Next: Symbolic or Functional Grounding
Up: EMuds
Previous: Introduction
Subsections
Philosophy of mind is concerned with all the deep questions regarding
the human mind, from freedom of will, personal identity or the
existence of the soul to the nature of behavior, emotions or
consciousness. With the advent of computing machines, the philosophy
of mind has extended to the question of what minds are in general, and
of whether human-made artifacts can possess minds, an interrogation
that has renewed the interest in the mind-body problem. The mind-body
problem, which is the question of how the physical world relates to
the mental (experienced) world, now even appears to be the central
question relating all problems concerning the mind. Although the roots
of this discipline can be found with the Greek philosophers such as
Plato and Aristotle, I start my introduction with the modern day
answers that have been given to the mind-body problem.
Descartes was the first modern day philosopher to answer this problem
with his theory of dualism [42,14], where he
supports the idea that the physical world, including plants, animals
and the human body, is governed purely by mechanical/physical laws,
whereas human beings ``contain'' some non-material mental
substance. Although these two elements of nature are distinct and
disjoint, they interact through the human body: sensations create
thoughts, and thoughts drive actions such as bodily motion.
This view of the world has encountered many different problems over
the years, essentially over the nature of the interaction between such
different objects as matter and thought stuff and has spurred two
different trends in philosophy that avoid this question. On one hand
the mentalist approach that holds that all of nature is purely mental
experience or thought stuff and on the other hand, the materialist
(physicalist) approach, more typical of occidental thinking, that holds
that all of nature is made of physical matter and that somehow, the
mind emerges as a feature of the complex interactions in the physical
substance that forms the human body.
The objective of providing a theory of the mind has lead philosophers
through various stages in materialism, starting with the behaviorist
theory [57] that is grounded on an epistemological
position: that the only way to observe mind properties is through the
behaviors exhibited by the subjects that experience them. The
behaviorist theory has thereby also founded the psychology paradigm of
behaviorism, but has now lost some of its attractiveness since it
fails to explain the mind in situations where there is the need of
more powerful observation tools than the strict study of behavioral
patterns. The typical thought experiments that have been used to
contradict behaviorism being those of super-actors: how is one to
explain the mental state of ``pain'' when someone could act as if he
felt no pain when he is indeed in pain or how can one distinguish
real pain from the pain that an actor simulates perfectly.
The identity theory [1] that followed reached into the
human brain to identify physical brain events with mental states. In
this theory, mental states or thoughts were explained in terms of
neurophysiological states in the brain and the theory introduced the
notion that the meaning of words for example must be stated in terms
of brain configurations. Although this theory remains the basis of
today's functionalism, it encountered problems in that it presupposes
that mental items or types of items can be characterized by
neurophysiological states once for all subjects. This precludes
the possibility that different creatures from humans have minds and
even that humans beings are able to function with different brain
processes. To solve this problem, functionalism [47]
introduces the idea that it is not a specific physical component that
characterizes a mental state, but the causal role a physical component
has in the organism that characterizes mental item types such as
``pain'' or ``pleasure''. This characterization differentiates
token-token identity, one individual being's current pain state
identified with his current neurophysiological state, to type-type
identity, identifying pain with the more abstract functional role that
some component may play in an organism.
The functionalist theory of mind is thus focussed on three levels of
description [38]:
- the physiological description of an entity at a given time,
corresponding to a physical state token;
- the functional description of this state, relative to the
entity's organism;
- the eventual mental description of such a state if it
instantiates a mental item type.
With functionalism, the philosophy of mind introduces a theory that is
clearly meant to accommodate other types of minds than the human mind,
it generally demystifies the idea of minds and states that Mind is
simply brain stuff in (inter)action. The levels of description of the
theory provide a framework that leads from the physical world to the
mental world in three steps, independently of the realization of an
entity in the physical world. On the first level, the emphasis is
given to a physical description of an entity, that can potentially
have any form and is independent of external interpretation. On the
second level, the functional description of the entity's current state
that depends only on this entity's internal organization. On the third
level, the qualitative description of the current state instantiated
by the entity as viewed by an observer in reference to himself. Of
these three levels, the first two are independent of the observer and
the third implies that an observer might not be able to ascribe a
mental description to an entity if he is not capable of similar mental
states (or at least understanding them), even though the entity might
have such a mental description.
