A New Algorithm for Generative Planning

Existing generative planners have two properties that one would like to avoid if possible. First, they use a single mechanism to solve problems both of action selection and of action sequencing, thereby failing to exploit recent progress on scheduling and sat-issability algorithms. Second, the context in which a subgoal is solved is governed in part by the solutions to other subgoals, as opposed to plans for the subgoals being developed in isolation and then merged to yield a plan for the conjunction. We present a reformulation of the planning problem that appears to avoid these diiculties, describing an algorithm that solves subgoals in isolation and then appeals to a separate NP-complete scheduling test to determine whether the actions that have been selected can be combined in a useful way.