Automated Construction of Graphical Decision Models from Multiple Knowledge Sources

Li Wenting

Abstract

Complicated probabilistic decision models can be illustrated graphically using Bayesian Network (BN) and Influence Diagram (ID). As a powerful tool, influence diagram is able to effectively represent and solve decision problems. However, in this dynamic world, multiple knowledge sources are often available. It is necessary to automatically construct graphical decision models that combine various sources of information together. With comprehensive understanding of the situation, it will facilitate the decision maker to make a sound decision.

Existing approaches only focus on Bayesian Network and single agent decision problem. Based on our literature survey, we propose an algorithm to qualitatively and graphically integrate single-agent influence diagrams. During this process, we identify the challenges in merging influence diagrams for both single-agent and multi-agent IDs. The basic idea of the methodology is to avoid cycles in combined IDs through the use of order value. Then we use a real-world application to better illustrate how the method works.

In the second part of the report, applicability of the method is illustrated through a case study in medicine. It demonstrates how combined influence diagram allows patients to make better-informed and sound decisions. In the last part of the research, the method is extended to multi-agent influence diagram that associates with game theory in many cases. Additional challenge and precautions in combining MAIDs are elaborated.