Case Study: Planning Courses of Action - Thinking as Computation - page 204

Sensing actions
Ordinary actions (the kind considered so far) change the state of the world. But sens-
ing actions, such as reading a thermometer or listening to a message, are not used
to change the world but to change what is known about the world. In many cases
of planning, information is available to achieve the goal, but that information only
becomes available via sensing actions after execution of the plan has begun.
Want to read more?
This chapter was about planning, a form of thinking that is fundamental to intelli-
gent behavior. The first research paper that can be clearly said to be about artificial
intelligence deals with this topic [47]. This seminal paper by John McCarthy (parts
of which date back to 1958) discusses reasoning about how to achieve goals using
actions, and introduces the monkey and bananas problem. It also presents for the
first time a representation based on situations and fluents.
Since the time of McCarthy's paper, researchers in the area of AI planning have
sought to expand the boundaries of what could be effectively handled by automated
planning. Part of this involved using less expressive representation formalisms that
appear to have better computational properties. But some of the research involved
using more expressive representations to capture a wider range of planning problems.
The state of the art of this research is well covered in a graduate-level textbook by
Ghallab, Nau, and Traverso [45]. The idea of using knowledge to filter the actions
considered at each step of the planning process is explored in a technical paper by
Bacchus and Kabanza [44].
While planning clearly plays an important role in our own intelligent behavior, it
seems implausible to imagine that we would constantly be planning from first prin-
ciples each time we need to act. So the question concerning what exactly a thinking
agent in a dynamic environment should be thinking about is what is studied in the
area of cognitive robotics [46]. Planning is involved, no doubt, but thinking about
action and change clearly needs to go beyond this, as discussed in the outstanding
(but technically demanding) book by Reiter [48].
The problem of how to think in general about what is and is not affected by an
action in a complex setting is called the frame problem . The solution outlined here (in
terms of successor-state clauses and the like) is due to Reiter [48].