An agent is an animate entity that is capable of doing something
on purpose. That definition is broad enough to include humans and
other animals, the subjects of verbs that express actions, and
the computerized robots and softbots. But it depends on other words
whose meanings are just as problematical: animate, capable, doing,
and purpose. The task of defining those words raises questions
that involve almost every other aspect of ontology.
For the primitive terms of any theory, circular definitions are inevitable.
As an example, Newton's famous equation F=ma appears to define
the force F in terms of the mass m and the acceleration a.
Yet that same equation could be used to define the mass in terms of the
force and the acceleration. Newton assumed that acceleration could be
independently defined in terms of space and time, but Einstein showed
that the structure of space and time itself depends on the mass
of the entities in it. The fundamental concepts of any subject
can only be defined implicitly by laws or axioms that express
a pattern of relationships among them. Closed-form definitions are
never possible for basic primitives.
Literally, an animate entity is one that has an anima or soul.
But anima is the Latin translation of Aristotle's word
psychê, which had a much broader meaning than the English
word soul. Aristotle defined a hierarchy ranging from
a vegetative psyche for plants to a rational psyche for humans.
The first question is whether Aristotle's hierarchy of psyches
can accommodate the modern robots and softbots.
The agent of a verb plays that role only as long as the action persists,
but an entity can also be considered an agent if it has the power
to perform some action whether or not it actually does.
Formalizing that notion of power raises questions about modality,
potentiality, dispositions, and counterfactuals that have
been discussed in philosophy for centuries.
The verb do sounds as simple as two other little verbs
be and have. But like those verbs, its dictionary entry
has one of the largest number of senses of any word in the English
language. A common feature of all those senses is causality and purpose:
some agent for some purpose causes some process to occur.
This feature not only creates a cyclic dependency of doing
on agent, it also introduces the notions of causality,
process, and occurrence.
In the top-level ontology, purpose
is defined as an intention of some agent that determines
the interaction of entities in a situation. That is consistent
with the definition of an agent as an entity that does something
on purpose, but the circularity makes it impossible to give
a closed-form definition of either term.
Psychology of Agents
Linguistically, an agent is an animate being that can perform some
action, and an action is an event that is initiated or carried out
by some animate being. The circularity in those definitions can be
broken by determining what characteristics of an animate being are
necessary for it to play the role of an agent. Then those features
can be generalized to a definition of agent that applies
to people, animals, robots, and certain kinds of computer programs.
The word animate comes from the Latin anima,
which means breath or soul. The medieval Scholastics used
anima as a translation of the Greek psychê,
which also means breath or soul. The basis for the modern terminology
is Aristotle's treatise Peri Psychês, which is called
De Anima in Latin or On the Soul in English.
Aristotle defined the psyche as the logos or principle
that determines what it is for something to be a living entity.
Instead of a single principle of the psyche that covered
all living things, Aristotle found
six related functions, which he arranged in a hierarchy: nutrition,
perception, desire, locomotion, imagery, and thought:
We must inquire for each kind of living thing, what is its psyche;
what is that of a plant, and what is that of a human or a beast.
The reason why the functions are arranged in this order must also
be considered. For without nutrition, there does not exist perception,
but in plants, nutrition is found without perception.
Again, without the sense of touch none of the other senses exists,
but touch exists without the others, for many animals have
neither vision nor hearing nor sense of smell.
And of those that can perceive, some have locomotion, while others
have not. Finally and most rarely, they have reason and thought.
Those mortal creatures that have reason have all the rest,
but not all those that have each of the others have reason;
some do not even have imagery, but others live by this alone.
The rational intellect requires a separate principle (logos).
An appropriate definition of each of these functions would be
the most appropriate for the psyche as well. [414b32]
Aristotle's hierarchy of functions was based on his extensive study
of the plants and animals known in his day. With his criteria,
he was the first to recognize that sponges were primitive animals rather
than plants. The subdivisions in the tree of Porphyry (Figure 1.1 in the
book Knowledge Representation),
are based on Aristotle's distinctions of animate/inanimate,
sensitive/insensitive, and rational/irrational.
Aristotle's hierarchy resembles the competence levels
that Rodney Brooks (1986) defined for mobile robots.
A robot is an AI system that receives signals
from the environment and acts on the environment in a way
that helps it to achieve some preestablished goals.
In what he called the subsumption architecture for mobile robots,
Brooks distinguished eight levels of competence, each with
increasingly more sophisticated goals and means for achieving them:
Each of these levels depends on and subsumes the competence
achieved by the earlier levels. Each level responds to signs, signals,
or stimuli from the input sensors and generates output for the motor
mechanisms. Yet the robot as a whole does not depend on a strict
control hierarchy. The first few levels by themselves could support
an insectlike intelligence that responds directly to immediate inputs
without doing abstract reasoning or planning. The higher levels
could inhibit the lower levels and take control for more sophisticated
or intelligent behavior, but the lower levels would still be capable
of automatic, reflexlike reactions to danger signals.
