Simplifications in Temporal Persistence: An Approach to the Intractable Domain Theory Problem in Explanation-Based Learning

摘要

In real-world domains, large amounts of knowledge are needed to adequately describe world behavior. With a complex domain theory, complete reasoning becomes a computationally intractable task. This is particularly true of knowledge-intensive learning techniques such as Explanation-Based Learning. This thesis describes an approach to problem-solving and learning designed to deal with computationally ill-behaved domains and first implementation of this approach. In this approach, the system learns by constructing and generalizing causal explanations of observed plans. By using simplifications when necessary, the system avoids the intractability of complete reasoning. However, this introduces the possibility of learning imperfect plans. In order to deal with this contigency the system monitors execution of these plans. When a discrepancy between the expected world state and the actual world state is detected, the system constructs an explanation for the discrepancy and uses this explanation to refine the faulty simplification. By using the real world to focus attention on incorrect simplifications, the system avoids the intractability of complete reasoning. (Author)