Results from UQ ultimately serve as decision support. Hence it is relevant to set the UQ analysis in the context of a formal decision analysis, to ensure the optimal choice of UQ methods and interpretation of results. This minisymposium focuses on such a combination of UQ with formal decision analysis methods. On the one hand, this includes the selection of metrics for UQ analysis based on decision-theoretic considerations. Examples include the choice of appropriate objective functions and decision-theoretic sensitivity measures. On the other hand, the minisymposium considers the integration of UQ in artificial intelligence applications, and more specifically sequential decision making algorithms, which are of increasing relevance in many fields of application.
10:30
- CANCELED - Global sensitivity analysis and reducing input uncertainty
Jeremy Oakley | University of Sheffield | United Kingdom
Show details
Author:
Jeremy Oakley | University of Sheffield | United Kingdom
Global sensitivity analysis (GSA) is a well-established tool in uncertainty quantification, and can provide valuable insight into how individual uncertain model inputs (or groups of model inputs) contribute to model output uncertainty. However, the application of GSA in decision-making is less well developed: various sensitivity measures can be difficult to interpret quantitatively, and may not directly relate to actions available to a decision maker. For example, a GSA might assess the value of eliminating uncertainty about an input, whereas a decision-maker might only have the means to achieve a modest reduction in input uncertainty, through further data collection. We will discuss the use of GSA to support decision-making in two scenarios. Firstly, we will consider the scenario of delaying a decision to collect more data, in an attempt to reduce output uncertainty via reduced input uncertainty. Secondly, we will consider the use of formal (and labor-intensive) expert judgement in quantifying model input uncertainty. In particular, for a model with a large number of uncertain inputs, it may not be feasible to elicit expert uncertainty about all them. We will discuss how GSA can help prioritise the use of expert judgement.
11:00
Information Density in the Global Sensitivity Analysis of Computer Experiments
Emanuele Borgonovo | Bocconi University | Italy
Show details
Authors:
Emanuele Borgonovo | Bocconi University | Italy
Elmar Plischke | Clausthal University of Technology | Germany
Gordon B. Hazen | Northwestern University | United States
Value of information is an ideal sensitivity measure for decision support. However, this sensitivity measure is, at times, difficult to interpret. We discuss the use of the complementary notion of information density as a tool to augment the insights derived from value of information. We show that information density defined under expected utility increase is proportional to information density for the value of buying information and for the value of selling information. We then discuss its application in the context of global sensitivity analysis of computer experiments. We study the definition and prove relevant properties. We discuss the insights delivered to the analyst. We address its computation from given data and provide numerical results for well known simulators.
11:30
Mitigating epistemic bias in Bayesian statistical model combination, with examples from natural disaster research
Mauricio Monsalve | Research Center for Integrated Disaster Risk Management (CIGIDEN) | Chile
Show details
Author:
Mauricio Monsalve | Research Center for Integrated Disaster Risk Management (CIGIDEN) | Chile
Epistemic uncertainty refers to the inadequacy of a model at describing or representing a given situation, which occurs because of the lack of knowledge that the modeler has about the situation being modeled. One way to reduce this type of uncertainty is to propose a collection of candidate models and then combine them statistically. Bayesian stacking or Bayesian model averaging does this by using a weighting mechanism that is consistent with probability theory. This helps translate some of the epistemic uncertainty into aleatoric uncertainty, which currently is much easier to address. However, this method is not perfect and might introduce bias. This talk will illustrate several ways in which this bias can be unintentionally introduced in the modeling and how these can be mitigated. Examples drawn from natural disaster research will be used to illustrate these points.
12:00
Communicating uncertainty for decision support
Thais Fonseca | University of Warwick | United Kingdom
Show details
Authors:
Martine J. Barons | University of Warwick | United Kingdom
Thais Fonseca | University of Warwick | United Kingdom
Communicating uncertain information to decision-makers in such a way that they can integrate the information into their decision-making is vital; we have shown that ignoring uncertainty can lead to recommending a suboptimal course of action. In order to be able to communicate the uncertainty, it first has to be identified. Since there are many kinds of uncertainty in systems that are other than profoundly simple, this in itself is no straightforward takes. Next, the sources of uncertainty must be combined in a transparent, defensible and robust manner so that the combined uncertainty can be quantified. The effects of this combined uncertainty on the decision question under consideration is then calculated. Finally, the range of possible decisions, their relative ranking against success criteria identified by the decision-maker along with the uncertainty must be communicated to the decision-maker. In anthropogenic decision-making the communication phase is of particular importance, because the human decision-maker may have a different skill set form those who generated the decision ranking. This, therefore, requires strongly interdisciplinary working and fusion of distinct vocabularies in order to be effective. We will be drawing on recent examples taken from nuclear emergency, food security, energy, digital preservation and security applications.