RULE-BASED VERSUS STRUCTURE-BASED MODELS FOR EXPLAINING AND GENERATING EXPERT BEHAVIOR

Abstract

Flexible representations are required in order to understand and generate expert behavior. While production rules with quantifiers can encode experiential knowledge, they often have assumptions implicit in them, making them brittle in problem scenarios where these assumptions do not hold. Qualitative models achieve flexibility by representing the domain entities and their interrelationships explicitly. However, in problem domains where assumptions underlying such models change periodically, it is necessary to be able to synthesize and maintain qualitative models in response to the changing assumptions. In this paper, we argue for a representation that contains partial model components that are synthesized into qualitative models containing entities and relationships relevant to the domain. The model components can be replaced and rearranged in response to changes in the task environment. We have found this "model constructor" to be useful in synthesizing models that explain and generate expert behavior, and have explored its ability to support decision-making in the problem domain of business resource planning, where reasoning is based on models that evolve in response to changing external conditions or internal policies.