Aperçu de cette version: 3
Uncertainty and risk are pervasive issues in decision making. With a wide range of causes and types of uncertainty, there are correspondingly many approaches to their treatment in decision analysis. Some are tackled through discussion and creativity techniques to help decision makers set the boundaries of their problem; others, are tackled through modelling techniques, e.g. probability, to reflect the randomness in the external world; yet others are approached through the use of sensitivity and robustness studies to explore the possible consequences of lack of precision in estimates and judgements. Where there is a need for much more work is in the interfaces between the approaches for dealing with different types of uncertainty. Thus the issues that are going to be addressed relate to the bringing together of specific expertises in aspects of handling uncertainty within decision modelling to build a more comprehensive overview and integrated methodology to tackle all the various uncertainties in a decision problem. Specific subjects which are going to be considered include:
- Formalisms for modelling uncertainty, with particular emphasis in decision making situations. Besides probability theory, several other formalisms have been developed in order to take into account situations of lack of knowledge or of ambiguous information (possibility theory, belief functions etc.). The issue here is how to understand, within a decision situation, which formalism could be more appropriate, and how to combine
them within a decision aiding process (how to discriminate between solutions which are highly preferred but assume unlikely events and solutions which are not preferred but robust with respect to uncertainty?). Qualitative Decision Theory results need to be enhanced here and combined with findings in preference modelling and the use of appropriate languages for this purpose.
- Sensitivity and Robustness. While it is known how to conduct specific sensitivity analyses on a class of models, it is still required to have more comprehensive methodologies for developing complete sensitivity analyses, especially in the case of large complex decision models. Moreover, it is necessary to further investigate the specific class of robustness issues in decision making: obtaining results which will be useful independently from variations which may occur in data and/or scenarios.
- Sequential Decision Making. With the increasing amount of data faced by decision makers and the increasing speed with which their decisions need to be made, it is often the case that decisions have to be taken online before having access to all of the relevant data. Sequential decisions are also important in uncertain domains in which decisions impact the environment and therefore the context for future decisions. It is important that robust policies
-policies with guaranteed performance when the nominal model deviates from the unknown true model- can be obtained. Robustness is an increasingly important issue in decision making, a major challenge being to obtain a robust policy for dynamic decision models.