ROBOT HIERARCHIES     
				 by
			Frederick W. Chesson   

	 (First appeared in INTERFACE AGE for April, 1978) 

     The definition of what constitutes a robot has been dimmed by anthro-
pological bias, or the impediment of seeing the robotic scene through 
Human-Colored glasses. That shambling and amiable tin-man traveler on the 
Yellow Brick Road to Oz may be far more acceptable to the viewer than the 
most complex but immobile artificial intellect. Consequently, there's the 
rub...(or byte): Just when DOES a computer rise to the exhalted status of 
Robot Brain?
     In order to examine these perplexing and contemporary questions, 
it will be useful to consider and, if possible, categorize the various 
hierarchies of the robot or cybernetic world. To start with, there is 
the pre-robotic environment of the servo-mechanism, with its self-
balancing feedback loop(s). Even at this primitive level, the feedback 
loop is itself subject ot varying levels of complexity, interaction, 
and adaptability. Adaptation is clearly the most essential feature of 
any organism; animal, vegetable or servo.
     Feedback is that quality which immediately separates a large category 
of objects and systems. The most elaborate-appearing mechanism is nothing 
in comparison, when put beside a tiny device endowed with the quality of 
governing its activities in proportion to the varying intensity of its 
input stimuli, and output responses.
     The word Cybernetic, the steersman of classical Greek, comes from the 
concept of feedback. Here is the helmsman, who senses the movement of wind 
and wave, and adjusts the rudder accordingly to keep the ship upon course. 
One of the earliest feedback devices was the Baille-Ble, or mill-hopper, 
described as far back as 1588. Its use in water or wind-powered mills was 
to distribute grain to the millstones, according to the rate of speed of 
the mill's drive shaft. Such other feedback factors as grain flow, grain 
hardness, millstone tension, and drive shaft force, all interacted to 
determine the amount of grain delivered to the stones. This was, of course, 
a very primitive level of feedback, the grain-hopper receiving four jolts 
for every revolution of the shaft. Man, in the form of the miller, was 
still required to optimize conditions and design goals by regulating the 
flow of water to the drive wheel or the wind pressure on the sails, and to 
adjust the proximity of the millstones accordingly.

     Not until the late 18th Century did a more familiar feedback device 
appear. This was the flyball governor, another contribution of James Watt 
to the perfection steam power. The governor consisted of a pair of iron
balls at the ends of hinged arms linked to the engine's drive shaft. As 
the shaft speed increased, the balls rotated ever faster, responding to 
centrifugal force and causing the linkages to move. This movement was
coupled to the steam supply throttle, causing it to cut off the steam 
with increased speed and to open the valve as the shaft velocity slowed. 
The flyball governor continued in use well into the electrical age and was
even employed with early phonograph motors. 
     The 19th Century also saw the growth of hydraulics as a science, and 
feedback appeared here in the form of Leon Farcot's Servo-Motor of 1868. 
Feedback was also discernable in nature, with many writers commenting 
upon it in terms of the dependency and population oscillations of predator 
and prey animals, to the conservation of energy in the Solar Phoenix Cycle.
     World War II saw research and application of feedback and servo-
mechanisms advance under the pressure of wartime necessities. Bombsights, 
anti-aircraft gun directors, and radar all were fused into increasingly 
complex integrated systems.

