UC San Diego

Automated planning is investigated for very large, uncertain, partially observable environments. Experimental analysis is conducted on two widely different applications : the classic RISK game and a simulation of an Urban Search and Rescue Operation (US&RO). Many aspects of automated planning are explored simultaneously to limit assumptions, problem domain requirements, and subject matter expertise. Challenges addressed are in search capability, implementation strategy (scalability, modularity and extensibility), and machine learning for continual improvement. A general architecture is designed for future applications, and includes a hierarchical framework, a well-defined set of functions, and a nine- step implementation strategy. Metrics are designed and implemented at seven levels of the hierarchy to address scalability, modularity and extensibility. Experimental results are analyzed for efficiency and effectiveness at three levels of simultaneous machine learning: plan generation (plan search and selection), planning cycle operation (incorporates plan-execution and -assessment as an iterative process), and planners' utility (measure of planners ability to accomplish a mission or win a game)