The obvious artificial candidate to validate this theory is the
digital computer. If it is possible to show that a computer with an
adequate design (hardware) and internal state (software) that can be
described physically and functionally can also be shown to possess some
form of mental description, then minds are purely emergent properties
of some kinds of material constructs.
In practice, computer science has for first objective the resolution
of problems with computers. The question of artificial minds is only
raised by the fact that some problems seem to require elaborate
cognitive faculties similar to those exhibited by humans if they are
to be solved. In this context, if computer science needs to reproduce
certain mental processes in order to solve some classes of problems
that humans can solve, either, how brains ``do'' minds must be
understood and imitated on computers or another method that is
functionally equivalent but works on computers must be found. The
current trend is to attempt the synthesis of some form of intelligence in
the form of computational models that include processes functionally
equivalent to those occurring in the brain. I will show later that
computers are instances of physical symbol systems, but the essential
comment that must be made here is that the underlying assumption of
the field of artificial intelligence is the physical symbol system
hypothesis [44] (PSSH), which states that a physical symbol
system has the necessary and sufficient means for general intelligent
action. That is, the PSSH (or an equivalent) is generally accepted as
a working hypothesis in the field and not as a problem to be
answered. Of course, if by practicing AI, strong evidence for the
correctness of this hypothesis can be exhibited, this will be
considered as encouraging results for the discipline.
Some additional assumptions are now deemed necessary for building
intelligent machines and I will introduce them shortly with a few key
points at the end of this section, but before that, it is necessary to
speak of the abstract machines we consider when speaking about
computers in general.
The notion of Physical Symbol System dates back to the origins of
computing, with the work of A. Turing or J. Von Neumann. It is linked
with the introduction of Turing machines that are used to define the
concept of algorithm by simple operations on symbols. The first
formalization of such systems by Newell and Simon [44] was
intended to provide a general category of systems that are capable of
intelligent behavior.
To formally define algorithms, Turing reduced operations of the kind that are
intuitively used as steps in an algorithm to simple mechanical
operations in an idealized machine. In his theory, the machines he
considers consist in a read/write head scanning the surface of a tape
divided into cells (see figure 2.1). The tape cells can
hold symbols from a predefined
finite alphabet and extends to infinity at least in one direction. The
head can move along the tape in both directions (to the ``right'' or to
the ``left''), reads symbols in the tape cells and can inscribe new
symbols in these cells. At any given time, the machine is in a state
chosen from a finite set of states and depending on this state and the
symbol read on the tape, it can decide to write a new symbol in the
current cell and eventually move to the left or the right on the tape.
Figure 2.1:
A Turing machine.
 |
Clearly, any Turing machine with its set of states, symbol alphabet
and current internal state instantiates an algorithm. The
Church-Turing thesis, on the other hand, states that any algorithm a
human being would consider can be represented as a Turing machine,
thus giving the equivalence between Turing machines and our intuitive
understanding of algorithms. To this day, this thesis holds and is
generally accepted by the scientific community as the right definition
of an algorithm.
Since the definition of Turing machines, it has been shown that many
formal systems are equivalent to Turing machines and thus define the
same notion of algorithm. Moreover, the modern day computers built
following the model of von Neumann are instanciations of the formal
Turing machines, but where some restrictions have been introduced,
such as a finite length tape. The theory applying to Turing machines
can generally be extended to other equivalent formal or physically
realized systems and Newell and Simon have
called all these systems symbol systems. They then postulate that
physical symbol systems are a category of systems that are capable
of intelligent behavior in their ``Physical Symbol System Hypothesis''
[44]. Their work then goes on to try and convince us that
this hypothesis is true and even that human beings are physical symbol
systems. Clarifying the work of Newell and Simon, Harnad
[22] gives the following definition for a symbol system.
A symbol system is:
- a set of arbitrary symbols called physical tokens
manipulated on the
basis of explicit rules, as well as strings of tokens built on
the atomic tokens
- where manipulation is based solely on the ``shape'' of
tokens (i.e. is syntactic)
- manipulation consists in combining via rules the atomic
or composite symbol tokens
- the syntax can be systematically assigned a meaning,
i.e. it is semantically interpretable
This description gives an intuitive view of what symbol systems
can do, although it must be said that there is nothing more here than
what was already in Turing machines. In fact, the foundational work in
artificial intelligence led by Newell and Simon on symbol systems has
even earned these systems the status of a ``symbolic model of mind''
[18] for psychology, where symbol strings capture
mental phenomena such as thoughts and beliefs. But it suffices to say
that such systems make explicit what computers can do and if we are to
work with computers, a symbol system provides a good abstraction of
these machines.
