Combining Associational and Causal Reasoning to Solve Interpretation and Planning Problems
Abstract
Efficiency and robustness are two desirable, but often conflicting, characteristics of problem solvers. This report presents an approach, called Generate, Test and Debug (GTD), that integrates associational and casual reasoning techniques to efficiently solve a wide class of interpretation and planning problems. The GTD paradigm generates an initial hypothesis using rules that associate features of the problem with events that can cause them. If the tester detects bugs in the hypothesis, it is debugged until a correct solution is produced. The debugger employs three domain-independent causal reasoning techniques: 1) it analyzes causal explanations produced by the tester to locate the assumptions underlying bugs in the hypothesis, 2) it regresses values back through the explanations to indicate the direction in which to change the assumptions, and 3) it replaces faulty assumptions based on a model of causality that explicitly represents time, persistence, and the effects of events. Our analysis of the GTD paradigm indicates that the generator's efficiency stems from its use of nearly independent associational rules. This enables it to construct hypotheses that are correct, or nearly so, without having to check for potential interactions between events. In contrast, the debugger achieves robustness by using causal models of how the world works to determine how events interact in the achievement of goal. We characterize domains for which GTD may be useful based on this analysis of the strengths and weakness of the two reasoning techniques.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 1988
- Accession Number
- ADA202184
Entities
People
- Reid G. Simmons
Organizations
- Massachusetts Institute of Technology