Minimal elicitation for decision problems described by bounded probability assessments

How to efficiently elicit relevant information from an expert when having to decide in an uncertain environment? In this paper, we study this problem when uncertainty is modeled by coherent upper previsions, a very general uncertainty model that includes many others as special cases. We propose an algorithmic elicitation protocol that explicitly takes into account the decision problem. The protocol relies on the range of all possible coherent bounds from the expert, and we establish some new results on how to compute these coherent ranges. Experiments show that our approach is efficient, in particular when queries concern pairwise differences between alternatives.