevitably belong to one of the 17 possi-
ble crystallographic groups first classified by the Russian scientist Evgraf
S. Fedorov. Considering the enormous wealth of latticeworks inherited
from the past, it is not surprising that examples of all 17 groups can be
found, whether peeping shyly through a palace window grille or pro-
claiming the sultan's grandeur in the audience hall.
Translations are just one of four possible symmetry operations that
can be done on a plane imprinted with a pattern. A reflection flips the
plane like some vast door about a line that lies within the plane itself.
This is essentially the looking-glass operation, in which letters get re-
versed and faces look slightly different. The third symmetry operation is
the rotation, and in it the plane is rotated through a specific angle about a
single point. The fourth and final symmetry is called a glide reflection. It
consists of a translation followed by a reflection about a line that is paral-
lel to the translation.
Symmetry operations (of whatever type) can be combined simply

by performing them in some sequence. An algebraic structure called a
group emerges from all of this. In a group the product of any two opera-
tions is an operation, and every operation has an inverse operation that
in effect undoes it. In addition a group has a so-called identity element
(the do-nothing operation). A group must also satisfy the associative law:
when three successive operations are carried out, it does not matter
whether the third operation follows the first two or whether the last two
follow the first one. Put this way the law sounds mildly idiotic, but only
because the associative law happens to be transparently true in the case
of symmetry operations on the plane. The symmetry group of the lat-
ticework in Figure 23.1 is called p6 in the international crystallographic
language. It is characterized by having rotations of 60 degrees about one
set of centers and rotations of 120 degrees about another set. Readers
will readily find the centers in the figure.
produced the design a few years ago by the method I shall outline
I
here. Later I discovered the same latticework in a book about Islamic art
of the medieval period. My deflation at not being first was more than
compensated for by the discovery itself; the method seemed confirmed.
Since then it has been my lot to "rediscover" other designs.
The method requires the would-be artisan to set up a grid of points.
The grid is limited to one of four types: triangular, square, rectangular or
hexagonal. Such grids are easy to lay out by ruler or compass: draw a
base line, then use the compass to mark off evenly spaced points. Square
and rectangular grids employ a right-angle construction to add new
points above and below the base line. Triangular and hexagonal grids re-
quire equilateral triangles.
The design under discussion began life as a triangular grid. A circle
was then drawn at each of the grid points. Here intuition made its first
entrance, because the size of the circle happens to be critical. In a mo-
ment I shall explain the role intuition may play in selecting the size.
Once the circles are all drawn, the amateur artisan selects points
evenly spaced around the circle. These points will anchor the lineal ele-
ments of the design. The position and number of points must reflect the
symmetry of the grid itself. In other words, the points should preserve
symmetry with respect to reflections and rotations. Any of the intended
symmetries should carry points on one circle onto points on another cir-
cle or on the same circle. In the example under construction the number
of points on each circle must be a multiple of 3. 1 chose 12 to give body to
the lattice. Since the underlying pattern was to have reflectional sym-
metry, only two positions for the points on the circles were possible. I
chose the position in which six of the points were closest to the sur-
rounding circles. As a general rule, whenever there is a choice, the best
decisions are those that harmonize with a symmetry already present.
In the next stage of construction one joins the points on each circle to
points on other circles. Here intuition makes a second entrance. The
possibilities appear to be so numerous that only intuition would seem to
serve. In fact, the combinatorial possibilities are greatly limited once
more by conditions of symmetry: if I join a certain point on circle A to
another on circle B, the symmetries of the pattern carry that connection
onto other points on circle A. Before one has drawn more than two lines
it may be necessary to erase the experiment and try another connection.
A kind of feedback loop binds this design phase with the earlier
choice of circle size. Once a seemingly satisfactory and consistent
scheme of interconnection has emerged, the results may look unauthen-
tic,not to mention ugly. The lines do not harmonize with the symmetry
of the pattern. In such a case it is usually obvious whether shrinking or
expanding the circles will produce connections that parallel the major
symmetries of the pattern. Here intuition may provide the leap of in-
sight, a kind of artistic "Aha!" experience. In the mind's eye one sud-
denly sees the lines generated by the new circles.
At this point in apprenticeship a certain excitement causes the com-
pass and ruler to quiver slightly. An amazement that is perhaps half ar-
tistic and half mathematical grips the holder of these instruments.
Should one take credit for an intuition that was merely acceding to ge-
ometry?
In any event, it is now time to pave the highways of thought by giving
them some width. No sophisticated system of roads should suffer traffic
lights, and so how can the angular freeways be interlaced? In true weav-
ing, after all, overs and unders alternate. Can the artisan be forced into
some kind of logical cul-de-sac where a road faces two consecutive un-
derpasses? Topology saves the day.
The following experiment in scribbling shows how. Draw a large
rectangle on a sheet of paper. Then scribble inside the rectangle, abiding
by only two rules:
1. If a line begins or ends, it must do so on the boundary of the rec-

