Probabilistic semantics for qualitative influences
Michael P. Wellman · National Conference on Artificial Intelligence · 1987
What's in an influence link? To answer this foundational question, I propose a semantics for qualitative influences: a positively influences b if and only if the posterior distribution for b given a increases with a in the sense of first-order stochastic dominance. By requiring that this condition hold in all contexts, we gain the ability to perform inference across chains of qualitative influences. Under sets of basic desiderata, the proposed definition is necessary as well as sufficient for this desirable computational property.