Planning with Loops: Some New Results

In AI planning, there has been an increasing interest in solving a class of problems, rather than individual problems, with a generalized notion of “plan.” One such generalization is plans with loops, i.e. programlike plans, whose execution on a specific problem in the class results in a sequential plan. Levesque’s KPLANNER falls into this paradigm: it generates robot programs, a type of loopy plan, that solves a range of parametrized planning problems. In this paper, we build on that work, and propose another plan representation along with a novel planning algorithm. We show that the new plan representation is more general than robot programs, and the new planner more efficient than KPLANNER.