Footprints of conditionals

Abstract

Probabilistic conditionals are a powerful means of representing commonsense and expert knowledge. By viewing probabilistic conditionals as an institution, we obtain a formalization of probabilistic conditionals as a logical system. Using the framework of institutions, we phrase a general representation problem that is closely related to the selection of preferred models. The problem of discovering probabilistic conditionals from data can be seen as an instance of the inverse representation problem, thereby considering knowledge discovery as an operation inverse to inductive knowledge representation. These concepts are illustrated using the well-known probabilistic principle of maximum entropy for which we sketch an approach to solve the inverse representation problem. © Springer-Verlag Berlin Heidelberg 2005.