The symbol system definition typically omits a vital component in the
study of artificial intelligence. It has long been assumed that a
system can be considered as a closed system, but this can never hold
in practice and I believe it is important to be aware that the whole
underlying assumptions made about the systems under study are based on
a conceptual shift of perspective that originates in the
nineteen-thirties with Ludwig Von Bertalanffy's General Systems Theory
(GST) [65] and the Cybernetics movement begun at the
Macy conferences from 1946 to 1956 [66]. The two fundamental
assertions that should be retained from this systemic enterprise are
adequately illustrated by the two following quotes:
When we try to pick-up anything by itself, we find it is attached to
everything in the universe. John Muir
The science of observed systems cannot be divorced from the science of
observing systems. Heinz Von Foerster
That is, contrary to the common methodology often used in the
physical sciences: 1) all systems must be considered open, there is no
such thing as a closed system, 2) since all systems are open, an
observer is always part of the system he is observing and plays a role
that affects the objects under his observation. This role of the
environment in the development of an entity is particularly important
to artificial intelligence.
It should be noted that although the popularity of GST and Cybernetics
has steeply fallen from the original enthusiasm they had
generated (fallen in disrepute even). These theories are the seeds
that allowed models such as Maturana and Varela's theory of
Autopoiesis and Cognition [39] or the general
field of Cognitive Science to grow [17], and these now play
a dominant role in many areas of modern research, necessarily
including artificial intelligence, but also philosophy of mind,
psychology, economics or management.
Maturana and Varela present their biological theory of cognition in
[39]. To introduce their ideas, many concepts and terms
are used, but essentially they emphasize the fact that living
organisms are structures able to sustain their unity in an environment
(via autopoïesis or self production) and situated in a
physical space. These organisms are instances of a category of
structures defined by their organization, with organization
used in the sense of internal relations and dynamics. It is by the
maintenance of this organization that an organism preserves its
autonomy with respect to the rest of the environment.
When an organism evolves in an environment, the actual changes that it
experiences are controlled by its structure as a result of the
coupling between the environment and the organism's structure through
senses. In relating an entity to an environment or other entities,
domains of interactions can be defined such as the domain of
relations (set of relations in which an entity can be
observed). In this context, Maturana and Varela state that
``Living systems are cognitive systems, and living as a process is a
process of cognition''. The necessary conditions for a system to be
cognitive are thus that the system exists within a physical space and
is structurally coupled to it (embodiment) and the cognition
itself is the behavior of an entity engaging in interaction with the
environment. Much prominence is given to the fact that the environment
does not transform a living organism, but that the structure of the
organism controls the changes it can undergo when subject to
environmental perturbations. The notion of representation is also refuted
by the theory, as an organism does not engage in the construction of
a model of the world, but is presented with perturbations of its
structure through its sensors, an experience that it can then actively
rebuild as needed. Note however, that I will continue using the term
representation even for such situations.
The influence this theory has over the field of computer science is in
the methodology as I will show in the chapter on agent systems. It is
with the concept of experiential enaction [63] that
the relation between the mind and the body is explained. In the
enactive view, the only relation between the mind and the body is to
be found in the nature of the experience that is acquired by engaging
in worldly activity with this body and mind. Therefore, to build
artificial minds one requires bodies that exist and experience the
world. And so, a new requisite for artificial intelligence can be
postulated by stating that intelligence is contingent on being embodied
in the world.
While this assumption appears essential to me, the restrictive view
often taken that the world in which an agent must be embodied can
only be the real world seems unnecessary. I like to assume that any
type of world can be host to problems that may be solved by entities
existing within them. That environments require some high level of
complexity and/or unpredictability for cognitive level processes to
evolve in the entities that inhabit them is not excluded, but even in
simpler spaces, there are many interesting questions that
remain to be answered.
Next: Symbolic or Functional Grounding
Up: EMuds
Previous: Introduction
Antony Robert
2000-02-11