Describing and Valuing Interventions That Observe or Control Decision Situations

Authors:
David Matheson;James E. Matheson
Affiliations:
SmartOrg, Inc., 855 Oak Grove Avenue, Suite 202, Menlo Park, California 94025;SmartOrg, Inc., 855 Oak Grove Avenue, Suite 202, Menlo Park, California 94025, and Department of Management Science and Engineering, Stanford University, Stanford, California 94305-4026
Venue:
Decision Analysis
Year:
2005

Citing 4
Cited 5

Decision theory in expert systems and artificial intelligence

International Journal of Approximate Reasoning
Intelligent decision systems

Intelligent decision systems
Structuring conditional relationships in influence diagrams

Operations Research
Decision-theoretic foundations for causal reasoning

Journal of Artificial Intelligence Research

From the Editor...

Decision Analysis
From the Editors...

Decision Analysis
From the Editor---Decisions over Time (Exploding Offers or Purchase Regret), in Game Settings (Embedded Nash Bargaining or Adversarial Games), and in Influence Diagrams

Decision Analysis
From the Editors---Games and Decisions in Reliability and Risk

Decision Analysis
From the Editor---Optimal Betting, Reducing Unnecessary Mammography in Breast Cancer Diagnosis, Product Line Design, and Value of Information

Decision Analysis

Quantified Score

Hi-index	0.00

Visualization

Abstract

The value of information and value of control calculations have long been two separate parts of a decision analyst's efforts to extract as much insight as possible from a decision model. This paper unifies these concepts as interventions that modify the structure of the original problem, which have two key properties, purity and quality. Purity is an idealization that leads to Howard canonical form, clarifies the definition of control intervention, and allows us to extend and correct the calculation of the value of control. Quality is a characteristic that leads to generic models of imperfect intervention, which, because of their equivalence to any pure intervention, prevent misguided recommendations when the value of a perfect intervention is high but the value of a somewhat imperfect intervention is low. Quality is a number between 0 and 1 that normalizes and allows comparison of imperfect interventions between applications having very different value scales.