Information Technology Reference
In-Depth Information
While these traits were introduced to allow the computational tractability of the model,
they are now recognized as being the most distinctive characteristics of BDI systems in
general for their ability to support rational, resource-bounded reasoning in dynamic and
uncertain domains.
Beliefs, for example, are essential since an agent has limited sensory ability and
needs to build up its knowledge of the world over time. In this sense, beliefs - usually
represented with first order logic predicates - serve as a cache with which the agent can
remember past events or other important information that could be costly to re-compute
from raw perceptual data or inferred logically.
The distinctions between goals and plans constitute an important source of modular-
ity that contributes to the agent's ability to cope with contingencies. The fundamental
observation is that goals, as compared to plans, are more stable in any application do-
main and multiple plans can be used/attempted to achieve the same goals. This also
allows examining the application domain in terms of what needs to be achieved, rather
than the types of behaviour that will lead to achieving it.
PRS implements a computationally tractable BDI model with the following simpli-
fying assumptions:
-
The system explicitly represents beliefs about the current state of the world as a
ground set of literals with no disjunctions or implications (as in STRIPS).
-
The system represents the information about the means of achieving certain future
world states and the options available to the agent as pre-compiled plans.
Each plan in the plan library can be described in the form of Event Condition Action
(ECA) rules
e
:
P
where
P
is the body of the plan,
e
is an event that triggers the
plan (the plan's post-conditions),
Ψ
←
is the context for which the plan can be applied
(which corresponds to the preconditions of the plan).
The body of each plan is a procedural description containing a particular sequence
of actions and tests that may be performed to achieve the plan's post-condition. Plans
are activated in response to the posting of new goals or upon reception of events. This
process consists of finding the plans whose invocation unifies with the active goals
or events and whose context unifies with the state contained in the agent's belief set.
Plans may also post new goal events, leading to the characteristic AND/OR, goal/plan
execution graphs. For a plan to succeed all the subgoals and actions of the plan must be
successful (AND); for a subgoal to succeed one of the plans to achieve it must succeed
(OR). When a plan step (an action or sub-goal) fails for some reason, this causes the
plan to fail, and an alternative applicable plan for its parent goal is tried. If there is no
alternative applicable plan, the parent goal fails, cascading the failure and search for
alternative plans one level up the goal-plan tree.
In PRS-like BDI systems, desires and goals are represented only in the transient form
of goal events (posted by the application), while the intentions to purse them is stored
implicitly in the stack of plans they triggered. This poses an obstacle to the effective
decoupling between plans and goals, forcing, for instance, the agent to drop goals for
which no feasible plan can be attempted at the time the goal is posted. Such an issue
is addressed in modern agent systems, such as Jadex, by incorporating explicit and
declarative goal representations into the agent interpreter in order to ease the definition
of goal deliberation strategies [3].
Ψ
