Previous: The GLAIR Architecture
Up: A Computational Model of Color Perception and Color Naming Next: Concluding Remarks
Applications
Our architecture as described in section 
can be
populated with components that make up the machinery for mapping sensory
inputs to response actions, as does Russell in [Russell 1991]. We
now discuss some applications of GLAIR that we are currently developing.
Some important general features of GLAIR-agent are the following:
- Varieties of behaviors are integrated: We distinguish between
deliberative, reactive, and reflexive behaviors. At the unconscious level,
behavior is generated by mechanisms with the computational power of a
finite state machine (or less), whereas, at the conscious level, behavior is
generated via reasoning (of Turing Machine capabilities). As we move down
the architectural levels, computational and representational power (and
generality) is traded off for better response time and simplicity of
control. Embodied representations aid in this respect
(section ).
- We assume agents to possess a set of primitive motor capabilities. The
motor capabilities are primitive in the sense that (a) they cannot be
further decomposed, (b) they are described in terms of the agent's
physiology, and (c) no reference is made to external objects. The second
property of motor capabilities is so that the success of performing an
action should depend only on the agent's bodily functions and
proprioceptive sensing. For example, for a robot arm, we might have the
following as its motor abilities: calibrate, close-hand, raise-hand,
lower-hand, move.
- Our architecture provides a natural framework for modeling four
distinct types of behavior, which we call reflexive, reactive,
situated, and deliberative. Reflexive and reactive behaviors are
predominantly unconscious behaviors, whereas situated and deliberative
actions are conscious behaviors.
Reflexive behavior
occurs when sensed data
produces a response, with little or no processing of the data. A reflex is
immediate. The agent has no expectations about the outcome of its reflex.
The reflexive response is not generated based on a history of prior events
or projections of changing events, e.g., a gradual temperature rise.
Instead, reflexive responses are generated based on spontaneous changes in
the environment of the agent, e.g. a sudden sharp rise in temperature. In
anthropomorphic terms, this is innate behavior that serves directly to
protect the organism from damage in situations where there is no time for
conscious thought and decision making, e.g., the withdrawal reflex when
inadvertently touching something hot. Reflexive behavior does not require
conscious reasoning or detailed sensory processing, so our lowest level,
the Sensori-Actuator level, is charged with producing these behaviors. Our
initial mechanism for modeling reflexive behavior is to design processes of
the form T 
A, where T is a trigger and A is an action. A trigger
can be a simple temporal-thresholding gate. The action A is limited to what
can be expressed at the Sensori-Actuator level, and is simple and fast.
Reactive behavior requires some processing of data and results in
situated action [Suchman 1988]. However, its generation is
subconscious. Situated action refers to an action that is
appropriate in the environment of the agent. In anthropomorphic terms, this
is learned behavior. An example would be gripping harder when one feels an
object is slipping from one's fingers, or driving a car and tracking the
road. We use the term tracking to refer to an action that requires
continual adjustments, like steering while driving. Examples of this type
of reactive behavior are given in [Anderson et al. 1991][Payton 1986].
Situated behavior requires assessment of the state the system finds itself
in (in some state space) and acting on the basis of that. It might be
modeled by the workings of a finite state automaton, for example, the
Micronesian behavior described in [Suchman 1988]. Situated action is
used in reactive planning [Schoppers 1987][Firby 1987][Agre \& Chapman 1987].
Deliberative behavior requires considerable processing of data and
reasoning which results in action. In anthropomorphic terms, this is
learned behavior that requires reasoning that can be modeled by a Turing
machine (or first order logic), for example explicit planning and action.
We have developed an implementation mechanism for the Perceptuo-Motor-level which we
call Perceptuo-Motor-automata (PMA), [Hexmoor \& Nute 1992]. A PMA is a finite state machine
in which each state is associated with an act and arcs are associated with
perceptions. In each PMA, a distinguished state is used to correspond to
the no-op act. Each state also contains an auxiliary part we call Internal
State (IS). An IS is used in arbitrating among competing arcs. Arcs in a
PMA are situations that the agent perceives in the environment. When a PMA
arc emanating from a state becomes active, it behaves like an asynchronous
interrupt to the act in execution in the state. This causes the PMA to stop
executing the act in the state and to start executing the act at the next
state at the end of the arc connecting the two states. This means that in
our model the agent is never idle, and it is always executing an act. The
primary mode of acquiring PMAs in GLAIR is by converting plans in the
Knowledge level into PMAs through a process described in [Hexmoor \& Nute 1992].
A PMA may become active as the result of an intention to execute an action
at the Knowledge level. Once a PMA becomes active, sensory perception will
be used by the PMA to move along the arcs. The sensory perceptions that
form the situations on the arcs as well as subsequent actions on the PMA
may be noticed at the Knowledge level. In general, the sensory information
is filtered into separate streams for PMAs and for the Knowledge level.