- Avoiding. Avoid contact with other objects, either
moving or stationary.
- Wandering. Wander around aimlessly without hitting
- Exploring. Look for places in the world that
seem reachable and head for them.
- Mapping. Build a map of the environment and
record the routes from one place to another.
- Noticing. Recognize changes in the environment
that require updates to the mental maps.
- Reasoning. Identify objects, reason about them,
and perform actions on them.
- Planning. Formulate and execute plans that involve
changing the environment in some desirable way.
- Anticipating. Reason about the behavior of other
objects, anticipate their actions, and modify plans accordingly.
The behavior of the lower levels depends primarily on immediate inputs.
The higher levels depend more heavily on internal representations,
such as maps of the environment, memories of previous inputs,
stored patterns for recognizing familiar objects, and
established habits for repeatable behaviors.
Every level responds to signs from the external environment and
from other internal levels, but there is an increase in complexity
from the automatic responses at the lower levels to
the knowledge-based reasoning at the higher levels.
Aristotle's levels may help to clarify and refine the
competence levels. Nutrition, which Brooks omitted, is necessary
for a robot to recharge its batteries; and desire or something
like it is necessary to determine goals for the robot at every level,
from the most primitive nutrition to the most sophisticated planning.
What distinguishes a software agent from an ordinary program
is a unifying principle that gives it a certain autonomy.
Following Aristotle, that principle may be called its psyche,
and its definition can be based on an appropriate definition
of each of its functions. The six functions of the psyche, which
Aristotle applied to living things from plants and insects to humans,
can serve as metaphors for the functions of artificial agents:
The notion of psyche with its hierarchy of functions
provides a framework for classifying agentive behavior.
The psyche of an agent is its functional organization, and its level
of sophistication depends on how much of the Aristotelian range
of function it is able to support. A formal definition
of the term agent might be based on a formalization
of the informal hierarchies proposed by Aristotle and Brooks.
But such a formalization would require a complete axiomatization
of all the top-level concepts
in the ontology, which Peirce said is a "labor for generations
of analysts, not for one."
- Nutrition. For a robot or embodied agent, nutrition is
the act of recharging its batteries or energy stores from time to time.
For a software agent, nutrition is the procurement of computer time
and storage space from a host system. A computer virus is a parasite
that steals the time and space; a more benign agent lives in a symbiotic
relationship with its host, providing useful services in exchange
for room and board.
- Perception. For a robot, perception depends on input
sensors and the ability to interpret the inputs. A television camera,
for example, may provide a stream of data; to see, however, the robot
must convert the data to a representation of objects in the environment.
For a software agent, perception requires access to input devices
of the host and the ability to interpret data from those devices.
- Desire. Aristotle's general word for desire is
orexis, which causes an agent to reach for what
is desired -- one that doesn't reach is anorexic.
He distinguished three aspects of desire:
appetite (epithymia), passion (thymos),
and will (boulêsis). He classified appetite and passion
as feelings shared with beasts and will as the result of rational thought.
In their psychology of agents, Moffat and Frijda (1995) made
a similar distinction between preference and will.
For a software agent, the built-in equivalent of appetite or passion
gives it a preference for certain kinds of states. Its will is
determined by a logically derived plan for reaching a preferred state.
- Locomotion. For mobile robots, locomotion is a basic
function that may be further divided into subfunctions, such as
Brooks's competence levels. Software agents, which operate in
some host system, may use the input/output devices of the host to explore
the environment, including anything reachable via computer networks.
- Imagery. Aristotle's term phantasia,
according to the Liddell and Scott dictionary, means the appearance,
presentation, or representation of images "whether immediate or
in memory, whether true or illusory." The processing of imagery
by computer is an active research topic in artificial intelligence.
Like Aristotle, many researchers believe that important aspects
of animal-level intelligence can be achieved by manipulating imagelike
data structures rather than propositions (Glasgow et al. 1995).
- Thought. Aristotle reserved the highest level of the
psyche for rational thought. His term for rational animal
was zôon logon echon (animal having logos).
For software agents, rational thought
corresponds to the deductive and planning capacity that transforms
the motivating forces of appetite and passion into will.
A rational agent must be able to perceive relevant aspects
of a situation, evaluate their desirability, and determine plans
for transforming the current situation into a more desirable one.
For further discussion about continuous processes, discrete processes,
and causal influences,
see the paper on Processes and Causality.
Send comments to John F. Sowa.