10:15 〜 10:30
[HSC05-06] Recent advances in value of information analysis and application to geological CO2 storage
キーワード：二酸化炭素地中貯留、不確実性下の意思決定、情報の価値、マルチレベルモンテカルロ法、ベイズ推定
The value of information (VOI) analysis provides an objective means for evaluating the cost-effectiveness of gaining additional information for decision making under uncertainty, and some attempts have been made in the literature to apply the VOI analysis to the context of geological CO2 storage. In recent years, mostly motivated by applications to health economic evaluations, there have been significant progresses in developing efficient computational methods to estimate the expected value of information, which have been shown to work well even for complicated nonlinear models consisting of many input random variables. Such developments can lower the barrier to use of the VOI analysis in practical operations of geological CO2 storage.
In this talk we start from an introduction to the VOI analysis; we define three classes of information (perfect information, partial perfect information, and sample information) and derive the expected value which each class of information brings to decision making under uncertainty. Here we confirm that the expected values of partial perfect information and sample information (abbreviated as EVPPI and EVSI, respectively) are defined as nested expectations mathematically, which make it difficult to construct an efficient estimator for them. Then we move on to an overview of recent developments in the methodology for the EVPPI and EVSI estimation. Finally, through a simple two-alternative decision problem on which site to store CO2, we demonstrate the following items: how three classes of information are related each other, how the expected value of information can be estimated efficiently in comparison with the classical method, and which points we should be cautious about in practical use.
In this talk we start from an introduction to the VOI analysis; we define three classes of information (perfect information, partial perfect information, and sample information) and derive the expected value which each class of information brings to decision making under uncertainty. Here we confirm that the expected values of partial perfect information and sample information (abbreviated as EVPPI and EVSI, respectively) are defined as nested expectations mathematically, which make it difficult to construct an efficient estimator for them. Then we move on to an overview of recent developments in the methodology for the EVPPI and EVSI estimation. Finally, through a simple two-alternative decision problem on which site to store CO2, we demonstrate the following items: how three classes of information are related each other, how the expected value of information can be estimated efficiently in comparison with the classical method, and which points we should be cautious about in practical use.