tangle.
2. No more than two lines (or parts of the same) may cross at any
point.
One can, for example, scribble something like the abstract-expressionist

composition shown in Figure 23.2. Lines can be curved or bent, repeat-
edly crossing over one another.
To convert the scribble into the cleverest knot imaginable, it is now
only necessary to follow the "over-under" rule: starting anywhere one
likes, simply follow one of the lines, repeating "Over, under,
over . . . "Of must make an overpass for the line being
course, one
followed when "Over" is said. By the same token, "Under" calls for an
underpass. Eventually one either returns to where one started or the
boundary of the rectangle is reached. En route the line may have crossed
itself. Amazingly, whenever one arrives at a previously processed cross-
ing, it already has the required structure. In other words, one never finds
that an overpass is called for at a crossing that has already been desig-
nated an underpass. Eventually all bridges have been built and the scrib-
ble takes on an appearance that is almost intelligent.
The over-under rule works because in a sense it must. The simplest
demonstration at a public level invokes a pleasant thought excursion.
The scribble divides the rectangle into many small regions, or pieces. It
turns out that the regions in the rectangle can each be given one of two
colors, so thatno two regions sharing a common boundary are assigned
the same color. (A convincing elementary proof of this property would
take at most a few paragraphs, but I hasten to the punch line.) Suppose
the regions are painted in this manner with, say, red and blue paint.
Driving along one of the roads toward a crossing, we would notice one of
Figure 23.2 The "over-under" rule in progress.

two things: either the region on our right would be red and the one on our
left would be blue or vice versa.
The recipe for crossings lurks in this simple observation. Traveling

along a given road, the mental construction crew will know whether to
build an overpass or underpass when they arrive at a crossing. If the re-
gion to the right is red, build an overpass; if it is blue, build an underpass.
The result is exactly thesame as (or possibly the "negative image" of)
the result if the crew had followed the over-under rule stated above.
With a pencil and a ruler the designing reader can easily thicken each
line of the latticework by drawing a parallel line on each side of it. At
bends there is some fiddling with the meetings of these lines, but the
project proceeds more or less mechanically. The thickening procedure
may ignore the crossings until they are complete. Equipped with an er-
aser, the artisan now attacks the crossings, invoking the over-under rule
as he or she goes. To create an overpass, one must erase the two seg-
ments of road edge that cross the road one travels. As soon as this is done
the overpass springs into existence. The road that was erased seems to
pass under the road one is on. Underpasses are created by the opposite
procedure. It is interesting that the interlacing procedure destroys all re-
flectional symmetries of the pattern; the reflected pattern may look the
same but underpasses and overpasses have been swapped.
The final stages in the creation of a design involve inking and color-
ing. A good pen will follow the ruler and produce an even line with occa-
sional interruptions at bends and crossings. A design of moderate com-
plexity may take an hour or more to ink, but what is an hour in the
timeless world of the artisan? There is time to think of other things dur-
ing this phase. It is perhaps a loose form of meditation.
When the design has been inked and all the pencil marks have been
erased, it can be colored directly. Because coloring does not always turn
out as well as one might like, it may be preferable to copy the inked origi-
nal and color the copy. In this way the original can produce many off-
spring, each more beautiful than the one before. I suggest using tempera
paint. It goes on evenly, produces minimal wrinkling of paper and is
available in virtually all colors, including gold and silver. Moreover,
water-based paints such as tempera seem to be repelled by xerographic
inks. This is fortunate because unless one has a very steady hand, over-
painting of ink lines seems inevitable. One wants all colored areas to be
finely edged with black in order to enhance contrast.
The colors one selects are of course a matter of personal choice. Au-
thentic latticeworks often employ dark and muted primaries for the re-
gions between the bands. Such treatment results in a retreating back-

