Ambient Intelligence Through Agile Agents - Ambient Intelligence for Scientific Discovery - page 288

of action. In AF-APL a commitment is comprised of: an agent identifier (the
agent for whom the commitment has been made), a start time (before which the
commitment should not be considered), a maintenance condition (which defines
the conditions under which the commitment should not be dropped), and an ac-
tivity (the course of action that the agent is committed to). Currently, an activity
may take one of three forms: (1) an action identifier (i.e. some primitive action
that the agent must perform), (2) a plan identifier (i.e. an identifier that can be
used to retrieve a partial plan from the agents plan library), and (3) an explicit
partial plan (i.e. partial plans can be directly encoded into a commitment).
Action identifiers are modeled as first-order structures where the parameters
may be used to customize the action. For example, the activity of one agent
informing another agent of something is realized through the inform(?agent,
?content) action. Here, the ?agent parameter refers to the agent to whom the
message is to be sent (their identifier), and the ?content parameter refers to the
content of the message. An example of an inform message can be seen in Fig.
2. Within AF-APL, actions are realized through the triggering of an associated
actuator unit . As with perceptor units, actuator units are associated with specific
agents as part of the agent program through the ACTUATOR keyword. Actions
can be combined into plans that form more complex behaviors using one or
more plan operators - currently there are four plan operators: sequence ( SEQ ),
parallel ( PAR ), unordered choice ( OR ), and ordered choice ( XOR ). Plans can be
stored within an agent's internal plan library, where they are distinguished from
one another by a unique plan identifier.
Finally, commitment rules describe the situations, encoded as a conjunction
of positive and negative belief literals, under which the agent should adopt a
given commitment. An implication operator ( ) delimits the situation and the

⇒

commitment.
AF-APL agent programs (actuators + perceptors + plans + initial mental
state + commitment rules) are executed upon a purpose-built agent interpreter.
Specifically, the agent program is loaded into appropriate data structures inside
the agent interpreter. The interpreter then manipulates these data structures
through a simple control cycle that encapsulates various axioms defined in the
associated logic of commitment. This cycle is comprised of three steps: (1) update
the agents' beliefs, (2) manage the agents' commitments, and (3) check whether
or not to migrate. It is invoked repeatedly for the lifetime of the agent (at least
whenever the agent is active).
By way of illustration, Fig. 2 presents a fragment of AF-APL code from a
WWW spider agent that provides a service in which it informs subscribed agents
of any new documents it finds. This is realized through two commitment rules
and one initial belief. The first commitment rule states that if the agent receives
a request to subscribe to a service (identified by the variable ?svc ), and the agent
believes that it provides the service, then it should commit to performing two
actions in parallel (specified by the PAR plan operator). The first action involves
the agent informing the requester that they have successfully subscribed to the
service, and the second action involves that adoption of a belief by the agent