Belief Desire Intention
BDI (Rao and Georgeff (1991)) is currently the more interesting approach in
the Multi Agent System framework; it is a theoretical model that describes
the functioning of a goal oriented system, i.e. a system that pro-actively
acts in the world in order to fulfil its desires. It has been implementes in
slightly different forms in many architectures. A BDI agent has a set of
beliefs (mainly about the environment); a set of desires and goals, that are
its own motivations; a repository of plans that are used in order to fulfil
the goals. All the plans for the same goals are in intrinsic concurrence,
and the choice of which one the agent will follow depends from contextual
conditions as well as some implicit anticipations of the future states of
the world. It exists a subgoal mechanism, because each plan can have some
intermediate goals that become active in parallel. The intention of a BDI
agent is the currently activated goal; the agent is committed to this goal
(that is slightly biased in order to avoid oscillations). Goals can be
activated both pro-actively, in order to fulfil a desire, or reactively, in
order to react to a perception.
In the classic BDI model, a centralised interpret chooses one among all the
eligible goals and selects a plan to fulfil it. At the contrary, AKIRA is
well suited for parallel and concurrent computation, so many goals and plans
can be fulfilled in parallel: the choice is no more a top-down, off-line and
static decision, but it is the result of the concurrent computation and the
activity of the agent in the world. This allows to take into account all the
multiple parallel constraints of the situation (contextual, motivational,
reactive), to model many well known psychological evidence (priming and
biasing effects, blending of different situations) and to have a faster
computation (without an off-line phase). (An example: we do not want to
fulfil separately the goals “go to x” and “be quick”; what we want is “go
quickly to x”).
More, AKIRA adds some powerful capabilities to the BDI model, namely the
ability to deal with expectations, anticipation, forecasts (mental entities
that allow the agent to “be a step in the future”); the ability to form
models of the environment and of other agents’ minds and use them in order
to understand their intentions and anticipate them (using FCMs in order to
“run the model”); a quantitative logic (fuzzy logic) for beliefs and goals
that allows very fast and precise approximate reasoning.
The parallel calculus of AKIRA together with those advanced capabilities
allow to extend the BDI model in order to build intelligent, goal-oriented
agents that are able able to act in real time within a dynamic environment.