ground that points up the latticework itself all the more prominently.
I have included another design that uses such a color scheme (Figure
23.3) .
The second design was produced by methods similar to the first. Ad-
vanced latticeworks have not only what might be called primary circles
of inflection but also secondary ones. Each five-pointed star in the sec-
ond design arises from such a secondary circle. I have included two ad-
ditional charming examples of latticework in merely skeletal form. In

one of them the primary and secondary circles are situated on different
centers of rotational symmetry; in the other they are concentric (Figure
23.4) Both latticeworks can be found in traditional settings.
.
Latticeworks are closely related to two-dimensional tessellations:

patterns produced by laying down multiple copies of a single shape
Figure 23.3 Islamic latticework.

Figure 23.4 Two latticeworks in skeletal form.
without overlap. Each latticework can be produced in the form of a sin-

gle, finite shape, more specifically a tile. On the tile appears a small part
of the overall design. If one has available a great many copies of the tile,
one can reproduce the latticework to any extent by laying down enough
tiles inthe proper way. An examination of any of the latticeworks shown
here will reveal a small area in each example that could act as a tile. Here,
in any event, lies the prospect for a larger view of any such pattern. Make
a tile of sorts on paper and make many copies of it. When the tiles are
colored, they lend themselves to a host of decorating ideas.
But what of the original latticework designs? How were they pro-
duced and what, for that matter, did they mean? Lisa Golombek has
studied the designs for many years, both on site and in her office at the
department of Middle East and Islamic studies at the University of Tor-
onto. She thinks some kind of geometric underpinning in the form of
grids and circles is the likely method. Undoubtedly individual craftsmen
employed variations on the basic technique. As for meaning, Golombek
takes issue with the usual explanation that geometric forms were used
because of the Islamic prohibition on figurate (human and animal)
forms. In her opinion the latticeworks represent a cosmic order that is a
hallmark of Islam. For one thing, latticeworks appear in the remains of
private residences along with figurative works. In view of the prohibi-
tion ignored, the cosmic order obviously appealed to the residents.
It was not altogether correct of me to say that computers have no
role to play in the production of latticeworks. Although it may be diffi-

cult to program a computer to make the kind of intuitive choices that

lead to beautiful patterns, it would be less difficult to write what might be
called a computer-aided design program. Such a program would lay out a
grid of points according to human choice. The human user might also se-
number of points and so on. When the user
lect the size of the circles, the
of such a program indicates what connections to try, the entire screen
would fill with the implication of his or her choice. In short, much of the
tedium could be removed. The latticework would emerge, one hopes,
from the printer.
Further Reading
Issam El-Said and Ay§e Parman. Geometric Concepts in Islamic Art. World of
Islam Festival Publishing Company, London, 1976.
Illustration
Text
Figure 1.4 Courtesy of Brian Silverman, Logo Systems, Montreal

Figure 4.1 Photo © 1989 by Bruce Frisch, Brooklyn, New York
Figures 6.2 and 6.4 Courtesy of Greg Turk
Figure 6.5 Courtesy of Odd Arild Olsen, Oslo, Norway
Figures 10.2 - 10.4 Courtesy of Mario Markus, Max Planck Institute for Nu-
trition, Hamburg, Germany
Figures 11.2 and 11.4 Courtesy of Michael Barnsley, Iterated Systems, Inc.,
Norcross, Georgia
Figure 19.5 Courtesy of Steffen Schindler and Ralf D. Tscheuschner, Insti-

tute for Theoretical Physics, University of Hamburg, Germany
Figure 22.1 Courtesy of Clifford A. Pickover, T. J. Watson IBM Research
Center, Yorktown Heights, New York
Color insert
Color Plate 1 Copyright ©

