Artificial Mind for Virtual Characters - Advancing Artificial Intelligence Through Biological Process Applications - page 191

inputs from the outer environment ( E outer ) in
order to build the inner environment ( E inner ).
Each SCA is a cognitive agent designed to cap-
ture specific environment aspects, filtering and
controlling their inputs so as not to overload
the composite agent if the outer environment is
rich and complex. The other set groups reactive
agents (RAs), each one responsible for promoting
a specific composite agent behavior, then coding
what behaviors the agent can execute. The inputs
to RAs come from E inner .
These Composite Agents use the “ sense-
update-act ” model (Wooldridge, and Jennings,
1995). This means that they first sense the envi-
ronment, then update their internal environmental
representation, and finally decide what action to
perform, before restarting the cycle. They can
sense only those aspects perceived by the active
SCAs, which are constantly updating an internal
environment representation ( E inner ). Combining
E inner information with goals and knowledge
about the context, composite agents can generate
appropriated actions.
Composite Agent Architecture is interesting
for artificial minds design because it matches some
of the characteristics enumerated earlier, like a
multi-modular structure and emergent behavior.
However, the “ sense-update-act ” model does
not seem convenient to the intended Perception-
Knowledge-Action relationship as previously
discussed. Instead, a two-steps “ motivation 
knowledge-based perception/action ” cycle would
be more appropriate. While in the first case the
environment information filtering is made at the
perception level, in the second, it is made at the
motivation level. So, attention is determined from
the outside in the composite agent architecture,
and from the inside in the artificial mind structure.
Although sometimes the environment can catch
our attention (usually when extraordinary stimuli
are present - like intense sounds or lights), most
part of the time our senses are guided by our goals.
Besides, information filtering is a potent way to
decrease computational processing. If the outside

determines it, the whole process may be slowed
down in complex environments. But the inside
has a constant structure that can be adequately
designed.
The multi-agent structure here proposed is
inspired in the composite agent architecture, but
functionalities are differently distributed. First of
all, there are three sets of sub-agents: Motivation
agents, Cognition agents and Execution agents.
Motivation Agents code the character needs
and desires, conveying the internal forces that
drive the character behavior. They compete for
priority and constitute the Inner Motivational
Environment (IME), which represents the current
emotional and affective state of the character.
Only those Motivation Agents with higher-level
of priority can act outside the IME by activating
specific pre-defined Cognition Agents that have
the knowledge to lower their activation level.
Cognition Agents are responsible for the
“ knowledge based perception/action ” step of
the iteration cycle. They store and represent the
character knowledge, constituting the character
cognitive structure. Like a motivational graph
node described previously, each one embodies
knowledge that can be activated by a certain
quantitative internal force. They can be compared
to Minsky's agents in the sense that they execute
some specific and particular task, and they are
related to each other in a hierarchical structure. In
terms of implementation techniques, they are not
necessarily restricted to rule sets. Their knowledge
may be represented by any AI paradigm. The
developer may choose whatever technique seems
best suited to the task in question. A Cognition
Agent uses the sense-update-act cycle but only
in a very restricted scope. It searches for relevant
information wherever it can find it (this is part of
its knowledge), changes its own state accordingly,
and then propagates the result to a specific agent,
which is also determined by the processing of
the encapsulated knowledge. It may be another
Cognition Agent or an Execution one (see next).
In the first case, the knowledge processing is