This project has the goal of developing planner-generalization techniques that can be used to modify AI planning algorithms to remove some of the restrictive assumptions found in classical approaches to AI planning, such as: perfect knowledge about actions and objects and history of the planning environment, static planning environment, instantaneous actions, discrete time, determinism, black-and-white solution criteria. A starting point for the project will be the PI's preliminary results on a method to "non-determinize" forward-chaining planners. Using both theoretical and empirical techniques, the project will explore ways to systematically generalize planning algorithms to deal with nondeterminism, actions with probabilistic effects, and temporal issues. Results of the project will provide theoretical and experimental underpinnings necessary to enable AI planning to better address the needs of real-world planning applications such as manufacturing planning and ship movement planning. It is intended that implementations of algorithms developed in this project will be made available as open-source software.