1993 by O. Christian Irgens, St. Albans, New
York; courtesy of The Computer Museum, Boston, Massachusetts
Color Plate 2 Top copyright © 1989, bottom copyright © 1991 by Bruce

Frisch, Brooklyn, New York
Color Plate 3 Courtesy of Greg Turk

Color Plate 4 Courtesy of Clifford A. Pickover, T. J. Watson IBM Research
Center, Yorktown Heights, New York
231
In
Acceleration table, in TREK program, 87 Lorenz, 93-99
Addition, neural network, 56 one-point, 101, 194
Affine transformation two-point, 101, 194
in IFS, 109-118 Augmented finite-state machine
for natural system, 120-128 (AFSM), 38
AFSM (augmented finite-state machine),
38 Back-propagation algorithm, 49-50,
Alexandrov, A. D., 160 53-55
Algorithm (See also Program) Barnsley, Michael F., 109, 112,
back-propagation, 49-50, 53-55 116-118, 119-120
BEAT, 186-187, 190 BASIC language, 74-78, 83-84, 197
BIRDIE, 72, 79-80 BEAT (algorithm), 186-187, 190
British Museum, 29 Bifurcation diagram, 193-194
CANON, 182-186 Billiard-ball computer, 30-31
CLOUD, 123-125 Binary principle, 14, 16
EAGLE, 72, 79-80 BIRDIE (algorithm), 72, 79-80
FACEBENDER, 210-214 Bolker, Ethan D., 161
GRIN, 154-155 Brain
hill-climbing, 55 human, 26-36, 67, 174
HOLE IN ONE, 71-78 infinite, 33-36
IFS, 123 of a tur-mite, 62-65
Lyapunov, 107-108 Braitenberg, Valentino, 43
Mark V. Shaney, 203 Brennan, Susan E., 205-215
POLARNET, 46, 48-56 British Museum Algorithm, 29
SE Q, 166-170 Brooks, Rodney, 37
SOLFEGGIO, 180-183
TREK, 83-92 CANON (algorithm), 182-186
TURMITE, 64-65 Caricature, by computer, 205-215
WAVE, 126-127 Cartesian coordinates, 49, 52-53
WEATHER WHEEL, 96-99 CAT (computerized axial tomography)
Alternating tripod gait, 41-43 scan, 148-157
Ammonoid, 219 Catastrophic pattern, 150-153
Analog multiplication, 25 Categorical pattern, 150-154
AND gate, 14, 18-23 Cauchy, Augustin-Louis, 162
in billiard-ball computer, 30-31 Cauchy's theorem, 162
Angle, Colin, 38 Chain
Apraphul, island of, 16 in catastrophic pattern, 152-153
Artificial intelligence, 26-36 Markov, in language program, 204
in I.Q. testing, 174 Chaos
Attila, a robot, 43-44 in logistic system, 101-108, 192
Attractor in music, 191
chaotic, 101-102 in weather system, 93-99
234 Index
Chaotic attractor, 101-102 Face, average, in caricature, 205-215

