Information Technology Reference
In-Depth Information
three parts: (1) physical actions in which the agent engages, (2) low-level be-
haviors, which are the agent's simple activities, and (3) high-level behaviors,
which combine low-level behaviors into high-level activities using more com-
plex reasoning. Because these activities are implemented according to what
makes sense from the agent's internal point of view, there is no necessary cor-
relation between the agent's behaviors and the behaviors we would like the user
to see in our agent.
But if the agent is to be narratively comprehensible, it may make more
sense to design the agent according to the desired user interpretation, i.e. mak-
ing the internal behaviors exactly those behaviors we want to communicate
to the user. Then, communicating what the agent does reduces to the prob-
lem of making sure that each of these behaviors is properly communicated.
For this reason, the Expressivator structures an agent not into physical actions
and problem-solving behaviors, but into signs and signifiers, or units of action
that are likely to be meaningful to the user. This structure involves three levels,
roughly corresponding to those of generic behavior-based AI: (1)
signs
,which
are small sets of physical actions that are likely to be interpreted in a particular
way by the user; (2)
low-level signifiers
, which combine signs, physical actions,
and mental actions to communicate particular immediate physical activities
to the user; and (3)
high-level signifiers
, which combine low-level signifiers to
communicate the agent's high-level activities.
There are several differences between these structural units and the default
behavior-based ones. Unlike physical actions and behaviors, signs and signi-
fiers focus on
what the user is likely to interpret
, rather than what the agent is
'actually' (i.e. internally) doing. In addition, signs and signifiers are
context-
dependent
; the same physical movements may lead to different signs or sig-
nifiers, depending on the context in which the actions are interpreted. Most
importantly, signs and signifiers carry an
explicit commitment
to communica-
tion; they require the agent designer to think about how the agent should be
interpreted and to provide visual cues to support that interpretation.
Signs and signifiers are not simply design constructs; they also have tech-
nical manifestations. Formally, a sign is a token the system produces after hav-
ing engaged in physical behavior that is likely to be interpreted in a particu-
lar way. This token consists of an arbitrary label and an optional set of argu-
ments. The label, such as “noticed possible insult”, is meaningful to the de-
signer, and represents how the designer expects that physical behavior to be
interpreted. The arguments (such as “would-be insulter is Wilma”) give more
information about the sign. This token is stored by the sign-management sys-
tem described below, so that the agent can use it to influence its subsequent
