Goodwin begins with a very simple working definition of agency: "an agent is an entity created to perform some task or set of tasks".
He divides agents into two classes, general and deliberative. He states that for agents in general, the only meaningful approach is the behavioural approach, that is, defining and describing agents in terms of what they do, without ascribing internal states of any kind. Deliberative agents are those which contain an internal model of the world (including the agent itself) and act based on the state of that model. The state of the model is modified over time according to sensory input and in response to the agent's own actions.
The fundamental properties of an agent in Goodwin's terms are that such an entity be capable, perceptive and successful. Capable means able to affect its environment in ways that will allow it to accomplish its task, that is, having suitable effectors. Perceptive means able to perceive its environment in ways that allow it to bring its effectors to bear in order to accomplish its task. Successful means effective in accomplishing its task. To these three, he adds reactive (meaning to be able to respond to events in the world swiftly enough to be successful) and reflexive (meaning responding in a stimulus/response fashion).
To the five general properties given above, Goodwin adds for deliberative agents predictive, interpretive, rational and sound. Predictive means able to predict how it can achieve its task, based on an internal model of the world (environment). Interpretive means able to correctly use its sensor readings, and rational means that the agent chooses to do those things that it predicts will achieve its goals. Sound is simply the state of being predictive, interpretive and rational.
Goodwin begins by defining these properties in a binary fashion
(an agent that is/isn't successful, is/isn't capable and so on)
then extends the definitions into a relative framework to allow
comparisons between agents. This is done by introducing the concept
of a utility function which allows "success' to be compared
- from this flow the other comparisons. For example, one agent
is more capable than another if it is more successful ("has
higher utility") in the same environment than the other.
While Goodwin implies that all agents exist in order to effect
changes of some sort in their environment, he never actually states
this.
A housebrick could be an agent within Goodwin's formalism: The
surfaces of the brick are its effectors. Its task is to not compress,
that is, its task is to hold other objects apart (presumably other
bricks). Its sensors are non-existent - sufficient, but degenerate.
It is reflexive in that it responds to the "stimulus"
of pressure by resisting collapse, though this is a somewhat long
bow! The brick could even be seen as a deliberative agent, with
a degenerate internal model of the world, though it seems (slightly)
more satisfying to consider the brick a reflexive agent.
The properties "successful" and to a lesser extent
"capable" provide admirably well for a sliding scale; Goodwin in fact
wishes his formalisation to provide a basis for comparing the performance
of agents performing the same tasks in the same environment. A
sliding scale needs extremes, and perhaps the brick is one extreme.
The only escape seems to be the tenuous requirement that "success
means total success" - since the brick performs its task
indefinitely, we have no way of knowing that it hasn't
succeeded unless it actually eventually succumbs to pressure and
collapses (or is removed from duty). Until then we can only say
"so far, so good...".
This same comment would appear to apply to any task falling into
the class of tasks Goodwin refers to as "tasks of maintenance",
that is, tasks where the agent must "maintain some condition
in the environment". The example Goodwin gives is an agent
controlling room temperature. This could be avoided by modifying
the formalisation so that such tasks are defined as sequences
of tasks, each to be achieved within a set period. However, it
seems to me that allowing a housebrick status as an agent, even
as a degenerate case, is a weakness in this formalism.
There seems to be some ambiguity in Goodwin's formalism regarding
the place of humans. Firstly, theological arguments aside, it
does not seem obvious that humans are created to accomplish specific
tasks (Goodwin's fundamental definition). Secondly, Goodwin specifically
limits his model to "simple deliberative agents" which
do not learn. That is, they can alter the state of their internal
model, but not the model itself. However in defining a non-deterministic
environment, Goodwin postulates agents with free will (presumably
humans) as part of that environment.
On the whole I suspect that Goodwin does not intend his formalism
to apply to humans, though humans are part of the environment
within which he expects agents to operate. The ambiguity arises
from a statement of purpose ("...created to perform...") without
any statement of the source of that purpose.
Comment
The relationship between agent and environment is unsatisfying in
Goodwin's formalism. Defining the environment as whatever was there
before the agent was built is uncomfortably akin to describing the art of
sculpture as removing all the marble that doesn't look like the subject.
Email me!
Last modified 16 December 1995, 23:45
© Copyright 1995 Karl Auer