Date of Award
Doctor of Philosophy (PhD)
AI Planning is an important research field. Heuristic search is the most commonly used method in solving planning problems. Despite recent advances in improving the quality of heuristics and devising better search strategies, the high computational cost of heuristic search remains a barrier that severely limits its application to real world problems. In this dissertation, we propose theories, algorithms and systems to accelerate heuristic search for AI planning.
We make four major contributions in this dissertation. First, we propose a state-space reduction method called Stratified Planning to accelerate heuristic search. Stratified Planning can be combined with any heuristic search to prune redundant paths in state space, without sacrificing the optimality and completeness of search algorithms.
Second, we propose a general theory for partial order reduction in planning. The proposed theory unifies previous reduction algorithms for planning, and ushers in new partial order reduction algorithms that can further accelerate heuristic search by pruning more nodes in state space than previously proposed algorithms.
Third, we study the local structure of state space and propose using random walks to accelerate plateau exploration for heuristic search. We also implement two state-of-the-art planners that perform competitively in the Seventh International Planning Competition.
Last, we utilize cloud computing to further accelerate search for planning. We propose a portfolio stochastic search algorithm that takes advantage of the cloud. We also implement a cloud-based planning system to which users can submit planning tasks and make full use of the computational resources provided by the cloud.
We push the state of the art in AI planning by developing theories and algorithms that can accelerate heuristic search for planning. We implement state-of-the-art planning systems that have strong speed and quality performance.
Lee Benham, Chenyang Lu, Yinjie Tang