Checker-jumping demonstration, FACEBENDER (algorithm), 210-214
133-134 Face space, in caricature, 209-210
Chinese room (in The Emperor's New Fedorov, Evgraf S., 223
Mind), 26-28 Feenstra, Marcel, 166, 170-171
Circle-grid method, 224-226 Fermi function, 48
Circular queue, in TREK program, 90 Fern fractal, 114-116
Cloud, IFS for, 120-122 Fiber, of neurodes, 61
CLOUD (algorithm),123-125 First-species imitation, 182
Collage, in IFS, 120-121, 126-127 Flexible framework, 158-164
Collage theorem, 116 Flexing nonconvex surface, 163
Computer Flip-flop, 16, 22, 22-23
billiard-ball, 30-31 Flynn, Anita, 43
caricature by, 205-215 Forced logistic system, 103-107
infinite, 35 Formula
rope-and-pulley, 16-25 golygon, 144
Tinkertoy, 7-15 logistic, 100-108
tone generator in, 179 Lyapunov, 102-103, 107
Computerized axial tomography (CAT), Fractal
148-157 created by IFS, 119-128
Connelly, Robert, 162-164 fern, 114-116
Connelly-Steffen surface, 163 in IFS, 109-118
Convection cell, in weather system, 94 in logistic system, 102
Convergence, of a neural network, 56 Fractal tennis, 112-115
Conway, John, 134 Framework
Core piece (of tinkertoy computer), 7-15 cube, 159-160
Cormack, Allan M., 148 flexible, 158-164
Cox, Donna J., 221 grid, 161
Crapo, Henry, 161 polyhedral, 160-164
Crapo-Bolker theorem, 161 rigid, 158-164
223-224
Crystallographic group, two-dimensional, 160-161
Cube framework, 159-160 Friction, in golf program, 79-80
Frye, Robert C, 46
Dandelin, Germinal, 131-133 Fuller, R. Buckminster, 160
Davidson, Arthur, 191-192, 195-197 Function
Dewdney, Jonathan, 158 Fermi, 48
Diagram, bifurcation, 193-194 hyperbolic tangent, 48
Dynamics, in weather systems, 93-99 sigmoidal, of a neuron, 47-54
EAGLE (algorithm), 72, 79-80 Gait, alternating tripod, 41-43

Electronic Arts, 80 Game board, 10-14
Ellis, Bruce, 202 Game tree, 10-11
Equation, logistic, in 191-198
music, Gardner, Martin, 140, 142-145
Error, in a neural network, 49-55 Gate
Evans, Thomas G., 171 AND, 14, 18-23, 30-31
program of, 172-174 OR, 8, 18-22
Eysenck, Hans J., 166 Gee, Erlyne, 8
Index 235
Generator Infinite brain, 33-36

linear congruential, in music, 180-181 Infinite computer, 35
tone, in a computer, 179 Infinitesimal flex, 159-160
Genghis, a robot, 38-42 Infinite structure, of universe, 26, 33-36
Glass, Philip,191 Infrared sensor, in a robot, 42
Gleick, James, 93 Input layer, of a learning network, 46-55
Gluck, Herman R., 162 Insectoids, 37-45
Godel, Kurt, 29 Intelligence
Godel's theorem, 29-30 26-36, 174
artificial,
Golombek, Lisa, 229 human, 26-36

Golygon, 139-145 Intelligence quotient (I.Q.), 165-176
formula for, 144 Internal table, of a Turing machine,
largest, 144 59-67
number of sides in, 140-144 Inverter, 8, 17-18
polyominid, 146 Islamic medieval design, 222-230
prime-sided, 146 Isogon, serial, 145
theorem, 142-144 Iterated-function system (IFS),
tiling by, 145-146 109-118, 119-128
Go-moku narabe, 13
Gorgonoid, 216-218 Knuth, Donald E., 144-145
Gould, Stephen Jay, 165, 174-175
Grid framework, 161 Language
GRIN (algorithm), 154-155 BASIC, 74-78, 83-84, 197
Groff, Jonathan N., 92 Turbo Pascal, 191, 196, 197
Guy, Richard K., 144-145 Latham, William, 221
Latticework, 222-230
Hardebeck, Edward, 8, 15 Learning curve, in a neural network,
Harmonic intervals, table of, 183 51-52
Hill-climbing algorithm, 55 Lewis, John, 221
Hillis, Daniel, 8 Linear congruential generator, in music,
HIQ program, 171 180-181
HOLE IN ONE (algorithm), 71-78 Linear-congruential music, 181
Honsberger, Ross, 131-138 Lissajous, Antoine, 220
Hounsfield, Sir Godfrey N., 148 Lissajous figure, 220-221
Human brain, 26-36, 67, 174 Loebner, Hugh G., 199
Human intelligence, 26-36 Loebner Prize Competition, 199-202,
Hyperbolic tangent function, 48 204
Logic unit, of rope-and-pulley computer,
I.Q. (intelligence quotient), 165-176 24
IFS (iterated-function system), 109-118 Logistic equation, in music, 191-198
IFS algorithm, 123 Logistic formula, 100-108
IFS coefficients for Sierpinski triangle, Logolygon, 146-147
117 Lorenz, Edward, 93-94
Illert, Chris, 219 Lorenz attractor, 93-99
Image storing, 117-118 Lorenzian water wheel, 93 - 99
Imitation, first-species, 182 Lyapunov, Aleksandr M., 100
In-betweening method, 207-208 Lyapunov algorithm, 107-108
236 Index
Lyapunov exponent, 100-108 Neurode, of a tur-mite, 61-64

