The flexible use of abstract knowledge in planning
Eric K. Jones · 1992
People are flexible and robust problem solvers: faced with problems that go beyond their past experience in a domain, they often find ways to cope. When domain-specific knowledge is inadequate, more general knowledge can often be brought to bear. In particular, problems that at first appears novel can often be redescribed in domain-unspecific terms and thereby revealed to be familiar at a more abstract level. Existing domain-specific knowledge may then be directly applicable to the redescribed problem. We call this approach to problem solving problem reformulation. This dissertation presents a case-based account of problem reformulation. We propose an architecture in which a case-based reasoner acts as an intelligent assistant to an automated planner. The case-based reasoner reformulates novel problems using abstract knowledge represented as culturally-shared cases. Whenever the planner lacks appropriate domain-specific knowledge, the case-based reasoner adapts an appropriate culturally-shared case to resolve the planner's difficulty. We focus on the adaptation phase of case-based reasoning. Adaptation takes as input a planner difficulty and a culturally-shared case that may resolve it, and proceeds in three stages. First, an abstract model of the planning process is used to transform the case into a planner data structure that may resolve the planner's difficulty. Second, the planner's problematic situation or initial state is redescribed to fit the abstract vocabulary of the proposed solution. Knowledge structures called viewing schemas are used to control redescription inference. Finally, the candidate solution is elaborated with relevant domain-specific knowledge that was inaccessible prior to redescription. The approach is implemented in scBRAINSTORMER, a planner that uses culturally-shared cases to help it plan. The system operates in the domain of political and military policy as it relates to terrorism. scBRAINSTORMER makes contributions in the areas of case-based reasoning, knowledge representation, automated knowledge acquisition, and model-based reasoning.