Classical AI planning programs have made very restrictive assumptions about their domains, particularly that the planning agent has complete control over, and complete information about, the world in which it operates. Relaxing these assumptions forces us to abandon the traditional planning paradigm of "build a plan that will provably achieve the goal, then execute it," since such a proof will almost never be achievable. We instead view the planning process as one of choosing a reasonable course of action based on what is currently known about the world, but being willing to adapt to new information gathered during execution. Planning, in other words, is a process of decision making under uncertainty, but the decisions must continually be re-evaluated.
We are pursuing this view of planning in a number of areas. One is the building of a system that concurrently plans and executes, thus has to coordinate the two processes. Another is the investigation of applying techniques from decision analysis to the planning problem. We are developing extensions to classical planning algorithms that manipulate probabilistic models of the world, that gather and exploit information about the world at run time, and that use rich utility models allowing the planner to reason about deadlines, partial goal satisfaction, and balance the benefit of achieving a goal against the cost of doing so.
We are applying these planners to a variety of problem domains including medical decision making and process planning.
Principal Investigator: Steve Hanks