Lyapunov formula, 102-103, 107 Nicklaus, Jack, 71
Numerical-sequence completion test,
Machine 166-170
augmented finite-state (AFSM), 38
millipede, 41 Olsen, Odd Arild, 65
Turing, 57-67 One-point attractor, 101, 194
Mandelbrot, Benoit B., 28, 119 Orbit, of spaceship, 81-92
Mandelbrot set, 28-29, 35 OR gate, 8, 18-22
Markov chain, in language program, 204 Over-under method, 226
Markus, Mario, 100-103, 103, 107
Mark V. Shaney (algorithm), 203 Pattern
Mark V. Shaney program, 199, 202-204 catastrophic, 150-153
Memory, in rope-and-pulley computer, categorical, 150-154
22-24 reconstruction of scanned, 152-157
Memory spindle (of Tinkertoy of a tur-mite, 61-67
computer), 7-15 PC Therapist program, 202
Method Pendulum method, 218
circle-grid, 224-226 Penrose, Roger, 26-32
in-betweening, 207-208 Period doubling, in logistic system, 195
over-under, 226 Periodic forcing, in logistic system, 103
pendulum, 218 Pickover, Clifford A., 216-221
pyramid, in numerical-sequence Pigeonhole principle, 135-136
completion test, 167 Pixel, in golf display, 74-75
worm-necklace, 218-219 Platonic reality, 28
Microminiature golf, 71-80 Polar coordinates, conversionof, 48-56
MIDI (Musical Instrument Digital POLARNET 48-56

(algorithm), 46,
Interface), 182 Polyhedral framework, 160-164
Millipede machine, 41 Polynomial, quadratic, 167-168
Minsky, Margaret, 8 Pressing, Jeff, 198
Mitchell, Don P., 202 Principle
Moon, Francis, 197 binary, 14, 16
Multiplexer, 20-22 pigeonhole, 135-136
Multiplication, neural network, 56 Problem
Music urn, 137-138
chaos in, 191 word, 34-36
linear-congruential, 181 Program (See also Algorithm)
linear congruential generator in, of Evans, Thomas G., 172
180-181 HI Q, 171
logistic equation in, 191-198 HOLE IN ONE, 71-78
Musical Instrument Digital Interface language, Markov chain in, 204
(MIDI), 182 Mark V. Shaney, 199, 202-204
PC Therapist, 202
Network, Rietman-Frye, 46-56 Racter, 202
Network of processors, in robotic TREK, 83-92
37-45
design, for a tur-mite, 64-65
Neural network, 46-56 weather simulation, 93-94
Index 237
Prompt, in golf program, 76 Surface, 163

