Success in applying AI-based planning systems to real domains requires sophisticated methods of knowledge acquisition. Both interactive and automated methods are required: interactive methods to aid the user in entering planning knowledge; and automated methods to verify the interactively developed knowledge and extract new knowledge from a variety of sources, including simulators, on-line databases, training exercises, and actual situations.
We describe two knowledge development tools for a crisis action planning system. The first tool is a graphical operator editor that enables users to develop new planning operators and revise existing ones. The operator editor provides type- and consistency checking and incorporates methods for ensuring the syntactic validity of the new knowledge. The second tool is a largely automated inductive learning system based on the PAGODA learning model [desJardins 1992] that learns from simulator feedback and from choices made by the user during planning.
[desJardins 1992] Marie desJardins. PAGODA: A Model for Autonomous Learning in Probabilistic Domains. UC Berkeley Ph.D. thesis, 1992. Available as UCB CS Dept. Technical Report 92/678.
Click to get a postscript version of this paper.