|
Planning and Plan Recognition
Plan recognition is a term used to refer
to the task of inferring the plan or plans of an intelligent
agent from observations of the agent's actions or the
effects of those actions. * It involves a mapping from a temporal
sequence of actions and their effects to an organization
of these actions and their effects into some plan representation
that identifies the goal of the plan together with the relation
between the components of the plan. In our investigation of planning
we have encountered a variety of representations of a planthe
most austere is simply a sequence of actions and the most involved
was the plan representation used in PLANX10. Obviously, the nature
of a plan recognition process will depend on assumptions concerning
the nature of the plan representation that must be inferred from
the observed actions.
There
are a variety of additional considerations that affect the nature
of a particular plan recognition process. As with planning, the
nature of plan recognition will depend on the reason for attempting
to infer the planner's intents as well as the relation between
planner and plan recognizer. For example, in the section on learning
we presented an abstraction of the relation between teacher and
learner. Here, if both actors follow their respective script
then the plans of each will be transparent to the other. This
is typical of cooperative interactions and, as we saw from the
work on learning the concepts of baseball, this transparency
of intent is also typical of much of the interaction that occurs
in competitive games. For example, there is typically little
doubt that a batter is attempting to get on base and that the
opposing team is trying to block that attempt. But there are
also situations where the actor is intentionally trying to deceive
observers of his actions. For example, con artists make their
living in this manner.
Yet
another consideration may affect the nature of the plan recognition
process. Most plans that we carry out are executed in the context
of other agents. Sometimes our plan depends on the assumption
that the actions of other agents will have no effect on out plan.
Other times, we rely on other agents forgoing certain actions
to accommodate our own plans, or to engage in certain actions
to further our own plans. In these contexts, where the plan recognizer
is also a planner; the plan recognition process may be a very
self-centered one - that is, he interest is only in understanding
the plans of others to the extent that this must be done to further
ones own actions.
Despite
the importance and ubiquity of plan recognition there has not
been a great deal of computational research on this problem.
It has been most consistently approached from the point of view
of person-machine interaction. Here a task oriented dialogue
between person and machine is often viewed as a context where
the machine must attempt to understand the user's actions and
responses in order to respond appropriately to the user's intent.
The
earliest computational work on plan recognition was done by Schmidt
and his colleagues. In their AI Journal paper available in this
section you will find that they chose to focus their investigation
on a situation where the actor:
- only carried out physical actions;
- neither attempted to hide nor
attempted to communicate intent;
- engaged in only one plan at
a time;
and the plan recognizer adopted
the role of a passive observer whose only intent was to infer
the plan that could explain the actions of the planner.
|