Proof, of golygon theorem, 143-144 Symmetry operations, in plane, 223-224
Pyramid method, in numerical-sequence Synapse, 47-56
completion test, 167 System
chaotic, self-similarity in, 107
167-168
Quadratic polynomial, forced logistic, 103-107
Quantum mechanics, 26-36 iterated-function (IFS), 109-118,
119- 128
Racter program, 202 logistic, chaos in, 101-108, 192,
Read head (of Tinkertoy computer), 7-15 195
Reagan, Ronald, 205 natural, affine transformation for,
Register, in rope-and-pulley computer, 120- 128
23-24
Rhodes, Gillian, 215 Table
Rietman, Edward A., 46 acceleration, in TREK program, 87
Rietman-Frye network, 46-56 of harmonic intervals, 183
Rigid framework, 158-164 internal, of a Turing machine,59-67
Rigidity theory, 158-164 Tape, of a Turing machine, 58-59
Robot, 37-44 Tessellation, two-dimensional, 228-229
Rope-and-pulley computer, 16-25 Test
I.Q. (intelligence quotient), 165-176
Sallows, Lee, 139-140, 144-145 numerical-sequence completion,
Scale, twelve-tone semitone, 197 166-170
Schrodinger, Erwin, 26 Turing, 27,199-202
Schrodinger's cat, 31-32 172-173
visual analogy,
Searle, John R., 27 The Emperor's New Mind, 26-32
Self-similarity Theorem
in chaotic system, 107 Cauchy's, 162
in natural structures, 119 collage, 116
Sensitivity to initial conditions, 94 Crapo-Bolker, 161
Sensor, infrared, in a robot, 42 Godel's, 29-30
SE Q (algorithm), 166-170 golygon, 143-144
Serial isogon, 145 transformation, for catastrophic
Shepard, Roger N., 215 pattern, 152
Sierpinski triangle, 110, 116-117 Thue, Axel, 34
Sigmoidal function, of a neuron, 47-54 Tic-tac-toe, 8-15
Silverman, Barry, 8 Tiling, by golygon, 145-146
Silverman, Brian, 8 Time constant, in weather simulation,
Sloan, Alan D., 118, 119 98
Solar energy, in game of Star Trek, 83, 89 Tinkertoy computer, 7-15
SOLFEGGIO 180-183
(algorithm), Tone generator, in a computer, 179
Spheres of Dandelin, 131-133 Topology
Squirt, a robot, 38, 43 employed in rigidity theory, 162
Star Trek, game 81-92
of, in latticework, 226-227
State vector collapse, 31-32 of trousers, 131
Steffen, Klaus, 163 Toroidal screen, in Star Trek game, 83
Subsumption architecture, 37-45 Training a neural network, 48-56
238 Index
Transformation, affine Twelve-tone semitone scale, 197

in IFS, 109-118 Two-dimensional framework, 160-161
for natural system, 120-128 Two-dimensional tessellation, 228-229
Transformation theorem, for Two-point attractor, 101, 194
catastrophic pattern, 152
TREK 83-92
(algorithm),
Urn problem, 137-138
TREK program, 83-92
Utas, Gregory, 184
acceleration table in, 87
circular queue in, 90
cursor-control keys in, 84-85 Vehicle, of Valentino Braitenberg, 43
Tscheuschner, Ralf D., 198 Visual analogy test, 172-173
Voltage sculpture, 216, 218
Tubanide, 217, 219
Turbo Pascal language, 191, 196, 197
Turing, Alan M., 57 Walking, robotic, 39-45
Turing machine, 57-67 Washer puzzle, 135-136
Turing test, 27, 199-202 Wave, IFS for, 125-127
Turk, Greg, 57 WAVE (algorithm), 126-127
Tur-mite, 57-67 Weather in a jar, 93-99
behavior of, 63 Weather simulation program, 93-94
brain of, 62-65 WEATHER WHEEL (algorithm), 96-99
neurode of, 61-64 Weintraub, Joseph, 202
pairs of, 65-66 Winston, Patrick Henry, 172
pattern of, 61-67 Word problem, 34-36
TURMITE (algorithm), 64-65 Worm-necklace method, 218-219
Behind the humorous, light-hearted
style of A. K. Dewdney lies a serious
computer scientist with much to say.
Whether he is writing on the ins and
outs of programming, the elegance of
mathematics, or the philosophical rami-
fications of the computer age, Dewdney
always delights and enlightens. With
THE TINKERTOY COMPUTER,
longtime fans and newcomers will have
a lot to do and to think about.
About the Author

A. K. (Kee) Dewdney, a computer
scientist at the University of Western
Ontario, began writing the "Computer

Recreations" column for Scientific
American in 1984, after Martin
Gardner's retirement. He now writes
a similar column for Algorithm.
THE
Dewdney's other books include
(NEW) TURING OMNIBUS (1993),
and two other collections of columns,
THE MAGIC MACHINE (1990) and
THE ARMCHAIR UNIVERSE
(1988).
f
the Tinkertoy Computer

and other m achinations
A. K. DEWDNEY
Can a contraption made of Tinkertoys win at tic-tac-toe?

