COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	.cb THE MINDS OF DOGS
C00014 ENDMK
C⊗;
.cb THE MINDS OF DOGS


	One of the major areas of study in %2cognology%1 (artificial
intelligence) is what information structures are required in order to
solve different kinds of problem.  We can put this in the language
of %2common sense psychology%1 as the question of what kind of mind
is required to achieve certain goals.  In this connection the minds
of animals are interesting.

	Consider the following thought-experiment with dogs:

.item←0
	#. There are four dog houses each having a light and a heater.

	#. These facilities can be operated by a dog using foot switches
or nose switches.

	#. The dog has been trained to obey commands "turn on the light",
"turn off the light", "turn on the heater" "turn off the heater".

	#. The effects of the switches in controlling the light and
heater are not constant, and the dog has learned to operate them
until the commanded result has been obtained.

	#. A cushion is fastened to the floor of one of the houses,
sometimes to one house and sometimes to another.  Other things being
equal, the dog favors the house with the cushion.

	#. When it is cold, the dog prefers a house with the heater
on, and when it is hot, he prefers a house with the heater off.

	We can ask whether the dog will control the temperature of the
house with the cushion according to its preference.  The null result would
be that if the dog is in the house with the cushion and it gets cold, he
will go to the warm house rather than turn on the heater in the house with
the cushion.

	The intent of the provision that the dog has been taught to
manipulate the switches until the commanded condition has been achieved
is to be sure that the dog has learned to achieve the result - not
merely to perform a specified action.

	Here are some questions:

.item←0
	Supposing that some dogs operate heaters for their comfort,
here are some questions that can be asked:

	#. Do dogs have preferences about lights?  There may be
gaps in the connection between preferences and the measures
an animal will take to achieve these preferences even when the action
is one the animal can be trained to perform.

	#. Does training a dog to achieve a condition insure that he
will bring about the condition himself when he prefers it?  Is it
immediate?  If not, can it be learned in general, so that it becomes
immediate for a new condition?  If not, can it be learned for each
specific condition?

	#. Does a dog learn something in order to please humans
more readily than for his own benefit?  How about wolves and coyotes?
Perhaps there is a general instinct to obey.  Often it seems that
a dog is watching its master trying to figure out what the master
wants.

	#. Suppose that a dog has first learned to request a human
to turn on the heater (assuming that dogs can learn this as they
learn to request to be let in or out).  How much will this interfere
with his subsequent learning to turn on the heater for himself?

	One may conjecture that living with humans interferes with
dogs' learning to do things for themselves.  Imagine that we lived
in close proximity with gods who did highly visible things for totally
inexplicable reasons and sometimes rewarded and punished us for reasons
that were sometimes explicable and sometimes were not.  It seems that
it would interfere with our ability to learn to achieve goals.  In fact,
we can perhaps characterize the original form of religion as the
hypothesis that there were such gods, and speculate that the hypothesis
did interfere with our acquiring competence in understanding and
controlling our environment.

	On this basis, we might conjecture that dogs or apes would
develop a more interesting culture if they lived in total ignorance
of humans, and if their enviroment had many useful causal relations
and that apparent miracles, e.g. a new object putting in an appearance
one fine day, happened rarely if at all.
This may be required for apes to reach their full intellectual
potential.  Perhaps a compromise could be developed that would
still allow apes to be taught sign language, but would minimize
miracles in their environment.
The best environment might not be natural.  A simpler environment
of geometrically shaped simply colored rooms might be better, but
once the apes are introduced to it, it shouldn't be changed except
by comprehensible processes.

	Another question about the minds of dogs arises from the following
experience.  I had a house with an %2atrium%1 in the middle, and the walls
enclosing the %2atrium%1 were of glass.  If food was placed outside a
closed glass door, neither my dog nor my cat would look across the ⊗atrium 
and notice that a door on the other side was open.  They would sometimes
find the door by searching, however.  Imagine the following formal experiment:
A dog can go around an obstacle two ways, each way leading to a door to
the goal chamber.  Will it learn to look across and see which door is open?
More generally, can it learn to look at a maze from above and run it efficiently?
What about looking at a map (adapted to the animal's visual system)?
What about apes?

	In general, it seems that testable hypotheses can be formulated about
what animals believe and what mental structures they have.
Much of the discussion in my
%2Ascribing Mental Qualities to Machines%1
is applicable to animals.

	Much behavior is evidently genetically coded.  A proper study of
this requires determing what are the "elements of behavior" in terms of
which complex behaviors are encoded.  Perhaps sequences of instinctual
behaviors in lower animals like spiders or bees could be investigated
in the manner in which the metabolic pathways of bacteria are investigated.
There one damages an organism genetically and investigates what steps
in a synthesis are broken.  Here one would damage the organism genetically
and investigate where the sequence of behavior is broken.

Notes of Dec. 19, 1978

1. Presumably dogs think about food a lot.  Do they imagine specific
foods or is their thought more primitive and general?  Can we test it.
Suppose we train dogs to request specific foods, i.e. unless the request
is specific he gets nothing.  Sometimes the dog is told that the requested
food is not available and he must request something else.  Sometimes when
he requests a specific food, he is confronted with a set of dishes.  Does
he actually eat what he requested or does he choose anew?

	I am optimistic that a set of hypotheses can be devised that will lead
to good experiments.  Such experiments would go beyond the maze type
psychology that is traditional.  Ascription of mental qualities will
be essential for this.

	The recent work (1981) on whether apes and other animals have
a sense of self is an example of a question of what concepts different
animals can use.  However, the following surprising result might be
obtained.  It has been discovered that gorillas are among the animals
not having self concepts.  It might be supposed that this is a fundamental
part of the psychology of an animal.  What if Penny Patterson were to
discover that Koko has no concept of herself, but were able to teach
Koko to have such concept?

Various revisions were made 1981 December.

John McCarthy

The source of this is DOGMIN[E77,JMC] at SU-AI.  Version pubbed at {time}
on {date}.