Solving Simple Planning Problems with More Inference and No Search

Many benchmark domains in AI planning including Blocks, Logistics, Gripper, Satellite, and others lack the interactions that characterize puzzles and can be solved non-optimally in low polynomial time. They are indeed easy problems for people, although as with many other problems in AI, not always easy for machines. In this paper, we address the question of whether simple problems such as these can be solved in a simple way, i.e., without search, by means of a domain-independent planner. We address this question empirically by extending the constraintbased planner CPT with additional domain-independent inference mechanisms. We show then for the first time that these and several other benchmark domains can be solved with no backtracks while performing only polynomial node operations. This is a remarkable finding in our view that suggests that the classes of problems that are solvable without search may be actually much broader than the classes that have been identified so far by work in Tractable Planning.