This research project advances artificial-intelligence research in the areas of planning and reinforcement learning. The primary goal of the research is to create theoretically justified planning algorithms with wide applicability to tasks including robot navigation and control, medical decision making, flexible manufacturing, communications-network monitoring, and space mission scheduling. The research takes a two-pronged approach to develop algorithms for solving large-scale domains derived from practical, real-world problems: careful formal analysis to aid in the construction and identification of justifiable algorithms, and the development and intensive empirical validation of algorithms. This strategy encourages the creation of algorithms that will be useful in solving practical tasks while avoiding ``over fitting'' (i.e., losing generality by tailoring algorithms to attributes of specific domains). New planning algorithms developed in the course of this research combine insights from the areas of planning and reinforcement learning to make it possible to solve larger and more difficult problems than could be addressed previously. The research is creating methods for approximately solving large control problems that could be used in engineering applications ranging from elevator design to the adaptive control of computer systems; this will help create systems that aremore efficient, safe, and reliable.
