On Thursday, September 19, we will have a LIRa session with Olivier Cailloux.
Everyone is cordially invited!
Speaker: Olivier Cailloux (University of Amsterdam)
Title: Preference modeling in multiple criteria decision aiding
Date and Time: Thursday, September 19, 2013, 15:30-17:30
Venue: Science Park 107, Room F1.15
Abstract:
In this presentation, I will describe what is called preference modeling in the multiple criteria decision aiding (MCDA) research field, with a special focus on validation. (I will not assume any prior knowledge of the field.)
MCDA is interested in situations where multiple points of view (criteria) are used to rate objects (alternatives). We want to come up with a global quality evaluation for each object, for example, a quality label for each alternative. In a setting where a committee wants to choose which research projects to fund, the alternatives would be the research projects, the criteria may consist in the interest of the project goal, the experience of the proposing team and the adequacy of the proposed research methodology with respect to the targeted goal.
Preference modeling is the activity which aims at building a model of the relation between the performances of evaluated alternatives on the criteria and the desired aggregated evaluation. This models the preferences of a given individual (decision maker). Several methods of preference modeling have been proposed in MCDA, I will present the most classical one.
Two problems make validation and justification of such methods difficult. First, they are not descriptive: they do not attempt to reflect faithfully how a decision maker normally behaves, but rather try to help him think about his preferences. Second, numerous experiments have shown that preferences are not precisely determined in the decision maker’s head prior to the start of the modeling process. This is referred to as the lability of the preferences. These two points make epistemology related to preference models very particular, compared to physical models of a natural phenomenon.
I will show how lability may be taken into account when modeling preferences, and that doing this can bring practical benefits.
I will finally come back to these epistemological difficulties and propose a way to falsify methods of preference modeling.