Explanation Structures for Complex Qualitative Behavior Graphs*

Substantial progress has been made in building and simulating qualitative models . However, the result s of qualitative simulation may be difficult to under stand and programs for explaining them have been fe w and have had a number of limitations . This paper describes a system, Expound, which provides causal natural language explanations of events in provabl y "faithful" abstractions of behaviors produced by qualitative simulation . Unlike previous explanation systems, Expound explains events rather than individua l states, producing richer and more comprehensive explanations . It also does not require numeric simulation, handles behavior branching, and explains change s other than changes in active processes or in the model . The main steps in building explanations and basic data structures produced are discussed .