Analyzing, Learning, and Shaping Planning Search Spaces

Introduction The complexity of a task faced by a planner depends strongly on the search space. The importance of the planner’s search space is reflected in the two main classes of classical AI planners: state-space planners, which explore the search space by considering totally ordered sequences of directly neighboring states of the world, and plan-space planners, which explore the search space by considering incomplete plans consisting of partially ordered actions. The search space also depends on the structure of subgoals (their interactions) and on the representation of knowledge used by the planner (e.g., the level of abstraction at which the planning is done or the expressiveness of the description language). Taking these factors into account can lead to finer-grained classification of planning domains and search spaces. Knowing the properties of a search space is useful when a human designer creates a new planner or decides which of the existing planners to use. However, knowledge of the properties of a search space may also be useful to the planner itself. In my research, I plan to explore which features of the search spaces can be automatically determined or learned by planners (during repeated executions), and how such information may be used by a planner to improve the search process (e.g., by choosing an appropriate planning technique, abstraction level, or representation). This research is still in the idea stage: I am in the process of developing my dissertation proposal, and plan to take the preliminary exam (proposal defense) in July, after the conference. Attending the doctoral consortium will give me the ideal opportunity to receive early feedback that can help guide my dissertation research. I also look forward to the opportunity to attend the conference in order to get a better sense of the current state of the field, and to meet other graduate students and senior researchers.