Can a computer create music?
Design golf courses on the screen?
Simulate the human brain?
Can mathematics really explain anything and everything?
For further elaboration on these and other provocative ques-

tions, read this latest collection of A. K. Dewdney's columns,
drawn from the pages of Scientific American and Algorithm^
For novice hackers and longtime afficionados alike, it is a
stimulating, fun-filled journey to the frontiers of computer
science.
Other books by A. K. Dewdney from

W. H. Freeman and Company:
THE (NEW) TURING OMNIBUS

66 Excursions in Computer Science (1993)
THE MAGIC MACHINE

A Handbook of Computer Sorcery (1990)
THE ARMCHAIR UNIVERSE

An Exploration of Computer Worlds (1988)
8
W. H. Freeman and Company
41 Madison Avenue
New York, NY 10010
20 Beaumont Street
Oxford OX1 2NQ, England

The Tinkertoy Computer and Other Machinations

Uploaded by

Copyright:

Available Formats

The Tinkertoy Computer and Other Machinations

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Tinkertoy Computer and Other Machinations

Uploaded by

Copyright:

Available Formats

Puzzles and Games/Computers $22.

the Tinkertoy Computer

Matter Computes. Matter Misbehaves.

Drawn from the pages of Scientific

ence: chaos and fractals, artificial intel-

digital computer (a rope-and-pulley cal-

culator dated circa 850 A.D.), as well as

But you won't just have to read about

Copyright © 1993 by W. H. Freeman and Company

No part of this book may be reproduced by any mechanical,

Printed in the United States of America

Theme One: Matter Computes 5

2 The Rope-and-Pulley Wonder 16

3 The Infinite Brain 26

4 Invasion of the Insectoids 37

6 Dance of the Tur-mites 57

Theme Two: Matter Misbehaves 69

8 Star Trek Dynamics 81

10 A Portrait of Chaos 100

1 1 Designer Fractals 109

12 The Fractal Workshop 1 19

Theme Three: Mathematics Matters 129

14 Golygon City 139

15 Scanning the Cat 148

16 Rigid Thinking 158

17 Automated Math 165

Theme Four: Computers Create 177

1 8 The Computer Composer 1 79

19 Chaos in A Major 191

20 Mark V. Shaney 199

21 Face Space 205

22 Voltage Sculptures 216

23 Latticeworks by Hand 222

Illustration Credits 231

The term "machinations" in this book's title

refers first to the marvellous machines called

During my career as a computer scientist, I

computer science a bit more fun and interest-

of my lecturing style. Moreover, I saw it as a

serious contextual disadvantage. If my regular

philosophical themes that arise from computing

MATTER COMPUTES: Almost any form of matter

MATTER MISBEHAVES: We can simulate a great

MATHEMATICS MATTERS: Explicitly or implicitly,

COMPUTERS CREATE: From a sociological point

To say that matter computes means, in part,

choice? The next chapter in this suite takes

(computers) yet, but we had toy construction sets.

those spools and sticks are enough to make anything.

Marvin Minsky, in preface to LogoWorks

ing a game Each spindle is a sequence of smooth spools

symmetry argument: the excluded boards are merely rotations or re-

1 The Tinkertoy Computer 1

Then a memory spindle was divided conceptually into nine consecutive

Tinkertoy computer require any other memory? Could spool-and-stick

The Tinkertoy computer can teach us resides in a

Actually, Tinkertoys are well suited to digital computing. For exam-

corresponding spools on a memory spindle. Such an operation is called

Figure 1.4 Edward Hardebeck helps to assemble the Tinkertoy com-

first appeared in Scientific American in April, 1988.