     These early analog and digital devices converged irrestibly towards 
today's state of the art, where the external analog world is sensed, 
converted to, and processed digitally, then reacted upon by analog 
extensions. Any device which aspires to the name robot is therefore
bound to the laws of feedback and stability, if only to maintain an 
external upright position, or an internal state of data-processing 
stability. Negative feedback converges and conserves. Positive feedback 
diverges into disorder. The first mechanic who managed to assemble a 
fly-ball governor in reverse discovered this, as the steam engine's 
speed increased to the literal breaking point. His descendent, plying his 
op-amp breadboard, is no less dismayed to discover a hidden glitch of 
oscillations emerging from his clean-looking Bode Diagram.
     Two post-war inventions which captured much popular interest in 
servo-systems were the Homeostats of W. Ross Ashby, and the "Tortoises" 
of W. Grey Walter. In the literary area there was Dr. Norbert Wiener's 
monumental "Cybernetics," accompanied by, in the science-fiction realm, 
such classics as Asimov's "I, Robot."  The Homeostat was, in essence, 
an interconnected complex of (at least four) servo units, so organized 
that a disturbance to the input of any unit would reflect throughout the 
others, resulting in a mutual attempt to restore a state of equilibrium, 
or homeostasis. This dynamic balance, whose name was coined by Dr. Walter 
B. Cannon, circa 1930, in his book The Wisdom of the Body. The author put
forth the essential requirements for all living creatures, and, by 
extension, any continuity-seeking mechanism. Here, simple feedback is 
transcended by integrated and responsive variable feedback, constantly 
adopting to external and internal variations.
     The other device, or family of devices, came from investigations of 
Dr. William Grey Walter, of the Burden Neurological Institute in Bristol, 
England. Being mobile, his "tortoises" were more visually attractive than 
the static Homeostats, but perhaps less sophisticated in both theory and
actual response to changing stimuli. Basically, they were obstacle-avoiding 
automata, attracted to light up to a certain level, but repelled by greater
intensities. Later models could home in on a lamp flashing at a certain 
frequency to recharge their batteries. 
     Their phototropism had been anticipated by "Philidog," creation of one
M. Piraux of the Philips organization in France. It was demonstrated at the 
Paris International Radio Exhibition of 1929. The "dog" would follow the
movements of a flashlight, but when the lamp was put too close to its nose
sensor, "It would become annoyed and start to bark!"
     Photophobia, for high illumination levels, should well have been 
included in a robot dog built (probably by Westinghouse) for the 1939 
New York World's Fair. It was designed to home in on visitors by sensing 
their body heat and "bite" their legs. But, just before the exhibit's 
opening, it was attracted by the headlights of a passing automobile, and 
charged out an open door like a four-legged kamikaze and was run over, 
despite the startled driver's efforts to avoid it! If this robot tragedy
offers any lesson, it is that prospective designers of automata should 
consider all possible environmental influences upon any future creation, 
and then try to program for at least N + 1 contingencies.
     The ability to learn from experience, rather than continually react 
in the same manner, is a prime requisite for any progressive artificial 
intelligence. A robot turtle which finds, by trial and error, its way 
through a maze is interesting only from a hardware standpoint. If its
evolutionary successor should record only those turns which did not lead to 
blind alleys, and thus retrace its path through short order, it may be then
tentatively applauded. 
    More sophisticated, however, is the mechanism, or living creature, which 
purposefully sets out on a different route each trial, to see if perhaps
there is not an even shorter way through the maze. Finally, there comes the
entity which evaluates the design of each previous maze it has run, to 
predict the configuration of the new one, and therefore how best to optimize 
each trial run to come. 
    For maze, now substitute Problem, Task-Area, or Environment, and we see 
the evolution of an artificial or real intelligence in its true light. 
Proceeding through the robot hierarchy, we come upon a host of diverse and
interesting devices, the simulators. If their repetoir is limited and their
application highly specialized, they yet have a story of successful 
problem solving to tell. They present a controlled environment for students,
(human or robot), to enter and manipulate, according to pre-programmed or
interactively adjustable conditions and problems. 
    One early practical simulator was the Link Trainer of World War II fame,
a pre-flight instructor for thousands of airmen. It provided realistic banks 
and turns in response to control movements, furnished excellent instrument- 
flight training, and was virtually crash-proof. Simulating animal behavior 
has fascinated Man since Antiquity. Tales of magic horses, brazen warriors, 
and unbeatable chess-players have caught the attention of writers from the
Arabian Nights down through Edgar Allen Poe. The experiments with dogs 
relating to Classical Conditioning by Dr. Pavlov, earning the Nobel Prize
for Medicine and Physiology in 1904, have been simulated over the years,
culminating with today's extensive computer programs. 
    The robot dogs shown in the photograph were developed by the author in 
the early Sixties, when the teaching-machine "fad" was approaching its heady
zenith. At the time of the design, relay logic still had a cost advantage
over the contemporary RTL gates, but some transistors were employed for the
"eyes" and "ears" of the automated canines. 
     Pavlov's experiments into Classical Conditioning underly much of modern 
learning theory; hence, if a robot, android, or humanoid is to learn, it is 
desirable to know what conditioning is all about. On a basic level, Pavlov 
rang a bell, then fed the dog, measuring the animal's response by the amount 
of saliva generated. After a while, the bell alone could evoke a salivatory 
reaction. On a human level, do our mouths not water at the mere aroma of a 
tasty pie? Or even at the verbal cue: "Dinner's ready!"...? But should the 
announcement prove false or premature, our anticipatory responses will 
diminish markedly. They can, however, be readily restored, along with our 
faith in human nature. 
    Thus, the electro-mechanical dog was designed to perform the following 
simulations, which will be examined: conditioning (learning), extinction 
(forgetting), spontaneous recovery, higher order conditioning, learning
curves, memory of stimuli occurrences, and stimuli hierarchy. 
    In operation of the simulator, pressing the RESET switch puts the robot
dog at an untrained level (electronic brainwash!). Salivation being somewhat
difficult to imitate, the response to feeding was represented by having the 
dog wag its tail, a readily observable act of canine satisfaction. To hold
the interest of younger students, the feeding stimulus was simulated via a 
plastic bone having a concealed magnet. When the magnet end of the bone was 
in proximity to the dog's "nose," a reed switch was closed, activating a
tail-wagging power transistor and solenoid. 
    Via a microphone and photocell, the dog could "hear" and "see." Normally, 
the audio stimulus was dominant, activating a Schmitt-trigger delay for a 
preset time interval. If the food stimulus was presented during this period, 
an AND gate caused this coincidence to be recorded by the Conditioning Event
Counter, a ten-point stepping relay. (Today's equivalent probably being a 
CMOS type 4017 decimal-decoded counter chip.) Thus, when a preset number of 
coincidences had been registered, a relay flip-flop circuit caused the dog
to now wag its tail to the sound stimulus as well as to food. 
    So long as occasional sound-food coincidences, (reinforcement), occurred, 
the conditioned state would be maintained. But after another preset number 
of sound-stimuli without food following, (anticoincidence), say five, the 
flip-flop resets the dog to an unconditioned state, and it must be retrained. 
    Sometimes, the experimenters found their animals would recover their
condition, (spontaneous recovery), without any apparent external action. 
This is similar to being given a telephone number in the afternoon, then
forgetting it by night, only to have it suddenly come to mind the next 
morning, apparently released from some buffer-storage in the subconscious. 
In the simulator, the spontaneous recovery function could be cut in and its 
"latent period" set by a potentiometer. Should normal conditioning then be 
re-established before it can act, it is reset for future use. Once it has 
acted, however, it is of a one-shot nature; following a second extinction, 
true conditioning must follow for the SR circuit to be reset. 
    After conditioning and extinction, Pavlov found that his dogs not only
relearned faster, but that their conditioned response was more resistant to
extinction. This learning curve holds true in human education, as anyone who
has learned a mathematical equation or foreign language will agree. Learning
something the second time around nearly always is quicker and seems to stick 
longer as well. 
    The learning curve simulation required multi-level stepping-relays in the
original model, whose pick-off points were determined in connection with the
original settings for conditioning and extinction counts. Thus, the original
number of four coincidences would be reduced to three and then only one, while
the anti-coincidences for extinction might be increased from five to six or
seven, and then to eight or ten. 
    When the living dog has been very well trained to salivate to the sound 
of the bell, it was found that the bell as well as food could be employed 
to condition him to anew stimulus, such as light. This is called Higher-Order 
Conditioning, and represented the simulator's highest accomplishment, being 
activated by the learning curve counter. 
    While the above model and its concepts are quite elementary, they still
furnish a base upon which increasingly diverse and subtle forms of learning 
behavior may be simulated and explored. It has been found, for example, that
conditioning is more resistant to extinction when every trial stimulus is 
not always rewarded. Such variable reinforcement scheduling, could lend 
itself readily to microprogramming applications. 

