Knowledge Acquisition for Visually Oriented Planning

Many planning tasks can be represented using mental models in which an expert manipulates objects from one state to another. This suggests a highly graphical knowledge acquisition tool where the expert is able to capture the visual intuition of the problem solving to facilitate the encoding of a domain knowledge base. By exploring knowledge acquisition for object manipulation domains, insight will be gained in how knowledge is acquired and represented for such visually oriented tasks. This thesis addresses graphical knowledge acquisition in visually oriented domains in the context of Prodigy, a general problem solving and planning architecture. The prototype system, called APPRENTICE, demonstrates the main ideas in the thesis. This system establishes the feasibility of a graphical interface to enhance the ability of the expert to develop factual domain knowledge objects, relations, and operators in multiple domains. The system has been evaluated in four studies. In the first study, 32 AI students used the system to build their own domains. In the second study, domains developed by different types of users were completed faster using graphical input than using textual input. The third study was a learning study in which a subject developed several domains in APPRENTICE. APPRENTICE and its techniques proved to be usable, flexible, and extendable.