SUMMARY
Over the last few years, the agent/multi-agent system paradigm is one
of the most relevant computing models. It has produced a lot of specic research
and applications. The objective of this computational model is to build
systems that will face situations showing some human being features, such
as intelligence, reactivity, pro-activity, . . . .
There are a lot of different kinds of agents. The work that has been done
is focused in agents working in an environment with critical temporal restrictions.
In these environments, there are some problems to be solved by the
agent before some time has passed, otherwise the consequences may be catastrophic.
It is fundamental for an agent like this to get an optimal usage of the
processor time. In fact, this is the most important resource in these systems.
That is the reason why it is relevant to get the agent to dedicate its processor
time to what is really necessary in each situation. The agent being able
to reason about its own reasoning process is fundamental to get this adaptation.
This is called meta-reasoning, and when used, it has to be aware that this
process also consumes processor time.
So, the purpose of this work is the study of the necessary capabilities to
incorporate the meta-reasoning ability to an agent with critical temporal restrictions.
Moreover, to incorporate it to an specic agent architecture, the ARTIS
agent architecture. The conclusion of this study was that the capabilities
needed to incorporate the ability of meta-reasoning to an agent with critical
temporal restrictions are: to detect signicant situations, to adapt its behaviour,
to adapt its reasoning process taking into account its temporal restrictions,
and learning.
To get an agent able to adapt, it must be able to detect the moment, the
situation, when this adaption must be carried out. This is the reason why the
detection of signicant situations is an essential capability, because without
it the adaption is meaningless.
Regarding the capability to adapt its behaviour, the agent will have several
behaviours, being able to activate, at each moment, the proper one to its
situation.
The following three concepts are dened as variable ones to allow the
agent to adapt its reasoning process taking into account its temporal restrictions:
Reactivity:
This concept is a basic feature of an agent. As it has been commented,
the new approach presented here is to consider this concept as variable,
dening a Reactivity Degree. This degree allows the agent to behave as
more reactive or more deliberative according to it.
Introspection:
Generic Psychology distinguishes between extrospection and introspection
as observation methods. The rst one is based on the sense-derived
contents analysis. The second one is based on the contents of the awareness
of the individual itself.
In a similar way, the present work raises this distinction when studying
the agent's deliberation process. This work distinguishes between
the deliberation motivated as an answer to environment changes and
the deliberation about the agent's own objectives (according to its proactivity).
In this way, the Introspection Degree is introduced as a value
between 0 and 1 indicating the split between the process time dedicated
to attend to environment changes and the process time dedicated to
deliberate, to the introspection. As the introspection degree's value gets
closer to 0, the agent becomes more sensitive to environment changes,
whilst as the introspection degree's value gets closer to 1, the agent becomes
more self-centered (more introspective).
Concentration:
This concept is dened, from Generic Psychology point of view, as the
capacity of focusing on some specic stimuli, avoiding the ones not related
to them, so that the attention is focalized and the reection is optimized.
On a similar way, Generic Psychology considers the attention
concept as the selective concentration of the mental activity that implies
an increase of the efciency over an specic task, as well as a perception
and cognition inhibition of the rest of activities.
So, the purpose of this capability to adapt the reasoning process is to apply
the preceding three concepts to the agent architecture. Thus, one way to
increase the efciency of the agent in the use of the process time is to be able
to focus such process; that is, to reduce the eld of its reasoning process to use
the available process time to what is relevant to each situation (to modify the
agent's concentration). In this way, an attention focus is dened as a subset of
the agent's believes that are signicant in an specic situation. As the concept
of concentration presents, these believes subsets will be used to optimize the
reasoning (and meta-reasoning) process by means of prioritizing the reasoning
about the active attention focus or focuses. The focalization mechanism
does not have to be reduced to only one focus, but there can exist several
attention focus at the same time, each one with a different attention degree
(according to the relative importance of this focus with the other active focuses).
Regarding to the last capability, learning, it must allow the agent to re-
ne, using its own experience, its behaviour parameters. The agent should be
even able to learn new ways of behaving to add to the ones specied by the
designer.
As it has been commented before, these capabilities forming the metareasoning
ability have been carried out in the ARTIS agent architecture. To
make it so, all the different models of such architecture (corresponding to
different abstraction levels) have been extended: formal model (abstract level
denition), user model (denition from the ARTIS agent designer's level) and
system model (denition from the directly executable level).