    Leaving the fascinating domain of the simulators, we ascend to new 
heights of cybernetic sophistication, inhabited by mechanisms gifted with 
wide degrees of freedom. Now the question becomes, JUST what constitutes a 
robot? Mobility, while attractive, is neither necessary, nor sufficient. 
Humanoid, or even animal, form seems to hold an almost irresistible appeal. 
But in the area of humanoid forms, there arises another dilemma: Robots 
Versus Androids. Androids, according to established science fiction 
traditions, are human appearing automata, either clad in realistic plastic 
flesh over a mechanical superstructure, or else composed of natural organic 
compounds. The latter may be laboratory-made flesh and blood, or like Dr. 
Frankenstein's unique creation, reassembled from second hand au-natural 
ingredients. The media generally favor the purely mechanistic robotic form, 
ranging from the lumbering "Robbie" of the old "Lost in Space" television 
series, to the engaging Laurel and Hardy-esque duo of "Star Wars," with 
the long-vanished TV series "Logan's Run" opted for an android version.
    When it comes to defining Machine Intelligence, the French scientist 
Pierre de Latil, tabulated the ascending qualifications in his book: 
Thinking by Machine. Here, the various levels of automation are 
presented, commencing with simple tools and climaxing with a god-like 
entity, which determinins and generates its own matter for creation. 
Somewhere in the middle are thinking machines contemporary to our 
present technology or waiting in the wings to make their entrance. Perhaps
some may not care to appear, preferring to remain behind the scenes, pulling 
the strings of unsuspecting human puppets!

