Value of Evidence on Influence Diagrams
This work addresses decision-making under uncertainty for normative expert systems, presenting incremental extensions to existing influence diagram methods.
The paper tackles the problem of measuring the value of experimentation in decision analysis by introducing evidence propagation operations and a concept of value of evidence on influence diagrams, enabling direct computation of outcome sensitivity, value of perfect information, and value of control.
In this paper, we introduce evidence propagation operations on influence diagrams and a concept of value of evidence, which measures the value of experimentation. Evidence propagation operations are critical for the computation of the value of evidence, general update and inference operations in normative expert systems which are based on the influence diagram (generalized Bayesian network) paradigm. The value of evidence allows us to compute directly an outcome sensitivity, a value of perfect information and a value of control which are used in decision analysis (the science of decision making under uncertainty). More specifically, the outcome sensitivity is the maximum difference among the values of evidence, the value of perfect information is the expected value of the values of evidence, and the value of control is the optimal value of the values of evidence. We also discuss an implementation and a relative computational efficiency issues related to the value of evidence and the value of perfect information.