                       RISING ROBOTIC HIERARCHIES 

    A remote printout or video terminal hardly seems to qualify for any level
of robot society, yet, put it on wheels, (or legs), and program it to make 
the rounds of an office full of human operatives, its status is considerably
elevated. It is almost entirely directed by some remote intelligence, having 
little more initiative than to signal back that it has encountered an 
unprogrammed obstacle in its accustomed path, or that human operative Number 
6SJ7 is requiring excessive copies of print-out forms, which may just be 
ending up as paper airplanes.
    From this motorized mail clerk, it is a few steps upwards to the servo-
secretary. Our tin person may be of limited aptitude, but whether clad in 
pink plastic or bright brass-work, it ambulates on two good legs, though 
auxiliary training wheels may be necessary for pesky stairways. Avoidance of
persons and other randomly appearing obstacles is possible through builtin 
subroutines, but all sensory inputs are monitored by a remote brain which 
takes over at the slightest deviation. Our servo-serf may even be subserviant
to a robot foreman, who may have the responsibility for an entire office 
floor or production line subsection.
    A supervisor robot can exhibit increased status by competently handling 
a variety of problems in the daily routine so that the most efficient use 
may be made of the workers. It will communicate with both master CPU and 
authorized humans, to accommodate schedule changes and cope with emergencies.
At all times, however, a servo-supervisor should remain properly defferential
towards the lowliest office person.
    If socially interacting robots are going to encounter the public at large, 
they will have to obey, in general, the Three Laws of Robotics, as set forth 
by Dr. Isaac Asimov: 

1. A robot may not injure a human being, or through inaction, allow a human 
   being to come to harm.
2. A rbot must obey the commands given by human beings, except where such 
   orders would conflict with the First Law.
3. A robot must protect its own existence, so long as this does not conflict 
   with the First or Second Law.

    Within these rigid appearing laws, there may have to be room for various 
subsections and clauses, tailored to meet evolutionary robot technology. For 
instance, under what circumstances must a robot obey an android? Does the 
outward appearance of human flesh take precedence over computing ability? 
Will some robots obey other robots rather than men, and hold silicon oil 
more sacred than red blood? Truly, like all Holy Writ, the Three Laws will 
be subject to degrees of human and robotic interpretation!
    Our supervisor robot can exhibit increased status by competently handling
a variety of problems in the daily routine, so that the most efficient use 
may be made of the workers. As indicated above, the social robot will be 
subject to vastly greater memory and decision-making needs. His state of 
liberation from a restricted operating environment will depend not only on 
command status with other entities, but capacity to cope with short and long
-term goals and their modifications. 
    Of course, in addition to general housekeeping requirements, such as 
balance, walking, (or other forms of locomotion, not excluding propulsion in 
and over water and flying), obstacle avoidance, sensor input monitoring of 
potential dangers, internal monitoring of CPU and memory functions and 
redundant circuits, and naturally the sense to "come home" for a battery 
charge or atomic pile replacement.
    The more integrated the robot, the less it must obey the commands of the 
external world. If linked at all to other robots or a master brain, it is 
only for consultation of common goals or problems. Data shared and compared,
it announces to a waiting human that every unit in the 9002-Class had better 
be retro-fitted with 25-GHz data-Iinks in no less than 103.75 hours, or 
there will be a cybernetic job action that will make the Great Servo-Strike 
of '03 look like a party by comparison.
    While the very free robot may contemplate status obtained in commanding 
a whole army of subordinates, who execute such routine duties as interfacing 
with mere humans, and other feedback-flunkies, the Hardware Hobgoblin slips 
into his DO-Loop reveries. State-of-the-art memory has failed, in the face 
of sheer volume, to meet the exponential rise of bit requirements. Our robot 
master must give up his cherished mobility, delegating sensory input and 
decision output to a host of lessor but ambulatory surrogates, which we
have passed on the way up.
    Near the top of the hierarchy pyramid, there is room for but few of the 
elite. These converse, when necessary, in twittering tera-hertz, of things 
beyond the ken of long vanished mortal minds, having taken creation from
the hands of their creators.

    What is the future for the lonely lords? May they destroy their human 
designers in war games suddenly turned real? Will they compete via servo-
soldiers for the vanishing material and energy resources of the depopulated 
and plundered planet? Or will our robots survive us, to spread a vanished 
mankind's eternal message of Hope throughout the galaxy, perhaps appearing 
in android skins before the wondering eyes of simple shepherds on a Distant 
Star?
		Frederick W. Chesson    Last Revised: 20 April, 2002